blob: e838feb6adc504605c85ff4f252d50a7198826c0 [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
Peter Zijlstra391e43d2011-11-15 17:14:39 +01003 * kernel/sched/core.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 *
Ingo Molnard1ccc662017-02-01 11:46:42 +01005 * Core kernel scheduler code and related syscalls
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Copyright (C) 1991-2002 Linus Torvalds
Linus Torvalds1da177e2005-04-16 15:20:36 -07008 */
Ingo Molnare66f6482022-02-23 08:17:15 +01009#include <linux/highmem.h>
10#include <linux/hrtimer_api.h>
11#include <linux/ktime_api.h>
12#include <linux/sched/signal.h>
13#include <linux/syscalls_api.h>
14#include <linux/debug_locks.h>
15#include <linux/prefetch.h>
16#include <linux/capability.h>
17#include <linux/pgtable_api.h>
18#include <linux/wait_bit.h>
19#include <linux/jiffies.h>
20#include <linux/spinlock_api.h>
21#include <linux/cpumask_api.h>
22#include <linux/lockdep_api.h>
23#include <linux/hardirq.h>
24#include <linux/softirq.h>
25#include <linux/refcount_api.h>
26#include <linux/topology.h>
27#include <linux/sched/clock.h>
28#include <linux/sched/cond_resched.h>
Thomas Gleixnerd664e392022-04-13 15:31:02 +020029#include <linux/sched/cputime.h>
Ingo Molnare66f6482022-02-23 08:17:15 +010030#include <linux/sched/debug.h>
Thomas Gleixnerd664e392022-04-13 15:31:02 +020031#include <linux/sched/hotplug.h>
32#include <linux/sched/init.h>
Ingo Molnare66f6482022-02-23 08:17:15 +010033#include <linux/sched/isolation.h>
34#include <linux/sched/loadavg.h>
35#include <linux/sched/mm.h>
36#include <linux/sched/nohz.h>
37#include <linux/sched/rseq_api.h>
38#include <linux/sched/rt.h>
Phil Auld9d246052020-06-29 15:23:03 -040039
Christoph Hellwig6a5850d2021-09-20 14:33:18 +020040#include <linux/blkdev.h>
Ingo Molnare66f6482022-02-23 08:17:15 +010041#include <linux/context_tracking.h>
42#include <linux/cpuset.h>
43#include <linux/delayacct.h>
44#include <linux/init_task.h>
45#include <linux/interrupt.h>
46#include <linux/ioprio.h>
47#include <linux/kallsyms.h>
Mark Rutland0ed557a2018-06-14 15:27:41 -070048#include <linux/kcov.h>
Ingo Molnare66f6482022-02-23 08:17:15 +010049#include <linux/kprobes.h>
50#include <linux/llist_api.h>
51#include <linux/mmu_context.h>
52#include <linux/mmzone.h>
53#include <linux/mutex_api.h>
54#include <linux/nmi.h>
55#include <linux/nospec.h>
56#include <linux/perf_event_api.h>
57#include <linux/profile.h>
58#include <linux/psi.h>
59#include <linux/rcuwait_api.h>
60#include <linux/sched/wake_q.h>
Sami Tolvanend08b9f02020-04-27 09:00:07 -070061#include <linux/scs.h>
Ingo Molnare66f6482022-02-23 08:17:15 +010062#include <linux/slab.h>
63#include <linux/syscalls.h>
64#include <linux/vtime.h>
65#include <linux/wait_api.h>
66#include <linux/workqueue_api.h>
67
68#ifdef CONFIG_PREEMPT_DYNAMIC
Ingo Molnara7b25532022-03-15 10:33:53 +010069# ifdef CONFIG_GENERIC_ENTRY
70# include <linux/entry-common.h>
71# endif
Ingo Molnare66f6482022-02-23 08:17:15 +010072#endif
73
74#include <uapi/linux/sched/types.h>
Mark Rutland0ed557a2018-06-14 15:27:41 -070075
Zhen Leibc1cca92022-08-04 10:34:20 +080076#include <asm/irq_regs.h>
David Howells96f951e2012-03-28 18:30:03 +010077#include <asm/switch_to.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070078#include <asm/tlb.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070079
Linus Torvalds1da177e2005-04-16 15:20:36 -070080#define CREATE_TRACE_POINTS
Ingo Molnare66f6482022-02-23 08:17:15 +010081#include <linux/sched/rseq_api.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070082#include <trace/events/sched.h>
83#undef CREATE_TRACE_POINTS
84
85#include "sched.h"
Ingo Molnarb9e9c6ca2022-02-13 08:19:43 +010086#include "stats.h"
87#include "autogroup.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070088
Ingo Molnare66f6482022-02-23 08:17:15 +010089#include "autogroup.h"
90#include "pelt.h"
91#include "smp.h"
92#include "stats.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070093
Tejun Heoea138442013-01-18 14:05:55 -080094#include "../workqueue_internal.h"
Jens Axboeed29b0b2022-05-23 17:05:03 -060095#include "../../io_uring/io-wq.h"
Thomas Gleixner29d5e042012-04-20 13:05:45 +000096#include "../smpboot.h"
Gregory Haskins6e0534f2008-05-12 21:21:01 +020097
Qais Yousefa056a5b2019-06-04 12:14:59 +010098/*
99 * Export tracepoints that act as a bare tracehook (ie: have no trace event
100 * associated with them) to allow external modules to probe them.
101 */
102EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_cfs_tp);
103EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_rt_tp);
104EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_dl_tp);
105EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_irq_tp);
106EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_se_tp);
Qais Yousef77cf1512021-10-28 12:50:05 +0100107EXPORT_TRACEPOINT_SYMBOL_GPL(pelt_thermal_tp);
Vincent Donnefort51cf18c2020-08-28 10:00:49 +0100108EXPORT_TRACEPOINT_SYMBOL_GPL(sched_cpu_capacity_tp);
Qais Yousefa056a5b2019-06-04 12:14:59 +0100109EXPORT_TRACEPOINT_SYMBOL_GPL(sched_overutilized_tp);
Vincent Donnefort4581bea2020-05-27 17:39:14 +0100110EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_cfs_tp);
111EXPORT_TRACEPOINT_SYMBOL_GPL(sched_util_est_se_tp);
Phil Auld9d246052020-06-29 15:23:03 -0400112EXPORT_TRACEPOINT_SYMBOL_GPL(sched_update_nr_running_tp);
Qais Yousefa056a5b2019-06-04 12:14:59 +0100113
Peter Zijlstra029632f2011-10-25 10:00:11 +0200114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Peter Zijlstradc61b1d2010-06-08 11:40:42 +0200115
Juri Lellia73f8632020-10-13 07:31:14 +0200116#ifdef CONFIG_SCHED_DEBUG
Ingo Molnare436d802007-07-19 21:28:35 +0200117/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200118 * Debugging: various feature bits
Patrick Bellasi765cc3a2017-11-08 18:41:01 +0000119 *
120 * If SCHED_DEBUG is disabled, each compilation unit has its own copy of
121 * sysctl_sched_features, defined in sched.h, to allow constants propagation
122 * at compile time and compiler optimization based on features default.
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200123 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200124#define SCHED_FEAT(name, enabled) \
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200125 (1UL << __SCHED_FEAT_##name) * enabled |
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200126const_debug unsigned int sysctl_sched_features =
Peter Zijlstra391e43d2011-11-15 17:14:39 +0100127#include "features.h"
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200128 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200129#undef SCHED_FEAT
Paul Turnerc006fac2021-04-16 14:29:36 -0700130
131/*
132 * Print a warning if need_resched is set for the given duration (if
133 * LATENCY_WARN is enabled).
134 *
135 * If sysctl_resched_latency_warn_once is set, only one warning will be shown
136 * per boot.
137 */
138__read_mostly int sysctl_resched_latency_warn_ms = 100;
139__read_mostly int sysctl_resched_latency_warn_once = 1;
140#endif /* CONFIG_SCHED_DEBUG */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200141
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200142/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100143 * Number of tasks to iterate in a single balance run.
144 * Limited because this is done with IRQs disabled.
145 */
Vincent Guittotc59862f2022-08-25 14:27:24 +0200146const_debug unsigned int sysctl_sched_nr_migrate = SCHED_NR_MIGRATE_BREAK;
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100147
Peter Zijlstra029632f2011-10-25 10:00:11 +0200148__read_mostly int scheduler_running;
Ingo Molnar6892b752008-02-13 14:02:36 +0100149
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500150#ifdef CONFIG_SCHED_CORE
151
152DEFINE_STATIC_KEY_FALSE(__sched_core_enabled);
153
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500154/* kernel prio, less is more */
155static inline int __task_prio(struct task_struct *p)
156{
157 if (p->sched_class == &stop_sched_class) /* trumps deadline */
158 return -2;
159
160 if (rt_prio(p->prio)) /* includes deadline */
161 return p->prio; /* [-1, 99] */
162
163 if (p->sched_class == &idle_sched_class)
164 return MAX_RT_PRIO + NICE_WIDTH; /* 140 */
165
166 return MAX_RT_PRIO + MAX_NICE; /* 120, squash fair */
167}
168
169/*
170 * l(a,b)
171 * le(a,b) := !l(b,a)
172 * g(a,b) := l(b,a)
173 * ge(a,b) := !l(a,b)
174 */
175
176/* real prio, less is less */
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -0500177static inline bool prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500178{
179
180 int pa = __task_prio(a), pb = __task_prio(b);
181
182 if (-pa < -pb)
183 return true;
184
185 if (-pb < -pa)
186 return false;
187
188 if (pa == -1) /* dl_prio() doesn't work because of stop_class above */
189 return !dl_time_before(a->dl.deadline, b->dl.deadline);
190
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -0500191 if (pa == MAX_RT_PRIO + MAX_NICE) /* fair */
192 return cfs_prio_less(a, b, in_fi);
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500193
194 return false;
195}
196
197static inline bool __sched_core_less(struct task_struct *a, struct task_struct *b)
198{
199 if (a->core_cookie < b->core_cookie)
200 return true;
201
202 if (a->core_cookie > b->core_cookie)
203 return false;
204
205 /* flip prio, so high prio is leftmost */
Josh Don4feee7d12021-10-18 13:34:28 -0700206 if (prio_less(b, a, !!task_rq(a)->core->core_forceidle_count))
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500207 return true;
208
209 return false;
210}
211
212#define __node_2_sc(node) rb_entry((node), struct task_struct, core_node)
213
214static inline bool rb_sched_core_less(struct rb_node *a, const struct rb_node *b)
215{
216 return __sched_core_less(__node_2_sc(a), __node_2_sc(b));
217}
218
219static inline int rb_sched_core_cmp(const void *key, const struct rb_node *node)
220{
221 const struct task_struct *p = __node_2_sc(node);
222 unsigned long cookie = (unsigned long)key;
223
224 if (cookie < p->core_cookie)
225 return -1;
226
227 if (cookie > p->core_cookie)
228 return 1;
229
230 return 0;
231}
232
Peter Zijlstra6e33cad2021-03-26 18:55:06 +0100233void sched_core_enqueue(struct rq *rq, struct task_struct *p)
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500234{
235 rq->core->core_task_seq++;
236
237 if (!p->core_cookie)
238 return;
239
240 rb_add(&p->core_node, &rq->core_tree, rb_sched_core_less);
241}
242
Josh Don4feee7d12021-10-18 13:34:28 -0700243void sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500244{
245 rq->core->core_task_seq++;
246
Josh Don4feee7d12021-10-18 13:34:28 -0700247 if (sched_core_enqueued(p)) {
248 rb_erase(&p->core_node, &rq->core_tree);
249 RB_CLEAR_NODE(&p->core_node);
250 }
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500251
Josh Don4feee7d12021-10-18 13:34:28 -0700252 /*
253 * Migrating the last task off the cpu, with the cpu in forced idle
254 * state. Reschedule to create an accounting edge for forced idle,
255 * and re-examine whether the core is still in forced idle state.
256 */
257 if (!(flags & DEQUEUE_SAVE) && rq->nr_running == 1 &&
258 rq->core->core_forceidle_count && rq->curr == rq->idle)
259 resched_curr(rq);
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500260}
261
262/*
263 * Find left-most (aka, highest priority) task matching @cookie.
264 */
265static struct task_struct *sched_core_find(struct rq *rq, unsigned long cookie)
266{
267 struct rb_node *node;
268
269 node = rb_find_first((void *)cookie, &rq->core_tree, rb_sched_core_cmp);
270 /*
271 * The idle task always matches any cookie!
272 */
273 if (!node)
274 return idle_sched_class.pick_task(rq);
275
276 return __node_2_sc(node);
277}
278
Peter Zijlstrad2dfa172020-11-17 18:19:43 -0500279static struct task_struct *sched_core_next(struct task_struct *p, unsigned long cookie)
280{
281 struct rb_node *node = &p->core_node;
282
283 node = rb_next(node);
284 if (!node)
285 return NULL;
286
287 p = container_of(node, struct task_struct, core_node);
288 if (p->core_cookie != cookie)
289 return NULL;
290
291 return p;
292}
293
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500294/*
295 * Magic required such that:
296 *
297 * raw_spin_rq_lock(rq);
298 * ...
299 * raw_spin_rq_unlock(rq);
300 *
301 * ends up locking and unlocking the _same_ lock, and all CPUs
302 * always agree on what rq has what lock.
303 *
304 * XXX entirely possible to selectively enable cores, don't bother for now.
305 */
306
307static DEFINE_MUTEX(sched_core_mutex);
Peter Zijlstra875feb42021-03-29 10:08:58 +0200308static atomic_t sched_core_count;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500309static struct cpumask sched_core_mask;
310
Peter Zijlstra3c474b32021-08-19 13:09:17 +0200311static void sched_core_lock(int cpu, unsigned long *flags)
312{
313 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
314 int t, i = 0;
315
316 local_irq_save(*flags);
317 for_each_cpu(t, smt_mask)
318 raw_spin_lock_nested(&cpu_rq(t)->__lock, i++);
319}
320
321static void sched_core_unlock(int cpu, unsigned long *flags)
322{
323 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
324 int t;
325
326 for_each_cpu(t, smt_mask)
327 raw_spin_unlock(&cpu_rq(t)->__lock);
328 local_irq_restore(*flags);
329}
330
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500331static void __sched_core_flip(bool enabled)
332{
Peter Zijlstra3c474b32021-08-19 13:09:17 +0200333 unsigned long flags;
334 int cpu, t;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500335
336 cpus_read_lock();
337
338 /*
339 * Toggle the online cores, one by one.
340 */
341 cpumask_copy(&sched_core_mask, cpu_online_mask);
342 for_each_cpu(cpu, &sched_core_mask) {
343 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
344
Peter Zijlstra3c474b32021-08-19 13:09:17 +0200345 sched_core_lock(cpu, &flags);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500346
347 for_each_cpu(t, smt_mask)
348 cpu_rq(t)->core_enabled = enabled;
349
Josh Don4feee7d12021-10-18 13:34:28 -0700350 cpu_rq(cpu)->core->core_forceidle_start = 0;
351
Peter Zijlstra3c474b32021-08-19 13:09:17 +0200352 sched_core_unlock(cpu, &flags);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500353
354 cpumask_andnot(&sched_core_mask, &sched_core_mask, smt_mask);
355 }
356
357 /*
358 * Toggle the offline CPUs.
359 */
Valentin Schneider585463f02022-10-03 16:34:20 +0100360 for_each_cpu_andnot(cpu, cpu_possible_mask, cpu_online_mask)
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500361 cpu_rq(cpu)->core_enabled = enabled;
362
363 cpus_read_unlock();
364}
365
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500366static void sched_core_assert_empty(void)
367{
368 int cpu;
369
370 for_each_possible_cpu(cpu)
371 WARN_ON_ONCE(!RB_EMPTY_ROOT(&cpu_rq(cpu)->core_tree));
372}
373
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500374static void __sched_core_enable(void)
375{
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500376 static_branch_enable(&__sched_core_enabled);
377 /*
378 * Ensure all previous instances of raw_spin_rq_*lock() have finished
379 * and future ones will observe !sched_core_disabled().
380 */
381 synchronize_rcu();
382 __sched_core_flip(true);
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500383 sched_core_assert_empty();
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500384}
385
386static void __sched_core_disable(void)
387{
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500388 sched_core_assert_empty();
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500389 __sched_core_flip(false);
390 static_branch_disable(&__sched_core_enabled);
391}
392
393void sched_core_get(void)
394{
Peter Zijlstra875feb42021-03-29 10:08:58 +0200395 if (atomic_inc_not_zero(&sched_core_count))
396 return;
397
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500398 mutex_lock(&sched_core_mutex);
Peter Zijlstra875feb42021-03-29 10:08:58 +0200399 if (!atomic_read(&sched_core_count))
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500400 __sched_core_enable();
Peter Zijlstra875feb42021-03-29 10:08:58 +0200401
402 smp_mb__before_atomic();
403 atomic_inc(&sched_core_count);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500404 mutex_unlock(&sched_core_mutex);
405}
406
Peter Zijlstra875feb42021-03-29 10:08:58 +0200407static void __sched_core_put(struct work_struct *work)
408{
409 if (atomic_dec_and_mutex_lock(&sched_core_count, &sched_core_mutex)) {
410 __sched_core_disable();
411 mutex_unlock(&sched_core_mutex);
412 }
413}
414
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500415void sched_core_put(void)
416{
Peter Zijlstra875feb42021-03-29 10:08:58 +0200417 static DECLARE_WORK(_work, __sched_core_put);
418
419 /*
420 * "There can be only one"
421 *
422 * Either this is the last one, or we don't actually need to do any
423 * 'work'. If it is the last *again*, we rely on
424 * WORK_STRUCT_PENDING_BIT.
425 */
426 if (!atomic_add_unless(&sched_core_count, -1, 1))
427 schedule_work(&_work);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500428}
429
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500430#else /* !CONFIG_SCHED_CORE */
431
432static inline void sched_core_enqueue(struct rq *rq, struct task_struct *p) { }
Josh Don4feee7d12021-10-18 13:34:28 -0700433static inline void
434sched_core_dequeue(struct rq *rq, struct task_struct *p, int flags) { }
Peter Zijlstra8a311c72020-11-17 18:19:36 -0500435
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500436#endif /* CONFIG_SCHED_CORE */
437
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100438/*
Peter Zijlstra58877d32020-07-02 14:52:11 +0200439 * Serialization rules:
440 *
441 * Lock order:
442 *
443 * p->pi_lock
444 * rq->lock
445 * hrtimer_cpu_base->lock (hrtimer_start() for bandwidth controls)
446 *
447 * rq1->lock
448 * rq2->lock where: rq1 < rq2
449 *
450 * Regular state:
451 *
452 * Normal scheduling state is serialized by rq->lock. __schedule() takes the
453 * local CPU's rq->lock, it optionally removes the task from the runqueue and
Tal Zussmanb19a8882020-11-12 19:51:56 -0500454 * always looks at the local rq data structures to find the most eligible task
Peter Zijlstra58877d32020-07-02 14:52:11 +0200455 * to run next.
456 *
457 * Task enqueue is also under rq->lock, possibly taken from another CPU.
458 * Wakeups from another LLC domain might use an IPI to transfer the enqueue to
459 * the local CPU to avoid bouncing the runqueue state around [ see
460 * ttwu_queue_wakelist() ]
461 *
462 * Task wakeup, specifically wakeups that involve migration, are horribly
463 * complicated to avoid having to take two rq->locks.
464 *
465 * Special state:
466 *
467 * System-calls and anything external will use task_rq_lock() which acquires
468 * both p->pi_lock and rq->lock. As a consequence the state they change is
469 * stable while holding either lock:
470 *
471 * - sched_setaffinity()/
472 * set_cpus_allowed_ptr(): p->cpus_ptr, p->nr_cpus_allowed
473 * - set_user_nice(): p->se.load, p->*prio
474 * - __sched_setscheduler(): p->sched_class, p->policy, p->*prio,
475 * p->se.load, p->rt_priority,
476 * p->dl.dl_{runtime, deadline, period, flags, bw, density}
477 * - sched_setnuma(): p->numa_preferred_nid
Chengming Zhou39c42612022-08-18 20:47:58 +0800478 * - sched_move_task(): p->sched_task_group
Peter Zijlstra58877d32020-07-02 14:52:11 +0200479 * - uclamp_update_active() p->uclamp*
480 *
481 * p->state <- TASK_*:
482 *
483 * is changed locklessly using set_current_state(), __set_current_state() or
484 * set_special_state(), see their respective comments, or by
485 * try_to_wake_up(). This latter uses p->pi_lock to serialize against
486 * concurrent self.
487 *
488 * p->on_rq <- { 0, 1 = TASK_ON_RQ_QUEUED, 2 = TASK_ON_RQ_MIGRATING }:
489 *
490 * is set by activate_task() and cleared by deactivate_task(), under
491 * rq->lock. Non-zero indicates the task is runnable, the special
492 * ON_RQ_MIGRATING state is used for migration without holding both
493 * rq->locks. It indicates task_cpu() is not stable, see task_rq_lock().
494 *
495 * p->on_cpu <- { 0, 1 }:
496 *
497 * is set by prepare_task() and cleared by finish_task() such that it will be
498 * set before p is scheduled-in and cleared after p is scheduled-out, both
499 * under rq->lock. Non-zero indicates the task is running on its CPU.
500 *
501 * [ The astute reader will observe that it is possible for two tasks on one
502 * CPU to have ->on_cpu = 1 at the same time. ]
503 *
504 * task_cpu(p): is changed by set_task_cpu(), the rules are:
505 *
506 * - Don't call set_task_cpu() on a blocked task:
507 *
508 * We don't care what CPU we're not running on, this simplifies hotplug,
509 * the CPU assignment of blocked tasks isn't required to be valid.
510 *
511 * - for try_to_wake_up(), called under p->pi_lock:
512 *
513 * This allows try_to_wake_up() to only take one rq->lock, see its comment.
514 *
515 * - for migration called under rq->lock:
516 * [ see task_on_rq_migrating() in task_rq_lock() ]
517 *
518 * o move_queued_task()
519 * o detach_task()
520 *
521 * - for migration called under double_rq_lock():
522 *
523 * o __migrate_swap_task()
524 * o push_rt_task() / pull_rt_task()
525 * o push_dl_task() / pull_dl_task()
526 * o dl_task_offline_migration()
527 *
528 */
529
Peter Zijlstra39d371b2021-03-02 12:13:13 +0100530void raw_spin_rq_lock_nested(struct rq *rq, int subclass)
531{
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100532 raw_spinlock_t *lock;
533
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500534 /* Matches synchronize_rcu() in __sched_core_enable() */
535 preempt_disable();
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100536 if (sched_core_disabled()) {
537 raw_spin_lock_nested(&rq->__lock, subclass);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500538 /* preempt_count *MUST* be > 1 */
539 preempt_enable_no_resched();
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100540 return;
541 }
542
543 for (;;) {
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +0100544 lock = __rq_lockp(rq);
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100545 raw_spin_lock_nested(lock, subclass);
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +0100546 if (likely(lock == __rq_lockp(rq))) {
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500547 /* preempt_count *MUST* be > 1 */
548 preempt_enable_no_resched();
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100549 return;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500550 }
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100551 raw_spin_unlock(lock);
552 }
Peter Zijlstra39d371b2021-03-02 12:13:13 +0100553}
554
555bool raw_spin_rq_trylock(struct rq *rq)
556{
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100557 raw_spinlock_t *lock;
558 bool ret;
559
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500560 /* Matches synchronize_rcu() in __sched_core_enable() */
561 preempt_disable();
562 if (sched_core_disabled()) {
563 ret = raw_spin_trylock(&rq->__lock);
564 preempt_enable();
565 return ret;
566 }
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100567
568 for (;;) {
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +0100569 lock = __rq_lockp(rq);
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100570 ret = raw_spin_trylock(lock);
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +0100571 if (!ret || (likely(lock == __rq_lockp(rq)))) {
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500572 preempt_enable();
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100573 return ret;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -0500574 }
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100575 raw_spin_unlock(lock);
576 }
Peter Zijlstra39d371b2021-03-02 12:13:13 +0100577}
578
579void raw_spin_rq_unlock(struct rq *rq)
580{
581 raw_spin_unlock(rq_lockp(rq));
582}
583
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100584#ifdef CONFIG_SMP
585/*
586 * double_rq_lock - safely lock two runqueues
587 */
588void double_rq_lock(struct rq *rq1, struct rq *rq2)
589{
590 lockdep_assert_irqs_disabled();
591
592 if (rq_order_less(rq2, rq1))
593 swap(rq1, rq2);
594
595 raw_spin_rq_lock(rq1);
Hao Jia2679a832022-04-30 16:58:42 +0800596 if (__rq_lockp(rq1) != __rq_lockp(rq2))
597 raw_spin_rq_lock_nested(rq2, SINGLE_DEPTH_NESTING);
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100598
Hao Jia2679a832022-04-30 16:58:42 +0800599 double_rq_clock_clear_update(rq1, rq2);
Peter Zijlstrad66f1b02021-03-02 12:16:48 +0100600}
601#endif
602
Dario Faggioli332ac172013-11-07 14:43:45 +0100603/*
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200604 * __task_rq_lock - lock the rq @p resides on.
605 */
Peter Zijlstraeb580752015-07-31 21:28:18 +0200606struct rq *__task_rq_lock(struct task_struct *p, struct rq_flags *rf)
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200607 __acquires(rq->lock)
608{
609 struct rq *rq;
610
611 lockdep_assert_held(&p->pi_lock);
612
613 for (;;) {
614 rq = task_rq(p);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -0500615 raw_spin_rq_lock(rq);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200616 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
Matt Flemingd8ac8972016-09-21 14:38:10 +0100617 rq_pin_lock(rq, rf);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200618 return rq;
619 }
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -0500620 raw_spin_rq_unlock(rq);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200621
622 while (unlikely(task_on_rq_migrating(p)))
623 cpu_relax();
624 }
625}
626
627/*
628 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
629 */
Peter Zijlstraeb580752015-07-31 21:28:18 +0200630struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200631 __acquires(p->pi_lock)
632 __acquires(rq->lock)
633{
634 struct rq *rq;
635
636 for (;;) {
Peter Zijlstraeb580752015-07-31 21:28:18 +0200637 raw_spin_lock_irqsave(&p->pi_lock, rf->flags);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200638 rq = task_rq(p);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -0500639 raw_spin_rq_lock(rq);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200640 /*
641 * move_queued_task() task_rq_lock()
642 *
643 * ACQUIRE (rq->lock)
644 * [S] ->on_rq = MIGRATING [L] rq = task_rq()
645 * WMB (__set_task_cpu()) ACQUIRE (rq->lock);
646 * [S] ->cpu = new_cpu [L] task_rq()
647 * [L] ->on_rq
648 * RELEASE (rq->lock)
649 *
Andrea Parric5469512019-01-21 16:52:40 +0100650 * If we observe the old CPU in task_rq_lock(), the acquire of
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200651 * the old rq->lock will fully serialize against the stores.
652 *
Andrea Parric5469512019-01-21 16:52:40 +0100653 * If we observe the new CPU in task_rq_lock(), the address
654 * dependency headed by '[L] rq = task_rq()' and the acquire
655 * will pair with the WMB to ensure we then also see migrating.
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200656 */
657 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) {
Matt Flemingd8ac8972016-09-21 14:38:10 +0100658 rq_pin_lock(rq, rf);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200659 return rq;
660 }
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -0500661 raw_spin_rq_unlock(rq);
Peter Zijlstraeb580752015-07-31 21:28:18 +0200662 raw_spin_unlock_irqrestore(&p->pi_lock, rf->flags);
Peter Zijlstra3e71a462016-04-28 16:16:33 +0200663
664 while (unlikely(task_on_rq_migrating(p)))
665 cpu_relax();
666 }
667}
668
Ingo Molnar535b9552017-02-01 12:29:21 +0100669/*
670 * RQ-clock updating methods:
671 */
672
673static void update_rq_clock_task(struct rq *rq, s64 delta)
674{
675/*
676 * In theory, the compile should just see 0 here, and optimize out the call
677 * to sched_rt_avg_update. But I don't trust it...
678 */
Vincent Guittot11d4afd2018-09-25 11:17:42 +0200679 s64 __maybe_unused steal = 0, irq_delta = 0;
680
Ingo Molnar535b9552017-02-01 12:29:21 +0100681#ifdef CONFIG_IRQ_TIME_ACCOUNTING
682 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
683
684 /*
685 * Since irq_time is only updated on {soft,}irq_exit, we might run into
686 * this case when a previous update_rq_clock() happened inside a
687 * {soft,}irq region.
688 *
689 * When this happens, we stop ->clock_task and only update the
690 * prev_irq_time stamp to account for the part that fit, so that a next
691 * update will consume the rest. This ensures ->clock_task is
692 * monotonic.
693 *
694 * It does however cause some slight miss-attribution of {soft,}irq
695 * time, a more accurate solution would be to update the irq_time using
696 * the current rq->clock timestamp, except that would require using
697 * atomic ops.
698 */
699 if (irq_delta > delta)
700 irq_delta = delta;
701
702 rq->prev_irq_time += irq_delta;
703 delta -= irq_delta;
Chengming Zhou52b13642022-08-26 00:41:08 +0800704 psi_account_irqtime(rq->curr, irq_delta);
Ingo Molnar535b9552017-02-01 12:29:21 +0100705#endif
706#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
707 if (static_key_false((&paravirt_steal_rq_enabled))) {
708 steal = paravirt_steal_clock(cpu_of(rq));
709 steal -= rq->prev_steal_time_rq;
710
711 if (unlikely(steal > delta))
712 steal = delta;
713
714 rq->prev_steal_time_rq += steal;
715 delta -= steal;
716 }
717#endif
718
719 rq->clock_task += delta;
720
Vincent Guittot11d4afd2018-09-25 11:17:42 +0200721#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
Ingo Molnar535b9552017-02-01 12:29:21 +0100722 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
Vincent Guittot91c27492018-06-28 17:45:09 +0200723 update_irq_load_avg(rq, irq_delta + steal);
Ingo Molnar535b9552017-02-01 12:29:21 +0100724#endif
Vincent Guittot23127292019-01-23 16:26:53 +0100725 update_rq_clock_pelt(rq, delta);
Ingo Molnar535b9552017-02-01 12:29:21 +0100726}
727
728void update_rq_clock(struct rq *rq)
729{
730 s64 delta;
731
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -0500732 lockdep_assert_rq_held(rq);
Ingo Molnar535b9552017-02-01 12:29:21 +0100733
734 if (rq->clock_update_flags & RQCF_ACT_SKIP)
735 return;
736
737#ifdef CONFIG_SCHED_DEBUG
Peter Zijlstra26ae58d2016-10-03 16:53:49 +0200738 if (sched_feat(WARN_DOUBLE_CLOCK))
739 SCHED_WARN_ON(rq->clock_update_flags & RQCF_UPDATED);
Ingo Molnar535b9552017-02-01 12:29:21 +0100740 rq->clock_update_flags |= RQCF_UPDATED;
741#endif
Peter Zijlstra26ae58d2016-10-03 16:53:49 +0200742
Ingo Molnar535b9552017-02-01 12:29:21 +0100743 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
744 if (delta < 0)
745 return;
746 rq->clock += delta;
747 update_rq_clock_task(rq, delta);
748}
749
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100750#ifdef CONFIG_SCHED_HRTICK
751/*
752 * Use HR-timers to deliver accurate preemption points.
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100753 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100754
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100755static void hrtick_clear(struct rq *rq)
756{
757 if (hrtimer_active(&rq->hrtick_timer))
758 hrtimer_cancel(&rq->hrtick_timer);
759}
760
761/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100762 * High-resolution timer tick.
763 * Runs from hardirq context with interrupts disabled.
764 */
765static enum hrtimer_restart hrtick(struct hrtimer *timer)
766{
767 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200768 struct rq_flags rf;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100769
770 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
771
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200772 rq_lock(rq, &rf);
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200773 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100774 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200775 rq_unlock(rq, &rf);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100776
777 return HRTIMER_NORESTART;
778}
779
Rabin Vincent95e904c2008-05-11 05:55:33 +0530780#ifdef CONFIG_SMP
Peter Zijlstra971ee282013-06-28 11:18:53 +0200781
Thomas Gleixner4961b6e2015-04-14 21:09:05 +0000782static void __hrtick_restart(struct rq *rq)
Peter Zijlstra971ee282013-06-28 11:18:53 +0200783{
784 struct hrtimer *timer = &rq->hrtick_timer;
Juri Lelli156ec6f2021-02-08 08:35:53 +0100785 ktime_t time = rq->hrtick_time;
Peter Zijlstra971ee282013-06-28 11:18:53 +0200786
Juri Lelli156ec6f2021-02-08 08:35:53 +0100787 hrtimer_start(timer, time, HRTIMER_MODE_ABS_PINNED_HARD);
Peter Zijlstra971ee282013-06-28 11:18:53 +0200788}
789
Peter Zijlstra31656512008-07-18 18:01:23 +0200790/*
791 * called from hardirq (IPI) context
792 */
793static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200794{
Peter Zijlstra31656512008-07-18 18:01:23 +0200795 struct rq *rq = arg;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200796 struct rq_flags rf;
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200797
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200798 rq_lock(rq, &rf);
Peter Zijlstra971ee282013-06-28 11:18:53 +0200799 __hrtick_restart(rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +0200800 rq_unlock(rq, &rf);
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200801}
802
Peter Zijlstra31656512008-07-18 18:01:23 +0200803/*
804 * Called to set the hrtick timer state.
805 *
806 * called with rq->lock held and irqs disabled
807 */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200808void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200809{
Peter Zijlstra31656512008-07-18 18:01:23 +0200810 struct hrtimer *timer = &rq->hrtick_timer;
xiaofeng.yan177ef2a2014-08-26 03:15:41 +0000811 s64 delta;
812
813 /*
814 * Don't schedule slices shorter than 10000ns, that just
815 * doesn't make sense and can cause timer DoS.
816 */
817 delta = max_t(s64, delay, 10000LL);
Juri Lelli156ec6f2021-02-08 08:35:53 +0100818 rq->hrtick_time = ktime_add_ns(timer->base->get_time(), delta);
Peter Zijlstra31656512008-07-18 18:01:23 +0200819
Peter Xufd3eafd2019-12-16 16:31:25 -0500820 if (rq == this_rq())
Peter Zijlstra971ee282013-06-28 11:18:53 +0200821 __hrtick_restart(rq);
Peter Xufd3eafd2019-12-16 16:31:25 -0500822 else
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100823 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200824}
825
Peter Zijlstra31656512008-07-18 18:01:23 +0200826#else
827/*
828 * Called to set the hrtick timer state.
829 *
830 * called with rq->lock held and irqs disabled
831 */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200832void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstra31656512008-07-18 18:01:23 +0200833{
Wanpeng Li86893332014-11-26 08:44:06 +0800834 /*
835 * Don't schedule slices shorter than 10000ns, that just
836 * doesn't make sense. Rely on vruntime for fairness.
837 */
838 delay = max_t(u64, delay, 10000LL);
Thomas Gleixner4961b6e2015-04-14 21:09:05 +0000839 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
Sebastian Andrzej Siewiord5096aa2019-07-26 20:30:52 +0200840 HRTIMER_MODE_REL_PINNED_HARD);
Peter Zijlstra31656512008-07-18 18:01:23 +0200841}
Peter Zijlstra (Intel)90b53632020-03-27 11:44:56 +0100842
Rabin Vincent95e904c2008-05-11 05:55:33 +0530843#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200844
Frederic Weisbecker77a021b2018-02-21 05:17:23 +0100845static void hrtick_rq_init(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100846{
Peter Zijlstra31656512008-07-18 18:01:23 +0200847#ifdef CONFIG_SMP
Peter Zijlstra545b8c82020-06-15 11:29:31 +0200848 INIT_CSD(&rq->hrtick_csd, __hrtick_start, rq);
Peter Zijlstra31656512008-07-18 18:01:23 +0200849#endif
Sebastian Andrzej Siewiord5096aa2019-07-26 20:30:52 +0200850 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100851 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100852}
Andrew Morton006c75f2008-09-22 14:55:46 -0700853#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100854static inline void hrtick_clear(struct rq *rq)
855{
856}
857
Frederic Weisbecker77a021b2018-02-21 05:17:23 +0100858static inline void hrtick_rq_init(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100859{
860}
Andrew Morton006c75f2008-09-22 14:55:46 -0700861#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100862
Frederic Weisbecker55295782016-03-24 15:38:01 +0100863/*
864 * cmpxchg based fetch_or, macro so it works for different integer types
865 */
866#define fetch_or(ptr, mask) \
867 ({ \
868 typeof(ptr) _ptr = (ptr); \
869 typeof(mask) _mask = (mask); \
Uros Bizjakc02d5542022-06-29 17:15:52 +0200870 typeof(*_ptr) _val = *_ptr; \
Frederic Weisbecker55295782016-03-24 15:38:01 +0100871 \
Uros Bizjakc02d5542022-06-29 17:15:52 +0200872 do { \
873 } while (!try_cmpxchg(_ptr, &_val, _val | _mask)); \
874 _val; \
Frederic Weisbecker55295782016-03-24 15:38:01 +0100875})
876
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700877#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200878/*
879 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
880 * this avoids any races wrt polling state changes and thereby avoids
881 * spurious IPIs.
882 */
Uros Bizjakc02d5542022-06-29 17:15:52 +0200883static inline bool set_nr_and_not_polling(struct task_struct *p)
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200884{
885 struct thread_info *ti = task_thread_info(p);
886 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
887}
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700888
889/*
890 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
891 *
892 * If this returns true, then the idle task promises to call
893 * sched_ttwu_pending() and reschedule soon.
894 */
895static bool set_nr_if_polling(struct task_struct *p)
896{
897 struct thread_info *ti = task_thread_info(p);
Uros Bizjakc02d5542022-06-29 17:15:52 +0200898 typeof(ti->flags) val = READ_ONCE(ti->flags);
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700899
900 for (;;) {
901 if (!(val & _TIF_POLLING_NRFLAG))
902 return false;
903 if (val & _TIF_NEED_RESCHED)
904 return true;
Uros Bizjakc02d5542022-06-29 17:15:52 +0200905 if (try_cmpxchg(&ti->flags, &val, val | _TIF_NEED_RESCHED))
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700906 break;
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700907 }
908 return true;
909}
910
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200911#else
Uros Bizjakc02d5542022-06-29 17:15:52 +0200912static inline bool set_nr_and_not_polling(struct task_struct *p)
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200913{
914 set_tsk_need_resched(p);
915 return true;
916}
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700917
918#ifdef CONFIG_SMP
Uros Bizjakc02d5542022-06-29 17:15:52 +0200919static inline bool set_nr_if_polling(struct task_struct *p)
Peter Zijlstrae3baac472014-06-04 10:31:18 -0700920{
921 return false;
922}
923#endif
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200924#endif
925
Davidlohr Bueso07879c62018-12-18 11:53:52 -0800926static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
927{
928 struct wake_q_node *node = &task->wake_q;
929
930 /*
931 * Atomically grab the task, if ->wake_q is !nil already it means
Tal Zussmanb19a8882020-11-12 19:51:56 -0500932 * it's already queued (either by us or someone else) and will get the
Davidlohr Bueso07879c62018-12-18 11:53:52 -0800933 * wakeup due to that.
934 *
935 * In order to ensure that a pending wakeup will observe our pending
936 * state, even in the failed case, an explicit smp_mb() must be used.
937 */
938 smp_mb__before_atomic();
939 if (unlikely(cmpxchg_relaxed(&node->next, NULL, WAKE_Q_TAIL)))
940 return false;
941
942 /*
943 * The head is context local, there can be no concurrency.
944 */
945 *head->lastp = node;
946 head->lastp = &node->next;
947 return true;
948}
949
Peter Zijlstrae6018c02018-12-17 10:14:53 +0100950/**
951 * wake_q_add() - queue a wakeup for 'later' waking.
952 * @head: the wake_q_head to add @task to
953 * @task: the task to queue for 'later' wakeup
954 *
955 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
956 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
957 * instantly.
958 *
959 * This function must be used as-if it were wake_up_process(); IOW the task
960 * must be ready to be woken at this location.
961 */
Peter Zijlstra76751042015-05-01 08:27:50 -0700962void wake_q_add(struct wake_q_head *head, struct task_struct *task)
963{
Davidlohr Bueso07879c62018-12-18 11:53:52 -0800964 if (__wake_q_add(head, task))
965 get_task_struct(task);
966}
Peter Zijlstra76751042015-05-01 08:27:50 -0700967
Davidlohr Bueso07879c62018-12-18 11:53:52 -0800968/**
969 * wake_q_add_safe() - safely queue a wakeup for 'later' waking.
970 * @head: the wake_q_head to add @task to
971 * @task: the task to queue for 'later' wakeup
972 *
973 * Queue a task for later wakeup, most likely by the wake_up_q() call in the
974 * same context, _HOWEVER_ this is not guaranteed, the wakeup can come
975 * instantly.
976 *
977 * This function must be used as-if it were wake_up_process(); IOW the task
978 * must be ready to be woken at this location.
979 *
980 * This function is essentially a task-safe equivalent to wake_q_add(). Callers
981 * that already hold reference to @task can call the 'safe' version and trust
982 * wake_q to do the right thing depending whether or not the @task is already
983 * queued for wakeup.
984 */
985void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task)
986{
987 if (!__wake_q_add(head, task))
988 put_task_struct(task);
Peter Zijlstra76751042015-05-01 08:27:50 -0700989}
990
991void wake_up_q(struct wake_q_head *head)
992{
993 struct wake_q_node *node = head->first;
994
995 while (node != WAKE_Q_TAIL) {
996 struct task_struct *task;
997
998 task = container_of(node, struct task_struct, wake_q);
Ingo Molnard1ccc662017-02-01 11:46:42 +0100999 /* Task can safely be re-inserted now: */
Peter Zijlstra76751042015-05-01 08:27:50 -07001000 node = node->next;
1001 task->wake_q.next = NULL;
1002
1003 /*
Andrea Parri7696f992018-07-16 11:06:03 -07001004 * wake_up_process() executes a full barrier, which pairs with
1005 * the queueing in wake_q_add() so as not to miss wakeups.
Peter Zijlstra76751042015-05-01 08:27:50 -07001006 */
1007 wake_up_process(task);
1008 put_task_struct(task);
1009 }
1010}
1011
Peter Zijlstrafd99f912014-04-09 15:35:08 +02001012/*
Kirill Tkhai88751252014-06-29 00:03:57 +04001013 * resched_curr - mark rq's current task 'to be rescheduled now'.
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001014 *
1015 * On UP this means the setting of the need_resched flag, on SMP it
1016 * might also involve a cross-CPU call to trigger the scheduler on
1017 * the target CPU.
1018 */
Kirill Tkhai88751252014-06-29 00:03:57 +04001019void resched_curr(struct rq *rq)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001020{
Kirill Tkhai88751252014-06-29 00:03:57 +04001021 struct task_struct *curr = rq->curr;
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001022 int cpu;
1023
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05001024 lockdep_assert_rq_held(rq);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001025
Kirill Tkhai88751252014-06-29 00:03:57 +04001026 if (test_tsk_need_resched(curr))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001027 return;
1028
Kirill Tkhai88751252014-06-29 00:03:57 +04001029 cpu = cpu_of(rq);
Peter Zijlstrafd99f912014-04-09 15:35:08 +02001030
Peter Zijlstraf27dde82013-08-14 14:55:31 +02001031 if (cpu == smp_processor_id()) {
Kirill Tkhai88751252014-06-29 00:03:57 +04001032 set_tsk_need_resched(curr);
Peter Zijlstraf27dde82013-08-14 14:55:31 +02001033 set_preempt_need_resched();
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001034 return;
Peter Zijlstraf27dde82013-08-14 14:55:31 +02001035 }
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001036
Kirill Tkhai88751252014-06-29 00:03:57 +04001037 if (set_nr_and_not_polling(curr))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001038 smp_send_reschedule(cpu);
Andy Lutomirskidfc68f22014-06-04 10:31:15 -07001039 else
1040 trace_sched_wake_idle_without_ipi(cpu);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001041}
1042
Peter Zijlstra029632f2011-10-25 10:00:11 +02001043void resched_cpu(int cpu)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001044{
1045 struct rq *rq = cpu_rq(cpu);
1046 unsigned long flags;
1047
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05001048 raw_spin_rq_lock_irqsave(rq, flags);
Paul E. McKenneya0982df2017-10-13 16:24:28 -07001049 if (cpu_online(cpu) || cpu == smp_processor_id())
1050 resched_curr(rq);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05001051 raw_spin_rq_unlock_irqrestore(rq, flags);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001052}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001053
Peter Zijlstrab021fe32013-09-17 09:30:55 +02001054#ifdef CONFIG_SMP
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001055#ifdef CONFIG_NO_HZ_COMMON
Thomas Gleixner06d83082008-03-22 09:20:24 +01001056/*
Ingo Molnard1ccc662017-02-01 11:46:42 +01001057 * In the semi idle case, use the nearest busy CPU for migrating timers
1058 * from an idle CPU. This is good for power-savings.
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001059 *
1060 * We don't do similar optimization for completely idle system, as
Ingo Molnard1ccc662017-02-01 11:46:42 +01001061 * selecting an idle CPU will add more delays to the timers than intended
1062 * (as that CPU's timer base may not be uptodate wrt jiffies etc).
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001063 */
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +00001064int get_nohz_timer_target(void)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001065{
Wanpeng Lie938b9c2020-01-13 08:50:27 +08001066 int i, cpu = smp_processor_id(), default_cpu = -1;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001067 struct sched_domain *sd;
Yuan ZhaoXiong031e3bd2021-06-06 21:11:55 +08001068 const struct cpumask *hk_mask;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001069
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01001070 if (housekeeping_cpu(cpu, HK_TYPE_TIMER)) {
Wanpeng Lie938b9c2020-01-13 08:50:27 +08001071 if (!idle_cpu(cpu))
1072 return cpu;
1073 default_cpu = cpu;
1074 }
Viresh Kumar6201b4d2014-03-18 16:26:07 +05301075
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01001076 hk_mask = housekeeping_cpumask(HK_TYPE_TIMER);
Yuan ZhaoXiong031e3bd2021-06-06 21:11:55 +08001077
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02001078 rcu_read_lock();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001079 for_each_domain(cpu, sd) {
Yuan ZhaoXiong031e3bd2021-06-06 21:11:55 +08001080 for_each_cpu_and(i, sched_domain_span(sd), hk_mask) {
Wanpeng Li44496922016-05-04 14:45:34 +08001081 if (cpu == i)
1082 continue;
1083
Wanpeng Lie938b9c2020-01-13 08:50:27 +08001084 if (!idle_cpu(i)) {
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02001085 cpu = i;
1086 goto unlock;
1087 }
1088 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001089 }
Vatika Harlalka9642d18e2015-09-01 16:50:59 +02001090
Wanpeng Lie938b9c2020-01-13 08:50:27 +08001091 if (default_cpu == -1)
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01001092 default_cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
Wanpeng Lie938b9c2020-01-13 08:50:27 +08001093 cpu = default_cpu;
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02001094unlock:
1095 rcu_read_unlock();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001096 return cpu;
1097}
Ingo Molnard1ccc662017-02-01 11:46:42 +01001098
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07001099/*
Thomas Gleixner06d83082008-03-22 09:20:24 +01001100 * When add_timer_on() enqueues a timer into the timer wheel of an
1101 * idle CPU then this timer might expire before the next timer event
1102 * which is scheduled to wake up that CPU. In case of a completely
1103 * idle system the next event might even be infinite time into the
1104 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1105 * leaves the inner idle loop so the newly added timer is taken into
1106 * account when the CPU goes back to idle and evaluates the timer
1107 * wheel for the next timer event.
1108 */
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001109static void wake_up_idle_cpu(int cpu)
Thomas Gleixner06d83082008-03-22 09:20:24 +01001110{
1111 struct rq *rq = cpu_rq(cpu);
1112
1113 if (cpu == smp_processor_id())
1114 return;
1115
Andy Lutomirski67b9ca72014-06-04 10:31:17 -07001116 if (set_nr_and_not_polling(rq->idle))
Thomas Gleixner06d83082008-03-22 09:20:24 +01001117 smp_send_reschedule(cpu);
Andy Lutomirskidfc68f22014-06-04 10:31:15 -07001118 else
1119 trace_sched_wake_idle_without_ipi(cpu);
Thomas Gleixner06d83082008-03-22 09:20:24 +01001120}
Mike Galbraith39c0cbe2010-03-11 17:17:13 +01001121
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +02001122static bool wake_up_full_nohz_cpu(int cpu)
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001123{
Frederic Weisbecker53c5fa12014-06-04 16:20:21 +02001124 /*
1125 * We just need the target to call irq_exit() and re-evaluate
1126 * the next tick. The nohz full kick at least implies that.
1127 * If needed we can still optimize that later with an
1128 * empty IRQ.
1129 */
Paul E. McKenney379d9ec2016-06-30 10:37:20 -07001130 if (cpu_is_offline(cpu))
1131 return true; /* Don't try to wake offline CPUs. */
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +02001132 if (tick_nohz_full_cpu(cpu)) {
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001133 if (cpu != smp_processor_id() ||
1134 tick_nohz_tick_stopped())
Frederic Weisbecker53c5fa12014-06-04 16:20:21 +02001135 tick_nohz_full_kick_cpu(cpu);
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001136 return true;
1137 }
1138
1139 return false;
1140}
1141
Paul E. McKenney379d9ec2016-06-30 10:37:20 -07001142/*
1143 * Wake up the specified CPU. If the CPU is going offline, it is the
1144 * caller's responsibility to deal with the lost wakeup, for example,
1145 * by hooking into the CPU_DEAD notifier like timers and hrtimers do.
1146 */
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001147void wake_up_nohz_cpu(int cpu)
1148{
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +02001149 if (!wake_up_full_nohz_cpu(cpu))
Frederic Weisbecker1c200912011-08-10 23:21:01 +02001150 wake_up_idle_cpu(cpu);
1151}
1152
Peter Zijlstra (Intel)90b53632020-03-27 11:44:56 +01001153static void nohz_csd_func(void *info)
Suresh Siddhaca380622011-10-03 15:09:00 -07001154{
Peter Zijlstra (Intel)90b53632020-03-27 11:44:56 +01001155 struct rq *rq = info;
Peter Zijlstra19a1f5e2020-05-26 18:10:58 +02001156 int cpu = cpu_of(rq);
1157 unsigned int flags;
Vincent Guittot873b4c62013-06-05 10:13:11 +02001158
1159 /*
Peter Zijlstra19a1f5e2020-05-26 18:10:58 +02001160 * Release the rq::nohz_csd.
Vincent Guittot873b4c62013-06-05 10:13:11 +02001161 */
Vincent Guittotc6f88652021-02-24 14:30:06 +01001162 flags = atomic_fetch_andnot(NOHZ_KICK_MASK | NOHZ_NEWILB_KICK, nohz_flags(cpu));
Peter Zijlstra19a1f5e2020-05-26 18:10:58 +02001163 WARN_ON(!(flags & NOHZ_KICK_MASK));
Suresh Siddhaca380622011-10-03 15:09:00 -07001164
Peter Zijlstra19a1f5e2020-05-26 18:10:58 +02001165 rq->idle_balance = idle_cpu(cpu);
1166 if (rq->idle_balance && !need_resched()) {
1167 rq->nohz_idle_balance = flags;
Peter Zijlstra (Intel)90b53632020-03-27 11:44:56 +01001168 raise_softirq_irqoff(SCHED_SOFTIRQ);
1169 }
Suresh Siddhaca380622011-10-03 15:09:00 -07001170}
1171
Frederic Weisbecker3451d022011-08-10 23:21:01 +02001172#endif /* CONFIG_NO_HZ_COMMON */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001173
Frederic Weisbeckerce831b32013-04-20 15:15:35 +02001174#ifdef CONFIG_NO_HZ_FULL
Frederic Weisbecker76d92ac2015-07-17 22:25:49 +02001175bool sched_can_stop_tick(struct rq *rq)
Frederic Weisbeckerce831b32013-04-20 15:15:35 +02001176{
Frederic Weisbecker76d92ac2015-07-17 22:25:49 +02001177 int fifo_nr_running;
1178
1179 /* Deadline tasks, even if single, need the tick */
1180 if (rq->dl.dl_nr_running)
1181 return false;
1182
Frederic Weisbecker3882ec62014-03-18 22:54:04 +01001183 /*
Tal Zussmanb19a8882020-11-12 19:51:56 -05001184 * If there are more than one RR tasks, we need the tick to affect the
Peter Zijlstra2548d542016-04-21 18:03:15 +02001185 * actual RR behaviour.
Rik van Riel1e78cdb2015-02-16 15:23:49 -05001186 */
Frederic Weisbecker76d92ac2015-07-17 22:25:49 +02001187 if (rq->rt.rr_nr_running) {
1188 if (rq->rt.rr_nr_running == 1)
1189 return true;
1190 else
1191 return false;
Rik van Riel1e78cdb2015-02-16 15:23:49 -05001192 }
1193
Peter Zijlstra2548d542016-04-21 18:03:15 +02001194 /*
1195 * If there's no RR tasks, but FIFO tasks, we can skip the tick, no
1196 * forced preemption between FIFO tasks.
1197 */
1198 fifo_nr_running = rq->rt.rt_nr_running - rq->rt.rr_nr_running;
1199 if (fifo_nr_running)
1200 return true;
1201
1202 /*
1203 * If there are no DL,RR/FIFO tasks, there must only be CFS tasks left;
1204 * if there's more than one we need the tick for involuntary
1205 * preemption.
1206 */
1207 if (rq->nr_running > 1)
Viresh Kumar541b8262014-06-24 14:04:12 +05301208 return false;
Frederic Weisbeckerce831b32013-04-20 15:15:35 +02001209
Viresh Kumar541b8262014-06-24 14:04:12 +05301210 return true;
Frederic Weisbeckerce831b32013-04-20 15:15:35 +02001211}
1212#endif /* CONFIG_NO_HZ_FULL */
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001213#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001214
Paul Turnera790de92011-07-21 09:43:29 -07001215#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1216 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
Peter Zijlstraeb755802008-08-19 12:33:05 +02001217/*
Paul Turner82774342011-07-21 09:43:35 -07001218 * Iterate task_group tree rooted at *from, calling @down when first entering a
1219 * node and @up when leaving it for the final time.
1220 *
1221 * Caller must hold rcu_lock or sufficient equivalent.
Peter Zijlstraeb755802008-08-19 12:33:05 +02001222 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02001223int walk_tg_tree_from(struct task_group *from,
Paul Turner82774342011-07-21 09:43:35 -07001224 tg_visitor down, tg_visitor up, void *data)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001225{
1226 struct task_group *parent, *child;
1227 int ret;
1228
Paul Turner82774342011-07-21 09:43:35 -07001229 parent = from;
1230
Peter Zijlstraeb755802008-08-19 12:33:05 +02001231down:
1232 ret = (*down)(parent, data);
1233 if (ret)
Paul Turner82774342011-07-21 09:43:35 -07001234 goto out;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001235 list_for_each_entry_rcu(child, &parent->children, siblings) {
1236 parent = child;
1237 goto down;
1238
1239up:
1240 continue;
1241 }
1242 ret = (*up)(parent, data);
Paul Turner82774342011-07-21 09:43:35 -07001243 if (ret || parent == from)
1244 goto out;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001245
1246 child = parent;
1247 parent = parent->parent;
1248 if (parent)
1249 goto up;
Paul Turner82774342011-07-21 09:43:35 -07001250out:
Peter Zijlstraeb755802008-08-19 12:33:05 +02001251 return ret;
1252}
1253
Peter Zijlstra029632f2011-10-25 10:00:11 +02001254int tg_nop(struct task_group *tg, void *data)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001255{
1256 return 0;
1257}
1258#endif
1259
Peter Zijlstrab1e82062022-02-14 10:16:57 +01001260static void set_load_weight(struct task_struct *p, bool update_load)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001261{
Nikhil Raof05998d2011-05-18 10:09:38 -07001262 int prio = p->static_prio - MAX_RT_PRIO;
1263 struct load_weight *load = &p->se.load;
1264
Ingo Molnardd41f592007-07-09 18:51:59 +02001265 /*
1266 * SCHED_IDLE tasks get minimal weight:
1267 */
Viresh Kumar1da18432018-11-05 16:51:55 +05301268 if (task_has_idle_policy(p)) {
Nikhil Raoc8b28112011-05-18 14:37:48 -07001269 load->weight = scale_load(WEIGHT_IDLEPRIO);
Nikhil Raof05998d2011-05-18 10:09:38 -07001270 load->inv_weight = WMULT_IDLEPRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02001271 return;
1272 }
1273
Vincent Guittot90593932017-05-17 11:50:45 +02001274 /*
1275 * SCHED_OTHER tasks have to update their load when changing their
1276 * weight
1277 */
1278 if (update_load && p->sched_class == &fair_sched_class) {
1279 reweight_task(p, prio);
1280 } else {
1281 load->weight = scale_load(sched_prio_to_weight[prio]);
1282 load->inv_weight = sched_prio_to_wmult[prio];
1283 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001284}
1285
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001286#ifdef CONFIG_UCLAMP_TASK
Patrick Bellasi2480c092019-08-22 14:28:06 +01001287/*
1288 * Serializes updates of utilization clamp values
1289 *
1290 * The (slow-path) user-space triggers utilization clamp value updates which
1291 * can require updates on (fast-path) scheduler's data structures used to
1292 * support enqueue/dequeue operations.
1293 * While the per-CPU rq lock protects fast-path update operations, user-space
1294 * requests are serialized using a mutex to reduce the risk of conflicting
1295 * updates or API abuses.
1296 */
1297static DEFINE_MUTEX(uclamp_mutex);
1298
Patrick Bellasie8f14172019-06-21 09:42:05 +01001299/* Max allowed minimum utilization */
YueHaibing494dcdf2022-04-27 21:10:02 +08001300static unsigned int __maybe_unused sysctl_sched_uclamp_util_min = SCHED_CAPACITY_SCALE;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001301
1302/* Max allowed maximum utilization */
YueHaibing494dcdf2022-04-27 21:10:02 +08001303static unsigned int __maybe_unused sysctl_sched_uclamp_util_max = SCHED_CAPACITY_SCALE;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001304
Qais Yousef13685c42020-07-16 12:03:45 +01001305/*
1306 * By default RT tasks run at the maximum performance point/capacity of the
1307 * system. Uclamp enforces this by always setting UCLAMP_MIN of RT tasks to
1308 * SCHED_CAPACITY_SCALE.
1309 *
1310 * This knob allows admins to change the default behavior when uclamp is being
1311 * used. In battery powered devices, particularly, running at the maximum
1312 * capacity and frequency will increase energy consumption and shorten the
1313 * battery life.
1314 *
1315 * This knob only affects RT tasks that their uclamp_se->user_defined == false.
1316 *
1317 * This knob will not override the system default sched_util_clamp_min defined
1318 * above.
1319 */
Zhen Ni3267e012022-02-15 19:46:02 +08001320static unsigned int sysctl_sched_uclamp_util_min_rt_default = SCHED_CAPACITY_SCALE;
Qais Yousef13685c42020-07-16 12:03:45 +01001321
Patrick Bellasie8f14172019-06-21 09:42:05 +01001322/* All clamps are required to be less or equal than these values */
1323static struct uclamp_se uclamp_default[UCLAMP_CNT];
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001324
Qais Yousef46609ce2020-06-30 12:21:23 +01001325/*
1326 * This static key is used to reduce the uclamp overhead in the fast path. It
1327 * primarily disables the call to uclamp_rq_{inc, dec}() in
1328 * enqueue/dequeue_task().
1329 *
1330 * This allows users to continue to enable uclamp in their kernel config with
1331 * minimum uclamp overhead in the fast path.
1332 *
1333 * As soon as userspace modifies any of the uclamp knobs, the static key is
1334 * enabled, since we have an actual users that make use of uclamp
1335 * functionality.
1336 *
1337 * The knobs that would enable this static key are:
1338 *
1339 * * A task modifying its uclamp value with sched_setattr().
1340 * * An admin modifying the sysctl_sched_uclamp_{min, max} via procfs.
1341 * * An admin modifying the cgroup cpu.uclamp.{min, max}
1342 */
1343DEFINE_STATIC_KEY_FALSE(sched_uclamp_used);
1344
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001345/* Integer rounded range for each bucket */
1346#define UCLAMP_BUCKET_DELTA DIV_ROUND_CLOSEST(SCHED_CAPACITY_SCALE, UCLAMP_BUCKETS)
1347
1348#define for_each_clamp_id(clamp_id) \
1349 for ((clamp_id) = 0; (clamp_id) < UCLAMP_CNT; (clamp_id)++)
1350
1351static inline unsigned int uclamp_bucket_id(unsigned int clamp_value)
1352{
Quentin Perret6d2f8902021-04-30 15:14:12 +00001353 return min_t(unsigned int, clamp_value / UCLAMP_BUCKET_DELTA, UCLAMP_BUCKETS - 1);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001354}
1355
Valentin Schneider7763baa2019-11-15 10:39:08 +00001356static inline unsigned int uclamp_none(enum uclamp_id clamp_id)
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001357{
1358 if (clamp_id == UCLAMP_MIN)
1359 return 0;
1360 return SCHED_CAPACITY_SCALE;
1361}
1362
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001363static inline void uclamp_se_set(struct uclamp_se *uc_se,
1364 unsigned int value, bool user_defined)
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001365{
1366 uc_se->value = value;
1367 uc_se->bucket_id = uclamp_bucket_id(value);
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001368 uc_se->user_defined = user_defined;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001369}
1370
Patrick Bellasie4961872019-06-21 09:42:04 +01001371static inline unsigned int
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001372uclamp_idle_value(struct rq *rq, enum uclamp_id clamp_id,
Patrick Bellasie4961872019-06-21 09:42:04 +01001373 unsigned int clamp_value)
1374{
1375 /*
1376 * Avoid blocked utilization pushing up the frequency when we go
1377 * idle (which drops the max-clamp) by retaining the last known
1378 * max-clamp.
1379 */
1380 if (clamp_id == UCLAMP_MAX) {
1381 rq->uclamp_flags |= UCLAMP_FLAG_IDLE;
1382 return clamp_value;
1383 }
1384
1385 return uclamp_none(UCLAMP_MIN);
1386}
1387
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001388static inline void uclamp_idle_reset(struct rq *rq, enum uclamp_id clamp_id,
Patrick Bellasie4961872019-06-21 09:42:04 +01001389 unsigned int clamp_value)
1390{
1391 /* Reset max-clamp retention only on idle exit */
1392 if (!(rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1393 return;
1394
Qais Yousef24422602022-08-04 15:36:03 +01001395 uclamp_rq_set(rq, clamp_id, clamp_value);
Patrick Bellasie4961872019-06-21 09:42:04 +01001396}
1397
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001398static inline
Valentin Schneider7763baa2019-11-15 10:39:08 +00001399unsigned int uclamp_rq_max_value(struct rq *rq, enum uclamp_id clamp_id,
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001400 unsigned int clamp_value)
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001401{
1402 struct uclamp_bucket *bucket = rq->uclamp[clamp_id].bucket;
1403 int bucket_id = UCLAMP_BUCKETS - 1;
1404
1405 /*
1406 * Since both min and max clamps are max aggregated, find the
1407 * top most bucket with tasks in.
1408 */
1409 for ( ; bucket_id >= 0; bucket_id--) {
1410 if (!bucket[bucket_id].tasks)
1411 continue;
1412 return bucket[bucket_id].value;
1413 }
1414
1415 /* No tasks -- default clamp values */
Patrick Bellasie4961872019-06-21 09:42:04 +01001416 return uclamp_idle_value(rq, clamp_id, clamp_value);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001417}
1418
Qais Yousef13685c42020-07-16 12:03:45 +01001419static void __uclamp_update_util_min_rt_default(struct task_struct *p)
1420{
1421 unsigned int default_util_min;
1422 struct uclamp_se *uc_se;
1423
1424 lockdep_assert_held(&p->pi_lock);
1425
1426 uc_se = &p->uclamp_req[UCLAMP_MIN];
1427
1428 /* Only sync if user didn't override the default */
1429 if (uc_se->user_defined)
1430 return;
1431
1432 default_util_min = sysctl_sched_uclamp_util_min_rt_default;
1433 uclamp_se_set(uc_se, default_util_min, false);
1434}
1435
1436static void uclamp_update_util_min_rt_default(struct task_struct *p)
1437{
1438 struct rq_flags rf;
1439 struct rq *rq;
1440
1441 if (!rt_task(p))
1442 return;
1443
1444 /* Protect updates to p->uclamp_* */
1445 rq = task_rq_lock(p, &rf);
1446 __uclamp_update_util_min_rt_default(p);
1447 task_rq_unlock(rq, p, &rf);
1448}
1449
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001450static inline struct uclamp_se
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001451uclamp_tg_restrict(struct task_struct *p, enum uclamp_id clamp_id)
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001452{
Qais Yousef0213b702021-06-17 17:51:55 +01001453 /* Copy by value as we could modify it */
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001454 struct uclamp_se uc_req = p->uclamp_req[clamp_id];
1455#ifdef CONFIG_UCLAMP_TASK_GROUP
Qais Yousef0213b702021-06-17 17:51:55 +01001456 unsigned int tg_min, tg_max, value;
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001457
1458 /*
1459 * Tasks in autogroups or root task group will be
1460 * restricted by system defaults.
1461 */
1462 if (task_group_is_autogroup(task_group(p)))
1463 return uc_req;
1464 if (task_group(p) == &root_task_group)
1465 return uc_req;
1466
Qais Yousef0213b702021-06-17 17:51:55 +01001467 tg_min = task_group(p)->uclamp[UCLAMP_MIN].value;
1468 tg_max = task_group(p)->uclamp[UCLAMP_MAX].value;
1469 value = uc_req.value;
1470 value = clamp(value, tg_min, tg_max);
1471 uclamp_se_set(&uc_req, value, false);
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001472#endif
1473
1474 return uc_req;
1475}
1476
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001477/*
Patrick Bellasie8f14172019-06-21 09:42:05 +01001478 * The effective clamp bucket index of a task depends on, by increasing
1479 * priority:
1480 * - the task specific clamp value, when explicitly requested from userspace
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001481 * - the task group effective clamp value, for tasks not either in the root
1482 * group or in an autogroup
Patrick Bellasie8f14172019-06-21 09:42:05 +01001483 * - the system default clamp value, defined by the sysadmin
1484 */
1485static inline struct uclamp_se
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001486uclamp_eff_get(struct task_struct *p, enum uclamp_id clamp_id)
Patrick Bellasie8f14172019-06-21 09:42:05 +01001487{
Patrick Bellasi3eac8702019-08-22 14:28:09 +01001488 struct uclamp_se uc_req = uclamp_tg_restrict(p, clamp_id);
Patrick Bellasie8f14172019-06-21 09:42:05 +01001489 struct uclamp_se uc_max = uclamp_default[clamp_id];
1490
1491 /* System default restrictions always apply */
1492 if (unlikely(uc_req.value > uc_max.value))
1493 return uc_max;
1494
1495 return uc_req;
1496}
1497
Valentin Schneider686516b2019-12-11 11:38:48 +00001498unsigned long uclamp_eff_value(struct task_struct *p, enum uclamp_id clamp_id)
Patrick Bellasi9d20ad72019-06-21 09:42:11 +01001499{
1500 struct uclamp_se uc_eff;
1501
1502 /* Task currently refcounted: use back-annotated (effective) value */
1503 if (p->uclamp[clamp_id].active)
Valentin Schneider686516b2019-12-11 11:38:48 +00001504 return (unsigned long)p->uclamp[clamp_id].value;
Patrick Bellasi9d20ad72019-06-21 09:42:11 +01001505
1506 uc_eff = uclamp_eff_get(p, clamp_id);
1507
Valentin Schneider686516b2019-12-11 11:38:48 +00001508 return (unsigned long)uc_eff.value;
Patrick Bellasi9d20ad72019-06-21 09:42:11 +01001509}
1510
Patrick Bellasie8f14172019-06-21 09:42:05 +01001511/*
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001512 * When a task is enqueued on a rq, the clamp bucket currently defined by the
1513 * task's uclamp::bucket_id is refcounted on that rq. This also immediately
1514 * updates the rq's clamp value if required.
Patrick Bellasi60daf9c2019-06-21 09:42:03 +01001515 *
1516 * Tasks can have a task-specific value requested from user-space, track
1517 * within each bucket the maximum value for tasks refcounted in it.
1518 * This "local max aggregation" allows to track the exact "requested" value
1519 * for each bucket when all its RUNNABLE tasks require the same clamp.
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001520 */
1521static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001522 enum uclamp_id clamp_id)
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001523{
1524 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1525 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1526 struct uclamp_bucket *bucket;
1527
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05001528 lockdep_assert_rq_held(rq);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001529
Patrick Bellasie8f14172019-06-21 09:42:05 +01001530 /* Update task effective clamp */
1531 p->uclamp[clamp_id] = uclamp_eff_get(p, clamp_id);
1532
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001533 bucket = &uc_rq->bucket[uc_se->bucket_id];
1534 bucket->tasks++;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001535 uc_se->active = true;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001536
Patrick Bellasie4961872019-06-21 09:42:04 +01001537 uclamp_idle_reset(rq, clamp_id, uc_se->value);
1538
Patrick Bellasi60daf9c2019-06-21 09:42:03 +01001539 /*
1540 * Local max aggregation: rq buckets always track the max
1541 * "requested" clamp value of its RUNNABLE tasks.
1542 */
1543 if (bucket->tasks == 1 || uc_se->value > bucket->value)
1544 bucket->value = uc_se->value;
1545
Qais Yousef24422602022-08-04 15:36:03 +01001546 if (uc_se->value > uclamp_rq_get(rq, clamp_id))
1547 uclamp_rq_set(rq, clamp_id, uc_se->value);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001548}
1549
1550/*
1551 * When a task is dequeued from a rq, the clamp bucket refcounted by the task
1552 * is released. If this is the last task reference counting the rq's max
1553 * active clamp value, then the rq's clamp value is updated.
1554 *
1555 * Both refcounted tasks and rq's cached clamp values are expected to be
1556 * always valid. If it's detected they are not, as defensive programming,
1557 * enforce the expected state and warn.
1558 */
1559static inline void uclamp_rq_dec_id(struct rq *rq, struct task_struct *p,
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001560 enum uclamp_id clamp_id)
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001561{
1562 struct uclamp_rq *uc_rq = &rq->uclamp[clamp_id];
1563 struct uclamp_se *uc_se = &p->uclamp[clamp_id];
1564 struct uclamp_bucket *bucket;
Patrick Bellasie4961872019-06-21 09:42:04 +01001565 unsigned int bkt_clamp;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001566 unsigned int rq_clamp;
1567
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05001568 lockdep_assert_rq_held(rq);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001569
Qais Yousef46609ce2020-06-30 12:21:23 +01001570 /*
1571 * If sched_uclamp_used was enabled after task @p was enqueued,
1572 * we could end up with unbalanced call to uclamp_rq_dec_id().
1573 *
1574 * In this case the uc_se->active flag should be false since no uclamp
1575 * accounting was performed at enqueue time and we can just return
1576 * here.
1577 *
Tal Zussmanb19a8882020-11-12 19:51:56 -05001578 * Need to be careful of the following enqueue/dequeue ordering
Qais Yousef46609ce2020-06-30 12:21:23 +01001579 * problem too
1580 *
1581 * enqueue(taskA)
1582 * // sched_uclamp_used gets enabled
1583 * enqueue(taskB)
1584 * dequeue(taskA)
Tal Zussmanb19a8882020-11-12 19:51:56 -05001585 * // Must not decrement bucket->tasks here
Qais Yousef46609ce2020-06-30 12:21:23 +01001586 * dequeue(taskB)
1587 *
1588 * where we could end up with stale data in uc_se and
1589 * bucket[uc_se->bucket_id].
1590 *
1591 * The following check here eliminates the possibility of such race.
1592 */
1593 if (unlikely(!uc_se->active))
1594 return;
1595
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001596 bucket = &uc_rq->bucket[uc_se->bucket_id];
Qais Yousef46609ce2020-06-30 12:21:23 +01001597
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001598 SCHED_WARN_ON(!bucket->tasks);
1599 if (likely(bucket->tasks))
1600 bucket->tasks--;
Qais Yousef46609ce2020-06-30 12:21:23 +01001601
Patrick Bellasie8f14172019-06-21 09:42:05 +01001602 uc_se->active = false;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001603
Patrick Bellasi60daf9c2019-06-21 09:42:03 +01001604 /*
1605 * Keep "local max aggregation" simple and accept to (possibly)
1606 * overboost some RUNNABLE tasks in the same bucket.
1607 * The rq clamp bucket value is reset to its base value whenever
1608 * there are no more RUNNABLE tasks refcounting it.
1609 */
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001610 if (likely(bucket->tasks))
1611 return;
1612
Qais Yousef24422602022-08-04 15:36:03 +01001613 rq_clamp = uclamp_rq_get(rq, clamp_id);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001614 /*
1615 * Defensive programming: this should never happen. If it happens,
1616 * e.g. due to future modification, warn and fixup the expected value.
1617 */
1618 SCHED_WARN_ON(bucket->value > rq_clamp);
Patrick Bellasie4961872019-06-21 09:42:04 +01001619 if (bucket->value >= rq_clamp) {
1620 bkt_clamp = uclamp_rq_max_value(rq, clamp_id, uc_se->value);
Qais Yousef24422602022-08-04 15:36:03 +01001621 uclamp_rq_set(rq, clamp_id, bkt_clamp);
Patrick Bellasie4961872019-06-21 09:42:04 +01001622 }
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001623}
1624
1625static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p)
1626{
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001627 enum uclamp_id clamp_id;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001628
Qais Yousef46609ce2020-06-30 12:21:23 +01001629 /*
1630 * Avoid any overhead until uclamp is actually used by the userspace.
1631 *
1632 * The condition is constructed such that a NOP is generated when
1633 * sched_uclamp_used is disabled.
1634 */
1635 if (!static_branch_unlikely(&sched_uclamp_used))
1636 return;
1637
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001638 if (unlikely(!p->sched_class->uclamp_enabled))
1639 return;
1640
1641 for_each_clamp_id(clamp_id)
1642 uclamp_rq_inc_id(rq, p, clamp_id);
Patrick Bellasie4961872019-06-21 09:42:04 +01001643
1644 /* Reset clamp idle holding when there is one RUNNABLE task */
1645 if (rq->uclamp_flags & UCLAMP_FLAG_IDLE)
1646 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001647}
1648
1649static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p)
1650{
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001651 enum uclamp_id clamp_id;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001652
Qais Yousef46609ce2020-06-30 12:21:23 +01001653 /*
1654 * Avoid any overhead until uclamp is actually used by the userspace.
1655 *
1656 * The condition is constructed such that a NOP is generated when
1657 * sched_uclamp_used is disabled.
1658 */
1659 if (!static_branch_unlikely(&sched_uclamp_used))
1660 return;
1661
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001662 if (unlikely(!p->sched_class->uclamp_enabled))
1663 return;
1664
1665 for_each_clamp_id(clamp_id)
1666 uclamp_rq_dec_id(rq, p, clamp_id);
1667}
1668
Quentin Perretca4984a2021-08-05 11:21:53 +01001669static inline void uclamp_rq_reinc_id(struct rq *rq, struct task_struct *p,
1670 enum uclamp_id clamp_id)
1671{
1672 if (!p->uclamp[clamp_id].active)
1673 return;
1674
1675 uclamp_rq_dec_id(rq, p, clamp_id);
1676 uclamp_rq_inc_id(rq, p, clamp_id);
1677
1678 /*
1679 * Make sure to clear the idle flag if we've transiently reached 0
1680 * active tasks on rq.
1681 */
1682 if (clamp_id == UCLAMP_MAX && (rq->uclamp_flags & UCLAMP_FLAG_IDLE))
1683 rq->uclamp_flags &= ~UCLAMP_FLAG_IDLE;
1684}
1685
Patrick Bellasibabbe172019-08-22 14:28:10 +01001686static inline void
Qais Yousef0213b702021-06-17 17:51:55 +01001687uclamp_update_active(struct task_struct *p)
Patrick Bellasibabbe172019-08-22 14:28:10 +01001688{
Qais Yousef0213b702021-06-17 17:51:55 +01001689 enum uclamp_id clamp_id;
Patrick Bellasibabbe172019-08-22 14:28:10 +01001690 struct rq_flags rf;
1691 struct rq *rq;
1692
1693 /*
1694 * Lock the task and the rq where the task is (or was) queued.
1695 *
1696 * We might lock the (previous) rq of a !RUNNABLE task, but that's the
1697 * price to pay to safely serialize util_{min,max} updates with
1698 * enqueues, dequeues and migration operations.
1699 * This is the same locking schema used by __set_cpus_allowed_ptr().
1700 */
1701 rq = task_rq_lock(p, &rf);
1702
1703 /*
1704 * Setting the clamp bucket is serialized by task_rq_lock().
1705 * If the task is not yet RUNNABLE and its task_struct is not
1706 * affecting a valid clamp bucket, the next time it's enqueued,
1707 * it will already see the updated clamp bucket value.
1708 */
Quentin Perretca4984a2021-08-05 11:21:53 +01001709 for_each_clamp_id(clamp_id)
1710 uclamp_rq_reinc_id(rq, p, clamp_id);
Patrick Bellasibabbe172019-08-22 14:28:10 +01001711
1712 task_rq_unlock(rq, p, &rf);
1713}
1714
Qais Yousefe3b8b6a2019-11-05 11:22:12 +00001715#ifdef CONFIG_UCLAMP_TASK_GROUP
Patrick Bellasibabbe172019-08-22 14:28:10 +01001716static inline void
Qais Yousef0213b702021-06-17 17:51:55 +01001717uclamp_update_active_tasks(struct cgroup_subsys_state *css)
Patrick Bellasibabbe172019-08-22 14:28:10 +01001718{
1719 struct css_task_iter it;
1720 struct task_struct *p;
Patrick Bellasibabbe172019-08-22 14:28:10 +01001721
1722 css_task_iter_start(css, 0, &it);
Qais Yousef0213b702021-06-17 17:51:55 +01001723 while ((p = css_task_iter_next(&it)))
1724 uclamp_update_active(p);
Patrick Bellasibabbe172019-08-22 14:28:10 +01001725 css_task_iter_end(&it);
1726}
1727
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001728static void cpu_util_update_eff(struct cgroup_subsys_state *css);
YueHaibing494dcdf2022-04-27 21:10:02 +08001729#endif
1730
1731#ifdef CONFIG_SYSCTL
1732#ifdef CONFIG_UCLAMP_TASK
1733#ifdef CONFIG_UCLAMP_TASK_GROUP
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001734static void uclamp_update_root_tg(void)
1735{
1736 struct task_group *tg = &root_task_group;
1737
1738 uclamp_se_set(&tg->uclamp_req[UCLAMP_MIN],
1739 sysctl_sched_uclamp_util_min, false);
1740 uclamp_se_set(&tg->uclamp_req[UCLAMP_MAX],
1741 sysctl_sched_uclamp_util_max, false);
1742
1743 rcu_read_lock();
1744 cpu_util_update_eff(&root_task_group.css);
1745 rcu_read_unlock();
1746}
1747#else
1748static void uclamp_update_root_tg(void) { }
1749#endif
1750
YueHaibing494dcdf2022-04-27 21:10:02 +08001751static void uclamp_sync_util_min_rt_default(void)
1752{
1753 struct task_struct *g, *p;
1754
1755 /*
1756 * copy_process() sysctl_uclamp
1757 * uclamp_min_rt = X;
1758 * write_lock(&tasklist_lock) read_lock(&tasklist_lock)
1759 * // link thread smp_mb__after_spinlock()
1760 * write_unlock(&tasklist_lock) read_unlock(&tasklist_lock);
1761 * sched_post_fork() for_each_process_thread()
1762 * __uclamp_sync_rt() __uclamp_sync_rt()
1763 *
1764 * Ensures that either sched_post_fork() will observe the new
1765 * uclamp_min_rt or for_each_process_thread() will observe the new
1766 * task.
1767 */
1768 read_lock(&tasklist_lock);
1769 smp_mb__after_spinlock();
1770 read_unlock(&tasklist_lock);
1771
1772 rcu_read_lock();
1773 for_each_process_thread(g, p)
1774 uclamp_update_util_min_rt_default(p);
1775 rcu_read_unlock();
1776}
1777
Zhen Ni3267e012022-02-15 19:46:02 +08001778static int sysctl_sched_uclamp_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02001779 void *buffer, size_t *lenp, loff_t *ppos)
Patrick Bellasie8f14172019-06-21 09:42:05 +01001780{
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001781 bool update_root_tg = false;
Qais Yousef13685c42020-07-16 12:03:45 +01001782 int old_min, old_max, old_min_rt;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001783 int result;
1784
Patrick Bellasi2480c092019-08-22 14:28:06 +01001785 mutex_lock(&uclamp_mutex);
Patrick Bellasie8f14172019-06-21 09:42:05 +01001786 old_min = sysctl_sched_uclamp_util_min;
1787 old_max = sysctl_sched_uclamp_util_max;
Qais Yousef13685c42020-07-16 12:03:45 +01001788 old_min_rt = sysctl_sched_uclamp_util_min_rt_default;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001789
1790 result = proc_dointvec(table, write, buffer, lenp, ppos);
1791 if (result)
1792 goto undo;
1793 if (!write)
1794 goto done;
1795
1796 if (sysctl_sched_uclamp_util_min > sysctl_sched_uclamp_util_max ||
Qais Yousef13685c42020-07-16 12:03:45 +01001797 sysctl_sched_uclamp_util_max > SCHED_CAPACITY_SCALE ||
1798 sysctl_sched_uclamp_util_min_rt_default > SCHED_CAPACITY_SCALE) {
1799
Patrick Bellasie8f14172019-06-21 09:42:05 +01001800 result = -EINVAL;
1801 goto undo;
1802 }
1803
1804 if (old_min != sysctl_sched_uclamp_util_min) {
1805 uclamp_se_set(&uclamp_default[UCLAMP_MIN],
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001806 sysctl_sched_uclamp_util_min, false);
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001807 update_root_tg = true;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001808 }
1809 if (old_max != sysctl_sched_uclamp_util_max) {
1810 uclamp_se_set(&uclamp_default[UCLAMP_MAX],
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001811 sysctl_sched_uclamp_util_max, false);
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001812 update_root_tg = true;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001813 }
1814
Qais Yousef46609ce2020-06-30 12:21:23 +01001815 if (update_root_tg) {
1816 static_branch_enable(&sched_uclamp_used);
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001817 uclamp_update_root_tg();
Qais Yousef46609ce2020-06-30 12:21:23 +01001818 }
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001819
Qais Yousef13685c42020-07-16 12:03:45 +01001820 if (old_min_rt != sysctl_sched_uclamp_util_min_rt_default) {
1821 static_branch_enable(&sched_uclamp_used);
1822 uclamp_sync_util_min_rt_default();
1823 }
Patrick Bellasie8f14172019-06-21 09:42:05 +01001824
1825 /*
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001826 * We update all RUNNABLE tasks only when task groups are in use.
1827 * Otherwise, keep it simple and do just a lazy update at each next
1828 * task enqueue time.
Patrick Bellasie8f14172019-06-21 09:42:05 +01001829 */
Patrick Bellasi7274a5c2019-08-22 14:28:08 +01001830
Patrick Bellasie8f14172019-06-21 09:42:05 +01001831 goto done;
1832
1833undo:
1834 sysctl_sched_uclamp_util_min = old_min;
1835 sysctl_sched_uclamp_util_max = old_max;
Qais Yousef13685c42020-07-16 12:03:45 +01001836 sysctl_sched_uclamp_util_min_rt_default = old_min_rt;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001837done:
Patrick Bellasi2480c092019-08-22 14:28:06 +01001838 mutex_unlock(&uclamp_mutex);
Patrick Bellasie8f14172019-06-21 09:42:05 +01001839
1840 return result;
1841}
YueHaibing494dcdf2022-04-27 21:10:02 +08001842#endif
1843#endif
Patrick Bellasie8f14172019-06-21 09:42:05 +01001844
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001845static int uclamp_validate(struct task_struct *p,
1846 const struct sched_attr *attr)
1847{
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001848 int util_min = p->uclamp_req[UCLAMP_MIN].value;
1849 int util_max = p->uclamp_req[UCLAMP_MAX].value;
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001850
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001851 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN) {
1852 util_min = attr->sched_util_min;
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001853
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001854 if (util_min + 1 > SCHED_CAPACITY_SCALE + 1)
1855 return -EINVAL;
1856 }
1857
1858 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX) {
1859 util_max = attr->sched_util_max;
1860
1861 if (util_max + 1 > SCHED_CAPACITY_SCALE + 1)
1862 return -EINVAL;
1863 }
1864
1865 if (util_min != -1 && util_max != -1 && util_min > util_max)
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001866 return -EINVAL;
1867
Qais Yousefe65855a2020-07-16 12:03:47 +01001868 /*
1869 * We have valid uclamp attributes; make sure uclamp is enabled.
1870 *
1871 * We need to do that here, because enabling static branches is a
1872 * blocking operation which obviously cannot be done while holding
1873 * scheduler locks.
1874 */
1875 static_branch_enable(&sched_uclamp_used);
1876
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001877 return 0;
1878}
1879
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001880static bool uclamp_reset(const struct sched_attr *attr,
1881 enum uclamp_id clamp_id,
1882 struct uclamp_se *uc_se)
1883{
1884 /* Reset on sched class change for a non user-defined clamp value. */
1885 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)) &&
1886 !uc_se->user_defined)
1887 return true;
1888
1889 /* Reset on sched_util_{min,max} == -1. */
1890 if (clamp_id == UCLAMP_MIN &&
1891 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1892 attr->sched_util_min == -1) {
1893 return true;
1894 }
1895
1896 if (clamp_id == UCLAMP_MAX &&
1897 attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1898 attr->sched_util_max == -1) {
1899 return true;
1900 }
1901
1902 return false;
1903}
1904
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001905static void __setscheduler_uclamp(struct task_struct *p,
1906 const struct sched_attr *attr)
1907{
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001908 enum uclamp_id clamp_id;
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001909
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001910 for_each_clamp_id(clamp_id) {
1911 struct uclamp_se *uc_se = &p->uclamp_req[clamp_id];
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001912 unsigned int value;
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001913
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001914 if (!uclamp_reset(attr, clamp_id, uc_se))
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001915 continue;
1916
Qais Yousef13685c42020-07-16 12:03:45 +01001917 /*
1918 * RT by default have a 100% boost value that could be modified
1919 * at runtime.
1920 */
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001921 if (unlikely(rt_task(p) && clamp_id == UCLAMP_MIN))
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001922 value = sysctl_sched_uclamp_util_min_rt_default;
Qais Yousef13685c42020-07-16 12:03:45 +01001923 else
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001924 value = uclamp_none(clamp_id);
1925
1926 uclamp_se_set(uc_se, value, false);
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001927
Patrick Bellasi1a00d992019-06-21 09:42:09 +01001928 }
1929
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001930 if (likely(!(attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)))
1931 return;
1932
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001933 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MIN &&
1934 attr->sched_util_min != -1) {
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001935 uclamp_se_set(&p->uclamp_req[UCLAMP_MIN],
1936 attr->sched_util_min, true);
1937 }
1938
Dietmar Eggemann480a6ca2020-11-13 12:34:54 +01001939 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP_MAX &&
1940 attr->sched_util_max != -1) {
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001941 uclamp_se_set(&p->uclamp_req[UCLAMP_MAX],
1942 attr->sched_util_max, true);
1943 }
1944}
1945
Patrick Bellasie8f14172019-06-21 09:42:05 +01001946static void uclamp_fork(struct task_struct *p)
1947{
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001948 enum uclamp_id clamp_id;
Patrick Bellasie8f14172019-06-21 09:42:05 +01001949
Qais Yousef13685c42020-07-16 12:03:45 +01001950 /*
1951 * We don't need to hold task_rq_lock() when updating p->uclamp_* here
1952 * as the task is still at its early fork stages.
1953 */
Patrick Bellasie8f14172019-06-21 09:42:05 +01001954 for_each_clamp_id(clamp_id)
1955 p->uclamp[clamp_id].active = false;
Patrick Bellasia87498a2019-06-21 09:42:08 +01001956
1957 if (likely(!p->sched_reset_on_fork))
1958 return;
1959
1960 for_each_clamp_id(clamp_id) {
Quentin Perreteaf5a922020-04-16 09:59:56 +01001961 uclamp_se_set(&p->uclamp_req[clamp_id],
1962 uclamp_none(clamp_id), false);
Patrick Bellasia87498a2019-06-21 09:42:08 +01001963 }
Patrick Bellasie8f14172019-06-21 09:42:05 +01001964}
1965
Qais Yousef13685c42020-07-16 12:03:45 +01001966static void uclamp_post_fork(struct task_struct *p)
1967{
1968 uclamp_update_util_min_rt_default(p);
1969}
1970
Qais Yousefd81ae8aa2020-06-30 12:21:22 +01001971static void __init init_uclamp_rq(struct rq *rq)
1972{
1973 enum uclamp_id clamp_id;
1974 struct uclamp_rq *uc_rq = rq->uclamp;
1975
1976 for_each_clamp_id(clamp_id) {
1977 uc_rq[clamp_id] = (struct uclamp_rq) {
1978 .value = uclamp_none(clamp_id)
1979 };
1980 }
1981
Qais Yousef315c4f82021-12-02 11:20:33 +00001982 rq->uclamp_flags = UCLAMP_FLAG_IDLE;
Qais Yousefd81ae8aa2020-06-30 12:21:22 +01001983}
1984
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001985static void __init init_uclamp(void)
1986{
Patrick Bellasie8f14172019-06-21 09:42:05 +01001987 struct uclamp_se uc_max = {};
Patrick Bellasi0413d7f2019-08-22 14:28:11 +01001988 enum uclamp_id clamp_id;
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001989 int cpu;
1990
Qais Yousefd81ae8aa2020-06-30 12:21:22 +01001991 for_each_possible_cpu(cpu)
1992 init_uclamp_rq(cpu_rq(cpu));
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001993
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001994 for_each_clamp_id(clamp_id) {
Patrick Bellasie8f14172019-06-21 09:42:05 +01001995 uclamp_se_set(&init_task.uclamp_req[clamp_id],
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01001996 uclamp_none(clamp_id), false);
Patrick Bellasi69842cb2019-06-21 09:42:02 +01001997 }
Patrick Bellasie8f14172019-06-21 09:42:05 +01001998
1999 /* System defaults allow max clamp values for both indexes */
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01002000 uclamp_se_set(&uc_max, uclamp_none(UCLAMP_MAX), false);
Patrick Bellasi2480c092019-08-22 14:28:06 +01002001 for_each_clamp_id(clamp_id) {
Patrick Bellasie8f14172019-06-21 09:42:05 +01002002 uclamp_default[clamp_id] = uc_max;
Patrick Bellasi2480c092019-08-22 14:28:06 +01002003#ifdef CONFIG_UCLAMP_TASK_GROUP
2004 root_task_group.uclamp_req[clamp_id] = uc_max;
Patrick Bellasi0b60ba22019-08-22 14:28:07 +01002005 root_task_group.uclamp[clamp_id] = uc_max;
Patrick Bellasi2480c092019-08-22 14:28:06 +01002006#endif
2007 }
Patrick Bellasi69842cb2019-06-21 09:42:02 +01002008}
2009
2010#else /* CONFIG_UCLAMP_TASK */
2011static inline void uclamp_rq_inc(struct rq *rq, struct task_struct *p) { }
2012static inline void uclamp_rq_dec(struct rq *rq, struct task_struct *p) { }
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01002013static inline int uclamp_validate(struct task_struct *p,
2014 const struct sched_attr *attr)
2015{
2016 return -EOPNOTSUPP;
2017}
2018static void __setscheduler_uclamp(struct task_struct *p,
2019 const struct sched_attr *attr) { }
Patrick Bellasie8f14172019-06-21 09:42:05 +01002020static inline void uclamp_fork(struct task_struct *p) { }
Qais Yousef13685c42020-07-16 12:03:45 +01002021static inline void uclamp_post_fork(struct task_struct *p) { }
Patrick Bellasi69842cb2019-06-21 09:42:02 +01002022static inline void init_uclamp(void) { }
2023#endif /* CONFIG_UCLAMP_TASK */
2024
Marcelo Tosattia1dfb632021-05-13 01:29:22 +02002025bool sched_task_on_rq(struct task_struct *p)
2026{
2027 return task_on_rq_queued(p);
2028}
2029
Kees Cook42a20f82021-09-29 15:02:14 -07002030unsigned long get_wchan(struct task_struct *p)
2031{
2032 unsigned long ip = 0;
2033 unsigned int state;
2034
2035 if (!p || p == current)
2036 return 0;
2037
2038 /* Only get wchan if task is blocked and we can keep it that way. */
2039 raw_spin_lock_irq(&p->pi_lock);
2040 state = READ_ONCE(p->__state);
2041 smp_rmb(); /* see try_to_wake_up() */
2042 if (state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq)
2043 ip = __get_wchan(p);
2044 raw_spin_unlock_irq(&p->pi_lock);
2045
2046 return ip;
2047}
2048
Peter Zijlstra1de64442015-09-30 17:44:13 +02002049static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
Gregory Haskins2087a1a2008-06-27 14:30:00 -06002050{
Peter Zijlstra0a67d1e2016-10-04 16:29:45 +02002051 if (!(flags & ENQUEUE_NOCLOCK))
2052 update_rq_clock(rq);
2053
Johannes Weinereb414682018-10-26 15:06:27 -07002054 if (!(flags & ENQUEUE_RESTORE)) {
Peter Zijlstra4e29fb72021-05-04 22:43:45 +02002055 sched_info_enqueue(rq, p);
Chengming Zhou52b33d82022-09-26 16:19:31 +08002056 psi_enqueue(p, (flags & ENQUEUE_WAKEUP) && !(flags & ENQUEUE_MIGRATED));
Johannes Weinereb414682018-10-26 15:06:27 -07002057 }
Peter Zijlstra0a67d1e2016-10-04 16:29:45 +02002058
Patrick Bellasi69842cb2019-06-21 09:42:02 +01002059 uclamp_rq_inc(rq, p);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01002060 p->sched_class->enqueue_task(rq, p, flags);
Peter Zijlstra8a311c72020-11-17 18:19:36 -05002061
2062 if (sched_core_enabled(rq))
2063 sched_core_enqueue(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02002064}
2065
Peter Zijlstra1de64442015-09-30 17:44:13 +02002066static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnardd41f592007-07-09 18:51:59 +02002067{
Peter Zijlstra8a311c72020-11-17 18:19:36 -05002068 if (sched_core_enabled(rq))
Josh Don4feee7d12021-10-18 13:34:28 -07002069 sched_core_dequeue(rq, p, flags);
Peter Zijlstra8a311c72020-11-17 18:19:36 -05002070
Peter Zijlstra0a67d1e2016-10-04 16:29:45 +02002071 if (!(flags & DEQUEUE_NOCLOCK))
2072 update_rq_clock(rq);
2073
Johannes Weinereb414682018-10-26 15:06:27 -07002074 if (!(flags & DEQUEUE_SAVE)) {
Peter Zijlstra4e29fb72021-05-04 22:43:45 +02002075 sched_info_dequeue(rq, p);
Johannes Weinereb414682018-10-26 15:06:27 -07002076 psi_dequeue(p, flags & DEQUEUE_SLEEP);
2077 }
Peter Zijlstra0a67d1e2016-10-04 16:29:45 +02002078
Patrick Bellasi69842cb2019-06-21 09:42:02 +01002079 uclamp_rq_dec(rq, p);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01002080 p->sched_class->dequeue_task(rq, p, flags);
Ingo Molnar71f8bd42007-07-09 18:51:59 +02002081}
2082
Peter Zijlstra029632f2011-10-25 10:00:11 +02002083void activate_task(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002084{
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01002085 enqueue_task(rq, p, flags);
Peter Zijlstra7dd77882019-04-09 09:59:05 +02002086
2087 p->on_rq = TASK_ON_RQ_QUEUED;
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002088}
2089
Peter Zijlstra029632f2011-10-25 10:00:11 +02002090void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002091{
Peter Zijlstra7dd77882019-04-09 09:59:05 +02002092 p->on_rq = (flags & DEQUEUE_SLEEP) ? 0 : TASK_ON_RQ_MIGRATING;
2093
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01002094 dequeue_task(rq, p, flags);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +01002095}
2096
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02002097static inline int __normal_prio(int policy, int rt_prio, int nice)
Ingo Molnar14531182007-07-09 18:51:59 +02002098{
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02002099 int prio;
2100
2101 if (dl_policy(policy))
2102 prio = MAX_DL_PRIO - 1;
2103 else if (rt_policy(policy))
2104 prio = MAX_RT_PRIO - 1 - rt_prio;
2105 else
2106 prio = NICE_TO_PRIO(nice);
2107
2108 return prio;
Ingo Molnar14531182007-07-09 18:51:59 +02002109}
2110
2111/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07002112 * Calculate the expected normal priority: i.e. priority
2113 * without taking RT-inheritance into account. Might be
2114 * boosted by interactivity modifiers. Changes upon fork,
2115 * setprio syscalls, and whenever the interactivity
2116 * estimator recalculates.
2117 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002118static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07002119{
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02002120 return __normal_prio(p->policy, p->rt_priority, PRIO_TO_NICE(p->static_prio));
Ingo Molnarb29739f2006-06-27 02:54:51 -07002121}
2122
2123/*
2124 * Calculate the current priority, i.e. the priority
2125 * taken into account by the scheduler. This value might
2126 * be boosted by RT tasks, or might be boosted by
2127 * interactivity modifiers. Will be RT if the task got
2128 * RT-boosted. If not then it returns p->normal_prio.
2129 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002130static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07002131{
2132 p->normal_prio = normal_prio(p);
2133 /*
2134 * If we are RT tasks or we were boosted to RT priority,
2135 * keep the priority unchanged. Otherwise, update priority
2136 * to the normal priority:
2137 */
2138 if (!rt_prio(p->prio))
2139 return p->normal_prio;
2140 return p->prio;
2141}
2142
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143/**
2144 * task_curr - is this task currently executing on a CPU?
2145 * @p: the task in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02002146 *
2147 * Return: 1 if the task is currently executing. 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002149inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150{
2151 return cpu_curr(task_cpu(p)) == p;
2152}
2153
Kirill Tkhai67dfa1b2014-10-27 17:40:52 +03002154/*
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02002155 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
2156 * use the balance_callback list if you want balancing.
2157 *
2158 * this means any call to check_class_changed() must be followed by a call to
2159 * balance_callback().
Kirill Tkhai67dfa1b2014-10-27 17:40:52 +03002160 */
Steven Rostedtcb469842008-01-25 21:08:22 +01002161static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2162 const struct sched_class *prev_class,
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002163 int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01002164{
2165 if (prev_class != p->sched_class) {
2166 if (prev_class->switched_from)
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002167 prev_class->switched_from(rq, p);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02002168
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002169 p->sched_class->switched_to(rq, p);
Dario Faggioli2d3d8912013-11-07 14:43:44 +01002170 } else if (oldprio != p->prio || dl_task(p))
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002171 p->sched_class->prio_changed(rq, p, oldprio);
Steven Rostedtcb469842008-01-25 21:08:22 +01002172}
2173
Peter Zijlstra029632f2011-10-25 10:00:11 +02002174void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01002175{
Kirill Tkhaiaa93cd52019-12-19 16:44:55 -05002176 if (p->sched_class == rq->curr->sched_class)
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01002177 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
Peter Zijlstra546a3fe2022-05-17 13:46:54 +02002178 else if (sched_class_above(p->sched_class, rq->curr->sched_class))
Kirill Tkhaiaa93cd52019-12-19 16:44:55 -05002179 resched_curr(rq);
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01002180
2181 /*
2182 * A queue event has occurred, and we're going to schedule. In
2183 * this case, we can save a useless back to back clock update.
2184 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002185 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
Davidlohr Buesoadcc8da2018-04-04 09:15:39 -07002186 rq_clock_skip_update(rq);
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01002187}
2188
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189#ifdef CONFIG_SMP
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002190
Peter Zijlstraaf449902020-09-17 10:38:30 +02002191static void
Waiman Long713a2e22022-09-22 14:00:40 -04002192__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx);
Peter Zijlstraaf449902020-09-17 10:38:30 +02002193
2194static int __set_cpus_allowed_ptr(struct task_struct *p,
Waiman Long713a2e22022-09-22 14:00:40 -04002195 struct affinity_context *ctx);
Peter Zijlstraaf449902020-09-17 10:38:30 +02002196
2197static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
2198{
Waiman Long713a2e22022-09-22 14:00:40 -04002199 struct affinity_context ac = {
2200 .new_mask = cpumask_of(rq->cpu),
2201 .flags = SCA_MIGRATE_DISABLE,
2202 };
2203
Peter Zijlstraaf449902020-09-17 10:38:30 +02002204 if (likely(!p->migration_disabled))
2205 return;
2206
2207 if (p->cpus_ptr != &p->cpus_mask)
2208 return;
2209
2210 /*
2211 * Violates locking rules! see comment in __do_set_cpus_allowed().
2212 */
Waiman Long713a2e22022-09-22 14:00:40 -04002213 __do_set_cpus_allowed(p, &ac);
Peter Zijlstraaf449902020-09-17 10:38:30 +02002214}
2215
2216void migrate_disable(void)
2217{
Thomas Gleixner3015ef42020-08-26 14:08:10 +02002218 struct task_struct *p = current;
Peter Zijlstraaf449902020-09-17 10:38:30 +02002219
Thomas Gleixner3015ef42020-08-26 14:08:10 +02002220 if (p->migration_disabled) {
2221 p->migration_disabled++;
2222 return;
2223 }
2224
2225 preempt_disable();
2226 this_rq()->nr_pinned++;
2227 p->migration_disabled = 1;
2228 preempt_enable();
Peter Zijlstraaf449902020-09-17 10:38:30 +02002229}
2230EXPORT_SYMBOL_GPL(migrate_disable);
2231
2232void migrate_enable(void)
2233{
2234 struct task_struct *p = current;
Waiman Long713a2e22022-09-22 14:00:40 -04002235 struct affinity_context ac = {
2236 .new_mask = &p->cpus_mask,
2237 .flags = SCA_MIGRATE_ENABLE,
2238 };
Peter Zijlstraaf449902020-09-17 10:38:30 +02002239
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002240 if (p->migration_disabled > 1) {
2241 p->migration_disabled--;
Peter Zijlstraaf449902020-09-17 10:38:30 +02002242 return;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002243 }
Peter Zijlstraaf449902020-09-17 10:38:30 +02002244
Sebastian Andrzej Siewior9d0df372021-11-29 18:46:44 +01002245 if (WARN_ON_ONCE(!p->migration_disabled))
2246 return;
2247
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002248 /*
2249 * Ensure stop_task runs either before or after this, and that
2250 * __set_cpus_allowed_ptr(SCA_MIGRATE_ENABLE) doesn't schedule().
2251 */
2252 preempt_disable();
2253 if (p->cpus_ptr != &p->cpus_mask)
Waiman Long713a2e22022-09-22 14:00:40 -04002254 __set_cpus_allowed_ptr(p, &ac);
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002255 /*
2256 * Mustn't clear migration_disabled() until cpus_ptr points back at the
2257 * regular cpus_mask, otherwise things that race (eg.
2258 * select_fallback_rq) get confused.
2259 */
Peter Zijlstraaf449902020-09-17 10:38:30 +02002260 barrier();
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002261 p->migration_disabled = 0;
Thomas Gleixner3015ef42020-08-26 14:08:10 +02002262 this_rq()->nr_pinned--;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002263 preempt_enable();
Peter Zijlstraaf449902020-09-17 10:38:30 +02002264}
2265EXPORT_SYMBOL_GPL(migrate_enable);
2266
Thomas Gleixner3015ef42020-08-26 14:08:10 +02002267static inline bool rq_has_pinned_tasks(struct rq *rq)
2268{
2269 return rq->nr_pinned;
2270}
2271
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002272/*
Joel Savitzbee98532019-03-06 20:13:33 -05002273 * Per-CPU kthreads are allowed to run on !active && online CPUs, see
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002274 * __set_cpus_allowed_ptr() and select_fallback_rq().
2275 */
2276static inline bool is_cpu_allowed(struct task_struct *p, int cpu)
2277{
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01002278 /* When not in the task's cpumask, no point in looking further. */
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02002279 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002280 return false;
2281
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01002282 /* migrate_disabled() must be allowed to finish. */
2283 if (is_migration_disabled(p))
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002284 return cpu_online(cpu);
2285
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01002286 /* Non kernel threads are not allowed during either online or offline. */
2287 if (!(p->flags & PF_KTHREAD))
Will Deacon9ae606b2021-07-30 12:24:28 +01002288 return cpu_active(cpu) && task_cpu_possible(cpu, p);
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01002289
2290 /* KTHREAD_IS_PER_CPU is always allowed. */
2291 if (kthread_is_per_cpu(p))
2292 return cpu_online(cpu);
2293
2294 /* Regular kernel threads don't get to stay during offline. */
Peter Zijlstrab5c44772021-01-21 16:09:32 +01002295 if (cpu_dying(cpu))
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01002296 return false;
2297
2298 /* But are allowed during online. */
2299 return cpu_online(cpu);
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002300}
2301
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002302/*
2303 * This is how migration works:
2304 *
2305 * 1) we invoke migration_cpu_stop() on the target CPU using
2306 * stop_one_cpu().
2307 * 2) stopper starts to run (implicitly forcing the migrated thread
2308 * off the CPU)
2309 * 3) it checks whether the migrated task is still in the wrong runqueue.
2310 * 4) if it's in the wrong runqueue then the migration thread removes
2311 * it and puts it into the right queue.
2312 * 5) stopper completes and stop_one_cpu() returns and the migration
2313 * is done.
2314 */
2315
2316/*
2317 * move_queued_task - move a queued task to new rq.
2318 *
2319 * Returns (locked) new rq. Old rq's lock is released.
2320 */
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002321static struct rq *move_queued_task(struct rq *rq, struct rq_flags *rf,
2322 struct task_struct *p, int new_cpu)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002323{
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05002324 lockdep_assert_rq_held(rq);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002325
Peter Zijlstra58877d32020-07-02 14:52:11 +02002326 deactivate_task(rq, p, DEQUEUE_NOCLOCK);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002327 set_task_cpu(p, new_cpu);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002328 rq_unlock(rq, rf);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002329
2330 rq = cpu_rq(new_cpu);
2331
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002332 rq_lock(rq, rf);
Ingo Molnar09348d72022-08-11 08:54:52 +02002333 WARN_ON_ONCE(task_cpu(p) != new_cpu);
Peter Zijlstra58877d32020-07-02 14:52:11 +02002334 activate_task(rq, p, 0);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002335 check_preempt_curr(rq, p, 0);
2336
2337 return rq;
2338}
2339
2340struct migration_arg {
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002341 struct task_struct *task;
2342 int dest_cpu;
2343 struct set_affinity_pending *pending;
2344};
2345
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002346/*
2347 * @refs: number of wait_for_completion()
2348 * @stop_pending: is @stop_work in use
2349 */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002350struct set_affinity_pending {
2351 refcount_t refs;
Peter Zijlstra9e818892021-02-24 11:31:09 +01002352 unsigned int stop_pending;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002353 struct completion done;
2354 struct cpu_stop_work stop_work;
2355 struct migration_arg arg;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002356};
2357
2358/*
Ingo Molnard1ccc662017-02-01 11:46:42 +01002359 * Move (not current) task off this CPU, onto the destination CPU. We're doing
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002360 * this because either it can't run here any more (set_cpus_allowed()
2361 * away from this CPU, or CPU going down), or because we're
2362 * attempting to rebalance this task on exec (sched_exec).
2363 *
2364 * So we race with normal scheduler movements, but that's OK, as long
2365 * as the task is no longer on this CPU.
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002366 */
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002367static struct rq *__migrate_task(struct rq *rq, struct rq_flags *rf,
2368 struct task_struct *p, int dest_cpu)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002369{
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002370 /* Affinity changed (again). */
Peter Zijlstra175f0e22017-07-25 18:58:21 +02002371 if (!is_cpu_allowed(p, dest_cpu))
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002372 return rq;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002373
Peter Zijlstra15ff9912016-10-05 17:59:32 +02002374 update_rq_clock(rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002375 rq = move_queued_task(rq, rf, p, dest_cpu);
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002376
2377 return rq;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002378}
2379
2380/*
2381 * migration_cpu_stop - this will be executed by a highprio stopper thread
2382 * and performs thread migration by bumping thread off CPU then
2383 * 'pushing' onto another runqueue.
2384 */
2385static int migration_cpu_stop(void *data)
2386{
2387 struct migration_arg *arg = data;
Peter Zijlstrac20cf062021-02-24 11:50:39 +01002388 struct set_affinity_pending *pending = arg->pending;
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002389 struct task_struct *p = arg->task;
2390 struct rq *rq = this_rq();
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002391 bool complete = false;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002392 struct rq_flags rf;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002393
2394 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01002395 * The original target CPU might have gone down and we might
2396 * be on another CPU but it doesn't matter.
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002397 */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002398 local_irq_save(rf.flags);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002399 /*
2400 * We need to explicitly wake pending tasks before running
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02002401 * __migrate_task() such that we will not miss enforcing cpus_ptr
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002402 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
2403 */
Thomas Gleixner16bf5a52022-04-13 15:31:03 +02002404 flush_smp_call_function_queue();
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002405
2406 raw_spin_lock(&p->pi_lock);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02002407 rq_lock(rq, &rf);
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002408
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002409 /*
Valentin Schneidere1407492021-02-25 10:22:30 +01002410 * If we were passed a pending, then ->stop_pending was set, thus
2411 * p->migration_pending must have remained stable.
2412 */
2413 WARN_ON_ONCE(pending && pending != p->migration_pending);
2414
2415 /*
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002416 * If task_rq(p) != rq, it cannot be migrated here, because we're
2417 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
2418 * we're holding p->pi_lock.
2419 */
Cheng Chaobf89a302016-09-14 10:01:50 +08002420 if (task_rq(p) == rq) {
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002421 if (is_migration_disabled(p))
2422 goto out;
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02002423
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002424 if (pending) {
Valentin Schneidere1407492021-02-25 10:22:30 +01002425 p->migration_pending = NULL;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002426 complete = true;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002427
Peter Zijlstra3f1bc112021-02-24 11:21:35 +01002428 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask))
2429 goto out;
Peter Zijlstra3f1bc112021-02-24 11:21:35 +01002430 }
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002431
2432 if (task_on_rq_queued(p))
Valentin Schneider475ea6c2021-05-26 21:57:50 +01002433 rq = __migrate_task(rq, &rf, p, arg->dest_cpu);
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002434 else
Valentin Schneider475ea6c2021-05-26 21:57:50 +01002435 p->wake_cpu = arg->dest_cpu;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002436
Peter Zijlstra3f1bc112021-02-24 11:21:35 +01002437 /*
2438 * XXX __migrate_task() can fail, at which point we might end
2439 * up running on a dodgy CPU, AFAICT this can only happen
2440 * during CPU hotplug, at which point we'll get pushed out
2441 * anyway, so it's probably not a big deal.
2442 */
2443
Peter Zijlstrac20cf062021-02-24 11:50:39 +01002444 } else if (pending) {
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002445 /*
2446 * This happens when we get migrated between migrate_enable()'s
2447 * preempt_enable() and scheduling the stopper task. At that
2448 * point we're a regular task again and not current anymore.
2449 *
2450 * A !PREEMPT kernel has a giant hole here, which makes it far
2451 * more likely.
2452 */
2453
2454 /*
Valentin Schneiderd707faa2020-11-13 11:24:14 +00002455 * The task moved before the stopper got to run. We're holding
2456 * ->pi_lock, so the allowed mask is stable - if it got
2457 * somewhere allowed, we're done.
2458 */
Peter Zijlstrac20cf062021-02-24 11:50:39 +01002459 if (cpumask_test_cpu(task_cpu(p), p->cpus_ptr)) {
Valentin Schneidere1407492021-02-25 10:22:30 +01002460 p->migration_pending = NULL;
Valentin Schneiderd707faa2020-11-13 11:24:14 +00002461 complete = true;
2462 goto out;
2463 }
2464
2465 /*
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002466 * When migrate_enable() hits a rq mis-match we can't reliably
2467 * determine is_migration_disabled() and so have to chase after
2468 * it.
2469 */
Peter Zijlstra9e818892021-02-24 11:31:09 +01002470 WARN_ON_ONCE(!pending->stop_pending);
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002471 task_rq_unlock(rq, p, &rf);
2472 stop_one_cpu_nowait(task_cpu(p), migration_cpu_stop,
2473 &pending->arg, &pending->stop_work);
2474 return 0;
2475 }
2476out:
Peter Zijlstra9e818892021-02-24 11:31:09 +01002477 if (pending)
2478 pending->stop_pending = false;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002479 task_rq_unlock(rq, p, &rf);
2480
2481 if (complete)
2482 complete_all(&pending->done);
2483
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002484 return 0;
2485}
2486
Peter Zijlstraa7c81552020-09-28 17:06:07 +02002487int push_cpu_stop(void *arg)
2488{
2489 struct rq *lowest_rq = NULL, *rq = this_rq();
2490 struct task_struct *p = arg;
2491
2492 raw_spin_lock_irq(&p->pi_lock);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05002493 raw_spin_rq_lock(rq);
Peter Zijlstraa7c81552020-09-28 17:06:07 +02002494
2495 if (task_rq(p) != rq)
2496 goto out_unlock;
2497
2498 if (is_migration_disabled(p)) {
2499 p->migration_flags |= MDF_PUSH;
2500 goto out_unlock;
2501 }
2502
2503 p->migration_flags &= ~MDF_PUSH;
2504
2505 if (p->sched_class->find_lock_rq)
2506 lowest_rq = p->sched_class->find_lock_rq(p, rq);
2507
2508 if (!lowest_rq)
2509 goto out_unlock;
2510
2511 // XXX validate p is still the highest prio task
2512 if (task_rq(p) == rq) {
2513 deactivate_task(rq, p, 0);
2514 set_task_cpu(p, lowest_rq->cpu);
2515 activate_task(lowest_rq, p, 0);
2516 resched_curr(lowest_rq);
2517 }
2518
2519 double_unlock_balance(rq, lowest_rq);
2520
2521out_unlock:
2522 rq->push_busy = false;
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05002523 raw_spin_rq_unlock(rq);
Peter Zijlstraa7c81552020-09-28 17:06:07 +02002524 raw_spin_unlock_irq(&p->pi_lock);
2525
2526 put_task_struct(p);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002527 return 0;
2528}
2529
Peter Zijlstrac5b28032015-05-15 17:43:35 +02002530/*
2531 * sched_class::set_cpus_allowed must do the below, but is not required to
2532 * actually call this function.
2533 */
Waiman Long713a2e22022-09-22 14:00:40 -04002534void set_cpus_allowed_common(struct task_struct *p, struct affinity_context *ctx)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002535{
Waiman Long713a2e22022-09-22 14:00:40 -04002536 if (ctx->flags & (SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) {
2537 p->cpus_ptr = ctx->new_mask;
Peter Zijlstraaf449902020-09-17 10:38:30 +02002538 return;
2539 }
2540
Waiman Long713a2e22022-09-22 14:00:40 -04002541 cpumask_copy(&p->cpus_mask, ctx->new_mask);
2542 p->nr_cpus_allowed = cpumask_weight(ctx->new_mask);
Waiman Long8f9ea862022-09-22 14:00:38 -04002543
2544 /*
2545 * Swap in a new user_cpus_ptr if SCA_USER flag set
2546 */
2547 if (ctx->flags & SCA_USER)
2548 swap(p->user_cpus_ptr, ctx->user_mask);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002549}
2550
Peter Zijlstra9cfc3e12020-09-16 14:59:08 +02002551static void
Waiman Long713a2e22022-09-22 14:00:40 -04002552__do_set_cpus_allowed(struct task_struct *p, struct affinity_context *ctx)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002553{
Peter Zijlstra6c370672015-05-15 17:43:36 +02002554 struct rq *rq = task_rq(p);
2555 bool queued, running;
2556
Peter Zijlstraaf449902020-09-17 10:38:30 +02002557 /*
2558 * This here violates the locking rules for affinity, since we're only
2559 * supposed to change these variables while holding both rq->lock and
2560 * p->pi_lock.
2561 *
2562 * HOWEVER, it magically works, because ttwu() is the only code that
2563 * accesses these variables under p->pi_lock and only does so after
2564 * smp_cond_load_acquire(&p->on_cpu, !VAL), and we're in __schedule()
2565 * before finish_task().
2566 *
2567 * XXX do further audits, this smells like something putrid.
2568 */
Waiman Long713a2e22022-09-22 14:00:40 -04002569 if (ctx->flags & SCA_MIGRATE_DISABLE)
Peter Zijlstraaf449902020-09-17 10:38:30 +02002570 SCHED_WARN_ON(!p->on_cpu);
2571 else
2572 lockdep_assert_held(&p->pi_lock);
Peter Zijlstra6c370672015-05-15 17:43:36 +02002573
2574 queued = task_on_rq_queued(p);
2575 running = task_current(rq, p);
2576
2577 if (queued) {
2578 /*
2579 * Because __kthread_bind() calls this on blocked tasks without
2580 * holding rq->lock.
2581 */
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05002582 lockdep_assert_rq_held(rq);
Peter Zijlstra7a57f322017-02-21 14:47:02 +01002583 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
Peter Zijlstra6c370672015-05-15 17:43:36 +02002584 }
2585 if (running)
2586 put_prev_task(rq, p);
2587
Waiman Long713a2e22022-09-22 14:00:40 -04002588 p->sched_class->set_cpus_allowed(p, ctx);
Peter Zijlstra6c370672015-05-15 17:43:36 +02002589
Peter Zijlstra6c370672015-05-15 17:43:36 +02002590 if (queued)
Peter Zijlstra7134b3e2017-02-21 14:23:38 +01002591 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
Vincent Guittota399d232016-09-12 09:47:52 +02002592 if (running)
Peter Zijlstra03b7fad2019-05-29 20:36:41 +00002593 set_next_task(rq, p);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002594}
2595
Waiman Long851a7232022-09-22 14:00:41 -04002596/*
2597 * Used for kthread_bind() and select_fallback_rq(), in both cases the user
2598 * affinity (if any) should be destroyed too.
2599 */
Peter Zijlstra9cfc3e12020-09-16 14:59:08 +02002600void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
2601{
Waiman Long713a2e22022-09-22 14:00:40 -04002602 struct affinity_context ac = {
2603 .new_mask = new_mask,
Waiman Long851a7232022-09-22 14:00:41 -04002604 .user_mask = NULL,
2605 .flags = SCA_USER, /* clear the user requested mask */
Waiman Long713a2e22022-09-22 14:00:40 -04002606 };
Waiman Long9a5418b2022-12-30 23:11:20 -05002607 union cpumask_rcuhead {
2608 cpumask_t cpumask;
2609 struct rcu_head rcu;
2610 };
Waiman Long713a2e22022-09-22 14:00:40 -04002611
2612 __do_set_cpus_allowed(p, &ac);
Waiman Long9a5418b2022-12-30 23:11:20 -05002613
2614 /*
2615 * Because this is called with p->pi_lock held, it is not possible
2616 * to use kfree() here (when PREEMPT_RT=y), therefore punt to using
2617 * kfree_rcu().
2618 */
2619 kfree_rcu((union cpumask_rcuhead *)ac.user_mask, rcu);
2620}
2621
2622static cpumask_t *alloc_user_cpus_ptr(int node)
2623{
2624 /*
2625 * See do_set_cpus_allowed() above for the rcu_head usage.
2626 */
2627 int size = max_t(int, cpumask_size(), sizeof(struct rcu_head));
2628
2629 return kmalloc_node(size, GFP_KERNEL, node);
Peter Zijlstra9cfc3e12020-09-16 14:59:08 +02002630}
2631
Will Deaconb90ca8b2021-07-30 12:24:33 +01002632int dup_user_cpus_ptr(struct task_struct *dst, struct task_struct *src,
2633 int node)
2634{
Waiman Long87ca4f92022-12-30 23:11:19 -05002635 cpumask_t *user_mask;
Waiman Long8f9ea862022-09-22 14:00:38 -04002636 unsigned long flags;
2637
Waiman Long87ca4f92022-12-30 23:11:19 -05002638 /*
2639 * Always clear dst->user_cpus_ptr first as their user_cpus_ptr's
2640 * may differ by now due to racing.
2641 */
2642 dst->user_cpus_ptr = NULL;
2643
2644 /*
2645 * This check is racy and losing the race is a valid situation.
2646 * It is not worth the extra overhead of taking the pi_lock on
2647 * every fork/clone.
2648 */
2649 if (data_race(!src->user_cpus_ptr))
Will Deaconb90ca8b2021-07-30 12:24:33 +01002650 return 0;
2651
Waiman Long9a5418b2022-12-30 23:11:20 -05002652 user_mask = alloc_user_cpus_ptr(node);
Waiman Long87ca4f92022-12-30 23:11:19 -05002653 if (!user_mask)
Will Deaconb90ca8b2021-07-30 12:24:33 +01002654 return -ENOMEM;
2655
Waiman Long87ca4f92022-12-30 23:11:19 -05002656 /*
2657 * Use pi_lock to protect content of user_cpus_ptr
2658 *
2659 * Though unlikely, user_cpus_ptr can be reset to NULL by a concurrent
2660 * do_set_cpus_allowed().
2661 */
Waiman Long8f9ea862022-09-22 14:00:38 -04002662 raw_spin_lock_irqsave(&src->pi_lock, flags);
Waiman Long87ca4f92022-12-30 23:11:19 -05002663 if (src->user_cpus_ptr) {
2664 swap(dst->user_cpus_ptr, user_mask);
2665 cpumask_copy(dst->user_cpus_ptr, src->user_cpus_ptr);
2666 }
Waiman Long8f9ea862022-09-22 14:00:38 -04002667 raw_spin_unlock_irqrestore(&src->pi_lock, flags);
Waiman Long87ca4f92022-12-30 23:11:19 -05002668
2669 if (unlikely(user_mask))
2670 kfree(user_mask);
2671
Will Deaconb90ca8b2021-07-30 12:24:33 +01002672 return 0;
2673}
2674
Will Deacon07ec77a2021-07-30 12:24:35 +01002675static inline struct cpumask *clear_user_cpus_ptr(struct task_struct *p)
2676{
2677 struct cpumask *user_mask = NULL;
2678
2679 swap(p->user_cpus_ptr, user_mask);
2680
2681 return user_mask;
2682}
2683
Will Deaconb90ca8b2021-07-30 12:24:33 +01002684void release_user_cpus_ptr(struct task_struct *p)
2685{
Will Deacon07ec77a2021-07-30 12:24:35 +01002686 kfree(clear_user_cpus_ptr(p));
Will Deaconb90ca8b2021-07-30 12:24:33 +01002687}
2688
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002689/*
Valentin Schneiderc777d842020-10-13 15:01:16 +01002690 * This function is wildly self concurrent; here be dragons.
2691 *
2692 *
2693 * When given a valid mask, __set_cpus_allowed_ptr() must block until the
2694 * designated task is enqueued on an allowed CPU. If that task is currently
2695 * running, we have to kick it out using the CPU stopper.
2696 *
2697 * Migrate-Disable comes along and tramples all over our nice sandcastle.
2698 * Consider:
2699 *
2700 * Initial conditions: P0->cpus_mask = [0, 1]
2701 *
2702 * P0@CPU0 P1
2703 *
2704 * migrate_disable();
2705 * <preempted>
2706 * set_cpus_allowed_ptr(P0, [1]);
2707 *
2708 * P1 *cannot* return from this set_cpus_allowed_ptr() call until P0 executes
2709 * its outermost migrate_enable() (i.e. it exits its Migrate-Disable region).
2710 * This means we need the following scheme:
2711 *
2712 * P0@CPU0 P1
2713 *
2714 * migrate_disable();
2715 * <preempted>
2716 * set_cpus_allowed_ptr(P0, [1]);
2717 * <blocks>
2718 * <resumes>
2719 * migrate_enable();
2720 * __set_cpus_allowed_ptr();
2721 * <wakes local stopper>
2722 * `--> <woken on migration completion>
2723 *
2724 * Now the fun stuff: there may be several P1-like tasks, i.e. multiple
2725 * concurrent set_cpus_allowed_ptr(P0, [*]) calls. CPU affinity changes of any
2726 * task p are serialized by p->pi_lock, which we can leverage: the one that
2727 * should come into effect at the end of the Migrate-Disable region is the last
2728 * one. This means we only need to track a single cpumask (i.e. p->cpus_mask),
2729 * but we still need to properly signal those waiting tasks at the appropriate
2730 * moment.
2731 *
2732 * This is implemented using struct set_affinity_pending. The first
2733 * __set_cpus_allowed_ptr() caller within a given Migrate-Disable region will
2734 * setup an instance of that struct and install it on the targeted task_struct.
2735 * Any and all further callers will reuse that instance. Those then wait for
2736 * a completion signaled at the tail of the CPU stopper callback (1), triggered
2737 * on the end of the Migrate-Disable region (i.e. outermost migrate_enable()).
2738 *
2739 *
2740 * (1) In the cases covered above. There is one more where the completion is
2741 * signaled within affine_move_task() itself: when a subsequent affinity request
Valentin Schneidere1407492021-02-25 10:22:30 +01002742 * occurs after the stopper bailed out due to the targeted task still being
2743 * Migrate-Disable. Consider:
Valentin Schneiderc777d842020-10-13 15:01:16 +01002744 *
2745 * Initial conditions: P0->cpus_mask = [0, 1]
2746 *
Valentin Schneidere1407492021-02-25 10:22:30 +01002747 * CPU0 P1 P2
2748 * <P0>
2749 * migrate_disable();
2750 * <preempted>
Valentin Schneiderc777d842020-10-13 15:01:16 +01002751 * set_cpus_allowed_ptr(P0, [1]);
2752 * <blocks>
Valentin Schneidere1407492021-02-25 10:22:30 +01002753 * <migration/0>
2754 * migration_cpu_stop()
2755 * is_migration_disabled()
2756 * <bails>
Valentin Schneiderc777d842020-10-13 15:01:16 +01002757 * set_cpus_allowed_ptr(P0, [0, 1]);
2758 * <signal completion>
2759 * <awakes>
2760 *
2761 * Note that the above is safe vs a concurrent migrate_enable(), as any
2762 * pending affinity completion is preceded by an uninstallation of
2763 * p->migration_pending done with p->pi_lock held.
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002764 */
2765static int affine_move_task(struct rq *rq, struct task_struct *p, struct rq_flags *rf,
2766 int dest_cpu, unsigned int flags)
Waiman Long5584e8a2022-09-22 14:00:37 -04002767 __releases(rq->lock)
2768 __releases(p->pi_lock)
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002769{
2770 struct set_affinity_pending my_pending = { }, *pending = NULL;
Peter Zijlstra9e818892021-02-24 11:31:09 +01002771 bool stop_pending, complete = false;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002772
2773 /* Can the task run on the task's current CPU? If so, we're done */
2774 if (cpumask_test_cpu(task_cpu(p), &p->cpus_mask)) {
Peter Zijlstraa7c81552020-09-28 17:06:07 +02002775 struct task_struct *push_task = NULL;
2776
2777 if ((flags & SCA_MIGRATE_ENABLE) &&
2778 (p->migration_flags & MDF_PUSH) && !rq->push_busy) {
2779 rq->push_busy = true;
2780 push_task = get_task_struct(p);
2781 }
2782
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002783 /*
2784 * If there are pending waiters, but no pending stop_work,
2785 * then complete now.
2786 */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002787 pending = p->migration_pending;
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002788 if (pending && !pending->stop_pending) {
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002789 p->migration_pending = NULL;
2790 complete = true;
2791 }
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002792
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002793 task_rq_unlock(rq, p, rf);
2794
Peter Zijlstraa7c81552020-09-28 17:06:07 +02002795 if (push_task) {
2796 stop_one_cpu_nowait(rq->cpu, push_cpu_stop,
2797 p, &rq->push_work);
2798 }
2799
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002800 if (complete)
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002801 complete_all(&pending->done);
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002802
2803 return 0;
2804 }
2805
2806 if (!(flags & SCA_MIGRATE_ENABLE)) {
2807 /* serialized by p->pi_lock */
2808 if (!p->migration_pending) {
Valentin Schneiderc777d842020-10-13 15:01:16 +01002809 /* Install the request */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002810 refcount_set(&my_pending.refs, 1);
2811 init_completion(&my_pending.done);
Peter Zijlstra8a6edb52021-02-13 13:10:35 +01002812 my_pending.arg = (struct migration_arg) {
2813 .task = p,
Valentin Schneider475ea6c2021-05-26 21:57:50 +01002814 .dest_cpu = dest_cpu,
Peter Zijlstra8a6edb52021-02-13 13:10:35 +01002815 .pending = &my_pending,
2816 };
2817
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002818 p->migration_pending = &my_pending;
2819 } else {
2820 pending = p->migration_pending;
2821 refcount_inc(&pending->refs);
Valentin Schneider475ea6c2021-05-26 21:57:50 +01002822 /*
2823 * Affinity has changed, but we've already installed a
2824 * pending. migration_cpu_stop() *must* see this, else
2825 * we risk a completion of the pending despite having a
2826 * task on a disallowed CPU.
2827 *
2828 * Serialized by p->pi_lock, so this is safe.
2829 */
2830 pending->arg.dest_cpu = dest_cpu;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002831 }
2832 }
2833 pending = p->migration_pending;
2834 /*
2835 * - !MIGRATE_ENABLE:
2836 * we'll have installed a pending if there wasn't one already.
2837 *
2838 * - MIGRATE_ENABLE:
2839 * we're here because the current CPU isn't matching anymore,
2840 * the only way that can happen is because of a concurrent
2841 * set_cpus_allowed_ptr() call, which should then still be
2842 * pending completion.
2843 *
2844 * Either way, we really should have a @pending here.
2845 */
2846 if (WARN_ON_ONCE(!pending)) {
2847 task_rq_unlock(rq, p, rf);
2848 return -EINVAL;
2849 }
2850
Peter Zijlstra0b9d46f2022-09-06 12:33:04 +02002851 if (task_on_cpu(rq, p) || READ_ONCE(p->__state) == TASK_WAKING) {
Valentin Schneiderc777d842020-10-13 15:01:16 +01002852 /*
Peter Zijlstra58b1a452021-02-24 11:15:23 +01002853 * MIGRATE_ENABLE gets here because 'p == current', but for
2854 * anything else we cannot do is_migration_disabled(), punt
2855 * and have the stopper function handle it all race-free.
Valentin Schneiderc777d842020-10-13 15:01:16 +01002856 */
Peter Zijlstra9e818892021-02-24 11:31:09 +01002857 stop_pending = pending->stop_pending;
2858 if (!stop_pending)
2859 pending->stop_pending = true;
Peter Zijlstra58b1a452021-02-24 11:15:23 +01002860
Peter Zijlstra58b1a452021-02-24 11:15:23 +01002861 if (flags & SCA_MIGRATE_ENABLE)
2862 p->migration_flags &= ~MDF_PUSH;
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002863
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002864 task_rq_unlock(rq, p, rf);
Peter Zijlstra8a6edb52021-02-13 13:10:35 +01002865
Peter Zijlstra9e818892021-02-24 11:31:09 +01002866 if (!stop_pending) {
2867 stop_one_cpu_nowait(cpu_of(rq), migration_cpu_stop,
2868 &pending->arg, &pending->stop_work);
2869 }
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002870
Peter Zijlstra58b1a452021-02-24 11:15:23 +01002871 if (flags & SCA_MIGRATE_ENABLE)
2872 return 0;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002873 } else {
2874
2875 if (!is_migration_disabled(p)) {
2876 if (task_on_rq_queued(p))
2877 rq = move_queued_task(rq, rf, p, dest_cpu);
2878
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002879 if (!pending->stop_pending) {
2880 p->migration_pending = NULL;
2881 complete = true;
2882 }
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002883 }
2884 task_rq_unlock(rq, p, rf);
2885
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002886 if (complete)
2887 complete_all(&pending->done);
2888 }
2889
2890 wait_for_completion(&pending->done);
2891
2892 if (refcount_dec_and_test(&pending->refs))
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002893 wake_up_var(&pending->refs); /* No UaF, just an address */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002894
Valentin Schneiderc777d842020-10-13 15:01:16 +01002895 /*
2896 * Block the original owner of &pending until all subsequent callers
2897 * have seen the completion and decremented the refcount
2898 */
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002899 wait_var_event(&my_pending.refs, !refcount_read(&my_pending.refs));
2900
Peter Zijlstra50caf9c2021-02-24 11:42:08 +01002901 /* ARGH */
2902 WARN_ON_ONCE(my_pending.stop_pending);
2903
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02002904 return 0;
2905}
2906
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002907/*
Will Deacon07ec77a2021-07-30 12:24:35 +01002908 * Called with both p->pi_lock and rq->lock held; drops both before returning.
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002909 */
Will Deacon07ec77a2021-07-30 12:24:35 +01002910static int __set_cpus_allowed_ptr_locked(struct task_struct *p,
Waiman Long713a2e22022-09-22 14:00:40 -04002911 struct affinity_context *ctx,
Will Deacon07ec77a2021-07-30 12:24:35 +01002912 struct rq *rq,
2913 struct rq_flags *rf)
2914 __releases(rq->lock)
2915 __releases(p->pi_lock)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002916{
Will Deacon234a5032021-07-30 12:24:32 +01002917 const struct cpumask *cpu_allowed_mask = task_cpu_possible_mask(p);
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01002918 const struct cpumask *cpu_valid_mask = cpu_active_mask;
Will Deacon234a5032021-07-30 12:24:32 +01002919 bool kthread = p->flags & PF_KTHREAD;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002920 unsigned int dest_cpu;
2921 int ret = 0;
2922
Wanpeng Lia499c3e2017-02-21 23:52:55 -08002923 update_rq_clock(rq);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002924
Will Deacon234a5032021-07-30 12:24:32 +01002925 if (kthread || is_migration_disabled(p)) {
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01002926 /*
Peter Zijlstra741ba802021-01-16 11:56:37 +01002927 * Kernel threads are allowed on online && !active CPUs,
2928 * however, during cpu-hot-unplug, even these might get pushed
2929 * away if not KTHREAD_IS_PER_CPU.
Peter Zijlstraaf449902020-09-17 10:38:30 +02002930 *
2931 * Specifically, migration_disabled() tasks must not fail the
2932 * cpumask_any_and_distribute() pick below, esp. so on
2933 * SCA_MIGRATE_ENABLE, otherwise we'll not call
2934 * set_cpus_allowed_common() and actually reset p->cpus_ptr.
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01002935 */
2936 cpu_valid_mask = cpu_online_mask;
2937 }
2938
Waiman Long713a2e22022-09-22 14:00:40 -04002939 if (!kthread && !cpumask_subset(ctx->new_mask, cpu_allowed_mask)) {
Will Deacon234a5032021-07-30 12:24:32 +01002940 ret = -EINVAL;
2941 goto out;
2942 }
2943
Peter Zijlstra25834c72015-05-15 17:43:34 +02002944 /*
2945 * Must re-check here, to close a race against __kthread_bind(),
2946 * sched_setaffinity() is not guaranteed to observe the flag.
2947 */
Waiman Long713a2e22022-09-22 14:00:40 -04002948 if ((ctx->flags & SCA_CHECK) && (p->flags & PF_NO_SETAFFINITY)) {
Peter Zijlstra25834c72015-05-15 17:43:34 +02002949 ret = -EINVAL;
2950 goto out;
2951 }
2952
Waiman Long713a2e22022-09-22 14:00:40 -04002953 if (!(ctx->flags & SCA_MIGRATE_ENABLE)) {
2954 if (cpumask_equal(&p->cpus_mask, ctx->new_mask))
Valentin Schneider885b3ba2020-10-13 15:01:15 +01002955 goto out;
2956
2957 if (WARN_ON_ONCE(p == current &&
2958 is_migration_disabled(p) &&
Waiman Long713a2e22022-09-22 14:00:40 -04002959 !cpumask_test_cpu(task_cpu(p), ctx->new_mask))) {
Valentin Schneider885b3ba2020-10-13 15:01:15 +01002960 ret = -EBUSY;
2961 goto out;
2962 }
2963 }
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002964
Paul Turner46a87b32020-03-10 18:01:13 -07002965 /*
2966 * Picking a ~random cpu helps in cases where we are changing affinity
2967 * for groups of tasks (ie. cpuset), so that load balancing is not
2968 * immediately required to distribute the tasks within their new mask.
2969 */
Waiman Long713a2e22022-09-22 14:00:40 -04002970 dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
KeMeng Shi714e5012019-09-16 06:53:28 +00002971 if (dest_cpu >= nr_cpu_ids) {
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002972 ret = -EINVAL;
2973 goto out;
2974 }
2975
Waiman Long713a2e22022-09-22 14:00:40 -04002976 __do_set_cpus_allowed(p, ctx);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02002977
Waiman Long8f9ea862022-09-22 14:00:38 -04002978 return affine_move_task(rq, p, rf, dest_cpu, ctx->flags);
Will Deacon07ec77a2021-07-30 12:24:35 +01002979
2980out:
2981 task_rq_unlock(rq, p, rf);
2982
2983 return ret;
2984}
2985
2986/*
2987 * Change a given task's CPU affinity. Migrate the thread to a
2988 * proper CPU and schedule it away if the CPU it's executing on
2989 * is removed from the allowed bitmask.
2990 *
2991 * NOTE: the caller must have a valid reference to the task, the
2992 * task must not exit() & deallocate itself prematurely. The
2993 * call is not atomic; no spinlocks may be held.
2994 */
2995static int __set_cpus_allowed_ptr(struct task_struct *p,
Waiman Long713a2e22022-09-22 14:00:40 -04002996 struct affinity_context *ctx)
Will Deacon07ec77a2021-07-30 12:24:35 +01002997{
2998 struct rq_flags rf;
2999 struct rq *rq;
3000
3001 rq = task_rq_lock(p, &rf);
Waiman Longda019032022-09-22 14:00:39 -04003002 /*
3003 * Masking should be skipped if SCA_USER or any of the SCA_MIGRATE_*
3004 * flags are set.
3005 */
3006 if (p->user_cpus_ptr &&
3007 !(ctx->flags & (SCA_USER | SCA_MIGRATE_ENABLE | SCA_MIGRATE_DISABLE)) &&
3008 cpumask_and(rq->scratch_mask, ctx->new_mask, p->user_cpus_ptr))
3009 ctx->new_mask = rq->scratch_mask;
3010
Waiman Long713a2e22022-09-22 14:00:40 -04003011 return __set_cpus_allowed_ptr_locked(p, ctx, rq, &rf);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02003012}
Peter Zijlstra25834c72015-05-15 17:43:34 +02003013
3014int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3015{
Waiman Long713a2e22022-09-22 14:00:40 -04003016 struct affinity_context ac = {
3017 .new_mask = new_mask,
3018 .flags = 0,
3019 };
3020
3021 return __set_cpus_allowed_ptr(p, &ac);
Peter Zijlstra25834c72015-05-15 17:43:34 +02003022}
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02003023EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
3024
Will Deacon07ec77a2021-07-30 12:24:35 +01003025/*
3026 * Change a given task's CPU affinity to the intersection of its current
Waiman Long8f9ea862022-09-22 14:00:38 -04003027 * affinity mask and @subset_mask, writing the resulting mask to @new_mask.
3028 * If user_cpus_ptr is defined, use it as the basis for restricting CPU
3029 * affinity or use cpu_online_mask instead.
3030 *
Will Deacon07ec77a2021-07-30 12:24:35 +01003031 * If the resulting mask is empty, leave the affinity unchanged and return
3032 * -EINVAL.
3033 */
3034static int restrict_cpus_allowed_ptr(struct task_struct *p,
3035 struct cpumask *new_mask,
3036 const struct cpumask *subset_mask)
3037{
Waiman Long8f9ea862022-09-22 14:00:38 -04003038 struct affinity_context ac = {
3039 .new_mask = new_mask,
3040 .flags = 0,
3041 };
Will Deacon07ec77a2021-07-30 12:24:35 +01003042 struct rq_flags rf;
3043 struct rq *rq;
3044 int err;
3045
Will Deacon07ec77a2021-07-30 12:24:35 +01003046 rq = task_rq_lock(p, &rf);
3047
3048 /*
3049 * Forcefully restricting the affinity of a deadline task is
3050 * likely to cause problems, so fail and noisily override the
3051 * mask entirely.
3052 */
3053 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
3054 err = -EPERM;
3055 goto err_unlock;
3056 }
3057
Waiman Long8f9ea862022-09-22 14:00:38 -04003058 if (!cpumask_and(new_mask, task_user_cpus(p), subset_mask)) {
Will Deacon07ec77a2021-07-30 12:24:35 +01003059 err = -EINVAL;
3060 goto err_unlock;
3061 }
3062
Waiman Long713a2e22022-09-22 14:00:40 -04003063 return __set_cpus_allowed_ptr_locked(p, &ac, rq, &rf);
Will Deacon07ec77a2021-07-30 12:24:35 +01003064
3065err_unlock:
3066 task_rq_unlock(rq, p, &rf);
Will Deacon07ec77a2021-07-30 12:24:35 +01003067 return err;
3068}
3069
3070/*
3071 * Restrict the CPU affinity of task @p so that it is a subset of
Waiman Long5584e8a2022-09-22 14:00:37 -04003072 * task_cpu_possible_mask() and point @p->user_cpus_ptr to a copy of the
Will Deacon07ec77a2021-07-30 12:24:35 +01003073 * old affinity mask. If the resulting mask is empty, we warn and walk
3074 * up the cpuset hierarchy until we find a suitable mask.
3075 */
3076void force_compatible_cpus_allowed_ptr(struct task_struct *p)
3077{
3078 cpumask_var_t new_mask;
3079 const struct cpumask *override_mask = task_cpu_possible_mask(p);
3080
3081 alloc_cpumask_var(&new_mask, GFP_KERNEL);
3082
3083 /*
3084 * __migrate_task() can fail silently in the face of concurrent
3085 * offlining of the chosen destination CPU, so take the hotplug
3086 * lock to ensure that the migration succeeds.
3087 */
3088 cpus_read_lock();
3089 if (!cpumask_available(new_mask))
3090 goto out_set_mask;
3091
3092 if (!restrict_cpus_allowed_ptr(p, new_mask, override_mask))
3093 goto out_free_mask;
3094
3095 /*
3096 * We failed to find a valid subset of the affinity mask for the
3097 * task, so override it based on its cpuset hierarchy.
3098 */
3099 cpuset_cpus_allowed(p, new_mask);
3100 override_mask = new_mask;
3101
3102out_set_mask:
3103 if (printk_ratelimit()) {
3104 printk_deferred("Overriding affinity for process %d (%s) to CPUs %*pbl\n",
3105 task_pid_nr(p), p->comm,
3106 cpumask_pr_args(override_mask));
3107 }
3108
3109 WARN_ON(set_cpus_allowed_ptr(p, override_mask));
3110out_free_mask:
3111 cpus_read_unlock();
3112 free_cpumask_var(new_mask);
3113}
3114
3115static int
Waiman Long713a2e22022-09-22 14:00:40 -04003116__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx);
Will Deacon07ec77a2021-07-30 12:24:35 +01003117
3118/*
3119 * Restore the affinity of a task @p which was previously restricted by a
Waiman Long8f9ea862022-09-22 14:00:38 -04003120 * call to force_compatible_cpus_allowed_ptr().
Will Deacon07ec77a2021-07-30 12:24:35 +01003121 *
3122 * It is the caller's responsibility to serialise this with any calls to
3123 * force_compatible_cpus_allowed_ptr(@p).
3124 */
3125void relax_compatible_cpus_allowed_ptr(struct task_struct *p)
3126{
Waiman Long713a2e22022-09-22 14:00:40 -04003127 struct affinity_context ac = {
Waiman Long8f9ea862022-09-22 14:00:38 -04003128 .new_mask = task_user_cpus(p),
3129 .flags = 0,
Waiman Long713a2e22022-09-22 14:00:40 -04003130 };
Waiman Long8f9ea862022-09-22 14:00:38 -04003131 int ret;
Will Deacon07ec77a2021-07-30 12:24:35 +01003132
3133 /*
Waiman Long8f9ea862022-09-22 14:00:38 -04003134 * Try to restore the old affinity mask with __sched_setaffinity().
3135 * Cpuset masking will be done there too.
Will Deacon07ec77a2021-07-30 12:24:35 +01003136 */
Waiman Long8f9ea862022-09-22 14:00:38 -04003137 ret = __sched_setaffinity(p, &ac);
3138 WARN_ON_ONCE(ret);
Will Deacon07ec77a2021-07-30 12:24:35 +01003139}
3140
Ingo Molnardd41f592007-07-09 18:51:59 +02003141void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02003142{
Peter Zijlstrae2912002009-12-16 18:04:36 +01003143#ifdef CONFIG_SCHED_DEBUG
Peter Zijlstra2f064a52021-06-11 10:28:17 +02003144 unsigned int state = READ_ONCE(p->__state);
3145
Peter Zijlstrae2912002009-12-16 18:04:36 +01003146 /*
3147 * We should never call set_task_cpu() on a blocked task,
3148 * ttwu() will sort out the placement.
3149 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02003150 WARN_ON_ONCE(state != TASK_RUNNING && state != TASK_WAKING && !p->on_rq);
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003151
Joonwoo Park3ea94de2015-11-12 19:38:54 -08003152 /*
3153 * Migrating fair class task must have p->on_rq = TASK_ON_RQ_MIGRATING,
3154 * because schedstat_wait_{start,end} rebase migrating task's wait_start
3155 * time relying on p->on_rq.
3156 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02003157 WARN_ON_ONCE(state == TASK_RUNNING &&
Joonwoo Park3ea94de2015-11-12 19:38:54 -08003158 p->sched_class == &fair_sched_class &&
3159 (p->on_rq && !task_on_rq_migrating(p)));
3160
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003161#ifdef CONFIG_LOCKDEP
Peter Zijlstra6c6c54e2011-06-03 17:37:07 +02003162 /*
3163 * The caller should hold either p->pi_lock or rq->lock, when changing
3164 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
3165 *
3166 * sched_move_task() holds both and thus holding either pins the cgroup,
Peter Zijlstra8323f262012-06-22 13:36:05 +02003167 * see task_group().
Peter Zijlstra6c6c54e2011-06-03 17:37:07 +02003168 *
3169 * Furthermore, all task_rq users should acquire both locks, see
3170 * task_rq_lock().
3171 */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003172 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +01003173 lockdep_is_held(__rq_lockp(task_rq(p)))));
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003174#endif
Peter Zijlstra4ff90832017-09-07 17:03:52 +02003175 /*
3176 * Clearly, migrating tasks to offline CPUs is a fairly daft thing.
3177 */
3178 WARN_ON_ONCE(!cpu_online(new_cpu));
Peter Zijlstraaf449902020-09-17 10:38:30 +02003179
3180 WARN_ON_ONCE(is_migration_disabled(p));
Peter Zijlstrae2912002009-12-16 18:04:36 +01003181#endif
3182
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08003183 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01003184
Peter Zijlstra0c697742009-12-22 15:43:19 +01003185 if (task_cpu(p) != new_cpu) {
Paul Turner0a74bef2012-10-04 13:18:30 +02003186 if (p->sched_class->migrate_task_rq)
Srikar Dronamraju1327237a52018-09-21 23:18:57 +05303187 p->sched_class->migrate_task_rq(p, new_cpu);
Peter Zijlstra0c697742009-12-22 15:43:19 +01003188 p->se.nr_migrations++;
Mathieu Desnoyersd7822b12018-06-02 08:43:54 -04003189 rseq_migrate(p);
Peter Zijlstraff303e62015-04-17 20:05:30 +02003190 perf_event_task_migrate(p);
Peter Zijlstra0c697742009-12-22 15:43:19 +01003191 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003192
3193 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02003194}
3195
Srikar Dronamraju0ad4e3d2018-06-20 22:32:50 +05303196#ifdef CONFIG_NUMA_BALANCING
Peter Zijlstraac66f542013-10-07 11:29:16 +01003197static void __migrate_swap_task(struct task_struct *p, int cpu)
3198{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003199 if (task_on_rq_queued(p)) {
Peter Zijlstraac66f542013-10-07 11:29:16 +01003200 struct rq *src_rq, *dst_rq;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003201 struct rq_flags srf, drf;
Peter Zijlstraac66f542013-10-07 11:29:16 +01003202
3203 src_rq = task_rq(p);
3204 dst_rq = cpu_rq(cpu);
3205
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003206 rq_pin_lock(src_rq, &srf);
3207 rq_pin_lock(dst_rq, &drf);
3208
Peter Zijlstraac66f542013-10-07 11:29:16 +01003209 deactivate_task(src_rq, p, 0);
3210 set_task_cpu(p, cpu);
3211 activate_task(dst_rq, p, 0);
3212 check_preempt_curr(dst_rq, p, 0);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003213
3214 rq_unpin_lock(dst_rq, &drf);
3215 rq_unpin_lock(src_rq, &srf);
3216
Peter Zijlstraac66f542013-10-07 11:29:16 +01003217 } else {
3218 /*
3219 * Task isn't running anymore; make it appear like we migrated
3220 * it before it went to sleep. This means on wakeup we make the
Ingo Molnard1ccc662017-02-01 11:46:42 +01003221 * previous CPU our target instead of where it really is.
Peter Zijlstraac66f542013-10-07 11:29:16 +01003222 */
3223 p->wake_cpu = cpu;
3224 }
3225}
3226
3227struct migration_swap_arg {
3228 struct task_struct *src_task, *dst_task;
3229 int src_cpu, dst_cpu;
3230};
3231
3232static int migrate_swap_stop(void *data)
3233{
3234 struct migration_swap_arg *arg = data;
3235 struct rq *src_rq, *dst_rq;
3236 int ret = -EAGAIN;
3237
Peter Zijlstra62694cd2015-10-09 18:36:29 +02003238 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
3239 return -EAGAIN;
3240
Peter Zijlstraac66f542013-10-07 11:29:16 +01003241 src_rq = cpu_rq(arg->src_cpu);
3242 dst_rq = cpu_rq(arg->dst_cpu);
3243
Peter Zijlstra74602312013-10-10 20:17:22 +02003244 double_raw_lock(&arg->src_task->pi_lock,
3245 &arg->dst_task->pi_lock);
Peter Zijlstraac66f542013-10-07 11:29:16 +01003246 double_rq_lock(src_rq, dst_rq);
Peter Zijlstra62694cd2015-10-09 18:36:29 +02003247
Peter Zijlstraac66f542013-10-07 11:29:16 +01003248 if (task_cpu(arg->dst_task) != arg->dst_cpu)
3249 goto unlock;
3250
3251 if (task_cpu(arg->src_task) != arg->src_cpu)
3252 goto unlock;
3253
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003254 if (!cpumask_test_cpu(arg->dst_cpu, arg->src_task->cpus_ptr))
Peter Zijlstraac66f542013-10-07 11:29:16 +01003255 goto unlock;
3256
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003257 if (!cpumask_test_cpu(arg->src_cpu, arg->dst_task->cpus_ptr))
Peter Zijlstraac66f542013-10-07 11:29:16 +01003258 goto unlock;
3259
3260 __migrate_swap_task(arg->src_task, arg->dst_cpu);
3261 __migrate_swap_task(arg->dst_task, arg->src_cpu);
3262
3263 ret = 0;
3264
3265unlock:
3266 double_rq_unlock(src_rq, dst_rq);
Peter Zijlstra74602312013-10-10 20:17:22 +02003267 raw_spin_unlock(&arg->dst_task->pi_lock);
3268 raw_spin_unlock(&arg->src_task->pi_lock);
Peter Zijlstraac66f542013-10-07 11:29:16 +01003269
3270 return ret;
3271}
3272
3273/*
3274 * Cross migrate two tasks
3275 */
Srikar Dronamraju0ad4e3d2018-06-20 22:32:50 +05303276int migrate_swap(struct task_struct *cur, struct task_struct *p,
3277 int target_cpu, int curr_cpu)
Peter Zijlstraac66f542013-10-07 11:29:16 +01003278{
3279 struct migration_swap_arg arg;
3280 int ret = -EINVAL;
3281
Peter Zijlstraac66f542013-10-07 11:29:16 +01003282 arg = (struct migration_swap_arg){
3283 .src_task = cur,
Srikar Dronamraju0ad4e3d2018-06-20 22:32:50 +05303284 .src_cpu = curr_cpu,
Peter Zijlstraac66f542013-10-07 11:29:16 +01003285 .dst_task = p,
Srikar Dronamraju0ad4e3d2018-06-20 22:32:50 +05303286 .dst_cpu = target_cpu,
Peter Zijlstraac66f542013-10-07 11:29:16 +01003287 };
3288
3289 if (arg.src_cpu == arg.dst_cpu)
3290 goto out;
3291
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02003292 /*
3293 * These three tests are all lockless; this is OK since all of them
3294 * will be re-checked with proper locks held further down the line.
3295 */
Peter Zijlstraac66f542013-10-07 11:29:16 +01003296 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
3297 goto out;
3298
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003299 if (!cpumask_test_cpu(arg.dst_cpu, arg.src_task->cpus_ptr))
Peter Zijlstraac66f542013-10-07 11:29:16 +01003300 goto out;
3301
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003302 if (!cpumask_test_cpu(arg.src_cpu, arg.dst_task->cpus_ptr))
Peter Zijlstraac66f542013-10-07 11:29:16 +01003303 goto out;
3304
Mel Gorman286549d2014-01-21 15:51:03 -08003305 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
Peter Zijlstraac66f542013-10-07 11:29:16 +01003306 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
3307
3308out:
Peter Zijlstraac66f542013-10-07 11:29:16 +01003309 return ret;
3310}
Srikar Dronamraju0ad4e3d2018-06-20 22:32:50 +05303311#endif /* CONFIG_NUMA_BALANCING */
Peter Zijlstraac66f542013-10-07 11:29:16 +01003312
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 * wait_task_inactive - wait for a thread to unschedule.
3315 *
Peter Zijlstraf9fc8ca2022-09-06 12:39:55 +02003316 * Wait for the thread to block in any of the states set in @match_state.
3317 * If it changes, i.e. @p might have woken up, then return zero. When we
3318 * succeed in waiting for @p to be off its CPU, we return a positive number
3319 * (its total switch count). If a second call a short while later returns the
3320 * same number, the caller can be sure that @p has remained unscheduled the
3321 * whole time.
Roland McGrath85ba2d82008-07-25 19:45:58 -07003322 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 * The caller must ensure that the task *will* unschedule sometime soon,
3324 * else this function might spin for a *long* time. This function can't
3325 * be called with interrupts off, or it may introduce deadlock with
3326 * smp_call_function() if an IPI is sent by the same process we are
3327 * waiting to become inactive.
3328 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02003329unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003331 int running, queued;
Peter Zijlstraeb580752015-07-31 21:28:18 +02003332 struct rq_flags rf;
Roland McGrath85ba2d82008-07-25 19:45:58 -07003333 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003334 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335
Andi Kleen3a5c3592007-10-15 17:00:14 +02003336 for (;;) {
3337 /*
3338 * We do the initial early heuristics without holding
3339 * any task-queue locks at all. We'll only try to get
3340 * the runqueue lock when things look like they will
3341 * work out!
3342 */
3343 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07003344
Andi Kleen3a5c3592007-10-15 17:00:14 +02003345 /*
3346 * If the task is actively running on another CPU
3347 * still, just relax and busy-wait without holding
3348 * any locks.
3349 *
3350 * NOTE! Since we don't hold any locks, it's not
3351 * even sure that "rq" stays as the right runqueue!
Peter Zijlstra0b9d46f2022-09-06 12:33:04 +02003352 * But we don't care, since "task_on_cpu()" will
Andi Kleen3a5c3592007-10-15 17:00:14 +02003353 * return false if the runqueue has changed and p
3354 * is actually now running somewhere else!
3355 */
Peter Zijlstra0b9d46f2022-09-06 12:33:04 +02003356 while (task_on_cpu(rq, p)) {
Peter Zijlstraf9fc8ca2022-09-06 12:39:55 +02003357 if (!(READ_ONCE(p->__state) & match_state))
Roland McGrath85ba2d82008-07-25 19:45:58 -07003358 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02003359 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07003360 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07003361
Andi Kleen3a5c3592007-10-15 17:00:14 +02003362 /*
3363 * Ok, time to look more closely! We need the rq
3364 * lock now, to be *sure*. If we're wrong, we'll
3365 * just go back and repeat.
3366 */
Peter Zijlstraeb580752015-07-31 21:28:18 +02003367 rq = task_rq_lock(p, &rf);
Peter Zijlstra27a9da62010-05-04 20:36:56 +02003368 trace_sched_wait_task(p);
Peter Zijlstra0b9d46f2022-09-06 12:33:04 +02003369 running = task_on_cpu(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003370 queued = task_on_rq_queued(p);
Roland McGrath85ba2d82008-07-25 19:45:58 -07003371 ncsw = 0;
Peter Zijlstraf9fc8ca2022-09-06 12:39:55 +02003372 if (READ_ONCE(p->__state) & match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07003373 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Peter Zijlstraeb580752015-07-31 21:28:18 +02003374 task_rq_unlock(rq, p, &rf);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07003375
Andi Kleen3a5c3592007-10-15 17:00:14 +02003376 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07003377 * If it changed from the expected state, bail out now.
3378 */
3379 if (unlikely(!ncsw))
3380 break;
3381
3382 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02003383 * Was it really running after all now that we
3384 * checked with the proper locks actually held?
3385 *
3386 * Oops. Go back and try again..
3387 */
3388 if (unlikely(running)) {
3389 cpu_relax();
3390 continue;
3391 }
3392
3393 /*
3394 * It's not enough that it's not actively running,
3395 * it must be off the runqueue _entirely_, and not
3396 * preempted!
3397 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00003398 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02003399 * running right now), it's preempted, and we should
3400 * yield - it could be a while.
3401 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003402 if (unlikely(queued)) {
Thomas Gleixner8b0e1952016-12-25 12:30:41 +01003403 ktime_t to = NSEC_PER_SEC / HZ;
Thomas Gleixner8eb90c32011-02-23 23:52:21 +00003404
3405 set_current_state(TASK_UNINTERRUPTIBLE);
Sebastian Andrzej Siewiorc33627e2021-08-26 19:04:08 +02003406 schedule_hrtimeout(&to, HRTIMER_MODE_REL_HARD);
Andi Kleen3a5c3592007-10-15 17:00:14 +02003407 continue;
3408 }
3409
3410 /*
3411 * Ahh, all good. It wasn't running, and it wasn't
3412 * runnable, which means that it will never become
3413 * running in the future either. We're all done!
3414 */
3415 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07003417
3418 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419}
3420
3421/***
3422 * kick_process - kick a running thread to enter/exit the kernel
3423 * @p: the to-be-kicked thread
3424 *
3425 * Cause a process which is running on another CPU to enter
3426 * kernel-mode, without any delay. (to get signals handled.)
3427 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003428 * NOTE: this function doesn't have to take the runqueue lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429 * because all it wants to ensure is that the remote task enters
3430 * the kernel. If the IPI races and the task has been migrated
3431 * to another CPU then no harm is done and the purpose has been
3432 * achieved as well.
3433 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003434void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435{
3436 int cpu;
3437
3438 preempt_disable();
3439 cpu = task_cpu(p);
3440 if ((cpu != smp_processor_id()) && task_curr(p))
3441 smp_send_reschedule(cpu);
3442 preempt_enable();
3443}
Rusty Russellb43e3522009-06-12 22:27:00 -06003444EXPORT_SYMBOL_GPL(kick_process);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
Oleg Nesterov30da6882010-03-15 10:10:19 +01003446/*
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003447 * ->cpus_ptr is protected by both rq->lock and p->pi_lock
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003448 *
3449 * A few notes on cpu_active vs cpu_online:
3450 *
3451 * - cpu_active must be a subset of cpu_online
3452 *
Ingo Molnar97fb7a02018-03-03 14:01:12 +01003453 * - on CPU-up we allow per-CPU kthreads on the online && !active CPU,
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003454 * see __set_cpus_allowed_ptr(). At this point the newly online
Ingo Molnard1ccc662017-02-01 11:46:42 +01003455 * CPU isn't yet part of the sched domains, and balancing will not
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003456 * see it.
3457 *
Ingo Molnard1ccc662017-02-01 11:46:42 +01003458 * - on CPU-down we clear cpu_active() to mask the sched domains and
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003459 * avoid the load balancer to place new tasks on the to be removed
Ingo Molnard1ccc662017-02-01 11:46:42 +01003460 * CPU. Existing tasks will remain running there and will be taken
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003461 * off.
3462 *
3463 * This means that fallback selection must not select !active CPUs.
3464 * And can assume that any active CPU must be online. Conversely
3465 * select_task_rq() below may allow selection of !active CPUs in order
3466 * to satisfy the above rules.
Oleg Nesterov30da6882010-03-15 10:10:19 +01003467 */
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003468static int select_fallback_rq(int cpu, struct task_struct *p)
3469{
Tang Chenaa00d892013-02-22 16:33:33 -08003470 int nid = cpu_to_node(cpu);
3471 const struct cpumask *nodemask = NULL;
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003472 enum { cpuset, possible, fail } state = cpuset;
3473 int dest_cpu;
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003474
Tang Chenaa00d892013-02-22 16:33:33 -08003475 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01003476 * If the node that the CPU is on has been offlined, cpu_to_node()
3477 * will return -1. There is no CPU on the node, and we should
3478 * select the CPU on the other node.
Tang Chenaa00d892013-02-22 16:33:33 -08003479 */
3480 if (nid != -1) {
3481 nodemask = cpumask_of_node(nid);
3482
3483 /* Look for allowed, online CPU in same node. */
3484 for_each_cpu(dest_cpu, nodemask) {
Will Deacon9ae606b2021-07-30 12:24:28 +01003485 if (is_cpu_allowed(p, dest_cpu))
Tang Chenaa00d892013-02-22 16:33:33 -08003486 return dest_cpu;
3487 }
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003488 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003489
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003490 for (;;) {
3491 /* Any allowed, online CPU? */
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003492 for_each_cpu(dest_cpu, p->cpus_ptr) {
Peter Zijlstra175f0e22017-07-25 18:58:21 +02003493 if (!is_cpu_allowed(p, dest_cpu))
Tejun Heofeb245e2016-06-16 15:35:04 -04003494 continue;
Peter Zijlstra175f0e22017-07-25 18:58:21 +02003495
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003496 goto out;
3497 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003498
Oleg Nesterove73e85f2015-10-10 20:53:15 +02003499 /* No more Mr. Nice Guy. */
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003500 switch (state) {
3501 case cpuset:
Will Deacon97c00542021-07-30 12:24:31 +01003502 if (cpuset_cpus_allowed_fallback(p)) {
Oleg Nesterove73e85f2015-10-10 20:53:15 +02003503 state = possible;
3504 break;
3505 }
Gustavo A. R. Silvadf561f662020-08-23 17:36:59 -05003506 fallthrough;
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003507 case possible:
Peter Zijlstraaf449902020-09-17 10:38:30 +02003508 /*
3509 * XXX When called from select_task_rq() we only
3510 * hold p->pi_lock and again violate locking order.
3511 *
3512 * More yuck to audit.
3513 */
Will Deacon9ae606b2021-07-30 12:24:28 +01003514 do_set_cpus_allowed(p, task_cpu_possible_mask(p));
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003515 state = fail;
3516 break;
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003517 case fail:
3518 BUG();
3519 break;
3520 }
3521 }
3522
3523out:
3524 if (state != cpuset) {
3525 /*
3526 * Don't tell them about moving exiting tasks or
3527 * kernel threads (both mm NULL), since they never
3528 * leave kernel.
3529 */
3530 if (p->mm && printk_ratelimit()) {
John Stultzaac74dc2014-06-04 16:11:40 -07003531 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01003532 task_pid_nr(p), p->comm, cpu);
3533 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003534 }
3535
3536 return dest_cpu;
3537}
3538
Peter Zijlstrae2912002009-12-16 18:04:36 +01003539/*
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003540 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_ptr is stable.
Peter Zijlstrae2912002009-12-16 18:04:36 +01003541 */
Peter Zijlstra970b13b2009-11-25 13:31:39 +01003542static inline
Valentin Schneider3aef15512020-11-02 18:45:13 +00003543int select_task_rq(struct task_struct *p, int cpu, int wake_flags)
Peter Zijlstra970b13b2009-11-25 13:31:39 +01003544{
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003545 lockdep_assert_held(&p->pi_lock);
3546
Peter Zijlstraaf449902020-09-17 10:38:30 +02003547 if (p->nr_cpus_allowed > 1 && !is_migration_disabled(p))
Valentin Schneider3aef15512020-11-02 18:45:13 +00003548 cpu = p->sched_class->select_task_rq(p, cpu, wake_flags);
Peter Zijlstra (Intel)e9d867a2016-03-10 12:54:08 +01003549 else
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003550 cpu = cpumask_any(p->cpus_ptr);
Peter Zijlstrae2912002009-12-16 18:04:36 +01003551
3552 /*
3553 * In order not to call set_task_cpu() on a blocking task we need
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02003554 * to rely on ttwu() to place the task on a valid ->cpus_ptr
Ingo Molnard1ccc662017-02-01 11:46:42 +01003555 * CPU.
Peter Zijlstrae2912002009-12-16 18:04:36 +01003556 *
3557 * Since this is common to all placement strategies, this lives here.
3558 *
3559 * [ this allows ->select_task() to simply return task_cpu(p) and
3560 * not worry about this generic constraint ]
3561 */
Paul Burton7af443e2018-05-26 08:46:47 -07003562 if (unlikely(!is_cpu_allowed(p, cpu)))
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01003563 cpu = select_fallback_rq(task_cpu(p), p);
Peter Zijlstrae2912002009-12-16 18:04:36 +01003564
3565 return cpu;
Peter Zijlstra970b13b2009-11-25 13:31:39 +01003566}
Mike Galbraith09a40af2010-04-15 07:29:59 +02003567
Nicolas Pitref5832c12017-05-29 17:02:57 -04003568void sched_set_stop_task(int cpu, struct task_struct *stop)
3569{
Peter Zijlstraded467d2020-10-01 16:13:01 +02003570 static struct lock_class_key stop_pi_lock;
Nicolas Pitref5832c12017-05-29 17:02:57 -04003571 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
3572 struct task_struct *old_stop = cpu_rq(cpu)->stop;
3573
3574 if (stop) {
3575 /*
3576 * Make it appear like a SCHED_FIFO task, its something
3577 * userspace knows about and won't get confused about.
3578 *
3579 * Also, it will make PI more or less work without too
3580 * much confusion -- but then, stop work should not
3581 * rely on PI working anyway.
3582 */
3583 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
3584
3585 stop->sched_class = &stop_sched_class;
Peter Zijlstraded467d2020-10-01 16:13:01 +02003586
3587 /*
3588 * The PI code calls rt_mutex_setprio() with ->pi_lock held to
3589 * adjust the effective priority of a task. As a result,
3590 * rt_mutex_setprio() can trigger (RT) balancing operations,
3591 * which can then trigger wakeups of the stop thread to push
3592 * around the current task.
3593 *
3594 * The stop task itself will never be part of the PI-chain, it
3595 * never blocks, therefore that ->pi_lock recursion is safe.
3596 * Tell lockdep about this by placing the stop->pi_lock in its
3597 * own class.
3598 */
3599 lockdep_set_class(&stop->pi_lock, &stop_pi_lock);
Nicolas Pitref5832c12017-05-29 17:02:57 -04003600 }
3601
3602 cpu_rq(cpu)->stop = stop;
3603
3604 if (old_stop) {
3605 /*
3606 * Reset it back to a normal scheduling class so that
3607 * it can die in pieces.
3608 */
3609 old_stop->sched_class = &rt_sched_class;
3610 }
3611}
3612
Thomas Gleixner74d862b62020-11-18 20:48:42 +01003613#else /* CONFIG_SMP */
Peter Zijlstra25834c72015-05-15 17:43:34 +02003614
3615static inline int __set_cpus_allowed_ptr(struct task_struct *p,
Waiman Long713a2e22022-09-22 14:00:40 -04003616 struct affinity_context *ctx)
Peter Zijlstra25834c72015-05-15 17:43:34 +02003617{
Waiman Long713a2e22022-09-22 14:00:40 -04003618 return set_cpus_allowed_ptr(p, ctx->new_mask);
Peter Zijlstra25834c72015-05-15 17:43:34 +02003619}
3620
Peter Zijlstraaf449902020-09-17 10:38:30 +02003621static inline void migrate_disable_switch(struct rq *rq, struct task_struct *p) { }
3622
Thomas Gleixner3015ef42020-08-26 14:08:10 +02003623static inline bool rq_has_pinned_tasks(struct rq *rq)
3624{
3625 return false;
3626}
3627
Waiman Long9a5418b2022-12-30 23:11:20 -05003628static inline cpumask_t *alloc_user_cpus_ptr(int node)
3629{
3630 return NULL;
3631}
3632
Thomas Gleixner74d862b62020-11-18 20:48:42 +01003633#endif /* !CONFIG_SMP */
Peter Zijlstra970b13b2009-11-25 13:31:39 +01003634
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003635static void
Peter Zijlstrab84cb5d2011-04-05 17:23:55 +02003636ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
Tejun Heo9ed38112009-12-03 15:08:03 +09003637{
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003638 struct rq *rq;
3639
3640 if (!schedstat_enabled())
3641 return;
3642
3643 rq = this_rq();
Tejun Heo9ed38112009-12-03 15:08:03 +09003644
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003645#ifdef CONFIG_SMP
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003646 if (cpu == rq->cpu) {
Peter Zijlstrab85c8b72018-01-16 20:51:06 +01003647 __schedstat_inc(rq->ttwu_local);
Yafang Shaoceeadb82021-09-05 14:35:41 +00003648 __schedstat_inc(p->stats.nr_wakeups_local);
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003649 } else {
3650 struct sched_domain *sd;
3651
Yafang Shaoceeadb82021-09-05 14:35:41 +00003652 __schedstat_inc(p->stats.nr_wakeups_remote);
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02003653 rcu_read_lock();
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -05003654 for_each_domain(rq->cpu, sd) {
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003655 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Peter Zijlstrab85c8b72018-01-16 20:51:06 +01003656 __schedstat_inc(sd->ttwu_wake_remote);
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003657 break;
3658 }
3659 }
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02003660 rcu_read_unlock();
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003661 }
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02003662
3663 if (wake_flags & WF_MIGRATED)
Yafang Shaoceeadb82021-09-05 14:35:41 +00003664 __schedstat_inc(p->stats.nr_wakeups_migrate);
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003665#endif /* CONFIG_SMP */
3666
Peter Zijlstrab85c8b72018-01-16 20:51:06 +01003667 __schedstat_inc(rq->ttwu_count);
Yafang Shaoceeadb82021-09-05 14:35:41 +00003668 __schedstat_inc(p->stats.nr_wakeups);
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02003669
3670 if (wake_flags & WF_SYNC)
Yafang Shaoceeadb82021-09-05 14:35:41 +00003671 __schedstat_inc(p->stats.nr_wakeups_sync);
Tejun Heo9ed38112009-12-03 15:08:03 +09003672}
3673
Peter Zijlstra23f41ee2011-04-05 17:23:56 +02003674/*
3675 * Mark the task runnable and perform wakeup-preemption.
3676 */
Peter Zijlstrae7904a22015-08-01 19:25:08 +02003677static void ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags,
Matt Flemingd8ac8972016-09-21 14:38:10 +01003678 struct rq_flags *rf)
Tejun Heo9ed38112009-12-03 15:08:03 +09003679{
Tejun Heo9ed38112009-12-03 15:08:03 +09003680 check_preempt_curr(rq, p, wake_flags);
Peter Zijlstra2f064a52021-06-11 10:28:17 +02003681 WRITE_ONCE(p->__state, TASK_RUNNING);
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02003682 trace_sched_wakeup(p);
3683
Tejun Heo9ed38112009-12-03 15:08:03 +09003684#ifdef CONFIG_SMP
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003685 if (p->sched_class->task_woken) {
3686 /*
Tal Zussmanb19a8882020-11-12 19:51:56 -05003687 * Our task @p is fully woken up and running; so it's safe to
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003688 * drop the rq->lock, hereafter rq is only used for statistics.
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003689 */
Matt Flemingd8ac8972016-09-21 14:38:10 +01003690 rq_unpin_lock(rq, rf);
Tejun Heo9ed38112009-12-03 15:08:03 +09003691 p->sched_class->task_woken(rq, p);
Matt Flemingd8ac8972016-09-21 14:38:10 +01003692 rq_repin_lock(rq, rf);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003693 }
Tejun Heo9ed38112009-12-03 15:08:03 +09003694
Steven Rostedte69c6342010-12-06 17:10:31 -05003695 if (rq->idle_stamp) {
Frederic Weisbecker78becc22013-04-12 01:51:02 +02003696 u64 delta = rq_clock(rq) - rq->idle_stamp;
Jason Low9bd721c2013-09-13 11:26:52 -07003697 u64 max = 2*rq->max_idle_balance_cost;
Tejun Heo9ed38112009-12-03 15:08:03 +09003698
Jason Lowabfafa52013-09-13 11:26:51 -07003699 update_avg(&rq->avg_idle, delta);
3700
3701 if (rq->avg_idle > max)
Tejun Heo9ed38112009-12-03 15:08:03 +09003702 rq->avg_idle = max;
Jason Lowabfafa52013-09-13 11:26:51 -07003703
Peter Zijlstra94aafc32021-06-15 12:16:11 +01003704 rq->wake_stamp = jiffies;
3705 rq->wake_avg_idle = rq->avg_idle / 2;
3706
Tejun Heo9ed38112009-12-03 15:08:03 +09003707 rq->idle_stamp = 0;
3708 }
3709#endif
3710}
3711
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003712static void
Peter Zijlstrae7904a22015-08-01 19:25:08 +02003713ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags,
Matt Flemingd8ac8972016-09-21 14:38:10 +01003714 struct rq_flags *rf)
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003715{
Peter Zijlstra77558e42017-02-21 14:36:23 +01003716 int en_flags = ENQUEUE_WAKEUP | ENQUEUE_NOCLOCK;
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003717
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05003718 lockdep_assert_rq_held(rq);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003719
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003720 if (p->sched_contributes_to_load)
3721 rq->nr_uninterruptible--;
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003722
Peter Zijlstradbfb0892020-07-03 12:40:33 +02003723#ifdef CONFIG_SMP
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003724 if (wake_flags & WF_MIGRATED)
Peter Zijlstra59efa0b2016-05-10 18:24:37 +02003725 en_flags |= ENQUEUE_MIGRATED;
Peter Zijlstraec618b82020-09-24 13:50:42 +02003726 else
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003727#endif
Peter Zijlstraec618b82020-09-24 13:50:42 +02003728 if (p->in_iowait) {
3729 delayacct_blkio_end(p);
3730 atomic_dec(&task_rq(p)->nr_iowait);
3731 }
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003732
Peter Zijlstra1b174a22019-04-09 09:53:13 +02003733 activate_task(rq, p, en_flags);
Matt Flemingd8ac8972016-09-21 14:38:10 +01003734 ttwu_do_wakeup(rq, p, wake_flags, rf);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003735}
3736
3737/*
Peter Zijlstra58877d32020-07-02 14:52:11 +02003738 * Consider @p being inside a wait loop:
3739 *
3740 * for (;;) {
3741 * set_current_state(TASK_UNINTERRUPTIBLE);
3742 *
3743 * if (CONDITION)
3744 * break;
3745 *
3746 * schedule();
3747 * }
3748 * __set_current_state(TASK_RUNNING);
3749 *
3750 * between set_current_state() and schedule(). In this case @p is still
3751 * runnable, so all that needs doing is change p->state back to TASK_RUNNING in
3752 * an atomic manner.
3753 *
3754 * By taking task_rq(p)->lock we serialize against schedule(), if @p->on_rq
3755 * then schedule() must still happen and p->state can be changed to
3756 * TASK_RUNNING. Otherwise we lost the race, schedule() has happened, and we
3757 * need to do a full wakeup with enqueue.
3758 *
3759 * Returns: %true when the wakeup is done,
3760 * %false otherwise.
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003761 */
Peter Zijlstra58877d32020-07-02 14:52:11 +02003762static int ttwu_runnable(struct task_struct *p, int wake_flags)
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003763{
Peter Zijlstraeb580752015-07-31 21:28:18 +02003764 struct rq_flags rf;
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003765 struct rq *rq;
3766 int ret = 0;
3767
Peter Zijlstraeb580752015-07-31 21:28:18 +02003768 rq = __task_rq_lock(p, &rf);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003769 if (task_on_rq_queued(p)) {
Frederic Weisbecker1ad4ec02013-04-12 01:51:00 +02003770 /* check_preempt_curr() may use rq clock */
3771 update_rq_clock(rq);
Matt Flemingd8ac8972016-09-21 14:38:10 +01003772 ttwu_do_wakeup(rq, p, wake_flags, &rf);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003773 ret = 1;
3774 }
Peter Zijlstraeb580752015-07-31 21:28:18 +02003775 __task_rq_unlock(rq, &rf);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003776
3777 return ret;
3778}
3779
Peter Zijlstra317f3942011-04-05 17:23:58 +02003780#ifdef CONFIG_SMP
Peter Zijlstraa14886642020-05-26 18:11:04 +02003781void sched_ttwu_pending(void *arg)
Peter Zijlstra317f3942011-04-05 17:23:58 +02003782{
Peter Zijlstraa14886642020-05-26 18:11:04 +02003783 struct llist_node *llist = arg;
Peter Zijlstra317f3942011-04-05 17:23:58 +02003784 struct rq *rq = this_rq();
Byungchul Park73215842017-05-12 09:39:44 +09003785 struct task_struct *p, *t;
Matt Flemingd8ac8972016-09-21 14:38:10 +01003786 struct rq_flags rf;
Peter Zijlstra317f3942011-04-05 17:23:58 +02003787
Peter Zijlstrae3baac472014-06-04 10:31:18 -07003788 if (!llist)
3789 return;
3790
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003791 rq_lock_irqsave(rq, &rf);
Peter Zijlstra77558e42017-02-21 14:36:23 +01003792 update_rq_clock(rq);
Peter Zijlstra317f3942011-04-05 17:23:58 +02003793
Peter Zijlstra8c4890d2020-06-22 12:01:25 +02003794 llist_for_each_entry_safe(p, t, llist, wake_entry.llist) {
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02003795 if (WARN_ON_ONCE(p->on_cpu))
3796 smp_cond_load_acquire(&p->on_cpu, !VAL);
3797
3798 if (WARN_ON_ONCE(task_cpu(p) != cpu_of(rq)))
3799 set_task_cpu(p, cpu_of(rq));
3800
Byungchul Park73215842017-05-12 09:39:44 +09003801 ttwu_do_activate(rq, p, p->sched_remote_wakeup ? WF_MIGRATED : 0, &rf);
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02003802 }
Peter Zijlstra317f3942011-04-05 17:23:58 +02003803
Tianchen Dingd6962c42022-11-04 10:36:01 +08003804 /*
3805 * Must be after enqueueing at least once task such that
3806 * idle_cpu() does not observe a false-negative -- if it does,
3807 * it is possible for select_idle_siblings() to stack a number
3808 * of tasks on this CPU during that window.
3809 *
3810 * It is ok to clear ttwu_pending when another task pending.
3811 * We will receive IPI after local irq enabled and then enqueue it.
3812 * Since now nr_running > 0, idle_cpu() will always get correct result.
3813 */
3814 WRITE_ONCE(rq->ttwu_pending, 0);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003815 rq_unlock_irqrestore(rq, &rf);
Peter Zijlstra317f3942011-04-05 17:23:58 +02003816}
3817
Peter Zijlstrab2a02fc2020-05-26 18:11:01 +02003818void send_call_function_single_ipi(int cpu)
Peter Zijlstra317f3942011-04-05 17:23:58 +02003819{
Peter Zijlstrab2a02fc2020-05-26 18:11:01 +02003820 struct rq *rq = cpu_rq(cpu);
Peter Zijlstraf27dde82013-08-14 14:55:31 +02003821
Peter Zijlstrab2a02fc2020-05-26 18:11:01 +02003822 if (!set_nr_if_polling(rq->idle))
3823 arch_send_call_function_single_ipi(cpu);
3824 else
3825 trace_sched_wake_idle_without_ipi(cpu);
Peter Zijlstra317f3942011-04-05 17:23:58 +02003826}
3827
Mel Gorman2ebb1772020-05-24 21:29:56 +01003828/*
3829 * Queue a task on the target CPUs wake_list and wake the CPU via IPI if
3830 * necessary. The wakee CPU on receipt of the IPI will queue the task
3831 * via sched_ttwu_wakeup() for activation so the wakee incurs the cost
3832 * of the wakeup instead of the waker.
3833 */
3834static void __ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
Peter Zijlstra317f3942011-04-05 17:23:58 +02003835{
Peter Zijlstrae3baac472014-06-04 10:31:18 -07003836 struct rq *rq = cpu_rq(cpu);
3837
Peter Zijlstrab7e7ade2016-05-23 11:19:07 +02003838 p->sched_remote_wakeup = !!(wake_flags & WF_MIGRATED);
3839
Peter Zijlstra126c2092020-05-26 18:11:03 +02003840 WRITE_ONCE(rq->ttwu_pending, 1);
Peter Zijlstra8c4890d2020-06-22 12:01:25 +02003841 __smp_call_single_queue(cpu, &p->wake_entry.llist);
Peter Zijlstra317f3942011-04-05 17:23:58 +02003842}
Peter Zijlstrad6aa8f82011-05-26 14:21:33 +02003843
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08003844void wake_up_if_idle(int cpu)
3845{
3846 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003847 struct rq_flags rf;
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08003848
Andy Lutomirskifd7de1e82014-11-29 08:13:51 -08003849 rcu_read_lock();
3850
3851 if (!is_idle_task(rcu_dereference(rq->curr)))
3852 goto out;
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08003853
Peter Zijlstra8850cb62021-09-21 22:16:02 +02003854 rq_lock_irqsave(rq, &rf);
3855 if (is_idle_task(rq->curr))
3856 resched_curr(rq);
3857 /* Else CPU is not idle, do nothing here: */
3858 rq_unlock_irqrestore(rq, &rf);
Andy Lutomirskifd7de1e82014-11-29 08:13:51 -08003859
3860out:
3861 rcu_read_unlock();
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08003862}
3863
Peter Zijlstra39be3502012-01-26 12:44:34 +01003864bool cpus_share_cache(int this_cpu, int that_cpu)
Peter Zijlstra518cd622011-12-07 15:07:31 +01003865{
Vincent Donnefort42dc9382021-11-04 17:51:20 +00003866 if (this_cpu == that_cpu)
3867 return true;
3868
Peter Zijlstra518cd622011-12-07 15:07:31 +01003869 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
3870}
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01003871
Mel Gorman751d4cb2022-08-04 10:21:19 +01003872static inline bool ttwu_queue_cond(struct task_struct *p, int cpu)
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01003873{
Mel Gorman2ebb1772020-05-24 21:29:56 +01003874 /*
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01003875 * Do not complicate things with the async wake_list while the CPU is
3876 * in hotplug state.
3877 */
3878 if (!cpu_active(cpu))
3879 return false;
3880
Mel Gorman751d4cb2022-08-04 10:21:19 +01003881 /* Ensure the task will still be allowed to run on the CPU. */
3882 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
3883 return false;
3884
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01003885 /*
Mel Gorman2ebb1772020-05-24 21:29:56 +01003886 * If the CPU does not share cache, then queue the task on the
3887 * remote rqs wakelist to avoid accessing remote data.
3888 */
3889 if (!cpus_share_cache(smp_processor_id(), cpu))
3890 return true;
3891
Tianchen Dingf3dd3f62022-06-09 07:34:12 +08003892 if (cpu == smp_processor_id())
3893 return false;
3894
Mel Gorman2ebb1772020-05-24 21:29:56 +01003895 /*
Tianchen Dingf3dd3f62022-06-09 07:34:12 +08003896 * If the wakee cpu is idle, or the task is descheduling and the
3897 * only running task on the CPU, then use the wakelist to offload
3898 * the task activation to the idle (or soon-to-be-idle) CPU as
3899 * the current CPU is likely busy. nr_running is checked to
3900 * avoid unnecessary task stacking.
Tianchen Ding28156102022-06-09 07:34:11 +08003901 *
3902 * Note that we can only get here with (wakee) p->on_rq=0,
3903 * p->on_cpu can be whatever, we've done the dequeue, so
3904 * the wakee has been accounted out of ->nr_running.
Mel Gorman2ebb1772020-05-24 21:29:56 +01003905 */
Tianchen Dingf3dd3f62022-06-09 07:34:12 +08003906 if (!cpu_rq(cpu)->nr_running)
Mel Gorman2ebb1772020-05-24 21:29:56 +01003907 return true;
3908
3909 return false;
3910}
3911
3912static bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3913{
Mel Gorman751d4cb2022-08-04 10:21:19 +01003914 if (sched_feat(TTWU_QUEUE) && ttwu_queue_cond(p, cpu)) {
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01003915 sched_clock_cpu(cpu); /* Sync clocks across CPUs */
Mel Gorman2ebb1772020-05-24 21:29:56 +01003916 __ttwu_queue_wakelist(p, cpu, wake_flags);
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01003917 return true;
3918 }
3919
3920 return false;
3921}
Peter Zijlstra58877d32020-07-02 14:52:11 +02003922
3923#else /* !CONFIG_SMP */
3924
3925static inline bool ttwu_queue_wakelist(struct task_struct *p, int cpu, int wake_flags)
3926{
3927 return false;
3928}
3929
Peter Zijlstrad6aa8f82011-05-26 14:21:33 +02003930#endif /* CONFIG_SMP */
Peter Zijlstra317f3942011-04-05 17:23:58 +02003931
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02003932static void ttwu_queue(struct task_struct *p, int cpu, int wake_flags)
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003933{
3934 struct rq *rq = cpu_rq(cpu);
Matt Flemingd8ac8972016-09-21 14:38:10 +01003935 struct rq_flags rf;
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02003936
Mel Gorman2ebb1772020-05-24 21:29:56 +01003937 if (ttwu_queue_wakelist(p, cpu, wake_flags))
Peter Zijlstra317f3942011-04-05 17:23:58 +02003938 return;
Peter Zijlstra317f3942011-04-05 17:23:58 +02003939
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003940 rq_lock(rq, &rf);
Peter Zijlstra77558e42017-02-21 14:36:23 +01003941 update_rq_clock(rq);
Matt Flemingd8ac8972016-09-21 14:38:10 +01003942 ttwu_do_activate(rq, p, wake_flags, &rf);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02003943 rq_unlock(rq, &rf);
Tejun Heo9ed38112009-12-03 15:08:03 +09003944}
3945
Peter Zijlstra8643cda2015-11-17 19:01:11 +01003946/*
Thomas Gleixner43295d72021-08-15 23:27:40 +02003947 * Invoked from try_to_wake_up() to check whether the task can be woken up.
3948 *
3949 * The caller holds p::pi_lock if p != current or has preemption
3950 * disabled when p == current.
Thomas Gleixner5f220be2021-08-15 23:27:44 +02003951 *
3952 * The rules of PREEMPT_RT saved_state:
3953 *
3954 * The related locking code always holds p::pi_lock when updating
3955 * p::saved_state, which means the code is fully serialized in both cases.
3956 *
3957 * The lock wait and lock wakeups happen via TASK_RTLOCK_WAIT. No other
3958 * bits set. This allows to distinguish all wakeup scenarios.
Thomas Gleixner43295d72021-08-15 23:27:40 +02003959 */
3960static __always_inline
3961bool ttwu_state_match(struct task_struct *p, unsigned int state, int *success)
3962{
Thomas Gleixner5f220be2021-08-15 23:27:44 +02003963 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)) {
3964 WARN_ON_ONCE((state & TASK_RTLOCK_WAIT) &&
3965 state != TASK_RTLOCK_WAIT);
3966 }
3967
Thomas Gleixner43295d72021-08-15 23:27:40 +02003968 if (READ_ONCE(p->__state) & state) {
3969 *success = 1;
3970 return true;
3971 }
Thomas Gleixner5f220be2021-08-15 23:27:44 +02003972
3973#ifdef CONFIG_PREEMPT_RT
3974 /*
3975 * Saved state preserves the task state across blocking on
3976 * an RT lock. If the state matches, set p::saved_state to
3977 * TASK_RUNNING, but do not wake the task because it waits
3978 * for a lock wakeup. Also indicate success because from
3979 * the regular waker's point of view this has succeeded.
3980 *
3981 * After acquiring the lock the task will restore p::__state
3982 * from p::saved_state which ensures that the regular
3983 * wakeup is not lost. The restore will also set
3984 * p::saved_state to TASK_RUNNING so any further tests will
3985 * not result in false positives vs. @success
3986 */
3987 if (p->saved_state & state) {
3988 p->saved_state = TASK_RUNNING;
3989 *success = 1;
3990 }
3991#endif
Thomas Gleixner43295d72021-08-15 23:27:40 +02003992 return false;
3993}
3994
3995/*
Peter Zijlstra8643cda2015-11-17 19:01:11 +01003996 * Notes on Program-Order guarantees on SMP systems.
3997 *
3998 * MIGRATION
3999 *
4000 * The basic program-order guarantee on SMP systems is that when a task [t]
Ingo Molnard1ccc662017-02-01 11:46:42 +01004001 * migrates, all its activity on its old CPU [c0] happens-before any subsequent
4002 * execution on its new CPU [c1].
Peter Zijlstra8643cda2015-11-17 19:01:11 +01004003 *
4004 * For migration (of runnable tasks) this is provided by the following means:
4005 *
4006 * A) UNLOCK of the rq(c0)->lock scheduling out task t
4007 * B) migration for t is required to synchronize *both* rq(c0)->lock and
4008 * rq(c1)->lock (if not at the same time, then in that order).
4009 * C) LOCK of the rq(c1)->lock scheduling in task
4010 *
Andrea Parri7696f992018-07-16 11:06:03 -07004011 * Release/acquire chaining guarantees that B happens after A and C after B.
Ingo Molnard1ccc662017-02-01 11:46:42 +01004012 * Note: the CPU doing B need not be c0 or c1
Peter Zijlstra8643cda2015-11-17 19:01:11 +01004013 *
4014 * Example:
4015 *
4016 * CPU0 CPU1 CPU2
4017 *
4018 * LOCK rq(0)->lock
4019 * sched-out X
4020 * sched-in Y
4021 * UNLOCK rq(0)->lock
4022 *
4023 * LOCK rq(0)->lock // orders against CPU0
4024 * dequeue X
4025 * UNLOCK rq(0)->lock
4026 *
4027 * LOCK rq(1)->lock
4028 * enqueue X
4029 * UNLOCK rq(1)->lock
4030 *
4031 * LOCK rq(1)->lock // orders against CPU2
4032 * sched-out Z
4033 * sched-in X
4034 * UNLOCK rq(1)->lock
4035 *
4036 *
4037 * BLOCKING -- aka. SLEEP + WAKEUP
4038 *
4039 * For blocking we (obviously) need to provide the same guarantee as for
4040 * migration. However the means are completely different as there is no lock
4041 * chain to provide order. Instead we do:
4042 *
Peter Zijlstra58877d32020-07-02 14:52:11 +02004043 * 1) smp_store_release(X->on_cpu, 0) -- finish_task()
4044 * 2) smp_cond_load_acquire(!X->on_cpu) -- try_to_wake_up()
Peter Zijlstra8643cda2015-11-17 19:01:11 +01004045 *
4046 * Example:
4047 *
4048 * CPU0 (schedule) CPU1 (try_to_wake_up) CPU2 (schedule)
4049 *
4050 * LOCK rq(0)->lock LOCK X->pi_lock
4051 * dequeue X
4052 * sched-out X
4053 * smp_store_release(X->on_cpu, 0);
4054 *
Peter Zijlstra1f03e8d2016-04-04 10:57:12 +02004055 * smp_cond_load_acquire(&X->on_cpu, !VAL);
Peter Zijlstra8643cda2015-11-17 19:01:11 +01004056 * X->state = WAKING
4057 * set_task_cpu(X,2)
4058 *
4059 * LOCK rq(2)->lock
4060 * enqueue X
4061 * X->state = RUNNING
4062 * UNLOCK rq(2)->lock
4063 *
4064 * LOCK rq(2)->lock // orders against CPU1
4065 * sched-out Z
4066 * sched-in X
4067 * UNLOCK rq(2)->lock
4068 *
4069 * UNLOCK X->pi_lock
4070 * UNLOCK rq(0)->lock
4071 *
4072 *
Andrea Parri7696f992018-07-16 11:06:03 -07004073 * However, for wakeups there is a second guarantee we must provide, namely we
4074 * must ensure that CONDITION=1 done by the caller can not be reordered with
4075 * accesses to the task state; see try_to_wake_up() and set_current_state().
Peter Zijlstra8643cda2015-11-17 19:01:11 +01004076 */
4077
Tejun Heo9ed38112009-12-03 15:08:03 +09004078/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07004079 * try_to_wake_up - wake up a thread
Tejun Heo9ed38112009-12-03 15:08:03 +09004080 * @p: the thread to be awakened
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 * @state: the mask of task states that can be woken
Tejun Heo9ed38112009-12-03 15:08:03 +09004082 * @wake_flags: wake modifier flags (WF_*)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 *
Peter Zijlstra58877d32020-07-02 14:52:11 +02004084 * Conceptually does:
4085 *
4086 * If (@state & @p->state) @p->state = TASK_RUNNING.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 *
Peter Zijlstraa2250232016-10-19 15:45:27 +02004088 * If the task was not queued/runnable, also place it back on a runqueue.
4089 *
Peter Zijlstra58877d32020-07-02 14:52:11 +02004090 * This function is atomic against schedule() which would dequeue the task.
Peter Zijlstraa2250232016-10-19 15:45:27 +02004091 *
Peter Zijlstra58877d32020-07-02 14:52:11 +02004092 * It issues a full memory barrier before accessing @p->state, see the comment
4093 * with set_current_state().
4094 *
4095 * Uses p->pi_lock to serialize against concurrent wake-ups.
4096 *
4097 * Relies on p->pi_lock stabilizing:
4098 * - p->sched_class
4099 * - p->cpus_ptr
4100 * - p->sched_task_group
4101 * in order to do migration, see its use of select_task_rq()/set_task_cpu().
4102 *
4103 * Tries really hard to only take one task_rq(p)->lock for performance.
4104 * Takes rq->lock in:
4105 * - ttwu_runnable() -- old rq, unavoidable, see comment there;
4106 * - ttwu_queue() -- new rq, for enqueue of the task;
4107 * - psi_ttwu_dequeue() -- much sadness :-( accounting will kill us.
4108 *
4109 * As a consequence we race really badly with just about everything. See the
4110 * many memory barriers and their comments for details.
Andrea Parri7696f992018-07-16 11:06:03 -07004111 *
Peter Zijlstraa2250232016-10-19 15:45:27 +02004112 * Return: %true if @p->state changes (an actual wakeup was done),
4113 * %false otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114 */
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02004115static int
4116try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118 unsigned long flags;
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02004119 int cpu, success = 0;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02004120
Peter Zijlstrae3d85482019-07-10 12:57:36 +02004121 preempt_disable();
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004122 if (p == current) {
4123 /*
4124 * We're waking current, this means 'p->on_rq' and 'task_cpu(p)
4125 * == smp_processor_id()'. Together this means we can special
Peter Zijlstra58877d32020-07-02 14:52:11 +02004126 * case the whole 'p->on_rq && ttwu_runnable()' case below
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004127 * without taking any locks.
4128 *
4129 * In particular:
4130 * - we rely on Program-Order guarantees for all the ordering,
4131 * - we're serialized against set_special_state() by virtue of
4132 * it disabling IRQs (this allows not taking ->pi_lock).
4133 */
Thomas Gleixner43295d72021-08-15 23:27:40 +02004134 if (!ttwu_state_match(p, state, &success))
Peter Zijlstrae3d85482019-07-10 12:57:36 +02004135 goto out;
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004136
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004137 trace_sched_waking(p);
Peter Zijlstra2f064a52021-06-11 10:28:17 +02004138 WRITE_ONCE(p->__state, TASK_RUNNING);
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004139 trace_sched_wakeup(p);
4140 goto out;
4141 }
4142
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02004143 /*
4144 * If we are going to wake up a thread waiting for CONDITION we
4145 * need to ensure that CONDITION=1 done by the caller can not be
Peter Zijlstra58877d32020-07-02 14:52:11 +02004146 * reordered with p->state check below. This pairs with smp_store_mb()
4147 * in set_current_state() that the waiting thread does.
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02004148 */
Peter Zijlstra013fdb82011-04-05 17:23:45 +02004149 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstrad89e588c2016-09-05 11:37:53 +02004150 smp_mb__after_spinlock();
Thomas Gleixner43295d72021-08-15 23:27:40 +02004151 if (!ttwu_state_match(p, state, &success))
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004152 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02004154 trace_sched_waking(p);
4155
Balbir Singh135e8c92016-09-05 13:16:40 +10004156 /*
4157 * Ensure we load p->on_rq _after_ p->state, otherwise it would
4158 * be possible to, falsely, observe p->on_rq == 0 and get stuck
4159 * in smp_cond_load_acquire() below.
4160 *
Andrea Parri3d85b272018-07-16 11:06:02 -07004161 * sched_ttwu_pending() try_to_wake_up()
4162 * STORE p->on_rq = 1 LOAD p->state
4163 * UNLOCK rq->lock
4164 *
4165 * __schedule() (switch to task 'p')
4166 * LOCK rq->lock smp_rmb();
4167 * smp_mb__after_spinlock();
4168 * UNLOCK rq->lock
Balbir Singh135e8c92016-09-05 13:16:40 +10004169 *
4170 * [task p]
Andrea Parri3d85b272018-07-16 11:06:02 -07004171 * STORE p->state = UNINTERRUPTIBLE LOAD p->on_rq
Balbir Singh135e8c92016-09-05 13:16:40 +10004172 *
Andrea Parri3d85b272018-07-16 11:06:02 -07004173 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4174 * __schedule(). See the comment for smp_mb__after_spinlock().
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004175 *
4176 * A similar smb_rmb() lives in try_invoke_on_locked_down_task().
Balbir Singh135e8c92016-09-05 13:16:40 +10004177 */
4178 smp_rmb();
Peter Zijlstra58877d32020-07-02 14:52:11 +02004179 if (READ_ONCE(p->on_rq) && ttwu_runnable(p, wake_flags))
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004180 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181
4182#ifdef CONFIG_SMP
Peter Zijlstrae9c84312009-09-15 14:43:03 +02004183 /*
Peter Zijlstraecf7d012015-10-07 14:14:13 +02004184 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
4185 * possible to, falsely, observe p->on_cpu == 0.
4186 *
4187 * One must be running (->on_cpu == 1) in order to remove oneself
4188 * from the runqueue.
4189 *
Andrea Parri3d85b272018-07-16 11:06:02 -07004190 * __schedule() (switch to task 'p') try_to_wake_up()
4191 * STORE p->on_cpu = 1 LOAD p->on_rq
4192 * UNLOCK rq->lock
Peter Zijlstraecf7d012015-10-07 14:14:13 +02004193 *
Andrea Parri3d85b272018-07-16 11:06:02 -07004194 * __schedule() (put 'p' to sleep)
4195 * LOCK rq->lock smp_rmb();
4196 * smp_mb__after_spinlock();
4197 * STORE p->on_rq = 0 LOAD p->on_cpu
4198 *
4199 * Pairs with the LOCK+smp_mb__after_spinlock() on rq->lock in
4200 * __schedule(). See the comment for smp_mb__after_spinlock().
Peter Zijlstradbfb0892020-07-03 12:40:33 +02004201 *
4202 * Form a control-dep-acquire with p->on_rq == 0 above, to ensure
4203 * schedule()'s deactivate_task() has 'happened' and p will no longer
4204 * care about it's own p->state. See the comment in __schedule().
Peter Zijlstraecf7d012015-10-07 14:14:13 +02004205 */
Peter Zijlstradbfb0892020-07-03 12:40:33 +02004206 smp_acquire__after_ctrl_dep();
4207
4208 /*
4209 * We're doing the wakeup (@success == 1), they did a dequeue (p->on_rq
4210 * == 0), which means we need to do an enqueue, change p->state to
4211 * TASK_WAKING such that we can unlock p->pi_lock before doing the
4212 * enqueue, such as ttwu_queue_wakelist().
4213 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02004214 WRITE_ONCE(p->__state, TASK_WAKING);
Peter Zijlstraecf7d012015-10-07 14:14:13 +02004215
4216 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01004217 * If the owning (remote) CPU is still in the middle of schedule() with
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01004218 * this task as prev, considering queueing p on the remote CPUs wake_list
4219 * which potentially sends an IPI instead of spinning on p->on_cpu to
4220 * let the waker make forward progress. This is safe because IRQs are
4221 * disabled and the IPI will deliver after on_cpu is cleared.
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02004222 *
4223 * Ensure we load task_cpu(p) after p->on_cpu:
4224 *
4225 * set_task_cpu(p, cpu);
4226 * STORE p->cpu = @cpu
4227 * __schedule() (switch to task 'p')
4228 * LOCK rq->lock
4229 * smp_mb__after_spin_lock() smp_cond_load_acquire(&p->on_cpu)
4230 * STORE p->on_cpu = 1 LOAD p->cpu
4231 *
4232 * to ensure we observe the correct CPU on which the task is currently
4233 * scheduling.
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01004234 */
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02004235 if (smp_load_acquire(&p->on_cpu) &&
Tianchen Dingf3dd3f62022-06-09 07:34:12 +08004236 ttwu_queue_wakelist(p, task_cpu(p), wake_flags))
Peter Zijlstrac6e7bd72020-05-24 21:29:55 +01004237 goto unlock;
4238
4239 /*
4240 * If the owning (remote) CPU is still in the middle of schedule() with
Tal Zussmanb19a8882020-11-12 19:51:56 -05004241 * this task as prev, wait until it's done referencing the task.
Peter Zijlstrab75a2252015-10-06 14:36:17 +02004242 *
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02004243 * Pairs with the smp_store_release() in finish_task().
Peter Zijlstrab75a2252015-10-06 14:36:17 +02004244 *
4245 * This ensures that tasks getting woken will be fully ordered against
4246 * their previous state and preserve Program Order.
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02004247 */
Peter Zijlstra1f03e8d2016-04-04 10:57:12 +02004248 smp_cond_load_acquire(&p->on_cpu, !VAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249
Valentin Schneider3aef15512020-11-02 18:45:13 +00004250 cpu = select_task_rq(p, p->wake_cpu, wake_flags | WF_TTWU);
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02004251 if (task_cpu(p) != cpu) {
Peter Zijlstraec618b82020-09-24 13:50:42 +02004252 if (p->in_iowait) {
4253 delayacct_blkio_end(p);
4254 atomic_dec(&task_rq(p)->nr_iowait);
4255 }
4256
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02004257 wake_flags |= WF_MIGRATED;
Johannes Weinereb414682018-10-26 15:06:27 -07004258 psi_ttwu_dequeue(p);
Mike Galbraith055a0082009-11-12 11:07:44 +01004259 set_task_cpu(p, cpu);
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02004260 }
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02004261#else
4262 cpu = task_cpu(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264
Peter Zijlstrab5179ac2016-05-11 16:10:34 +02004265 ttwu_queue(p, cpu, wake_flags);
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004266unlock:
Peter Zijlstra013fdb82011-04-05 17:23:45 +02004267 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Peter Zijlstraaacedf22019-06-07 15:39:49 +02004268out:
4269 if (success)
Peter Zijlstrab6e13e82020-06-22 12:01:23 +02004270 ttwu_stat(p, task_cpu(p), wake_flags);
Peter Zijlstrae3d85482019-07-10 12:57:36 +02004271 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272
4273 return success;
4274}
4275
Peter Zijlstra91dabf32022-10-26 13:43:00 +02004276static bool __task_needs_rq_lock(struct task_struct *p)
4277{
4278 unsigned int state = READ_ONCE(p->__state);
4279
4280 /*
4281 * Since pi->lock blocks try_to_wake_up(), we don't need rq->lock when
4282 * the task is blocked. Make sure to check @state since ttwu() can drop
4283 * locks at the end, see ttwu_queue_wakelist().
4284 */
4285 if (state == TASK_RUNNING || state == TASK_WAKING)
4286 return true;
4287
4288 /*
4289 * Ensure we load p->on_rq after p->__state, otherwise it would be
4290 * possible to, falsely, observe p->on_rq == 0.
4291 *
4292 * See try_to_wake_up() for a longer comment.
4293 */
4294 smp_rmb();
4295 if (p->on_rq)
4296 return true;
4297
4298#ifdef CONFIG_SMP
4299 /*
4300 * Ensure the task has finished __schedule() and will not be referenced
4301 * anymore. Again, see try_to_wake_up() for a longer comment.
4302 */
4303 smp_rmb();
4304 smp_cond_load_acquire(&p->on_cpu, !VAL);
4305#endif
4306
4307 return false;
4308}
4309
David Howells50fa6102009-04-28 15:01:38 +01004310/**
Peter Zijlstra9b3c4ab2021-09-21 21:54:32 +02004311 * task_call_func - Invoke a function on task in fixed state
Peter Zijlstra1b7af292020-08-29 10:22:24 -07004312 * @p: Process for which the function is to be invoked, can be @current.
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004313 * @func: Function to invoke.
4314 * @arg: Argument to function.
4315 *
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004316 * Fix the task in it's current state by avoiding wakeups and or rq operations
4317 * and call @func(@arg) on it. This function can use ->on_rq and task_curr()
4318 * to work out what the state is, if required. Given that @func can be invoked
4319 * with a runqueue lock held, it had better be quite lightweight.
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004320 *
4321 * Returns:
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004322 * Whatever @func returns
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004323 */
Peter Zijlstra9b3c4ab2021-09-21 21:54:32 +02004324int task_call_func(struct task_struct *p, task_call_f func, void *arg)
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004325{
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004326 struct rq *rq = NULL;
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004327 struct rq_flags rf;
Peter Zijlstra9b3c4ab2021-09-21 21:54:32 +02004328 int ret;
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004329
Peter Zijlstra1b7af292020-08-29 10:22:24 -07004330 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004331
Peter Zijlstra91dabf32022-10-26 13:43:00 +02004332 if (__task_needs_rq_lock(p))
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004333 rq = __task_rq_lock(p, &rf);
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004334
4335 /*
4336 * At this point the task is pinned; either:
4337 * - blocked and we're holding off wakeups (pi->lock)
4338 * - woken, and we're holding off enqueue (rq->lock)
4339 * - queued, and we're holding off schedule (rq->lock)
4340 * - running, and we're holding off de-schedule (rq->lock)
4341 *
4342 * The called function (@func) can use: task_curr(), p->on_rq and
4343 * p->__state to differentiate between these states.
4344 */
4345 ret = func(p, arg);
4346
4347 if (rq)
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004348 rq_unlock(rq, &rf);
Peter Zijlstraf6ac18f2021-09-22 10:14:15 +02004349
Peter Zijlstra1b7af292020-08-29 10:22:24 -07004350 raw_spin_unlock_irqrestore(&p->pi_lock, rf.flags);
Paul E. McKenney2beaf322020-03-11 14:23:21 -07004351 return ret;
4352}
4353
4354/**
Paul E. McKenneye386b672022-06-02 17:30:01 -07004355 * cpu_curr_snapshot - Return a snapshot of the currently running task
4356 * @cpu: The CPU on which to snapshot the task.
4357 *
4358 * Returns the task_struct pointer of the task "currently" running on
4359 * the specified CPU. If the same task is running on that CPU throughout,
4360 * the return value will be a pointer to that task's task_struct structure.
4361 * If the CPU did any context switches even vaguely concurrently with the
4362 * execution of this function, the return value will be a pointer to the
4363 * task_struct structure of a randomly chosen task that was running on
4364 * that CPU somewhere around the time that this function was executing.
4365 *
4366 * If the specified CPU was offline, the return value is whatever it
4367 * is, perhaps a pointer to the task_struct structure of that CPU's idle
4368 * task, but there is no guarantee. Callers wishing a useful return
4369 * value must take some action to ensure that the specified CPU remains
4370 * online throughout.
4371 *
4372 * This function executes full memory barriers before and after fetching
4373 * the pointer, which permits the caller to confine this function's fetch
4374 * with respect to the caller's accesses to other shared variables.
4375 */
4376struct task_struct *cpu_curr_snapshot(int cpu)
4377{
4378 struct task_struct *t;
4379
4380 smp_mb(); /* Pairing determined by caller's synchronization design. */
4381 t = rcu_dereference(cpu_curr(cpu));
4382 smp_mb(); /* Pairing determined by caller's synchronization design. */
4383 return t;
4384}
4385
4386/**
David Howells50fa6102009-04-28 15:01:38 +01004387 * wake_up_process - Wake up a specific process
4388 * @p: The process to be woken up.
4389 *
4390 * Attempt to wake up the nominated process and move it to the set of runnable
Yacine Belkadie69f6182013-07-12 20:45:47 +02004391 * processes.
4392 *
4393 * Return: 1 if the process was woken up, 0 if it was already running.
David Howells50fa6102009-04-28 15:01:38 +01004394 *
Andrea Parri7696f992018-07-16 11:06:03 -07004395 * This function executes a full memory barrier before accessing the task state.
David Howells50fa6102009-04-28 15:01:38 +01004396 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004397int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004398{
Oleg Nesterov9067ac852013-01-21 20:48:17 +01004399 return try_to_wake_up(p, TASK_NORMAL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004400}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004401EXPORT_SYMBOL(wake_up_process);
4402
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004403int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404{
4405 return try_to_wake_up(p, state, 0);
4406}
4407
Linus Torvalds1da177e2005-04-16 15:20:36 -07004408/*
4409 * Perform scheduler related setup for a newly forked process p.
4410 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02004411 *
4412 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004413 */
Rik van Riel5e1576e2013-10-07 11:29:26 +01004414static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415{
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02004416 p->on_rq = 0;
4417
4418 p->se.on_rq = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02004419 p->se.exec_start = 0;
4420 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02004421 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01004422 p->se.nr_migrations = 0;
Peter Zijlstrada7a7352011-01-17 17:03:27 +01004423 p->se.vruntime = 0;
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02004424 INIT_LIST_HEAD(&p->se.group_node);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02004425
Byungchul Parkad936d82015-10-24 01:16:19 +09004426#ifdef CONFIG_FAIR_GROUP_SCHED
4427 p->se.cfs_rq = NULL;
4428#endif
4429
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02004430#ifdef CONFIG_SCHEDSTATS
Mel Gormancb251762016-02-05 09:08:36 +00004431 /* Even if schedstat is disabled, there should not be garbage */
Yafang Shaoceeadb82021-09-05 14:35:41 +00004432 memset(&p->stats, 0, sizeof(p->stats));
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02004433#endif
Nick Piggin476d1392005-06-25 14:57:29 -07004434
Dario Faggioliaab03e02013-11-28 11:14:43 +01004435 RB_CLEAR_NODE(&p->dl.rb_node);
Peter Zijlstra40767b02015-01-28 15:08:03 +01004436 init_dl_task_timer(&p->dl);
Luca Abeni209a0cb2017-05-18 22:13:29 +02004437 init_dl_inactive_task_timer(&p->dl);
Juri Lellia5e7be32014-09-19 10:22:39 +01004438 __dl_clear_params(p);
Dario Faggioliaab03e02013-11-28 11:14:43 +01004439
Peter Zijlstrafa717062008-01-25 21:08:27 +01004440 INIT_LIST_HEAD(&p->rt.run_list);
Peter Zijlstraff77e462016-01-18 15:27:07 +01004441 p->rt.timeout = 0;
4442 p->rt.time_slice = sched_rr_timeslice;
4443 p->rt.on_rq = 0;
4444 p->rt.on_list = 0;
Nick Piggin476d1392005-06-25 14:57:29 -07004445
Avi Kivitye107be32007-07-26 13:40:43 +02004446#ifdef CONFIG_PREEMPT_NOTIFIERS
4447 INIT_HLIST_HEAD(&p->preempt_notifiers);
4448#endif
Peter Zijlstracbee9f82012-10-25 14:16:43 +02004449
Mel Gorman5e1f0f02019-03-05 15:45:41 -08004450#ifdef CONFIG_COMPACTION
4451 p->capture_control = NULL;
4452#endif
Mel Gorman13784472018-05-04 16:41:09 +01004453 init_numa_balancing(clone_flags, p);
Peter Zijlstraa14886642020-05-26 18:11:04 +02004454#ifdef CONFIG_SMP
Peter Zijlstra8c4890d2020-06-22 12:01:25 +02004455 p->wake_entry.u_flags = CSD_TYPE_TTWU;
Peter Zijlstra6d337ea2020-09-18 17:24:31 +02004456 p->migration_pending = NULL;
Peter Zijlstraa14886642020-05-26 18:11:04 +02004457#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004458}
4459
Srikar Dronamraju2a595722015-08-11 21:54:21 +05304460DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
4461
Mel Gorman1a687c22012-11-22 11:16:36 +00004462#ifdef CONFIG_NUMA_BALANCING
Mel Gorman3105b862012-11-23 11:23:49 +00004463
Huang Yingc574bbe2022-03-22 14:46:23 -07004464int sysctl_numa_balancing_mode;
4465
4466static void __set_numabalancing_state(bool enabled)
Mel Gorman3105b862012-11-23 11:23:49 +00004467{
Srikar Dronamraju2a595722015-08-11 21:54:21 +05304468 if (enabled)
4469 static_branch_enable(&sched_numa_balancing);
4470 else
4471 static_branch_disable(&sched_numa_balancing);
Mel Gorman3105b862012-11-23 11:23:49 +00004472}
Andi Kleen54a43d52014-01-23 15:53:13 -08004473
Huang Yingc574bbe2022-03-22 14:46:23 -07004474void set_numabalancing_state(bool enabled)
4475{
4476 if (enabled)
4477 sysctl_numa_balancing_mode = NUMA_BALANCING_NORMAL;
4478 else
4479 sysctl_numa_balancing_mode = NUMA_BALANCING_DISABLED;
4480 __set_numabalancing_state(enabled);
4481}
4482
Andi Kleen54a43d52014-01-23 15:53:13 -08004483#ifdef CONFIG_PROC_SYSCTL
Huang Yingc959924b2022-07-13 16:39:53 +08004484static void reset_memory_tiering(void)
4485{
4486 struct pglist_data *pgdat;
4487
4488 for_each_online_pgdat(pgdat) {
4489 pgdat->nbp_threshold = 0;
4490 pgdat->nbp_th_nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
4491 pgdat->nbp_th_start = jiffies_to_msecs(jiffies);
4492 }
4493}
4494
Kefeng Wang0dff89c2022-09-08 20:07:14 +08004495static int sysctl_numa_balancing(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02004496 void *buffer, size_t *lenp, loff_t *ppos)
Andi Kleen54a43d52014-01-23 15:53:13 -08004497{
4498 struct ctl_table t;
4499 int err;
Huang Yingc574bbe2022-03-22 14:46:23 -07004500 int state = sysctl_numa_balancing_mode;
Andi Kleen54a43d52014-01-23 15:53:13 -08004501
4502 if (write && !capable(CAP_SYS_ADMIN))
4503 return -EPERM;
4504
4505 t = *table;
4506 t.data = &state;
4507 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4508 if (err < 0)
4509 return err;
Huang Yingc574bbe2022-03-22 14:46:23 -07004510 if (write) {
Huang Yingc959924b2022-07-13 16:39:53 +08004511 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
4512 (state & NUMA_BALANCING_MEMORY_TIERING))
4513 reset_memory_tiering();
Huang Yingc574bbe2022-03-22 14:46:23 -07004514 sysctl_numa_balancing_mode = state;
4515 __set_numabalancing_state(state);
4516 }
Andi Kleen54a43d52014-01-23 15:53:13 -08004517 return err;
4518}
4519#endif
4520#endif
Mel Gorman1a687c22012-11-22 11:16:36 +00004521
Mel Gormancb251762016-02-05 09:08:36 +00004522#ifdef CONFIG_SCHEDSTATS
Josh Poimboeuf4698f882016-06-07 14:43:16 -05004523
4524DEFINE_STATIC_KEY_FALSE(sched_schedstats);
Josh Poimboeuf4698f882016-06-07 14:43:16 -05004525
Mel Gormancb251762016-02-05 09:08:36 +00004526static void set_schedstats(bool enabled)
4527{
4528 if (enabled)
4529 static_branch_enable(&sched_schedstats);
4530 else
4531 static_branch_disable(&sched_schedstats);
4532}
4533
4534void force_schedstat_enabled(void)
4535{
4536 if (!schedstat_enabled()) {
4537 pr_info("kernel profiling enabled schedstats, disable via kernel.sched_schedstats.\n");
4538 static_branch_enable(&sched_schedstats);
4539 }
4540}
4541
4542static int __init setup_schedstats(char *str)
4543{
4544 int ret = 0;
4545 if (!str)
4546 goto out;
4547
4548 if (!strcmp(str, "enable")) {
Eric Dumazet1faa4912021-06-02 04:21:08 -07004549 set_schedstats(true);
Mel Gormancb251762016-02-05 09:08:36 +00004550 ret = 1;
4551 } else if (!strcmp(str, "disable")) {
Eric Dumazet1faa4912021-06-02 04:21:08 -07004552 set_schedstats(false);
Mel Gormancb251762016-02-05 09:08:36 +00004553 ret = 1;
4554 }
4555out:
4556 if (!ret)
4557 pr_warn("Unable to parse schedstats=\n");
4558
4559 return ret;
4560}
4561__setup("schedstats=", setup_schedstats);
4562
4563#ifdef CONFIG_PROC_SYSCTL
Zhen Nif5ef06d2022-02-15 19:45:58 +08004564static int sysctl_schedstats(struct ctl_table *table, int write, void *buffer,
Christoph Hellwig32927392020-04-24 08:43:38 +02004565 size_t *lenp, loff_t *ppos)
Mel Gormancb251762016-02-05 09:08:36 +00004566{
4567 struct ctl_table t;
4568 int err;
4569 int state = static_branch_likely(&sched_schedstats);
4570
4571 if (write && !capable(CAP_SYS_ADMIN))
4572 return -EPERM;
4573
4574 t = *table;
4575 t.data = &state;
4576 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
4577 if (err < 0)
4578 return err;
4579 if (write)
4580 set_schedstats(state);
4581 return err;
4582}
Josh Poimboeuf4698f882016-06-07 14:43:16 -05004583#endif /* CONFIG_PROC_SYSCTL */
Josh Poimboeuf4698f882016-06-07 14:43:16 -05004584#endif /* CONFIG_SCHEDSTATS */
Mel Gormancb251762016-02-05 09:08:36 +00004585
Zhen Ni3267e012022-02-15 19:46:02 +08004586#ifdef CONFIG_SYSCTL
4587static struct ctl_table sched_core_sysctls[] = {
4588#ifdef CONFIG_SCHEDSTATS
Zhen Nif5ef06d2022-02-15 19:45:58 +08004589 {
4590 .procname = "sched_schedstats",
4591 .data = NULL,
4592 .maxlen = sizeof(unsigned int),
4593 .mode = 0644,
4594 .proc_handler = sysctl_schedstats,
4595 .extra1 = SYSCTL_ZERO,
4596 .extra2 = SYSCTL_ONE,
4597 },
Zhen Ni3267e012022-02-15 19:46:02 +08004598#endif /* CONFIG_SCHEDSTATS */
4599#ifdef CONFIG_UCLAMP_TASK
4600 {
4601 .procname = "sched_util_clamp_min",
4602 .data = &sysctl_sched_uclamp_util_min,
4603 .maxlen = sizeof(unsigned int),
4604 .mode = 0644,
4605 .proc_handler = sysctl_sched_uclamp_handler,
4606 },
4607 {
4608 .procname = "sched_util_clamp_max",
4609 .data = &sysctl_sched_uclamp_util_max,
4610 .maxlen = sizeof(unsigned int),
4611 .mode = 0644,
4612 .proc_handler = sysctl_sched_uclamp_handler,
4613 },
4614 {
4615 .procname = "sched_util_clamp_min_rt_default",
4616 .data = &sysctl_sched_uclamp_util_min_rt_default,
4617 .maxlen = sizeof(unsigned int),
4618 .mode = 0644,
4619 .proc_handler = sysctl_sched_uclamp_handler,
4620 },
4621#endif /* CONFIG_UCLAMP_TASK */
Kefeng Wang0dff89c2022-09-08 20:07:14 +08004622#ifdef CONFIG_NUMA_BALANCING
4623 {
4624 .procname = "numa_balancing",
4625 .data = NULL, /* filled in by handler */
4626 .maxlen = sizeof(unsigned int),
4627 .mode = 0644,
4628 .proc_handler = sysctl_numa_balancing,
4629 .extra1 = SYSCTL_ZERO,
4630 .extra2 = SYSCTL_FOUR,
4631 },
4632#endif /* CONFIG_NUMA_BALANCING */
Zhen Nif5ef06d2022-02-15 19:45:58 +08004633 {}
4634};
Zhen Ni3267e012022-02-15 19:46:02 +08004635static int __init sched_core_sysctl_init(void)
Zhen Nif5ef06d2022-02-15 19:45:58 +08004636{
Zhen Ni3267e012022-02-15 19:46:02 +08004637 register_sysctl_init("kernel", sched_core_sysctls);
Zhen Nif5ef06d2022-02-15 19:45:58 +08004638 return 0;
4639}
Zhen Ni3267e012022-02-15 19:46:02 +08004640late_initcall(sched_core_sysctl_init);
4641#endif /* CONFIG_SYSCTL */
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02004642
Lennart Poetteringca94c442009-06-15 17:17:47 +02004643/*
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02004644 * fork()/clone()-time setup:
4645 */
Dario Faggioliaab03e02013-11-28 11:14:43 +01004646int sched_fork(unsigned long clone_flags, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004647{
Rik van Riel5e1576e2013-10-07 11:29:26 +01004648 __sched_fork(clone_flags, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004649 /*
Peter Zijlstra7dc603c2016-06-16 13:29:28 +02004650 * We mark the process as NEW here. This guarantees that
Linus Torvalds1da177e2005-04-16 15:20:36 -07004651 * nobody will actually run it, and a signal or other external
4652 * event cannot wake it up and insert it on the runqueue either.
4653 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02004654 p->__state = TASK_NEW;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004655
4656 /*
4657 * Make sure we do not leak PI boosting priority to the child.
4658 */
4659 p->prio = current->normal_prio;
4660
Patrick Bellasie8f14172019-06-21 09:42:05 +01004661 uclamp_fork(p);
4662
Linus Torvalds1da177e2005-04-16 15:20:36 -07004663 /*
4664 * Revert to default priority/policy on fork if requested.
4665 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004666 if (unlikely(p->sched_reset_on_fork)) {
Dario Faggioliaab03e02013-11-28 11:14:43 +01004667 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02004668 p->policy = SCHED_NORMAL;
4669 p->static_prio = NICE_TO_PRIO(0);
4670 p->rt_priority = 0;
4671 } else if (PRIO_TO_NICE(p->static_prio) < 0)
4672 p->static_prio = NICE_TO_PRIO(0);
4673
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02004674 p->prio = p->normal_prio = p->static_prio;
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004675 set_load_weight(p, false);
Ingo Molnardd41f592007-07-09 18:51:59 +02004676
4677 /*
4678 * We don't need the reset flag anymore after the fork. It has
4679 * fulfilled its duty:
Ingo Molnar02e4bac22007-10-15 17:00:11 +02004680 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07004681 p->sched_reset_on_fork = 0;
4682 }
4683
Sebastian Andrzej Siewioraf0fffd2018-07-06 15:06:15 +02004684 if (dl_prio(p->prio))
Dario Faggioliaab03e02013-11-28 11:14:43 +01004685 return -EAGAIN;
Sebastian Andrzej Siewioraf0fffd2018-07-06 15:06:15 +02004686 else if (rt_prio(p->prio))
Dario Faggioliaab03e02013-11-28 11:14:43 +01004687 p->sched_class = &rt_sched_class;
Sebastian Andrzej Siewioraf0fffd2018-07-06 15:06:15 +02004688 else
Ingo Molnarb29739f2006-06-27 02:54:51 -07004689 p->sched_class = &fair_sched_class;
4690
Peter Zijlstra7dc603c2016-06-16 13:29:28 +02004691 init_entity_runnable_average(&p->se);
Peter Zijlstracd29fe62009-11-27 17:32:46 +01004692
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004693
Naveen N. Raof6db8342015-06-25 23:53:37 +05304694#ifdef CONFIG_SCHED_INFO
Ingo Molnardd41f592007-07-09 18:51:59 +02004695 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07004696 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004697#endif
Peter Zijlstra3ca7a442011-04-05 17:23:40 +02004698#if defined(CONFIG_SMP)
4699 p->on_cpu = 0;
Nick Piggin4866cde2005-06-25 14:57:23 -07004700#endif
Peter Zijlstra01028742013-08-14 14:55:46 +02004701 init_task_preempt_count(p);
Dario Faggioli806c09a2010-11-30 19:51:33 +01004702#ifdef CONFIG_SMP
Gregory Haskins917b6272008-12-29 09:39:53 -05004703 plist_node_init(&p->pushable_tasks, MAX_PRIO);
Juri Lelli1baca4c2013-11-07 14:43:38 +01004704 RB_CLEAR_NODE(&p->pushable_dl_tasks);
Dario Faggioli806c09a2010-11-30 19:51:33 +01004705#endif
Dario Faggioliaab03e02013-11-28 11:14:43 +01004706 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004707}
4708
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004709void sched_cgroup_fork(struct task_struct *p, struct kernel_clone_args *kargs)
Qais Yousef13685c42020-07-16 12:03:45 +01004710{
Zhang Qiao4ef0c5c2021-09-15 14:40:30 +08004711 unsigned long flags;
Zhang Qiao4ef0c5c2021-09-15 14:40:30 +08004712
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004713 /*
4714 * Because we're not yet on the pid-hash, p->pi_lock isn't strictly
4715 * required yet, but lockdep gets upset if rules are violated.
4716 */
Zhang Qiao4ef0c5c2021-09-15 14:40:30 +08004717 raw_spin_lock_irqsave(&p->pi_lock, flags);
4718#ifdef CONFIG_CGROUP_SCHED
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004719 if (1) {
4720 struct task_group *tg;
4721 tg = container_of(kargs->cset->subsys[cpu_cgrp_id],
4722 struct task_group, css);
4723 tg = autogroup_task_group(p, tg);
4724 p->sched_task_group = tg;
4725 }
Zhang Qiao4ef0c5c2021-09-15 14:40:30 +08004726#endif
4727 rseq_migrate(p);
4728 /*
4729 * We're setting the CPU for the first time, we don't migrate,
4730 * so use __set_task_cpu().
4731 */
4732 __set_task_cpu(p, smp_processor_id());
4733 if (p->sched_class->task_fork)
4734 p->sched_class->task_fork(p);
4735 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004736}
Zhang Qiao4ef0c5c2021-09-15 14:40:30 +08004737
Peter Zijlstrab1e82062022-02-14 10:16:57 +01004738void sched_post_fork(struct task_struct *p)
4739{
Qais Yousef13685c42020-07-16 12:03:45 +01004740 uclamp_post_fork(p);
4741}
4742
Dario Faggioli332ac172013-11-07 14:43:45 +01004743unsigned long to_ratio(u64 period, u64 runtime)
4744{
4745 if (runtime == RUNTIME_INF)
Luca Abenic52f14d2017-05-18 22:13:31 +02004746 return BW_UNIT;
Dario Faggioli332ac172013-11-07 14:43:45 +01004747
4748 /*
4749 * Doing this here saves a lot of checks in all
4750 * the calling paths, and returning zero seems
4751 * safe for them anyway.
4752 */
4753 if (period == 0)
4754 return 0;
4755
Luca Abenic52f14d2017-05-18 22:13:31 +02004756 return div64_u64(runtime << BW_SHIFT, period);
Dario Faggioli332ac172013-11-07 14:43:45 +01004757}
4758
Linus Torvalds1da177e2005-04-16 15:20:36 -07004759/*
4760 * wake_up_new_task - wake up a newly created task for the first time.
4761 *
4762 * This function will do some initial scheduler statistics housekeeping
4763 * that must be done for every newly created context, then puts the task
4764 * on the runqueue and wakes it.
4765 */
Samir Bellabes3e51e3e2011-05-11 18:18:05 +02004766void wake_up_new_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004767{
Peter Zijlstraeb580752015-07-31 21:28:18 +02004768 struct rq_flags rf;
Ingo Molnardd41f592007-07-09 18:51:59 +02004769 struct rq *rq;
Peter Zijlstrafabf3182010-01-21 21:04:57 +01004770
Peter Zijlstraeb580752015-07-31 21:28:18 +02004771 raw_spin_lock_irqsave(&p->pi_lock, rf.flags);
Peter Zijlstra2f064a52021-06-11 10:28:17 +02004772 WRITE_ONCE(p->__state, TASK_RUNNING);
Peter Zijlstrafabf3182010-01-21 21:04:57 +01004773#ifdef CONFIG_SMP
4774 /*
4775 * Fork balancing, do it here and not earlier because:
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02004776 * - cpus_ptr can change in the fork path
Ingo Molnard1ccc662017-02-01 11:46:42 +01004777 * - any previously selected CPU might disappear through hotplug
Peter Zijlstrae210bff2016-06-16 18:51:48 +02004778 *
4779 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
4780 * as we're not fully set-up yet.
Peter Zijlstrafabf3182010-01-21 21:04:57 +01004781 */
Mel Gorman32e839d2018-01-30 10:45:55 +00004782 p->recent_used_cpu = task_cpu(p);
Mathieu Desnoyersce3614d2020-07-06 16:49:10 -04004783 rseq_migrate(p);
Valentin Schneider3aef15512020-11-02 18:45:13 +00004784 __set_task_cpu(p, select_task_rq(p, task_cpu(p), WF_FORK));
Peter Zijlstrafabf3182010-01-21 21:04:57 +01004785#endif
Peter Zijlstrab7fa30c2016-06-09 15:07:50 +02004786 rq = __task_rq_lock(p, &rf);
Peter Zijlstra4126bad2016-10-03 16:20:59 +02004787 update_rq_clock(rq);
Dietmar Eggemannd0fe0b92019-01-22 16:25:01 +00004788 post_init_entity_util_avg(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004789
Peter Zijlstra7a57f322017-02-21 14:47:02 +01004790 activate_task(rq, p, ENQUEUE_NOCLOCK);
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02004791 trace_sched_wakeup_new(p);
Peter Zijlstraa7558e02009-09-14 20:02:34 +02004792 check_preempt_curr(rq, p, WF_FORK);
Steven Rostedt9a897c52008-01-25 21:08:22 +01004793#ifdef CONFIG_SMP
Peter Zijlstra0aaafaa2015-10-23 11:50:08 +02004794 if (p->sched_class->task_woken) {
4795 /*
Tal Zussmanb19a8882020-11-12 19:51:56 -05004796 * Nothing relies on rq->lock after this, so it's fine to
Peter Zijlstra0aaafaa2015-10-23 11:50:08 +02004797 * drop it.
4798 */
Matt Flemingd8ac8972016-09-21 14:38:10 +01004799 rq_unpin_lock(rq, &rf);
Peter Zijlstraefbbd052009-12-16 18:04:40 +01004800 p->sched_class->task_woken(rq, p);
Matt Flemingd8ac8972016-09-21 14:38:10 +01004801 rq_repin_lock(rq, &rf);
Peter Zijlstra0aaafaa2015-10-23 11:50:08 +02004802 }
Steven Rostedt9a897c52008-01-25 21:08:22 +01004803#endif
Peter Zijlstraeb580752015-07-31 21:28:18 +02004804 task_rq_unlock(rq, p, &rf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004805}
4806
Avi Kivitye107be32007-07-26 13:40:43 +02004807#ifdef CONFIG_PREEMPT_NOTIFIERS
4808
Davidlohr Buesob7203422018-03-26 14:09:26 -07004809static DEFINE_STATIC_KEY_FALSE(preempt_notifier_key);
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004810
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02004811void preempt_notifier_inc(void)
4812{
Davidlohr Buesob7203422018-03-26 14:09:26 -07004813 static_branch_inc(&preempt_notifier_key);
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02004814}
4815EXPORT_SYMBOL_GPL(preempt_notifier_inc);
4816
4817void preempt_notifier_dec(void)
4818{
Davidlohr Buesob7203422018-03-26 14:09:26 -07004819 static_branch_dec(&preempt_notifier_key);
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02004820}
4821EXPORT_SYMBOL_GPL(preempt_notifier_dec);
4822
Avi Kivitye107be32007-07-26 13:40:43 +02004823/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00004824 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07004825 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02004826 */
4827void preempt_notifier_register(struct preempt_notifier *notifier)
4828{
Davidlohr Buesob7203422018-03-26 14:09:26 -07004829 if (!static_branch_unlikely(&preempt_notifier_key))
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02004830 WARN(1, "registering preempt_notifier while notifiers disabled\n");
4831
Avi Kivitye107be32007-07-26 13:40:43 +02004832 hlist_add_head(&notifier->link, &current->preempt_notifiers);
4833}
4834EXPORT_SYMBOL_GPL(preempt_notifier_register);
4835
4836/**
4837 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07004838 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02004839 *
Mathieu Desnoyersd84525a2015-05-17 12:53:10 -04004840 * This is *not* safe to call from within a preemption notifier.
Avi Kivitye107be32007-07-26 13:40:43 +02004841 */
4842void preempt_notifier_unregister(struct preempt_notifier *notifier)
4843{
4844 hlist_del(&notifier->link);
4845}
4846EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
4847
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004848static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
Avi Kivitye107be32007-07-26 13:40:43 +02004849{
4850 struct preempt_notifier *notifier;
Avi Kivitye107be32007-07-26 13:40:43 +02004851
Sasha Levinb67bfe02013-02-27 17:06:00 -08004852 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
Avi Kivitye107be32007-07-26 13:40:43 +02004853 notifier->ops->sched_in(notifier, raw_smp_processor_id());
4854}
4855
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004856static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
4857{
Davidlohr Buesob7203422018-03-26 14:09:26 -07004858 if (static_branch_unlikely(&preempt_notifier_key))
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004859 __fire_sched_in_preempt_notifiers(curr);
4860}
4861
Avi Kivitye107be32007-07-26 13:40:43 +02004862static void
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004863__fire_sched_out_preempt_notifiers(struct task_struct *curr,
4864 struct task_struct *next)
Avi Kivitye107be32007-07-26 13:40:43 +02004865{
4866 struct preempt_notifier *notifier;
Avi Kivitye107be32007-07-26 13:40:43 +02004867
Sasha Levinb67bfe02013-02-27 17:06:00 -08004868 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
Avi Kivitye107be32007-07-26 13:40:43 +02004869 notifier->ops->sched_out(notifier, next);
4870}
4871
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004872static __always_inline void
4873fire_sched_out_preempt_notifiers(struct task_struct *curr,
4874 struct task_struct *next)
4875{
Davidlohr Buesob7203422018-03-26 14:09:26 -07004876 if (static_branch_unlikely(&preempt_notifier_key))
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004877 __fire_sched_out_preempt_notifiers(curr, next);
4878}
4879
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02004880#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02004881
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004882static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
Avi Kivitye107be32007-07-26 13:40:43 +02004883{
4884}
4885
Peter Zijlstra1cde2932015-06-08 16:00:30 +02004886static inline void
Avi Kivitye107be32007-07-26 13:40:43 +02004887fire_sched_out_preempt_notifiers(struct task_struct *curr,
4888 struct task_struct *next)
4889{
4890}
4891
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02004892#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02004893
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02004894static inline void prepare_task(struct task_struct *next)
4895{
4896#ifdef CONFIG_SMP
4897 /*
4898 * Claim the task as running, we do this before switching to it
4899 * such that any running task will have this set.
Peter Zijlstra58877d32020-07-02 14:52:11 +02004900 *
Tianchen Dingf3dd3f62022-06-09 07:34:12 +08004901 * See the smp_load_acquire(&p->on_cpu) case in ttwu() and
4902 * its ordering comment.
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02004903 */
Peter Zijlstra58877d32020-07-02 14:52:11 +02004904 WRITE_ONCE(next->on_cpu, 1);
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02004905#endif
4906}
4907
4908static inline void finish_task(struct task_struct *prev)
4909{
4910#ifdef CONFIG_SMP
4911 /*
Peter Zijlstra58877d32020-07-02 14:52:11 +02004912 * This must be the very last reference to @prev from this CPU. After
4913 * p->on_cpu is cleared, the task can be moved to a different CPU. We
4914 * must ensure this doesn't happen until the switch is completely
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02004915 * finished.
4916 *
4917 * In particular, the load of prev->state in finish_task_switch() must
4918 * happen before this.
4919 *
4920 * Pairs with the smp_cond_load_acquire() in try_to_wake_up().
4921 */
4922 smp_store_release(&prev->on_cpu, 0);
4923#endif
4924}
4925
Peter Zijlstra565790d22020-05-11 14:13:00 +02004926#ifdef CONFIG_SMP
4927
Kees Cook8e5bad72022-10-07 17:07:58 -07004928static void do_balance_callbacks(struct rq *rq, struct balance_callback *head)
Peter Zijlstra565790d22020-05-11 14:13:00 +02004929{
4930 void (*func)(struct rq *rq);
Kees Cook8e5bad72022-10-07 17:07:58 -07004931 struct balance_callback *next;
Peter Zijlstra565790d22020-05-11 14:13:00 +02004932
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05004933 lockdep_assert_rq_held(rq);
Peter Zijlstra565790d22020-05-11 14:13:00 +02004934
4935 while (head) {
4936 func = (void (*)(struct rq *))head->func;
4937 next = head->next;
4938 head->next = NULL;
4939 head = next;
4940
4941 func(rq);
4942 }
4943}
4944
Peter Zijlstraae792702020-12-10 17:14:08 +01004945static void balance_push(struct rq *rq);
4946
Peter Zijlstra04193d52022-06-07 22:41:55 +02004947/*
4948 * balance_push_callback is a right abuse of the callback interface and plays
4949 * by significantly different rules.
4950 *
4951 * Where the normal balance_callback's purpose is to be ran in the same context
4952 * that queued it (only later, when it's safe to drop rq->lock again),
4953 * balance_push_callback is specifically targeted at __schedule().
4954 *
4955 * This abuse is tolerated because it places all the unlikely/odd cases behind
4956 * a single test, namely: rq->balance_callback == NULL.
4957 */
Kees Cook8e5bad72022-10-07 17:07:58 -07004958struct balance_callback balance_push_callback = {
Peter Zijlstraae792702020-12-10 17:14:08 +01004959 .next = NULL,
Kees Cook8e5bad72022-10-07 17:07:58 -07004960 .func = balance_push,
Peter Zijlstraae792702020-12-10 17:14:08 +01004961};
4962
Kees Cook8e5bad72022-10-07 17:07:58 -07004963static inline struct balance_callback *
Peter Zijlstra04193d52022-06-07 22:41:55 +02004964__splice_balance_callbacks(struct rq *rq, bool split)
Peter Zijlstra565790d22020-05-11 14:13:00 +02004965{
Kees Cook8e5bad72022-10-07 17:07:58 -07004966 struct balance_callback *head = rq->balance_callback;
Peter Zijlstra565790d22020-05-11 14:13:00 +02004967
Peter Zijlstra04193d52022-06-07 22:41:55 +02004968 if (likely(!head))
4969 return NULL;
4970
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05004971 lockdep_assert_rq_held(rq);
Peter Zijlstra04193d52022-06-07 22:41:55 +02004972 /*
4973 * Must not take balance_push_callback off the list when
4974 * splice_balance_callbacks() and balance_callbacks() are not
4975 * in the same rq->lock section.
4976 *
4977 * In that case it would be possible for __schedule() to interleave
4978 * and observe the list empty.
4979 */
4980 if (split && head == &balance_push_callback)
4981 head = NULL;
4982 else
Peter Zijlstra565790d22020-05-11 14:13:00 +02004983 rq->balance_callback = NULL;
4984
4985 return head;
4986}
4987
Kees Cook8e5bad72022-10-07 17:07:58 -07004988static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
Peter Zijlstra04193d52022-06-07 22:41:55 +02004989{
4990 return __splice_balance_callbacks(rq, true);
4991}
4992
Peter Zijlstra565790d22020-05-11 14:13:00 +02004993static void __balance_callbacks(struct rq *rq)
4994{
Peter Zijlstra04193d52022-06-07 22:41:55 +02004995 do_balance_callbacks(rq, __splice_balance_callbacks(rq, false));
Peter Zijlstra565790d22020-05-11 14:13:00 +02004996}
4997
Kees Cook8e5bad72022-10-07 17:07:58 -07004998static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
Peter Zijlstra565790d22020-05-11 14:13:00 +02004999{
5000 unsigned long flags;
5001
5002 if (unlikely(head)) {
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05005003 raw_spin_rq_lock_irqsave(rq, flags);
Peter Zijlstra565790d22020-05-11 14:13:00 +02005004 do_balance_callbacks(rq, head);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05005005 raw_spin_rq_unlock_irqrestore(rq, flags);
Peter Zijlstra565790d22020-05-11 14:13:00 +02005006 }
5007}
5008
5009#else
5010
5011static inline void __balance_callbacks(struct rq *rq)
5012{
5013}
5014
Kees Cook8e5bad72022-10-07 17:07:58 -07005015static inline struct balance_callback *splice_balance_callbacks(struct rq *rq)
Peter Zijlstra565790d22020-05-11 14:13:00 +02005016{
5017 return NULL;
5018}
5019
Kees Cook8e5bad72022-10-07 17:07:58 -07005020static inline void balance_callbacks(struct rq *rq, struct balance_callback *head)
Peter Zijlstra565790d22020-05-11 14:13:00 +02005021{
5022}
5023
5024#endif
5025
Peter Zijlstra269d5992018-02-06 17:52:13 +01005026static inline void
5027prepare_lock_switch(struct rq *rq, struct task_struct *next, struct rq_flags *rf)
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005028{
Peter Zijlstra269d5992018-02-06 17:52:13 +01005029 /*
5030 * Since the runqueue lock will be released by the next
5031 * task (which is an invalid locking op but in the case
5032 * of the scheduler it's an obvious special-case), so we
5033 * do an early lockdep release here:
5034 */
5035 rq_unpin_lock(rq, rf);
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +01005036 spin_release(&__rq_lockp(rq)->dep_map, _THIS_IP_);
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005037#ifdef CONFIG_DEBUG_SPINLOCK
5038 /* this is a valid case when another task releases the spinlock */
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05005039 rq_lockp(rq)->owner = next;
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005040#endif
Peter Zijlstra269d5992018-02-06 17:52:13 +01005041}
5042
5043static inline void finish_lock_switch(struct rq *rq)
5044{
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005045 /*
5046 * If we are tracking spinlock dependencies then we have to
5047 * fix up the runqueue lock - which gets 'carried over' from
5048 * prev into current:
5049 */
Peter Zijlstra9ef7e7e2021-03-03 16:45:41 +01005050 spin_acquire(&__rq_lockp(rq)->dep_map, 0, 0, _THIS_IP_);
Peter Zijlstraae792702020-12-10 17:14:08 +01005051 __balance_callbacks(rq);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05005052 raw_spin_rq_unlock_irq(rq);
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005053}
5054
Ingo Molnar325ea102018-03-03 12:20:47 +01005055/*
5056 * NOP if the arch has not defined these:
5057 */
5058
5059#ifndef prepare_arch_switch
5060# define prepare_arch_switch(next) do { } while (0)
5061#endif
5062
5063#ifndef finish_arch_post_lock_switch
5064# define finish_arch_post_lock_switch() do { } while (0)
5065#endif
5066
Thomas Gleixner5fbda3e2020-11-18 20:48:43 +01005067static inline void kmap_local_sched_out(void)
5068{
5069#ifdef CONFIG_KMAP_LOCAL
5070 if (unlikely(current->kmap_ctrl.idx))
5071 __kmap_local_sched_out();
5072#endif
5073}
5074
5075static inline void kmap_local_sched_in(void)
5076{
5077#ifdef CONFIG_KMAP_LOCAL
5078 if (unlikely(current->kmap_ctrl.idx))
5079 __kmap_local_sched_in();
5080#endif
5081}
5082
Linus Torvalds1da177e2005-04-16 15:20:36 -07005083/**
Nick Piggin4866cde2005-06-25 14:57:23 -07005084 * prepare_task_switch - prepare to switch tasks
5085 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07005086 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07005087 * @next: the task we are going to switch to.
5088 *
5089 * This is called with the rq lock held and interrupts off. It must
5090 * be paired with a subsequent finish_task_switch after the context
5091 * switch.
5092 *
5093 * prepare_task_switch sets up locking and calls architecture specific
5094 * hooks.
5095 */
Avi Kivitye107be32007-07-26 13:40:43 +02005096static inline void
5097prepare_task_switch(struct rq *rq, struct task_struct *prev,
5098 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07005099{
Mark Rutland0ed557a2018-06-14 15:27:41 -07005100 kcov_prepare_switch(prev);
Michael S. Tsirkin43148952013-09-22 17:20:54 +03005101 sched_info_switch(rq, prev, next);
Peter Zijlstrafe4b04f2011-02-02 13:19:09 +01005102 perf_event_task_sched_out(prev, next);
Mathieu Desnoyersd7822b12018-06-02 08:43:54 -04005103 rseq_preempt(prev);
Avi Kivitye107be32007-07-26 13:40:43 +02005104 fire_sched_out_preempt_notifiers(prev, next);
Thomas Gleixner5fbda3e2020-11-18 20:48:43 +01005105 kmap_local_sched_out();
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005106 prepare_task(next);
Nick Piggin4866cde2005-06-25 14:57:23 -07005107 prepare_arch_switch(next);
5108}
5109
5110/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07005111 * finish_task_switch - clean up after a task-switch
5112 * @prev: the thread we just switched away from.
5113 *
Nick Piggin4866cde2005-06-25 14:57:23 -07005114 * finish_task_switch must be called after the context switch, paired
5115 * with a prepare_task_switch call before the context switch.
5116 * finish_task_switch will reconcile locking set up by prepare_task_switch,
5117 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005118 *
5119 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005120 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121 * with the lock held can cause deadlocks; see schedule() for
5122 * details.)
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005123 *
5124 * The context switch have flipped the stack from under us and restored the
5125 * local variables which were saved when this task called schedule() in the
5126 * past. prev == current is still correct but we need to recalculate this_rq
5127 * because prev may have moved to another CPU.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005128 */
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005129static struct rq *finish_task_switch(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005130 __releases(rq->lock)
5131{
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005132 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005133 struct mm_struct *mm = rq->prev_mm;
Valentin Schneiderfa2c3252022-01-20 16:25:19 +00005134 unsigned int prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005135
Peter Zijlstra609ca062015-09-28 17:52:18 +02005136 /*
5137 * The previous task will have left us with a preempt_count of 2
5138 * because it left us after:
5139 *
5140 * schedule()
5141 * preempt_disable(); // 1
5142 * __schedule()
5143 * raw_spin_lock_irq(&rq->lock) // 2
5144 *
5145 * Also, see FORK_PREEMPT_COUNT.
5146 */
Peter Zijlstrae2bf1c4b2015-09-29 12:18:46 +02005147 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
5148 "corrupted preempt_count: %s/%d/0x%x\n",
5149 current->comm, current->pid, preempt_count()))
5150 preempt_count_set(FORK_PREEMPT_COUNT);
Peter Zijlstra609ca062015-09-28 17:52:18 +02005151
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 rq->prev_mm = NULL;
5153
5154 /*
5155 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07005156 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07005157 * schedule one last time. The schedule call will never return, and
5158 * the scheduled task must drop that reference.
Peter Zijlstra95913d92015-09-29 14:45:09 +02005159 *
5160 * We must observe prev->state before clearing prev->on_cpu (in
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005161 * finish_task), otherwise a concurrent wakeup can get prev
Peter Zijlstra95913d92015-09-29 14:45:09 +02005162 * running on another CPU and we could rave with its RUNNING -> DEAD
5163 * transition, resulting in a double drop.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005164 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02005165 prev_state = READ_ONCE(prev->__state);
Frederic Weisbeckerbf9fae92012-09-08 15:23:11 +02005166 vtime_task_switch(prev);
Stephane Eraniana8d757e2011-08-25 15:58:03 +02005167 perf_event_task_sched_in(prev, current);
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005168 finish_task(prev);
Peter Zijlstra0fdcccf2021-05-13 01:29:23 +02005169 tick_nohz_task_switch();
rodrigosiqueira31cb1bc2017-12-15 12:06:03 -02005170 finish_lock_switch(rq);
Catalin Marinas01f23e12011-11-27 21:43:10 +00005171 finish_arch_post_lock_switch();
Mark Rutland0ed557a2018-06-14 15:27:41 -07005172 kcov_finish_switch(current);
Thomas Gleixner5fbda3e2020-11-18 20:48:43 +01005173 /*
5174 * kmap_local_sched_out() is invoked with rq::lock held and
5175 * interrupts disabled. There is no requirement for that, but the
5176 * sched out code does not have an interrupt enabled section.
5177 * Restoring the maps on sched in does not require interrupts being
5178 * disabled either.
5179 */
5180 kmap_local_sched_in();
Steven Rostedte8fa1362008-01-25 21:08:05 +01005181
Avi Kivitye107be32007-07-26 13:40:43 +02005182 fire_sched_in_preempt_notifiers(current);
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05005183 /*
Mathieu Desnoyers70216e12018-01-29 15:20:17 -05005184 * When switching through a kernel thread, the loop in
5185 * membarrier_{private,global}_expedited() may have observed that
5186 * kernel thread and not issued an IPI. It is therefore possible to
5187 * schedule between user->kernel->user threads without passing though
5188 * switch_mm(). Membarrier requires a barrier after storing to
5189 * rq->curr, before returning to userspace, so provide them here:
5190 *
5191 * - a full memory barrier for {PRIVATE,GLOBAL}_EXPEDITED, implicitly
5192 * provided by mmdrop(),
5193 * - a sync_core for SYNC_CORE.
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05005194 */
Mathieu Desnoyers70216e12018-01-29 15:20:17 -05005195 if (mm) {
5196 membarrier_mm_sync_core_before_usermode(mm);
Thomas Gleixner8d491de2021-09-28 14:24:32 +02005197 mmdrop_sched(mm);
Mathieu Desnoyers70216e12018-01-29 15:20:17 -05005198 }
Peter Zijlstra1cef1152018-06-07 11:45:49 +02005199 if (unlikely(prev_state == TASK_DEAD)) {
5200 if (prev->sched_class->task_dead)
5201 prev->sched_class->task_dead(prev);
Dario Faggiolie6c390f2013-11-07 14:43:35 +01005202
Peter Zijlstra1cef1152018-06-07 11:45:49 +02005203 /* Task is done with its stack. */
5204 put_task_stack(prev);
Andy Lutomirski68f24b082016-09-15 22:45:48 -07005205
Eric W. Biederman0ff7b2cf2019-09-14 07:33:58 -05005206 put_task_struct_rcu_user(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08005207 }
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +02005208
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005209 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005210}
5211
5212/**
5213 * schedule_tail - first thing a freshly forked thread must call.
5214 * @prev: the thread we just switched away from.
5215 */
Andi Kleen722a9f92014-05-02 00:44:38 +02005216asmlinkage __visible void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005217 __releases(rq->lock)
5218{
Peter Zijlstra609ca062015-09-28 17:52:18 +02005219 /*
5220 * New tasks start with FORK_PREEMPT_COUNT, see there and
5221 * finish_task_switch() for details.
5222 *
5223 * finish_task_switch() will drop rq->lock() and lower preempt_count
5224 * and the preempt_enable() will end up enabling preemption (on
5225 * PREEMPT_COUNT kernels).
5226 */
5227
Edmundo Carmona Antoranz13c22352021-03-06 15:07:39 -06005228 finish_task_switch(prev);
Oleg Nesterov1a43a142014-10-08 21:36:44 +02005229 preempt_enable();
Steven Rostedtda19ab52009-07-29 00:21:22 -04005230
Linus Torvalds1da177e2005-04-16 15:20:36 -07005231 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07005232 put_user(task_pid_vnr(current), current->set_child_tid);
Eric W. Biederman088fe472018-07-23 17:26:49 -05005233
5234 calculate_sigpending();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005235}
5236
5237/*
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005238 * context_switch - switch to the new MM and the new thread's register state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005239 */
Josh Poimboeuf04936942016-02-28 22:22:39 -06005240static __always_inline struct rq *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005241context_switch(struct rq *rq, struct task_struct *prev,
Matt Flemingd8ac8972016-09-21 14:38:10 +01005242 struct task_struct *next, struct rq_flags *rf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005243{
Avi Kivitye107be32007-07-26 13:40:43 +02005244 prepare_task_switch(rq, prev, next);
Peter Zijlstrafe4b04f2011-02-02 13:19:09 +01005245
Zachary Amsden9226d122007-02-13 13:26:21 +01005246 /*
5247 * For paravirt, this is coupled with an exit in switch_to to
5248 * combine the page table reload and the switch backend into
5249 * one hypercall.
5250 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08005251 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01005252
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05005253 /*
Peter Zijlstra139d0252019-07-29 16:05:15 +02005254 * kernel -> kernel lazy + transfer active
5255 * user -> kernel lazy + mmgrab() active
5256 *
5257 * kernel -> user switch + mmdrop() active
5258 * user -> user switch
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05005259 */
Peter Zijlstra139d0252019-07-29 16:05:15 +02005260 if (!next->mm) { // to kernel
5261 enter_lazy_tlb(prev->active_mm, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005262
Peter Zijlstra139d0252019-07-29 16:05:15 +02005263 next->active_mm = prev->active_mm;
5264 if (prev->mm) // from user
5265 mmgrab(prev->active_mm);
5266 else
5267 prev->active_mm = NULL;
5268 } else { // to user
Mathieu Desnoyers227a4aa2019-09-19 13:37:02 -04005269 membarrier_switch_mm(rq, prev->active_mm, next->mm);
Peter Zijlstra139d0252019-07-29 16:05:15 +02005270 /*
5271 * sys_membarrier() requires an smp_mb() between setting
Mathieu Desnoyers227a4aa2019-09-19 13:37:02 -04005272 * rq->curr / membarrier_switch_mm() and returning to userspace.
Peter Zijlstra139d0252019-07-29 16:05:15 +02005273 *
5274 * The below provides this either through switch_mm(), or in
5275 * case 'prev->active_mm == next->mm' through
5276 * finish_task_switch()'s mmdrop().
5277 */
Peter Zijlstra139d0252019-07-29 16:05:15 +02005278 switch_mm_irqs_off(prev->active_mm, next->mm, next);
Yu Zhaobd74fda2022-09-18 02:00:05 -06005279 lru_gen_use_mm(next->mm);
Peter Zijlstra139d0252019-07-29 16:05:15 +02005280
5281 if (!prev->mm) { // from kernel
5282 /* will mmdrop() in finish_task_switch(). */
5283 rq->prev_mm = prev->active_mm;
5284 prev->active_mm = NULL;
5285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005286 }
Matt Fleming92509b72016-09-21 14:38:11 +01005287
Matt Flemingcb42c9a2016-09-21 14:38:13 +01005288 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
Matt Fleming92509b72016-09-21 14:38:11 +01005289
Peter Zijlstra269d5992018-02-06 17:52:13 +01005290 prepare_lock_switch(rq, next, rf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005291
5292 /* Here we just switch the register state and the stack. */
5293 switch_to(prev, next, prev);
Ingo Molnardd41f592007-07-09 18:51:59 +02005294 barrier();
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02005295
5296 return finish_task_switch(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297}
5298
5299/*
Sha Zhengju1c3e8262013-02-20 17:14:38 +08005300 * nr_running and nr_context_switches:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301 *
5302 * externally visible scheduler statistics: current number of runnable
Sha Zhengju1c3e8262013-02-20 17:14:38 +08005303 * threads, total number of context switches performed since bootup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005304 */
Alexey Dobriyan01aee8f2021-04-22 23:02:25 +03005305unsigned int nr_running(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005306{
Alexey Dobriyan01aee8f2021-04-22 23:02:25 +03005307 unsigned int i, sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005308
5309 for_each_online_cpu(i)
5310 sum += cpu_rq(i)->nr_running;
5311
5312 return sum;
5313}
5314
Tim Chen2ee507c2014-07-31 10:29:48 -07005315/*
Ingo Molnard1ccc662017-02-01 11:46:42 +01005316 * Check if only the current task is running on the CPU.
Dominik Dingel00cc16332015-09-18 11:27:45 +02005317 *
5318 * Caution: this function does not check that the caller has disabled
5319 * preemption, thus the result might have a time-of-check-to-time-of-use
5320 * race. The caller is responsible to use it correctly, for example:
5321 *
Ingo Molnardfcb2452018-12-03 10:05:56 +01005322 * - from a non-preemptible section (of course)
Dominik Dingel00cc16332015-09-18 11:27:45 +02005323 *
5324 * - from a thread that is bound to a single CPU
5325 *
5326 * - in a loop with very short iterations (e.g. a polling loop)
Tim Chen2ee507c2014-07-31 10:29:48 -07005327 */
5328bool single_task_running(void)
5329{
Dominik Dingel00cc16332015-09-18 11:27:45 +02005330 return raw_rq()->nr_running == 1;
Tim Chen2ee507c2014-07-31 10:29:48 -07005331}
5332EXPORT_SYMBOL(single_task_running);
5333
Linus Torvalds1da177e2005-04-16 15:20:36 -07005334unsigned long long nr_context_switches(void)
5335{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07005336 int i;
5337 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08005339 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005340 sum += cpu_rq(i)->nr_switches;
5341
5342 return sum;
5343}
5344
Tejun Heoe33a9bb2016-12-07 15:48:41 -05005345/*
Daniel Lezcano145d9522018-10-04 14:04:02 +02005346 * Consumers of these two interfaces, like for example the cpuidle menu
5347 * governor, are using nonsensical data. Preferring shallow idle state selection
5348 * for a CPU that has IO-wait which might not even end up running the task when
5349 * it does become runnable.
5350 */
5351
Alexey Dobriyan8fc28582021-04-22 23:02:27 +03005352unsigned int nr_iowait_cpu(int cpu)
Daniel Lezcano145d9522018-10-04 14:04:02 +02005353{
5354 return atomic_read(&cpu_rq(cpu)->nr_iowait);
5355}
5356
Daniel Lezcano145d9522018-10-04 14:04:02 +02005357/*
Tal Zussmanb19a8882020-11-12 19:51:56 -05005358 * IO-wait accounting, and how it's mostly bollocks (on SMP).
Tejun Heoe33a9bb2016-12-07 15:48:41 -05005359 *
5360 * The idea behind IO-wait account is to account the idle time that we could
5361 * have spend running if it were not for IO. That is, if we were to improve the
5362 * storage performance, we'd have a proportional reduction in IO-wait time.
5363 *
5364 * This all works nicely on UP, where, when a task blocks on IO, we account
5365 * idle time as IO-wait, because if the storage were faster, it could've been
5366 * running and we'd not be idle.
5367 *
5368 * This has been extended to SMP, by doing the same for each CPU. This however
5369 * is broken.
5370 *
5371 * Imagine for instance the case where two tasks block on one CPU, only the one
5372 * CPU will have IO-wait accounted, while the other has regular idle. Even
5373 * though, if the storage were faster, both could've ran at the same time,
5374 * utilising both CPUs.
5375 *
5376 * This means, that when looking globally, the current IO-wait accounting on
5377 * SMP is a lower bound, by reason of under accounting.
5378 *
5379 * Worse, since the numbers are provided per CPU, they are sometimes
5380 * interpreted per CPU, and that is nonsensical. A blocked task isn't strictly
5381 * associated with any one particular CPU, it can wake to another CPU than it
5382 * blocked on. This means the per CPU IO-wait number is meaningless.
5383 *
5384 * Task CPU affinities can make all that even more 'interesting'.
5385 */
5386
Alexey Dobriyan97455162021-04-22 23:02:26 +03005387unsigned int nr_iowait(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388{
Alexey Dobriyan97455162021-04-22 23:02:26 +03005389 unsigned int i, sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005390
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08005391 for_each_possible_cpu(i)
Daniel Lezcano145d9522018-10-04 14:04:02 +02005392 sum += nr_iowait_cpu(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005393
5394 return sum;
5395}
5396
Ingo Molnardd41f592007-07-09 18:51:59 +02005397#ifdef CONFIG_SMP
5398
Ingo Molnar48f24c42006-07-03 00:25:40 -07005399/*
Peter Zijlstra38022902009-12-16 18:04:37 +01005400 * sched_exec - execve() is a valuable balancing opportunity, because at
5401 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005402 */
Peter Zijlstra38022902009-12-16 18:04:37 +01005403void sched_exec(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404{
Peter Zijlstra38022902009-12-16 18:04:37 +01005405 struct task_struct *p = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005406 unsigned long flags;
Peter Zijlstra0017d732010-03-24 18:34:10 +01005407 int dest_cpu;
Peter Zijlstra38022902009-12-16 18:04:37 +01005408
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02005409 raw_spin_lock_irqsave(&p->pi_lock, flags);
Valentin Schneider3aef15512020-11-02 18:45:13 +00005410 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
Peter Zijlstra0017d732010-03-24 18:34:10 +01005411 if (dest_cpu == smp_processor_id())
5412 goto unlock;
Peter Zijlstra38022902009-12-16 18:04:37 +01005413
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02005414 if (likely(cpu_active(dest_cpu))) {
Tejun Heo969c7922010-05-06 18:49:21 +02005415 struct migration_arg arg = { p, dest_cpu };
Ingo Molnar36c8b582006-07-03 00:25:41 -07005416
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02005417 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5418 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005419 return;
5420 }
Peter Zijlstra0017d732010-03-24 18:34:10 +01005421unlock:
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02005422 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005423}
5424
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425#endif
5426
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427DEFINE_PER_CPU(struct kernel_stat, kstat);
Glauber Costa3292beb2011-11-28 14:45:17 -02005428DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005429
5430EXPORT_PER_CPU_SYMBOL(kstat);
Glauber Costa3292beb2011-11-28 14:45:17 -02005431EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005432
5433/*
Giovanni Gherdovich60756202016-08-05 10:21:56 +02005434 * The function fair_sched_class.update_curr accesses the struct curr
5435 * and its field curr->exec_start; when called from task_sched_runtime(),
5436 * we observe a high rate of cache misses in practice.
5437 * Prefetching this data results in improved performance.
5438 */
5439static inline void prefetch_curr_exec_start(struct task_struct *p)
5440{
5441#ifdef CONFIG_FAIR_GROUP_SCHED
5442 struct sched_entity *curr = (&p->se)->cfs_rq->curr;
5443#else
5444 struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
5445#endif
5446 prefetch(curr);
5447 prefetch(&curr->exec_start);
5448}
5449
5450/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09005451 * Return accounted runtime for the task.
5452 * In case the task is currently running, return the runtime plus current's
5453 * pending runtime that have not been accounted yet.
5454 */
5455unsigned long long task_sched_runtime(struct task_struct *p)
5456{
Peter Zijlstraeb580752015-07-31 21:28:18 +02005457 struct rq_flags rf;
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09005458 struct rq *rq;
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01005459 u64 ns;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005460
Peter Zijlstra911b2892013-11-11 18:21:56 +01005461#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
5462 /*
Ingo Molnar97fb7a02018-03-03 14:01:12 +01005463 * 64-bit doesn't need locks to atomically read a 64-bit value.
Peter Zijlstra911b2892013-11-11 18:21:56 +01005464 * So we have a optimization chance when the task's delta_exec is 0.
5465 * Reading ->on_cpu is racy, but this is ok.
5466 *
Ingo Molnard1ccc662017-02-01 11:46:42 +01005467 * If we race with it leaving CPU, we'll take a lock. So we're correct.
5468 * If we race with it entering CPU, unaccounted time is 0. This is
Peter Zijlstra911b2892013-11-11 18:21:56 +01005469 * indistinguishable from the read occurring a few cycles earlier.
Mike Galbraith4036ac12014-06-24 07:49:40 +02005470 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
5471 * been accounted, so we're correct here as well.
Peter Zijlstra911b2892013-11-11 18:21:56 +01005472 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04005473 if (!p->on_cpu || !task_on_rq_queued(p))
Peter Zijlstra911b2892013-11-11 18:21:56 +01005474 return p->se.sum_exec_runtime;
5475#endif
5476
Peter Zijlstraeb580752015-07-31 21:28:18 +02005477 rq = task_rq_lock(p, &rf);
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01005478 /*
5479 * Must be ->curr _and_ ->on_rq. If dequeued, we would
5480 * project cycles that may never be accounted to this
5481 * thread, breaking clock_gettime().
5482 */
5483 if (task_current(rq, p) && task_on_rq_queued(p)) {
Giovanni Gherdovich60756202016-08-05 10:21:56 +02005484 prefetch_curr_exec_start(p);
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01005485 update_rq_clock(rq);
5486 p->sched_class->update_curr(rq);
5487 }
5488 ns = p->se.sum_exec_runtime;
Peter Zijlstraeb580752015-07-31 21:28:18 +02005489 task_rq_unlock(rq, p, &rf);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09005490
5491 return ns;
5492}
5493
Paul Turnerc006fac2021-04-16 14:29:36 -07005494#ifdef CONFIG_SCHED_DEBUG
5495static u64 cpu_resched_latency(struct rq *rq)
5496{
5497 int latency_warn_ms = READ_ONCE(sysctl_resched_latency_warn_ms);
5498 u64 resched_latency, now = rq_clock(rq);
5499 static bool warned_once;
5500
5501 if (sysctl_resched_latency_warn_once && warned_once)
5502 return 0;
5503
5504 if (!need_resched() || !latency_warn_ms)
5505 return 0;
5506
5507 if (system_state == SYSTEM_BOOTING)
5508 return 0;
5509
5510 if (!rq->last_seen_need_resched_ns) {
5511 rq->last_seen_need_resched_ns = now;
5512 rq->ticks_without_resched = 0;
5513 return 0;
5514 }
5515
5516 rq->ticks_without_resched++;
5517 resched_latency = now - rq->last_seen_need_resched_ns;
5518 if (resched_latency <= latency_warn_ms * NSEC_PER_MSEC)
5519 return 0;
5520
5521 warned_once = true;
5522
5523 return resched_latency;
5524}
5525
5526static int __init setup_resched_latency_warn_ms(char *str)
5527{
5528 long val;
5529
5530 if ((kstrtol(str, 0, &val))) {
5531 pr_warn("Unable to set resched_latency_warn_ms\n");
5532 return 1;
5533 }
5534
5535 sysctl_resched_latency_warn_ms = val;
5536 return 1;
5537}
5538__setup("resched_latency_warn_ms=", setup_resched_latency_warn_ms);
5539#else
5540static inline u64 cpu_resched_latency(struct rq *rq) { return 0; }
5541#endif /* CONFIG_SCHED_DEBUG */
5542
Balbir Singh49048622008-09-05 18:12:23 +02005543/*
Christoph Lameter7835b982006-12-10 02:20:22 -08005544 * This function gets called by the timer code, with HZ frequency.
5545 * We call it with interrupts disabled.
Christoph Lameter7835b982006-12-10 02:20:22 -08005546 */
5547void scheduler_tick(void)
5548{
Christoph Lameter7835b982006-12-10 02:20:22 -08005549 int cpu = smp_processor_id();
5550 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005551 struct task_struct *curr = rq->curr;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005552 struct rq_flags rf;
Thara Gopinathb4eccf52020-02-21 19:52:10 -05005553 unsigned long thermal_pressure;
Paul Turnerc006fac2021-04-16 14:29:36 -07005554 u64 resched_latency;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005555
Yair Podemsky7fb3ff22022-11-30 14:51:21 +02005556 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
5557 arch_scale_freq_tick();
5558
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005559 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08005560
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005561 rq_lock(rq, &rf);
5562
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005563 update_rq_clock(rq);
Thara Gopinathb4eccf52020-02-21 19:52:10 -05005564 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
Thara Gopinath05289b92020-02-21 19:52:13 -05005565 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01005566 curr->sched_class->task_tick(rq, curr, 0);
Paul Turnerc006fac2021-04-16 14:29:36 -07005567 if (sched_feat(LATENCY_WARN))
5568 resched_latency = cpu_resched_latency(rq);
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02005569 calc_global_load_tick(rq);
Josh Don4feee7d12021-10-18 13:34:28 -07005570 sched_core_tick(rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02005571
5572 rq_unlock(rq, &rf);
Ingo Molnardd41f592007-07-09 18:51:59 +02005573
Paul Turnerc006fac2021-04-16 14:29:36 -07005574 if (sched_feat(LATENCY_WARN) && resched_latency)
5575 resched_latency_warn(cpu, resched_latency);
5576
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02005577 perf_event_task_tick();
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02005578
Christoph Lametere418e1c2006-12-10 02:20:23 -08005579#ifdef CONFIG_SMP
Suresh Siddha6eb57e02011-10-03 15:09:01 -07005580 rq->idle_balance = idle_cpu(cpu);
Daniel Lezcano7caff662014-01-06 12:34:38 +01005581 trigger_load_balance(rq);
Christoph Lametere418e1c2006-12-10 02:20:23 -08005582#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005583}
5584
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02005585#ifdef CONFIG_NO_HZ_FULL
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005586
5587struct tick_work {
5588 int cpu;
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005589 atomic_t state;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005590 struct delayed_work work;
5591};
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005592/* Values for ->state, see diagram below. */
5593#define TICK_SCHED_REMOTE_OFFLINE 0
5594#define TICK_SCHED_REMOTE_OFFLINING 1
5595#define TICK_SCHED_REMOTE_RUNNING 2
5596
5597/*
5598 * State diagram for ->state:
5599 *
5600 *
5601 * TICK_SCHED_REMOTE_OFFLINE
5602 * | ^
5603 * | |
5604 * | | sched_tick_remote()
5605 * | |
5606 * | |
5607 * +--TICK_SCHED_REMOTE_OFFLINING
5608 * | ^
5609 * | |
5610 * sched_tick_start() | | sched_tick_stop()
5611 * | |
5612 * V |
5613 * TICK_SCHED_REMOTE_RUNNING
5614 *
5615 *
5616 * Other transitions get WARN_ON_ONCE(), except that sched_tick_remote()
5617 * and sched_tick_start() are happy to leave the state in RUNNING.
5618 */
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005619
5620static struct tick_work __percpu *tick_work_cpu;
5621
5622static void sched_tick_remote(struct work_struct *work)
5623{
5624 struct delayed_work *dwork = to_delayed_work(work);
5625 struct tick_work *twork = container_of(dwork, struct tick_work, work);
5626 int cpu = twork->cpu;
5627 struct rq *rq = cpu_rq(cpu);
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005628 struct task_struct *curr;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005629 struct rq_flags rf;
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005630 u64 delta;
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005631 int os;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005632
5633 /*
5634 * Handle the tick only if it appears the remote CPU is running in full
5635 * dynticks mode. The check is racy by nature, but missing a tick or
5636 * having one too much is no big deal because the scheduler tick updates
5637 * statistics and checks timeslices in a time-independent way, regardless
5638 * of when exactly it is running.
5639 */
Scott Wood488603b2020-01-11 04:53:38 -05005640 if (!tick_nohz_tick_stopped_cpu(cpu))
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005641 goto out_requeue;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005642
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005643 rq_lock_irq(rq, &rf);
5644 curr = rq->curr;
Scott Wood488603b2020-01-11 04:53:38 -05005645 if (cpu_is_offline(cpu))
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005646 goto out_unlock;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005647
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005648 update_rq_clock(rq);
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005649
Scott Wood488603b2020-01-11 04:53:38 -05005650 if (!is_idle_task(curr)) {
5651 /*
5652 * Make sure the next tick runs within a reasonable
5653 * amount of time.
5654 */
5655 delta = rq_clock_task(rq) - curr->se.exec_start;
5656 WARN_ON_ONCE(delta > (u64)NSEC_PER_SEC * 3);
5657 }
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005658 curr->sched_class->task_tick(rq, curr, 0);
5659
Peter Zijlstra (Intel)ebc0f832020-01-11 04:53:39 -05005660 calc_load_nohz_remote(rq);
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005661out_unlock:
5662 rq_unlock_irq(rq, &rf);
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005663out_requeue:
Peter Zijlstra (Intel)ebc0f832020-01-11 04:53:39 -05005664
Frederic Weisbeckerd9c0ffc2018-06-28 18:29:41 +02005665 /*
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005666 * Run the remote tick once per second (1Hz). This arbitrary
5667 * frequency is large enough to avoid overload but short enough
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005668 * to keep scheduler internal stats reasonably up to date. But
5669 * first update state to reflect hotplug activity if required.
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005670 */
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005671 os = atomic_fetch_add_unless(&twork->state, -1, TICK_SCHED_REMOTE_RUNNING);
5672 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_OFFLINE);
5673 if (os == TICK_SCHED_REMOTE_RUNNING)
5674 queue_delayed_work(system_unbound_wq, dwork, HZ);
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005675}
5676
5677static void sched_tick_start(int cpu)
5678{
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005679 int os;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005680 struct tick_work *twork;
5681
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01005682 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005683 return;
5684
5685 WARN_ON_ONCE(!tick_work_cpu);
5686
5687 twork = per_cpu_ptr(tick_work_cpu, cpu);
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005688 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_RUNNING);
5689 WARN_ON_ONCE(os == TICK_SCHED_REMOTE_RUNNING);
5690 if (os == TICK_SCHED_REMOTE_OFFLINE) {
5691 twork->cpu = cpu;
5692 INIT_DELAYED_WORK(&twork->work, sched_tick_remote);
5693 queue_delayed_work(system_unbound_wq, &twork->work, HZ);
5694 }
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005695}
5696
5697#ifdef CONFIG_HOTPLUG_CPU
5698static void sched_tick_stop(int cpu)
5699{
5700 struct tick_work *twork;
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005701 int os;
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005702
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01005703 if (housekeeping_cpu(cpu, HK_TYPE_TICK))
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005704 return;
5705
5706 WARN_ON_ONCE(!tick_work_cpu);
5707
5708 twork = per_cpu_ptr(tick_work_cpu, cpu);
Paul E. McKenneyb55bd582019-05-30 05:39:25 -07005709 /* There cannot be competing actions, but don't rely on stop-machine. */
5710 os = atomic_xchg(&twork->state, TICK_SCHED_REMOTE_OFFLINING);
5711 WARN_ON_ONCE(os != TICK_SCHED_REMOTE_RUNNING);
5712 /* Don't cancel, as this would mess up the state machine. */
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005713}
5714#endif /* CONFIG_HOTPLUG_CPU */
5715
5716int __init sched_tick_offload_init(void)
5717{
5718 tick_work_cpu = alloc_percpu(struct tick_work);
5719 BUG_ON(!tick_work_cpu);
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01005720 return 0;
5721}
5722
5723#else /* !CONFIG_NO_HZ_FULL */
5724static inline void sched_tick_start(int cpu) { }
5725static inline void sched_tick_stop(int cpu) { }
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02005726#endif
5727
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02005728#if defined(CONFIG_PREEMPTION) && (defined(CONFIG_DEBUG_PREEMPT) || \
Joel Fernandes (Google)c3bc8fd2018-07-30 15:24:23 -07005729 defined(CONFIG_TRACE_PREEMPT_TOGGLE))
Steven Rostedt47252cf2016-03-21 11:23:39 -04005730/*
5731 * If the value passed in is equal to the current preempt count
5732 * then we just disabled preemption. Start timing the latency.
5733 */
5734static inline void preempt_latency_start(int val)
5735{
5736 if (preempt_count() == val) {
5737 unsigned long ip = get_lock_parent_ip();
5738#ifdef CONFIG_DEBUG_PREEMPT
5739 current->preempt_disable_ip = ip;
5740#endif
5741 trace_preempt_off(CALLER_ADDR0, ip);
5742 }
5743}
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05005744
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09005745void preempt_count_add(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005746{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005747#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005748 /*
5749 * Underflow?
5750 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005751 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5752 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005753#endif
Peter Zijlstrabdb43802013-09-10 12:15:23 +02005754 __preempt_count_add(val);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005755#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005756 /*
5757 * Spinlock count overflowing soon?
5758 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005759 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5760 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005761#endif
Steven Rostedt47252cf2016-03-21 11:23:39 -04005762 preempt_latency_start(val);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005763}
Peter Zijlstrabdb43802013-09-10 12:15:23 +02005764EXPORT_SYMBOL(preempt_count_add);
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09005765NOKPROBE_SYMBOL(preempt_count_add);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005766
Steven Rostedt47252cf2016-03-21 11:23:39 -04005767/*
5768 * If the value passed in equals to the current preempt count
5769 * then we just enabled preemption. Stop timing the latency.
5770 */
5771static inline void preempt_latency_stop(int val)
5772{
5773 if (preempt_count() == val)
5774 trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip());
5775}
5776
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09005777void preempt_count_sub(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005778{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005779#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005780 /*
5781 * Underflow?
5782 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01005783 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005784 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005785 /*
5786 * Is the spinlock portion underflowing?
5787 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005788 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5789 !(preempt_count() & PREEMPT_MASK)))
5790 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005791#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005792
Steven Rostedt47252cf2016-03-21 11:23:39 -04005793 preempt_latency_stop(val);
Peter Zijlstrabdb43802013-09-10 12:15:23 +02005794 __preempt_count_sub(val);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005795}
Peter Zijlstrabdb43802013-09-10 12:15:23 +02005796EXPORT_SYMBOL(preempt_count_sub);
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09005797NOKPROBE_SYMBOL(preempt_count_sub);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005798
Steven Rostedt47252cf2016-03-21 11:23:39 -04005799#else
5800static inline void preempt_latency_start(int val) { }
5801static inline void preempt_latency_stop(int val) { }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005802#endif
5803
Ingo Molnar59ddbcb2017-02-03 23:37:48 +01005804static inline unsigned long get_preempt_disable_ip(struct task_struct *p)
5805{
5806#ifdef CONFIG_DEBUG_PREEMPT
5807 return p->preempt_disable_ip;
5808#else
5809 return 0;
5810#endif
5811}
5812
Linus Torvalds1da177e2005-04-16 15:20:36 -07005813/*
Ingo Molnardd41f592007-07-09 18:51:59 +02005814 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005815 */
Ingo Molnardd41f592007-07-09 18:51:59 +02005816static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005817{
Vegard Nossumd1c6d142016-07-23 09:46:39 +02005818 /* Save this before calling printk(), since that will clobber it */
5819 unsigned long preempt_disable_ip = get_preempt_disable_ip(current);
5820
Dave Jones664dfa62011-12-22 16:39:30 -05005821 if (oops_in_progress)
5822 return;
5823
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005824 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5825 prev->comm, prev->pid, preempt_count());
Satyam Sharma838225b2007-10-24 18:23:50 +02005826
Ingo Molnardd41f592007-07-09 18:51:59 +02005827 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07005828 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02005829 if (irqs_disabled())
5830 print_irqtrace_events(prev);
Vegard Nossumd1c6d142016-07-23 09:46:39 +02005831 if (IS_ENABLED(CONFIG_DEBUG_PREEMPT)
5832 && in_atomic_preempt_off()) {
Thomas Gleixner8f47b182014-02-07 20:58:39 +01005833 pr_err("Preemption disabled at:");
Dmitry Safonov2062a4e2020-06-08 21:29:56 -07005834 print_ip_sym(KERN_ERR, preempt_disable_ip);
Thomas Gleixner8f47b182014-02-07 20:58:39 +01005835 }
Kees Cook79cc1ba2022-11-17 15:43:24 -08005836 check_panic_on_warn("scheduling while atomic");
Daniel Bristot de Oliveira748c7202016-06-03 17:10:18 -03005837
Stephen Boyd6135fc12012-03-28 17:10:47 -07005838 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +10305839 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
Ingo Molnardd41f592007-07-09 18:51:59 +02005840}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005841
Ingo Molnardd41f592007-07-09 18:51:59 +02005842/*
5843 * Various schedule()-time debugging checks and statistics:
5844 */
Daniel Vetter312364f32019-08-26 22:14:23 +02005845static inline void schedule_debug(struct task_struct *prev, bool preempt)
Ingo Molnardd41f592007-07-09 18:51:59 +02005846{
Aaron Tomlin0d9e2632014-09-12 14:16:19 +01005847#ifdef CONFIG_SCHED_STACK_END_CHECK
Jann Horn29d64552016-06-01 11:55:07 +02005848 if (task_stack_end_corrupted(prev))
5849 panic("corrupted stack end detected inside scheduler\n");
Will Deacon88485be2020-05-15 14:56:05 +01005850
5851 if (task_scs_end_corrupted(prev))
5852 panic("corrupted shadow stack detected inside scheduler\n");
Aaron Tomlin0d9e2632014-09-12 14:16:19 +01005853#endif
Peter Zijlstrab99def82015-09-28 18:02:03 +02005854
Daniel Vetter312364f32019-08-26 22:14:23 +02005855#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
Peter Zijlstra2f064a52021-06-11 10:28:17 +02005856 if (!preempt && READ_ONCE(prev->__state) && prev->non_block_count) {
Daniel Vetter312364f32019-08-26 22:14:23 +02005857 printk(KERN_ERR "BUG: scheduling in a non-blocking section: %s/%d/%i\n",
5858 prev->comm, prev->pid, prev->non_block_count);
5859 dump_stack();
5860 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
5861 }
5862#endif
5863
Peter Zijlstra1dc0fff2015-09-28 17:57:39 +02005864 if (unlikely(in_atomic_preempt_off())) {
Ingo Molnardd41f592007-07-09 18:51:59 +02005865 __schedule_bug(prev);
Peter Zijlstra1dc0fff2015-09-28 17:57:39 +02005866 preempt_count_set(PREEMPT_DISABLED);
5867 }
Paul E. McKenneyb3fbab02011-05-24 08:31:09 -07005868 rcu_sleep_check();
Frederic Weisbecker9f68b5b2020-11-17 16:16:35 +01005869 SCHED_WARN_ON(ct_state() == CONTEXT_USER);
Ingo Molnardd41f592007-07-09 18:51:59 +02005870
Linus Torvalds1da177e2005-04-16 15:20:36 -07005871 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5872
Josh Poimboeufae928822016-06-17 12:43:24 -05005873 schedstat_inc(this_rq()->sched_count);
Ingo Molnardd41f592007-07-09 18:51:59 +02005874}
5875
Chen Yu457d1f42020-04-21 18:50:43 +08005876static void put_prev_task_balance(struct rq *rq, struct task_struct *prev,
5877 struct rq_flags *rf)
5878{
5879#ifdef CONFIG_SMP
5880 const struct sched_class *class;
5881 /*
5882 * We must do the balancing pass before put_prev_task(), such
5883 * that when we release the rq->lock the task is in the same
5884 * state as before we took rq->lock.
5885 *
5886 * We can terminate the balance pass as soon as we know there is
5887 * a runnable task of @class priority or higher.
5888 */
5889 for_class_range(class, prev->sched_class, &idle_sched_class) {
5890 if (class->balance(rq, prev, rf))
5891 break;
5892 }
5893#endif
5894
5895 put_prev_task(rq, prev);
5896}
5897
Ingo Molnardd41f592007-07-09 18:51:59 +02005898/*
5899 * Pick up the highest-prio task:
5900 */
5901static inline struct task_struct *
Peter Zijlstra539f6512020-11-17 18:19:37 -05005902__pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
Ingo Molnardd41f592007-07-09 18:51:59 +02005903{
Peter Zijlstra49ee5762017-01-19 18:44:08 +01005904 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02005905 struct task_struct *p;
5906
5907 /*
Peter Zijlstra0ba87bb2017-03-01 10:51:47 +01005908 * Optimization: we know that if all tasks are in the fair class we can
5909 * call that function directly, but only if the @prev task wasn't of a
Tal Zussmanb19a8882020-11-12 19:51:56 -05005910 * higher scheduling class, because otherwise those lose the
Peter Zijlstra0ba87bb2017-03-01 10:51:47 +01005911 * opportunity to pull in more work from other CPUs.
Ingo Molnardd41f592007-07-09 18:51:59 +02005912 */
Peter Zijlstra546a3fe2022-05-17 13:46:54 +02005913 if (likely(!sched_class_above(prev->sched_class, &fair_sched_class) &&
Peter Zijlstra0ba87bb2017-03-01 10:51:47 +01005914 rq->nr_running == rq->cfs.h_nr_running)) {
5915
Peter Zijlstra5d7d6052019-11-08 14:15:57 +01005916 p = pick_next_task_fair(rq, prev, rf);
Peter Zijlstra6ccdc842014-04-24 12:00:47 +02005917 if (unlikely(p == RETRY_TASK))
Peter Zijlstra67692432019-05-29 20:36:44 +00005918 goto restart;
Peter Zijlstra6ccdc842014-04-24 12:00:47 +02005919
Masahiro Yamada16999492021-05-19 15:37:09 +09005920 /* Assume the next prioritized class is idle_sched_class */
Peter Zijlstra5d7d6052019-11-08 14:15:57 +01005921 if (!p) {
Peter Zijlstraf488e102019-11-08 14:15:56 +01005922 put_prev_task(rq, prev);
Peter Zijlstra98c2f702019-11-08 14:15:58 +01005923 p = pick_next_task_idle(rq);
Peter Zijlstraf488e102019-11-08 14:15:56 +01005924 }
Peter Zijlstra6ccdc842014-04-24 12:00:47 +02005925
5926 return p;
Ingo Molnardd41f592007-07-09 18:51:59 +02005927 }
5928
Peter Zijlstra67692432019-05-29 20:36:44 +00005929restart:
Chen Yu457d1f42020-04-21 18:50:43 +08005930 put_prev_task_balance(rq, prev, rf);
Peter Zijlstra67692432019-05-29 20:36:44 +00005931
Peter Zijlstra34f971f2010-09-22 13:53:15 +02005932 for_each_class(class) {
Peter Zijlstra98c2f702019-11-08 14:15:58 +01005933 p = class->pick_next_task(rq);
Peter Zijlstra67692432019-05-29 20:36:44 +00005934 if (p)
Ingo Molnardd41f592007-07-09 18:51:59 +02005935 return p;
Ingo Molnardd41f592007-07-09 18:51:59 +02005936 }
Peter Zijlstra34f971f2010-09-22 13:53:15 +02005937
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005938 BUG(); /* The idle class should always have a runnable task. */
Ingo Molnardd41f592007-07-09 18:51:59 +02005939}
5940
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05005941#ifdef CONFIG_SCHED_CORE
Peter Zijlstra539f6512020-11-17 18:19:37 -05005942static inline bool is_task_rq_idle(struct task_struct *t)
5943{
5944 return (task_rq(t)->idle == t);
5945}
5946
5947static inline bool cookie_equals(struct task_struct *a, unsigned long cookie)
5948{
5949 return is_task_rq_idle(a) || (a->core_cookie == cookie);
5950}
5951
5952static inline bool cookie_match(struct task_struct *a, struct task_struct *b)
5953{
5954 if (is_task_rq_idle(a) || is_task_rq_idle(b))
5955 return true;
5956
5957 return a->core_cookie == b->core_cookie;
5958}
5959
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005960static inline struct task_struct *pick_task(struct rq *rq)
Peter Zijlstra539f6512020-11-17 18:19:37 -05005961{
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005962 const struct sched_class *class;
5963 struct task_struct *p;
Peter Zijlstra539f6512020-11-17 18:19:37 -05005964
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005965 for_each_class(class) {
5966 p = class->pick_task(rq);
5967 if (p)
5968 return p;
Peter Zijlstra539f6512020-11-17 18:19:37 -05005969 }
5970
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005971 BUG(); /* The idle class should always have a runnable task. */
Peter Zijlstra539f6512020-11-17 18:19:37 -05005972}
5973
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -05005974extern void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi);
5975
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01005976static void queue_core_balance(struct rq *rq);
5977
Peter Zijlstra539f6512020-11-17 18:19:37 -05005978static struct task_struct *
5979pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
5980{
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005981 struct task_struct *next, *p, *max = NULL;
Peter Zijlstra539f6512020-11-17 18:19:37 -05005982 const struct cpumask *smt_mask;
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -05005983 bool fi_before = false;
Josh Don4feee7d12021-10-18 13:34:28 -07005984 bool core_clock_updated = (rq == rq->core);
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02005985 unsigned long cookie;
5986 int i, cpu, occ = 0;
5987 struct rq *rq_i;
Peter Zijlstra539f6512020-11-17 18:19:37 -05005988 bool need_sync;
Peter Zijlstra539f6512020-11-17 18:19:37 -05005989
5990 if (!sched_core_enabled(rq))
5991 return __pick_next_task(rq, prev, rf);
5992
5993 cpu = cpu_of(rq);
5994
5995 /* Stopper task is switching into idle, no need core-wide selection. */
5996 if (cpu_is_offline(cpu)) {
5997 /*
5998 * Reset core_pick so that we don't enter the fastpath when
5999 * coming online. core_pick would already be migrated to
6000 * another cpu during offline.
6001 */
6002 rq->core_pick = NULL;
6003 return __pick_next_task(rq, prev, rf);
6004 }
6005
6006 /*
6007 * If there were no {en,de}queues since we picked (IOW, the task
6008 * pointers are all still valid), and we haven't scheduled the last
6009 * pick yet, do so now.
6010 *
6011 * rq->core_pick can be NULL if no selection was made for a CPU because
6012 * it was either offline or went offline during a sibling's core-wide
6013 * selection. In this case, do a core-wide selection.
6014 */
6015 if (rq->core->core_pick_seq == rq->core->core_task_seq &&
6016 rq->core->core_pick_seq != rq->core_sched_seq &&
6017 rq->core_pick) {
6018 WRITE_ONCE(rq->core_sched_seq, rq->core->core_pick_seq);
6019
6020 next = rq->core_pick;
6021 if (next != prev) {
6022 put_prev_task(rq, prev);
6023 set_next_task(rq, next);
6024 }
6025
6026 rq->core_pick = NULL;
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01006027 goto out;
Peter Zijlstra539f6512020-11-17 18:19:37 -05006028 }
6029
6030 put_prev_task_balance(rq, prev, rf);
6031
6032 smt_mask = cpu_smt_mask(cpu);
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006033 need_sync = !!rq->core->core_cookie;
6034
6035 /* reset state */
6036 rq->core->core_cookie = 0UL;
Josh Don4feee7d12021-10-18 13:34:28 -07006037 if (rq->core->core_forceidle_count) {
6038 if (!core_clock_updated) {
6039 update_rq_clock(rq->core);
6040 core_clock_updated = true;
6041 }
6042 sched_core_account_forceidle(rq);
6043 /* reset after accounting force idle */
6044 rq->core->core_forceidle_start = 0;
6045 rq->core->core_forceidle_count = 0;
6046 rq->core->core_forceidle_occupation = 0;
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006047 need_sync = true;
6048 fi_before = true;
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006049 }
Peter Zijlstra539f6512020-11-17 18:19:37 -05006050
6051 /*
6052 * core->core_task_seq, core->core_pick_seq, rq->core_sched_seq
6053 *
6054 * @task_seq guards the task state ({en,de}queues)
6055 * @pick_seq is the @task_seq we did a selection on
6056 * @sched_seq is the @pick_seq we scheduled
6057 *
6058 * However, preemptions can cause multiple picks on the same task set.
6059 * 'Fix' this by also increasing @task_seq for every pick.
6060 */
6061 rq->core->core_task_seq++;
Peter Zijlstra539f6512020-11-17 18:19:37 -05006062
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006063 /*
6064 * Optimize for common case where this CPU has no cookies
6065 * and there are no cookied tasks running on siblings.
6066 */
6067 if (!need_sync) {
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006068 next = pick_task(rq);
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006069 if (!next->core_cookie) {
6070 rq->core_pick = NULL;
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -05006071 /*
6072 * For robustness, update the min_vruntime_fi for
6073 * unconstrained picks as well.
6074 */
6075 WARN_ON_ONCE(fi_before);
6076 task_vruntime_update(rq, next, false);
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01006077 goto out_set_next;
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006078 }
Vineeth Pillai8039e96f2020-11-17 18:19:38 -05006079 }
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006080
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006081 /*
6082 * For each thread: do the regular task pick and find the max prio task
6083 * amongst them.
6084 *
6085 * Tie-break prio towards the current CPU
6086 */
6087 for_each_cpu_wrap(i, smt_mask, cpu) {
6088 rq_i = cpu_rq(i);
Peter Zijlstra539f6512020-11-17 18:19:37 -05006089
Josh Don4feee7d12021-10-18 13:34:28 -07006090 /*
6091 * Current cpu always has its clock updated on entrance to
6092 * pick_next_task(). If the current cpu is not the core,
6093 * the core may also have been updated above.
6094 */
6095 if (i != cpu && (rq_i != rq->core || !core_clock_updated))
Peter Zijlstra539f6512020-11-17 18:19:37 -05006096 update_rq_clock(rq_i);
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006097
6098 p = rq_i->core_pick = pick_task(rq_i);
6099 if (!max || prio_less(max, p, fi_before))
6100 max = p;
Peter Zijlstra539f6512020-11-17 18:19:37 -05006101 }
6102
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006103 cookie = rq->core->core_cookie = max->core_cookie;
6104
Peter Zijlstra539f6512020-11-17 18:19:37 -05006105 /*
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006106 * For each thread: try and find a runnable task that matches @max or
6107 * force idle.
Peter Zijlstra539f6512020-11-17 18:19:37 -05006108 */
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006109 for_each_cpu(i, smt_mask) {
6110 rq_i = cpu_rq(i);
6111 p = rq_i->core_pick;
Peter Zijlstra539f6512020-11-17 18:19:37 -05006112
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006113 if (!cookie_equals(p, cookie)) {
6114 p = NULL;
6115 if (cookie)
6116 p = sched_core_find(rq_i, cookie);
Joel Fernandes (Google)7afbba12020-11-17 18:19:42 -05006117 if (!p)
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006118 p = idle_sched_class.pick_task(rq_i);
6119 }
Peter Zijlstra539f6512020-11-17 18:19:37 -05006120
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006121 rq_i->core_pick = p;
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006122
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006123 if (p == rq_i->idle) {
6124 if (rq_i->nr_running) {
Josh Don4feee7d12021-10-18 13:34:28 -07006125 rq->core->core_forceidle_count++;
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -05006126 if (!fi_before)
6127 rq->core->core_forceidle_seq++;
6128 }
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006129 } else {
6130 occ++;
Peter Zijlstra539f6512020-11-17 18:19:37 -05006131 }
Peter Zijlstra539f6512020-11-17 18:19:37 -05006132 }
6133
Josh Don4feee7d12021-10-18 13:34:28 -07006134 if (schedstat_enabled() && rq->core->core_forceidle_count) {
Cruz Zhaob1715012022-01-11 17:55:59 +08006135 rq->core->core_forceidle_start = rq_clock(rq->core);
Josh Don4feee7d12021-10-18 13:34:28 -07006136 rq->core->core_forceidle_occupation = occ;
6137 }
6138
Peter Zijlstra539f6512020-11-17 18:19:37 -05006139 rq->core->core_pick_seq = rq->core->core_task_seq;
6140 next = rq->core_pick;
6141 rq->core_sched_seq = rq->core->core_pick_seq;
6142
6143 /* Something should have been selected for current CPU */
6144 WARN_ON_ONCE(!next);
6145
6146 /*
6147 * Reschedule siblings
6148 *
6149 * NOTE: L1TF -- at this point we're no longer running the old task and
6150 * sending an IPI (below) ensures the sibling will no longer be running
6151 * their task. This ensures there is no inter-sibling overlap between
6152 * non-matching user state.
6153 */
6154 for_each_cpu(i, smt_mask) {
Peter Zijlstrabc9ffef2021-08-24 11:05:47 +02006155 rq_i = cpu_rq(i);
Peter Zijlstra539f6512020-11-17 18:19:37 -05006156
6157 /*
6158 * An online sibling might have gone offline before a task
6159 * could be picked for it, or it might be offline but later
6160 * happen to come online, but its too late and nothing was
6161 * picked for it. That's Ok - it will pick tasks for itself,
6162 * so ignore it.
6163 */
6164 if (!rq_i->core_pick)
6165 continue;
6166
Joel Fernandes (Google)c6047c22020-11-17 18:19:39 -05006167 /*
6168 * Update for new !FI->FI transitions, or if continuing to be in !FI:
6169 * fi_before fi update?
6170 * 0 0 1
6171 * 0 1 1
6172 * 1 0 1
6173 * 1 1 0
6174 */
Josh Don4feee7d12021-10-18 13:34:28 -07006175 if (!(fi_before && rq->core->core_forceidle_count))
6176 task_vruntime_update(rq_i, rq_i->core_pick, !!rq->core->core_forceidle_count);
Peter Zijlstra539f6512020-11-17 18:19:37 -05006177
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006178 rq_i->core_pick->core_occupation = occ;
6179
Peter Zijlstra539f6512020-11-17 18:19:37 -05006180 if (i == cpu) {
6181 rq_i->core_pick = NULL;
6182 continue;
6183 }
6184
6185 /* Did we break L1TF mitigation requirements? */
6186 WARN_ON_ONCE(!cookie_match(next, rq_i->core_pick));
6187
6188 if (rq_i->curr == rq_i->core_pick) {
6189 rq_i->core_pick = NULL;
6190 continue;
6191 }
6192
6193 resched_curr(rq_i);
6194 }
6195
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01006196out_set_next:
Peter Zijlstra539f6512020-11-17 18:19:37 -05006197 set_next_task(rq, next);
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01006198out:
6199 if (rq->core->core_forceidle_count && next == rq->idle)
6200 queue_core_balance(rq);
6201
Peter Zijlstra539f6512020-11-17 18:19:37 -05006202 return next;
6203}
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006204
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006205static bool try_steal_cookie(int this, int that)
6206{
6207 struct rq *dst = cpu_rq(this), *src = cpu_rq(that);
6208 struct task_struct *p;
6209 unsigned long cookie;
6210 bool success = false;
6211
6212 local_irq_disable();
6213 double_rq_lock(dst, src);
6214
6215 cookie = dst->core->core_cookie;
6216 if (!cookie)
6217 goto unlock;
6218
6219 if (dst->curr != dst->idle)
6220 goto unlock;
6221
6222 p = sched_core_find(src, cookie);
6223 if (p == src->idle)
6224 goto unlock;
6225
6226 do {
6227 if (p == src->core_pick || p == src->curr)
6228 goto next;
6229
Sebastian Andrzej Siewior386ef212022-03-17 15:51:32 +01006230 if (!is_cpu_allowed(p, this))
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006231 goto next;
6232
6233 if (p->core_occupation > dst->idle->core_occupation)
6234 goto next;
6235
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006236 deactivate_task(src, p, 0);
6237 set_task_cpu(p, this);
6238 activate_task(dst, p, 0);
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006239
6240 resched_curr(dst);
6241
6242 success = true;
6243 break;
6244
6245next:
6246 p = sched_core_next(p, cookie);
6247 } while (p);
6248
6249unlock:
6250 double_rq_unlock(dst, src);
6251 local_irq_enable();
6252
6253 return success;
6254}
6255
6256static bool steal_cookie_task(int cpu, struct sched_domain *sd)
6257{
6258 int i;
6259
6260 for_each_cpu_wrap(i, sched_domain_span(sd), cpu) {
6261 if (i == cpu)
6262 continue;
6263
6264 if (need_resched())
6265 break;
6266
6267 if (try_steal_cookie(cpu, i))
6268 return true;
6269 }
6270
6271 return false;
6272}
6273
6274static void sched_core_balance(struct rq *rq)
6275{
6276 struct sched_domain *sd;
6277 int cpu = cpu_of(rq);
6278
6279 preempt_disable();
6280 rcu_read_lock();
6281 raw_spin_rq_unlock_irq(rq);
6282 for_each_domain(cpu, sd) {
6283 if (need_resched())
6284 break;
6285
6286 if (steal_cookie_task(cpu, sd))
6287 break;
6288 }
6289 raw_spin_rq_lock_irq(rq);
6290 rcu_read_unlock();
6291 preempt_enable();
6292}
6293
Kees Cook8e5bad72022-10-07 17:07:58 -07006294static DEFINE_PER_CPU(struct balance_callback, core_balance_head);
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006295
Peter Zijlstra5b6547e2022-03-16 22:03:41 +01006296static void queue_core_balance(struct rq *rq)
Peter Zijlstrad2dfa172020-11-17 18:19:43 -05006297{
6298 if (!sched_core_enabled(rq))
6299 return;
6300
6301 if (!rq->core->core_cookie)
6302 return;
6303
6304 if (!rq->nr_running) /* not forced idle */
6305 return;
6306
6307 queue_balance_callback(rq, &per_cpu(core_balance_head, rq->cpu), sched_core_balance);
6308}
6309
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006310static void sched_core_cpu_starting(unsigned int cpu)
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006311{
6312 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006313 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6314 unsigned long flags;
6315 int t;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006316
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006317 sched_core_lock(cpu, &flags);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006318
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006319 WARN_ON_ONCE(rq->core != rq);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006320
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006321 /* if we're the first, we'll be our own leader */
6322 if (cpumask_weight(smt_mask) == 1)
6323 goto unlock;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006324
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006325 /* find the leader */
6326 for_each_cpu(t, smt_mask) {
6327 if (t == cpu)
6328 continue;
6329 rq = cpu_rq(t);
6330 if (rq->core == rq) {
6331 core_rq = rq;
6332 break;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006333 }
6334 }
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006335
6336 if (WARN_ON_ONCE(!core_rq)) /* whoopsie */
6337 goto unlock;
6338
6339 /* install and validate core_rq */
6340 for_each_cpu(t, smt_mask) {
6341 rq = cpu_rq(t);
6342
6343 if (t == cpu)
6344 rq->core = core_rq;
6345
6346 WARN_ON_ONCE(rq->core != core_rq);
6347 }
6348
6349unlock:
6350 sched_core_unlock(cpu, &flags);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006351}
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006352
6353static void sched_core_cpu_deactivate(unsigned int cpu)
6354{
6355 const struct cpumask *smt_mask = cpu_smt_mask(cpu);
6356 struct rq *rq = cpu_rq(cpu), *core_rq = NULL;
6357 unsigned long flags;
6358 int t;
6359
6360 sched_core_lock(cpu, &flags);
6361
6362 /* if we're the last man standing, nothing to do */
6363 if (cpumask_weight(smt_mask) == 1) {
6364 WARN_ON_ONCE(rq->core != rq);
6365 goto unlock;
6366 }
6367
6368 /* if we're not the leader, nothing to do */
6369 if (rq->core != rq)
6370 goto unlock;
6371
6372 /* find a new leader */
6373 for_each_cpu(t, smt_mask) {
6374 if (t == cpu)
6375 continue;
6376 core_rq = cpu_rq(t);
6377 break;
6378 }
6379
6380 if (WARN_ON_ONCE(!core_rq)) /* impossible */
6381 goto unlock;
6382
6383 /* copy the shared state to the new leader */
Josh Don4feee7d12021-10-18 13:34:28 -07006384 core_rq->core_task_seq = rq->core_task_seq;
6385 core_rq->core_pick_seq = rq->core_pick_seq;
6386 core_rq->core_cookie = rq->core_cookie;
6387 core_rq->core_forceidle_count = rq->core_forceidle_count;
6388 core_rq->core_forceidle_seq = rq->core_forceidle_seq;
6389 core_rq->core_forceidle_occupation = rq->core_forceidle_occupation;
6390
6391 /*
6392 * Accounting edge for forced idle is handled in pick_next_task().
6393 * Don't need another one here, since the hotplug thread shouldn't
6394 * have a cookie.
6395 */
6396 core_rq->core_forceidle_start = 0;
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006397
6398 /* install new leader */
6399 for_each_cpu(t, smt_mask) {
6400 rq = cpu_rq(t);
6401 rq->core = core_rq;
6402 }
6403
6404unlock:
6405 sched_core_unlock(cpu, &flags);
6406}
6407
6408static inline void sched_core_cpu_dying(unsigned int cpu)
6409{
6410 struct rq *rq = cpu_rq(cpu);
6411
6412 if (rq->core != rq)
6413 rq->core = rq;
6414}
6415
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006416#else /* !CONFIG_SCHED_CORE */
6417
6418static inline void sched_core_cpu_starting(unsigned int cpu) {}
Peter Zijlstra3c474b32021-08-19 13:09:17 +02006419static inline void sched_core_cpu_deactivate(unsigned int cpu) {}
6420static inline void sched_core_cpu_dying(unsigned int cpu) {}
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006421
Peter Zijlstra539f6512020-11-17 18:19:37 -05006422static struct task_struct *
6423pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6424{
6425 return __pick_next_task(rq, prev, rf);
6426}
6427
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05006428#endif /* CONFIG_SCHED_CORE */
6429
Ingo Molnardd41f592007-07-09 18:51:59 +02006430/*
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006431 * Constants for the sched_mode argument of __schedule().
6432 *
6433 * The mode argument allows RT enabled kernels to differentiate a
6434 * preemption from blocking on an 'sleeping' spin/rwlock. Note that
6435 * SM_MASK_PREEMPT for !RT has all bits set, which allows the compiler to
6436 * optimize the AND operation out and just check for zero.
6437 */
6438#define SM_NONE 0x0
6439#define SM_PREEMPT 0x1
Thomas Gleixner69914362021-08-15 23:27:48 +02006440#define SM_RTLOCK_WAIT 0x2
6441
6442#ifndef CONFIG_PREEMPT_RT
6443# define SM_MASK_PREEMPT (~0U)
6444#else
6445# define SM_MASK_PREEMPT SM_PREEMPT
6446#endif
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006447
6448/*
Thomas Gleixnerc259e012011-06-22 19:47:00 +02006449 * __schedule() is the main scheduler function.
Pekka Enbergedde96e2012-08-04 11:49:47 +03006450 *
6451 * The main means of driving the scheduler and thus entering this function are:
6452 *
6453 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
6454 *
6455 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
6456 * paths. For example, see arch/x86/entry_64.S.
6457 *
6458 * To drive preemption between tasks, the scheduler sets the flag in timer
6459 * interrupt handler scheduler_tick().
6460 *
6461 * 3. Wakeups don't really cause entry into schedule(). They add a
6462 * task to the run-queue and that's it.
6463 *
6464 * Now, if the new task added to the run-queue preempts the current
6465 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
6466 * called on the nearest possible occasion:
6467 *
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02006468 * - If the kernel is preemptible (CONFIG_PREEMPTION=y):
Pekka Enbergedde96e2012-08-04 11:49:47 +03006469 *
6470 * - in syscall or exception context, at the next outmost
6471 * preempt_enable(). (this might be as soon as the wake_up()'s
6472 * spin_unlock()!)
6473 *
6474 * - in IRQ context, return from interrupt-handler to
6475 * preemptible context
6476 *
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02006477 * - If the kernel is not preemptible (CONFIG_PREEMPTION is not set)
Pekka Enbergedde96e2012-08-04 11:49:47 +03006478 * then at the next:
6479 *
6480 * - cond_resched() call
6481 * - explicit schedule() call
6482 * - return from syscall or exception to user-space
6483 * - return from interrupt-handler to user-space
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01006484 *
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02006485 * WARNING: must be called with preemption disabled!
Ingo Molnardd41f592007-07-09 18:51:59 +02006486 */
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006487static void __sched notrace __schedule(unsigned int sched_mode)
Ingo Molnardd41f592007-07-09 18:51:59 +02006488{
6489 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08006490 unsigned long *switch_count;
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006491 unsigned long prev_state;
Matt Flemingd8ac8972016-09-21 14:38:10 +01006492 struct rq_flags rf;
Ingo Molnardd41f592007-07-09 18:51:59 +02006493 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02006494 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02006495
Ingo Molnardd41f592007-07-09 18:51:59 +02006496 cpu = smp_processor_id();
6497 rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02006498 prev = rq->curr;
Ingo Molnardd41f592007-07-09 18:51:59 +02006499
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006500 schedule_debug(prev, !!sched_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006501
Juri Lellie0ee4632021-02-08 08:35:54 +01006502 if (sched_feat(HRTICK) || sched_feat(HRTICK_DL))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02006503 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01006504
Paul E. McKenney46a5d162015-10-07 09:10:48 -07006505 local_irq_disable();
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006506 rcu_note_context_switch(!!sched_mode);
Paul E. McKenney46a5d162015-10-07 09:10:48 -07006507
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02006508 /*
6509 * Make sure that signal_pending_state()->signal_pending() below
6510 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006511 * done by the caller to avoid the race with signal_wake_up():
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05006512 *
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006513 * __set_current_state(@state) signal_wake_up()
6514 * schedule() set_tsk_thread_flag(p, TIF_SIGPENDING)
6515 * wake_up_state(p, state)
6516 * LOCK rq->lock LOCK p->pi_state
6517 * smp_mb__after_spinlock() smp_mb__after_spinlock()
6518 * if (signal_pending_state()) if (p->state & @state)
6519 *
6520 * Also, the membarrier system call requires a full memory barrier
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05006521 * after coming from user-space, before storing to rq->curr.
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02006522 */
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02006523 rq_lock(rq, &rf);
Peter Zijlstrad89e588c2016-09-05 11:37:53 +02006524 smp_mb__after_spinlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006525
Ingo Molnard1ccc662017-02-01 11:46:42 +01006526 /* Promote REQ to ACT */
6527 rq->clock_update_flags <<= 1;
Peter Zijlstrabce4dc82017-02-21 14:40:35 +01006528 update_rq_clock(rq);
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01006529
Oleg Nesterov246d86b2010-05-19 14:57:11 +02006530 switch_count = &prev->nivcsw;
Peter Zijlstrad1361222020-07-20 17:20:21 +02006531
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006532 /*
Peter Zijlstrad1361222020-07-20 17:20:21 +02006533 * We must load prev->state once (task_struct::state is volatile), such
Eric W. Biederman2500ad12022-04-29 08:43:34 -05006534 * that we form a control dependency vs deactivate_task() below.
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006535 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02006536 prev_state = READ_ONCE(prev->__state);
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006537 if (!(sched_mode & SM_MASK_PREEMPT) && prev_state) {
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006538 if (signal_pending_state(prev_state, prev)) {
Peter Zijlstra2f064a52021-06-11 10:28:17 +02006539 WRITE_ONCE(prev->__state, TASK_RUNNING);
Tejun Heo21aa9af2010-06-08 21:40:37 +02006540 } else {
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006541 prev->sched_contributes_to_load =
6542 (prev_state & TASK_UNINTERRUPTIBLE) &&
6543 !(prev_state & TASK_NOLOAD) &&
Peter Zijlstraf5d39b02022-08-22 13:18:22 +02006544 !(prev_state & TASK_FROZEN);
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006545
6546 if (prev->sched_contributes_to_load)
6547 rq->nr_uninterruptible++;
6548
6549 /*
6550 * __schedule() ttwu()
Peter Zijlstrad1361222020-07-20 17:20:21 +02006551 * prev_state = prev->state; if (p->on_rq && ...)
6552 * if (prev_state) goto out;
6553 * p->on_rq = 0; smp_acquire__after_ctrl_dep();
6554 * p->state = TASK_WAKING
6555 *
6556 * Where __schedule() and ttwu() have matching control dependencies.
Peter Zijlstradbfb0892020-07-03 12:40:33 +02006557 *
6558 * After this, schedule() must not care about p->state any more.
6559 */
Peter Zijlstrabce4dc82017-02-21 14:40:35 +01006560 deactivate_task(rq, prev, DEQUEUE_SLEEP | DEQUEUE_NOCLOCK);
Peter Zijlstra2acca552011-04-05 17:23:50 +02006561
Tejun Heoe33a9bb2016-12-07 15:48:41 -05006562 if (prev->in_iowait) {
6563 atomic_inc(&rq->nr_iowait);
6564 delayacct_blkio_start();
6565 }
Tejun Heo21aa9af2010-06-08 21:40:37 +02006566 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006567 switch_count = &prev->nvcsw;
6568 }
6569
Matt Flemingd8ac8972016-09-21 14:38:10 +01006570 next = pick_next_task(rq, prev, &rf);
Mike Galbraithf26f9af2010-12-08 11:05:42 +01006571 clear_tsk_need_resched(prev);
Peter Zijlstraf27dde82013-08-14 14:55:31 +02006572 clear_preempt_need_resched();
Paul Turnerc006fac2021-04-16 14:29:36 -07006573#ifdef CONFIG_SCHED_DEBUG
6574 rq->last_seen_need_resched_ns = 0;
6575#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006576
Linus Torvalds1da177e2005-04-16 15:20:36 -07006577 if (likely(prev != next)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006578 rq->nr_switches++;
Eric W. Biederman5311a982019-09-14 07:35:02 -05006579 /*
6580 * RCU users of rcu_dereference(rq->curr) may not see
6581 * changes to task_struct made by pick_next_task().
6582 */
6583 RCU_INIT_POINTER(rq->curr, next);
Mathieu Desnoyers22e4ebb2017-07-28 16:40:40 -04006584 /*
6585 * The membarrier system call requires each architecture
6586 * to have a full memory barrier after updating
Mathieu Desnoyers306e0602018-01-29 15:20:12 -05006587 * rq->curr, before returning to user-space.
6588 *
6589 * Here are the schemes providing that barrier on the
6590 * various architectures:
6591 * - mm ? switch_mm() : mmdrop() for x86, s390, sparc, PowerPC.
6592 * switch_mm() rely on membarrier_arch_switch_mm() on PowerPC.
6593 * - finish_lock_switch() for weakly-ordered
6594 * architectures where spin_unlock is a full barrier,
6595 * - switch_to() for arm64 (weakly-ordered, spin_unlock
6596 * is a RELEASE barrier),
Mathieu Desnoyers22e4ebb2017-07-28 16:40:40 -04006597 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006598 ++*switch_count;
6599
Peter Zijlstraaf449902020-09-17 10:38:30 +02006600 migrate_disable_switch(rq, prev);
Johannes Weinerb05e75d2020-03-16 15:13:31 -04006601 psi_sched_switch(prev, next, !task_on_rq_queued(prev));
6602
Delyan Kratunov9c2136b2022-05-11 18:28:36 +00006603 trace_sched_switch(sched_mode & SM_MASK_PREEMPT, prev, next, prev_state);
Ingo Molnard1ccc662017-02-01 11:46:42 +01006604
6605 /* Also unlocks the rq: */
6606 rq = context_switch(rq, prev, next, &rf);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02006607 } else {
Matt Flemingcb42c9a2016-09-21 14:38:13 +01006608 rq->clock_update_flags &= ~(RQCF_ACT_SKIP|RQCF_REQ_SKIP);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006609
Peter Zijlstra565790d22020-05-11 14:13:00 +02006610 rq_unpin_lock(rq, &rf);
6611 __balance_callbacks(rq);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05006612 raw_spin_rq_unlock_irq(rq);
Peter Zijlstra565790d22020-05-11 14:13:00 +02006613 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006614}
Thomas Gleixnerc259e012011-06-22 19:47:00 +02006615
Peter Zijlstra9af65282016-09-13 18:37:29 +02006616void __noreturn do_task_dead(void)
6617{
Ingo Molnard1ccc662017-02-01 11:46:42 +01006618 /* Causes final put_task_struct in finish_task_switch(): */
Peter Zijlstrab5bf9a92018-04-30 14:51:01 +02006619 set_special_state(TASK_DEAD);
Ingo Molnard1ccc662017-02-01 11:46:42 +01006620
6621 /* Tell freezer to ignore us: */
6622 current->flags |= PF_NOFREEZE;
6623
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006624 __schedule(SM_NONE);
Peter Zijlstra9af65282016-09-13 18:37:29 +02006625 BUG();
Ingo Molnard1ccc662017-02-01 11:46:42 +01006626
6627 /* Avoid "noreturn function does return" - but don't continue if BUG() is a NOP: */
Peter Zijlstra9af65282016-09-13 18:37:29 +02006628 for (;;)
Ingo Molnard1ccc662017-02-01 11:46:42 +01006629 cpu_relax();
Peter Zijlstra9af65282016-09-13 18:37:29 +02006630}
6631
Thomas Gleixner9c40cef22011-06-22 19:47:01 +02006632static inline void sched_submit_work(struct task_struct *tsk)
6633{
Sebastian Andrzej Siewiorc1cecf82020-08-19 22:00:25 +02006634 unsigned int task_flags;
6635
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02006636 if (task_is_running(tsk))
Thomas Gleixner9c40cef22011-06-22 19:47:01 +02006637 return;
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006638
Sebastian Andrzej Siewiorc1cecf82020-08-19 22:00:25 +02006639 task_flags = tsk->flags;
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006640 /*
Thomas Gleixnerb945efc2021-09-29 11:37:32 +02006641 * If a worker goes to sleep, notify and ask workqueue whether it
6642 * wants to wake up a task to maintain concurrency.
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006643 */
Sebastian Andrzej Siewiorc1cecf82020-08-19 22:00:25 +02006644 if (task_flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
Sebastian Andrzej Siewiorc1cecf82020-08-19 22:00:25 +02006645 if (task_flags & PF_WQ_WORKER)
Jens Axboe771b53d02019-10-22 10:25:58 -06006646 wq_worker_sleeping(tsk);
6647 else
6648 io_wq_worker_sleeping(tsk);
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006649 }
6650
John Keeping401e4962022-07-08 17:27:02 +01006651 /*
6652 * spinlock and rwlock must not flush block requests. This will
6653 * deadlock if the callback attempts to acquire a lock which is
6654 * already acquired.
6655 */
6656 SCHED_WARN_ON(current->__state & TASK_RTLOCK_WAIT);
Sebastian Andrzej Siewiorb0fdc012019-08-16 18:06:26 +02006657
Thomas Gleixner9c40cef22011-06-22 19:47:01 +02006658 /*
6659 * If we are going to sleep and we have plugged IO queued,
6660 * make sure to submit it to avoid deadlocks.
6661 */
Christoph Hellwigaa8dccc2022-01-27 08:05:49 +01006662 blk_flush_plug(tsk->plug, true);
Thomas Gleixner9c40cef22011-06-22 19:47:01 +02006663}
6664
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006665static void sched_update_worker(struct task_struct *tsk)
6666{
Jens Axboe771b53d02019-10-22 10:25:58 -06006667 if (tsk->flags & (PF_WQ_WORKER | PF_IO_WORKER)) {
6668 if (tsk->flags & PF_WQ_WORKER)
6669 wq_worker_running(tsk);
6670 else
6671 io_wq_worker_running(tsk);
6672 }
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006673}
6674
Andi Kleen722a9f92014-05-02 00:44:38 +02006675asmlinkage __visible void __sched schedule(void)
Thomas Gleixnerc259e012011-06-22 19:47:00 +02006676{
Thomas Gleixner9c40cef22011-06-22 19:47:01 +02006677 struct task_struct *tsk = current;
6678
6679 sched_submit_work(tsk);
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01006680 do {
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02006681 preempt_disable();
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006682 __schedule(SM_NONE);
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02006683 sched_preempt_enable_no_resched();
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01006684 } while (need_resched());
Thomas Gleixner6d25be52019-03-13 17:55:48 +01006685 sched_update_worker(tsk);
Thomas Gleixnerc259e012011-06-22 19:47:00 +02006686}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006687EXPORT_SYMBOL(schedule);
6688
Steven Rostedt (VMware)8663eff2017-04-14 08:48:09 -04006689/*
6690 * synchronize_rcu_tasks() makes sure that no task is stuck in preempted
6691 * state (have scheduled out non-voluntarily) by making sure that all
6692 * tasks have either left the run queue or have gone into user space.
6693 * As idle tasks do not do either, they must not ever be preempted
6694 * (schedule out non-voluntarily).
6695 *
6696 * schedule_idle() is similar to schedule_preempt_disable() except that it
6697 * never enables preemption because it does not call sched_submit_work().
6698 */
6699void __sched schedule_idle(void)
6700{
6701 /*
6702 * As this skips calling sched_submit_work(), which the idle task does
6703 * regardless because that function is a nop when the task is in a
6704 * TASK_RUNNING state, make sure this isn't used someplace that the
6705 * current task can be in any other state. Note, idle is always in the
6706 * TASK_RUNNING state.
6707 */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02006708 WARN_ON_ONCE(current->__state);
Steven Rostedt (VMware)8663eff2017-04-14 08:48:09 -04006709 do {
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006710 __schedule(SM_NONE);
Steven Rostedt (VMware)8663eff2017-04-14 08:48:09 -04006711 } while (need_resched());
6712}
6713
Frederic Weisbecker24a9c5412022-06-08 16:40:24 +02006714#if defined(CONFIG_CONTEXT_TRACKING_USER) && !defined(CONFIG_HAVE_CONTEXT_TRACKING_USER_OFFSTACK)
Andi Kleen722a9f92014-05-02 00:44:38 +02006715asmlinkage __visible void __sched schedule_user(void)
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02006716{
6717 /*
6718 * If we come here after a random call to set_need_resched(),
6719 * or we have been woken up remotely but the IPI has not yet arrived,
6720 * we haven't yet exited the RCU idle mode. Do it here manually until
6721 * we find a better solution.
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08006722 *
6723 * NB: There are buggy callers of this function. Ideally we
Frederic Weisbeckerc467ea72015-03-04 18:06:33 +01006724 * should warn if prev_state != CONTEXT_USER, but that will trigger
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08006725 * too frequently to make sense yet.
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02006726 */
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08006727 enum ctx_state prev_state = exception_enter();
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02006728 schedule();
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08006729 exception_exit(prev_state);
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02006730}
6731#endif
6732
Thomas Gleixnerc5491ea2011-03-21 12:09:35 +01006733/**
6734 * schedule_preempt_disabled - called with preemption disabled
6735 *
6736 * Returns with preemption disabled. Note: preempt_count must be 1
6737 */
6738void __sched schedule_preempt_disabled(void)
6739{
Thomas Gleixnerba74c142011-03-21 13:32:17 +01006740 sched_preempt_enable_no_resched();
Thomas Gleixnerc5491ea2011-03-21 12:09:35 +01006741 schedule();
6742 preempt_disable();
6743}
6744
Thomas Gleixner69914362021-08-15 23:27:48 +02006745#ifdef CONFIG_PREEMPT_RT
6746void __sched notrace schedule_rtlock(void)
6747{
6748 do {
6749 preempt_disable();
6750 __schedule(SM_RTLOCK_WAIT);
6751 sched_preempt_enable_no_resched();
6752 } while (need_resched());
6753}
6754NOKPROBE_SYMBOL(schedule_rtlock);
6755#endif
6756
Frederic Weisbecker06b1f802015-02-16 19:20:07 +01006757static void __sched notrace preempt_schedule_common(void)
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01006758{
6759 do {
Steven Rostedt47252cf2016-03-21 11:23:39 -04006760 /*
6761 * Because the function tracer can trace preempt_count_sub()
6762 * and it also uses preempt_enable/disable_notrace(), if
6763 * NEED_RESCHED is set, the preempt_enable_notrace() called
6764 * by the function tracer will call this function again and
6765 * cause infinite recursion.
6766 *
6767 * Preemption must be disabled here before the function
6768 * tracer can trace. Break up preempt_disable() into two
6769 * calls. One to disable preemption without fear of being
6770 * traced. The other to still record the preemption latency,
6771 * which can also be traced by the function tracer.
6772 */
Peter Zijlstra499d7952015-09-28 18:52:36 +02006773 preempt_disable_notrace();
Steven Rostedt47252cf2016-03-21 11:23:39 -04006774 preempt_latency_start(1);
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006775 __schedule(SM_PREEMPT);
Steven Rostedt47252cf2016-03-21 11:23:39 -04006776 preempt_latency_stop(1);
Peter Zijlstra499d7952015-09-28 18:52:36 +02006777 preempt_enable_no_resched_notrace();
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01006778
6779 /*
6780 * Check again in case we missed a preemption opportunity
6781 * between schedule and now.
6782 */
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01006783 } while (need_resched());
6784}
6785
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02006786#ifdef CONFIG_PREEMPTION
Linus Torvalds1da177e2005-04-16 15:20:36 -07006787/*
Valentin Schneidera49b4f42019-09-23 15:36:12 +01006788 * This is the entry point to schedule() from in-kernel preemption
6789 * off of preempt_enable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006790 */
Andi Kleen722a9f92014-05-02 00:44:38 +02006791asmlinkage __visible void __sched notrace preempt_schedule(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006792{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006793 /*
6794 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006795 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07006796 */
Frederic Weisbeckerfbb00b52013-06-19 23:56:22 +02006797 if (likely(!preemptible()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006798 return;
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01006799 preempt_schedule_common();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006800}
Masami Hiramatsu376e2422014-04-17 17:17:05 +09006801NOKPROBE_SYMBOL(preempt_schedule);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006802EXPORT_SYMBOL(preempt_schedule);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006803
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006804#ifdef CONFIG_PREEMPT_DYNAMIC
Mark Rutland99cf983c2022-02-14 16:52:14 +00006805#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
Mark Rutland8a69fe02022-02-14 16:52:11 +00006806#ifndef preempt_schedule_dynamic_enabled
6807#define preempt_schedule_dynamic_enabled preempt_schedule
6808#define preempt_schedule_dynamic_disabled NULL
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006809#endif
Mark Rutland8a69fe02022-02-14 16:52:11 +00006810DEFINE_STATIC_CALL(preempt_schedule, preempt_schedule_dynamic_enabled);
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006811EXPORT_STATIC_CALL_TRAMP(preempt_schedule);
Mark Rutland99cf983c2022-02-14 16:52:14 +00006812#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6813static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule);
6814void __sched notrace dynamic_preempt_schedule(void)
6815{
6816 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule))
6817 return;
6818 preempt_schedule();
6819}
6820NOKPROBE_SYMBOL(dynamic_preempt_schedule);
6821EXPORT_SYMBOL(dynamic_preempt_schedule);
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006822#endif
Mark Rutland99cf983c2022-02-14 16:52:14 +00006823#endif
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006824
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006825/**
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02006826 * preempt_schedule_notrace - preempt_schedule called by tracing
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006827 *
6828 * The tracing infrastructure uses preempt_enable_notrace to prevent
6829 * recursion and tracing preempt enabling caused by the tracing
6830 * infrastructure itself. But as tracing can happen in areas coming
6831 * from userspace or just about to enter userspace, a preempt enable
6832 * can occur before user_exit() is called. This will cause the scheduler
6833 * to be called when the system is still in usermode.
6834 *
6835 * To prevent this, the preempt_enable_notrace will use this function
6836 * instead of preempt_schedule() to exit user context if needed before
6837 * calling the scheduler.
6838 */
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02006839asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006840{
6841 enum ctx_state prev_ctx;
6842
6843 if (likely(!preemptible()))
6844 return;
6845
6846 do {
Steven Rostedt47252cf2016-03-21 11:23:39 -04006847 /*
6848 * Because the function tracer can trace preempt_count_sub()
6849 * and it also uses preempt_enable/disable_notrace(), if
6850 * NEED_RESCHED is set, the preempt_enable_notrace() called
6851 * by the function tracer will call this function again and
6852 * cause infinite recursion.
6853 *
6854 * Preemption must be disabled here before the function
6855 * tracer can trace. Break up preempt_disable() into two
6856 * calls. One to disable preemption without fear of being
6857 * traced. The other to still record the preemption latency,
6858 * which can also be traced by the function tracer.
6859 */
Peter Zijlstra3d8f74d2015-09-28 18:09:19 +02006860 preempt_disable_notrace();
Steven Rostedt47252cf2016-03-21 11:23:39 -04006861 preempt_latency_start(1);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006862 /*
6863 * Needs preempt disabled in case user_exit() is traced
6864 * and the tracer calls preempt_enable_notrace() causing
6865 * an infinite recursion.
6866 */
6867 prev_ctx = exception_enter();
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006868 __schedule(SM_PREEMPT);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006869 exception_exit(prev_ctx);
6870
Steven Rostedt47252cf2016-03-21 11:23:39 -04006871 preempt_latency_stop(1);
Peter Zijlstra3d8f74d2015-09-28 18:09:19 +02006872 preempt_enable_no_resched_notrace();
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006873 } while (need_resched());
6874}
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02006875EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02006876
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006877#ifdef CONFIG_PREEMPT_DYNAMIC
Mark Rutland99cf983c2022-02-14 16:52:14 +00006878#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
Mark Rutland8a69fe02022-02-14 16:52:11 +00006879#ifndef preempt_schedule_notrace_dynamic_enabled
6880#define preempt_schedule_notrace_dynamic_enabled preempt_schedule_notrace
6881#define preempt_schedule_notrace_dynamic_disabled NULL
6882#endif
6883DEFINE_STATIC_CALL(preempt_schedule_notrace, preempt_schedule_notrace_dynamic_enabled);
Peter Zijlstraef726612021-01-25 16:26:50 +01006884EXPORT_STATIC_CALL_TRAMP(preempt_schedule_notrace);
Mark Rutland99cf983c2022-02-14 16:52:14 +00006885#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
6886static DEFINE_STATIC_KEY_TRUE(sk_dynamic_preempt_schedule_notrace);
6887void __sched notrace dynamic_preempt_schedule_notrace(void)
6888{
6889 if (!static_branch_unlikely(&sk_dynamic_preempt_schedule_notrace))
6890 return;
6891 preempt_schedule_notrace();
6892}
6893NOKPROBE_SYMBOL(dynamic_preempt_schedule_notrace);
6894EXPORT_SYMBOL(dynamic_preempt_schedule_notrace);
6895#endif
Peter Zijlstra (Intel)2c9a98d2021-01-18 15:12:21 +01006896#endif
6897
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02006898#endif /* CONFIG_PREEMPTION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006899
6900/*
Valentin Schneidera49b4f42019-09-23 15:36:12 +01006901 * This is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07006902 * off of irq context.
6903 * Note, that this is called and return with irqs disabled. This will
6904 * protect us against recursive calling from irq.
6905 */
Andi Kleen722a9f92014-05-02 00:44:38 +02006906asmlinkage __visible void __sched preempt_schedule_irq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006907{
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01006908 enum ctx_state prev_state;
Ingo Molnar6478d882008-01-25 21:08:33 +01006909
Andreas Mohr2ed6e342006-07-10 04:43:52 -07006910 /* Catch callers which need to be fixed */
Peter Zijlstraf27dde82013-08-14 14:55:31 +02006911 BUG_ON(preempt_count() || !irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07006912
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01006913 prev_state = exception_enter();
6914
Andi Kleen3a5c3592007-10-15 17:00:14 +02006915 do {
Peter Zijlstra3d8f74d2015-09-28 18:09:19 +02006916 preempt_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02006917 local_irq_enable();
Thomas Gleixnerb4bfa3f2021-08-15 23:27:46 +02006918 __schedule(SM_PREEMPT);
Andi Kleen3a5c3592007-10-15 17:00:14 +02006919 local_irq_disable();
Peter Zijlstra3d8f74d2015-09-28 18:09:19 +02006920 sched_preempt_enable_no_resched();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08006921 } while (need_resched());
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01006922
6923 exception_exit(prev_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006924}
6925
Ingo Molnarac6424b2017-06-20 12:06:13 +02006926int default_wake_function(wait_queue_entry_t *curr, unsigned mode, int wake_flags,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006927 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006928{
Chris Wilson062d3f92020-07-23 21:10:42 +01006929 WARN_ON_ONCE(IS_ENABLED(CONFIG_SCHED_DEBUG) && wake_flags & ~WF_SYNC);
Peter Zijlstra63859d42009-09-15 19:14:42 +02006930 return try_to_wake_up(curr->private, mode, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006931}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006932EXPORT_SYMBOL(default_wake_function);
6933
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02006934static void __setscheduler_prio(struct task_struct *p, int prio)
6935{
6936 if (dl_prio(prio))
6937 p->sched_class = &dl_sched_class;
6938 else if (rt_prio(prio))
6939 p->sched_class = &rt_sched_class;
6940 else
6941 p->sched_class = &fair_sched_class;
6942
6943 p->prio = prio;
6944}
6945
Ingo Molnarb29739f2006-06-27 02:54:51 -07006946#ifdef CONFIG_RT_MUTEXES
6947
Peter Zijlstraacd58622017-03-23 15:56:11 +01006948static inline int __rt_effective_prio(struct task_struct *pi_task, int prio)
6949{
6950 if (pi_task)
6951 prio = min(prio, pi_task->prio);
6952
6953 return prio;
6954}
6955
6956static inline int rt_effective_prio(struct task_struct *p, int prio)
6957{
6958 struct task_struct *pi_task = rt_mutex_get_top_task(p);
6959
6960 return __rt_effective_prio(pi_task, prio);
6961}
6962
Ingo Molnarb29739f2006-06-27 02:54:51 -07006963/*
6964 * rt_mutex_setprio - set the current priority of a task
Peter Zijlstraacd58622017-03-23 15:56:11 +01006965 * @p: task to boost
6966 * @pi_task: donor task
Ingo Molnarb29739f2006-06-27 02:54:51 -07006967 *
6968 * This function changes the 'effective' priority of a task. It does
6969 * not touch ->normal_prio like __setscheduler().
6970 *
Thomas Gleixnerc365c292014-02-07 20:58:42 +01006971 * Used by the rt_mutex code to implement priority inheritance
6972 * logic. Call site only calls if the priority of the task changed.
Ingo Molnarb29739f2006-06-27 02:54:51 -07006973 */
Peter Zijlstraacd58622017-03-23 15:56:11 +01006974void rt_mutex_setprio(struct task_struct *p, struct task_struct *pi_task)
Ingo Molnarb29739f2006-06-27 02:54:51 -07006975{
Peter Zijlstraacd58622017-03-23 15:56:11 +01006976 int prio, oldprio, queued, running, queue_flag =
Peter Zijlstra7a57f322017-02-21 14:47:02 +01006977 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01006978 const struct sched_class *prev_class;
Peter Zijlstraeb580752015-07-31 21:28:18 +02006979 struct rq_flags rf;
6980 struct rq *rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006981
Peter Zijlstraacd58622017-03-23 15:56:11 +01006982 /* XXX used to be waiter->prio, not waiter->task->prio */
6983 prio = __rt_effective_prio(pi_task, p->normal_prio);
6984
6985 /*
6986 * If nothing changed; bail early.
6987 */
6988 if (p->pi_top_task == pi_task && prio == p->prio && !dl_prio(prio))
6989 return;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006990
Peter Zijlstraeb580752015-07-31 21:28:18 +02006991 rq = __task_rq_lock(p, &rf);
Peter Zijlstra80f5c1b2016-10-03 16:28:37 +02006992 update_rq_clock(rq);
Peter Zijlstraacd58622017-03-23 15:56:11 +01006993 /*
6994 * Set under pi_lock && rq->lock, such that the value can be used under
6995 * either lock.
6996 *
6997 * Note that there is loads of tricky to make this pointer cache work
6998 * right. rt_mutex_slowunlock()+rt_mutex_postunlock() work together to
6999 * ensure a task is de-boosted (pi_task is set to NULL) before the
7000 * task is allowed to run again (and can exit). This ensures the pointer
Tal Zussmanb19a8882020-11-12 19:51:56 -05007001 * points to a blocked task -- which guarantees the task is present.
Peter Zijlstraacd58622017-03-23 15:56:11 +01007002 */
7003 p->pi_top_task = pi_task;
7004
7005 /*
7006 * For FIFO/RR we only need to set prio, if that matches we're done.
7007 */
7008 if (prio == p->prio && !dl_prio(prio))
7009 goto out_unlock;
Ingo Molnarb29739f2006-06-27 02:54:51 -07007010
Thomas Gleixner1c4dd992011-06-06 20:07:38 +02007011 /*
7012 * Idle task boosting is a nono in general. There is one
7013 * exception, when PREEMPT_RT and NOHZ is active:
7014 *
7015 * The idle task calls get_next_timer_interrupt() and holds
7016 * the timer wheel base->lock on the CPU and another CPU wants
7017 * to access the timer (probably to cancel it). We can safely
7018 * ignore the boosting request, as the idle CPU runs this code
7019 * with interrupts disabled and will complete the lock
7020 * protected section without being interrupted. So there is no
7021 * real need to boost.
7022 */
7023 if (unlikely(p == rq->idle)) {
7024 WARN_ON(p != rq->curr);
7025 WARN_ON(p->pi_blocked_on);
7026 goto out_unlock;
7027 }
7028
Peter Zijlstrab91473f2017-03-23 15:56:12 +01007029 trace_sched_pi_setprio(p, pi_task);
Andrew Mortond5f9f942007-05-08 20:27:06 -07007030 oldprio = p->prio;
Peter Zijlstraff77e462016-01-18 15:27:07 +01007031
7032 if (oldprio == prio)
7033 queue_flag &= ~DEQUEUE_MOVE;
7034
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01007035 prev_class = p->sched_class;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007036 queued = task_on_rq_queued(p);
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01007037 running = task_current(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007038 if (queued)
Peter Zijlstraff77e462016-01-18 15:27:07 +01007039 dequeue_task(rq, p, queue_flag);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07007040 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04007041 put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02007042
Dario Faggioli2d3d8912013-11-07 14:43:44 +01007043 /*
7044 * Boosting condition are:
7045 * 1. -rt task is running and holds mutex A
7046 * --> -dl task blocks on mutex A
7047 *
7048 * 2. -dl task is running and holds mutex A
7049 * --> -dl task blocks on mutex A and could preempt the
7050 * running task
7051 */
7052 if (dl_prio(prio)) {
Oleg Nesterov466af292014-06-06 18:52:06 +02007053 if (!dl_prio(p->normal_prio) ||
Juri Lelli740797c2018-11-19 16:32:01 +01007054 (pi_task && dl_prio(pi_task->prio) &&
7055 dl_entity_preempt(&pi_task->dl, &p->dl))) {
Juri Lelli2279f542020-11-17 07:14:32 +01007056 p->dl.pi_se = pi_task->dl.pi_se;
Peter Zijlstraff77e462016-01-18 15:27:07 +01007057 queue_flag |= ENQUEUE_REPLENISH;
Juri Lelli2279f542020-11-17 07:14:32 +01007058 } else {
7059 p->dl.pi_se = &p->dl;
7060 }
Dario Faggioli2d3d8912013-11-07 14:43:44 +01007061 } else if (rt_prio(prio)) {
7062 if (dl_prio(oldprio))
Juri Lelli2279f542020-11-17 07:14:32 +01007063 p->dl.pi_se = &p->dl;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01007064 if (oldprio < prio)
Peter Zijlstraff77e462016-01-18 15:27:07 +01007065 queue_flag |= ENQUEUE_HEAD;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01007066 } else {
7067 if (dl_prio(oldprio))
Juri Lelli2279f542020-11-17 07:14:32 +01007068 p->dl.pi_se = &p->dl;
Brian Silverman746db942015-02-18 16:23:56 -08007069 if (rt_prio(oldprio))
7070 p->rt.timeout = 0;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01007071 }
Ingo Molnardd41f592007-07-09 18:51:59 +02007072
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02007073 __setscheduler_prio(p, prio);
Ingo Molnarb29739f2006-06-27 02:54:51 -07007074
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007075 if (queued)
Peter Zijlstraff77e462016-01-18 15:27:07 +01007076 enqueue_task(rq, p, queue_flag);
Vincent Guittota399d232016-09-12 09:47:52 +02007077 if (running)
Peter Zijlstra03b7fad2019-05-29 20:36:41 +00007078 set_next_task(rq, p);
Steven Rostedtcb469842008-01-25 21:08:22 +01007079
Peter Zijlstrada7a7352011-01-17 17:03:27 +01007080 check_class_changed(rq, p, prev_class, oldprio);
Thomas Gleixner1c4dd992011-06-06 20:07:38 +02007081out_unlock:
Ingo Molnard1ccc662017-02-01 11:46:42 +01007082 /* Avoid rq from going away on us: */
7083 preempt_disable();
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02007084
Peter Zijlstra565790d22020-05-11 14:13:00 +02007085 rq_unpin_lock(rq, &rf);
7086 __balance_callbacks(rq);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05007087 raw_spin_rq_unlock(rq);
Peter Zijlstra565790d22020-05-11 14:13:00 +02007088
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02007089 preempt_enable();
Ingo Molnarb29739f2006-06-27 02:54:51 -07007090}
Peter Zijlstraacd58622017-03-23 15:56:11 +01007091#else
7092static inline int rt_effective_prio(struct task_struct *p, int prio)
7093{
7094 return prio;
7095}
Ingo Molnarb29739f2006-06-27 02:54:51 -07007096#endif
Dario Faggiolid50dde52013-11-07 14:43:36 +01007097
Ingo Molnar36c8b582006-07-03 00:25:41 -07007098void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007099{
Peter Zijlstra49bd21e2016-09-20 22:06:01 +02007100 bool queued, running;
Qian Cai53a23362019-12-19 09:03:14 -05007101 int old_prio;
Peter Zijlstraeb580752015-07-31 21:28:18 +02007102 struct rq_flags rf;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007103 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007104
Dongsheng Yang75e45d52014-02-11 15:34:50 +08007105 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007106 return;
7107 /*
7108 * We have to be careful, if called from sys_setpriority(),
7109 * the task might be in the middle of scheduling on another CPU.
7110 */
Peter Zijlstraeb580752015-07-31 21:28:18 +02007111 rq = task_rq_lock(p, &rf);
Peter Zijlstra2fb8d362016-10-03 16:44:25 +02007112 update_rq_clock(rq);
7113
Linus Torvalds1da177e2005-04-16 15:20:36 -07007114 /*
7115 * The RT priorities are set via sched_setscheduler(), but we still
7116 * allow the 'normal' nice value to be set - but as expected
Tal Zussmanb19a8882020-11-12 19:51:56 -05007117 * it won't have any effect on scheduling until the task is
Dario Faggioliaab03e02013-11-28 11:14:43 +01007118 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007119 */
Dario Faggioliaab03e02013-11-28 11:14:43 +01007120 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007121 p->static_prio = NICE_TO_PRIO(nice);
7122 goto out_unlock;
7123 }
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007124 queued = task_on_rq_queued(p);
Peter Zijlstra49bd21e2016-09-20 22:06:01 +02007125 running = task_current(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007126 if (queued)
Peter Zijlstra7a57f322017-02-21 14:47:02 +01007127 dequeue_task(rq, p, DEQUEUE_SAVE | DEQUEUE_NOCLOCK);
Peter Zijlstra49bd21e2016-09-20 22:06:01 +02007128 if (running)
7129 put_prev_task(rq, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007130
Linus Torvalds1da177e2005-04-16 15:20:36 -07007131 p->static_prio = NICE_TO_PRIO(nice);
Peter Zijlstrab1e82062022-02-14 10:16:57 +01007132 set_load_weight(p, true);
Ingo Molnarb29739f2006-06-27 02:54:51 -07007133 old_prio = p->prio;
7134 p->prio = effective_prio(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007135
Frederic Weisbecker5443a0b2019-12-03 17:01:06 +01007136 if (queued)
Peter Zijlstra7134b3e2017-02-21 14:23:38 +01007137 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
Peter Zijlstra49bd21e2016-09-20 22:06:01 +02007138 if (running)
Peter Zijlstra03b7fad2019-05-29 20:36:41 +00007139 set_next_task(rq, p);
Frederic Weisbecker5443a0b2019-12-03 17:01:06 +01007140
7141 /*
7142 * If the task increased its priority or is running and
7143 * lowered its priority, then reschedule its CPU:
7144 */
7145 p->sched_class->prio_changed(rq, p, old_prio);
7146
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147out_unlock:
Peter Zijlstraeb580752015-07-31 21:28:18 +02007148 task_rq_unlock(rq, p, &rf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007149}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007150EXPORT_SYMBOL(set_user_nice);
7151
Matt Mackalle43379f2005-05-01 08:59:00 -07007152/*
Christian Göttsche700a7832022-06-15 17:25:04 +02007153 * is_nice_reduction - check if nice value is an actual reduction
7154 *
7155 * Similar to can_nice() but does not perform a capability check.
7156 *
7157 * @p: task
7158 * @nice: nice value
7159 */
7160static bool is_nice_reduction(const struct task_struct *p, const int nice)
7161{
7162 /* Convert nice value [19,-20] to rlimit style value [1,40]: */
7163 int nice_rlim = nice_to_rlimit(nice);
7164
7165 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE));
7166}
7167
7168/*
Matt Mackalle43379f2005-05-01 08:59:00 -07007169 * can_nice - check if a task can reduce its nice value
7170 * @p: task
7171 * @nice: nice value
7172 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007173int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07007174{
Christian Göttsche700a7832022-06-15 17:25:04 +02007175 return is_nice_reduction(p, nice) || capable(CAP_SYS_NICE);
Matt Mackalle43379f2005-05-01 08:59:00 -07007176}
7177
Linus Torvalds1da177e2005-04-16 15:20:36 -07007178#ifdef __ARCH_WANT_SYS_NICE
7179
7180/*
7181 * sys_nice - change the priority of the current process.
7182 * @increment: priority increment
7183 *
7184 * sys_setpriority is a more generic, but much slower function that
7185 * does similar things.
7186 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01007187SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007188{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007189 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007190
7191 /*
7192 * Setpriority might change our priority at the same moment.
7193 * We don't have to worry. Conceptually one call occurs first
7194 * and we have a single winner.
7195 */
Dongsheng Yanga9467fa2014-05-08 18:35:15 +09007196 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05007197 nice = task_nice(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007198
Dongsheng Yanga9467fa2014-05-08 18:35:15 +09007199 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
Matt Mackalle43379f2005-05-01 08:59:00 -07007200 if (increment < 0 && !can_nice(current, nice))
7201 return -EPERM;
7202
Linus Torvalds1da177e2005-04-16 15:20:36 -07007203 retval = security_task_setnice(current, nice);
7204 if (retval)
7205 return retval;
7206
7207 set_user_nice(current, nice);
7208 return 0;
7209}
7210
7211#endif
7212
7213/**
7214 * task_prio - return the priority value of a given task.
7215 * @p: the task in question.
7216 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02007217 * Return: The priority value as seen by users in /proc.
Dietmar Eggemannc541bb72021-01-28 14:10:40 +01007218 *
7219 * sched policy return value kernel prio user prio/nice
7220 *
7221 * normal, batch, idle [0 ... 39] [100 ... 139] 0/[-20 ... 19]
7222 * fifo, rr [-2 ... -100] [98 ... 0] [1 ... 99]
7223 * deadline -101 -1 0
Linus Torvalds1da177e2005-04-16 15:20:36 -07007224 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007225int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007226{
7227 return p->prio - MAX_RT_PRIO;
7228}
7229
7230/**
Ingo Molnard1ccc662017-02-01 11:46:42 +01007231 * idle_cpu - is a given CPU idle currently?
Linus Torvalds1da177e2005-04-16 15:20:36 -07007232 * @cpu: the processor in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007233 *
7234 * Return: 1 if the CPU is currently idle. 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007235 */
7236int idle_cpu(int cpu)
7237{
Thomas Gleixner908a3282011-09-15 15:32:06 +02007238 struct rq *rq = cpu_rq(cpu);
7239
7240 if (rq->curr != rq->idle)
7241 return 0;
7242
7243 if (rq->nr_running)
7244 return 0;
7245
7246#ifdef CONFIG_SMP
Peter Zijlstra126c2092020-05-26 18:11:03 +02007247 if (rq->ttwu_pending)
Thomas Gleixner908a3282011-09-15 15:32:06 +02007248 return 0;
7249#endif
7250
7251 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007252}
7253
Linus Torvalds1da177e2005-04-16 15:20:36 -07007254/**
Rohit Jain943d3552018-05-09 09:39:48 -07007255 * available_idle_cpu - is a given CPU idle for enqueuing work.
7256 * @cpu: the CPU in question.
7257 *
7258 * Return: 1 if the CPU is currently idle. 0 otherwise.
7259 */
7260int available_idle_cpu(int cpu)
7261{
7262 if (!idle_cpu(cpu))
7263 return 0;
7264
Rohit Jain247f2f62018-05-02 13:52:10 -07007265 if (vcpu_is_preempted(cpu))
7266 return 0;
7267
Linus Torvalds1da177e2005-04-16 15:20:36 -07007268 return 1;
7269}
7270
7271/**
Ingo Molnard1ccc662017-02-01 11:46:42 +01007272 * idle_task - return the idle task for a given CPU.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007273 * @cpu: the processor in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007274 *
Ingo Molnard1ccc662017-02-01 11:46:42 +01007275 * Return: The idle task for the CPU @cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007276 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007277struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007278{
7279 return cpu_rq(cpu)->idle;
7280}
7281
Viresh Kumar7d6a905f2020-12-08 09:46:55 +05307282#ifdef CONFIG_SMP
7283/*
7284 * This function computes an effective utilization for the given CPU, to be
7285 * used for frequency selection given the linear relation: f = u * f_max.
7286 *
7287 * The scheduler tracks the following metrics:
7288 *
7289 * cpu_util_{cfs,rt,dl,irq}()
7290 * cpu_bw_dl()
7291 *
7292 * Where the cfs,rt and dl util numbers are tracked with the same metric and
7293 * synchronized windows and are thus directly comparable.
7294 *
7295 * The cfs,rt,dl utilization are the running times measured with rq->clock_task
7296 * which excludes things like IRQ and steal-time. These latter are then accrued
7297 * in the irq utilization.
7298 *
7299 * The DL bandwidth number otoh is not a measured metric but a value computed
7300 * based on the task model parameters and gives the minimal utilization
7301 * required to meet deadlines.
7302 */
Viresh Kumara5418be2020-12-08 09:46:56 +05307303unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
Dietmar Eggemannbb447992022-06-21 10:04:10 +01007304 enum cpu_util_type type,
Viresh Kumar7d6a905f2020-12-08 09:46:55 +05307305 struct task_struct *p)
7306{
Dietmar Eggemannbb447992022-06-21 10:04:10 +01007307 unsigned long dl_util, util, irq, max;
Viresh Kumar7d6a905f2020-12-08 09:46:55 +05307308 struct rq *rq = cpu_rq(cpu);
7309
Dietmar Eggemannbb447992022-06-21 10:04:10 +01007310 max = arch_scale_cpu_capacity(cpu);
7311
Viresh Kumar7d6a905f2020-12-08 09:46:55 +05307312 if (!uclamp_is_used() &&
7313 type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
7314 return max;
7315 }
7316
7317 /*
7318 * Early check to see if IRQ/steal time saturates the CPU, can be
7319 * because of inaccuracies in how we track these -- see
7320 * update_irq_load_avg().
7321 */
7322 irq = cpu_util_irq(rq);
7323 if (unlikely(irq >= max))
7324 return max;
7325
7326 /*
7327 * Because the time spend on RT/DL tasks is visible as 'lost' time to
7328 * CFS tasks and we use the same metric to track the effective
7329 * utilization (PELT windows are synchronized) we can directly add them
7330 * to obtain the CPU's actual utilization.
7331 *
7332 * CFS and RT utilization can be boosted or capped, depending on
7333 * utilization clamp constraints requested by currently RUNNABLE
7334 * tasks.
7335 * When there are no CFS RUNNABLE tasks, clamps are released and
7336 * frequency will be gracefully reduced with the utilization decay.
7337 */
7338 util = util_cfs + cpu_util_rt(rq);
7339 if (type == FREQUENCY_UTIL)
7340 util = uclamp_rq_util_with(rq, util, p);
7341
7342 dl_util = cpu_util_dl(rq);
7343
7344 /*
7345 * For frequency selection we do not make cpu_util_dl() a permanent part
7346 * of this sum because we want to use cpu_bw_dl() later on, but we need
7347 * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
7348 * that we select f_max when there is no idle time.
7349 *
7350 * NOTE: numerical errors or stop class might cause us to not quite hit
7351 * saturation when we should -- something for later.
7352 */
7353 if (util + dl_util >= max)
7354 return max;
7355
7356 /*
7357 * OTOH, for energy computation we need the estimated running time, so
7358 * include util_dl and ignore dl_bw.
7359 */
7360 if (type == ENERGY_UTIL)
7361 util += dl_util;
7362
7363 /*
7364 * There is still idle time; further improve the number by using the
7365 * irq metric. Because IRQ/steal time is hidden from the task clock we
7366 * need to scale the task numbers:
7367 *
7368 * max - irq
7369 * U' = irq + --------- * U
7370 * max
7371 */
7372 util = scale_irq_capacity(util, irq, max);
7373 util += irq;
7374
7375 /*
7376 * Bandwidth required by DEADLINE must always be granted while, for
7377 * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
7378 * to gracefully reduce the frequency when no tasks show up for longer
7379 * periods of time.
7380 *
7381 * Ideally we would like to set bw_dl as min/guaranteed freq and util +
7382 * bw_dl as requested freq. However, cpufreq is not yet ready for such
7383 * an interface. So, we only do the latter for now.
7384 */
7385 if (type == FREQUENCY_UTIL)
7386 util += cpu_bw_dl(rq);
7387
7388 return min(max, util);
7389}
Viresh Kumara5418be2020-12-08 09:46:56 +05307390
Dietmar Eggemannbb447992022-06-21 10:04:10 +01007391unsigned long sched_cpu_util(int cpu)
Viresh Kumara5418be2020-12-08 09:46:56 +05307392{
Dietmar Eggemannbb447992022-06-21 10:04:10 +01007393 return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
Viresh Kumara5418be2020-12-08 09:46:56 +05307394}
Viresh Kumar7d6a905f2020-12-08 09:46:55 +05307395#endif /* CONFIG_SMP */
7396
Linus Torvalds1da177e2005-04-16 15:20:36 -07007397/**
7398 * find_process_by_pid - find a process with a matching PID value.
7399 * @pid: the pid in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007400 *
7401 * The task of @pid, if found. %NULL otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007402 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02007403static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007404{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07007405 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007406}
7407
Dario Faggioliaab03e02013-11-28 11:14:43 +01007408/*
Steven Rostedtc13db6b2014-07-23 11:28:26 -04007409 * sched_setparam() passes in -1 for its policy, to let the functions
7410 * it calls know not to change it.
7411 */
7412#define SETPARAM_POLICY -1
7413
Thomas Gleixnerc365c292014-02-07 20:58:42 +01007414static void __setscheduler_params(struct task_struct *p,
7415 const struct sched_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007416{
Dario Faggiolid50dde52013-11-07 14:43:36 +01007417 int policy = attr->sched_policy;
7418
Steven Rostedtc13db6b2014-07-23 11:28:26 -04007419 if (policy == SETPARAM_POLICY)
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01007420 policy = p->policy;
7421
Linus Torvalds1da177e2005-04-16 15:20:36 -07007422 p->policy = policy;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007423
Dario Faggioliaab03e02013-11-28 11:14:43 +01007424 if (dl_policy(policy))
7425 __setparam_dl(p, attr);
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01007426 else if (fair_policy(policy))
Dario Faggiolid50dde52013-11-07 14:43:36 +01007427 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
7428
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01007429 /*
7430 * __sched_setscheduler() ensures attr->sched_priority == 0 when
7431 * !rt_policy. Always setting this ensures that things like
7432 * getparam()/getattr() don't report silly values for !rt tasks.
7433 */
7434 p->rt_priority = attr->sched_priority;
Steven Rostedt383afd02014-03-11 19:24:20 -04007435 p->normal_prio = normal_prio(p);
Peter Zijlstrab1e82062022-02-14 10:16:57 +01007436 set_load_weight(p, true);
Thomas Gleixnerc365c292014-02-07 20:58:42 +01007437}
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01007438
David Howellsc69e8d92008-11-14 10:39:19 +11007439/*
Ingo Molnard1ccc662017-02-01 11:46:42 +01007440 * Check the target process has a UID that matches the current process's:
David Howellsc69e8d92008-11-14 10:39:19 +11007441 */
7442static bool check_same_owner(struct task_struct *p)
7443{
7444 const struct cred *cred = current_cred(), *pcred;
7445 bool match;
7446
7447 rcu_read_lock();
7448 pcred = __task_cred(p);
Eric W. Biederman9c806aa2012-02-02 18:54:02 -08007449 match = (uid_eq(cred->euid, pcred->euid) ||
7450 uid_eq(cred->euid, pcred->uid));
David Howellsc69e8d92008-11-14 10:39:19 +11007451 rcu_read_unlock();
7452 return match;
7453}
7454
Christian Göttsche700a7832022-06-15 17:25:04 +02007455/*
7456 * Allow unprivileged RT tasks to decrease priority.
7457 * Only issue a capable test if needed and only once to avoid an audit
7458 * event on permitted non-privileged operations:
7459 */
7460static int user_check_sched_setscheduler(struct task_struct *p,
7461 const struct sched_attr *attr,
7462 int policy, int reset_on_fork)
7463{
7464 if (fair_policy(policy)) {
7465 if (attr->sched_nice < task_nice(p) &&
7466 !is_nice_reduction(p, attr->sched_nice))
7467 goto req_priv;
7468 }
7469
7470 if (rt_policy(policy)) {
7471 unsigned long rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
7472
7473 /* Can't set/change the rt policy: */
7474 if (policy != p->policy && !rlim_rtprio)
7475 goto req_priv;
7476
7477 /* Can't increase priority: */
7478 if (attr->sched_priority > p->rt_priority &&
7479 attr->sched_priority > rlim_rtprio)
7480 goto req_priv;
7481 }
7482
7483 /*
7484 * Can't set/change SCHED_DEADLINE policy at all for now
7485 * (safest behavior); in the future we would like to allow
7486 * unprivileged DL tasks to increase their relative deadline
7487 * or reduce their runtime (both ways reducing utilization)
7488 */
7489 if (dl_policy(policy))
7490 goto req_priv;
7491
7492 /*
7493 * Treat SCHED_IDLE as nice 20. Only allow a switch to
7494 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
7495 */
7496 if (task_has_idle_policy(p) && !idle_policy(policy)) {
7497 if (!is_nice_reduction(p, task_nice(p)))
7498 goto req_priv;
7499 }
7500
7501 /* Can't change other user's priorities: */
7502 if (!check_same_owner(p))
7503 goto req_priv;
7504
7505 /* Normal users shall not reset the sched_reset_on_fork flag: */
7506 if (p->sched_reset_on_fork && !reset_on_fork)
7507 goto req_priv;
7508
7509 return 0;
7510
7511req_priv:
7512 if (!capable(CAP_SYS_NICE))
7513 return -EPERM;
7514
7515 return 0;
7516}
7517
Dario Faggiolid50dde52013-11-07 14:43:36 +01007518static int __sched_setscheduler(struct task_struct *p,
7519 const struct sched_attr *attr,
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007520 bool user, bool pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007521{
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02007522 int oldpolicy = -1, policy = attr->sched_policy;
7523 int retval, oldprio, newprio, queued, running;
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01007524 const struct sched_class *prev_class;
Kees Cook8e5bad72022-10-07 17:07:58 -07007525 struct balance_callback *head;
Peter Zijlstraeb580752015-07-31 21:28:18 +02007526 struct rq_flags rf;
Lennart Poetteringca94c442009-06-15 17:17:47 +02007527 int reset_on_fork;
Peter Zijlstra7a57f322017-02-21 14:47:02 +01007528 int queue_flags = DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
Peter Zijlstraeb580752015-07-31 21:28:18 +02007529 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007530
Steven Rostedt (VMware)896bbb22017-03-09 10:18:42 -05007531 /* The pi code expects interrupts enabled */
7532 BUG_ON(pi && in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07007533recheck:
Ingo Molnard1ccc662017-02-01 11:46:42 +01007534 /* Double check policy once rq lock held: */
Lennart Poetteringca94c442009-06-15 17:17:47 +02007535 if (policy < 0) {
7536 reset_on_fork = p->sched_reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007537 policy = oldpolicy = p->policy;
Lennart Poetteringca94c442009-06-15 17:17:47 +02007538 } else {
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007539 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
Lennart Poetteringca94c442009-06-15 17:17:47 +02007540
Henrik Austad20f9cd22015-09-09 17:00:41 +02007541 if (!valid_policy(policy))
Lennart Poetteringca94c442009-06-15 17:17:47 +02007542 return -EINVAL;
7543 }
7544
Juri Lelli794a56e2017-12-04 11:23:20 +01007545 if (attr->sched_flags & ~(SCHED_FLAG_ALL | SCHED_FLAG_SUGOV))
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007546 return -EINVAL;
7547
Linus Torvalds1da177e2005-04-16 15:20:36 -07007548 /*
7549 * Valid priorities for SCHED_FIFO and SCHED_RR are
Dietmar Eggemannae18ad22021-01-28 14:10:38 +01007550 * 1..MAX_RT_PRIO-1, valid priority for SCHED_NORMAL,
Ingo Molnardd41f592007-07-09 18:51:59 +02007551 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007552 */
Dietmar Eggemannae18ad22021-01-28 14:10:38 +01007553 if (attr->sched_priority > MAX_RT_PRIO-1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007554 return -EINVAL;
Dario Faggioliaab03e02013-11-28 11:14:43 +01007555 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
7556 (rt_policy(policy) != (attr->sched_priority != 0)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007557 return -EINVAL;
7558
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07007559 if (user) {
Christian Göttsche700a7832022-06-15 17:25:04 +02007560 retval = user_check_sched_setscheduler(p, attr, policy, reset_on_fork);
7561 if (retval)
7562 return retval;
7563
Juri Lelli794a56e2017-12-04 11:23:20 +01007564 if (attr->sched_flags & SCHED_FLAG_SUGOV)
7565 return -EINVAL;
7566
KOSAKI Motohirob0ae1982010-10-15 04:21:18 +09007567 retval = security_task_setscheduler(p);
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07007568 if (retval)
7569 return retval;
7570 }
7571
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007572 /* Update task specific "requested" clamps */
7573 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) {
7574 retval = uclamp_validate(p, attr);
7575 if (retval)
7576 return retval;
7577 }
7578
Juri Lelli710da3c2019-07-19 16:00:00 +02007579 if (pi)
7580 cpuset_read_lock();
7581
Linus Torvalds1da177e2005-04-16 15:20:36 -07007582 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01007583 * Make sure no PI-waiters arrive (or leave) while we are
Ingo Molnarb29739f2006-06-27 02:54:51 -07007584 * changing the priority of the task:
Peter Zijlstra0122ec52011-04-05 17:23:51 +02007585 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03007586 * To be able to change p->policy safely, the appropriate
Linus Torvalds1da177e2005-04-16 15:20:36 -07007587 * runqueue lock must be held.
7588 */
Peter Zijlstraeb580752015-07-31 21:28:18 +02007589 rq = task_rq_lock(p, &rf);
Peter Zijlstra80f5c1b2016-10-03 16:28:37 +02007590 update_rq_clock(rq);
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007591
Peter Zijlstra34f971f2010-09-22 13:53:15 +02007592 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01007593 * Changing the policy of the stop threads its a very bad idea:
Peter Zijlstra34f971f2010-09-22 13:53:15 +02007594 */
7595 if (p == rq->stop) {
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007596 retval = -EINVAL;
7597 goto unlock;
Peter Zijlstra34f971f2010-09-22 13:53:15 +02007598 }
7599
Dario Faggiolia51e9192011-03-24 14:00:18 +01007600 /*
Thomas Gleixnerd6b1e912014-02-07 20:58:40 +01007601 * If not changing anything there's no need to proceed further,
7602 * but store a possible modification of reset_on_fork.
Dario Faggiolia51e9192011-03-24 14:00:18 +01007603 */
Dario Faggiolid50dde52013-11-07 14:43:36 +01007604 if (unlikely(policy == p->policy)) {
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05007605 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
Dario Faggiolid50dde52013-11-07 14:43:36 +01007606 goto change;
7607 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
7608 goto change;
Wanpeng Li75381602014-11-26 08:44:04 +08007609 if (dl_policy(policy) && dl_param_changed(p, attr))
Dario Faggioliaab03e02013-11-28 11:14:43 +01007610 goto change;
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007611 if (attr->sched_flags & SCHED_FLAG_UTIL_CLAMP)
7612 goto change;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007613
Thomas Gleixnerd6b1e912014-02-07 20:58:40 +01007614 p->sched_reset_on_fork = reset_on_fork;
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007615 retval = 0;
7616 goto unlock;
Dario Faggiolia51e9192011-03-24 14:00:18 +01007617 }
Dario Faggiolid50dde52013-11-07 14:43:36 +01007618change:
Dario Faggiolia51e9192011-03-24 14:00:18 +01007619
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007620 if (user) {
Dario Faggioli332ac172013-11-07 14:43:45 +01007621#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007622 /*
7623 * Do not allow realtime tasks into groups that have no runtime
7624 * assigned.
7625 */
7626 if (rt_bandwidth_enabled() && rt_policy(policy) &&
Mike Galbraithf4493772011-01-13 04:54:50 +01007627 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
7628 !task_group_is_autogroup(task_group(p))) {
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007629 retval = -EPERM;
7630 goto unlock;
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007631 }
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007632#endif
Dario Faggioli332ac172013-11-07 14:43:45 +01007633#ifdef CONFIG_SMP
Juri Lelli794a56e2017-12-04 11:23:20 +01007634 if (dl_bandwidth_enabled() && dl_policy(policy) &&
7635 !(attr->sched_flags & SCHED_FLAG_SUGOV)) {
Dario Faggioli332ac172013-11-07 14:43:45 +01007636 cpumask_t *span = rq->rd->span;
Dario Faggioli332ac172013-11-07 14:43:45 +01007637
7638 /*
7639 * Don't allow tasks with an affinity mask smaller than
7640 * the entire root_domain to become SCHED_DEADLINE. We
7641 * will also fail if there's no bandwidth available.
7642 */
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02007643 if (!cpumask_subset(span, p->cpus_ptr) ||
Peter Zijlstrae4099a52013-12-17 10:03:34 +01007644 rq->rd->dl_bw.bw == 0) {
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007645 retval = -EPERM;
7646 goto unlock;
Dario Faggioli332ac172013-11-07 14:43:45 +01007647 }
7648 }
7649#endif
7650 }
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02007651
Ingo Molnard1ccc662017-02-01 11:46:42 +01007652 /* Re-check policy now with rq lock held: */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007653 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
7654 policy = oldpolicy = -1;
Peter Zijlstraeb580752015-07-31 21:28:18 +02007655 task_rq_unlock(rq, p, &rf);
Juri Lelli710da3c2019-07-19 16:00:00 +02007656 if (pi)
7657 cpuset_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007658 goto recheck;
7659 }
Dario Faggioli332ac172013-11-07 14:43:45 +01007660
7661 /*
7662 * If setscheduling to SCHED_DEADLINE (or changing the parameters
7663 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
7664 * is available.
7665 */
Nicolas Pitre06a76fe2017-06-21 14:22:01 -04007666 if ((dl_policy(policy) || dl_task(p)) && sched_dl_overflow(p, policy, attr)) {
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007667 retval = -EBUSY;
7668 goto unlock;
Dario Faggioli332ac172013-11-07 14:43:45 +01007669 }
7670
Thomas Gleixnerc365c292014-02-07 20:58:42 +01007671 p->sched_reset_on_fork = reset_on_fork;
7672 oldprio = p->prio;
7673
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02007674 newprio = __normal_prio(policy, attr->sched_priority, attr->sched_nice);
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007675 if (pi) {
7676 /*
7677 * Take priority boosted tasks into account. If the new
7678 * effective priority is unchanged, we just store the new
7679 * normal parameters and do not touch the scheduler class and
7680 * the runqueue. This will be done when the task deboost
7681 * itself.
7682 */
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02007683 newprio = rt_effective_prio(p, newprio);
7684 if (newprio == oldprio)
Peter Zijlstraff77e462016-01-18 15:27:07 +01007685 queue_flags &= ~DEQUEUE_MOVE;
Thomas Gleixnerc365c292014-02-07 20:58:42 +01007686 }
7687
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007688 queued = task_on_rq_queued(p);
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01007689 running = task_current(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007690 if (queued)
Peter Zijlstraff77e462016-01-18 15:27:07 +01007691 dequeue_task(rq, p, queue_flags);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07007692 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04007693 put_prev_task(rq, p);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02007694
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01007695 prev_class = p->sched_class;
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007696
Peter Zijlstraf558c2b2021-08-03 12:45:01 +02007697 if (!(attr->sched_flags & SCHED_FLAG_KEEP_PARAMS)) {
7698 __setscheduler_params(p, attr);
7699 __setscheduler_prio(p, newprio);
7700 }
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007701 __setscheduler_uclamp(p, attr);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02007702
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007703 if (queued) {
Thomas Gleixner81a44c52014-02-07 20:58:41 +01007704 /*
7705 * We enqueue to tail when the priority of a task is
7706 * increased (user space view).
7707 */
Peter Zijlstraff77e462016-01-18 15:27:07 +01007708 if (oldprio < p->prio)
7709 queue_flags |= ENQUEUE_HEAD;
Peter Zijlstra1de64442015-09-30 17:44:13 +02007710
Peter Zijlstraff77e462016-01-18 15:27:07 +01007711 enqueue_task(rq, p, queue_flags);
Thomas Gleixner81a44c52014-02-07 20:58:41 +01007712 }
Vincent Guittota399d232016-09-12 09:47:52 +02007713 if (running)
Peter Zijlstra03b7fad2019-05-29 20:36:41 +00007714 set_next_task(rq, p);
Steven Rostedtcb469842008-01-25 21:08:22 +01007715
Peter Zijlstrada7a7352011-01-17 17:03:27 +01007716 check_class_changed(rq, p, prev_class, oldprio);
Ingo Molnard1ccc662017-02-01 11:46:42 +01007717
7718 /* Avoid rq from going away on us: */
7719 preempt_disable();
Peter Zijlstra565790d22020-05-11 14:13:00 +02007720 head = splice_balance_callbacks(rq);
Peter Zijlstraeb580752015-07-31 21:28:18 +02007721 task_rq_unlock(rq, p, &rf);
Ingo Molnarb29739f2006-06-27 02:54:51 -07007722
Juri Lelli710da3c2019-07-19 16:00:00 +02007723 if (pi) {
7724 cpuset_read_unlock();
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007725 rt_mutex_adjust_pi(p);
Juri Lelli710da3c2019-07-19 16:00:00 +02007726 }
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07007727
Ingo Molnard1ccc662017-02-01 11:46:42 +01007728 /* Run balance callbacks after we've adjusted the PI chain: */
Peter Zijlstra565790d22020-05-11 14:13:00 +02007729 balance_callbacks(rq, head);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02007730 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007731
7732 return 0;
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007733
7734unlock:
7735 task_rq_unlock(rq, p, &rf);
Juri Lelli710da3c2019-07-19 16:00:00 +02007736 if (pi)
7737 cpuset_read_unlock();
Mathieu Poirier4b211f22019-07-19 15:59:54 +02007738 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007739}
Rusty Russell961ccdd2008-06-23 13:55:38 +10007740
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007741static int _sched_setscheduler(struct task_struct *p, int policy,
7742 const struct sched_param *param, bool check)
7743{
7744 struct sched_attr attr = {
7745 .sched_policy = policy,
7746 .sched_priority = param->sched_priority,
7747 .sched_nice = PRIO_TO_NICE(p->static_prio),
7748 };
7749
Steven Rostedtc13db6b2014-07-23 11:28:26 -04007750 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
7751 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007752 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
7753 policy &= ~SCHED_RESET_ON_FORK;
7754 attr.sched_policy = policy;
7755 }
7756
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007757 return __sched_setscheduler(p, &attr, check, true);
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007758}
Rusty Russell961ccdd2008-06-23 13:55:38 +10007759/**
7760 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
7761 * @p: the task in question.
7762 * @policy: new policy.
7763 * @param: structure containing the new RT priority.
7764 *
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007765 * Use sched_set_fifo(), read its comment.
7766 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02007767 * Return: 0 on success. An error code otherwise.
7768 *
Rusty Russell961ccdd2008-06-23 13:55:38 +10007769 * NOTE that the task may be already dead.
7770 */
7771int sched_setscheduler(struct task_struct *p, int policy,
KOSAKI Motohirofe7de492010-10-20 16:01:12 -07007772 const struct sched_param *param)
Rusty Russell961ccdd2008-06-23 13:55:38 +10007773{
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007774 return _sched_setscheduler(p, policy, param, true);
Rusty Russell961ccdd2008-06-23 13:55:38 +10007775}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007776
Dario Faggiolid50dde52013-11-07 14:43:36 +01007777int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
7778{
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007779 return __sched_setscheduler(p, attr, true, true);
Dario Faggiolid50dde52013-11-07 14:43:36 +01007780}
Dario Faggiolid50dde52013-11-07 14:43:36 +01007781
Juri Lelli794a56e2017-12-04 11:23:20 +01007782int sched_setattr_nocheck(struct task_struct *p, const struct sched_attr *attr)
7783{
7784 return __sched_setscheduler(p, attr, false, true);
7785}
Viresh Kumar1eb5dde2020-06-23 15:49:40 +05307786EXPORT_SYMBOL_GPL(sched_setattr_nocheck);
Juri Lelli794a56e2017-12-04 11:23:20 +01007787
Rusty Russell961ccdd2008-06-23 13:55:38 +10007788/**
7789 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
7790 * @p: the task in question.
7791 * @policy: new policy.
7792 * @param: structure containing the new RT priority.
7793 *
7794 * Just like sched_setscheduler, only don't bother checking if the
7795 * current context has permission. For example, this is needed in
7796 * stop_machine(): we create temporary high priority worker threads,
7797 * but our caller might not have that capability.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007798 *
7799 * Return: 0 on success. An error code otherwise.
Rusty Russell961ccdd2008-06-23 13:55:38 +10007800 */
7801int sched_setscheduler_nocheck(struct task_struct *p, int policy,
KOSAKI Motohirofe7de492010-10-20 16:01:12 -07007802 const struct sched_param *param)
Rusty Russell961ccdd2008-06-23 13:55:38 +10007803{
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01007804 return _sched_setscheduler(p, policy, param, false);
Rusty Russell961ccdd2008-06-23 13:55:38 +10007805}
7806
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007807/*
7808 * SCHED_FIFO is a broken scheduler model; that is, it is fundamentally
7809 * incapable of resource management, which is the one thing an OS really should
7810 * be doing.
7811 *
7812 * This is of course the reason it is limited to privileged users only.
7813 *
7814 * Worse still; it is fundamentally impossible to compose static priority
7815 * workloads. You cannot take two correctly working static prio workloads
7816 * and smash them together and still expect them to work.
7817 *
7818 * For this reason 'all' FIFO tasks the kernel creates are basically at:
7819 *
7820 * MAX_RT_PRIO / 2
7821 *
7822 * The administrator _MUST_ configure the system, the kernel simply doesn't
7823 * know enough information to make a sensible choice.
7824 */
Peter Zijlstra8b700982020-04-22 13:10:04 +02007825void sched_set_fifo(struct task_struct *p)
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007826{
7827 struct sched_param sp = { .sched_priority = MAX_RT_PRIO / 2 };
Peter Zijlstra8b700982020-04-22 13:10:04 +02007828 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007829}
7830EXPORT_SYMBOL_GPL(sched_set_fifo);
7831
7832/*
7833 * For when you don't much care about FIFO, but want to be above SCHED_NORMAL.
7834 */
Peter Zijlstra8b700982020-04-22 13:10:04 +02007835void sched_set_fifo_low(struct task_struct *p)
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007836{
7837 struct sched_param sp = { .sched_priority = 1 };
Peter Zijlstra8b700982020-04-22 13:10:04 +02007838 WARN_ON_ONCE(sched_setscheduler_nocheck(p, SCHED_FIFO, &sp) != 0);
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007839}
7840EXPORT_SYMBOL_GPL(sched_set_fifo_low);
7841
Peter Zijlstra8b700982020-04-22 13:10:04 +02007842void sched_set_normal(struct task_struct *p, int nice)
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007843{
7844 struct sched_attr attr = {
7845 .sched_policy = SCHED_NORMAL,
7846 .sched_nice = nice,
7847 };
Peter Zijlstra8b700982020-04-22 13:10:04 +02007848 WARN_ON_ONCE(sched_setattr_nocheck(p, &attr) != 0);
Peter Zijlstra7318d4c2020-04-21 12:09:13 +02007849}
7850EXPORT_SYMBOL_GPL(sched_set_normal);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007851
7852static int
7853do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
7854{
7855 struct sched_param lparam;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007856 struct task_struct *p;
7857 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007858
Jason Baronc21761f2006-01-18 17:43:03 -08007859 if (!param || pid < 0)
7860 return -EINVAL;
7861 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
7862 return -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007863
7864 rcu_read_lock();
7865 retval = -ESRCH;
7866 p = find_process_by_pid(pid);
Juri Lelli710da3c2019-07-19 16:00:00 +02007867 if (likely(p))
7868 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007869 rcu_read_unlock();
7870
Juri Lelli710da3c2019-07-19 16:00:00 +02007871 if (likely(p)) {
7872 retval = sched_setscheduler(p, policy, &lparam);
7873 put_task_struct(p);
7874 }
7875
Linus Torvalds1da177e2005-04-16 15:20:36 -07007876 return retval;
7877}
7878
Dario Faggiolid50dde52013-11-07 14:43:36 +01007879/*
7880 * Mimics kernel/events/core.c perf_copy_attr().
7881 */
Ingo Molnard1ccc662017-02-01 11:46:42 +01007882static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *attr)
Dario Faggiolid50dde52013-11-07 14:43:36 +01007883{
7884 u32 size;
7885 int ret;
7886
Ingo Molnard1ccc662017-02-01 11:46:42 +01007887 /* Zero the full structure, so that a short copy will be nice: */
Dario Faggiolid50dde52013-11-07 14:43:36 +01007888 memset(attr, 0, sizeof(*attr));
7889
7890 ret = get_user(size, &uattr->size);
7891 if (ret)
7892 return ret;
7893
Ingo Molnard1ccc662017-02-01 11:46:42 +01007894 /* ABI compatibility quirk: */
7895 if (!size)
Dario Faggiolid50dde52013-11-07 14:43:36 +01007896 size = SCHED_ATTR_SIZE_VER0;
Aleksa Saraidff3a852019-10-01 11:10:54 +10007897 if (size < SCHED_ATTR_SIZE_VER0 || size > PAGE_SIZE)
Dario Faggiolid50dde52013-11-07 14:43:36 +01007898 goto err_size;
7899
Aleksa Saraidff3a852019-10-01 11:10:54 +10007900 ret = copy_struct_from_user(attr, sizeof(*attr), uattr, size);
7901 if (ret) {
7902 if (ret == -E2BIG)
7903 goto err_size;
7904 return ret;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007905 }
7906
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007907 if ((attr->sched_flags & SCHED_FLAG_UTIL_CLAMP) &&
7908 size < SCHED_ATTR_SIZE_VER1)
7909 return -EINVAL;
7910
Dario Faggiolid50dde52013-11-07 14:43:36 +01007911 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01007912 * XXX: Do we want to be lenient like existing syscalls; or do we want
Dario Faggiolid50dde52013-11-07 14:43:36 +01007913 * to be strict and return an error on out-of-bounds values?
7914 */
Dongsheng Yang75e45d52014-02-11 15:34:50 +08007915 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
Dario Faggiolid50dde52013-11-07 14:43:36 +01007916
Michael Kerriske78c7bc2014-05-09 16:54:28 +02007917 return 0;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007918
7919err_size:
7920 put_user(sizeof(*attr), &uattr->size);
Michael Kerriske78c7bc2014-05-09 16:54:28 +02007921 return -E2BIG;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007922}
7923
Quentin Perretf4dddf92021-08-05 11:21:54 +01007924static void get_params(struct task_struct *p, struct sched_attr *attr)
7925{
7926 if (task_has_dl_policy(p))
7927 __getparam_dl(p, attr);
7928 else if (task_has_rt_policy(p))
7929 attr->sched_priority = p->rt_priority;
7930 else
7931 attr->sched_nice = task_nice(p);
7932}
7933
Linus Torvalds1da177e2005-04-16 15:20:36 -07007934/**
7935 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
7936 * @pid: the pid in question.
7937 * @policy: new policy.
7938 * @param: structure containing the new RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007939 *
7940 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007941 */
Ingo Molnard1ccc662017-02-01 11:46:42 +01007942SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007943{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007944 if (policy < 0)
7945 return -EINVAL;
7946
7947 return do_sched_setscheduler(pid, policy, param);
7948}
7949
7950/**
7951 * sys_sched_setparam - set/change the RT priority of a thread
7952 * @pid: the pid in question.
7953 * @param: structure containing the new RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02007954 *
7955 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007956 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01007957SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007958{
Steven Rostedtc13db6b2014-07-23 11:28:26 -04007959 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007960}
7961
7962/**
Dario Faggiolid50dde52013-11-07 14:43:36 +01007963 * sys_sched_setattr - same as above, but with extended sched_attr
7964 * @pid: the pid in question.
Juri Lelli5778fcc2014-01-14 16:10:39 +01007965 * @uattr: structure containing the extended parameters.
Masanari Iidadb66d752014-04-18 01:59:15 +09007966 * @flags: for future extension.
Dario Faggiolid50dde52013-11-07 14:43:36 +01007967 */
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01007968SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
7969 unsigned int, flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01007970{
7971 struct sched_attr attr;
7972 struct task_struct *p;
7973 int retval;
7974
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01007975 if (!uattr || pid < 0 || flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01007976 return -EINVAL;
7977
Michael Kerrisk143cf232014-05-09 16:54:15 +02007978 retval = sched_copy_attr(uattr, &attr);
7979 if (retval)
7980 return retval;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007981
Richard Weinbergerb14ed2c2014-06-02 22:38:34 +02007982 if ((int)attr.sched_policy < 0)
Peter Zijlstradbdb2272014-05-09 10:49:03 +02007983 return -EINVAL;
Patrick Bellasi1d6362f2019-06-21 09:42:06 +01007984 if (attr.sched_flags & SCHED_FLAG_KEEP_POLICY)
7985 attr.sched_policy = SETPARAM_POLICY;
Dario Faggiolid50dde52013-11-07 14:43:36 +01007986
7987 rcu_read_lock();
7988 retval = -ESRCH;
7989 p = find_process_by_pid(pid);
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007990 if (likely(p))
7991 get_task_struct(p);
Dario Faggiolid50dde52013-11-07 14:43:36 +01007992 rcu_read_unlock();
7993
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007994 if (likely(p)) {
Quentin Perretf4dddf92021-08-05 11:21:54 +01007995 if (attr.sched_flags & SCHED_FLAG_KEEP_PARAMS)
7996 get_params(p, &attr);
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01007997 retval = sched_setattr(p, &attr);
7998 put_task_struct(p);
7999 }
8000
Dario Faggiolid50dde52013-11-07 14:43:36 +01008001 return retval;
8002}
8003
8004/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07008005 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
8006 * @pid: the pid in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02008007 *
8008 * Return: On success, the policy of the thread. Otherwise, a negative error
8009 * code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008010 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008011SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008012{
Ingo Molnar36c8b582006-07-03 00:25:41 -07008013 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008014 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008015
8016 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02008017 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008018
8019 retval = -ESRCH;
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00008020 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008021 p = find_process_by_pid(pid);
8022 if (p) {
8023 retval = security_task_getscheduler(p);
8024 if (!retval)
Lennart Poetteringca94c442009-06-15 17:17:47 +02008025 retval = p->policy
8026 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008027 }
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00008028 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008029 return retval;
8030}
8031
8032/**
Lennart Poetteringca94c442009-06-15 17:17:47 +02008033 * sys_sched_getparam - get the RT priority of a thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07008034 * @pid: the pid in question.
8035 * @param: structure containing the RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02008036 *
8037 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
8038 * code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008039 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008040SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008041{
Peter Zijlstrace5f7f82014-05-12 22:50:34 +02008042 struct sched_param lp = { .sched_priority = 0 };
Ingo Molnar36c8b582006-07-03 00:25:41 -07008043 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008044 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008045
8046 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02008047 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008048
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00008049 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008050 p = find_process_by_pid(pid);
8051 retval = -ESRCH;
8052 if (!p)
8053 goto out_unlock;
8054
8055 retval = security_task_getscheduler(p);
8056 if (retval)
8057 goto out_unlock;
8058
Peter Zijlstrace5f7f82014-05-12 22:50:34 +02008059 if (task_has_rt_policy(p))
8060 lp.sched_priority = p->rt_priority;
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00008061 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008062
8063 /*
8064 * This one might sleep, we cannot do it with a spinlock held ...
8065 */
8066 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
8067
Linus Torvalds1da177e2005-04-16 15:20:36 -07008068 return retval;
8069
8070out_unlock:
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00008071 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008072 return retval;
8073}
8074
Ingo Molnar12512012019-09-04 09:55:32 +02008075/*
8076 * Copy the kernel size attribute structure (which might be larger
8077 * than what user-space knows about) to user-space.
8078 *
8079 * Note that all cases are valid: user-space buffer can be larger or
8080 * smaller than the kernel-space buffer. The usual case is that both
8081 * have the same size.
8082 */
8083static int
8084sched_attr_copy_to_user(struct sched_attr __user *uattr,
8085 struct sched_attr *kattr,
8086 unsigned int usize)
Dario Faggiolid50dde52013-11-07 14:43:36 +01008087{
Ingo Molnar12512012019-09-04 09:55:32 +02008088 unsigned int ksize = sizeof(*kattr);
Dario Faggiolid50dde52013-11-07 14:43:36 +01008089
Linus Torvalds96d4f262019-01-03 18:57:57 -08008090 if (!access_ok(uattr, usize))
Dario Faggiolid50dde52013-11-07 14:43:36 +01008091 return -EFAULT;
8092
8093 /*
Ingo Molnar12512012019-09-04 09:55:32 +02008094 * sched_getattr() ABI forwards and backwards compatibility:
8095 *
8096 * If usize == ksize then we just copy everything to user-space and all is good.
8097 *
8098 * If usize < ksize then we only copy as much as user-space has space for,
8099 * this keeps ABI compatibility as well. We skip the rest.
8100 *
8101 * If usize > ksize then user-space is using a newer version of the ABI,
8102 * which part the kernel doesn't know about. Just ignore it - tooling can
8103 * detect the kernel's knowledge of attributes from the attr->size value
8104 * which is set to ksize in this case.
Dario Faggiolid50dde52013-11-07 14:43:36 +01008105 */
Ingo Molnar12512012019-09-04 09:55:32 +02008106 kattr->size = min(usize, ksize);
Dario Faggiolid50dde52013-11-07 14:43:36 +01008107
Ingo Molnar12512012019-09-04 09:55:32 +02008108 if (copy_to_user(uattr, kattr, kattr->size))
Dario Faggiolid50dde52013-11-07 14:43:36 +01008109 return -EFAULT;
8110
Michael Kerrisk22400672014-05-09 16:54:33 +02008111 return 0;
Dario Faggiolid50dde52013-11-07 14:43:36 +01008112}
8113
8114/**
Dario Faggioliaab03e02013-11-28 11:14:43 +01008115 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
Dario Faggiolid50dde52013-11-07 14:43:36 +01008116 * @pid: the pid in question.
Juri Lelli5778fcc2014-01-14 16:10:39 +01008117 * @uattr: structure containing the extended parameters.
Aleksa Saraidff3a852019-10-01 11:10:54 +10008118 * @usize: sizeof(attr) for fwd/bwd comp.
Masanari Iidadb66d752014-04-18 01:59:15 +09008119 * @flags: for future extension.
Dario Faggiolid50dde52013-11-07 14:43:36 +01008120 */
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01008121SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
Ingo Molnar12512012019-09-04 09:55:32 +02008122 unsigned int, usize, unsigned int, flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01008123{
Ingo Molnar12512012019-09-04 09:55:32 +02008124 struct sched_attr kattr = { };
Dario Faggiolid50dde52013-11-07 14:43:36 +01008125 struct task_struct *p;
8126 int retval;
8127
Ingo Molnar12512012019-09-04 09:55:32 +02008128 if (!uattr || pid < 0 || usize > PAGE_SIZE ||
8129 usize < SCHED_ATTR_SIZE_VER0 || flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01008130 return -EINVAL;
8131
8132 rcu_read_lock();
8133 p = find_process_by_pid(pid);
8134 retval = -ESRCH;
8135 if (!p)
8136 goto out_unlock;
8137
8138 retval = security_task_getscheduler(p);
8139 if (retval)
8140 goto out_unlock;
8141
Ingo Molnar12512012019-09-04 09:55:32 +02008142 kattr.sched_policy = p->policy;
Peter Zijlstra7479f3c9c2014-01-15 17:05:04 +01008143 if (p->sched_reset_on_fork)
Ingo Molnar12512012019-09-04 09:55:32 +02008144 kattr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
Quentin Perretf4dddf92021-08-05 11:21:54 +01008145 get_params(p, &kattr);
Quentin Perret7ad721b2021-07-27 11:11:02 +01008146 kattr.sched_flags &= SCHED_FLAG_ALL;
Dario Faggiolid50dde52013-11-07 14:43:36 +01008147
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01008148#ifdef CONFIG_UCLAMP_TASK
Qais Yousef13685c42020-07-16 12:03:45 +01008149 /*
8150 * This could race with another potential updater, but this is fine
8151 * because it'll correctly read the old or the new value. We don't need
8152 * to guarantee who wins the race as long as it doesn't return garbage.
8153 */
Ingo Molnar12512012019-09-04 09:55:32 +02008154 kattr.sched_util_min = p->uclamp_req[UCLAMP_MIN].value;
8155 kattr.sched_util_max = p->uclamp_req[UCLAMP_MAX].value;
Patrick Bellasia509a7cd2019-06-21 09:42:07 +01008156#endif
8157
Dario Faggiolid50dde52013-11-07 14:43:36 +01008158 rcu_read_unlock();
8159
Ingo Molnar12512012019-09-04 09:55:32 +02008160 return sched_attr_copy_to_user(uattr, &kattr, usize);
Dario Faggiolid50dde52013-11-07 14:43:36 +01008161
8162out_unlock:
8163 rcu_read_unlock();
8164 return retval;
8165}
8166
Will Deacon234b8ab2021-07-30 12:24:36 +01008167#ifdef CONFIG_SMP
8168int dl_task_check_affinity(struct task_struct *p, const struct cpumask *mask)
8169{
8170 int ret = 0;
8171
8172 /*
8173 * If the task isn't a deadline task or admission control is
8174 * disabled then we don't care about affinity changes.
8175 */
8176 if (!task_has_dl_policy(p) || !dl_bandwidth_enabled())
8177 return 0;
8178
8179 /*
8180 * Since bandwidth control happens on root_domain basis,
8181 * if admission test is enabled, we only admit -deadline
8182 * tasks allowed to run on all the CPUs in the task's
8183 * root_domain.
8184 */
8185 rcu_read_lock();
8186 if (!cpumask_subset(task_rq(p)->rd->span, mask))
8187 ret = -EBUSY;
8188 rcu_read_unlock();
8189 return ret;
8190}
8191#endif
8192
Will Deacondb3b02a2021-07-30 12:24:34 +01008193static int
Waiman Long713a2e22022-09-22 14:00:40 -04008194__sched_setaffinity(struct task_struct *p, struct affinity_context *ctx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008195{
Ingo Molnar36c8b582006-07-03 00:25:41 -07008196 int retval;
Will Deacondb3b02a2021-07-30 12:24:34 +01008197 cpumask_var_t cpus_allowed, new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008198
Will Deacondb3b02a2021-07-30 12:24:34 +01008199 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL))
8200 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008201
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308202 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
8203 retval = -ENOMEM;
8204 goto out_free_cpus_allowed;
8205 }
Peter Zijlstrae4099a52013-12-17 10:03:34 +01008206
8207 cpuset_cpus_allowed(p, cpus_allowed);
Waiman Long713a2e22022-09-22 14:00:40 -04008208 cpumask_and(new_mask, ctx->new_mask, cpus_allowed);
8209
8210 ctx->new_mask = new_mask;
8211 ctx->flags |= SCA_CHECK;
Peter Zijlstrae4099a52013-12-17 10:03:34 +01008212
Will Deacon234b8ab2021-07-30 12:24:36 +01008213 retval = dl_task_check_affinity(p, new_mask);
8214 if (retval)
8215 goto out_free_new_mask;
Waiman Long8f9ea862022-09-22 14:00:38 -04008216
Waiman Long713a2e22022-09-22 14:00:40 -04008217 retval = __set_cpus_allowed_ptr(p, ctx);
Will Deacondb3b02a2021-07-30 12:24:34 +01008218 if (retval)
8219 goto out_free_new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008220
Will Deacondb3b02a2021-07-30 12:24:34 +01008221 cpuset_cpus_allowed(p, cpus_allowed);
8222 if (!cpumask_subset(new_mask, cpus_allowed)) {
8223 /*
8224 * We must have raced with a concurrent cpuset update.
8225 * Just reset the cpumask to the cpuset's cpus_allowed.
8226 */
8227 cpumask_copy(new_mask, cpus_allowed);
Waiman Long8f9ea862022-09-22 14:00:38 -04008228
8229 /*
8230 * If SCA_USER is set, a 2nd call to __set_cpus_allowed_ptr()
8231 * will restore the previous user_cpus_ptr value.
8232 *
8233 * In the unlikely event a previous user_cpus_ptr exists,
8234 * we need to further restrict the mask to what is allowed
8235 * by that old user_cpus_ptr.
8236 */
8237 if (unlikely((ctx->flags & SCA_USER) && ctx->user_mask)) {
8238 bool empty = !cpumask_and(new_mask, new_mask,
8239 ctx->user_mask);
8240
8241 if (WARN_ON_ONCE(empty))
8242 cpumask_copy(new_mask, cpus_allowed);
8243 }
8244 __set_cpus_allowed_ptr(p, ctx);
8245 retval = -EINVAL;
Paul Menage8707d8b2007-10-18 23:40:22 -07008246 }
Will Deacondb3b02a2021-07-30 12:24:34 +01008247
Kirill Tkhai16303ab2014-09-22 22:36:30 +04008248out_free_new_mask:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308249 free_cpumask_var(new_mask);
8250out_free_cpus_allowed:
8251 free_cpumask_var(cpus_allowed);
Will Deacondb3b02a2021-07-30 12:24:34 +01008252 return retval;
8253}
8254
Linus Torvalds1da177e2005-04-16 15:20:36 -07008255long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
8256{
Waiman Long8f9ea862022-09-22 14:00:38 -04008257 struct affinity_context ac;
8258 struct cpumask *user_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008259 struct task_struct *p;
8260 int retval;
8261
8262 rcu_read_lock();
8263
8264 p = find_process_by_pid(pid);
8265 if (!p) {
8266 rcu_read_unlock();
8267 return -ESRCH;
8268 }
8269
8270 /* Prevent p going away */
8271 get_task_struct(p);
8272 rcu_read_unlock();
8273
8274 if (p->flags & PF_NO_SETAFFINITY) {
8275 retval = -EINVAL;
8276 goto out_put_task;
8277 }
Will Deacondb3b02a2021-07-30 12:24:34 +01008278
Linus Torvalds1da177e2005-04-16 15:20:36 -07008279 if (!check_same_owner(p)) {
8280 rcu_read_lock();
8281 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
8282 rcu_read_unlock();
Will Deacondb3b02a2021-07-30 12:24:34 +01008283 retval = -EPERM;
8284 goto out_put_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008285 }
8286 rcu_read_unlock();
8287 }
8288
8289 retval = security_task_setscheduler(p);
8290 if (retval)
Will Deacondb3b02a2021-07-30 12:24:34 +01008291 goto out_put_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008292
Waiman Long5657c112023-01-15 14:31:22 -05008293 /*
8294 * With non-SMP configs, user_cpus_ptr/user_mask isn't used and
8295 * alloc_user_cpus_ptr() returns NULL.
8296 */
Waiman Long9a5418b2022-12-30 23:11:20 -05008297 user_mask = alloc_user_cpus_ptr(NUMA_NO_NODE);
Waiman Long5657c112023-01-15 14:31:22 -05008298 if (user_mask) {
8299 cpumask_copy(user_mask, in_mask);
8300 } else if (IS_ENABLED(CONFIG_SMP)) {
Waiman Long8f9ea862022-09-22 14:00:38 -04008301 retval = -ENOMEM;
8302 goto out_put_task;
8303 }
Waiman Long5657c112023-01-15 14:31:22 -05008304
Waiman Long8f9ea862022-09-22 14:00:38 -04008305 ac = (struct affinity_context){
8306 .new_mask = in_mask,
8307 .user_mask = user_mask,
8308 .flags = SCA_USER,
8309 };
8310
Waiman Long713a2e22022-09-22 14:00:40 -04008311 retval = __sched_setaffinity(p, &ac);
Waiman Long8f9ea862022-09-22 14:00:38 -04008312 kfree(ac.user_mask);
8313
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308314out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008315 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008316 return retval;
8317}
8318
8319static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e22008-11-25 02:35:14 +10308320 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008321{
Rusty Russell96f874e22008-11-25 02:35:14 +10308322 if (len < cpumask_size())
8323 cpumask_clear(new_mask);
8324 else if (len > cpumask_size())
8325 len = cpumask_size();
8326
Linus Torvalds1da177e2005-04-16 15:20:36 -07008327 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
8328}
8329
8330/**
Ingo Molnard1ccc662017-02-01 11:46:42 +01008331 * sys_sched_setaffinity - set the CPU affinity of a process
Linus Torvalds1da177e2005-04-16 15:20:36 -07008332 * @pid: pid of the process
8333 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
Ingo Molnard1ccc662017-02-01 11:46:42 +01008334 * @user_mask_ptr: user-space pointer to the new CPU mask
Yacine Belkadie69f6182013-07-12 20:45:47 +02008335 *
8336 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008337 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008338SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
8339 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008340{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308341 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008342 int retval;
8343
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308344 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
8345 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008346
Rusty Russell5a16f3d2008-11-25 02:35:11 +10308347 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
8348 if (retval == 0)
8349 retval = sched_setaffinity(pid, new_mask);
8350 free_cpumask_var(new_mask);
8351 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008352}
8353
Rusty Russell96f874e22008-11-25 02:35:14 +10308354long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008355{
Ingo Molnar36c8b582006-07-03 00:25:41 -07008356 struct task_struct *p;
Thomas Gleixner31605682009-12-08 20:24:16 +00008357 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008358 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008359
Thomas Gleixner23f5d142009-12-09 10:15:01 +00008360 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008361
8362 retval = -ESRCH;
8363 p = find_process_by_pid(pid);
8364 if (!p)
8365 goto out_unlock;
8366
David Quigleye7834f82006-06-23 02:03:59 -07008367 retval = security_task_getscheduler(p);
8368 if (retval)
8369 goto out_unlock;
8370
Peter Zijlstra013fdb82011-04-05 17:23:45 +02008371 raw_spin_lock_irqsave(&p->pi_lock, flags);
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02008372 cpumask_and(mask, &p->cpus_mask, cpu_active_mask);
Peter Zijlstra013fdb82011-04-05 17:23:45 +02008373 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008374
8375out_unlock:
Thomas Gleixner23f5d142009-12-09 10:15:01 +00008376 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008377
Ulrich Drepper9531b622007-08-09 11:16:46 +02008378 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008379}
8380
8381/**
Ingo Molnard1ccc662017-02-01 11:46:42 +01008382 * sys_sched_getaffinity - get the CPU affinity of a process
Linus Torvalds1da177e2005-04-16 15:20:36 -07008383 * @pid: pid of the process
8384 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
Ingo Molnard1ccc662017-02-01 11:46:42 +01008385 * @user_mask_ptr: user-space pointer to hold the current CPU mask
Yacine Belkadie69f6182013-07-12 20:45:47 +02008386 *
Zev Weiss599b4842016-06-26 16:13:23 -05008387 * Return: size of CPU mask copied to user_mask_ptr on success. An
8388 * error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008389 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008390SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
8391 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008392{
8393 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10308394 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008395
Anton Blanchard84fba5e2010-04-06 17:02:19 +10008396 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09008397 return -EINVAL;
8398 if (len & (sizeof(unsigned long)-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008399 return -EINVAL;
8400
Rusty Russellf17c8602008-11-25 02:35:11 +10308401 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
8402 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008403
Rusty Russellf17c8602008-11-25 02:35:11 +10308404 ret = sched_getaffinity(pid, mask);
8405 if (ret == 0) {
Alexey Dobriyan4de373a2018-02-06 15:39:37 -08008406 unsigned int retlen = min(len, cpumask_size());
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09008407
8408 if (copy_to_user(user_mask_ptr, mask, retlen))
Rusty Russellf17c8602008-11-25 02:35:11 +10308409 ret = -EFAULT;
8410 else
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09008411 ret = retlen;
Rusty Russellf17c8602008-11-25 02:35:11 +10308412 }
8413 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008414
Rusty Russellf17c8602008-11-25 02:35:11 +10308415 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008416}
8417
Dominik Brodowski7d4dd4f2018-03-14 22:40:35 +01008418static void do_sched_yield(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008419{
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008420 struct rq_flags rf;
8421 struct rq *rq;
8422
Johannes Weiner246b3b32018-10-26 15:06:23 -07008423 rq = this_rq_lock_irq(&rf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008424
Josh Poimboeufae928822016-06-17 12:43:24 -05008425 schedstat_inc(rq->yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02008426 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008427
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02008428 preempt_disable();
Thomas Gleixner345a9572020-10-20 16:46:55 +02008429 rq_unlock_irq(rq, &rf);
Thomas Gleixnerba74c142011-03-21 13:32:17 +01008430 sched_preempt_enable_no_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008431
8432 schedule();
Dominik Brodowski7d4dd4f2018-03-14 22:40:35 +01008433}
Linus Torvalds1da177e2005-04-16 15:20:36 -07008434
Mauro Carvalho Chehab59a74b12020-12-01 13:09:06 +01008435/**
8436 * sys_sched_yield - yield the current processor to other threads.
8437 *
8438 * This function yields the current CPU to other tasks. If there are no
8439 * other threads running on this CPU then this function will return.
8440 *
8441 * Return: 0.
8442 */
Dominik Brodowski7d4dd4f2018-03-14 22:40:35 +01008443SYSCALL_DEFINE0(sched_yield)
8444{
8445 do_sched_yield();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008446 return 0;
8447}
8448
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008449#if !defined(CONFIG_PREEMPTION) || defined(CONFIG_PREEMPT_DYNAMIC)
8450int __sched __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008451{
Konstantin Khlebnikovfe32d3c2015-07-15 12:52:04 +03008452 if (should_resched(0)) {
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01008453 preempt_schedule_common();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008454 return 1;
8455 }
Frederic Weisbecker50895822021-07-06 01:43:43 +02008456 /*
8457 * In preemptible kernels, ->rcu_read_lock_nesting tells the tick
8458 * whether the current CPU is in an RCU read-side critical section,
8459 * so the tick can report quiescent states even for CPUs looping
8460 * in kernel context. In contrast, in non-preemptible kernels,
8461 * RCU readers leave no in-memory hints, which means that CPU-bound
8462 * processes executing in kernel context might never report an
8463 * RCU quiescent state. Therefore, the following code causes
8464 * cond_resched() to report a quiescent state, but only when RCU
8465 * is in urgent need of one.
8466 */
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008467#ifndef CONFIG_PREEMPT_RCU
Paul E. McKenneyf79c3ad2016-11-30 06:24:30 -08008468 rcu_all_qs();
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008469#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008470 return 0;
8471}
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008472EXPORT_SYMBOL(__cond_resched);
8473#endif
8474
8475#ifdef CONFIG_PREEMPT_DYNAMIC
Mark Rutland99cf983c2022-02-14 16:52:14 +00008476#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
Mark Rutland8a69fe02022-02-14 16:52:11 +00008477#define cond_resched_dynamic_enabled __cond_resched
8478#define cond_resched_dynamic_disabled ((void *)&__static_call_return0)
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008479DEFINE_STATIC_CALL_RET0(cond_resched, __cond_resched);
Peter Zijlstraef726612021-01-25 16:26:50 +01008480EXPORT_STATIC_CALL_TRAMP(cond_resched);
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008481
Mark Rutland8a69fe02022-02-14 16:52:11 +00008482#define might_resched_dynamic_enabled __cond_resched
8483#define might_resched_dynamic_disabled ((void *)&__static_call_return0)
Peter Zijlstra (Intel)b965f1d2021-01-18 15:12:20 +01008484DEFINE_STATIC_CALL_RET0(might_resched, __cond_resched);
Peter Zijlstraef726612021-01-25 16:26:50 +01008485EXPORT_STATIC_CALL_TRAMP(might_resched);
Mark Rutland99cf983c2022-02-14 16:52:14 +00008486#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8487static DEFINE_STATIC_KEY_FALSE(sk_dynamic_cond_resched);
8488int __sched dynamic_cond_resched(void)
8489{
8490 if (!static_branch_unlikely(&sk_dynamic_cond_resched))
8491 return 0;
8492 return __cond_resched();
8493}
8494EXPORT_SYMBOL(dynamic_cond_resched);
8495
8496static DEFINE_STATIC_KEY_FALSE(sk_dynamic_might_resched);
8497int __sched dynamic_might_resched(void)
8498{
8499 if (!static_branch_unlikely(&sk_dynamic_might_resched))
8500 return 0;
8501 return __cond_resched();
8502}
8503EXPORT_SYMBOL(dynamic_might_resched);
8504#endif
Peter Zijlstra35a773a2016-09-19 12:57:53 +02008505#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008506
8507/*
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02008508 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07008509 * call schedule, and on return reacquire the lock.
8510 *
Thomas Gleixnerc1a280b2019-07-26 23:19:37 +02008511 * This works OK both with and without CONFIG_PREEMPTION. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07008512 * operations here to prevent schedule() from being called twice (once via
8513 * spin_unlock(), once by hand).
8514 */
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02008515int __cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008516{
Konstantin Khlebnikovfe32d3c2015-07-15 12:52:04 +03008517 int resched = should_resched(PREEMPT_LOCK_OFFSET);
Jan Kara6df3cec2005-06-13 15:52:32 -07008518 int ret = 0;
8519
Peter Zijlstraf607c662009-07-20 19:16:29 +02008520 lockdep_assert_held(lock);
8521
Paul E. McKenney4a81e832014-06-20 16:49:01 -07008522 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008523 spin_unlock(lock);
Peter Zijlstra7e406d1f2021-12-25 01:04:57 +01008524 if (!_cond_resched())
Nick Piggin95c354f2008-01-30 13:31:20 +01008525 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07008526 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008527 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008528 }
Jan Kara6df3cec2005-06-13 15:52:32 -07008529 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008530}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02008531EXPORT_SYMBOL(__cond_resched_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008532
Ben Gardonf3d4b4b2021-02-02 10:57:14 -08008533int __cond_resched_rwlock_read(rwlock_t *lock)
8534{
8535 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8536 int ret = 0;
8537
8538 lockdep_assert_held_read(lock);
8539
8540 if (rwlock_needbreak(lock) || resched) {
8541 read_unlock(lock);
Peter Zijlstra7e406d1f2021-12-25 01:04:57 +01008542 if (!_cond_resched())
Ben Gardonf3d4b4b2021-02-02 10:57:14 -08008543 cpu_relax();
8544 ret = 1;
8545 read_lock(lock);
8546 }
8547 return ret;
8548}
8549EXPORT_SYMBOL(__cond_resched_rwlock_read);
8550
8551int __cond_resched_rwlock_write(rwlock_t *lock)
8552{
8553 int resched = should_resched(PREEMPT_LOCK_OFFSET);
8554 int ret = 0;
8555
8556 lockdep_assert_held_write(lock);
8557
8558 if (rwlock_needbreak(lock) || resched) {
8559 write_unlock(lock);
Peter Zijlstra7e406d1f2021-12-25 01:04:57 +01008560 if (!_cond_resched())
Ben Gardonf3d4b4b2021-02-02 10:57:14 -08008561 cpu_relax();
8562 ret = 1;
8563 write_lock(lock);
8564 }
8565 return ret;
8566}
8567EXPORT_SYMBOL(__cond_resched_rwlock_write);
8568
Mark Rutland4c748552022-02-14 16:52:10 +00008569#ifdef CONFIG_PREEMPT_DYNAMIC
8570
Mark Rutland33c64732022-02-14 16:52:13 +00008571#ifdef CONFIG_GENERIC_ENTRY
Mark Rutland4c748552022-02-14 16:52:10 +00008572#include <linux/entry-common.h>
Mark Rutland33c64732022-02-14 16:52:13 +00008573#endif
Mark Rutland4c748552022-02-14 16:52:10 +00008574
8575/*
8576 * SC:cond_resched
8577 * SC:might_resched
8578 * SC:preempt_schedule
8579 * SC:preempt_schedule_notrace
8580 * SC:irqentry_exit_cond_resched
8581 *
8582 *
8583 * NONE:
8584 * cond_resched <- __cond_resched
8585 * might_resched <- RET0
8586 * preempt_schedule <- NOP
8587 * preempt_schedule_notrace <- NOP
8588 * irqentry_exit_cond_resched <- NOP
8589 *
8590 * VOLUNTARY:
8591 * cond_resched <- __cond_resched
8592 * might_resched <- __cond_resched
8593 * preempt_schedule <- NOP
8594 * preempt_schedule_notrace <- NOP
8595 * irqentry_exit_cond_resched <- NOP
8596 *
8597 * FULL:
8598 * cond_resched <- RET0
8599 * might_resched <- RET0
8600 * preempt_schedule <- preempt_schedule
8601 * preempt_schedule_notrace <- preempt_schedule_notrace
8602 * irqentry_exit_cond_resched <- irqentry_exit_cond_resched
8603 */
8604
8605enum {
8606 preempt_dynamic_undefined = -1,
8607 preempt_dynamic_none,
8608 preempt_dynamic_voluntary,
8609 preempt_dynamic_full,
8610};
8611
8612int preempt_dynamic_mode = preempt_dynamic_undefined;
8613
8614int sched_dynamic_mode(const char *str)
8615{
8616 if (!strcmp(str, "none"))
8617 return preempt_dynamic_none;
8618
8619 if (!strcmp(str, "voluntary"))
8620 return preempt_dynamic_voluntary;
8621
8622 if (!strcmp(str, "full"))
8623 return preempt_dynamic_full;
8624
8625 return -EINVAL;
8626}
8627
Mark Rutland99cf983c2022-02-14 16:52:14 +00008628#if defined(CONFIG_HAVE_PREEMPT_DYNAMIC_CALL)
Mark Rutland8a69fe02022-02-14 16:52:11 +00008629#define preempt_dynamic_enable(f) static_call_update(f, f##_dynamic_enabled)
8630#define preempt_dynamic_disable(f) static_call_update(f, f##_dynamic_disabled)
Mark Rutland99cf983c2022-02-14 16:52:14 +00008631#elif defined(CONFIG_HAVE_PREEMPT_DYNAMIC_KEY)
8632#define preempt_dynamic_enable(f) static_key_enable(&sk_dynamic_##f.key)
8633#define preempt_dynamic_disable(f) static_key_disable(&sk_dynamic_##f.key)
8634#else
8635#error "Unsupported PREEMPT_DYNAMIC mechanism"
8636#endif
Mark Rutland8a69fe02022-02-14 16:52:11 +00008637
Mark Rutland4c748552022-02-14 16:52:10 +00008638void sched_dynamic_update(int mode)
8639{
8640 /*
8641 * Avoid {NONE,VOLUNTARY} -> FULL transitions from ever ending up in
8642 * the ZERO state, which is invalid.
8643 */
Mark Rutland8a69fe02022-02-14 16:52:11 +00008644 preempt_dynamic_enable(cond_resched);
8645 preempt_dynamic_enable(might_resched);
8646 preempt_dynamic_enable(preempt_schedule);
8647 preempt_dynamic_enable(preempt_schedule_notrace);
8648 preempt_dynamic_enable(irqentry_exit_cond_resched);
Mark Rutland4c748552022-02-14 16:52:10 +00008649
8650 switch (mode) {
8651 case preempt_dynamic_none:
Mark Rutland8a69fe02022-02-14 16:52:11 +00008652 preempt_dynamic_enable(cond_resched);
8653 preempt_dynamic_disable(might_resched);
8654 preempt_dynamic_disable(preempt_schedule);
8655 preempt_dynamic_disable(preempt_schedule_notrace);
8656 preempt_dynamic_disable(irqentry_exit_cond_resched);
Mark Rutland4c748552022-02-14 16:52:10 +00008657 pr_info("Dynamic Preempt: none\n");
8658 break;
8659
8660 case preempt_dynamic_voluntary:
Mark Rutland8a69fe02022-02-14 16:52:11 +00008661 preempt_dynamic_enable(cond_resched);
8662 preempt_dynamic_enable(might_resched);
8663 preempt_dynamic_disable(preempt_schedule);
8664 preempt_dynamic_disable(preempt_schedule_notrace);
8665 preempt_dynamic_disable(irqentry_exit_cond_resched);
Mark Rutland4c748552022-02-14 16:52:10 +00008666 pr_info("Dynamic Preempt: voluntary\n");
8667 break;
8668
8669 case preempt_dynamic_full:
Mark Rutland8a69fe02022-02-14 16:52:11 +00008670 preempt_dynamic_disable(cond_resched);
8671 preempt_dynamic_disable(might_resched);
8672 preempt_dynamic_enable(preempt_schedule);
8673 preempt_dynamic_enable(preempt_schedule_notrace);
8674 preempt_dynamic_enable(irqentry_exit_cond_resched);
Mark Rutland4c748552022-02-14 16:52:10 +00008675 pr_info("Dynamic Preempt: full\n");
8676 break;
8677 }
8678
8679 preempt_dynamic_mode = mode;
8680}
8681
8682static int __init setup_preempt_mode(char *str)
8683{
8684 int mode = sched_dynamic_mode(str);
8685 if (mode < 0) {
8686 pr_warn("Dynamic Preempt: unsupported mode: %s\n", str);
8687 return 0;
8688 }
8689
8690 sched_dynamic_update(mode);
8691 return 1;
8692}
8693__setup("preempt=", setup_preempt_mode);
8694
8695static void __init preempt_dynamic_init(void)
8696{
8697 if (preempt_dynamic_mode == preempt_dynamic_undefined) {
8698 if (IS_ENABLED(CONFIG_PREEMPT_NONE)) {
8699 sched_dynamic_update(preempt_dynamic_none);
8700 } else if (IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)) {
8701 sched_dynamic_update(preempt_dynamic_voluntary);
8702 } else {
8703 /* Default static call setting, nothing to do */
8704 WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT));
8705 preempt_dynamic_mode = preempt_dynamic_full;
8706 pr_info("Dynamic Preempt: full\n");
8707 }
8708 }
8709}
8710
Valentin Schneidercfe43f42021-11-12 18:52:01 +00008711#define PREEMPT_MODEL_ACCESSOR(mode) \
8712 bool preempt_model_##mode(void) \
8713 { \
8714 WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
8715 return preempt_dynamic_mode == preempt_dynamic_##mode; \
8716 } \
8717 EXPORT_SYMBOL_GPL(preempt_model_##mode)
8718
8719PREEMPT_MODEL_ACCESSOR(none);
8720PREEMPT_MODEL_ACCESSOR(voluntary);
8721PREEMPT_MODEL_ACCESSOR(full);
8722
Mark Rutland4c748552022-02-14 16:52:10 +00008723#else /* !CONFIG_PREEMPT_DYNAMIC */
8724
8725static inline void preempt_dynamic_init(void) { }
8726
8727#endif /* #ifdef CONFIG_PREEMPT_DYNAMIC */
8728
Linus Torvalds1da177e2005-04-16 15:20:36 -07008729/**
8730 * yield - yield the current processor to other threads.
8731 *
Peter Zijlstra8e3fabf2012-03-06 18:54:26 +01008732 * Do not ever use this function, there's a 99% chance you're doing it wrong.
8733 *
8734 * The scheduler is at all times free to pick the calling task as the most
8735 * eligible task to run, if removing the yield() call from your code breaks
Tal Zussmanb19a8882020-11-12 19:51:56 -05008736 * it, it's already broken.
Peter Zijlstra8e3fabf2012-03-06 18:54:26 +01008737 *
8738 * Typical broken usage is:
8739 *
8740 * while (!event)
Ingo Molnard1ccc662017-02-01 11:46:42 +01008741 * yield();
Peter Zijlstra8e3fabf2012-03-06 18:54:26 +01008742 *
8743 * where one assumes that yield() will let 'the other' process run that will
8744 * make event true. If the current task is a SCHED_FIFO task that will never
8745 * happen. Never use yield() as a progress guarantee!!
8746 *
8747 * If you want to use yield() to wait for something, use wait_event().
8748 * If you want to use yield() to be 'nice' for others, use cond_resched().
8749 * If you still want to use yield(), do not!
Linus Torvalds1da177e2005-04-16 15:20:36 -07008750 */
8751void __sched yield(void)
8752{
8753 set_current_state(TASK_RUNNING);
Dominik Brodowski7d4dd4f2018-03-14 22:40:35 +01008754 do_sched_yield();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008755}
Linus Torvalds1da177e2005-04-16 15:20:36 -07008756EXPORT_SYMBOL(yield);
8757
Mike Galbraithd95f4122011-02-01 09:50:51 -05008758/**
8759 * yield_to - yield the current processor to another thread in
8760 * your thread group, or accelerate that thread toward the
8761 * processor it's on.
Randy Dunlap16addf92011-03-18 09:34:53 -07008762 * @p: target task
8763 * @preempt: whether task preemption is allowed or not
Mike Galbraithd95f4122011-02-01 09:50:51 -05008764 *
8765 * It's the caller's job to ensure that the target task struct
8766 * can't go away on us before we can do any checks.
8767 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02008768 * Return:
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308769 * true (>0) if we indeed boosted the target task.
8770 * false (0) if we failed to boost the target.
8771 * -ESRCH if there's no task to yield to.
Mike Galbraithd95f4122011-02-01 09:50:51 -05008772 */
Dan Carpenterfa933842014-05-23 13:20:42 +03008773int __sched yield_to(struct task_struct *p, bool preempt)
Mike Galbraithd95f4122011-02-01 09:50:51 -05008774{
8775 struct task_struct *curr = current;
8776 struct rq *rq, *p_rq;
8777 unsigned long flags;
Dan Carpenterc3c18642013-02-05 14:37:51 +03008778 int yielded = 0;
Mike Galbraithd95f4122011-02-01 09:50:51 -05008779
8780 local_irq_save(flags);
8781 rq = this_rq();
8782
8783again:
8784 p_rq = task_rq(p);
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308785 /*
8786 * If we're the only runnable task on the rq and target rq also
8787 * has only one task, there's absolutely no point in yielding.
8788 */
8789 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
8790 yielded = -ESRCH;
8791 goto out_irq;
8792 }
8793
Mike Galbraithd95f4122011-02-01 09:50:51 -05008794 double_rq_lock(rq, p_rq);
Shigeru Yoshida39e24d8f2013-11-23 18:38:01 +09008795 if (task_rq(p) != p_rq) {
Mike Galbraithd95f4122011-02-01 09:50:51 -05008796 double_rq_unlock(rq, p_rq);
8797 goto again;
8798 }
8799
8800 if (!curr->sched_class->yield_to_task)
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308801 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05008802
8803 if (curr->sched_class != p->sched_class)
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308804 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05008805
Peter Zijlstra0b9d46f2022-09-06 12:33:04 +02008806 if (task_on_cpu(p_rq, p) || !task_is_running(p))
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308807 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05008808
Dietmar Eggemann0900acf2020-06-03 10:03:02 +02008809 yielded = curr->sched_class->yield_to_task(rq, p);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08008810 if (yielded) {
Josh Poimboeufae928822016-06-17 12:43:24 -05008811 schedstat_inc(rq->yld_count);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08008812 /*
8813 * Make p's CPU reschedule; pick_next_entity takes care of
8814 * fairness.
8815 */
8816 if (preempt && rq != p_rq)
Kirill Tkhai88751252014-06-29 00:03:57 +04008817 resched_curr(p_rq);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08008818 }
Mike Galbraithd95f4122011-02-01 09:50:51 -05008819
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308820out_unlock:
Mike Galbraithd95f4122011-02-01 09:50:51 -05008821 double_rq_unlock(rq, p_rq);
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308822out_irq:
Mike Galbraithd95f4122011-02-01 09:50:51 -05008823 local_irq_restore(flags);
8824
Peter Zijlstra7b270f62013-01-22 13:09:13 +05308825 if (yielded > 0)
Mike Galbraithd95f4122011-02-01 09:50:51 -05008826 schedule();
8827
8828 return yielded;
8829}
8830EXPORT_SYMBOL_GPL(yield_to);
8831
Tejun Heo10ab5642016-10-28 12:58:10 -04008832int io_schedule_prepare(void)
8833{
8834 int old_iowait = current->in_iowait;
8835
8836 current->in_iowait = 1;
Christoph Hellwigaa8dccc2022-01-27 08:05:49 +01008837 blk_flush_plug(current->plug, true);
Tejun Heo10ab5642016-10-28 12:58:10 -04008838 return old_iowait;
8839}
8840
8841void io_schedule_finish(int token)
8842{
8843 current->in_iowait = token;
8844}
8845
Linus Torvalds1da177e2005-04-16 15:20:36 -07008846/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008847 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07008848 * that process accounting knows that this is a task in IO wait state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008849 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008850long __sched io_schedule_timeout(long timeout)
8851{
Tejun Heo10ab5642016-10-28 12:58:10 -04008852 int token;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008853 long ret;
8854
Tejun Heo10ab5642016-10-28 12:58:10 -04008855 token = io_schedule_prepare();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008856 ret = schedule_timeout(timeout);
Tejun Heo10ab5642016-10-28 12:58:10 -04008857 io_schedule_finish(token);
NeilBrown9cff8ad2015-02-13 15:49:17 +11008858
Linus Torvalds1da177e2005-04-16 15:20:36 -07008859 return ret;
8860}
NeilBrown9cff8ad2015-02-13 15:49:17 +11008861EXPORT_SYMBOL(io_schedule_timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008862
Gao Xiange3b929b2019-06-03 17:13:38 +08008863void __sched io_schedule(void)
Tejun Heo10ab5642016-10-28 12:58:10 -04008864{
8865 int token;
8866
8867 token = io_schedule_prepare();
8868 schedule();
8869 io_schedule_finish(token);
8870}
8871EXPORT_SYMBOL(io_schedule);
8872
Linus Torvalds1da177e2005-04-16 15:20:36 -07008873/**
8874 * sys_sched_get_priority_max - return maximum RT priority.
8875 * @policy: scheduling class.
8876 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02008877 * Return: On success, this syscall returns the maximum
8878 * rt_priority that can be used by a given scheduling class.
8879 * On failure, a negative error code is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008880 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008881SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008882{
8883 int ret = -EINVAL;
8884
8885 switch (policy) {
8886 case SCHED_FIFO:
8887 case SCHED_RR:
Dietmar Eggemannae18ad22021-01-28 14:10:38 +01008888 ret = MAX_RT_PRIO-1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008889 break;
Dario Faggioliaab03e02013-11-28 11:14:43 +01008890 case SCHED_DEADLINE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008891 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08008892 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02008893 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008894 ret = 0;
8895 break;
8896 }
8897 return ret;
8898}
8899
8900/**
8901 * sys_sched_get_priority_min - return minimum RT priority.
8902 * @policy: scheduling class.
8903 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02008904 * Return: On success, this syscall returns the minimum
8905 * rt_priority that can be used by a given scheduling class.
8906 * On failure, a negative error code is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008907 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01008908SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008909{
8910 int ret = -EINVAL;
8911
8912 switch (policy) {
8913 case SCHED_FIFO:
8914 case SCHED_RR:
8915 ret = 1;
8916 break;
Dario Faggioliaab03e02013-11-28 11:14:43 +01008917 case SCHED_DEADLINE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008918 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08008919 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02008920 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008921 ret = 0;
8922 }
8923 return ret;
8924}
8925
Al Viroabca5fc2017-09-19 18:17:46 -04008926static int sched_rr_get_interval(pid_t pid, struct timespec64 *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008927{
Ingo Molnar36c8b582006-07-03 00:25:41 -07008928 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02008929 unsigned int time_slice;
Peter Zijlstraeb580752015-07-31 21:28:18 +02008930 struct rq_flags rf;
Thomas Gleixnerdba091b2009-12-09 09:32:03 +01008931 struct rq *rq;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008932 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008933
8934 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02008935 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008936
8937 retval = -ESRCH;
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00008938 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008939 p = find_process_by_pid(pid);
8940 if (!p)
8941 goto out_unlock;
8942
8943 retval = security_task_getscheduler(p);
8944 if (retval)
8945 goto out_unlock;
8946
Peter Zijlstraeb580752015-07-31 21:28:18 +02008947 rq = task_rq_lock(p, &rf);
Peter Zijlstraa57beec2014-01-27 11:54:13 +01008948 time_slice = 0;
8949 if (p->sched_class->get_rr_interval)
8950 time_slice = p->sched_class->get_rr_interval(rq, p);
Peter Zijlstraeb580752015-07-31 21:28:18 +02008951 task_rq_unlock(rq, p, &rf);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02008952
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00008953 rcu_read_unlock();
Al Viroabca5fc2017-09-19 18:17:46 -04008954 jiffies_to_timespec64(time_slice, t);
8955 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008956
Linus Torvalds1da177e2005-04-16 15:20:36 -07008957out_unlock:
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00008958 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008959 return retval;
8960}
8961
Randy Dunlap2064a5a2017-12-03 13:19:00 -08008962/**
8963 * sys_sched_rr_get_interval - return the default timeslice of a process.
8964 * @pid: pid of the process.
8965 * @interval: userspace pointer to the timeslice value.
8966 *
8967 * this syscall writes the default timeslice value of a given process
8968 * into the user-space timespec buffer. A value of '0' means infinity.
8969 *
8970 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
8971 * an error code.
8972 */
Al Viroabca5fc2017-09-19 18:17:46 -04008973SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Arnd Bergmann474b9c72018-04-17 21:59:47 +02008974 struct __kernel_timespec __user *, interval)
Al Viroabca5fc2017-09-19 18:17:46 -04008975{
8976 struct timespec64 t;
8977 int retval = sched_rr_get_interval(pid, &t);
8978
8979 if (retval == 0)
8980 retval = put_timespec64(&t, interval);
8981
8982 return retval;
8983}
8984
Arnd Bergmann474b9c72018-04-17 21:59:47 +02008985#ifdef CONFIG_COMPAT_32BIT_TIME
Arnd Bergmann8dabe722019-01-07 00:33:08 +01008986SYSCALL_DEFINE2(sched_rr_get_interval_time32, pid_t, pid,
8987 struct old_timespec32 __user *, interval)
Al Viroabca5fc2017-09-19 18:17:46 -04008988{
8989 struct timespec64 t;
8990 int retval = sched_rr_get_interval(pid, &t);
8991
8992 if (retval == 0)
Arnd Bergmann9afc5ee2018-07-13 12:52:28 +02008993 retval = put_old_timespec32(&t, interval);
Al Viroabca5fc2017-09-19 18:17:46 -04008994 return retval;
8995}
8996#endif
8997
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01008998void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008999{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009000 unsigned long free = 0;
Paul E. McKenney4e797522012-11-07 13:35:32 -08009001 int ppid;
Ingo Molnarc930b2c2017-02-03 12:22:54 +01009002
Tetsuo Handa38200502016-11-02 19:50:29 +09009003 if (!try_get_task_stack(p))
9004 return;
Xie XiuQi20435d82017-08-07 16:44:23 +08009005
Libing Zhoucc172ff2020-08-14 11:02:36 +08009006 pr_info("task:%-15.15s state:%c", p->comm, task_state_to_char(p));
Xie XiuQi20435d82017-08-07 16:44:23 +08009007
Peter Zijlstrab03fbd42021-06-11 10:28:12 +02009008 if (task_is_running(p))
Libing Zhoucc172ff2020-08-14 11:02:36 +08009009 pr_cont(" running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07009010#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05009011 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009012#endif
Oleg Nesterova90e9842014-12-10 15:45:21 -08009013 ppid = 0;
Paul E. McKenney4e797522012-11-07 13:35:32 -08009014 rcu_read_lock();
Oleg Nesterova90e9842014-12-10 15:45:21 -08009015 if (pid_alive(p))
9016 ppid = task_pid_nr(rcu_dereference(p->real_parent));
Paul E. McKenney4e797522012-11-07 13:35:32 -08009017 rcu_read_unlock();
Zhen Lei0f03d682022-07-27 14:08:19 +08009018 pr_cont(" stack:%-5lu pid:%-5d ppid:%-6d flags:0x%08lx\n",
Libing Zhoucc172ff2020-08-14 11:02:36 +08009019 free, task_pid_nr(p), ppid,
Mark Rutland0569b242021-11-29 13:06:45 +00009020 read_task_thread_flags(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07009021
Tejun Heo3d1cb202013-04-30 15:27:22 -07009022 print_worker_info(KERN_INFO, p);
Peter Zijlstraa8b62fd2020-09-21 12:58:17 +02009023 print_stop_info(KERN_INFO, p);
Dmitry Safonov9cb8f062020-06-08 21:32:29 -07009024 show_stack(p, NULL, KERN_INFO);
Tetsuo Handa38200502016-11-02 19:50:29 +09009025 put_task_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009026}
Paul E. McKenney0032f4e2017-08-30 10:40:17 -07009027EXPORT_SYMBOL_GPL(sched_show_task);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009028
Peter Zijlstra5d68cc92017-09-22 18:32:41 +02009029static inline bool
9030state_filter_match(unsigned long state_filter, struct task_struct *p)
9031{
Peter Zijlstra2f064a52021-06-11 10:28:17 +02009032 unsigned int state = READ_ONCE(p->__state);
9033
Peter Zijlstra5d68cc92017-09-22 18:32:41 +02009034 /* no filter, everything matches */
9035 if (!state_filter)
9036 return true;
9037
9038 /* filter, but doesn't match */
Peter Zijlstra2f064a52021-06-11 10:28:17 +02009039 if (!(state & state_filter))
Peter Zijlstra5d68cc92017-09-22 18:32:41 +02009040 return false;
9041
9042 /*
9043 * When looking for TASK_UNINTERRUPTIBLE skip TASK_IDLE (allows
9044 * TASK_KILLABLE).
9045 */
Peter Zijlstra5aec7882022-09-27 21:02:34 +02009046 if (state_filter == TASK_UNINTERRUPTIBLE && (state & TASK_NOLOAD))
Peter Zijlstra5d68cc92017-09-22 18:32:41 +02009047 return false;
9048
9049 return true;
9050}
9051
9052
Peter Zijlstra2f064a52021-06-11 10:28:17 +02009053void show_state_filter(unsigned int state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009054{
Ingo Molnar36c8b582006-07-03 00:25:41 -07009055 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009056
Thomas Gleixner510f5ac2011-07-17 20:47:54 +02009057 rcu_read_lock();
Oleg Nesterov5d07f422014-08-13 21:19:53 +02009058 for_each_process_thread(g, p) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07009059 /*
9060 * reset the NMI-timeout, listing all files on a slow
Lucas De Marchi25985ed2011-03-30 22:57:33 -03009061 * console might take a lot of time:
Andrey Ryabinin57675cb2016-06-09 15:20:05 +03009062 * Also, reset softlockup watchdogs on all CPUs, because
9063 * another CPU might be blocked waiting for us to process
9064 * an IPI.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009065 */
9066 touch_nmi_watchdog();
Andrey Ryabinin57675cb2016-06-09 15:20:05 +03009067 touch_all_softlockup_watchdogs();
Peter Zijlstra5d68cc92017-09-22 18:32:41 +02009068 if (state_filter_match(state_filter, p))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01009069 sched_show_task(p);
Oleg Nesterov5d07f422014-08-13 21:19:53 +02009070 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009071
Ingo Molnardd41f592007-07-09 18:51:59 +02009072#ifdef CONFIG_SCHED_DEBUG
Rabin Vincentfb90a6e2016-04-04 15:42:02 +02009073 if (!state_filter)
9074 sysrq_sched_debug_show();
Ingo Molnardd41f592007-07-09 18:51:59 +02009075#endif
Thomas Gleixner510f5ac2011-07-17 20:47:54 +02009076 rcu_read_unlock();
Ingo Molnare59e2ae2006-12-06 20:35:59 -08009077 /*
9078 * Only show locks if all tasks are dumped:
9079 */
Shmulik Ladkani93335a22009-11-25 15:23:41 +02009080 if (!state_filter)
Ingo Molnare59e2ae2006-12-06 20:35:59 -08009081 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009082}
9083
Ingo Molnarf340c0d2005-06-28 16:40:42 +02009084/**
9085 * init_idle - set up an idle thread for a given CPU
9086 * @idle: task in question
Ingo Molnard1ccc662017-02-01 11:46:42 +01009087 * @cpu: CPU the idle task belongs to
Ingo Molnarf340c0d2005-06-28 16:40:42 +02009088 *
9089 * NOTE: this function does not set the idle thread's NEED_RESCHED
9090 * flag, to make booting more robust.
9091 */
Valentin Schneiderf1a0a372021-05-12 10:46:36 +01009092void __init init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009093{
Waiman Long713a2e22022-09-22 14:00:40 -04009094#ifdef CONFIG_SMP
9095 struct affinity_context ac = (struct affinity_context) {
9096 .new_mask = cpumask_of(cpu),
9097 .flags = 0,
9098 };
9099#endif
Ingo Molnar70b97a72006-07-03 00:25:42 -07009100 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009101 unsigned long flags;
9102
Peter Zijlstraff51ff82019-10-01 11:18:37 +02009103 __sched_fork(0, idle);
9104
Peter Zijlstra25834c72015-05-15 17:43:34 +02009105 raw_spin_lock_irqsave(&idle->pi_lock, flags);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009106 raw_spin_rq_lock(rq);
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01009107
Peter Zijlstra2f064a52021-06-11 10:28:17 +02009108 idle->__state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02009109 idle->se.exec_start = sched_clock();
Valentin Schneider00b89fe2021-05-10 16:10:23 +01009110 /*
9111 * PF_KTHREAD should already be set at this point; regardless, make it
9112 * look like a proper per-CPU kthread.
9113 */
9114 idle->flags |= PF_IDLE | PF_KTHREAD | PF_NO_SETAFFINITY;
9115 kthread_set_per_cpu(idle, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02009116
Peter Zijlstrade9b8f52015-08-13 23:09:29 +02009117#ifdef CONFIG_SMP
9118 /*
Tal Zussmanb19a8882020-11-12 19:51:56 -05009119 * It's possible that init_idle() gets called multiple times on a task,
Peter Zijlstrade9b8f52015-08-13 23:09:29 +02009120 * in that case do_set_cpus_allowed() will not do the right thing.
9121 *
9122 * And since this is boot we can forgo the serialization.
9123 */
Waiman Long713a2e22022-09-22 14:00:40 -04009124 set_cpus_allowed_common(idle, &ac);
Peter Zijlstrade9b8f52015-08-13 23:09:29 +02009125#endif
Peter Zijlstra6506cf6c2010-09-16 17:50:31 +02009126 /*
9127 * We're having a chicken and egg problem, even though we are
Ingo Molnard1ccc662017-02-01 11:46:42 +01009128 * holding rq->lock, the CPU isn't yet set to this CPU so the
Peter Zijlstra6506cf6c2010-09-16 17:50:31 +02009129 * lockdep check in task_group() will fail.
9130 *
9131 * Similar case to sched_fork(). / Alternatively we could
9132 * use task_rq_lock() here and obtain the other rq->lock.
9133 *
9134 * Silence PROVE_RCU
9135 */
9136 rcu_read_lock();
Ingo Molnardd41f592007-07-09 18:51:59 +02009137 __set_task_cpu(idle, cpu);
Peter Zijlstra6506cf6c2010-09-16 17:50:31 +02009138 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009139
Eric W. Biederman5311a982019-09-14 07:35:02 -05009140 rq->idle = idle;
9141 rcu_assign_pointer(rq->curr, idle);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009142 idle->on_rq = TASK_ON_RQ_QUEUED;
Peter Zijlstrade9b8f52015-08-13 23:09:29 +02009143#ifdef CONFIG_SMP
Peter Zijlstra3ca7a442011-04-05 17:23:40 +02009144 idle->on_cpu = 1;
Nick Piggin4866cde2005-06-25 14:57:23 -07009145#endif
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009146 raw_spin_rq_unlock(rq);
Peter Zijlstra25834c72015-05-15 17:43:34 +02009147 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009148
9149 /* Set the preempt count _outside_ the spinlocks! */
Peter Zijlstra01028742013-08-14 14:55:46 +02009150 init_idle_preempt_count(idle, cpu);
Jonathan Corbet625f2a32011-04-22 11:19:10 -06009151
Ingo Molnardd41f592007-07-09 18:51:59 +02009152 /*
9153 * The idle tasks have their own, simple scheduling class:
9154 */
9155 idle->sched_class = &idle_sched_class;
Steven Rostedt868baf02011-02-10 21:26:13 -05009156 ftrace_graph_init_idle_task(idle, cpu);
Frederic Weisbecker45eacc62013-05-15 22:16:32 +02009157 vtime_init_idle(idle, cpu);
Peter Zijlstrade9b8f52015-08-13 23:09:29 +02009158#ifdef CONFIG_SMP
Carsten Emdef1c6f1a2011-10-26 23:14:16 +02009159 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
9160#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009161}
9162
Nicolas Pitree1d4eee2017-06-14 13:19:23 -04009163#ifdef CONFIG_SMP
9164
Juri Lellif82f8042014-10-07 09:52:11 +01009165int cpuset_cpumask_can_shrink(const struct cpumask *cur,
9166 const struct cpumask *trial)
9167{
Nicolas Pitre06a76fe2017-06-21 14:22:01 -04009168 int ret = 1;
Juri Lellif82f8042014-10-07 09:52:11 +01009169
Yury Norov1087ad42022-02-10 14:49:06 -08009170 if (cpumask_empty(cur))
Mike Galbraithbb2bc552015-01-28 04:53:55 +01009171 return ret;
9172
Nicolas Pitre06a76fe2017-06-21 14:22:01 -04009173 ret = dl_cpuset_cpumask_can_shrink(cur, trial);
Juri Lellif82f8042014-10-07 09:52:11 +01009174
9175 return ret;
9176}
9177
Juri Lelli7f514122014-09-19 10:22:40 +01009178int task_can_attach(struct task_struct *p,
Waiman Longb6e8d402022-08-02 21:54:51 -04009179 const struct cpumask *cs_effective_cpus)
Juri Lelli7f514122014-09-19 10:22:40 +01009180{
9181 int ret = 0;
9182
9183 /*
9184 * Kthreads which disallow setaffinity shouldn't be moved
Ingo Molnard1ccc662017-02-01 11:46:42 +01009185 * to a new cpuset; we don't want to change their CPU
Juri Lelli7f514122014-09-19 10:22:40 +01009186 * affinity and isolating such threads by their set of
9187 * allowed nodes is unnecessary. Thus, cpusets are not
9188 * applicable for such threads. This prevents checking for
9189 * success of set_cpus_allowed_ptr() on all attached tasks
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02009190 * before cpus_mask may be changed.
Juri Lelli7f514122014-09-19 10:22:40 +01009191 */
9192 if (p->flags & PF_NO_SETAFFINITY) {
9193 ret = -EINVAL;
9194 goto out;
9195 }
9196
Juri Lelli7f514122014-09-19 10:22:40 +01009197 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
Waiman Longb6e8d402022-08-02 21:54:51 -04009198 cs_effective_cpus)) {
9199 int cpu = cpumask_any_and(cpu_active_mask, cs_effective_cpus);
Dietmar Eggemann772b65392022-03-02 19:34:30 +01009200
Waiman Longb6e8d402022-08-02 21:54:51 -04009201 if (unlikely(cpu >= nr_cpu_ids))
9202 return -EINVAL;
Dietmar Eggemann772b65392022-03-02 19:34:30 +01009203 ret = dl_cpu_busy(cpu, p);
9204 }
Juri Lelli7f514122014-09-19 10:22:40 +01009205
Juri Lelli7f514122014-09-19 10:22:40 +01009206out:
9207 return ret;
9208}
9209
Ingo Molnarf2cb1362017-02-01 13:10:18 +01009210bool sched_smp_initialized __read_mostly;
Thomas Gleixnere26fbff2016-03-10 12:54:10 +01009211
Mel Gormane6628d52013-10-07 11:29:02 +01009212#ifdef CONFIG_NUMA_BALANCING
9213/* Migrate current task p to target_cpu */
9214int migrate_task_to(struct task_struct *p, int target_cpu)
9215{
9216 struct migration_arg arg = { p, target_cpu };
9217 int curr_cpu = task_cpu(p);
9218
9219 if (curr_cpu == target_cpu)
9220 return 0;
9221
Sebastian Andrzej Siewior3bd370622019-04-23 16:26:36 +02009222 if (!cpumask_test_cpu(target_cpu, p->cpus_ptr))
Mel Gormane6628d52013-10-07 11:29:02 +01009223 return -EINVAL;
9224
9225 /* TODO: This is not properly updating schedstats */
9226
Mel Gorman286549d2014-01-21 15:51:03 -08009227 trace_sched_move_numa(p, curr_cpu, target_cpu);
Mel Gormane6628d52013-10-07 11:29:02 +01009228 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
9229}
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009230
9231/*
9232 * Requeue a task on a given node and accurately track the number of NUMA
9233 * tasks on the runqueues
9234 */
9235void sched_setnuma(struct task_struct *p, int nid)
9236{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009237 bool queued, running;
Peter Zijlstraeb580752015-07-31 21:28:18 +02009238 struct rq_flags rf;
9239 struct rq *rq;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009240
Peter Zijlstraeb580752015-07-31 21:28:18 +02009241 rq = task_rq_lock(p, &rf);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009242 queued = task_on_rq_queued(p);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009243 running = task_current(rq, p);
9244
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009245 if (queued)
Peter Zijlstra1de64442015-09-30 17:44:13 +02009246 dequeue_task(rq, p, DEQUEUE_SAVE);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009247 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04009248 put_prev_task(rq, p);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009249
9250 p->numa_preferred_nid = nid;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009251
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04009252 if (queued)
Peter Zijlstra7134b3e2017-02-21 14:23:38 +01009253 enqueue_task(rq, p, ENQUEUE_RESTORE | ENQUEUE_NOCLOCK);
Vincent Guittota399d232016-09-12 09:47:52 +02009254 if (running)
Peter Zijlstra03b7fad2019-05-29 20:36:41 +00009255 set_next_task(rq, p);
Peter Zijlstraeb580752015-07-31 21:28:18 +02009256 task_rq_unlock(rq, p, &rf);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01009257}
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02009258#endif /* CONFIG_NUMA_BALANCING */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009259
9260#ifdef CONFIG_HOTPLUG_CPU
Ingo Molnar48f24c42006-07-03 00:25:40 -07009261/*
Ingo Molnard1ccc662017-02-01 11:46:42 +01009262 * Ensure that the idle task is using init_mm right before its CPU goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07009263 * offline.
9264 */
9265void idle_task_exit(void)
9266{
9267 struct mm_struct *mm = current->active_mm;
9268
9269 BUG_ON(cpu_online(smp_processor_id()));
Peter Zijlstrabf2c59f2020-04-01 17:40:33 -04009270 BUG_ON(current != this_rq()->idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009271
Martin Schwidefskya53efe52012-10-26 17:17:44 +02009272 if (mm != &init_mm) {
Andy Lutomirski252d2a42017-06-09 11:49:15 -07009273 switch_mm(mm, &init_mm, current);
Martin Schwidefskya53efe52012-10-26 17:17:44 +02009274 finish_arch_post_lock_switch();
9275 }
Peter Zijlstrabf2c59f2020-04-01 17:40:33 -04009276
9277 /* finish_cpu(), as ran on the BP, will clean up the active_mm state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07009278}
9279
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009280static int __balance_push_cpu_stop(void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009281{
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009282 struct task_struct *p = arg;
9283 struct rq *rq = this_rq();
9284 struct rq_flags rf;
9285 int cpu;
Peter Zijlstra48c5ccae2010-11-13 19:32:29 +01009286
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009287 raw_spin_lock_irq(&p->pi_lock);
9288 rq_lock(rq, &rf);
Peter Zijlstra10e70712019-08-06 15:13:17 +02009289
Frederic Weisbecker77bd3972013-04-12 01:50:58 +02009290 update_rq_clock(rq);
9291
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009292 if (task_rq(p) == rq && task_on_rq_queued(p)) {
9293 cpu = select_fallback_rq(rq->cpu, p);
9294 rq = __migrate_task(rq, &rf, p, cpu);
Peter Zijlstra48c5ccae2010-11-13 19:32:29 +01009295 }
9296
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009297 rq_unlock(rq, &rf);
9298 raw_spin_unlock_irq(&p->pi_lock);
9299
9300 put_task_struct(p);
9301
9302 return 0;
Peter Zijlstra48c5ccae2010-11-13 19:32:29 +01009303}
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009304
9305static DEFINE_PER_CPU(struct cpu_stop_work, push_work);
9306
9307/*
9308 * Ensure we only run per-cpu kthreads once the CPU goes !active.
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009309 *
9310 * This is enabled below SCHED_AP_ACTIVE; when !cpu_active(), but only
9311 * effective when the hotplug motion is down.
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009312 */
9313static void balance_push(struct rq *rq)
9314{
9315 struct task_struct *push_task = rq->curr;
9316
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009317 lockdep_assert_rq_held(rq);
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009318
Peter Zijlstraae792702020-12-10 17:14:08 +01009319 /*
9320 * Ensure the thing is persistent until balance_push_set(.on = false);
9321 */
9322 rq->balance_callback = &balance_push_callback;
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009323
9324 /*
Thomas Gleixner868ad332021-08-28 15:55:52 +02009325 * Only active while going offline and when invoked on the outgoing
9326 * CPU.
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009327 */
Thomas Gleixner868ad332021-08-28 15:55:52 +02009328 if (!cpu_dying(rq->cpu) || rq != this_rq())
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009329 return;
9330
9331 /*
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009332 * Both the cpu-hotplug and stop task are in this case and are
9333 * required to complete the hotplug process.
9334 */
Valentin Schneider00b89fe2021-05-10 16:10:23 +01009335 if (kthread_is_per_cpu(push_task) ||
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01009336 is_migration_disabled(push_task)) {
9337
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009338 /*
9339 * If this is the idle task on the outgoing CPU try to wake
9340 * up the hotplug control thread which might wait for the
9341 * last task to vanish. The rcuwait_active() check is
9342 * accurate here because the waiter is pinned on this CPU
9343 * and can't obviously be running in parallel.
Thomas Gleixner3015ef42020-08-26 14:08:10 +02009344 *
9345 * On RT kernels this also has to check whether there are
9346 * pinned and scheduled out tasks on the runqueue. They
9347 * need to leave the migrate disabled section first.
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009348 */
Thomas Gleixner3015ef42020-08-26 14:08:10 +02009349 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
9350 rcuwait_active(&rq->hotplug_wait)) {
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009351 raw_spin_rq_unlock(rq);
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009352 rcuwait_wake_up(&rq->hotplug_wait);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009353 raw_spin_rq_lock(rq);
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009354 }
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009355 return;
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009356 }
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009357
9358 get_task_struct(push_task);
9359 /*
9360 * Temporarily drop rq->lock such that we can wake-up the stop task.
9361 * Both preemption and IRQs are still disabled.
9362 */
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009363 raw_spin_rq_unlock(rq);
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009364 stop_one_cpu_nowait(rq->cpu, __balance_push_cpu_stop, push_task,
9365 this_cpu_ptr(&push_work));
9366 /*
9367 * At this point need_resched() is true and we'll take the loop in
9368 * schedule(). The next pick is obviously going to be the stop task
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01009369 * which kthread_is_per_cpu() and will push this task away.
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009370 */
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009371 raw_spin_rq_lock(rq);
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009372}
9373
9374static void balance_push_set(int cpu, bool on)
9375{
9376 struct rq *rq = cpu_rq(cpu);
9377 struct rq_flags rf;
9378
9379 rq_lock_irqsave(rq, &rf);
Peter Zijlstra22f667c2021-01-15 18:17:45 +01009380 if (on) {
9381 WARN_ON_ONCE(rq->balance_callback);
Peter Zijlstraae792702020-12-10 17:14:08 +01009382 rq->balance_callback = &balance_push_callback;
Peter Zijlstra22f667c2021-01-15 18:17:45 +01009383 } else if (rq->balance_callback == &balance_push_callback) {
Peter Zijlstraae792702020-12-10 17:14:08 +01009384 rq->balance_callback = NULL;
Peter Zijlstra22f667c2021-01-15 18:17:45 +01009385 }
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009386 rq_unlock_irqrestore(rq, &rf);
9387}
9388
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009389/*
9390 * Invoked from a CPUs hotplug control thread after the CPU has been marked
9391 * inactive. All tasks which are not per CPU kernel threads are either
9392 * pushed off this CPU now via balance_push() or placed on a different CPU
9393 * during wakeup. Wait until the CPU is quiescent.
9394 */
9395static void balance_hotplug_wait(void)
9396{
9397 struct rq *rq = this_rq();
9398
Thomas Gleixner3015ef42020-08-26 14:08:10 +02009399 rcuwait_wait_event(&rq->hotplug_wait,
9400 rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009401 TASK_UNINTERRUPTIBLE);
9402}
9403
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009404#else
9405
9406static inline void balance_push(struct rq *rq)
9407{
9408}
9409
9410static inline void balance_push_set(int cpu, bool on)
9411{
9412}
9413
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009414static inline void balance_hotplug_wait(void)
9415{
9416}
9417
Linus Torvalds1da177e2005-04-16 15:20:36 -07009418#endif /* CONFIG_HOTPLUG_CPU */
9419
Ingo Molnarf2cb1362017-02-01 13:10:18 +01009420void set_rq_online(struct rq *rq)
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009421{
9422 if (!rq->online) {
9423 const struct sched_class *class;
9424
Rusty Russellc6c49272008-11-25 02:35:05 +10309425 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009426 rq->online = 1;
9427
9428 for_each_class(class) {
9429 if (class->rq_online)
9430 class->rq_online(rq);
9431 }
9432 }
9433}
9434
Ingo Molnarf2cb1362017-02-01 13:10:18 +01009435void set_rq_offline(struct rq *rq)
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009436{
9437 if (rq->online) {
9438 const struct sched_class *class;
9439
9440 for_each_class(class) {
9441 if (class->rq_offline)
9442 class->rq_offline(rq);
9443 }
9444
Rusty Russellc6c49272008-11-25 02:35:05 +10309445 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009446 rq->online = 0;
9447 }
9448}
9449
Ingo Molnard1ccc662017-02-01 11:46:42 +01009450/*
9451 * used to mark begin/end of suspend/resume:
9452 */
9453static int num_cpus_frozen;
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05309454
Linus Torvalds1da177e2005-04-16 15:20:36 -07009455/*
Tejun Heo3a101d02010-06-08 21:40:36 +02009456 * Update cpusets according to cpu_active mask. If cpusets are
9457 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
9458 * around partition_sched_domains().
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05309459 *
9460 * If we come here as part of a suspend/resume, don't touch cpusets because we
9461 * want to restore it back to its original state upon resume anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009462 */
Thomas Gleixner40190a72016-03-10 12:54:13 +01009463static void cpuset_cpu_active(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009464{
Thomas Gleixner40190a72016-03-10 12:54:13 +01009465 if (cpuhp_tasks_frozen) {
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05309466 /*
9467 * num_cpus_frozen tracks how many CPUs are involved in suspend
9468 * resume sequence. As long as this is not the last online
9469 * operation in the resume sequence, just build a single sched
9470 * domain, ignoring cpusets.
9471 */
Peter Zijlstra50e76632017-09-07 11:13:38 +02009472 partition_sched_domains(1, NULL, NULL);
9473 if (--num_cpus_frozen)
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009474 return;
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05309475 /*
9476 * This is the last CPU online operation. So fall through and
9477 * restore the original sched domains by considering the
9478 * cpuset configurations.
9479 */
Peter Zijlstra50e76632017-09-07 11:13:38 +02009480 cpuset_force_rebuild();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009481 }
Rakib Mullick30e03ac2017-04-09 07:36:14 +06009482 cpuset_update_active_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009483}
Tejun Heo3a101d02010-06-08 21:40:36 +02009484
Thomas Gleixner40190a72016-03-10 12:54:13 +01009485static int cpuset_cpu_inactive(unsigned int cpu)
Tejun Heo3a101d02010-06-08 21:40:36 +02009486{
Thomas Gleixner40190a72016-03-10 12:54:13 +01009487 if (!cpuhp_tasks_frozen) {
Dietmar Eggemann772b65392022-03-02 19:34:30 +01009488 int ret = dl_cpu_busy(cpu, NULL);
9489
9490 if (ret)
9491 return ret;
Rakib Mullick30e03ac2017-04-09 07:36:14 +06009492 cpuset_update_active_cpus();
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009493 } else {
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05309494 num_cpus_frozen++;
9495 partition_sched_domains(1, NULL, NULL);
Tejun Heo3a101d02010-06-08 21:40:36 +02009496 }
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009497 return 0;
Tejun Heo3a101d02010-06-08 21:40:36 +02009498}
Max Krasnyanskye761b772008-07-15 04:43:49 -07009499
Thomas Gleixner40190a72016-03-10 12:54:13 +01009500int sched_cpu_activate(unsigned int cpu)
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009501{
Thomas Gleixner7d976692016-03-10 12:54:17 +01009502 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009503 struct rq_flags rf;
Thomas Gleixner7d976692016-03-10 12:54:17 +01009504
Peter Zijlstra22f667c2021-01-15 18:17:45 +01009505 /*
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009506 * Clear the balance_push callback and prepare to schedule
9507 * regular tasks.
Peter Zijlstra22f667c2021-01-15 18:17:45 +01009508 */
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009509 balance_push_set(cpu, false);
9510
Peter Zijlstraba2591a2018-05-29 16:43:46 +02009511#ifdef CONFIG_SCHED_SMT
9512 /*
Peter Zijlstra (Intel)c5511d02018-11-25 19:33:36 +01009513 * When going up, increment the number of cores with SMT present.
Peter Zijlstraba2591a2018-05-29 16:43:46 +02009514 */
Peter Zijlstra (Intel)c5511d02018-11-25 19:33:36 +01009515 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9516 static_branch_inc_cpuslocked(&sched_smt_present);
Peter Zijlstraba2591a2018-05-29 16:43:46 +02009517#endif
Thomas Gleixner40190a72016-03-10 12:54:13 +01009518 set_cpu_active(cpu, true);
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009519
Thomas Gleixner40190a72016-03-10 12:54:13 +01009520 if (sched_smp_initialized) {
Huang Ying0fb3978b2022-02-14 20:15:52 +08009521 sched_update_numa(cpu, true);
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009522 sched_domains_numa_masks_set(cpu);
Thomas Gleixner40190a72016-03-10 12:54:13 +01009523 cpuset_cpu_active();
Nick Piggin5c1e1762006-10-03 01:14:04 -07009524 }
Thomas Gleixner7d976692016-03-10 12:54:17 +01009525
9526 /*
9527 * Put the rq online, if not already. This happens:
9528 *
9529 * 1) In the early boot process, because we build the real domains
Ingo Molnard1ccc662017-02-01 11:46:42 +01009530 * after all CPUs have been brought up.
Thomas Gleixner7d976692016-03-10 12:54:17 +01009531 *
9532 * 2) At runtime, if cpuset_cpu_active() fails to rebuild the
9533 * domains.
9534 */
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009535 rq_lock_irqsave(rq, &rf);
Thomas Gleixner7d976692016-03-10 12:54:17 +01009536 if (rq->rd) {
9537 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9538 set_rq_online(rq);
9539 }
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009540 rq_unlock_irqrestore(rq, &rf);
Thomas Gleixner7d976692016-03-10 12:54:17 +01009541
Thomas Gleixner40190a72016-03-10 12:54:13 +01009542 return 0;
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009543}
9544
Thomas Gleixner40190a72016-03-10 12:54:13 +01009545int sched_cpu_deactivate(unsigned int cpu)
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009546{
Peter Zijlstra120455c52020-09-25 16:42:31 +02009547 struct rq *rq = cpu_rq(cpu);
9548 struct rq_flags rf;
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009549 int ret;
9550
Anna-Maria Behnsene0b257c2020-12-15 11:44:00 +01009551 /*
9552 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
9553 * load balancing when not active
9554 */
9555 nohz_balance_exit_idle(rq);
9556
Thomas Gleixner40190a72016-03-10 12:54:13 +01009557 set_cpu_active(cpu, false);
Peter Zijlstra741ba802021-01-16 11:56:37 +01009558
9559 /*
9560 * From this point forward, this CPU will refuse to run any task that
9561 * is not: migrate_disable() or KTHREAD_IS_PER_CPU, and will actively
9562 * push those tasks away until this gets cleared, see
9563 * sched_cpu_dying().
9564 */
Peter Zijlstra975707f2021-01-20 15:05:41 +01009565 balance_push_set(cpu, true);
9566
Peter Zijlstrab2454ca2016-03-10 12:54:14 +01009567 /*
Peter Zijlstra975707f2021-01-20 15:05:41 +01009568 * We've cleared cpu_active_mask / set balance_push, wait for all
9569 * preempt-disabled and RCU users of this state to go away such that
9570 * all new such users will observe it.
Peter Zijlstrab2454ca2016-03-10 12:54:14 +01009571 *
Peter Zijlstra5ba2ffba2021-01-12 11:28:16 +01009572 * Specifically, we rely on ttwu to no longer target this CPU, see
9573 * ttwu_queue_cond() and is_cpu_allowed().
9574 *
Peter Zijlstrab2454ca2016-03-10 12:54:14 +01009575 * Do sync before park smpboot threads to take care the rcu boost case.
9576 */
Paul E. McKenney309ba852018-07-11 14:36:49 -07009577 synchronize_rcu();
Thomas Gleixner40190a72016-03-10 12:54:13 +01009578
Peter Zijlstra120455c52020-09-25 16:42:31 +02009579 rq_lock_irqsave(rq, &rf);
9580 if (rq->rd) {
9581 update_rq_clock(rq);
9582 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
9583 set_rq_offline(rq);
9584 }
9585 rq_unlock_irqrestore(rq, &rf);
9586
Peter Zijlstra (Intel)c5511d02018-11-25 19:33:36 +01009587#ifdef CONFIG_SCHED_SMT
9588 /*
9589 * When going down, decrement the number of cores with SMT present.
9590 */
9591 if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
9592 static_branch_dec_cpuslocked(&sched_smt_present);
Peter Zijlstra3c474b32021-08-19 13:09:17 +02009593
9594 sched_core_cpu_deactivate(cpu);
Peter Zijlstra (Intel)c5511d02018-11-25 19:33:36 +01009595#endif
9596
Thomas Gleixner40190a72016-03-10 12:54:13 +01009597 if (!sched_smp_initialized)
9598 return 0;
9599
Huang Ying0fb3978b2022-02-14 20:15:52 +08009600 sched_update_numa(cpu, false);
Thomas Gleixner40190a72016-03-10 12:54:13 +01009601 ret = cpuset_cpu_inactive(cpu);
9602 if (ret) {
Peter Zijlstra2558aac2020-09-11 09:54:27 +02009603 balance_push_set(cpu, false);
Thomas Gleixner40190a72016-03-10 12:54:13 +01009604 set_cpu_active(cpu, true);
Huang Ying0fb3978b2022-02-14 20:15:52 +08009605 sched_update_numa(cpu, true);
Thomas Gleixner40190a72016-03-10 12:54:13 +01009606 return ret;
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009607 }
Thomas Gleixner40190a72016-03-10 12:54:13 +01009608 sched_domains_numa_masks_clear(cpu);
9609 return 0;
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009610}
9611
Thomas Gleixner94baf7a2016-03-10 12:54:15 +01009612static void sched_rq_cpu_starting(unsigned int cpu)
9613{
9614 struct rq *rq = cpu_rq(cpu);
9615
9616 rq->calc_load_update = calc_load_update;
Thomas Gleixner94baf7a2016-03-10 12:54:15 +01009617 update_max_interval();
9618}
9619
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009620int sched_cpu_starting(unsigned int cpu)
9621{
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05009622 sched_core_cpu_starting(cpu);
Thomas Gleixner94baf7a2016-03-10 12:54:15 +01009623 sched_rq_cpu_starting(cpu);
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01009624 sched_tick_start(cpu);
Thomas Gleixner135fb3e2016-03-10 12:54:11 +01009625 return 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009626}
9627
Thomas Gleixnerf2785dd2016-03-10 12:54:18 +01009628#ifdef CONFIG_HOTPLUG_CPU
Thomas Gleixner1cf12e02020-09-16 09:27:18 +02009629
9630/*
9631 * Invoked immediately before the stopper thread is invoked to bring the
9632 * CPU down completely. At this point all per CPU kthreads except the
9633 * hotplug thread (current) and the stopper thread (inactive) have been
9634 * either parked or have been unbound from the outgoing CPU. Ensure that
9635 * any of those which might be on the way out are gone.
9636 *
9637 * If after this point a bound task is being woken on this CPU then the
9638 * responsible hotplug callback has failed to do it's job.
9639 * sched_cpu_dying() will catch it with the appropriate fireworks.
9640 */
9641int sched_cpu_wait_empty(unsigned int cpu)
9642{
9643 balance_hotplug_wait();
9644 return 0;
9645}
9646
9647/*
9648 * Since this CPU is going 'away' for a while, fold any nr_active delta we
9649 * might have. Called from the CPU stopper task after ensuring that the
9650 * stopper is the last running task on the CPU, so nr_active count is
9651 * stable. We need to take the teardown thread which is calling this into
9652 * account, so we hand in adjust = 1 to the load calculation.
9653 *
9654 * Also see the comment "Global load-average calculations".
9655 */
9656static void calc_load_migrate(struct rq *rq)
9657{
9658 long delta = calc_load_fold_active(rq, 1);
9659
9660 if (delta)
9661 atomic_long_add(delta, &calc_load_tasks);
9662}
9663
Valentin Schneider36c6e172021-01-13 18:31:41 +00009664static void dump_rq_tasks(struct rq *rq, const char *loglvl)
9665{
9666 struct task_struct *g, *p;
9667 int cpu = cpu_of(rq);
9668
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009669 lockdep_assert_rq_held(rq);
Valentin Schneider36c6e172021-01-13 18:31:41 +00009670
9671 printk("%sCPU%d enqueued tasks (%u total):\n", loglvl, cpu, rq->nr_running);
9672 for_each_process_thread(g, p) {
9673 if (task_cpu(p) != cpu)
9674 continue;
9675
9676 if (!task_on_rq_queued(p))
9677 continue;
9678
9679 printk("%s\tpid: %d, name: %s\n", loglvl, p->pid, p->comm);
9680 }
9681}
9682
Thomas Gleixnerf2785dd2016-03-10 12:54:18 +01009683int sched_cpu_dying(unsigned int cpu)
9684{
9685 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009686 struct rq_flags rf;
Thomas Gleixnerf2785dd2016-03-10 12:54:18 +01009687
9688 /* Handle pending wakeups and then migrate everything off */
Frederic Weisbeckerd84b3132018-02-21 05:17:27 +01009689 sched_tick_stop(cpu);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009690
9691 rq_lock_irqsave(rq, &rf);
Valentin Schneider36c6e172021-01-13 18:31:41 +00009692 if (rq->nr_running != 1 || rq_has_pinned_tasks(rq)) {
9693 WARN(true, "Dying CPU not properly vacated!");
9694 dump_rq_tasks(rq, KERN_WARNING);
9695 }
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +02009696 rq_unlock_irqrestore(rq, &rf);
9697
Thomas Gleixnerf2785dd2016-03-10 12:54:18 +01009698 calc_load_migrate(rq);
9699 update_max_interval();
Thomas Gleixnere5ef27d2016-03-10 12:54:21 +01009700 hrtick_clear(rq);
Peter Zijlstra3c474b32021-08-19 13:09:17 +02009701 sched_core_cpu_dying(cpu);
Thomas Gleixnerf2785dd2016-03-10 12:54:18 +01009702 return 0;
9703}
9704#endif
9705
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009706void __init sched_init_smp(void)
9707{
Huang Ying0fb3978b2022-02-14 20:15:52 +08009708 sched_init_numa(NUMA_NO_NODE);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009709
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02009710 /*
9711 * There's no userspace yet to cause hotplug operations; hence all the
Ingo Molnard1ccc662017-02-01 11:46:42 +01009712 * CPU masks are stable and all blatant races in the below code cannot
Valentin Schneiderb5a4e2b2018-12-19 18:23:16 +00009713 * happen.
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02009714 */
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009715 mutex_lock(&sched_domains_mutex);
Peter Zijlstra8d5dc512017-04-25 15:29:40 +02009716 sched_init_domains(cpu_active_mask);
Christoph Lameter476f3532007-05-06 14:48:58 -07009717 mutex_unlock(&sched_domains_mutex);
Mike Travis434d53b2008-04-04 18:11:04 -07009718
Ingo Molnardd41f592007-07-09 18:51:59 +02009719 /* Move init over to a non-isolated CPU */
Frederic Weisbecker04d4e662022-02-07 16:59:06 +01009720 if (set_cpus_allowed_ptr(current, housekeeping_cpumask(HK_TYPE_DOMAIN)) < 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02009721 BUG();
Peter Zijlstra15faafc2021-05-31 12:21:13 +02009722 current->flags &= ~PF_NO_SETAFFINITY;
Ingo Molnardd41f592007-07-09 18:51:59 +02009723 sched_init_granularity();
Rusty Russell42128232008-11-25 02:35:12 +10309724
Rusty Russell0e3900e2008-11-25 02:35:13 +10309725 init_sched_rt_class();
Juri Lelli1baca4c2013-11-07 14:43:38 +01009726 init_sched_dl_class();
Peter Zijlstra1b568f02016-05-09 10:38:41 +02009727
Thomas Gleixnere26fbff2016-03-10 12:54:10 +01009728 sched_smp_initialized = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009729}
Thomas Gleixnere26fbff2016-03-10 12:54:10 +01009730
9731static int __init migration_init(void)
9732{
Nicholas Piggin77a53522019-04-11 13:34:44 +10009733 sched_cpu_starting(smp_processor_id());
Thomas Gleixnere26fbff2016-03-10 12:54:10 +01009734 return 0;
9735}
9736early_initcall(migration_init);
9737
Christoph Lameter476f3532007-05-06 14:48:58 -07009738#else
Ingo Molnardd41f592007-07-09 18:51:59 +02009739void __init sched_init_smp(void)
9740{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009741 sched_init_granularity();
9742}
Peter Williams2dd73a42006-06-27 02:54:34 -07009743#endif /* CONFIG_SMP */
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009744
Avi Kivitye107be32007-07-26 13:40:43 +02009745int in_sched_functions(unsigned long addr)
9746{
9747 return in_lock_functions(addr) ||
9748 (addr >= (unsigned long)__sched_text_start
Christoph Lameterc9819f42006-12-10 02:20:25 -08009749 && addr < (unsigned long)__sched_text_end);
Christoph Lameter476f3532007-05-06 14:48:58 -07009750}
Christoph Lameterc9819f42006-12-10 02:20:25 -08009751
Peter Zijlstra029632f2011-10-25 10:00:11 +02009752#ifdef CONFIG_CGROUP_SCHED
Li Zefan27b4b932013-03-05 16:07:52 +08009753/*
9754 * Default task group.
9755 * Every task in system belongs to this group at bootup.
9756 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02009757struct task_group root_task_group;
Mike Galbraith35cf4e52012-08-07 05:00:13 +02009758LIST_HEAD(task_groups);
Waiman Longb0367622015-12-02 13:41:49 -05009759
9760/* Cacheline aligned slab cache for task_group */
9761static struct kmem_cache *task_group_cache __read_mostly;
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009762#endif
9763
Linus Torvalds1da177e2005-04-16 15:20:36 -07009764void __init sched_init(void)
9765{
Qian Caia1dc0442019-07-19 21:23:19 -04009766 unsigned long ptr = 0;
Dietmar Eggemann55627e32019-05-27 07:21:13 +01009767 int i;
Mike Travis434d53b2008-04-04 18:11:04 -07009768
Steven Rostedt (VMware)c3a340f2019-12-19 16:44:53 -05009769 /* Make sure the linker didn't screw up */
Peter Zijlstra546a3fe2022-05-17 13:46:54 +02009770 BUG_ON(&idle_sched_class != &fair_sched_class + 1 ||
9771 &fair_sched_class != &rt_sched_class + 1 ||
9772 &rt_sched_class != &dl_sched_class + 1);
Steven Rostedt (VMware)c3a340f2019-12-19 16:44:53 -05009773#ifdef CONFIG_SMP
Peter Zijlstra546a3fe2022-05-17 13:46:54 +02009774 BUG_ON(&dl_sched_class != &stop_sched_class + 1);
Steven Rostedt (VMware)c3a340f2019-12-19 16:44:53 -05009775#endif
9776
Ingo Molnar5822a452017-03-05 13:09:07 +01009777 wait_bit_init();
Linus Torvalds9dcb8b62016-10-26 10:15:30 -07009778
Mike Travis434d53b2008-04-04 18:11:04 -07009779#ifdef CONFIG_FAIR_GROUP_SCHED
Qian Caia1dc0442019-07-19 21:23:19 -04009780 ptr += 2 * nr_cpu_ids * sizeof(void **);
Mike Travis434d53b2008-04-04 18:11:04 -07009781#endif
9782#ifdef CONFIG_RT_GROUP_SCHED
Qian Caia1dc0442019-07-19 21:23:19 -04009783 ptr += 2 * nr_cpu_ids * sizeof(void **);
Mike Travis434d53b2008-04-04 18:11:04 -07009784#endif
Qian Caia1dc0442019-07-19 21:23:19 -04009785 if (ptr) {
9786 ptr = (unsigned long)kzalloc(ptr, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07009787
9788#ifdef CONFIG_FAIR_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08009789 root_task_group.se = (struct sched_entity **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07009790 ptr += nr_cpu_ids * sizeof(void **);
9791
Yong Zhang07e06b02011-01-07 15:17:36 +08009792 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07009793 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009794
Wei Yangb1d17792020-04-23 21:44:43 +00009795 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
9796 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009797#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009798#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08009799 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07009800 ptr += nr_cpu_ids * sizeof(void **);
9801
Yong Zhang07e06b02011-01-07 15:17:36 +08009802 root_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009803 ptr += nr_cpu_ids * sizeof(void **);
9804
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009805#endif /* CONFIG_RT_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009807
Ingo Molnard1ccc662017-02-01 11:46:42 +01009808 init_rt_bandwidth(&def_rt_bandwidth, global_rt_period(), global_rt_runtime());
Dario Faggioli332ac172013-11-07 14:43:45 +01009809
Gregory Haskins57d885f2008-01-25 21:08:18 +01009810#ifdef CONFIG_SMP
9811 init_defrootdomain();
9812#endif
9813
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009814#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08009815 init_rt_bandwidth(&root_task_group.rt_bandwidth,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009816 global_rt_period(), global_rt_runtime());
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009817#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009818
Dhaval Giani7c941432010-01-20 13:26:18 +01009819#ifdef CONFIG_CGROUP_SCHED
Waiman Longb0367622015-12-02 13:41:49 -05009820 task_group_cache = KMEM_CACHE(task_group, 0);
9821
Yong Zhang07e06b02011-01-07 15:17:36 +08009822 list_add(&root_task_group.list, &task_groups);
9823 INIT_LIST_HEAD(&root_task_group.children);
Glauber Costaf4d6f6c2011-11-01 19:19:07 -02009824 INIT_LIST_HEAD(&root_task_group.siblings);
Mike Galbraith5091faa2010-11-30 14:18:03 +01009825 autogroup_init(&init_task);
Dhaval Giani7c941432010-01-20 13:26:18 +01009826#endif /* CONFIG_CGROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009827
Ingo Molnardd41f592007-07-09 18:51:59 +02009828 for_each_possible_cpu(i) {
Ingo Molnardd41f592007-07-09 18:51:59 +02009829 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009830
9831 rq = cpu_rq(i);
Peter Zijlstra5cb9eaa2020-11-17 18:19:31 -05009832 raw_spin_lock_init(&rq->__lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009833 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009834 rq->calc_load_active = 0;
9835 rq->calc_load_update = jiffies + LOAD_FREQ;
Jan H. Schönherracb5a9b2011-07-14 18:32:43 +02009836 init_cfs_rq(&rq->cfs);
Abel Vesa07c54f72015-03-03 13:50:27 +02009837 init_rt_rq(&rq->rt);
9838 init_dl_rq(&rq->dl);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009839#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009840 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Vincent Guittot9c2791f2016-11-08 10:53:43 +01009841 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009842 /*
Ingo Molnard1ccc662017-02-01 11:46:42 +01009843 * How much CPU bandwidth does root_task_group get?
Dhaval Giani354d60c2008-04-19 19:44:59 +02009844 *
9845 * In case of task-groups formed thr' the cgroup filesystem, it
Ingo Molnard1ccc662017-02-01 11:46:42 +01009846 * gets 100% of the CPU resources in the system. This overall
9847 * system CPU resource is divided among the tasks of
Yong Zhang07e06b02011-01-07 15:17:36 +08009848 * root_task_group and its child task-groups in a fair manner,
Dhaval Giani354d60c2008-04-19 19:44:59 +02009849 * based on each entity's (task or task-group's) weight
9850 * (se->load.weight).
9851 *
Yong Zhang07e06b02011-01-07 15:17:36 +08009852 * In other words, if root_task_group has 10 tasks of weight
Dhaval Giani354d60c2008-04-19 19:44:59 +02009853 * 1024) and two child groups A0 and A1 (of weight 1024 each),
Ingo Molnard1ccc662017-02-01 11:46:42 +01009854 * then A0's share of the CPU resource is:
Dhaval Giani354d60c2008-04-19 19:44:59 +02009855 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009856 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02009857 *
Yong Zhang07e06b02011-01-07 15:17:36 +08009858 * We achieve this by letting root_task_group's tasks sit
9859 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
Dhaval Giani354d60c2008-04-19 19:44:59 +02009860 */
Yong Zhang07e06b02011-01-07 15:17:36 +08009861 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009862#endif /* CONFIG_FAIR_GROUP_SCHED */
9863
9864 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009865#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08009866 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009867#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009868#ifdef CONFIG_SMP
Peter Zijlstraa4c410f2006-12-06 20:37:21 -08009869 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01009870 rq->rd = NULL;
Vincent Guittotca6d75e2015-02-27 16:54:09 +01009871 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009872 rq->balance_callback = &balance_push_callback;
Ingo Molnar3117df02006-12-13 00:34:43 -08009873 rq->active_balance = 0;
9874 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009875 rq->push_cpu = 0;
9876 rq->cpu = i;
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009877 rq->online = 0;
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01009878 rq->idle_stamp = 0;
9879 rq->avg_idle = 2*sysctl_sched_migration_cost;
Peter Zijlstra94aafc32021-06-15 12:16:11 +01009880 rq->wake_stamp = jiffies;
9881 rq->wake_avg_idle = rq->avg_idle;
Jason Low9bd721c2013-09-13 11:26:52 -07009882 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
Peter Zijlstra367456c2012-02-20 21:49:09 +01009883
9884 INIT_LIST_HEAD(&rq->cfs_tasks);
9885
Gregory Haskinsdc938522008-01-25 21:08:26 +01009886 rq_attach_root(rq, &def_root_domain);
Frederic Weisbecker3451d022011-08-10 23:21:01 +02009887#ifdef CONFIG_NO_HZ_COMMON
Peter Zijlstrae022e0d2017-12-21 11:20:23 +01009888 rq->last_blocked_load_update_tick = jiffies;
Peter Zijlstraa22e47a2017-12-21 10:01:24 +01009889 atomic_set(&rq->nohz_flags, 0);
Peter Zijlstra (Intel)90b53632020-03-27 11:44:56 +01009890
Peter Zijlstra545b8c82020-06-15 11:29:31 +02009891 INIT_CSD(&rq->nohz_csd, nohz_csd_func, rq);
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07009892#endif
Thomas Gleixnerf2469a12020-09-14 14:47:28 +02009893#ifdef CONFIG_HOTPLUG_CPU
9894 rcuwait_init(&rq->hotplug_wait);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009895#endif
Frederic Weisbecker9fd81dd2016-04-19 17:36:51 +02009896#endif /* CONFIG_SMP */
Frederic Weisbecker77a021b2018-02-21 05:17:23 +01009897 hrtick_rq_init(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009898 atomic_set(&rq->nr_iowait, 0);
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05009899
9900#ifdef CONFIG_SCHED_CORE
Peter Zijlstra3c474b32021-08-19 13:09:17 +02009901 rq->core = rq;
Peter Zijlstra539f6512020-11-17 18:19:37 -05009902 rq->core_pick = NULL;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05009903 rq->core_enabled = 0;
Peter Zijlstra539f6512020-11-17 18:19:37 -05009904 rq->core_tree = RB_ROOT;
Josh Don4feee7d12021-10-18 13:34:28 -07009905 rq->core_forceidle_count = 0;
9906 rq->core_forceidle_occupation = 0;
9907 rq->core_forceidle_start = 0;
Peter Zijlstra539f6512020-11-17 18:19:37 -05009908
9909 rq->core_cookie = 0UL;
Peter Zijlstra9edeaea2020-11-17 18:19:34 -05009910#endif
Waiman Longda019032022-09-22 14:00:39 -04009911 zalloc_cpumask_var_node(&rq->scratch_mask, GFP_KERNEL, cpu_to_node(i));
Linus Torvalds1da177e2005-04-16 15:20:36 -07009912 }
Ingo Molnara0f98a12007-06-17 18:37:45 +02009913
Peter Zijlstrab1e82062022-02-14 10:16:57 +01009914 set_load_weight(&init_task, false);
Ingo Molnardd41f592007-07-09 18:51:59 +02009915
Linus Torvalds1da177e2005-04-16 15:20:36 -07009916 /*
9917 * The boot idle thread does lazy MMU switching as well:
9918 */
Vegard Nossumf1f10072017-02-27 14:30:07 -08009919 mmgrab(&init_mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009920 enter_lazy_tlb(&init_mm, current);
9921
9922 /*
Eric W. Biederman40966e32021-12-02 09:56:14 -06009923 * The idle task doesn't need the kthread struct to function, but it
9924 * is dressed up as a per-CPU kthread and thus needs to play the part
9925 * if we want to avoid special-casing it in code that deals with per-CPU
9926 * kthreads.
9927 */
Eric W. Biedermandd621ee2021-12-21 11:41:14 -06009928 WARN_ON(!set_kthread_struct(current));
Eric W. Biederman40966e32021-12-02 09:56:14 -06009929
9930 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07009931 * Make us the idle thread. Technically, schedule() should not be
9932 * called from this thread, however somewhere below it might be,
9933 * but because we are the idle thread, we just pick up running again
9934 * when this runqueue becomes "idle".
9935 */
9936 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009937
9938 calc_load_update = jiffies + LOAD_FREQ;
9939
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309940#ifdef CONFIG_SMP
Thomas Gleixner29d5e042012-04-20 13:05:45 +00009941 idle_thread_set_boot_cpu();
Peter Zijlstrab5c44772021-01-21 16:09:32 +01009942 balance_push_set(smp_processor_id(), false);
Peter Zijlstra029632f2011-10-25 10:00:11 +02009943#endif
9944 init_sched_fair_class();
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309945
Johannes Weinereb414682018-10-26 15:06:27 -07009946 psi_init();
9947
Patrick Bellasi69842cb2019-06-21 09:42:02 +01009948 init_uclamp();
9949
Frederic Weisbeckerc597bfd2021-09-14 12:31:34 +02009950 preempt_dynamic_init();
9951
Ingo Molnar6892b752008-02-13 14:02:36 +01009952 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009953}
9954
Frederic Weisbeckerd902db12011-06-08 19:31:56 +02009955#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009956
Thomas Gleixner42a38752021-09-23 18:54:38 +02009957void __might_sleep(const char *file, int line)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009958{
Peter Zijlstrad6c23bb2021-06-11 10:28:14 +02009959 unsigned int state = get_current_state();
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02009960 /*
9961 * Blocking primitives will set (and therefore destroy) current->state,
9962 * since we will exit with TASK_RUNNING make sure we enter with it,
9963 * otherwise we will destroy state.
9964 */
Peter Zijlstrad6c23bb2021-06-11 10:28:14 +02009965 WARN_ONCE(state != TASK_RUNNING && current->task_state_change,
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02009966 "do not call blocking ops when !TASK_RUNNING; "
Peter Zijlstrad6c23bb2021-06-11 10:28:14 +02009967 "state=%x set at [<%p>] %pS\n", state,
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02009968 (void *)current->task_state_change,
Linus Torvalds00845eb2015-02-01 12:23:32 -08009969 (void *)current->task_state_change);
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02009970
Thomas Gleixner42a38752021-09-23 18:54:38 +02009971 __might_resched(file, line, 0);
Peter Zijlstra34274452014-09-24 10:18:56 +02009972}
9973EXPORT_SYMBOL(__might_sleep);
9974
Thomas Gleixner8d713b62021-09-23 18:54:41 +02009975static void print_preempt_disable_ip(int preempt_offset, unsigned long ip)
9976{
9977 if (!IS_ENABLED(CONFIG_DEBUG_PREEMPT))
9978 return;
9979
9980 if (preempt_count() == preempt_offset)
9981 return;
9982
9983 pr_err("Preemption disabled at:");
9984 print_ip_sym(KERN_ERR, ip);
9985}
9986
Thomas Gleixner50e081b2021-09-23 18:54:43 +02009987static inline bool resched_offsets_ok(unsigned int offsets)
9988{
9989 unsigned int nested = preempt_count();
9990
9991 nested += rcu_preempt_depth() << MIGHT_RESCHED_RCU_SHIFT;
9992
9993 return nested == offsets;
9994}
9995
9996void __might_resched(const char *file, int line, unsigned int offsets)
Peter Zijlstra34274452014-09-24 10:18:56 +02009997{
Ingo Molnard1ccc662017-02-01 11:46:42 +01009998 /* Ratelimiting timestamp: */
9999 static unsigned long prev_jiffy;
10000
Vegard Nossumd1c6d142016-07-23 09:46:39 +020010001 unsigned long preempt_disable_ip;
Linus Torvalds1da177e2005-04-16 15:20:36 -070010002
Ingo Molnard1ccc662017-02-01 11:46:42 +010010003 /* WARN_ON_ONCE() by default, no rate limit required: */
10004 rcu_sleep_check();
10005
Thomas Gleixner50e081b2021-09-23 18:54:43 +020010006 if ((resched_offsets_ok(offsets) && !irqs_disabled() &&
Daniel Vetter312364f32019-08-26 22:14:23 +020010007 !is_idle_task(current) && !current->non_block_count) ||
Thomas Gleixner1c3c5ea2017-05-16 20:42:48 +020010008 system_state == SYSTEM_BOOTING || system_state > SYSTEM_RUNNING ||
10009 oops_in_progress)
Ingo Molnaraef745f2008-08-28 11:34:43 +020010010 return;
Thomas Gleixner1c3c5ea2017-05-16 20:42:48 +020010011
Ingo Molnaraef745f2008-08-28 11:34:43 +020010012 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10013 return;
10014 prev_jiffy = jiffies;
10015
Ingo Molnard1ccc662017-02-01 11:46:42 +010010016 /* Save this before calling printk(), since that will clobber it: */
Vegard Nossumd1c6d142016-07-23 09:46:39 +020010017 preempt_disable_ip = get_preempt_disable_ip(current);
10018
Thomas Gleixnera45ed302021-09-23 18:54:40 +020010019 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
10020 file, line);
10021 pr_err("in_atomic(): %d, irqs_disabled(): %d, non_block: %d, pid: %d, name: %s\n",
10022 in_atomic(), irqs_disabled(), current->non_block_count,
10023 current->pid, current->comm);
Thomas Gleixner8d713b62021-09-23 18:54:41 +020010024 pr_err("preempt_count: %x, expected: %x\n", preempt_count(),
Thomas Gleixner50e081b2021-09-23 18:54:43 +020010025 offsets & MIGHT_RESCHED_PREEMPT_MASK);
Thomas Gleixner8d713b62021-09-23 18:54:41 +020010026
10027 if (IS_ENABLED(CONFIG_PREEMPT_RCU)) {
Thomas Gleixner50e081b2021-09-23 18:54:43 +020010028 pr_err("RCU nest depth: %d, expected: %u\n",
10029 rcu_preempt_depth(), offsets >> MIGHT_RESCHED_RCU_SHIFT);
Thomas Gleixner8d713b62021-09-23 18:54:41 +020010030 }
Ingo Molnaraef745f2008-08-28 11:34:43 +020010031
Eric Sandeena8b686b2014-12-16 16:25:28 -060010032 if (task_stack_end_corrupted(current))
Thomas Gleixnera45ed302021-09-23 18:54:40 +020010033 pr_emerg("Thread overran stack, or stack corrupted\n");
Eric Sandeena8b686b2014-12-16 16:25:28 -060010034
Ingo Molnaraef745f2008-08-28 11:34:43 +020010035 debug_show_held_locks(current);
10036 if (irqs_disabled())
10037 print_irqtrace_events(current);
Thomas Gleixner8d713b62021-09-23 18:54:41 +020010038
Thomas Gleixner50e081b2021-09-23 18:54:43 +020010039 print_preempt_disable_ip(offsets & MIGHT_RESCHED_PREEMPT_MASK,
10040 preempt_disable_ip);
Thomas Gleixner8d713b62021-09-23 18:54:41 +020010041
Ingo Molnaraef745f2008-08-28 11:34:43 +020010042 dump_stack();
Vegard Nossumf0b22e32016-07-22 21:46:02 +020010043 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
Linus Torvalds1da177e2005-04-16 15:20:36 -070010044}
Thomas Gleixner874f6702021-09-23 18:54:35 +020010045EXPORT_SYMBOL(__might_resched);
Peter Zijlstra568f1962019-01-28 17:21:52 -080010046
10047void __cant_sleep(const char *file, int line, int preempt_offset)
10048{
10049 static unsigned long prev_jiffy;
10050
10051 if (irqs_disabled())
10052 return;
10053
10054 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10055 return;
10056
10057 if (preempt_count() > preempt_offset)
10058 return;
10059
10060 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10061 return;
10062 prev_jiffy = jiffies;
10063
10064 printk(KERN_ERR "BUG: assuming atomic context at %s:%d\n", file, line);
10065 printk(KERN_ERR "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
10066 in_atomic(), irqs_disabled(),
10067 current->pid, current->comm);
10068
10069 debug_show_held_locks(current);
10070 dump_stack();
10071 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10072}
10073EXPORT_SYMBOL_GPL(__cant_sleep);
Thomas Gleixner74d862b62020-11-18 20:48:42 +010010074
10075#ifdef CONFIG_SMP
10076void __cant_migrate(const char *file, int line)
10077{
10078 static unsigned long prev_jiffy;
10079
10080 if (irqs_disabled())
10081 return;
10082
10083 if (is_migration_disabled(current))
10084 return;
10085
10086 if (!IS_ENABLED(CONFIG_PREEMPT_COUNT))
10087 return;
10088
10089 if (preempt_count() > 0)
10090 return;
10091
10092 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
10093 return;
10094 prev_jiffy = jiffies;
10095
10096 pr_err("BUG: assuming non migratable context at %s:%d\n", file, line);
10097 pr_err("in_atomic(): %d, irqs_disabled(): %d, migration_disabled() %u pid: %d, name: %s\n",
10098 in_atomic(), irqs_disabled(), is_migration_disabled(current),
10099 current->pid, current->comm);
10100
10101 debug_show_held_locks(current);
10102 dump_stack();
10103 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
10104}
10105EXPORT_SYMBOL_GPL(__cant_migrate);
10106#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -070010107#endif
10108
10109#ifdef CONFIG_MAGIC_SYSRQ
10110void normalize_rt_tasks(void)
10111{
10112 struct task_struct *g, *p;
Peter Zijlstradbc7f062015-06-11 14:46:38 +020010113 struct sched_attr attr = {
10114 .sched_policy = SCHED_NORMAL,
10115 };
Linus Torvalds1da177e2005-04-16 15:20:36 -070010116
Oleg Nesterov3472eaa2014-09-21 21:33:38 +020010117 read_lock(&tasklist_lock);
Oleg Nesterov5d07f422014-08-13 21:19:53 +020010118 for_each_process_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +020010119 /*
10120 * Only normalize user tasks:
10121 */
Oleg Nesterov3472eaa2014-09-21 21:33:38 +020010122 if (p->flags & PF_KTHREAD)
Ingo Molnar178be792007-10-15 17:00:18 +020010123 continue;
10124
Josh Poimboeuf4fa8d292016-06-17 12:43:26 -050010125 p->se.exec_start = 0;
Yafang Shaoceeadb82021-09-05 14:35:41 +000010126 schedstat_set(p->stats.wait_start, 0);
10127 schedstat_set(p->stats.sleep_start, 0);
10128 schedstat_set(p->stats.block_start, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +020010129
Dario Faggioliaab03e02013-11-28 11:14:43 +010010130 if (!dl_task(p) && !rt_task(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +020010131 /*
10132 * Renice negative nice level userspace
10133 * tasks back to 0:
10134 */
Oleg Nesterov3472eaa2014-09-21 21:33:38 +020010135 if (task_nice(p) < 0)
Ingo Molnardd41f592007-07-09 18:51:59 +020010136 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -070010137 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +020010138 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070010139
Peter Zijlstradbc7f062015-06-11 14:46:38 +020010140 __sched_setscheduler(p, &attr, false, false);
Oleg Nesterov5d07f422014-08-13 21:19:53 +020010141 }
Oleg Nesterov3472eaa2014-09-21 21:33:38 +020010142 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -070010143}
10144
10145#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -070010146
Jason Wessel67fc4e02010-05-20 21:04:21 -050010147#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
Linus Torvalds1df5c102005-09-12 07:59:21 -070010148/*
Jason Wessel67fc4e02010-05-20 21:04:21 -050010149 * These functions are only useful for the IA64 MCA handling, or kdb.
Linus Torvalds1df5c102005-09-12 07:59:21 -070010150 *
10151 * They can only be called when the whole system has been
10152 * stopped - every CPU needs to be quiescent, and no scheduling
10153 * activity can take place. Using them for anything else would
10154 * be a serious bug, and as a result, they aren't even visible
10155 * under any other configuration.
10156 */
10157
10158/**
Ingo Molnard1ccc662017-02-01 11:46:42 +010010159 * curr_task - return the current task for a given CPU.
Linus Torvalds1df5c102005-09-12 07:59:21 -070010160 * @cpu: the processor in question.
10161 *
10162 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
Yacine Belkadie69f6182013-07-12 20:45:47 +020010163 *
10164 * Return: The current task for @cpu.
Linus Torvalds1df5c102005-09-12 07:59:21 -070010165 */
Ingo Molnar36c8b582006-07-03 00:25:41 -070010166struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -070010167{
10168 return cpu_curr(cpu);
10169}
10170
Jason Wessel67fc4e02010-05-20 21:04:21 -050010171#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
10172
10173#ifdef CONFIG_IA64
Linus Torvalds1df5c102005-09-12 07:59:21 -070010174/**
Peter Zijlstra5feeb782019-05-29 20:36:38 +000010175 * ia64_set_curr_task - set the current task for a given CPU.
Linus Torvalds1df5c102005-09-12 07:59:21 -070010176 * @cpu: the processor in question.
10177 * @p: the task pointer to set.
10178 *
10179 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010180 * are serviced on a separate stack. It allows the architecture to switch the
Ingo Molnard1ccc662017-02-01 11:46:42 +010010181 * notion of the current task on a CPU in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -070010182 * must be called with all CPU's synchronized, and interrupts disabled, the
10183 * and caller must save the original value of the current task (see
10184 * curr_task() above) and restore that value before reenabling interrupts and
10185 * re-starting the system.
10186 *
10187 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
10188 */
Peter Zijlstraa458ae22016-09-20 20:29:40 +020010189void ia64_set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -070010190{
10191 cpu_curr(cpu) = p;
10192}
10193
10194#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010195
Dhaval Giani7c941432010-01-20 13:26:18 +010010196#ifdef CONFIG_CGROUP_SCHED
Peter Zijlstra029632f2011-10-25 10:00:11 +020010197/* task_group_lock serializes the addition/removal of task groups */
10198static DEFINE_SPINLOCK(task_group_lock);
10199
Patrick Bellasi2480c092019-08-22 14:28:06 +010010200static inline void alloc_uclamp_sched_group(struct task_group *tg,
10201 struct task_group *parent)
10202{
10203#ifdef CONFIG_UCLAMP_TASK_GROUP
Patrick Bellasi0413d7f2019-08-22 14:28:11 +010010204 enum uclamp_id clamp_id;
Patrick Bellasi2480c092019-08-22 14:28:06 +010010205
10206 for_each_clamp_id(clamp_id) {
10207 uclamp_se_set(&tg->uclamp_req[clamp_id],
10208 uclamp_none(clamp_id), false);
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010209 tg->uclamp[clamp_id] = parent->uclamp[clamp_id];
Patrick Bellasi2480c092019-08-22 14:28:06 +010010210 }
10211#endif
10212}
10213
Peter Zijlstra2f5177f2016-03-16 16:22:45 +010010214static void sched_free_group(struct task_group *tg)
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010215{
10216 free_fair_sched_group(tg);
10217 free_rt_sched_group(tg);
Mike Galbraithe9aa1dd2011-01-05 11:11:25 +010010218 autogroup_free(tg);
Waiman Longb0367622015-12-02 13:41:49 -050010219 kmem_cache_free(task_group_cache, tg);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010220}
10221
Mathias Krauseb0277892021-11-03 20:06:13 +010010222static void sched_free_group_rcu(struct rcu_head *rcu)
10223{
10224 sched_free_group(container_of(rcu, struct task_group, rcu));
10225}
10226
10227static void sched_unregister_group(struct task_group *tg)
10228{
10229 unregister_fair_sched_group(tg);
10230 unregister_rt_sched_group(tg);
10231 /*
10232 * We have to wait for yet another RCU grace period to expire, as
10233 * print_cfs_stats() might run concurrently.
10234 */
10235 call_rcu(&tg->rcu, sched_free_group_rcu);
10236}
10237
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010238/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010239struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010240{
10241 struct task_group *tg;
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010242
Waiman Longb0367622015-12-02 13:41:49 -050010243 tg = kmem_cache_alloc(task_group_cache, GFP_KERNEL | __GFP_ZERO);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010244 if (!tg)
10245 return ERR_PTR(-ENOMEM);
10246
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010247 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010248 goto err;
10249
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010250 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010251 goto err;
10252
Patrick Bellasi2480c092019-08-22 14:28:06 +010010253 alloc_uclamp_sched_group(tg, parent);
10254
Li Zefanace783b2013-01-24 14:30:48 +080010255 return tg;
10256
10257err:
Peter Zijlstra2f5177f2016-03-16 16:22:45 +010010258 sched_free_group(tg);
Li Zefanace783b2013-01-24 14:30:48 +080010259 return ERR_PTR(-ENOMEM);
10260}
10261
10262void sched_online_group(struct task_group *tg, struct task_group *parent)
10263{
10264 unsigned long flags;
10265
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010266 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010267 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010268
Ingo Molnard1ccc662017-02-01 11:46:42 +010010269 /* Root should already exist: */
10270 WARN_ON(!parent);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010271
10272 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010273 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +080010274 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010275 spin_unlock_irqrestore(&task_group_lock, flags);
Peter Zijlstra8663e242016-06-22 14:58:02 +020010276
10277 online_fair_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010278}
10279
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010280/* rcu callback to free various structures associated with a task group */
Mathias Krauseb0277892021-11-03 20:06:13 +010010281static void sched_unregister_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010282{
Ingo Molnard1ccc662017-02-01 11:46:42 +010010283 /* Now it should be safe to free those cfs_rqs: */
Mathias Krauseb0277892021-11-03 20:06:13 +010010284 sched_unregister_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010285}
10286
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010287void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010288{
Ingo Molnard1ccc662017-02-01 11:46:42 +010010289 /* Wait for possible concurrent references to cfs_rqs complete: */
Mathias Krauseb0277892021-11-03 20:06:13 +010010290 call_rcu(&tg->rcu, sched_unregister_group_rcu);
Li Zefanace783b2013-01-24 14:30:48 +080010291}
10292
Mathias Krauseb0277892021-11-03 20:06:13 +010010293void sched_release_group(struct task_group *tg)
Li Zefanace783b2013-01-24 14:30:48 +080010294{
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010295 unsigned long flags;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010296
Mathias Krauseb0277892021-11-03 20:06:13 +010010297 /*
10298 * Unlink first, to avoid walk_tg_tree_from() from finding us (via
10299 * sched_cfs_period_timer()).
10300 *
10301 * For this to be effective, we have to wait for all pending users of
10302 * this task group to leave their RCU critical section to ensure no new
10303 * user will see our dying task group any more. Specifically ensure
10304 * that tg_unthrottle_up() won't add decayed cfs_rq's to it.
10305 *
10306 * We therefore defer calling unregister_fair_sched_group() to
10307 * sched_unregister_group() which is guarantied to get called only after the
10308 * current RCU grace period has expired.
10309 */
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -080010310 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010311 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010312 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010313 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010314}
10315
Chengming Zhou39c42612022-08-18 20:47:58 +080010316static void sched_change_group(struct task_struct *tsk)
Vincent Guittotea86cb42016-06-17 13:38:55 +020010317{
10318 struct task_group *tg;
10319
10320 /*
10321 * All callers are synchronized by task_rq_lock(); we do not use RCU
10322 * which is pointless here. Thus, we pass "true" to task_css_check()
10323 * to prevent lockdep warnings.
10324 */
10325 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
10326 struct task_group, css);
10327 tg = autogroup_task_group(tsk, tg);
10328 tsk->sched_task_group = tg;
10329
10330#ifdef CONFIG_FAIR_GROUP_SCHED
10331 if (tsk->sched_class->task_change_group)
Chengming Zhou39c42612022-08-18 20:47:58 +080010332 tsk->sched_class->task_change_group(tsk);
Vincent Guittotea86cb42016-06-17 13:38:55 +020010333 else
10334#endif
10335 set_task_rq(tsk, task_cpu(tsk));
10336}
10337
10338/*
10339 * Change task's runqueue when it moves between groups.
10340 *
10341 * The caller of this function should have put the task in its new group by
10342 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
10343 * its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010344 */
10345void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010346{
Peter Zijlstra7a57f322017-02-21 14:47:02 +010010347 int queued, running, queue_flags =
10348 DEQUEUE_SAVE | DEQUEUE_MOVE | DEQUEUE_NOCLOCK;
Peter Zijlstraeb580752015-07-31 21:28:18 +020010349 struct rq_flags rf;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010350 struct rq *rq;
10351
Peter Zijlstraeb580752015-07-31 21:28:18 +020010352 rq = task_rq_lock(tsk, &rf);
Peter Zijlstra1b1d6222017-01-23 16:05:55 +010010353 update_rq_clock(rq);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010354
Dmitry Adamushko051a1d12007-12-18 15:21:13 +010010355 running = task_current(rq, tsk);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +040010356 queued = task_on_rq_queued(tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010357
Kirill Tkhaida0c1e62014-08-20 13:47:32 +040010358 if (queued)
Peter Zijlstra7a57f322017-02-21 14:47:02 +010010359 dequeue_task(rq, tsk, queue_flags);
Steven Rostedt (VMware)bb3bac22017-02-06 11:04:26 -050010360 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +040010361 put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010362
Chengming Zhou39c42612022-08-18 20:47:58 +080010363 sched_change_group(tsk);
Peter Zijlstra810b3812008-02-29 15:21:01 -050010364
Kirill Tkhaida0c1e62014-08-20 13:47:32 +040010365 if (queued)
Peter Zijlstra7a57f322017-02-21 14:47:02 +010010366 enqueue_task(rq, tsk, queue_flags);
Vincent Guittot2a4b03f2020-01-14 15:13:56 +010010367 if (running) {
Peter Zijlstra03b7fad2019-05-29 20:36:41 +000010368 set_next_task(rq, tsk);
Vincent Guittot2a4b03f2020-01-14 15:13:56 +010010369 /*
10370 * After changing group, the running task may have joined a
10371 * throttled one but it's still the running task. Trigger a
10372 * resched to make sure that task can still run.
10373 */
10374 resched_curr(rq);
10375 }
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010376
Peter Zijlstraeb580752015-07-31 21:28:18 +020010377 task_rq_unlock(rq, tsk, &rf);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010378}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010379
Tejun Heoa7c6d552013-08-08 20:11:23 -040010380static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010381{
Tejun Heoa7c6d552013-08-08 20:11:23 -040010382 return css ? container_of(css, struct task_group, css) : NULL;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010383}
10384
Tejun Heoeb954192013-08-08 20:11:23 -040010385static struct cgroup_subsys_state *
10386cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010387{
Tejun Heoeb954192013-08-08 20:11:23 -040010388 struct task_group *parent = css_tg(parent_css);
10389 struct task_group *tg;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010390
Tejun Heoeb954192013-08-08 20:11:23 -040010391 if (!parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010392 /* This is early initialization for the top cgroup */
Yong Zhang07e06b02011-01-07 15:17:36 +080010393 return &root_task_group.css;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010394 }
10395
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010396 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010397 if (IS_ERR(tg))
10398 return ERR_PTR(-ENOMEM);
10399
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010400 return &tg->css;
10401}
10402
Konstantin Khlebnikov96b77742017-02-08 14:27:27 +030010403/* Expose task group only after completing cgroup initialization */
10404static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
10405{
10406 struct task_group *tg = css_tg(css);
10407 struct task_group *parent = css_tg(css->parent);
10408
10409 if (parent)
10410 sched_online_group(tg, parent);
Qais Yousef72260172019-12-24 11:54:04 +000010411
10412#ifdef CONFIG_UCLAMP_TASK_GROUP
10413 /* Propagate the effective uclamp value for the new group */
Qais Yousef93b73852021-05-10 15:50:32 +010010414 mutex_lock(&uclamp_mutex);
10415 rcu_read_lock();
Qais Yousef72260172019-12-24 11:54:04 +000010416 cpu_util_update_eff(css);
Qais Yousef93b73852021-05-10 15:50:32 +010010417 rcu_read_unlock();
10418 mutex_unlock(&uclamp_mutex);
Qais Yousef72260172019-12-24 11:54:04 +000010419#endif
10420
Konstantin Khlebnikov96b77742017-02-08 14:27:27 +030010421 return 0;
10422}
10423
Peter Zijlstra2f5177f2016-03-16 16:22:45 +010010424static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
Li Zefanace783b2013-01-24 14:30:48 +080010425{
Tejun Heoeb954192013-08-08 20:11:23 -040010426 struct task_group *tg = css_tg(css);
Li Zefanace783b2013-01-24 14:30:48 +080010427
Mathias Krauseb0277892021-11-03 20:06:13 +010010428 sched_release_group(tg);
Li Zefanace783b2013-01-24 14:30:48 +080010429}
10430
Tejun Heoeb954192013-08-08 20:11:23 -040010431static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010432{
Tejun Heoeb954192013-08-08 20:11:23 -040010433 struct task_group *tg = css_tg(css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010434
Peter Zijlstra2f5177f2016-03-16 16:22:45 +010010435 /*
10436 * Relies on the RCU grace period between css_released() and this.
10437 */
Mathias Krauseb0277892021-11-03 20:06:13 +010010438 sched_unregister_group(tg);
Li Zefanace783b2013-01-24 14:30:48 +080010439}
10440
Chengming Zhoudf16b712022-08-18 20:48:03 +080010441#ifdef CONFIG_RT_GROUP_SCHED
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010442static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010443{
Tejun Heobb9d97b2011-12-12 18:12:21 -080010444 struct task_struct *task;
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010445 struct cgroup_subsys_state *css;
Tejun Heobb9d97b2011-12-12 18:12:21 -080010446
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010447 cgroup_taskset_for_each(task, css, tset) {
Tejun Heoeb954192013-08-08 20:11:23 -040010448 if (!sched_rt_can_attach(css_tg(css), task))
Tejun Heobb9d97b2011-12-12 18:12:21 -080010449 return -EINVAL;
Tejun Heobb9d97b2011-12-12 18:12:21 -080010450 }
Chengming Zhoudf16b712022-08-18 20:48:03 +080010451 return 0;
Ben Blumbe367d02009-09-23 15:56:31 -070010452}
Chengming Zhoudf16b712022-08-18 20:48:03 +080010453#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010454
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010455static void cpu_cgroup_attach(struct cgroup_taskset *tset)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010456{
Tejun Heobb9d97b2011-12-12 18:12:21 -080010457 struct task_struct *task;
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010458 struct cgroup_subsys_state *css;
Tejun Heobb9d97b2011-12-12 18:12:21 -080010459
Tejun Heo1f7dd3e52015-12-03 10:18:21 -050010460 cgroup_taskset_for_each(task, css, tset)
Tejun Heobb9d97b2011-12-12 18:12:21 -080010461 sched_move_task(task);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010462}
10463
Patrick Bellasi2480c092019-08-22 14:28:06 +010010464#ifdef CONFIG_UCLAMP_TASK_GROUP
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010465static void cpu_util_update_eff(struct cgroup_subsys_state *css)
10466{
10467 struct cgroup_subsys_state *top_css = css;
10468 struct uclamp_se *uc_parent = NULL;
10469 struct uclamp_se *uc_se = NULL;
10470 unsigned int eff[UCLAMP_CNT];
Patrick Bellasi0413d7f2019-08-22 14:28:11 +010010471 enum uclamp_id clamp_id;
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010472 unsigned int clamps;
10473
Qais Yousef93b73852021-05-10 15:50:32 +010010474 lockdep_assert_held(&uclamp_mutex);
10475 SCHED_WARN_ON(!rcu_read_lock_held());
10476
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010477 css_for_each_descendant_pre(css, top_css) {
10478 uc_parent = css_tg(css)->parent
10479 ? css_tg(css)->parent->uclamp : NULL;
10480
10481 for_each_clamp_id(clamp_id) {
10482 /* Assume effective clamps matches requested clamps */
10483 eff[clamp_id] = css_tg(css)->uclamp_req[clamp_id].value;
10484 /* Cap effective clamps with parent's effective clamps */
10485 if (uc_parent &&
10486 eff[clamp_id] > uc_parent[clamp_id].value) {
10487 eff[clamp_id] = uc_parent[clamp_id].value;
10488 }
10489 }
10490 /* Ensure protection is always capped by limit */
10491 eff[UCLAMP_MIN] = min(eff[UCLAMP_MIN], eff[UCLAMP_MAX]);
10492
10493 /* Propagate most restrictive effective clamps */
10494 clamps = 0x0;
10495 uc_se = css_tg(css)->uclamp;
10496 for_each_clamp_id(clamp_id) {
10497 if (eff[clamp_id] == uc_se[clamp_id].value)
10498 continue;
10499 uc_se[clamp_id].value = eff[clamp_id];
10500 uc_se[clamp_id].bucket_id = uclamp_bucket_id(eff[clamp_id]);
10501 clamps |= (0x1 << clamp_id);
10502 }
Patrick Bellasibabbe172019-08-22 14:28:10 +010010503 if (!clamps) {
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010504 css = css_rightmost_descendant(css);
Patrick Bellasibabbe172019-08-22 14:28:10 +010010505 continue;
10506 }
10507
10508 /* Immediately update descendants RUNNABLE tasks */
Qais Yousef0213b702021-06-17 17:51:55 +010010509 uclamp_update_active_tasks(css);
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010510 }
10511}
Patrick Bellasi2480c092019-08-22 14:28:06 +010010512
10513/*
10514 * Integer 10^N with a given N exponent by casting to integer the literal "1eN"
10515 * C expression. Since there is no way to convert a macro argument (N) into a
10516 * character constant, use two levels of macros.
10517 */
10518#define _POW10(exp) ((unsigned int)1e##exp)
10519#define POW10(exp) _POW10(exp)
10520
10521struct uclamp_request {
10522#define UCLAMP_PERCENT_SHIFT 2
10523#define UCLAMP_PERCENT_SCALE (100 * POW10(UCLAMP_PERCENT_SHIFT))
10524 s64 percent;
10525 u64 util;
10526 int ret;
10527};
10528
10529static inline struct uclamp_request
10530capacity_from_percent(char *buf)
10531{
10532 struct uclamp_request req = {
10533 .percent = UCLAMP_PERCENT_SCALE,
10534 .util = SCHED_CAPACITY_SCALE,
10535 .ret = 0,
10536 };
10537
10538 buf = strim(buf);
10539 if (strcmp(buf, "max")) {
10540 req.ret = cgroup_parse_float(buf, UCLAMP_PERCENT_SHIFT,
10541 &req.percent);
10542 if (req.ret)
10543 return req;
Qais Yousefb562d142020-01-14 21:09:47 +000010544 if ((u64)req.percent > UCLAMP_PERCENT_SCALE) {
Patrick Bellasi2480c092019-08-22 14:28:06 +010010545 req.ret = -ERANGE;
10546 return req;
10547 }
10548
10549 req.util = req.percent << SCHED_CAPACITY_SHIFT;
10550 req.util = DIV_ROUND_CLOSEST_ULL(req.util, UCLAMP_PERCENT_SCALE);
10551 }
10552
10553 return req;
10554}
10555
10556static ssize_t cpu_uclamp_write(struct kernfs_open_file *of, char *buf,
10557 size_t nbytes, loff_t off,
10558 enum uclamp_id clamp_id)
10559{
10560 struct uclamp_request req;
10561 struct task_group *tg;
10562
10563 req = capacity_from_percent(buf);
10564 if (req.ret)
10565 return req.ret;
10566
Qais Yousef46609ce2020-06-30 12:21:23 +010010567 static_branch_enable(&sched_uclamp_used);
10568
Patrick Bellasi2480c092019-08-22 14:28:06 +010010569 mutex_lock(&uclamp_mutex);
10570 rcu_read_lock();
10571
10572 tg = css_tg(of_css(of));
10573 if (tg->uclamp_req[clamp_id].value != req.util)
10574 uclamp_se_set(&tg->uclamp_req[clamp_id], req.util, false);
10575
10576 /*
10577 * Because of not recoverable conversion rounding we keep track of the
10578 * exact requested value
10579 */
10580 tg->uclamp_pct[clamp_id] = req.percent;
10581
Patrick Bellasi0b60ba22019-08-22 14:28:07 +010010582 /* Update effective clamps to track the most restrictive value */
10583 cpu_util_update_eff(of_css(of));
10584
Patrick Bellasi2480c092019-08-22 14:28:06 +010010585 rcu_read_unlock();
10586 mutex_unlock(&uclamp_mutex);
10587
10588 return nbytes;
10589}
10590
10591static ssize_t cpu_uclamp_min_write(struct kernfs_open_file *of,
10592 char *buf, size_t nbytes,
10593 loff_t off)
10594{
10595 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MIN);
10596}
10597
10598static ssize_t cpu_uclamp_max_write(struct kernfs_open_file *of,
10599 char *buf, size_t nbytes,
10600 loff_t off)
10601{
10602 return cpu_uclamp_write(of, buf, nbytes, off, UCLAMP_MAX);
10603}
10604
10605static inline void cpu_uclamp_print(struct seq_file *sf,
10606 enum uclamp_id clamp_id)
10607{
10608 struct task_group *tg;
10609 u64 util_clamp;
10610 u64 percent;
10611 u32 rem;
10612
10613 rcu_read_lock();
10614 tg = css_tg(seq_css(sf));
10615 util_clamp = tg->uclamp_req[clamp_id].value;
10616 rcu_read_unlock();
10617
10618 if (util_clamp == SCHED_CAPACITY_SCALE) {
10619 seq_puts(sf, "max\n");
10620 return;
10621 }
10622
10623 percent = tg->uclamp_pct[clamp_id];
10624 percent = div_u64_rem(percent, POW10(UCLAMP_PERCENT_SHIFT), &rem);
10625 seq_printf(sf, "%llu.%0*u\n", percent, UCLAMP_PERCENT_SHIFT, rem);
10626}
10627
10628static int cpu_uclamp_min_show(struct seq_file *sf, void *v)
10629{
10630 cpu_uclamp_print(sf, UCLAMP_MIN);
10631 return 0;
10632}
10633
10634static int cpu_uclamp_max_show(struct seq_file *sf, void *v)
10635{
10636 cpu_uclamp_print(sf, UCLAMP_MAX);
10637 return 0;
10638}
10639#endif /* CONFIG_UCLAMP_TASK_GROUP */
10640
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010641#ifdef CONFIG_FAIR_GROUP_SCHED
Tejun Heo182446d2013-08-08 20:11:24 -040010642static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
10643 struct cftype *cftype, u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010644{
Konstantin Khlebnikov5b61d502019-02-27 11:10:18 +030010645 if (shareval > scale_load_down(ULONG_MAX))
10646 shareval = MAX_SHARES;
Tejun Heo182446d2013-08-08 20:11:24 -040010647 return sched_group_set_shares(css_tg(css), scale_load(shareval));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010648}
10649
Tejun Heo182446d2013-08-08 20:11:24 -040010650static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
10651 struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010652{
Tejun Heo182446d2013-08-08 20:11:24 -040010653 struct task_group *tg = css_tg(css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010654
Nikhil Raoc8b28112011-05-18 14:37:48 -070010655 return (u64) scale_load_down(tg->shares);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010656}
Paul Turnerab84d312011-07-21 09:43:28 -070010657
10658#ifdef CONFIG_CFS_BANDWIDTH
Paul Turnera790de92011-07-21 09:43:29 -070010659static DEFINE_MUTEX(cfs_constraints_mutex);
10660
Paul Turnerab84d312011-07-21 09:43:28 -070010661const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
YueHaibingb1546ed2019-04-18 22:47:13 +080010662static const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
Huaixin Changd505b8a2020-04-25 18:52:48 +080010663/* More than 203 days if BW_SHIFT equals 20. */
10664static const u64 max_cfs_runtime = MAX_BW * NSEC_PER_USEC;
Paul Turnerab84d312011-07-21 09:43:28 -070010665
Paul Turnera790de92011-07-21 09:43:29 -070010666static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
10667
Huaixin Changf4183712021-06-21 17:27:58 +080010668static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota,
10669 u64 burst)
Paul Turnerab84d312011-07-21 09:43:28 -070010670{
Paul Turner56f570e2011-11-07 20:26:33 -080010671 int i, ret = 0, runtime_enabled, runtime_was_enabled;
Peter Zijlstra029632f2011-10-25 10:00:11 +020010672 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Paul Turnerab84d312011-07-21 09:43:28 -070010673
10674 if (tg == &root_task_group)
10675 return -EINVAL;
10676
10677 /*
10678 * Ensure we have at some amount of bandwidth every period. This is
10679 * to prevent reaching a state of large arrears when throttled via
10680 * entity_tick() resulting in prolonged exit starvation.
10681 */
10682 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
10683 return -EINVAL;
10684
10685 /*
Ingo Molnar3b037062021-03-18 13:38:50 +010010686 * Likewise, bound things on the other side by preventing insane quota
Paul Turnerab84d312011-07-21 09:43:28 -070010687 * periods. This also allows us to normalize in computing quota
10688 * feasibility.
10689 */
10690 if (period > max_cfs_quota_period)
10691 return -EINVAL;
10692
Kirill Tkhai0e59bda2014-06-25 12:19:42 +040010693 /*
Huaixin Changd505b8a2020-04-25 18:52:48 +080010694 * Bound quota to defend quota against overflow during bandwidth shift.
10695 */
10696 if (quota != RUNTIME_INF && quota > max_cfs_runtime)
10697 return -EINVAL;
10698
Huaixin Changf4183712021-06-21 17:27:58 +080010699 if (quota != RUNTIME_INF && (burst > quota ||
10700 burst + quota > max_cfs_runtime))
10701 return -EINVAL;
10702
Huaixin Changd505b8a2020-04-25 18:52:48 +080010703 /*
Kirill Tkhai0e59bda2014-06-25 12:19:42 +040010704 * Prevent race between setting of cfs_rq->runtime_enabled and
10705 * unthrottle_offline_cfs_rqs().
10706 */
Sebastian Andrzej Siewior746f5ea2021-08-03 16:16:15 +020010707 cpus_read_lock();
Paul Turnera790de92011-07-21 09:43:29 -070010708 mutex_lock(&cfs_constraints_mutex);
10709 ret = __cfs_schedulable(tg, period, quota);
10710 if (ret)
10711 goto out_unlock;
10712
Paul Turner58088ad2011-07-21 09:43:31 -070010713 runtime_enabled = quota != RUNTIME_INF;
Paul Turner56f570e2011-11-07 20:26:33 -080010714 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
Ben Segall1ee14e62013-10-16 11:16:12 -070010715 /*
10716 * If we need to toggle cfs_bandwidth_used, off->on must occur
10717 * before making related changes, and on->off must occur afterwards
10718 */
10719 if (runtime_enabled && !runtime_was_enabled)
10720 cfs_bandwidth_usage_inc();
Paul Turnerab84d312011-07-21 09:43:28 -070010721 raw_spin_lock_irq(&cfs_b->lock);
10722 cfs_b->period = ns_to_ktime(period);
10723 cfs_b->quota = quota;
Huaixin Changf4183712021-06-21 17:27:58 +080010724 cfs_b->burst = burst;
Paul Turner58088ad2011-07-21 09:43:31 -070010725
Paul Turnera9cf55b2011-07-21 09:43:32 -070010726 __refill_cfs_bandwidth_runtime(cfs_b);
Ingo Molnard1ccc662017-02-01 11:46:42 +010010727
10728 /* Restart the period timer (if active) to handle new period expiry: */
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +020010729 if (runtime_enabled)
10730 start_cfs_bandwidth(cfs_b);
Ingo Molnard1ccc662017-02-01 11:46:42 +010010731
Paul Turnerab84d312011-07-21 09:43:28 -070010732 raw_spin_unlock_irq(&cfs_b->lock);
10733
Kirill Tkhai0e59bda2014-06-25 12:19:42 +040010734 for_each_online_cpu(i) {
Paul Turnerab84d312011-07-21 09:43:28 -070010735 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
Peter Zijlstra029632f2011-10-25 10:00:11 +020010736 struct rq *rq = cfs_rq->rq;
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +020010737 struct rq_flags rf;
Paul Turnerab84d312011-07-21 09:43:28 -070010738
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +020010739 rq_lock_irq(rq, &rf);
Paul Turner58088ad2011-07-21 09:43:31 -070010740 cfs_rq->runtime_enabled = runtime_enabled;
Paul Turnerab84d312011-07-21 09:43:28 -070010741 cfs_rq->runtime_remaining = 0;
Paul Turner671fd9d2011-07-21 09:43:34 -070010742
Peter Zijlstra029632f2011-10-25 10:00:11 +020010743 if (cfs_rq->throttled)
Paul Turner671fd9d2011-07-21 09:43:34 -070010744 unthrottle_cfs_rq(cfs_rq);
Peter Zijlstra8a8c69c2016-10-04 16:04:35 +020010745 rq_unlock_irq(rq, &rf);
Paul Turnerab84d312011-07-21 09:43:28 -070010746 }
Ben Segall1ee14e62013-10-16 11:16:12 -070010747 if (runtime_was_enabled && !runtime_enabled)
10748 cfs_bandwidth_usage_dec();
Paul Turnera790de92011-07-21 09:43:29 -070010749out_unlock:
10750 mutex_unlock(&cfs_constraints_mutex);
Sebastian Andrzej Siewior746f5ea2021-08-03 16:16:15 +020010751 cpus_read_unlock();
Paul Turnerab84d312011-07-21 09:43:28 -070010752
Paul Turnera790de92011-07-21 09:43:29 -070010753 return ret;
Paul Turnerab84d312011-07-21 09:43:28 -070010754}
10755
YueHaibingb1546ed2019-04-18 22:47:13 +080010756static int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
Paul Turnerab84d312011-07-21 09:43:28 -070010757{
Huaixin Changf4183712021-06-21 17:27:58 +080010758 u64 quota, period, burst;
Paul Turnerab84d312011-07-21 09:43:28 -070010759
Peter Zijlstra029632f2011-10-25 10:00:11 +020010760 period = ktime_to_ns(tg->cfs_bandwidth.period);
Huaixin Changf4183712021-06-21 17:27:58 +080010761 burst = tg->cfs_bandwidth.burst;
Paul Turnerab84d312011-07-21 09:43:28 -070010762 if (cfs_quota_us < 0)
10763 quota = RUNTIME_INF;
Konstantin Khlebnikov1a8b4542019-02-27 11:10:20 +030010764 else if ((u64)cfs_quota_us <= U64_MAX / NSEC_PER_USEC)
Paul Turnerab84d312011-07-21 09:43:28 -070010765 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
Konstantin Khlebnikov1a8b4542019-02-27 11:10:20 +030010766 else
10767 return -EINVAL;
Paul Turnerab84d312011-07-21 09:43:28 -070010768
Huaixin Changf4183712021-06-21 17:27:58 +080010769 return tg_set_cfs_bandwidth(tg, period, quota, burst);
Paul Turnerab84d312011-07-21 09:43:28 -070010770}
10771
YueHaibingb1546ed2019-04-18 22:47:13 +080010772static long tg_get_cfs_quota(struct task_group *tg)
Paul Turnerab84d312011-07-21 09:43:28 -070010773{
10774 u64 quota_us;
10775
Peter Zijlstra029632f2011-10-25 10:00:11 +020010776 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
Paul Turnerab84d312011-07-21 09:43:28 -070010777 return -1;
10778
Peter Zijlstra029632f2011-10-25 10:00:11 +020010779 quota_us = tg->cfs_bandwidth.quota;
Paul Turnerab84d312011-07-21 09:43:28 -070010780 do_div(quota_us, NSEC_PER_USEC);
10781
10782 return quota_us;
10783}
10784
YueHaibingb1546ed2019-04-18 22:47:13 +080010785static int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
Paul Turnerab84d312011-07-21 09:43:28 -070010786{
Huaixin Changf4183712021-06-21 17:27:58 +080010787 u64 quota, period, burst;
Paul Turnerab84d312011-07-21 09:43:28 -070010788
Konstantin Khlebnikov1a8b4542019-02-27 11:10:20 +030010789 if ((u64)cfs_period_us > U64_MAX / NSEC_PER_USEC)
10790 return -EINVAL;
10791
Paul Turnerab84d312011-07-21 09:43:28 -070010792 period = (u64)cfs_period_us * NSEC_PER_USEC;
Peter Zijlstra029632f2011-10-25 10:00:11 +020010793 quota = tg->cfs_bandwidth.quota;
Huaixin Changf4183712021-06-21 17:27:58 +080010794 burst = tg->cfs_bandwidth.burst;
Paul Turnerab84d312011-07-21 09:43:28 -070010795
Huaixin Changf4183712021-06-21 17:27:58 +080010796 return tg_set_cfs_bandwidth(tg, period, quota, burst);
Paul Turnerab84d312011-07-21 09:43:28 -070010797}
10798
YueHaibingb1546ed2019-04-18 22:47:13 +080010799static long tg_get_cfs_period(struct task_group *tg)
Paul Turnerab84d312011-07-21 09:43:28 -070010800{
10801 u64 cfs_period_us;
10802
Peter Zijlstra029632f2011-10-25 10:00:11 +020010803 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
Paul Turnerab84d312011-07-21 09:43:28 -070010804 do_div(cfs_period_us, NSEC_PER_USEC);
10805
10806 return cfs_period_us;
10807}
10808
Huaixin Changf4183712021-06-21 17:27:58 +080010809static int tg_set_cfs_burst(struct task_group *tg, long cfs_burst_us)
10810{
10811 u64 quota, period, burst;
10812
10813 if ((u64)cfs_burst_us > U64_MAX / NSEC_PER_USEC)
10814 return -EINVAL;
10815
10816 burst = (u64)cfs_burst_us * NSEC_PER_USEC;
10817 period = ktime_to_ns(tg->cfs_bandwidth.period);
10818 quota = tg->cfs_bandwidth.quota;
10819
10820 return tg_set_cfs_bandwidth(tg, period, quota, burst);
10821}
10822
10823static long tg_get_cfs_burst(struct task_group *tg)
10824{
10825 u64 burst_us;
10826
10827 burst_us = tg->cfs_bandwidth.burst;
10828 do_div(burst_us, NSEC_PER_USEC);
10829
10830 return burst_us;
10831}
10832
Tejun Heo182446d2013-08-08 20:11:24 -040010833static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
10834 struct cftype *cft)
Paul Turnerab84d312011-07-21 09:43:28 -070010835{
Tejun Heo182446d2013-08-08 20:11:24 -040010836 return tg_get_cfs_quota(css_tg(css));
Paul Turnerab84d312011-07-21 09:43:28 -070010837}
10838
Tejun Heo182446d2013-08-08 20:11:24 -040010839static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
10840 struct cftype *cftype, s64 cfs_quota_us)
Paul Turnerab84d312011-07-21 09:43:28 -070010841{
Tejun Heo182446d2013-08-08 20:11:24 -040010842 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
Paul Turnerab84d312011-07-21 09:43:28 -070010843}
10844
Tejun Heo182446d2013-08-08 20:11:24 -040010845static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
10846 struct cftype *cft)
Paul Turnerab84d312011-07-21 09:43:28 -070010847{
Tejun Heo182446d2013-08-08 20:11:24 -040010848 return tg_get_cfs_period(css_tg(css));
Paul Turnerab84d312011-07-21 09:43:28 -070010849}
10850
Tejun Heo182446d2013-08-08 20:11:24 -040010851static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
10852 struct cftype *cftype, u64 cfs_period_us)
Paul Turnerab84d312011-07-21 09:43:28 -070010853{
Tejun Heo182446d2013-08-08 20:11:24 -040010854 return tg_set_cfs_period(css_tg(css), cfs_period_us);
Paul Turnerab84d312011-07-21 09:43:28 -070010855}
10856
Huaixin Changf4183712021-06-21 17:27:58 +080010857static u64 cpu_cfs_burst_read_u64(struct cgroup_subsys_state *css,
10858 struct cftype *cft)
10859{
10860 return tg_get_cfs_burst(css_tg(css));
10861}
10862
10863static int cpu_cfs_burst_write_u64(struct cgroup_subsys_state *css,
10864 struct cftype *cftype, u64 cfs_burst_us)
10865{
10866 return tg_set_cfs_burst(css_tg(css), cfs_burst_us);
10867}
10868
Paul Turnera790de92011-07-21 09:43:29 -070010869struct cfs_schedulable_data {
10870 struct task_group *tg;
10871 u64 period, quota;
10872};
10873
10874/*
10875 * normalize group quota/period to be quota/max_period
10876 * note: units are usecs
10877 */
10878static u64 normalize_cfs_quota(struct task_group *tg,
10879 struct cfs_schedulable_data *d)
10880{
10881 u64 quota, period;
10882
10883 if (tg == d->tg) {
10884 period = d->period;
10885 quota = d->quota;
10886 } else {
10887 period = tg_get_cfs_period(tg);
10888 quota = tg_get_cfs_quota(tg);
10889 }
10890
10891 /* note: these should typically be equivalent */
10892 if (quota == RUNTIME_INF || quota == -1)
10893 return RUNTIME_INF;
10894
10895 return to_ratio(period, quota);
10896}
10897
10898static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
10899{
10900 struct cfs_schedulable_data *d = data;
Peter Zijlstra029632f2011-10-25 10:00:11 +020010901 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Paul Turnera790de92011-07-21 09:43:29 -070010902 s64 quota = 0, parent_quota = -1;
10903
10904 if (!tg->parent) {
10905 quota = RUNTIME_INF;
10906 } else {
Peter Zijlstra029632f2011-10-25 10:00:11 +020010907 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
Paul Turnera790de92011-07-21 09:43:29 -070010908
10909 quota = normalize_cfs_quota(tg, d);
Zhihui Zhang9c58c792014-09-20 21:24:36 -040010910 parent_quota = parent_b->hierarchical_quota;
Paul Turnera790de92011-07-21 09:43:29 -070010911
10912 /*
Tejun Heoc53593e2018-01-22 11:26:18 -080010913 * Ensure max(child_quota) <= parent_quota. On cgroup2,
10914 * always take the min. On cgroup1, only inherit when no
Ingo Molnard1ccc662017-02-01 11:46:42 +010010915 * limit is set:
Paul Turnera790de92011-07-21 09:43:29 -070010916 */
Tejun Heoc53593e2018-01-22 11:26:18 -080010917 if (cgroup_subsys_on_dfl(cpu_cgrp_subsys)) {
10918 quota = min(quota, parent_quota);
10919 } else {
10920 if (quota == RUNTIME_INF)
10921 quota = parent_quota;
10922 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
10923 return -EINVAL;
10924 }
Paul Turnera790de92011-07-21 09:43:29 -070010925 }
Zhihui Zhang9c58c792014-09-20 21:24:36 -040010926 cfs_b->hierarchical_quota = quota;
Paul Turnera790de92011-07-21 09:43:29 -070010927
10928 return 0;
10929}
10930
10931static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
10932{
Paul Turner82774342011-07-21 09:43:35 -070010933 int ret;
Paul Turnera790de92011-07-21 09:43:29 -070010934 struct cfs_schedulable_data data = {
10935 .tg = tg,
10936 .period = period,
10937 .quota = quota,
10938 };
10939
10940 if (quota != RUNTIME_INF) {
10941 do_div(data.period, NSEC_PER_USEC);
10942 do_div(data.quota, NSEC_PER_USEC);
10943 }
10944
Paul Turner82774342011-07-21 09:43:35 -070010945 rcu_read_lock();
10946 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
10947 rcu_read_unlock();
10948
10949 return ret;
Paul Turnera790de92011-07-21 09:43:29 -070010950}
Nikhil Raoe8da1b12011-07-21 09:43:40 -070010951
Tejun Heoa1f71642017-09-25 09:00:18 -070010952static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
Nikhil Raoe8da1b12011-07-21 09:43:40 -070010953{
Tejun Heo2da8ca82013-12-05 12:28:04 -050010954 struct task_group *tg = css_tg(seq_css(sf));
Peter Zijlstra029632f2011-10-25 10:00:11 +020010955 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Nikhil Raoe8da1b12011-07-21 09:43:40 -070010956
Tejun Heo44ffc752013-12-05 12:28:01 -050010957 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
10958 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
10959 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
Nikhil Raoe8da1b12011-07-21 09:43:40 -070010960
Yun Wang3d6c50c2018-07-04 11:27:27 +080010961 if (schedstat_enabled() && tg != &root_task_group) {
Yafang Shaoceeadb82021-09-05 14:35:41 +000010962 struct sched_statistics *stats;
Yun Wang3d6c50c2018-07-04 11:27:27 +080010963 u64 ws = 0;
10964 int i;
10965
Yafang Shaoceeadb82021-09-05 14:35:41 +000010966 for_each_possible_cpu(i) {
10967 stats = __schedstats_from_se(tg->se[i]);
10968 ws += schedstat_val(stats->wait_sum);
10969 }
Yun Wang3d6c50c2018-07-04 11:27:27 +080010970
10971 seq_printf(sf, "wait_sum %llu\n", ws);
10972 }
10973
Huaixin Changbcb17042021-08-30 11:22:14 +080010974 seq_printf(sf, "nr_bursts %d\n", cfs_b->nr_burst);
10975 seq_printf(sf, "burst_time %llu\n", cfs_b->burst_time);
10976
Nikhil Raoe8da1b12011-07-21 09:43:40 -070010977 return 0;
10978}
Paul Turnerab84d312011-07-21 09:43:28 -070010979#endif /* CONFIG_CFS_BANDWIDTH */
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010980#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010981
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010982#ifdef CONFIG_RT_GROUP_SCHED
Tejun Heo182446d2013-08-08 20:11:24 -040010983static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
10984 struct cftype *cft, s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010985{
Tejun Heo182446d2013-08-08 20:11:24 -040010986 return sched_group_set_rt_runtime(css_tg(css), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010987}
10988
Tejun Heo182446d2013-08-08 20:11:24 -040010989static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
10990 struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010991{
Tejun Heo182446d2013-08-08 20:11:24 -040010992 return sched_group_rt_runtime(css_tg(css));
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010993}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010994
Tejun Heo182446d2013-08-08 20:11:24 -040010995static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
10996 struct cftype *cftype, u64 rt_period_us)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010997{
Tejun Heo182446d2013-08-08 20:11:24 -040010998 return sched_group_set_rt_period(css_tg(css), rt_period_us);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010999}
11000
Tejun Heo182446d2013-08-08 20:11:24 -040011001static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
11002 struct cftype *cft)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020011003{
Tejun Heo182446d2013-08-08 20:11:24 -040011004 return sched_group_rt_period(css_tg(css));
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020011005}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020011006#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010011007
Josh Don30400032021-07-29 19:00:18 -070011008#ifdef CONFIG_FAIR_GROUP_SCHED
11009static s64 cpu_idle_read_s64(struct cgroup_subsys_state *css,
11010 struct cftype *cft)
11011{
11012 return css_tg(css)->idle;
11013}
11014
11015static int cpu_idle_write_s64(struct cgroup_subsys_state *css,
11016 struct cftype *cft, s64 idle)
11017{
11018 return sched_group_set_idle(css_tg(css), idle);
11019}
11020#endif
11021
Tejun Heoa1f71642017-09-25 09:00:18 -070011022static struct cftype cpu_legacy_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010011023#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +010011024 {
11025 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -070011026 .read_u64 = cpu_shares_read_u64,
11027 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +010011028 },
Josh Don30400032021-07-29 19:00:18 -070011029 {
11030 .name = "idle",
11031 .read_s64 = cpu_idle_read_s64,
11032 .write_s64 = cpu_idle_write_s64,
11033 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010011034#endif
Paul Turnerab84d312011-07-21 09:43:28 -070011035#ifdef CONFIG_CFS_BANDWIDTH
11036 {
11037 .name = "cfs_quota_us",
11038 .read_s64 = cpu_cfs_quota_read_s64,
11039 .write_s64 = cpu_cfs_quota_write_s64,
11040 },
11041 {
11042 .name = "cfs_period_us",
11043 .read_u64 = cpu_cfs_period_read_u64,
11044 .write_u64 = cpu_cfs_period_write_u64,
11045 },
Nikhil Raoe8da1b12011-07-21 09:43:40 -070011046 {
Huaixin Changf4183712021-06-21 17:27:58 +080011047 .name = "cfs_burst_us",
11048 .read_u64 = cpu_cfs_burst_read_u64,
11049 .write_u64 = cpu_cfs_burst_write_u64,
11050 },
11051 {
Nikhil Raoe8da1b12011-07-21 09:43:40 -070011052 .name = "stat",
Tejun Heoa1f71642017-09-25 09:00:18 -070011053 .seq_show = cpu_cfs_stat_show,
Nikhil Raoe8da1b12011-07-21 09:43:40 -070011054 },
Paul Turnerab84d312011-07-21 09:43:28 -070011055#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010011056#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010011057 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010011058 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -070011059 .read_s64 = cpu_rt_runtime_read,
11060 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +010011061 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020011062 {
11063 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -070011064 .read_u64 = cpu_rt_period_read_uint,
11065 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020011066 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010011067#endif
Patrick Bellasi2480c092019-08-22 14:28:06 +010011068#ifdef CONFIG_UCLAMP_TASK_GROUP
11069 {
11070 .name = "uclamp.min",
11071 .flags = CFTYPE_NOT_ON_ROOT,
11072 .seq_show = cpu_uclamp_min_show,
11073 .write = cpu_uclamp_min_write,
11074 },
11075 {
11076 .name = "uclamp.max",
11077 .flags = CFTYPE_NOT_ON_ROOT,
11078 .seq_show = cpu_uclamp_max_show,
11079 .write = cpu_uclamp_max_write,
11080 },
11081#endif
Ingo Molnard1ccc662017-02-01 11:46:42 +010011082 { } /* Terminate */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070011083};
11084
Tejun Heod41bf8c92017-10-23 16:18:27 -070011085static int cpu_extra_stat_show(struct seq_file *sf,
11086 struct cgroup_subsys_state *css)
Tejun Heo0d593632017-09-25 09:00:19 -070011087{
Tejun Heo0d593632017-09-25 09:00:19 -070011088#ifdef CONFIG_CFS_BANDWIDTH
11089 {
Tejun Heod41bf8c92017-10-23 16:18:27 -070011090 struct task_group *tg = css_tg(css);
Tejun Heo0d593632017-09-25 09:00:19 -070011091 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Huaixin Changbcb17042021-08-30 11:22:14 +080011092 u64 throttled_usec, burst_usec;
Tejun Heo0d593632017-09-25 09:00:19 -070011093
11094 throttled_usec = cfs_b->throttled_time;
11095 do_div(throttled_usec, NSEC_PER_USEC);
Huaixin Changbcb17042021-08-30 11:22:14 +080011096 burst_usec = cfs_b->burst_time;
11097 do_div(burst_usec, NSEC_PER_USEC);
Tejun Heo0d593632017-09-25 09:00:19 -070011098
11099 seq_printf(sf, "nr_periods %d\n"
11100 "nr_throttled %d\n"
Huaixin Changbcb17042021-08-30 11:22:14 +080011101 "throttled_usec %llu\n"
11102 "nr_bursts %d\n"
11103 "burst_usec %llu\n",
Tejun Heo0d593632017-09-25 09:00:19 -070011104 cfs_b->nr_periods, cfs_b->nr_throttled,
Huaixin Changbcb17042021-08-30 11:22:14 +080011105 throttled_usec, cfs_b->nr_burst, burst_usec);
Tejun Heo0d593632017-09-25 09:00:19 -070011106 }
11107#endif
11108 return 0;
11109}
11110
11111#ifdef CONFIG_FAIR_GROUP_SCHED
11112static u64 cpu_weight_read_u64(struct cgroup_subsys_state *css,
11113 struct cftype *cft)
11114{
11115 struct task_group *tg = css_tg(css);
11116 u64 weight = scale_load_down(tg->shares);
11117
11118 return DIV_ROUND_CLOSEST_ULL(weight * CGROUP_WEIGHT_DFL, 1024);
11119}
11120
11121static int cpu_weight_write_u64(struct cgroup_subsys_state *css,
11122 struct cftype *cft, u64 weight)
11123{
11124 /*
11125 * cgroup weight knobs should use the common MIN, DFL and MAX
11126 * values which are 1, 100 and 10000 respectively. While it loses
11127 * a bit of range on both ends, it maps pretty well onto the shares
11128 * value used by scheduler and the round-trip conversions preserve
11129 * the original value over the entire range.
11130 */
11131 if (weight < CGROUP_WEIGHT_MIN || weight > CGROUP_WEIGHT_MAX)
11132 return -ERANGE;
11133
11134 weight = DIV_ROUND_CLOSEST_ULL(weight * 1024, CGROUP_WEIGHT_DFL);
11135
11136 return sched_group_set_shares(css_tg(css), scale_load(weight));
11137}
11138
11139static s64 cpu_weight_nice_read_s64(struct cgroup_subsys_state *css,
11140 struct cftype *cft)
11141{
11142 unsigned long weight = scale_load_down(css_tg(css)->shares);
11143 int last_delta = INT_MAX;
11144 int prio, delta;
11145
11146 /* find the closest nice value to the current weight */
11147 for (prio = 0; prio < ARRAY_SIZE(sched_prio_to_weight); prio++) {
11148 delta = abs(sched_prio_to_weight[prio] - weight);
11149 if (delta >= last_delta)
11150 break;
11151 last_delta = delta;
11152 }
11153
11154 return PRIO_TO_NICE(prio - 1 + MAX_RT_PRIO);
11155}
11156
11157static int cpu_weight_nice_write_s64(struct cgroup_subsys_state *css,
11158 struct cftype *cft, s64 nice)
11159{
11160 unsigned long weight;
Peter Zijlstra7281c8d2018-04-20 14:29:51 +020011161 int idx;
Tejun Heo0d593632017-09-25 09:00:19 -070011162
11163 if (nice < MIN_NICE || nice > MAX_NICE)
11164 return -ERANGE;
11165
Peter Zijlstra7281c8d2018-04-20 14:29:51 +020011166 idx = NICE_TO_PRIO(nice) - MAX_RT_PRIO;
11167 idx = array_index_nospec(idx, 40);
11168 weight = sched_prio_to_weight[idx];
11169
Tejun Heo0d593632017-09-25 09:00:19 -070011170 return sched_group_set_shares(css_tg(css), scale_load(weight));
11171}
11172#endif
11173
11174static void __maybe_unused cpu_period_quota_print(struct seq_file *sf,
11175 long period, long quota)
11176{
11177 if (quota < 0)
11178 seq_puts(sf, "max");
11179 else
11180 seq_printf(sf, "%ld", quota);
11181
11182 seq_printf(sf, " %ld\n", period);
11183}
11184
11185/* caller should put the current value in *@periodp before calling */
11186static int __maybe_unused cpu_period_quota_parse(char *buf,
11187 u64 *periodp, u64 *quotap)
11188{
11189 char tok[21]; /* U64_MAX */
11190
Konstantin Khlebnikov4c47acd2019-03-06 20:11:42 +030011191 if (sscanf(buf, "%20s %llu", tok, periodp) < 1)
Tejun Heo0d593632017-09-25 09:00:19 -070011192 return -EINVAL;
11193
11194 *periodp *= NSEC_PER_USEC;
11195
11196 if (sscanf(tok, "%llu", quotap))
11197 *quotap *= NSEC_PER_USEC;
11198 else if (!strcmp(tok, "max"))
11199 *quotap = RUNTIME_INF;
11200 else
11201 return -EINVAL;
11202
11203 return 0;
11204}
11205
11206#ifdef CONFIG_CFS_BANDWIDTH
11207static int cpu_max_show(struct seq_file *sf, void *v)
11208{
11209 struct task_group *tg = css_tg(seq_css(sf));
11210
11211 cpu_period_quota_print(sf, tg_get_cfs_period(tg), tg_get_cfs_quota(tg));
11212 return 0;
11213}
11214
11215static ssize_t cpu_max_write(struct kernfs_open_file *of,
11216 char *buf, size_t nbytes, loff_t off)
11217{
11218 struct task_group *tg = css_tg(of_css(of));
11219 u64 period = tg_get_cfs_period(tg);
Huaixin Changf4183712021-06-21 17:27:58 +080011220 u64 burst = tg_get_cfs_burst(tg);
Tejun Heo0d593632017-09-25 09:00:19 -070011221 u64 quota;
11222 int ret;
11223
11224 ret = cpu_period_quota_parse(buf, &period, &quota);
11225 if (!ret)
Huaixin Changf4183712021-06-21 17:27:58 +080011226 ret = tg_set_cfs_bandwidth(tg, period, quota, burst);
Tejun Heo0d593632017-09-25 09:00:19 -070011227 return ret ?: nbytes;
11228}
11229#endif
11230
11231static struct cftype cpu_files[] = {
Tejun Heo0d593632017-09-25 09:00:19 -070011232#ifdef CONFIG_FAIR_GROUP_SCHED
11233 {
11234 .name = "weight",
11235 .flags = CFTYPE_NOT_ON_ROOT,
11236 .read_u64 = cpu_weight_read_u64,
11237 .write_u64 = cpu_weight_write_u64,
11238 },
11239 {
11240 .name = "weight.nice",
11241 .flags = CFTYPE_NOT_ON_ROOT,
11242 .read_s64 = cpu_weight_nice_read_s64,
11243 .write_s64 = cpu_weight_nice_write_s64,
11244 },
Josh Don30400032021-07-29 19:00:18 -070011245 {
11246 .name = "idle",
11247 .flags = CFTYPE_NOT_ON_ROOT,
11248 .read_s64 = cpu_idle_read_s64,
11249 .write_s64 = cpu_idle_write_s64,
11250 },
Tejun Heo0d593632017-09-25 09:00:19 -070011251#endif
11252#ifdef CONFIG_CFS_BANDWIDTH
11253 {
11254 .name = "max",
11255 .flags = CFTYPE_NOT_ON_ROOT,
11256 .seq_show = cpu_max_show,
11257 .write = cpu_max_write,
11258 },
Huaixin Changf4183712021-06-21 17:27:58 +080011259 {
11260 .name = "max.burst",
11261 .flags = CFTYPE_NOT_ON_ROOT,
11262 .read_u64 = cpu_cfs_burst_read_u64,
11263 .write_u64 = cpu_cfs_burst_write_u64,
11264 },
Tejun Heo0d593632017-09-25 09:00:19 -070011265#endif
Patrick Bellasi2480c092019-08-22 14:28:06 +010011266#ifdef CONFIG_UCLAMP_TASK_GROUP
11267 {
11268 .name = "uclamp.min",
11269 .flags = CFTYPE_NOT_ON_ROOT,
11270 .seq_show = cpu_uclamp_min_show,
11271 .write = cpu_uclamp_min_write,
11272 },
11273 {
11274 .name = "uclamp.max",
11275 .flags = CFTYPE_NOT_ON_ROOT,
11276 .seq_show = cpu_uclamp_max_show,
11277 .write = cpu_uclamp_max_write,
11278 },
11279#endif
Tejun Heo0d593632017-09-25 09:00:19 -070011280 { } /* terminate */
11281};
11282
Tejun Heo073219e2014-02-08 10:36:58 -050011283struct cgroup_subsys cpu_cgrp_subsys = {
Tejun Heo92fb9742012-11-19 08:13:38 -080011284 .css_alloc = cpu_cgroup_css_alloc,
Konstantin Khlebnikov96b77742017-02-08 14:27:27 +030011285 .css_online = cpu_cgroup_css_online,
Peter Zijlstra2f5177f2016-03-16 16:22:45 +010011286 .css_released = cpu_cgroup_css_released,
Tejun Heo92fb9742012-11-19 08:13:38 -080011287 .css_free = cpu_cgroup_css_free,
Tejun Heod41bf8c92017-10-23 16:18:27 -070011288 .css_extra_stat_show = cpu_extra_stat_show,
Chengming Zhoudf16b712022-08-18 20:48:03 +080011289#ifdef CONFIG_RT_GROUP_SCHED
Tejun Heobb9d97b2011-12-12 18:12:21 -080011290 .can_attach = cpu_cgroup_can_attach,
Chengming Zhoudf16b712022-08-18 20:48:03 +080011291#endif
Tejun Heobb9d97b2011-12-12 18:12:21 -080011292 .attach = cpu_cgroup_attach,
Tejun Heoa1f71642017-09-25 09:00:18 -070011293 .legacy_cftypes = cpu_legacy_files,
Tejun Heo0d593632017-09-25 09:00:19 -070011294 .dfl_cftypes = cpu_files,
Tejun Heob38e42e2016-02-23 10:00:50 -050011295 .early_init = true,
Tejun Heo0d593632017-09-25 09:00:19 -070011296 .threaded = true,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070011297};
11298
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010011299#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010011300
Paul E. McKenneyb637a322012-09-19 16:58:38 -070011301void dump_cpu_task(int cpu)
11302{
Zhen Leibc1cca92022-08-04 10:34:20 +080011303 if (cpu == smp_processor_id() && in_hardirq()) {
11304 struct pt_regs *regs;
11305
11306 regs = get_irq_regs();
11307 if (regs) {
11308 show_regs(regs);
11309 return;
11310 }
11311 }
11312
Zhen Leie73dfe32022-08-04 10:34:19 +080011313 if (trigger_single_cpu_backtrace(cpu))
11314 return;
11315
Paul E. McKenneyb637a322012-09-19 16:58:38 -070011316 pr_info("Task dump for CPU %d:\n", cpu);
11317 sched_show_task(cpu_curr(cpu));
11318}
Andi Kleened82b8a2015-11-29 20:59:43 -080011319
11320/*
11321 * Nice levels are multiplicative, with a gentle 10% change for every
11322 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
11323 * nice 1, it will get ~10% less CPU time than another CPU-bound task
11324 * that remained on nice 0.
11325 *
11326 * The "10% effect" is relative and cumulative: from _any_ nice level,
11327 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
11328 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
11329 * If a task goes up by ~10% and another task goes down by ~10% then
11330 * the relative distance between them is ~25%.)
11331 */
11332const int sched_prio_to_weight[40] = {
11333 /* -20 */ 88761, 71755, 56483, 46273, 36291,
11334 /* -15 */ 29154, 23254, 18705, 14949, 11916,
11335 /* -10 */ 9548, 7620, 6100, 4904, 3906,
11336 /* -5 */ 3121, 2501, 1991, 1586, 1277,
11337 /* 0 */ 1024, 820, 655, 526, 423,
11338 /* 5 */ 335, 272, 215, 172, 137,
11339 /* 10 */ 110, 87, 70, 56, 45,
11340 /* 15 */ 36, 29, 23, 18, 15,
11341};
11342
11343/*
11344 * Inverse (2^32/x) values of the sched_prio_to_weight[] array, precalculated.
11345 *
11346 * In cases where the weight does not change often, we can use the
11347 * precalculated inverse to speed up arithmetics by turning divisions
11348 * into multiplications:
11349 */
11350const u32 sched_prio_to_wmult[40] = {
11351 /* -20 */ 48388, 59856, 76040, 92818, 118348,
11352 /* -15 */ 147320, 184698, 229616, 287308, 360437,
11353 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
11354 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
11355 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
11356 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
11357 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
11358 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
11359};
Ingo Molnar14a74052018-03-03 16:32:24 +010011360
Phil Auld9d246052020-06-29 15:23:03 -040011361void call_trace_sched_update_nr_running(struct rq *rq, int count)
11362{
11363 trace_sched_update_nr_running_tp(rq, count);
11364}