blob: 18cceeecce35eec872455c9f5e8f83028ffbbaa6 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * kernel/sched.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
Ingo Molnarc31f2e82007-07-09 18:52:01 +020019 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
Ingo Molnarb9131762008-01-25 21:08:19 +010025 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
Linus Torvalds1da177e2005-04-16 15:20:36 -070027 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020033#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/highmem.h>
35#include <linux/smp_lock.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080038#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/completion.h>
40#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070041#include <linux/debug_locks.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020042#include <linux/perf_event.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/security.h>
44#include <linux/notifier.h>
45#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080046#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080047#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#include <linux/blkdev.h>
49#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070050#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051#include <linux/smp.h>
52#include <linux/threads.h>
53#include <linux/timer.h>
54#include <linux/rcupdate.h>
55#include <linux/cpu.h>
56#include <linux/cpuset.h>
57#include <linux/percpu.h>
58#include <linux/kthread.h>
Alexey Dobriyanb5aadf72008-10-06 13:23:43 +040059#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020061#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062#include <linux/syscalls.h>
63#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070064#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080065#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070066#include <linux/delayacct.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020067#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020068#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010069#include <linux/hrtimer.h>
Reynes Philippe30914a52008-03-17 16:19:05 -070070#include <linux/tick.h>
Peter Zijlstraf00b45c2008-04-19 19:45:00 +020071#include <linux/debugfs.h>
72#include <linux/ctype.h>
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +020073#include <linux/ftrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Eric Dumazet5517d862007-05-08 00:32:57 -070075#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020076#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Gregory Haskins6e0534f2008-05-12 21:21:01 +020078#include "sched_cpupri.h"
79
Steven Rostedta8d154b2009-04-10 09:36:00 -040080#define CREATE_TRACE_POINTS
Steven Rostedtad8d75f2009-04-14 19:39:12 -040081#include <trace/events/sched.h>
Steven Rostedta8d154b2009-04-10 09:36:00 -040082
Linus Torvalds1da177e2005-04-16 15:20:36 -070083/*
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 * and back.
87 */
88#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91
92/*
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
96 */
97#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100
101/*
Ingo Molnard7876a02008-01-25 21:08:19 +0100102 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100104#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200106#define NICE_0_LOAD SCHED_LOAD_SCALE
107#define NICE_0_SHIFT SCHED_LOAD_SHIFT
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109/*
110 * These are the 'tuning knobs' of the scheduler:
111 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 * Timeslices get refilled after they expire.
114 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700116
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200117/*
118 * single value that denotes runtime == period, ie unlimited time.
119 */
120#define RUNTIME_INF ((u64)~0ULL)
121
Ingo Molnare05606d2007-07-09 18:51:59 +0200122static inline int rt_policy(int policy)
123{
Roel Kluin3f33a7c2008-05-13 23:44:11 +0200124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
Ingo Molnare05606d2007-07-09 18:51:59 +0200125 return 1;
126 return 0;
127}
128
129static inline int task_has_rt_policy(struct task_struct *p)
130{
131 return rt_policy(p->policy);
132}
133
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200135 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200137struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
140};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200142struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100143 /* nests inside the rq lock: */
Thomas Gleixner0986b112009-11-17 15:32:06 +0100144 raw_spinlock_t rt_runtime_lock;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100145 ktime_t rt_period;
146 u64 rt_runtime;
147 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200148};
149
150static struct rt_bandwidth def_rt_bandwidth;
151
152static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
153
154static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
155{
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
158 ktime_t now;
159 int overrun;
160 int idle = 0;
161
162 for (;;) {
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
165
166 if (!overrun)
167 break;
168
169 idle = do_sched_rt_period_timer(rt_b, overrun);
170 }
171
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
173}
174
175static
176void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
177{
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
180
Thomas Gleixner0986b112009-11-17 15:32:06 +0100181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200182
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200186}
187
Krzysztof Heltc8bfff62008-09-05 23:46:19 +0200188static inline int rt_bandwidth_enabled(void)
189{
190 return sysctl_sched_rt_runtime >= 0;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200191}
192
193static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
194{
195 ktime_t now;
196
Hiroshi Shimamotocac64d02009-02-25 09:59:26 -0800197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200198 return;
199
200 if (hrtimer_active(&rt_b->rt_period_timer))
201 return;
202
Thomas Gleixner0986b112009-11-17 15:32:06 +0100203 raw_spin_lock(&rt_b->rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200204 for (;;) {
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100205 unsigned long delta;
206 ktime_t soft, hard;
207
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200208 if (hrtimer_active(&rt_b->rt_period_timer))
209 break;
210
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100213
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +0530218 HRTIMER_MODE_ABS_PINNED, 0);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200219 }
Thomas Gleixner0986b112009-11-17 15:32:06 +0100220 raw_spin_unlock(&rt_b->rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200221}
222
223#ifdef CONFIG_RT_GROUP_SCHED
224static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
225{
226 hrtimer_cancel(&rt_b->rt_period_timer);
227}
228#endif
229
Heiko Carstens712555e2008-04-28 11:33:07 +0200230/*
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
233 */
234static DEFINE_MUTEX(sched_domains_mutex);
235
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100236#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200237
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700238#include <linux/cgroup.h>
239
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200240struct cfs_rq;
241
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100242static LIST_HEAD(task_groups);
243
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200244/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200245struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100246#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700247 struct cgroup_subsys_state css;
248#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100249
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530250#ifdef CONFIG_USER_SCHED
251 uid_t uid;
252#endif
253
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100254#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100260#endif
261
262#ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
265
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200266 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100267#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100268
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100269 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100270 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200271
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200275};
276
Dhaval Giani354d60c2008-04-19 19:44:59 +0200277#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200278
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530279/* Helper function to pass uid information to create_sched_user() */
280void set_tg_uid(struct user_struct *user)
281{
282 user->tg->uid = user->uid;
283}
284
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200285/*
286 * Root task group.
Anirban Sinha84e9dab2009-08-28 22:40:43 -0700287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200289 */
290struct task_group root_task_group;
291
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100292#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200293/* Default task group's sched entity on each cpu */
294static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295/* Default task group's cfs_rq on each cpu */
Linus Torvaldsada3fa12009-09-15 09:39:44 -0700296static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200297#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100298
299#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100300static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
Tejun Heo1871e522009-10-29 22:34:13 +0900301static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq_var);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200302#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200303#else /* !CONFIG_USER_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200304#define root_task_group init_task_group
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200305#endif /* CONFIG_USER_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100306
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100307/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100308 * a task group's cpu shares.
309 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100310static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100311
Cyrill Gorcunove9036b32009-10-26 22:24:14 +0300312#ifdef CONFIG_FAIR_GROUP_SCHED
313
Peter Zijlstra57310a92009-03-09 13:56:21 +0100314#ifdef CONFIG_SMP
315static int root_task_group_empty(void)
316{
317 return list_empty(&root_task_group.children);
318}
319#endif
320
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100321#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100322# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200323#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100324# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200325#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200326
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800327/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
334 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200335#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800336#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200337
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100338static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100339#endif
340
341/* Default task group.
342 * Every task in system belong to this group at bootup.
343 */
Mike Travis434d53b2008-04-04 18:11:04 -0700344struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200345
346/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200347static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200348{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200349 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200350
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100351#ifdef CONFIG_USER_SCHED
David Howellsc69e8d92008-11-14 10:39:19 +1100352 rcu_read_lock();
353 tg = __task_cred(p)->user->tg;
354 rcu_read_unlock();
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100355#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200358#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100359 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200360#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200361 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200362}
363
364/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100365static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200366{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100367#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100370#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100371
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100372#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100375#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200376}
377
378#else
379
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100380static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Peter Zijlstra83378262008-06-27 13:41:37 +0200381static inline struct task_group *task_group(struct task_struct *p)
382{
383 return NULL;
384}
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200385
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100386#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200387
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200388/* CFS-related fields in a runqueue */
389struct cfs_rq {
390 struct load_weight load;
391 unsigned long nr_running;
392
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200393 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200394 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200395
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200398
399 struct list_head tasks;
400 struct list_head *balance_iterator;
401
402 /*
403 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200404 * It is set to NULL otherwise (i.e when none are currently running).
405 */
Peter Zijlstra47932412008-11-04 21:25:09 +0100406 struct sched_entity *curr, *next, *last;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200407
Peter Zijlstra5ac5c4d2008-11-10 10:46:32 +0100408 unsigned int nr_spread_over;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200409
Ingo Molnar62160e32007-10-15 17:00:03 +0200410#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
412
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100413 /*
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
417 *
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
420 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200423
424#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200425 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200426 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200427 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200428 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200429
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200430 /*
431 * h_load = weight * f(tg)
432 *
433 * Where f(tg) is the recursive weight fraction assigned to
434 * this group.
435 */
436 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200437
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200438 /*
439 * this cpu's part of tg->shares
440 */
441 unsigned long shares;
Peter Zijlstraf1d239f2008-06-27 13:41:38 +0200442
443 /*
444 * load.weight at the time we set shares
445 */
446 unsigned long rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200447#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200448#endif
449};
450
451/* Real-Time classes' related field in a runqueue: */
452struct rt_rq {
453 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100454 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100455#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -0500456 struct {
457 int curr; /* highest queued rt task prio */
Gregory Haskins398a1532009-01-14 09:10:04 -0500458#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500459 int next; /* next highest */
Gregory Haskins398a1532009-01-14 09:10:04 -0500460#endif
Gregory Haskinse864c492008-12-29 09:39:49 -0500461 } highest_prio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100462#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100463#ifdef CONFIG_SMP
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100464 unsigned long rt_nr_migratory;
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200465 unsigned long rt_nr_total;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100466 int overloaded;
Gregory Haskins917b6272008-12-29 09:39:53 -0500467 struct plist_head pushable_tasks;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100468#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100469 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100470 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200471 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100472 /* Nests inside the rq lock: */
Thomas Gleixner0986b112009-11-17 15:32:06 +0100473 raw_spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100474
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100475#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100476 unsigned long rt_nr_boosted;
477
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100478 struct rq *rq;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
482#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200483};
484
Gregory Haskins57d885f2008-01-25 21:08:18 +0100485#ifdef CONFIG_SMP
486
487/*
488 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100491 * exclusive cpuset is created, we also create and attach a new root-domain
492 * object.
493 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100494 */
495struct root_domain {
496 atomic_t refcount;
Rusty Russellc6c49272008-11-25 02:35:05 +1030497 cpumask_var_t span;
498 cpumask_var_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100499
Ingo Molnar0eab9142008-01-25 21:08:19 +0100500 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
503 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030504 cpumask_var_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100505 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200506#ifdef CONFIG_SMP
507 struct cpupri cpupri;
508#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100509};
510
Gregory Haskinsdc938522008-01-25 21:08:26 +0100511/*
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
514 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100515static struct root_domain def_root_domain;
516
517#endif
518
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200519/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 * This is the main, per-CPU runqueue data structure.
521 *
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
525 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700526struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200527 /* runqueue lock: */
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100528 raw_spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
533 */
534 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700537#ifdef CONFIG_NO_HZ
538 unsigned char in_nohz_recently;
539#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200542 unsigned long nr_load_updates;
543 u64 nr_switches;
544
545 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100546 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100547
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200548#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100551#endif
552#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100553 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556 /*
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
561 */
562 unsigned long nr_uninterruptible;
563
Ingo Molnar36c8b582006-07-03 00:25:41 -0700564 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800565 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200567
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200568 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200569
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 atomic_t nr_iowait;
571
572#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100573 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 struct sched_domain *sd;
575
Henrik Austada0a522c2009-02-13 20:35:45 +0100576 unsigned char idle_at_tick;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 /* For active balancing */
Gregory Haskins3f029d32009-07-29 11:08:47 -0400578 int post_schedule;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 int active_balance;
580 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200581 /* cpu of this runqueue: */
582 int cpu;
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400583 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
Peter Zijlstraa8a51d52008-06-27 13:41:26 +0200585 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
Ingo Molnar36c8b582006-07-03 00:25:41 -0700587 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 struct list_head migration_queue;
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200589
590 u64 rt_avg;
591 u64 age_stamp;
Mike Galbraith1b9508f2009-11-04 17:53:50 +0100592 u64 idle_stamp;
593 u64 avg_idle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594#endif
595
Thomas Gleixnerdce48a82009-04-11 10:43:41 +0200596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
599
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100600#ifdef CONFIG_SCHED_HRTICK
Peter Zijlstra31656512008-07-18 18:01:23 +0200601#ifdef CONFIG_SMP
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
604#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100605 struct hrtimer hrtick_timer;
606#endif
607
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608#ifdef CONFIG_SCHEDSTATS
609 /* latency stats */
610 struct sched_info rq_sched_info;
Ken Chen9c2c4802008-12-16 23:41:22 -0800611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
614 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200615 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
617 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
622 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200625
626 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200627 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628#endif
629};
630
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700631static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Peter Zijlstra7d478722009-09-14 19:55:44 +0200633static inline
634void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnardd41f592007-07-09 18:51:59 +0200635{
Peter Zijlstra7d478722009-09-14 19:55:44 +0200636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
Ingo Molnardd41f592007-07-09 18:51:59 +0200637}
638
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700639static inline int cpu_of(struct rq *rq)
640{
641#ifdef CONFIG_SMP
642 return rq->cpu;
643#else
644 return 0;
645#endif
646}
647
Ingo Molnar20d315d2007-07-09 18:51:58 +0200648/*
Nick Piggin674311d2005-06-25 14:57:27 -0700649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700650 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700651 *
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
654 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700655#define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
658#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659#define this_rq() (&__get_cpu_var(runqueues))
660#define task_rq(p) cpu_rq(task_cpu(p))
661#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
Hitoshi Mitake54d35f22009-06-29 14:44:57 +0900662#define raw_rq() (&__raw_get_cpu_var(runqueues))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Ingo Molnaraa9c4c02008-12-17 14:10:57 +0100664inline void update_rq_clock(struct rq *rq)
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200665{
666 rq->clock = sched_clock_cpu(cpu_of(rq));
667}
668
Ingo Molnare436d802007-07-19 21:28:35 +0200669/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 */
672#ifdef CONFIG_SCHED_DEBUG
673# define const_debug __read_mostly
674#else
675# define const_debug static const
676#endif
677
Ingo Molnar017730c2008-05-12 21:20:52 +0200678/**
679 * runqueue_is_locked
Randy Dunlape17b38b2009-10-11 19:12:00 -0700680 * @cpu: the processor in question.
Ingo Molnar017730c2008-05-12 21:20:52 +0200681 *
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
685 */
Andrew Morton89f19f02009-09-19 11:55:44 -0700686int runqueue_is_locked(int cpu)
Ingo Molnar017730c2008-05-12 21:20:52 +0200687{
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
Ingo Molnar017730c2008-05-12 21:20:52 +0200689}
690
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200691/*
692 * Debugging: various feature bits
693 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200694
695#define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
697
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200698enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200699#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200700};
701
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200702#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200703
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200704#define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
706
707const_debug unsigned int sysctl_sched_features =
708#include "sched_features.h"
709 0;
710
711#undef SCHED_FEAT
712
713#ifdef CONFIG_SCHED_DEBUG
714#define SCHED_FEAT(name, enabled) \
715 #name ,
716
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700717static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200718#include "sched_features.h"
719 NULL
720};
721
722#undef SCHED_FEAT
723
Li Zefan34f3a812008-10-30 15:23:32 +0800724static int sched_feat_show(struct seq_file *m, void *v)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200725{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200726 int i;
727
728 for (i = 0; sched_feat_names[i]; i++) {
Li Zefan34f3a812008-10-30 15:23:32 +0800729 if (!(sysctl_sched_features & (1UL << i)))
730 seq_puts(m, "NO_");
731 seq_printf(m, "%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200732 }
Li Zefan34f3a812008-10-30 15:23:32 +0800733 seq_puts(m, "\n");
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200734
Li Zefan34f3a812008-10-30 15:23:32 +0800735 return 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200736}
737
738static ssize_t
739sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
741{
742 char buf[64];
743 char *cmp = buf;
744 int neg = 0;
745 int i;
746
747 if (cnt > 63)
748 cnt = 63;
749
750 if (copy_from_user(&buf, ubuf, cnt))
751 return -EFAULT;
752
753 buf[cnt] = 0;
754
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200755 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200756 neg = 1;
757 cmp += 3;
758 }
759
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
762
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 if (neg)
765 sysctl_sched_features &= ~(1UL << i);
766 else
767 sysctl_sched_features |= (1UL << i);
768 break;
769 }
770 }
771
772 if (!sched_feat_names[i])
773 return -EINVAL;
774
Jan Blunck42994722009-11-20 17:40:37 +0100775 *ppos += cnt;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200776
777 return cnt;
778}
779
Li Zefan34f3a812008-10-30 15:23:32 +0800780static int sched_feat_open(struct inode *inode, struct file *filp)
781{
782 return single_open(filp, sched_feat_show, NULL);
783}
784
Alexey Dobriyan828c0952009-10-01 15:43:56 -0700785static const struct file_operations sched_feat_fops = {
Li Zefan34f3a812008-10-30 15:23:32 +0800786 .open = sched_feat_open,
787 .write = sched_feat_write,
788 .read = seq_read,
789 .llseek = seq_lseek,
790 .release = single_release,
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200791};
792
793static __init int sched_init_debug(void)
794{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200795 debugfs_create_file("sched_features", 0644, NULL, NULL,
796 &sched_feat_fops);
797
798 return 0;
799}
800late_initcall(sched_init_debug);
801
802#endif
803
804#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200805
806/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
809 */
810const_debug unsigned int sysctl_sched_nr_migrate = 32;
811
812/*
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200813 * ratelimit for updating the group shares.
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200814 * default: 0.25ms
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200815 */
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200816unsigned int sysctl_sched_shares_ratelimit = 250000;
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +0100817unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200818
819/*
Peter Zijlstraffda12a2008-10-17 19:27:02 +0200820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
822 * default: 4
823 */
824unsigned int sysctl_sched_shares_thresh = 4;
825
826/*
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200827 * period over which we average the RT time consumption, measured
828 * in ms.
829 *
830 * default: 1s
831 */
832const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
833
834/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100835 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100836 * default: 1s
837 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100838unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100839
Ingo Molnar6892b752008-02-13 14:02:36 +0100840static __read_mostly int scheduler_running;
841
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100842/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100843 * part of the period that we allow rt tasks to run in us.
844 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100845 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100846int sysctl_sched_rt_runtime = 950000;
847
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200848static inline u64 global_rt_period(void)
849{
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
851}
852
853static inline u64 global_rt_runtime(void)
854{
roel kluine26873b2008-07-22 16:51:15 -0400855 if (sysctl_sched_rt_runtime < 0)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200856 return RUNTIME_INF;
857
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
859}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100860
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700862# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700864#ifndef finish_arch_switch
865# define finish_arch_switch(prev) do { } while (0)
866#endif
867
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100868static inline int task_current(struct rq *rq, struct task_struct *p)
869{
870 return rq->curr == p;
871}
872
Nick Piggin4866cde2005-06-25 14:57:23 -0700873#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700874static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700875{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100876 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700877}
878
Ingo Molnar70b97a72006-07-03 00:25:42 -0700879static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700880{
881}
882
Ingo Molnar70b97a72006-07-03 00:25:42 -0700883static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700884{
Ingo Molnarda04c032005-09-13 11:17:59 +0200885#ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100887 rq->lock.owner = current;
Ingo Molnarda04c032005-09-13 11:17:59 +0200888#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700889 /*
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
892 * prev into current:
893 */
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100896 raw_spin_unlock_irq(&rq->lock);
Nick Piggin4866cde2005-06-25 14:57:23 -0700897}
898
899#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700900static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700901{
902#ifdef CONFIG_SMP
903 return p->oncpu;
904#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100905 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700906#endif
907}
908
Ingo Molnar70b97a72006-07-03 00:25:42 -0700909static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700910{
911#ifdef CONFIG_SMP
912 /*
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
915 * here.
916 */
917 next->oncpu = 1;
918#endif
919#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100920 raw_spin_unlock_irq(&rq->lock);
Nick Piggin4866cde2005-06-25 14:57:23 -0700921#else
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100922 raw_spin_unlock(&rq->lock);
Nick Piggin4866cde2005-06-25 14:57:23 -0700923#endif
924}
925
Ingo Molnar70b97a72006-07-03 00:25:42 -0700926static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700927{
928#ifdef CONFIG_SMP
929 /*
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
932 * finished.
933 */
934 smp_wmb();
935 prev->oncpu = 0;
936#endif
937#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 local_irq_enable();
939#endif
940}
941#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700942
943/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
946 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700947static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700948 __acquires(rq->lock)
949{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200950 for (;;) {
951 struct rq *rq = task_rq(p);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100952 raw_spin_lock(&rq->lock);
Andi Kleen3a5c3592007-10-15 17:00:14 +0200953 if (likely(rq == task_rq(p)))
954 return rq;
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100955 raw_spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700956 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700957}
958
959/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100961 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 * explicitly disabling preemption.
963 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700964static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 __acquires(rq->lock)
966{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700967 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968
Andi Kleen3a5c3592007-10-15 17:00:14 +0200969 for (;;) {
970 local_irq_save(*flags);
971 rq = task_rq(p);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100972 raw_spin_lock(&rq->lock);
Andi Kleen3a5c3592007-10-15 17:00:14 +0200973 if (likely(rq == task_rq(p)))
974 return rq;
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100975 raw_spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977}
978
Oleg Nesterovad474ca2008-11-10 15:39:30 +0100979void task_rq_unlock_wait(struct task_struct *p)
980{
981 struct rq *rq = task_rq(p);
982
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100984 raw_spin_unlock_wait(&rq->lock);
Oleg Nesterovad474ca2008-11-10 15:39:30 +0100985}
986
Alexey Dobriyana9957442007-10-15 17:00:13 +0200987static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700988 __releases(rq->lock)
989{
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100990 raw_spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700991}
992
Ingo Molnar70b97a72006-07-03 00:25:42 -0700993static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 __releases(rq->lock)
995{
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100996 raw_spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997}
998
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -08001000 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001002static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 __acquires(rq->lock)
1004{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001005 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006
1007 local_irq_disable();
1008 rq = this_rq();
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001009 raw_spin_lock(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010
1011 return rq;
1012}
1013
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001014#ifdef CONFIG_SCHED_HRTICK
1015/*
1016 * Use HR-timers to deliver accurate preemption points.
1017 *
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * reschedule event.
1021 *
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * rq->lock.
1024 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001025
1026/*
1027 * Use hrtick when:
1028 * - enabled by features
1029 * - hrtimer is actually high res
1030 */
1031static inline int hrtick_enabled(struct rq *rq)
1032{
1033 if (!sched_feat(HRTICK))
1034 return 0;
Ingo Molnarba420592008-07-20 11:02:06 +02001035 if (!cpu_active(cpu_of(rq)))
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001036 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1038}
1039
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001040static void hrtick_clear(struct rq *rq)
1041{
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1044}
1045
1046/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1049 */
1050static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051{
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001056 raw_spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001057 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001059 raw_spin_unlock(&rq->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001060
1061 return HRTIMER_NORESTART;
1062}
1063
Rabin Vincent95e904c2008-05-11 05:55:33 +05301064#ifdef CONFIG_SMP
Peter Zijlstra31656512008-07-18 18:01:23 +02001065/*
1066 * called from hardirq (IPI) context
1067 */
1068static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001069{
Peter Zijlstra31656512008-07-18 18:01:23 +02001070 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001071
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001072 raw_spin_lock(&rq->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001075 raw_spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001076}
1077
Peter Zijlstra31656512008-07-18 18:01:23 +02001078/*
1079 * Called to set the hrtick timer state.
1080 *
1081 * called with rq->lock held and irqs disabled
1082 */
1083static void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001084{
Peter Zijlstra31656512008-07-18 18:01:23 +02001085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001087
Arjan van de Vencc584b22008-09-01 15:02:30 -07001088 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +02001089
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
Peter Zijlstra6e275632009-02-25 13:59:48 +01001093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001094 rq->hrtick_csd_pending = 1;
1095 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001096}
1097
1098static int
1099hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100{
1101 int cpu = (int)(long)hcpu;
1102
1103 switch (action) {
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD:
1109 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +02001110 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001111 return NOTIFY_OK;
1112 }
1113
1114 return NOTIFY_DONE;
1115}
1116
Rakib Mullickfa748202008-09-22 14:55:45 -07001117static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001118{
1119 hotcpu_notifier(hotplug_hrtick, 0);
1120}
Peter Zijlstra31656512008-07-18 18:01:23 +02001121#else
1122/*
1123 * Called to set the hrtick timer state.
1124 *
1125 * called with rq->lock held and irqs disabled
1126 */
1127static void hrtick_start(struct rq *rq, u64 delay)
1128{
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +01001129 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +05301130 HRTIMER_MODE_REL_PINNED, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001131}
1132
Andrew Morton006c75f2008-09-22 14:55:46 -07001133static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +02001134{
1135}
Rabin Vincent95e904c2008-05-11 05:55:33 +05301136#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001137
1138static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001139{
Peter Zijlstra31656512008-07-18 18:01:23 +02001140#ifdef CONFIG_SMP
1141 rq->hrtick_csd_pending = 0;
1142
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1146#endif
1147
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001150}
Andrew Morton006c75f2008-09-22 14:55:46 -07001151#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001152static inline void hrtick_clear(struct rq *rq)
1153{
1154}
1155
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001156static inline void init_rq_hrtick(struct rq *rq)
1157{
1158}
1159
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001160static inline void init_hrtick(void)
1161{
1162}
Andrew Morton006c75f2008-09-22 14:55:46 -07001163#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001164
Ingo Molnar1b9f19c22007-07-09 18:51:59 +02001165/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001166 * resched_task - mark a task 'to be rescheduled now'.
1167 *
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1170 * the target CPU.
1171 */
1172#ifdef CONFIG_SMP
1173
1174#ifndef tsk_is_polling
1175#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176#endif
1177
Peter Zijlstra31656512008-07-18 18:01:23 +02001178static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001179{
1180 int cpu;
1181
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001182 assert_raw_spin_locked(&task_rq(p)->lock);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001183
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001184 if (test_tsk_need_resched(p))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001185 return;
1186
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001187 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001188
1189 cpu = task_cpu(p);
1190 if (cpu == smp_processor_id())
1191 return;
1192
1193 /* NEED_RESCHED must be visible before we test polling */
1194 smp_mb();
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1197}
1198
1199static void resched_cpu(int cpu)
1200{
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1203
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001204 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001205 return;
1206 resched_task(cpu_curr(cpu));
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001207 raw_spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001208}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001209
1210#ifdef CONFIG_NO_HZ
1211/*
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1220 */
1221void wake_up_idle_cpu(int cpu)
1222{
1223 struct rq *rq = cpu_rq(cpu);
1224
1225 if (cpu == smp_processor_id())
1226 return;
1227
1228 /*
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1234 */
1235 if (rq->curr != rq->idle)
1236 return;
1237
1238 /*
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1242 */
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001243 set_tsk_need_resched(rq->idle);
Thomas Gleixner06d83082008-03-22 09:20:24 +01001244
1245 /* NEED_RESCHED must be visible before we test polling */
1246 smp_mb();
1247 if (!tsk_is_polling(rq->idle))
1248 smp_send_reschedule(cpu);
1249}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001250#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001251
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001252static u64 sched_avg_period(void)
1253{
1254 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1255}
1256
1257static void sched_avg_update(struct rq *rq)
1258{
1259 s64 period = sched_avg_period();
1260
1261 while ((s64)(rq->clock - rq->age_stamp) > period) {
1262 rq->age_stamp += period;
1263 rq->rt_avg /= 2;
1264 }
1265}
1266
1267static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1268{
1269 rq->rt_avg += rt_delta;
1270 sched_avg_update(rq);
1271}
1272
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001273#else /* !CONFIG_SMP */
Peter Zijlstra31656512008-07-18 18:01:23 +02001274static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001275{
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001276 assert_raw_spin_locked(&task_rq(p)->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001277 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001278}
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001279
1280static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1281{
1282}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001283#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001284
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001285#if BITS_PER_LONG == 32
1286# define WMULT_CONST (~0UL)
1287#else
1288# define WMULT_CONST (1UL << 32)
1289#endif
1290
1291#define WMULT_SHIFT 32
1292
Ingo Molnar194081e2007-08-09 11:16:51 +02001293/*
1294 * Shift right and round:
1295 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001296#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001297
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001298/*
1299 * delta *= weight / lw
1300 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001301static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001302calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303 struct load_weight *lw)
1304{
1305 u64 tmp;
1306
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001307 if (!lw->inv_weight) {
1308 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309 lw->inv_weight = 1;
1310 else
1311 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312 / (lw->weight+1);
1313 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001314
1315 tmp = (u64)delta_exec * weight;
1316 /*
1317 * Check whether we'd overflow the 64-bit multiplication:
1318 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001319 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001320 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001321 WMULT_SHIFT/2);
1322 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001323 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001324
Ingo Molnarecf691d2007-08-02 17:41:40 +02001325 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001326}
1327
Ingo Molnar10919852007-10-15 17:00:04 +02001328static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001329{
1330 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001331 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001332}
1333
Ingo Molnar10919852007-10-15 17:00:04 +02001334static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001335{
1336 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001337 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001338}
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001345 * scaled version of the new time slice allocation that they receive on time
1346 * slice expiry etc.
1347 */
1348
Peter Zijlstracce7ade2009-01-15 14:53:37 +01001349#define WEIGHT_IDLEPRIO 3
1350#define WMULT_IDLEPRIO 1431655765
Ingo Molnardd41f592007-07-09 18:51:59 +02001351
1352/*
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1357 *
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001363 */
1364static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001373};
1374
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001375/*
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 *
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1381 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001382static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001391};
Peter Williams2dd73a42006-06-27 02:54:34 -07001392
Ingo Molnardd41f592007-07-09 18:51:59 +02001393static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1394
1395/*
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1399 */
1400struct rq_iterator {
1401 void *arg;
1402 struct task_struct *(*start)(void *);
1403 struct task_struct *(*next)(void *);
1404};
1405
Peter Williamse1d14842007-10-24 18:23:51 +02001406#ifdef CONFIG_SMP
1407static unsigned long
1408balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 unsigned long max_load_move, struct sched_domain *sd,
1410 enum cpu_idle_type idle, int *all_pinned,
1411 int *this_best_prio, struct rq_iterator *iterator);
1412
1413static int
1414iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 struct sched_domain *sd, enum cpu_idle_type idle,
1416 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001417#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001418
Bharata B Raoef12fef2009-03-31 10:02:22 +05301419/* Time spent by the tasks of the cpu accounting group executing in ... */
1420enum cpuacct_stat_index {
1421 CPUACCT_STAT_USER, /* ... user mode */
1422 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1423
1424 CPUACCT_STAT_NSTATS,
1425};
1426
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001427#ifdef CONFIG_CGROUP_CPUACCT
1428static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
Bharata B Raoef12fef2009-03-31 10:02:22 +05301429static void cpuacct_update_stats(struct task_struct *tsk,
1430 enum cpuacct_stat_index idx, cputime_t val);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001431#else
1432static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
Bharata B Raoef12fef2009-03-31 10:02:22 +05301433static inline void cpuacct_update_stats(struct task_struct *tsk,
1434 enum cpuacct_stat_index idx, cputime_t val) {}
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001435#endif
1436
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001437static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1438{
1439 update_load_add(&rq->load, load);
1440}
1441
1442static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1443{
1444 update_load_sub(&rq->load, load);
1445}
1446
Ingo Molnar7940ca32008-08-19 13:40:47 +02001447#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001448typedef int (*tg_visitor)(struct task_group *, void *);
1449
1450/*
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1453 */
1454static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1455{
1456 struct task_group *parent, *child;
1457 int ret;
1458
1459 rcu_read_lock();
1460 parent = &root_task_group;
1461down:
1462 ret = (*down)(parent, data);
1463 if (ret)
1464 goto out_unlock;
1465 list_for_each_entry_rcu(child, &parent->children, siblings) {
1466 parent = child;
1467 goto down;
1468
1469up:
1470 continue;
1471 }
1472 ret = (*up)(parent, data);
1473 if (ret)
1474 goto out_unlock;
1475
1476 child = parent;
1477 parent = parent->parent;
1478 if (parent)
1479 goto up;
1480out_unlock:
1481 rcu_read_unlock();
1482
1483 return ret;
1484}
1485
1486static int tg_nop(struct task_group *tg, void *data)
1487{
1488 return 0;
1489}
1490#endif
1491
Gregory Haskinse7693a32008-01-25 21:08:09 +01001492#ifdef CONFIG_SMP
Peter Zijlstraf5f08f32009-09-10 13:35:28 +02001493/* Used instead of source_load when we know the type == 0 */
1494static unsigned long weighted_cpuload(const int cpu)
1495{
1496 return cpu_rq(cpu)->load.weight;
1497}
1498
1499/*
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1502 *
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1505 */
1506static unsigned long source_load(int cpu, int type)
1507{
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1510
1511 if (type == 0 || !sched_feat(LB_BIAS))
1512 return total;
1513
1514 return min(rq->cpu_load[type-1], total);
1515}
1516
1517/*
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1520 */
1521static unsigned long target_load(int cpu, int type)
1522{
1523 struct rq *rq = cpu_rq(cpu);
1524 unsigned long total = weighted_cpuload(cpu);
1525
1526 if (type == 0 || !sched_feat(LB_BIAS))
1527 return total;
1528
1529 return max(rq->cpu_load[type-1], total);
1530}
1531
Peter Zijlstraae154be2009-09-10 14:40:57 +02001532static struct sched_group *group_of(int cpu)
1533{
1534 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1535
1536 if (!sd)
1537 return NULL;
1538
1539 return sd->groups;
1540}
1541
1542static unsigned long power_of(int cpu)
1543{
1544 struct sched_group *group = group_of(cpu);
1545
1546 if (!group)
1547 return SCHED_LOAD_SCALE;
1548
1549 return group->cpu_power;
1550}
1551
Gregory Haskinse7693a32008-01-25 21:08:09 +01001552static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001553
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001554static unsigned long cpu_avg_load_per_task(int cpu)
1555{
1556 struct rq *rq = cpu_rq(cpu);
Ingo Molnaraf6d5962008-11-29 20:45:15 +01001557 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001558
Steven Rostedt4cd42622008-11-26 21:04:24 -05001559 if (nr_running)
1560 rq->avg_load_per_task = rq->load.weight / nr_running;
Balbir Singha2d47772008-11-12 16:19:00 +05301561 else
1562 rq->avg_load_per_task = 0;
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001563
1564 return rq->avg_load_per_task;
1565}
1566
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001567#ifdef CONFIG_FAIR_GROUP_SCHED
1568
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001569static __read_mostly unsigned long *update_shares_data;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001570
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001571static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1572
1573/*
1574 * Calculate and set the cpu's group shares.
1575 */
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001576static void update_group_shares_cpu(struct task_group *tg, int cpu,
1577 unsigned long sd_shares,
1578 unsigned long sd_rq_weight,
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001579 unsigned long *usd_rq_weight)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001580{
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001581 unsigned long shares, rq_weight;
Peter Zijlstraa50042782009-07-27 14:04:49 +02001582 int boost = 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001583
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001584 rq_weight = usd_rq_weight[cpu];
Peter Zijlstraa50042782009-07-27 14:04:49 +02001585 if (!rq_weight) {
1586 boost = 1;
1587 rq_weight = NICE_0_LOAD;
1588 }
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001589
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001590 /*
Peter Zijlstraa8af7242009-08-21 13:58:54 +02001591 * \Sum_j shares_j * rq_weight_i
1592 * shares_i = -----------------------------
1593 * \Sum_j rq_weight_j
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001594 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001595 shares = (sd_shares * rq_weight) / sd_rq_weight;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001596 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001597
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001598 if (abs(shares - tg->se[cpu]->load.weight) >
1599 sysctl_sched_shares_thresh) {
1600 struct rq *rq = cpu_rq(cpu);
1601 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001602
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001603 raw_spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001604 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
Peter Zijlstraa50042782009-07-27 14:04:49 +02001605 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001606 __set_se_shares(tg->se[cpu], shares);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001607 raw_spin_unlock_irqrestore(&rq->lock, flags);
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001608 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001609}
1610
1611/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001615 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001616static int tg_shares_up(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001617{
Peter Zijlstracd8ad402009-12-03 18:00:07 +01001618 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001619 unsigned long *usd_rq_weight;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001620 struct sched_domain *sd = data;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001621 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001622 int i;
1623
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001624 if (!tg->se[0])
1625 return 0;
1626
1627 local_irq_save(flags);
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001628 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001629
Rusty Russell758b2cd2008-11-25 02:35:04 +10301630 for_each_cpu(i, sched_domain_span(sd)) {
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001631 weight = tg->cfs_rq[i]->load.weight;
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001632 usd_rq_weight[i] = weight;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001633
Peter Zijlstracd8ad402009-12-03 18:00:07 +01001634 rq_weight += weight;
Ken Chenec4e0e22008-11-18 22:41:57 -08001635 /*
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1639 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001640 if (!weight)
1641 weight = NICE_0_LOAD;
1642
Peter Zijlstracd8ad402009-12-03 18:00:07 +01001643 sum_weight += weight;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001644 shares += tg->cfs_rq[i]->shares;
1645 }
1646
Peter Zijlstracd8ad402009-12-03 18:00:07 +01001647 if (!rq_weight)
1648 rq_weight = sum_weight;
1649
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001650 if ((!shares && rq_weight) || shares > tg->shares)
1651 shares = tg->shares;
1652
1653 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1654 shares = tg->shares;
1655
Rusty Russell758b2cd2008-11-25 02:35:04 +10301656 for_each_cpu(i, sched_domain_span(sd))
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001657 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001658
1659 local_irq_restore(flags);
Peter Zijlstraeb755802008-08-19 12:33:05 +02001660
1661 return 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001662}
1663
1664/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001668 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001669static int tg_load_down(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001670{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001671 unsigned long load;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001672 long cpu = (long)data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001673
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001674 if (!tg->parent) {
1675 load = cpu_rq(cpu)->load.weight;
1676 } else {
1677 load = tg->parent->cfs_rq[cpu]->h_load;
1678 load *= tg->cfs_rq[cpu]->shares;
1679 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1680 }
1681
1682 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001683
Peter Zijlstraeb755802008-08-19 12:33:05 +02001684 return 0;
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001685}
1686
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001687static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001688{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001689 s64 elapsed;
1690 u64 now;
1691
1692 if (root_task_group_empty())
1693 return;
1694
1695 now = cpu_clock(raw_smp_processor_id());
1696 elapsed = now - sd->last_update;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001697
1698 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1699 sd->last_update = now;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001700 walk_tg_tree(tg_nop, tg_shares_up, sd);
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001701 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001702}
1703
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001704static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1705{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001706 if (root_task_group_empty())
1707 return;
1708
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001709 raw_spin_unlock(&rq->lock);
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001710 update_shares(sd);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001711 raw_spin_lock(&rq->lock);
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001712}
1713
Peter Zijlstraeb755802008-08-19 12:33:05 +02001714static void update_h_load(long cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001715{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001716 if (root_task_group_empty())
1717 return;
1718
Peter Zijlstraeb755802008-08-19 12:33:05 +02001719 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001720}
1721
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001722#else
1723
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001724static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001725{
1726}
1727
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001728static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1729{
1730}
1731
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001732#endif
1733
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001734#ifdef CONFIG_PREEMPT
1735
Peter Zijlstrab78bb862009-09-15 14:23:18 +02001736static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1737
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001738/*
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations. This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below. However, it
1744 * also adds more overhead and therefore may reduce throughput.
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001745 */
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001746static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1747 __releases(this_rq->lock)
1748 __acquires(busiest->lock)
1749 __acquires(this_rq->lock)
1750{
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001751 raw_spin_unlock(&this_rq->lock);
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001752 double_rq_lock(this_rq, busiest);
1753
1754 return 1;
1755}
1756
1757#else
1758/*
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry. This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1764 */
1765static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001766 __releases(this_rq->lock)
1767 __acquires(busiest->lock)
1768 __acquires(this_rq->lock)
1769{
1770 int ret = 0;
1771
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001772 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001773 if (busiest < this_rq) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001774 raw_spin_unlock(&this_rq->lock);
1775 raw_spin_lock(&busiest->lock);
1776 raw_spin_lock_nested(&this_rq->lock,
1777 SINGLE_DEPTH_NESTING);
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001778 ret = 1;
1779 } else
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001780 raw_spin_lock_nested(&busiest->lock,
1781 SINGLE_DEPTH_NESTING);
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001782 }
1783 return ret;
1784}
1785
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001786#endif /* CONFIG_PREEMPT */
1787
1788/*
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1790 */
1791static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1792{
1793 if (unlikely(!irqs_disabled())) {
1794 /* printk() doesn't work good under rq->lock */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001795 raw_spin_unlock(&this_rq->lock);
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001796 BUG_ON(1);
1797 }
1798
1799 return _double_lock_balance(this_rq, busiest);
1800}
1801
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001802static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1803 __releases(busiest->lock)
1804{
Thomas Gleixner05fa7852009-11-17 14:28:38 +01001805 raw_spin_unlock(&busiest->lock);
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1807}
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001808#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001809
1810#ifdef CONFIG_FAIR_GROUP_SCHED
1811static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812{
Vegard Nossum30432092008-06-27 21:35:50 +02001813#ifdef CONFIG_SMP
Ingo Molnar34e83e82008-06-27 15:42:36 +02001814 cfs_rq->shares = shares;
1815#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001816}
1817#endif
1818
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02001819static void calc_load_account_active(struct rq *this_rq);
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01001820static void update_sysctl(void);
Christian Ehrhardtacb4a842009-11-30 12:16:48 +01001821static int get_update_sysctl_factor(void);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02001822
Peter Zijlstracd29fe62009-11-27 17:32:46 +01001823static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1824{
1825 set_task_rq(p, cpu);
1826#ifdef CONFIG_SMP
1827 /*
1828 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1829 * successfuly executed on another CPU. We must ensure that updates of
1830 * per-task data have been completed by this moment.
1831 */
1832 smp_wmb();
1833 task_thread_info(p)->cpu = cpu;
1834#endif
1835}
Gregory Haskinse7693a32008-01-25 21:08:09 +01001836
Ingo Molnardd41f592007-07-09 18:51:59 +02001837#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001838#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001839#include "sched_fair.c"
1840#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001841#ifdef CONFIG_SCHED_DEBUG
1842# include "sched_debug.c"
1843#endif
1844
1845#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001846#define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001848
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001849static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001850{
1851 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001852}
1853
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001854static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001855{
1856 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001857}
1858
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001859static void set_load_weight(struct task_struct *p)
1860{
1861 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001862 p->se.load.weight = prio_to_weight[0] * 2;
1863 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1864 return;
1865 }
1866
1867 /*
1868 * SCHED_IDLE tasks get minimal weight:
1869 */
1870 if (p->policy == SCHED_IDLE) {
1871 p->se.load.weight = WEIGHT_IDLEPRIO;
1872 p->se.load.inv_weight = WMULT_IDLEPRIO;
1873 return;
1874 }
1875
1876 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001878}
1879
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001880static void update_avg(u64 *avg, u64 sample)
1881{
1882 s64 diff = sample - *avg;
1883 *avg += diff >> 3;
1884}
1885
Ingo Molnar8159f872007-08-09 11:16:49 +02001886static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001887{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001888 if (wakeup)
1889 p->se.start_runtime = p->se.sum_exec_runtime;
1890
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001891 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001892 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001893 p->se.on_rq = 1;
1894}
1895
Ingo Molnar69be72c2007-08-09 11:16:49 +02001896static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001897{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001898 if (sleep) {
1899 if (p->se.last_wakeup) {
1900 update_avg(&p->se.avg_overlap,
1901 p->se.sum_exec_runtime - p->se.last_wakeup);
1902 p->se.last_wakeup = 0;
1903 } else {
1904 update_avg(&p->se.avg_wakeup,
1905 sysctl_sched_wakeup_granularity);
1906 }
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001907 }
1908
Ankita Garg46ac22b2008-07-01 14:30:06 +05301909 sched_info_dequeued(p);
Ingo Molnarf02231e2007-08-09 11:16:48 +02001910 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001911 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001912}
1913
1914/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001915 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001916 */
Ingo Molnar14531182007-07-09 18:51:59 +02001917static inline int __normal_prio(struct task_struct *p)
1918{
Ingo Molnardd41f592007-07-09 18:51:59 +02001919 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001920}
1921
1922/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001923 * Calculate the expected normal priority: i.e. priority
1924 * without taking RT-inheritance into account. Might be
1925 * boosted by interactivity modifiers. Changes upon fork,
1926 * setprio syscalls, and whenever the interactivity
1927 * estimator recalculates.
1928 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001929static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001930{
1931 int prio;
1932
Ingo Molnare05606d2007-07-09 18:51:59 +02001933 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001934 prio = MAX_RT_PRIO-1 - p->rt_priority;
1935 else
1936 prio = __normal_prio(p);
1937 return prio;
1938}
1939
1940/*
1941 * Calculate the current priority, i.e. the priority
1942 * taken into account by the scheduler. This value might
1943 * be boosted by RT tasks, or might be boosted by
1944 * interactivity modifiers. Will be RT if the task got
1945 * RT-boosted. If not then it returns p->normal_prio.
1946 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001947static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001948{
1949 p->normal_prio = normal_prio(p);
1950 /*
1951 * If we are RT tasks or we were boosted to RT priority,
1952 * keep the priority unchanged. Otherwise, update priority
1953 * to the normal priority:
1954 */
1955 if (!rt_prio(p->prio))
1956 return p->normal_prio;
1957 return p->prio;
1958}
1959
1960/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001961 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001963static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001965 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001966 rq->nr_uninterruptible--;
1967
Ingo Molnar8159f872007-08-09 11:16:49 +02001968 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001969 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970}
1971
1972/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 * deactivate_task - remove a task from the runqueue.
1974 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001975static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001977 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001978 rq->nr_uninterruptible++;
1979
Ingo Molnar69be72c2007-08-09 11:16:49 +02001980 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001981 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982}
1983
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984/**
1985 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question.
1987 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001988inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989{
1990 return cpu_curr(task_cpu(p)) == p;
1991}
1992
Steven Rostedtcb469842008-01-25 21:08:22 +01001993static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1994 const struct sched_class *prev_class,
1995 int oldprio, int running)
1996{
1997 if (prev_class != p->sched_class) {
1998 if (prev_class->switched_from)
1999 prev_class->switched_from(rq, p, running);
2000 p->sched_class->switched_to(rq, p, running);
2001 } else
2002 p->sched_class->prio_changed(rq, p, oldprio, running);
2003}
2004
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01002005/**
2006 * kthread_bind - bind a just-created kthread to a cpu.
Randy Dunlap968c8642009-11-06 15:31:08 -08002007 * @p: thread created by kthread_create().
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01002008 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2009 *
2010 * Description: This function is equivalent to set_cpus_allowed(),
2011 * except that @cpu doesn't need to be online, and the thread must be
2012 * stopped (i.e., just returned from kthread_create()).
2013 *
2014 * Function lives here instead of kthread.c because it messes with
2015 * scheduler internals which require locking.
2016 */
2017void kthread_bind(struct task_struct *p, unsigned int cpu)
2018{
2019 struct rq *rq = cpu_rq(cpu);
2020 unsigned long flags;
2021
2022 /* Must have done schedule() in kthread() before we set_task_cpu */
2023 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2024 WARN_ON(1);
2025 return;
2026 }
2027
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002028 raw_spin_lock_irqsave(&rq->lock, flags);
Mike Galbraith055a0082009-11-12 11:07:44 +01002029 update_rq_clock(rq);
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01002030 set_task_cpu(p, cpu);
2031 p->cpus_allowed = cpumask_of_cpu(cpu);
2032 p->rt.nr_cpus_allowed = 1;
2033 p->flags |= PF_THREAD_BOUND;
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002034 raw_spin_unlock_irqrestore(&rq->lock, flags);
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01002035}
2036EXPORT_SYMBOL(kthread_bind);
2037
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038#ifdef CONFIG_SMP
Ingo Molnarcc367732007-10-15 17:00:18 +02002039/*
2040 * Is this task likely cache-hot:
2041 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002042static int
Ingo Molnarcc367732007-10-15 17:00:18 +02002043task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2044{
2045 s64 delta;
2046
Ingo Molnarf540a602008-03-15 17:10:34 +01002047 /*
2048 * Buddy candidates are cache hot:
2049 */
Mike Galbraithf685cea2009-10-23 23:09:22 +02002050 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
Peter Zijlstra47932412008-11-04 21:25:09 +01002051 (&p->se == cfs_rq_of(&p->se)->next ||
2052 &p->se == cfs_rq_of(&p->se)->last))
Ingo Molnarf540a602008-03-15 17:10:34 +01002053 return 1;
2054
Ingo Molnarcc367732007-10-15 17:00:18 +02002055 if (p->sched_class != &fair_sched_class)
2056 return 0;
2057
Ingo Molnar6bc16652007-10-15 17:00:18 +02002058 if (sysctl_sched_migration_cost == -1)
2059 return 1;
2060 if (sysctl_sched_migration_cost == 0)
2061 return 0;
2062
Ingo Molnarcc367732007-10-15 17:00:18 +02002063 delta = now - p->se.exec_start;
2064
2065 return delta < (s64)sysctl_sched_migration_cost;
2066}
2067
2068
Ingo Molnardd41f592007-07-09 18:51:59 +02002069void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02002070{
Ingo Molnardd41f592007-07-09 18:51:59 +02002071 int old_cpu = task_cpu(p);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002072 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2073 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002074
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08002075 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01002076
Ingo Molnarcc367732007-10-15 17:00:18 +02002077 if (old_cpu != new_cpu) {
Ingo Molnar6c594c22008-12-14 12:34:15 +01002078 p->se.nr_migrations++;
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002079 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
Peter Zijlstrae5289d42009-06-19 13:22:51 +02002080 1, 1, NULL, 0);
Ingo Molnar6c594c22008-12-14 12:34:15 +01002081 }
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002082 p->se.vruntime -= old_cfsrq->min_vruntime -
2083 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02002084
2085 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02002086}
2087
Ingo Molnar70b97a72006-07-03 00:25:42 -07002088struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090
Ingo Molnar36c8b582006-07-03 00:25:41 -07002091 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 int dest_cpu;
2093
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002095};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096
2097/*
2098 * The task's runqueue lock must be held.
2099 * Returns true if you have to wait for migration thread.
2100 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002101static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002102migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002104 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105
2106 /*
2107 * If the task is not on a runqueue (and not running), then
2108 * it is sufficient to simply update the task's cpu field.
2109 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002110 if (!p->se.on_rq && !task_running(rq, p)) {
Mike Galbraith055a0082009-11-12 11:07:44 +01002111 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 set_task_cpu(p, dest_cpu);
2113 return 0;
2114 }
2115
2116 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 req->task = p;
2118 req->dest_cpu = dest_cpu;
2119 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002120
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 return 1;
2122}
2123
2124/*
Markus Metzgera26b89f2009-04-03 16:43:34 +02002125 * wait_task_context_switch - wait for a thread to complete at least one
2126 * context switch.
2127 *
2128 * @p must not be current.
2129 */
2130void wait_task_context_switch(struct task_struct *p)
2131{
2132 unsigned long nvcsw, nivcsw, flags;
2133 int running;
2134 struct rq *rq;
2135
2136 nvcsw = p->nvcsw;
2137 nivcsw = p->nivcsw;
2138 for (;;) {
2139 /*
2140 * The runqueue is assigned before the actual context
2141 * switch. We need to take the runqueue lock.
2142 *
2143 * We could check initially without the lock but it is
2144 * very likely that we need to take the lock in every
2145 * iteration.
2146 */
2147 rq = task_rq_lock(p, &flags);
2148 running = task_running(rq, p);
2149 task_rq_unlock(rq, &flags);
2150
2151 if (likely(!running))
2152 break;
2153 /*
2154 * The switch count is incremented before the actual
2155 * context switch. We thus wait for two switches to be
2156 * sure at least one completed.
2157 */
2158 if ((p->nvcsw - nvcsw) > 1)
2159 break;
2160 if ((p->nivcsw - nivcsw) > 1)
2161 break;
2162
2163 cpu_relax();
2164 }
2165}
2166
2167/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 * wait_task_inactive - wait for a thread to unschedule.
2169 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07002170 * If @match_state is nonzero, it's the @p->state value just checked and
2171 * not expected to change. If it changes, i.e. @p might have woken up,
2172 * then return zero. When we succeed in waiting for @p to be off its CPU,
2173 * we return a positive number (its total switch count). If a second call
2174 * a short while later returns the same number, the caller can be sure that
2175 * @p has remained unscheduled the whole time.
2176 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 * The caller must ensure that the task *will* unschedule sometime soon,
2178 * else this function might spin for a *long* time. This function can't
2179 * be called with interrupts off, or it may introduce deadlock with
2180 * smp_call_function() if an IPI is sent by the same process we are
2181 * waiting to become inactive.
2182 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002183unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184{
2185 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002186 int running, on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002187 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002188 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189
Andi Kleen3a5c3592007-10-15 17:00:14 +02002190 for (;;) {
2191 /*
2192 * We do the initial early heuristics without holding
2193 * any task-queue locks at all. We'll only try to get
2194 * the runqueue lock when things look like they will
2195 * work out!
2196 */
2197 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002198
Andi Kleen3a5c3592007-10-15 17:00:14 +02002199 /*
2200 * If the task is actively running on another CPU
2201 * still, just relax and busy-wait without holding
2202 * any locks.
2203 *
2204 * NOTE! Since we don't hold any locks, it's not
2205 * even sure that "rq" stays as the right runqueue!
2206 * But we don't care, since "task_running()" will
2207 * return false if the runqueue has changed and p
2208 * is actually now running somewhere else!
2209 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002210 while (task_running(rq, p)) {
2211 if (match_state && unlikely(p->state != match_state))
2212 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02002213 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07002214 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002215
Andi Kleen3a5c3592007-10-15 17:00:14 +02002216 /*
2217 * Ok, time to look more closely! We need the rq
2218 * lock now, to be *sure*. If we're wrong, we'll
2219 * just go back and repeat.
2220 */
2221 rq = task_rq_lock(p, &flags);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002222 trace_sched_wait_task(rq, p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02002223 running = task_running(rq, p);
2224 on_rq = p->se.on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002225 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07002226 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07002227 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Andi Kleen3a5c3592007-10-15 17:00:14 +02002228 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002229
Andi Kleen3a5c3592007-10-15 17:00:14 +02002230 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07002231 * If it changed from the expected state, bail out now.
2232 */
2233 if (unlikely(!ncsw))
2234 break;
2235
2236 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02002237 * Was it really running after all now that we
2238 * checked with the proper locks actually held?
2239 *
2240 * Oops. Go back and try again..
2241 */
2242 if (unlikely(running)) {
2243 cpu_relax();
2244 continue;
2245 }
2246
2247 /*
2248 * It's not enough that it's not actively running,
2249 * it must be off the runqueue _entirely_, and not
2250 * preempted!
2251 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00002252 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02002253 * running right now), it's preempted, and we should
2254 * yield - it could be a while.
2255 */
2256 if (unlikely(on_rq)) {
2257 schedule_timeout_uninterruptible(1);
2258 continue;
2259 }
2260
2261 /*
2262 * Ahh, all good. It wasn't running, and it wasn't
2263 * runnable, which means that it will never become
2264 * running in the future either. We're all done!
2265 */
2266 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07002268
2269 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270}
2271
2272/***
2273 * kick_process - kick a running thread to enter/exit the kernel
2274 * @p: the to-be-kicked thread
2275 *
2276 * Cause a process which is running on another CPU to enter
2277 * kernel-mode, without any delay. (to get signals handled.)
2278 *
2279 * NOTE: this function doesnt have to take the runqueue lock,
2280 * because all it wants to ensure is that the remote task enters
2281 * the kernel. If the IPI races and the task has been migrated
2282 * to another CPU then no harm is done and the purpose has been
2283 * achieved as well.
2284 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002285void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286{
2287 int cpu;
2288
2289 preempt_disable();
2290 cpu = task_cpu(p);
2291 if ((cpu != smp_processor_id()) && task_curr(p))
2292 smp_send_reschedule(cpu);
2293 preempt_enable();
2294}
Rusty Russellb43e3522009-06-12 22:27:00 -06002295EXPORT_SYMBOL_GPL(kick_process);
Nick Piggin476d1392005-06-25 14:57:29 -07002296#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297
Thomas Gleixner0793a612008-12-04 20:12:29 +01002298/**
2299 * task_oncpu_function_call - call a function on the cpu on which a task runs
2300 * @p: the task to evaluate
2301 * @func: the function to be called
2302 * @info: the function call argument
2303 *
2304 * Calls the function @func when the task is currently running. This might
2305 * be on the current CPU, which just calls the function directly
2306 */
2307void task_oncpu_function_call(struct task_struct *p,
2308 void (*func) (void *info), void *info)
2309{
2310 int cpu;
2311
2312 preempt_disable();
2313 cpu = task_cpu(p);
2314 if (task_curr(p))
2315 smp_call_function_single(cpu, func, info, 1);
2316 preempt_enable();
2317}
2318
Peter Zijlstra970b13b2009-11-25 13:31:39 +01002319#ifdef CONFIG_SMP
2320static inline
2321int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2322{
2323 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2324}
2325#endif
2326
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327/***
2328 * try_to_wake_up - wake up a thread
2329 * @p: the to-be-woken-up thread
2330 * @state: the mask of task states that can be woken
2331 * @sync: do a synchronous wakeup?
2332 *
2333 * Put it on the run-queue if it's not already there. The "current"
2334 * thread is always on the run-queue (except when the actual
2335 * re-schedule is in progress), and as such you're allowed to do
2336 * the simpler "current->state = TASK_RUNNING" to mark yourself
2337 * runnable without the overhead of this.
2338 *
2339 * returns failure only if the task is already active.
2340 */
Peter Zijlstra7d478722009-09-14 19:55:44 +02002341static int try_to_wake_up(struct task_struct *p, unsigned int state,
2342 int wake_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343{
Ingo Molnarcc367732007-10-15 17:00:18 +02002344 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 unsigned long flags;
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002346 struct rq *rq, *orig_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347
Ingo Molnarb85d0662008-03-16 20:03:22 +01002348 if (!sched_feat(SYNC_WAKEUPS))
Peter Zijlstra7d478722009-09-14 19:55:44 +02002349 wake_flags &= ~WF_SYNC;
Ingo Molnarb85d0662008-03-16 20:03:22 +01002350
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002351 this_cpu = get_cpu();
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002352
Linus Torvalds04e2f172008-02-23 18:05:03 -08002353 smp_wmb();
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002354 rq = orig_rq = task_rq_lock(p, &flags);
Mike Galbraith03e89e42008-12-16 08:45:30 +01002355 update_rq_clock(rq);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002356 if (!(p->state & state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 goto out;
2358
Ingo Molnardd41f592007-07-09 18:51:59 +02002359 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 goto out_running;
2361
2362 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002363 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
2365#ifdef CONFIG_SMP
2366 if (unlikely(task_running(rq, p)))
2367 goto out_activate;
2368
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002369 /*
2370 * In order to handle concurrent wakeups and release the rq->lock
2371 * we put the task in TASK_WAKING state.
Ingo Molnareb240732009-09-16 21:09:13 +02002372 *
2373 * First fix up the nr_uninterruptible count:
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002374 */
Ingo Molnareb240732009-09-16 21:09:13 +02002375 if (task_contributes_to_load(p))
2376 rq->nr_uninterruptible--;
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002377 p->state = TASK_WAKING;
Peter Zijlstraab19cb22009-11-27 15:44:43 +01002378 __task_rq_unlock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379
Peter Zijlstra970b13b2009-11-25 13:31:39 +01002380 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
Peter Zijlstraab19cb22009-11-27 15:44:43 +01002381 if (cpu != orig_cpu)
Mike Galbraith055a0082009-11-12 11:07:44 +01002382 set_task_cpu(p, cpu);
Peter Zijlstraab19cb22009-11-27 15:44:43 +01002383
2384 rq = __task_rq_lock(p);
2385 update_rq_clock(rq);
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002386
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002387 WARN_ON(p->state != TASK_WAKING);
2388 cpu = task_cpu(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Gregory Haskinse7693a32008-01-25 21:08:09 +01002390#ifdef CONFIG_SCHEDSTATS
2391 schedstat_inc(rq, ttwu_count);
2392 if (cpu == this_cpu)
2393 schedstat_inc(rq, ttwu_local);
2394 else {
2395 struct sched_domain *sd;
2396 for_each_domain(this_cpu, sd) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Gregory Haskinse7693a32008-01-25 21:08:09 +01002398 schedstat_inc(sd, ttwu_wake_remote);
2399 break;
2400 }
2401 }
2402 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002403#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002404
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405out_activate:
2406#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002407 schedstat_inc(p, se.nr_wakeups);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002408 if (wake_flags & WF_SYNC)
Ingo Molnarcc367732007-10-15 17:00:18 +02002409 schedstat_inc(p, se.nr_wakeups_sync);
2410 if (orig_cpu != cpu)
2411 schedstat_inc(p, se.nr_wakeups_migrate);
2412 if (cpu == this_cpu)
2413 schedstat_inc(p, se.nr_wakeups_local);
2414 else
2415 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnardd41f592007-07-09 18:51:59 +02002416 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 success = 1;
2418
Peter Zijlstra831451a2009-01-14 12:39:18 +01002419 /*
2420 * Only attribute actual wakeups done by this task.
2421 */
2422 if (!in_interrupt()) {
2423 struct sched_entity *se = &current->se;
2424 u64 sample = se->sum_exec_runtime;
2425
2426 if (se->last_wakeup)
2427 sample -= se->last_wakeup;
2428 else
2429 sample -= se->start_runtime;
2430 update_avg(&se->avg_wakeup, sample);
2431
2432 se->last_wakeup = se->sum_exec_runtime;
2433 }
2434
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435out_running:
Peter Zijlstra468a15b2008-12-16 08:07:03 +01002436 trace_sched_wakeup(rq, p, success);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002437 check_preempt_curr(rq, p, wake_flags);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002438
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002440#ifdef CONFIG_SMP
2441 if (p->sched_class->task_wake_up)
2442 p->sched_class->task_wake_up(rq, p);
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01002443
2444 if (unlikely(rq->idle_stamp)) {
2445 u64 delta = rq->clock - rq->idle_stamp;
2446 u64 max = 2*sysctl_sched_migration_cost;
2447
2448 if (delta > max)
2449 rq->avg_idle = max;
2450 else
2451 update_avg(&rq->avg_idle, delta);
2452 rq->idle_stamp = 0;
2453 }
Steven Rostedt9a897c52008-01-25 21:08:22 +01002454#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455out:
2456 task_rq_unlock(rq, &flags);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002457 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458
2459 return success;
2460}
2461
David Howells50fa6102009-04-28 15:01:38 +01002462/**
2463 * wake_up_process - Wake up a specific process
2464 * @p: The process to be woken up.
2465 *
2466 * Attempt to wake up the nominated process and move it to the set of runnable
2467 * processes. Returns 1 if the process was woken up, 0 if it was already
2468 * running.
2469 *
2470 * It may be assumed that this function implies a write memory barrier before
2471 * changing the task state if and only if any tasks are woken up.
2472 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002473int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002475 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477EXPORT_SYMBOL(wake_up_process);
2478
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002479int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480{
2481 return try_to_wake_up(p, state, 0);
2482}
2483
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484/*
2485 * Perform scheduler related setup for a newly forked process p.
2486 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002487 *
2488 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002490static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491{
Ingo Molnardd41f592007-07-09 18:51:59 +02002492 p->se.exec_start = 0;
2493 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002494 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002495 p->se.nr_migrations = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002496 p->se.last_wakeup = 0;
2497 p->se.avg_overlap = 0;
Peter Zijlstra831451a2009-01-14 12:39:18 +01002498 p->se.start_runtime = 0;
2499 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002500
2501#ifdef CONFIG_SCHEDSTATS
Lucas De Marchi77935272009-07-09 13:57:20 +02002502 p->se.wait_start = 0;
2503 p->se.wait_max = 0;
2504 p->se.wait_count = 0;
2505 p->se.wait_sum = 0;
2506
2507 p->se.sleep_start = 0;
2508 p->se.sleep_max = 0;
2509 p->se.sum_sleep_runtime = 0;
2510
2511 p->se.block_start = 0;
2512 p->se.block_max = 0;
2513 p->se.exec_max = 0;
2514 p->se.slice_max = 0;
2515
2516 p->se.nr_migrations_cold = 0;
2517 p->se.nr_failed_migrations_affine = 0;
2518 p->se.nr_failed_migrations_running = 0;
2519 p->se.nr_failed_migrations_hot = 0;
2520 p->se.nr_forced_migrations = 0;
Lucas De Marchi77935272009-07-09 13:57:20 +02002521
2522 p->se.nr_wakeups = 0;
2523 p->se.nr_wakeups_sync = 0;
2524 p->se.nr_wakeups_migrate = 0;
2525 p->se.nr_wakeups_local = 0;
2526 p->se.nr_wakeups_remote = 0;
2527 p->se.nr_wakeups_affine = 0;
2528 p->se.nr_wakeups_affine_attempts = 0;
2529 p->se.nr_wakeups_passive = 0;
2530 p->se.nr_wakeups_idle = 0;
2531
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002532#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002533
Peter Zijlstrafa717062008-01-25 21:08:27 +01002534 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002535 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002536 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002537
Avi Kivitye107be32007-07-26 13:40:43 +02002538#ifdef CONFIG_PREEMPT_NOTIFIERS
2539 INIT_HLIST_HEAD(&p->preempt_notifiers);
2540#endif
2541
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 /*
2543 * We mark the process as running here, but have not actually
2544 * inserted it onto the runqueue yet. This guarantees that
2545 * nobody will actually run it, and a signal or other external
2546 * event cannot wake it up and insert it on the runqueue either.
2547 */
2548 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002549}
2550
2551/*
2552 * fork()/clone()-time setup:
2553 */
2554void sched_fork(struct task_struct *p, int clone_flags)
2555{
2556 int cpu = get_cpu();
2557
2558 __sched_fork(p);
2559
Ingo Molnarb29739f2006-06-27 02:54:51 -07002560 /*
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002561 * Revert to default priority/policy on fork if requested.
2562 */
2563 if (unlikely(p->sched_reset_on_fork)) {
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002564 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002565 p->policy = SCHED_NORMAL;
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002566 p->normal_prio = p->static_prio;
2567 }
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002568
Mike Galbraith6c697bd2009-06-17 10:48:02 +02002569 if (PRIO_TO_NICE(p->static_prio) < 0) {
2570 p->static_prio = NICE_TO_PRIO(0);
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002571 p->normal_prio = p->static_prio;
Mike Galbraith6c697bd2009-06-17 10:48:02 +02002572 set_load_weight(p);
2573 }
2574
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002575 /*
2576 * We don't need the reset flag anymore after the fork. It has
2577 * fulfilled its duty:
2578 */
2579 p->sched_reset_on_fork = 0;
2580 }
Lennart Poetteringca94c442009-06-15 17:17:47 +02002581
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002582 /*
2583 * Make sure we do not leak PI boosting priority to the child.
2584 */
2585 p->prio = current->normal_prio;
2586
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002587 if (!rt_prio(p->prio))
2588 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002589
Peter Zijlstracd29fe62009-11-27 17:32:46 +01002590 if (p->sched_class->task_fork)
2591 p->sched_class->task_fork(p);
2592
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002593#ifdef CONFIG_SMP
Peter Zijlstra970b13b2009-11-25 13:31:39 +01002594 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002595#endif
2596 set_task_cpu(p, cpu);
2597
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002598#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002599 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002600 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002602#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002603 p->oncpu = 0;
2604#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002606 /* Want to start with kernel preemption disabled. */
Al Viroa1261f542005-11-13 16:06:55 -08002607 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608#endif
Gregory Haskins917b6272008-12-29 09:39:53 -05002609 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2610
Nick Piggin476d1392005-06-25 14:57:29 -07002611 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612}
2613
2614/*
2615 * wake_up_new_task - wake up a newly created task for the first time.
2616 *
2617 * This function will do some initial scheduler statistics housekeeping
2618 * that must be done for every newly created context, then puts the task
2619 * on the runqueue and wakes it.
2620 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002621void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622{
2623 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002624 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625
2626 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002628 update_rq_clock(rq);
Peter Zijlstracd29fe62009-11-27 17:32:46 +01002629 activate_task(rq, p, 0);
Ingo Molnarc71dd422008-12-19 01:09:51 +01002630 trace_sched_wakeup_new(rq, p, 1);
Peter Zijlstraa7558e02009-09-14 20:02:34 +02002631 check_preempt_curr(rq, p, WF_FORK);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002632#ifdef CONFIG_SMP
2633 if (p->sched_class->task_wake_up)
2634 p->sched_class->task_wake_up(rq, p);
2635#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002636 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637}
2638
Avi Kivitye107be32007-07-26 13:40:43 +02002639#ifdef CONFIG_PREEMPT_NOTIFIERS
2640
2641/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00002642 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07002643 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002644 */
2645void preempt_notifier_register(struct preempt_notifier *notifier)
2646{
2647 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2648}
2649EXPORT_SYMBOL_GPL(preempt_notifier_register);
2650
2651/**
2652 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002653 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002654 *
2655 * This is safe to call from within a preemption notifier.
2656 */
2657void preempt_notifier_unregister(struct preempt_notifier *notifier)
2658{
2659 hlist_del(&notifier->link);
2660}
2661EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2662
2663static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2664{
2665 struct preempt_notifier *notifier;
2666 struct hlist_node *node;
2667
2668 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2670}
2671
2672static void
2673fire_sched_out_preempt_notifiers(struct task_struct *curr,
2674 struct task_struct *next)
2675{
2676 struct preempt_notifier *notifier;
2677 struct hlist_node *node;
2678
2679 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2680 notifier->ops->sched_out(notifier, next);
2681}
2682
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002683#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002684
2685static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2686{
2687}
2688
2689static void
2690fire_sched_out_preempt_notifiers(struct task_struct *curr,
2691 struct task_struct *next)
2692{
2693}
2694
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002695#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002696
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002698 * prepare_task_switch - prepare to switch tasks
2699 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002700 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002701 * @next: the task we are going to switch to.
2702 *
2703 * This is called with the rq lock held and interrupts off. It must
2704 * be paired with a subsequent finish_task_switch after the context
2705 * switch.
2706 *
2707 * prepare_task_switch sets up locking and calls architecture specific
2708 * hooks.
2709 */
Avi Kivitye107be32007-07-26 13:40:43 +02002710static inline void
2711prepare_task_switch(struct rq *rq, struct task_struct *prev,
2712 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002713{
Avi Kivitye107be32007-07-26 13:40:43 +02002714 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002715 prepare_lock_switch(rq, next);
2716 prepare_arch_switch(next);
2717}
2718
2719/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002721 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 * @prev: the thread we just switched away from.
2723 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002724 * finish_task_switch must be called after the context switch, paired
2725 * with a prepare_task_switch call before the context switch.
2726 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2727 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 *
2729 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002730 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 * with the lock held can cause deadlocks; see schedule() for
2732 * details.)
2733 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002734static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 __releases(rq->lock)
2736{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002738 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739
2740 rq->prev_mm = NULL;
2741
2742 /*
2743 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002744 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002745 * schedule one last time. The schedule call will never return, and
2746 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002747 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 * still held, otherwise prev could be scheduled on another cpu, die
2749 * there before we look at prev->state, and then the reference would
2750 * be dropped twice.
2751 * Manfred Spraul <manfred@colorfullife.com>
2752 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002753 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002754 finish_arch_switch(prev);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002755 perf_event_task_sched_in(current, cpu_of(rq));
Nick Piggin4866cde2005-06-25 14:57:23 -07002756 finish_lock_switch(rq, prev);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002757
Avi Kivitye107be32007-07-26 13:40:43 +02002758 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 if (mm)
2760 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002761 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002762 /*
2763 * Remove function-return probe instances associated with this
2764 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002765 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002766 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002768 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769}
2770
Gregory Haskins3f029d32009-07-29 11:08:47 -04002771#ifdef CONFIG_SMP
2772
2773/* assumes rq->lock is held */
2774static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2775{
2776 if (prev->sched_class->pre_schedule)
2777 prev->sched_class->pre_schedule(rq, prev);
2778}
2779
2780/* rq->lock is NOT held, but preemption is disabled */
2781static inline void post_schedule(struct rq *rq)
2782{
2783 if (rq->post_schedule) {
2784 unsigned long flags;
2785
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002786 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins3f029d32009-07-29 11:08:47 -04002787 if (rq->curr->sched_class->post_schedule)
2788 rq->curr->sched_class->post_schedule(rq);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002789 raw_spin_unlock_irqrestore(&rq->lock, flags);
Gregory Haskins3f029d32009-07-29 11:08:47 -04002790
2791 rq->post_schedule = 0;
2792 }
2793}
2794
2795#else
2796
2797static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2798{
2799}
2800
2801static inline void post_schedule(struct rq *rq)
2802{
2803}
2804
2805#endif
2806
Linus Torvalds1da177e2005-04-16 15:20:36 -07002807/**
2808 * schedule_tail - first thing a freshly forked thread must call.
2809 * @prev: the thread we just switched away from.
2810 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002811asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 __releases(rq->lock)
2813{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002814 struct rq *rq = this_rq();
2815
Nick Piggin4866cde2005-06-25 14:57:23 -07002816 finish_task_switch(rq, prev);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002817
Gregory Haskins3f029d32009-07-29 11:08:47 -04002818 /*
2819 * FIXME: do we need to worry about rq being invalidated by the
2820 * task_switch?
2821 */
2822 post_schedule(rq);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002823
Nick Piggin4866cde2005-06-25 14:57:23 -07002824#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2825 /* In this case, finish_task_switch does not reenable preemption */
2826 preempt_enable();
2827#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002829 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830}
2831
2832/*
2833 * context_switch - switch to the new MM and the new
2834 * thread's register state.
2835 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002836static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002837context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002838 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839{
Ingo Molnardd41f592007-07-09 18:51:59 +02002840 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841
Avi Kivitye107be32007-07-26 13:40:43 +02002842 prepare_task_switch(rq, prev, next);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002843 trace_sched_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002844 mm = next->mm;
2845 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002846 /*
2847 * For paravirt, this is coupled with an exit in switch_to to
2848 * combine the page table reload and the switch backend into
2849 * one hypercall.
2850 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08002851 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01002852
Tim Blechmann710390d2009-11-24 11:55:27 +01002853 if (likely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 next->active_mm = oldmm;
2855 atomic_inc(&oldmm->mm_count);
2856 enter_lazy_tlb(oldmm, next);
2857 } else
2858 switch_mm(oldmm, mm, next);
2859
Tim Blechmann710390d2009-11-24 11:55:27 +01002860 if (likely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 rq->prev_mm = oldmm;
2863 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002864 /*
2865 * Since the runqueue lock will be released by the next
2866 * task (which is an invalid locking op but in the case
2867 * of the scheduler it's an obvious special-case), so we
2868 * do an early lockdep release here:
2869 */
2870#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002871 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002872#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
2874 /* Here we just switch the register state and the stack. */
2875 switch_to(prev, next, prev);
2876
Ingo Molnardd41f592007-07-09 18:51:59 +02002877 barrier();
2878 /*
2879 * this_rq must be evaluated again because prev may have moved
2880 * CPUs since it called schedule(), thus the 'rq' on its stack
2881 * frame will be invalid.
2882 */
2883 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884}
2885
2886/*
2887 * nr_running, nr_uninterruptible and nr_context_switches:
2888 *
2889 * externally visible scheduler statistics: current number of runnable
2890 * threads, current number of uninterruptible-sleeping threads, total
2891 * number of context switches performed since bootup.
2892 */
2893unsigned long nr_running(void)
2894{
2895 unsigned long i, sum = 0;
2896
2897 for_each_online_cpu(i)
2898 sum += cpu_rq(i)->nr_running;
2899
2900 return sum;
2901}
2902
2903unsigned long nr_uninterruptible(void)
2904{
2905 unsigned long i, sum = 0;
2906
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002907 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 sum += cpu_rq(i)->nr_uninterruptible;
2909
2910 /*
2911 * Since we read the counters lockless, it might be slightly
2912 * inaccurate. Do not allow it to go below zero though:
2913 */
2914 if (unlikely((long)sum < 0))
2915 sum = 0;
2916
2917 return sum;
2918}
2919
2920unsigned long long nr_context_switches(void)
2921{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002922 int i;
2923 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002925 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 sum += cpu_rq(i)->nr_switches;
2927
2928 return sum;
2929}
2930
2931unsigned long nr_iowait(void)
2932{
2933 unsigned long i, sum = 0;
2934
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002935 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2937
2938 return sum;
2939}
2940
Arjan van de Ven69d25872009-09-21 17:04:08 -07002941unsigned long nr_iowait_cpu(void)
2942{
2943 struct rq *this = this_rq();
2944 return atomic_read(&this->nr_iowait);
2945}
2946
2947unsigned long this_cpu_load(void)
2948{
2949 struct rq *this = this_rq();
2950 return this->cpu_load[0];
2951}
2952
2953
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002954/* Variables and functions for calc_load */
2955static atomic_long_t calc_load_tasks;
2956static unsigned long calc_load_update;
2957unsigned long avenrun[3];
2958EXPORT_SYMBOL(avenrun);
2959
Thomas Gleixner2d024942009-05-02 20:08:52 +02002960/**
2961 * get_avenrun - get the load average array
2962 * @loads: pointer to dest load array
2963 * @offset: offset to add
2964 * @shift: shift count to shift the result left
2965 *
2966 * These values are estimates at best, so no need for locking.
2967 */
2968void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2969{
2970 loads[0] = (avenrun[0] + offset) << shift;
2971 loads[1] = (avenrun[1] + offset) << shift;
2972 loads[2] = (avenrun[2] + offset) << shift;
2973}
2974
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002975static unsigned long
2976calc_load(unsigned long load, unsigned long exp, unsigned long active)
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002977{
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002978 load *= exp;
2979 load += active * (FIXED_1 - exp);
2980 return load >> FSHIFT;
2981}
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002982
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002983/*
2984 * calc_load - update the avenrun load estimates 10 ticks after the
2985 * CPUs have updated calc_load_tasks.
2986 */
2987void calc_global_load(void)
2988{
2989 unsigned long upd = calc_load_update + 10;
2990 long active;
2991
2992 if (time_before(jiffies, upd))
2993 return;
2994
2995 active = atomic_long_read(&calc_load_tasks);
2996 active = active > 0 ? active * FIXED_1 : 0;
2997
2998 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2999 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3000 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3001
3002 calc_load_update += LOAD_FREQ;
3003}
3004
3005/*
3006 * Either called from update_cpu_load() or from a cpu going idle
3007 */
3008static void calc_load_account_active(struct rq *this_rq)
3009{
3010 long nr_active, delta;
3011
3012 nr_active = this_rq->nr_running;
3013 nr_active += (long) this_rq->nr_uninterruptible;
3014
3015 if (nr_active != this_rq->calc_load_active) {
3016 delta = nr_active - this_rq->calc_load_active;
3017 this_rq->calc_load_active = nr_active;
3018 atomic_long_add(delta, &calc_load_tasks);
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003019 }
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003020}
3021
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003023 * Update rq->cpu_load[] statistics. This function is usually called every
3024 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07003025 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003026static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003027{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02003028 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02003029 int i, scale;
3030
3031 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02003032
3033 /* Update our load: */
3034 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3035 unsigned long old_load, new_load;
3036
3037 /* scale is effectively 1 << i now, and >> i divides by scale */
3038
3039 old_load = this_rq->cpu_load[i];
3040 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02003041 /*
3042 * Round up the averaging division if load is increasing. This
3043 * prevents us from getting stuck on 9 if the load is 10, for
3044 * example.
3045 */
3046 if (new_load > old_load)
3047 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003048 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3049 }
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02003050
3051 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3052 this_rq->calc_load_update += LOAD_FREQ;
3053 calc_load_account_active(this_rq);
3054 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003055}
3056
Ingo Molnardd41f592007-07-09 18:51:59 +02003057#ifdef CONFIG_SMP
3058
Ingo Molnar48f24c42006-07-03 00:25:40 -07003059/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 * double_rq_lock - safely lock two runqueues
3061 *
3062 * Note this does not disable interrupts like task_rq_lock,
3063 * you need to do so manually before calling.
3064 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003065static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 __acquires(rq1->lock)
3067 __acquires(rq2->lock)
3068{
Kirill Korotaev054b9102006-12-10 02:20:11 -08003069 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 if (rq1 == rq2) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003071 raw_spin_lock(&rq1->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 __acquire(rq2->lock); /* Fake it out ;) */
3073 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003074 if (rq1 < rq2) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003075 raw_spin_lock(&rq1->lock);
3076 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 } else {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003078 raw_spin_lock(&rq2->lock);
3079 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 }
3081 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003082 update_rq_clock(rq1);
3083 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084}
3085
3086/*
3087 * double_rq_unlock - safely unlock two runqueues
3088 *
3089 * Note this does not restore interrupts like task_rq_unlock,
3090 * you need to do so manually after calling.
3091 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003092static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 __releases(rq1->lock)
3094 __releases(rq2->lock)
3095{
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003096 raw_spin_unlock(&rq1->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097 if (rq1 != rq2)
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003098 raw_spin_unlock(&rq2->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 else
3100 __release(rq2->lock);
3101}
3102
3103/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 * If dest_cpu is allowed for this process, migrate the task to it.
3105 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003106 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107 * the cpu_allowed mask is restored.
3108 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003109static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003111 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003113 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
3115 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e22008-11-25 02:35:14 +10303116 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
Max Krasnyanskye761b772008-07-15 04:43:49 -07003117 || unlikely(!cpu_active(dest_cpu)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 goto out;
3119
3120 /* force the process onto the specified CPU */
3121 if (migrate_task(p, dest_cpu, &req)) {
3122 /* Need to wait for migration thread (might exit: take ref). */
3123 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003124
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125 get_task_struct(mt);
3126 task_rq_unlock(rq, &flags);
3127 wake_up_process(mt);
3128 put_task_struct(mt);
3129 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07003130
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 return;
3132 }
3133out:
3134 task_rq_unlock(rq, &flags);
3135}
3136
3137/*
Nick Piggin476d1392005-06-25 14:57:29 -07003138 * sched_exec - execve() is a valuable balancing opportunity, because at
3139 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 */
3141void sched_exec(void)
3142{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 int new_cpu, this_cpu = get_cpu();
Peter Zijlstra970b13b2009-11-25 13:31:39 +01003144 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003146 if (new_cpu != this_cpu)
3147 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148}
3149
3150/*
3151 * pull_task - move a task from a remote runqueue to the local runqueue.
3152 * Both runqueues must be locked.
3153 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003154static void pull_task(struct rq *src_rq, struct task_struct *p,
3155 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003157 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003159 activate_task(this_rq, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02003160 check_preempt_curr(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161}
3162
3163/*
3164 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3165 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08003166static
Ingo Molnar70b97a72006-07-03 00:25:42 -07003167int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003168 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003169 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170{
Luis Henriques708dc512009-03-16 19:59:02 +00003171 int tsk_cache_hot = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 /*
3173 * We do not migrate tasks that are:
3174 * 1) running (obviously), or
3175 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3176 * 3) are cache-hot on their current CPU.
3177 */
Rusty Russell96f874e22008-11-25 02:35:14 +10303178 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003179 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003181 }
Nick Piggin81026792005-06-25 14:57:07 -07003182 *all_pinned = 0;
3183
Ingo Molnarcc367732007-10-15 17:00:18 +02003184 if (task_running(rq, p)) {
3185 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07003186 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003187 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188
Ingo Molnarda84d962007-10-15 17:00:18 +02003189 /*
3190 * Aggressive migration if:
3191 * 1) task is cache cold, or
3192 * 2) too many balance attempts have failed.
3193 */
3194
Luis Henriques708dc512009-03-16 19:59:02 +00003195 tsk_cache_hot = task_hot(p, rq->clock, sd);
3196 if (!tsk_cache_hot ||
3197 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003198#ifdef CONFIG_SCHEDSTATS
Luis Henriques708dc512009-03-16 19:59:02 +00003199 if (tsk_cache_hot) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003200 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02003201 schedstat_inc(p, se.nr_forced_migrations);
3202 }
Ingo Molnarda84d962007-10-15 17:00:18 +02003203#endif
3204 return 1;
3205 }
3206
Luis Henriques708dc512009-03-16 19:59:02 +00003207 if (tsk_cache_hot) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003208 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02003209 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211 return 1;
3212}
3213
Peter Williamse1d14842007-10-24 18:23:51 +02003214static unsigned long
3215balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3216 unsigned long max_load_move, struct sched_domain *sd,
3217 enum cpu_idle_type idle, int *all_pinned,
3218 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003219{
Peter Zijlstra051c6762008-06-27 13:41:31 +02003220 int loops = 0, pulled = 0, pinned = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02003221 struct task_struct *p;
3222 long rem_load_move = max_load_move;
3223
Peter Williamse1d14842007-10-24 18:23:51 +02003224 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02003225 goto out;
3226
3227 pinned = 1;
3228
3229 /*
3230 * Start the load-balancing iterator:
3231 */
3232 p = iterator->start(iterator->arg);
3233next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003234 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02003235 goto out;
Peter Zijlstra051c6762008-06-27 13:41:31 +02003236
3237 if ((p->se.load.weight >> 1) > rem_load_move ||
Ingo Molnardd41f592007-07-09 18:51:59 +02003238 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003239 p = iterator->next(iterator->arg);
3240 goto next;
3241 }
3242
3243 pull_task(busiest, p, this_rq, this_cpu);
3244 pulled++;
3245 rem_load_move -= p->se.load.weight;
3246
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003247#ifdef CONFIG_PREEMPT
3248 /*
3249 * NEWIDLE balancing is a source of latency, so preemptible kernels
3250 * will stop after the first task is pulled to minimize the critical
3251 * section.
3252 */
3253 if (idle == CPU_NEWLY_IDLE)
3254 goto out;
3255#endif
3256
Ingo Molnardd41f592007-07-09 18:51:59 +02003257 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003258 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003259 */
Peter Williamse1d14842007-10-24 18:23:51 +02003260 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003261 if (p->prio < *this_best_prio)
3262 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003263 p = iterator->next(iterator->arg);
3264 goto next;
3265 }
3266out:
3267 /*
Peter Williamse1d14842007-10-24 18:23:51 +02003268 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02003269 * so we can safely collect pull_task() stats here rather than
3270 * inside pull_task().
3271 */
3272 schedstat_add(sd, lb_gained[idle], pulled);
3273
3274 if (all_pinned)
3275 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003276
3277 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003278}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003279
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280/*
Peter Williams43010652007-08-09 11:16:46 +02003281 * move_tasks tries to move up to max_load_move weighted load from busiest to
3282 * this_rq, as part of a balancing operation within domain "sd".
3283 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 *
3285 * Called with both runqueues locked.
3286 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003287static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003288 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003289 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003290 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003292 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003293 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003294 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295
Ingo Molnardd41f592007-07-09 18:51:59 +02003296 do {
Peter Williams43010652007-08-09 11:16:46 +02003297 total_load_moved +=
3298 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003299 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003300 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003301 class = class->next;
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003302
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003303#ifdef CONFIG_PREEMPT
3304 /*
3305 * NEWIDLE balancing is a source of latency, so preemptible
3306 * kernels will stop after the first task is pulled to minimize
3307 * the critical section.
3308 */
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003309 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3310 break;
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003311#endif
Peter Williams43010652007-08-09 11:16:46 +02003312 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313
Peter Williams43010652007-08-09 11:16:46 +02003314 return total_load_moved > 0;
3315}
3316
Peter Williamse1d14842007-10-24 18:23:51 +02003317static int
3318iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3319 struct sched_domain *sd, enum cpu_idle_type idle,
3320 struct rq_iterator *iterator)
3321{
3322 struct task_struct *p = iterator->start(iterator->arg);
3323 int pinned = 0;
3324
3325 while (p) {
3326 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3327 pull_task(busiest, p, this_rq, this_cpu);
3328 /*
3329 * Right now, this is only the second place pull_task()
3330 * is called, so we can safely collect pull_task()
3331 * stats here rather than inside pull_task().
3332 */
3333 schedstat_inc(sd, lb_gained[idle]);
3334
3335 return 1;
3336 }
3337 p = iterator->next(iterator->arg);
3338 }
3339
3340 return 0;
3341}
3342
Peter Williams43010652007-08-09 11:16:46 +02003343/*
3344 * move_one_task tries to move exactly one task from busiest to this_rq, as
3345 * part of active balancing operations within "domain".
3346 * Returns 1 if successful and 0 otherwise.
3347 *
3348 * Called with both runqueues locked.
3349 */
3350static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3351 struct sched_domain *sd, enum cpu_idle_type idle)
3352{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003353 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003354
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003355 for_each_class(class) {
Peter Williamse1d14842007-10-24 18:23:51 +02003356 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003357 return 1;
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003358 }
Peter Williams43010652007-08-09 11:16:46 +02003359
3360 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361}
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303362/********** Helpers for find_busiest_group ************************/
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003363/*
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303364 * sd_lb_stats - Structure to store the statistics of a sched_domain
3365 * during load balancing.
3366 */
3367struct sd_lb_stats {
3368 struct sched_group *busiest; /* Busiest group in this sd */
3369 struct sched_group *this; /* Local group in this sd */
3370 unsigned long total_load; /* Total load of all groups in sd */
3371 unsigned long total_pwr; /* Total power of all groups in sd */
3372 unsigned long avg_load; /* Average load across all groups in sd */
3373
3374 /** Statistics of this group */
3375 unsigned long this_load;
3376 unsigned long this_load_per_task;
3377 unsigned long this_nr_running;
3378
3379 /* Statistics of the busiest group */
3380 unsigned long max_load;
3381 unsigned long busiest_load_per_task;
3382 unsigned long busiest_nr_running;
3383
3384 int group_imb; /* Is there imbalance in this sd */
3385#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3386 int power_savings_balance; /* Is powersave balance needed for this sd */
3387 struct sched_group *group_min; /* Least loaded group in sd */
3388 struct sched_group *group_leader; /* Group which relieves group_min */
3389 unsigned long min_load_per_task; /* load_per_task in group_min */
3390 unsigned long leader_nr_running; /* Nr running of group_leader */
3391 unsigned long min_nr_running; /* Nr running of group_min */
3392#endif
3393};
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394
3395/*
Gautham R Shenoy381be782009-03-25 14:43:46 +05303396 * sg_lb_stats - stats of a sched_group required for load_balancing
3397 */
3398struct sg_lb_stats {
3399 unsigned long avg_load; /*Avg load across the CPUs of the group */
3400 unsigned long group_load; /* Total load over the CPUs of the group */
3401 unsigned long sum_nr_running; /* Nr tasks running in the group */
3402 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3403 unsigned long group_capacity;
3404 int group_imb; /* Is there an imbalance in the group ? */
3405};
3406
3407/**
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303408 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3409 * @group: The group whose first cpu is to be returned.
3410 */
3411static inline unsigned int group_first_cpu(struct sched_group *group)
3412{
3413 return cpumask_first(sched_group_cpus(group));
3414}
3415
3416/**
3417 * get_sd_load_idx - Obtain the load index for a given sched domain.
3418 * @sd: The sched_domain whose load_idx is to be obtained.
3419 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3420 */
3421static inline int get_sd_load_idx(struct sched_domain *sd,
3422 enum cpu_idle_type idle)
3423{
3424 int load_idx;
3425
3426 switch (idle) {
3427 case CPU_NOT_IDLE:
3428 load_idx = sd->busy_idx;
3429 break;
3430
3431 case CPU_NEWLY_IDLE:
3432 load_idx = sd->newidle_idx;
3433 break;
3434 default:
3435 load_idx = sd->idle_idx;
3436 break;
3437 }
3438
3439 return load_idx;
3440}
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303441
3442
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303443#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3444/**
3445 * init_sd_power_savings_stats - Initialize power savings statistics for
3446 * the given sched_domain, during load balancing.
3447 *
3448 * @sd: Sched domain whose power-savings statistics are to be initialized.
3449 * @sds: Variable containing the statistics for sd.
3450 * @idle: Idle status of the CPU at which we're performing load-balancing.
3451 */
3452static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3453 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3454{
3455 /*
3456 * Busy processors will not participate in power savings
3457 * balance.
3458 */
3459 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3460 sds->power_savings_balance = 0;
3461 else {
3462 sds->power_savings_balance = 1;
3463 sds->min_nr_running = ULONG_MAX;
3464 sds->leader_nr_running = 0;
3465 }
3466}
3467
3468/**
3469 * update_sd_power_savings_stats - Update the power saving stats for a
3470 * sched_domain while performing load balancing.
3471 *
3472 * @group: sched_group belonging to the sched_domain under consideration.
3473 * @sds: Variable containing the statistics of the sched_domain
3474 * @local_group: Does group contain the CPU for which we're performing
3475 * load balancing ?
3476 * @sgs: Variable containing the statistics of the group.
3477 */
3478static inline void update_sd_power_savings_stats(struct sched_group *group,
3479 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3480{
3481
3482 if (!sds->power_savings_balance)
3483 return;
3484
3485 /*
3486 * If the local group is idle or completely loaded
3487 * no need to do power savings balance at this domain
3488 */
3489 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3490 !sds->this_nr_running))
3491 sds->power_savings_balance = 0;
3492
3493 /*
3494 * If a group is already running at full capacity or idle,
3495 * don't include that group in power savings calculations
3496 */
3497 if (!sds->power_savings_balance ||
3498 sgs->sum_nr_running >= sgs->group_capacity ||
3499 !sgs->sum_nr_running)
3500 return;
3501
3502 /*
3503 * Calculate the group which has the least non-idle load.
3504 * This is the group from where we need to pick up the load
3505 * for saving power
3506 */
3507 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3508 (sgs->sum_nr_running == sds->min_nr_running &&
3509 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3510 sds->group_min = group;
3511 sds->min_nr_running = sgs->sum_nr_running;
3512 sds->min_load_per_task = sgs->sum_weighted_load /
3513 sgs->sum_nr_running;
3514 }
3515
3516 /*
3517 * Calculate the group which is almost near its
3518 * capacity but still has some space to pick up some load
3519 * from other group and save more power
3520 */
Gautham R Shenoyd899a782009-09-02 16:59:10 +05303521 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303522 return;
3523
3524 if (sgs->sum_nr_running > sds->leader_nr_running ||
3525 (sgs->sum_nr_running == sds->leader_nr_running &&
3526 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3527 sds->group_leader = group;
3528 sds->leader_nr_running = sgs->sum_nr_running;
3529 }
3530}
3531
3532/**
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003533 * check_power_save_busiest_group - see if there is potential for some power-savings balance
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303534 * @sds: Variable containing the statistics of the sched_domain
3535 * under consideration.
3536 * @this_cpu: Cpu at which we're currently performing load-balancing.
3537 * @imbalance: Variable to store the imbalance.
3538 *
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003539 * Description:
3540 * Check if we have potential to perform some power-savings balance.
3541 * If yes, set the busiest group to be the least loaded group in the
3542 * sched_domain, so that it's CPUs can be put to idle.
3543 *
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303544 * Returns 1 if there is potential to perform power-savings balance.
3545 * Else returns 0.
3546 */
3547static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3548 int this_cpu, unsigned long *imbalance)
3549{
3550 if (!sds->power_savings_balance)
3551 return 0;
3552
3553 if (sds->this != sds->group_leader ||
3554 sds->group_leader == sds->group_min)
3555 return 0;
3556
3557 *imbalance = sds->min_load_per_task;
3558 sds->busiest = sds->group_min;
3559
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303560 return 1;
3561
3562}
3563#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3564static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3565 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3566{
3567 return;
3568}
3569
3570static inline void update_sd_power_savings_stats(struct sched_group *group,
3571 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3572{
3573 return;
3574}
3575
3576static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3577 int this_cpu, unsigned long *imbalance)
3578{
3579 return 0;
3580}
3581#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3582
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003583
3584unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3585{
3586 return SCHED_LOAD_SCALE;
3587}
3588
3589unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3590{
3591 return default_scale_freq_power(sd, cpu);
3592}
3593
3594unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003595{
3596 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3597 unsigned long smt_gain = sd->smt_gain;
3598
3599 smt_gain /= weight;
3600
3601 return smt_gain;
3602}
3603
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003604unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3605{
3606 return default_scale_smt_power(sd, cpu);
3607}
3608
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003609unsigned long scale_rt_power(int cpu)
3610{
3611 struct rq *rq = cpu_rq(cpu);
3612 u64 total, available;
3613
3614 sched_avg_update(rq);
3615
3616 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3617 available = total - rq->rt_avg;
3618
3619 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3620 total = SCHED_LOAD_SCALE;
3621
3622 total >>= SCHED_LOAD_SHIFT;
3623
3624 return div_u64(available, total);
3625}
3626
Peter Zijlstraab292302009-09-01 10:34:36 +02003627static void update_cpu_power(struct sched_domain *sd, int cpu)
3628{
3629 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3630 unsigned long power = SCHED_LOAD_SCALE;
3631 struct sched_group *sdg = sd->groups;
Peter Zijlstraab292302009-09-01 10:34:36 +02003632
Peter Zijlstra8e6598af2009-09-03 13:20:03 +02003633 if (sched_feat(ARCH_POWER))
3634 power *= arch_scale_freq_power(sd, cpu);
3635 else
3636 power *= default_scale_freq_power(sd, cpu);
3637
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003638 power >>= SCHED_LOAD_SHIFT;
Peter Zijlstraab292302009-09-01 10:34:36 +02003639
3640 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
Peter Zijlstra8e6598af2009-09-03 13:20:03 +02003641 if (sched_feat(ARCH_POWER))
3642 power *= arch_scale_smt_power(sd, cpu);
3643 else
3644 power *= default_scale_smt_power(sd, cpu);
3645
Peter Zijlstraab292302009-09-01 10:34:36 +02003646 power >>= SCHED_LOAD_SHIFT;
3647 }
3648
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003649 power *= scale_rt_power(cpu);
3650 power >>= SCHED_LOAD_SHIFT;
3651
3652 if (!power)
3653 power = 1;
Peter Zijlstraab292302009-09-01 10:34:36 +02003654
Peter Zijlstra18a38852009-09-01 10:34:39 +02003655 sdg->cpu_power = power;
Peter Zijlstraab292302009-09-01 10:34:36 +02003656}
3657
3658static void update_group_power(struct sched_domain *sd, int cpu)
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003659{
3660 struct sched_domain *child = sd->child;
3661 struct sched_group *group, *sdg = sd->groups;
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003662 unsigned long power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003663
3664 if (!child) {
Peter Zijlstraab292302009-09-01 10:34:36 +02003665 update_cpu_power(sd, cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003666 return;
3667 }
3668
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003669 power = 0;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003670
3671 group = child->groups;
3672 do {
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003673 power += group->cpu_power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003674 group = group->next;
3675 } while (group != child->groups);
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003676
3677 sdg->cpu_power = power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003678}
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303679
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303680/**
3681 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
Randy Dunlape17b38b2009-10-11 19:12:00 -07003682 * @sd: The sched_domain whose statistics are to be updated.
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303683 * @group: sched_group whose statistics are to be updated.
3684 * @this_cpu: Cpu for which load balance is currently performed.
3685 * @idle: Idle status of this_cpu
3686 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3687 * @sd_idle: Idle status of the sched_domain containing group.
3688 * @local_group: Does group contain this_cpu.
3689 * @cpus: Set of cpus considered for load balancing.
3690 * @balance: Should we balance.
3691 * @sgs: variable to hold the statistics for this group.
3692 */
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003693static inline void update_sg_lb_stats(struct sched_domain *sd,
3694 struct sched_group *group, int this_cpu,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303695 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3696 int local_group, const struct cpumask *cpus,
3697 int *balance, struct sg_lb_stats *sgs)
3698{
3699 unsigned long load, max_cpu_load, min_cpu_load;
3700 int i;
3701 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3702 unsigned long sum_avg_load_per_task;
3703 unsigned long avg_load_per_task;
3704
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003705 if (local_group) {
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303706 balance_cpu = group_first_cpu(group);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003707 if (balance_cpu == this_cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003708 update_group_power(sd, this_cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003709 }
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303710
3711 /* Tally up the load of all CPUs in the group */
3712 sum_avg_load_per_task = avg_load_per_task = 0;
3713 max_cpu_load = 0;
3714 min_cpu_load = ~0UL;
3715
3716 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3717 struct rq *rq = cpu_rq(i);
3718
3719 if (*sd_idle && rq->nr_running)
3720 *sd_idle = 0;
3721
3722 /* Bias balancing toward cpus of our domain */
3723 if (local_group) {
3724 if (idle_cpu(i) && !first_idle_cpu) {
3725 first_idle_cpu = 1;
3726 balance_cpu = i;
3727 }
3728
3729 load = target_load(i, load_idx);
3730 } else {
3731 load = source_load(i, load_idx);
3732 if (load > max_cpu_load)
3733 max_cpu_load = load;
3734 if (min_cpu_load > load)
3735 min_cpu_load = load;
3736 }
3737
3738 sgs->group_load += load;
3739 sgs->sum_nr_running += rq->nr_running;
3740 sgs->sum_weighted_load += weighted_cpuload(i);
3741
3742 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3743 }
3744
3745 /*
3746 * First idle cpu or the first cpu(busiest) in this sched group
3747 * is eligible for doing load balancing at this and above
3748 * domains. In the newly idle case, we will allow all the cpu's
3749 * to do the newly idle load balance.
3750 */
3751 if (idle != CPU_NEWLY_IDLE && local_group &&
3752 balance_cpu != this_cpu && balance) {
3753 *balance = 0;
3754 return;
3755 }
3756
3757 /* Adjust by relative CPU power of the group */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003758 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303759
3760
3761 /*
3762 * Consider the group unbalanced when the imbalance is larger
3763 * than the average weight of two tasks.
3764 *
3765 * APZ: with cgroup the avg task weight can vary wildly and
3766 * might not be a suitable number - should we keep a
3767 * normalized nr_running number somewhere that negates
3768 * the hierarchy?
3769 */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003770 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3771 group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303772
3773 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3774 sgs->group_imb = 1;
3775
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003776 sgs->group_capacity =
Peter Zijlstra18a38852009-09-01 10:34:39 +02003777 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303778}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303780/**
3781 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3782 * @sd: sched_domain whose statistics are to be updated.
3783 * @this_cpu: Cpu for which load balance is currently performed.
3784 * @idle: Idle status of this_cpu
3785 * @sd_idle: Idle status of the sched_domain containing group.
3786 * @cpus: Set of cpus considered for load balancing.
3787 * @balance: Should we balance.
3788 * @sds: variable to hold the statistics for this sched_domain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303790static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3791 enum cpu_idle_type idle, int *sd_idle,
3792 const struct cpumask *cpus, int *balance,
3793 struct sd_lb_stats *sds)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794{
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003795 struct sched_domain *child = sd->child;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303796 struct sched_group *group = sd->groups;
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303797 struct sg_lb_stats sgs;
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003798 int load_idx, prefer_sibling = 0;
3799
3800 if (child && child->flags & SD_PREFER_SIBLING)
3801 prefer_sibling = 1;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303802
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303803 init_sd_power_savings_stats(sd, sds, idle);
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303804 load_idx = get_sd_load_idx(sd, idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805
3806 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 int local_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808
Rusty Russell758b2cd2008-11-25 02:35:04 +10303809 local_group = cpumask_test_cpu(this_cpu,
3810 sched_group_cpus(group));
Gautham R Shenoy381be782009-03-25 14:43:46 +05303811 memset(&sgs, 0, sizeof(sgs));
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003812 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303813 local_group, cpus, balance, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303815 if (local_group && balance && !(*balance))
3816 return;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003817
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303818 sds->total_load += sgs.group_load;
Peter Zijlstra18a38852009-09-01 10:34:39 +02003819 sds->total_pwr += group->cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003821 /*
3822 * In case the child domain prefers tasks go to siblings
3823 * first, lower the group capacity to one so that we'll try
3824 * and move all the excess tasks away.
3825 */
3826 if (prefer_sibling)
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003827 sgs.group_capacity = min(sgs.group_capacity, 1UL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828
Linus Torvalds1da177e2005-04-16 15:20:36 -07003829 if (local_group) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303830 sds->this_load = sgs.avg_load;
3831 sds->this = group;
3832 sds->this_nr_running = sgs.sum_nr_running;
3833 sds->this_load_per_task = sgs.sum_weighted_load;
3834 } else if (sgs.avg_load > sds->max_load &&
Gautham R Shenoy381be782009-03-25 14:43:46 +05303835 (sgs.sum_nr_running > sgs.group_capacity ||
3836 sgs.group_imb)) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303837 sds->max_load = sgs.avg_load;
3838 sds->busiest = group;
3839 sds->busiest_nr_running = sgs.sum_nr_running;
3840 sds->busiest_load_per_task = sgs.sum_weighted_load;
3841 sds->group_imb = sgs.group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003843
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303844 update_sd_power_savings_stats(group, sds, local_group, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 group = group->next;
3846 } while (group != sd->groups);
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303847}
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303848
3849/**
3850 * fix_small_imbalance - Calculate the minor imbalance that exists
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303851 * amongst the groups of a sched_domain, during
3852 * load balancing.
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303853 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3854 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3855 * @imbalance: Variable to store the imbalance.
3856 */
3857static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3858 int this_cpu, unsigned long *imbalance)
3859{
3860 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3861 unsigned int imbn = 2;
3862
3863 if (sds->this_nr_running) {
3864 sds->this_load_per_task /= sds->this_nr_running;
3865 if (sds->busiest_load_per_task >
3866 sds->this_load_per_task)
3867 imbn = 1;
3868 } else
3869 sds->this_load_per_task =
3870 cpu_avg_load_per_task(this_cpu);
3871
3872 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3873 sds->busiest_load_per_task * imbn) {
3874 *imbalance = sds->busiest_load_per_task;
3875 return;
3876 }
3877
3878 /*
3879 * OK, we don't have enough imbalance to justify moving tasks,
3880 * however we may be able to increase total CPU power used by
3881 * moving them.
3882 */
3883
Peter Zijlstra18a38852009-09-01 10:34:39 +02003884 pwr_now += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303885 min(sds->busiest_load_per_task, sds->max_load);
Peter Zijlstra18a38852009-09-01 10:34:39 +02003886 pwr_now += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303887 min(sds->this_load_per_task, sds->this_load);
3888 pwr_now /= SCHED_LOAD_SCALE;
3889
3890 /* Amount of load we'd subtract */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003891 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3892 sds->busiest->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303893 if (sds->max_load > tmp)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003894 pwr_move += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303895 min(sds->busiest_load_per_task, sds->max_load - tmp);
3896
3897 /* Amount of load we'd add */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003898 if (sds->max_load * sds->busiest->cpu_power <
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303899 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003900 tmp = (sds->max_load * sds->busiest->cpu_power) /
3901 sds->this->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303902 else
Peter Zijlstra18a38852009-09-01 10:34:39 +02003903 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3904 sds->this->cpu_power;
3905 pwr_move += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303906 min(sds->this_load_per_task, sds->this_load + tmp);
3907 pwr_move /= SCHED_LOAD_SCALE;
3908
3909 /* Move if we gain throughput */
3910 if (pwr_move > pwr_now)
3911 *imbalance = sds->busiest_load_per_task;
3912}
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303913
3914/**
3915 * calculate_imbalance - Calculate the amount of imbalance present within the
3916 * groups of a given sched_domain during load balance.
3917 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3918 * @this_cpu: Cpu for which currently load balance is being performed.
3919 * @imbalance: The variable to store the imbalance.
3920 */
3921static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3922 unsigned long *imbalance)
3923{
3924 unsigned long max_pull;
3925 /*
3926 * In the presence of smp nice balancing, certain scenarios can have
3927 * max load less than avg load(as we skip the groups at or below
3928 * its cpu_power, while calculating max_load..)
3929 */
3930 if (sds->max_load < sds->avg_load) {
3931 *imbalance = 0;
3932 return fix_small_imbalance(sds, this_cpu, imbalance);
3933 }
3934
3935 /* Don't want to pull so many tasks that a group would go idle */
3936 max_pull = min(sds->max_load - sds->avg_load,
3937 sds->max_load - sds->busiest_load_per_task);
3938
3939 /* How much load to actually move to equalise the imbalance */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003940 *imbalance = min(max_pull * sds->busiest->cpu_power,
3941 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303942 / SCHED_LOAD_SCALE;
3943
3944 /*
3945 * if *imbalance is less than the average load per runnable task
3946 * there is no gaurantee that any tasks will be moved so we'll have
3947 * a think about bumping its value to force at least one task to be
3948 * moved
3949 */
3950 if (*imbalance < sds->busiest_load_per_task)
3951 return fix_small_imbalance(sds, this_cpu, imbalance);
3952
3953}
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303954/******* find_busiest_group() helpers end here *********************/
3955
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303956/**
3957 * find_busiest_group - Returns the busiest group within the sched_domain
3958 * if there is an imbalance. If there isn't an imbalance, and
3959 * the user has opted for power-savings, it returns a group whose
3960 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3961 * such a group exists.
3962 *
3963 * Also calculates the amount of weighted load which should be moved
3964 * to restore balance.
3965 *
3966 * @sd: The sched_domain whose busiest group is to be returned.
3967 * @this_cpu: The cpu for which load balancing is currently being performed.
3968 * @imbalance: Variable which stores amount of weighted load which should
3969 * be moved to restore balance/put a group to idle.
3970 * @idle: The idle status of this_cpu.
3971 * @sd_idle: The idleness of sd
3972 * @cpus: The set of CPUs under consideration for load-balancing.
3973 * @balance: Pointer to a variable indicating if this_cpu
3974 * is the appropriate cpu to perform load balancing at this_level.
3975 *
3976 * Returns: - the busiest group if imbalance exists.
3977 * - If no imbalance and user has opted for power-savings balance,
3978 * return the least loaded group whose CPUs can be
3979 * put to idle by rebalancing its tasks onto our group.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 */
3981static struct sched_group *
3982find_busiest_group(struct sched_domain *sd, int this_cpu,
3983 unsigned long *imbalance, enum cpu_idle_type idle,
3984 int *sd_idle, const struct cpumask *cpus, int *balance)
3985{
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303986 struct sd_lb_stats sds;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303988 memset(&sds, 0, sizeof(sds));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303990 /*
3991 * Compute the various statistics relavent for load balancing at
3992 * this level.
3993 */
3994 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3995 balance, &sds);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003996
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303997 /* Cases where imbalance does not exist from POV of this_cpu */
3998 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3999 * at this level.
4000 * 2) There is no busy sibling group to pull from.
4001 * 3) This group is the busiest group.
4002 * 4) This group is more busy than the avg busieness at this
4003 * sched_domain.
4004 * 5) The imbalance is within the specified limit.
4005 * 6) Any rebalance would lead to ping-pong
4006 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05304007 if (balance && !(*balance))
4008 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304010 if (!sds.busiest || sds.busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 goto out_balanced;
4012
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304013 if (sds.this_load >= sds.max_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014 goto out_balanced;
4015
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304016 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304018 if (sds.this_load >= sds.avg_load)
4019 goto out_balanced;
4020
4021 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 goto out_balanced;
4023
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304024 sds.busiest_load_per_task /= sds.busiest_nr_running;
4025 if (sds.group_imb)
4026 sds.busiest_load_per_task =
4027 min(sds.busiest_load_per_task, sds.avg_load);
Ken Chen908a7c12007-10-17 16:55:11 +02004028
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 /*
4030 * We're trying to get all the cpus to the average_load, so we don't
4031 * want to push ourselves above the average load, nor do we wish to
4032 * reduce the max loaded cpu below the average load, as either of these
4033 * actions would just result in more rebalancing later, and ping-pong
4034 * tasks around. Thus we look for the minimum possible imbalance.
4035 * Negative imbalances (*we* are more loaded than anyone else) will
4036 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004037 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 * appear as very large values with unsigned longs.
4039 */
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304040 if (sds.max_load <= sds.busiest_load_per_task)
Peter Williams2dd73a42006-06-27 02:54:34 -07004041 goto out_balanced;
4042
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05304043 /* Looks like there is an imbalance. Compute it */
4044 calculate_imbalance(&sds, this_cpu, imbalance);
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304045 return sds.busiest;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046
4047out_balanced:
Gautham R Shenoyc071df12009-03-25 14:44:22 +05304048 /*
4049 * There is no obvious imbalance. But check if we can do some balancing
4050 * to save power.
4051 */
4052 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4053 return sds.busiest;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004054ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 *imbalance = 0;
4056 return NULL;
4057}
4058
4059/*
4060 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4061 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004062static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004063find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Rusty Russell96f874e22008-11-25 02:35:14 +10304064 unsigned long imbalance, const struct cpumask *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004066 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07004067 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 int i;
4069
Rusty Russell758b2cd2008-11-25 02:35:04 +10304070 for_each_cpu(i, sched_group_cpus(group)) {
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004071 unsigned long power = power_of(i);
4072 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
Ingo Molnardd41f592007-07-09 18:51:59 +02004073 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004074
Rusty Russell96f874e22008-11-25 02:35:14 +10304075 if (!cpumask_test_cpu(i, cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004076 continue;
4077
Ingo Molnar48f24c42006-07-03 00:25:40 -07004078 rq = cpu_rq(i);
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004079 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4080 wl /= power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004082 if (capacity && rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07004083 continue;
4084
Ingo Molnardd41f592007-07-09 18:51:59 +02004085 if (wl > max_load) {
4086 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004087 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088 }
4089 }
4090
4091 return busiest;
4092}
4093
4094/*
Nick Piggin77391d72005-06-25 14:57:30 -07004095 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4096 * so long as it is large enough.
4097 */
4098#define MAX_PINNED_INTERVAL 512
4099
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304100/* Working cpumask for load_balance and load_balance_newidle. */
4101static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4102
Nick Piggin77391d72005-06-25 14:57:30 -07004103/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4105 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004107static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004108 struct sched_domain *sd, enum cpu_idle_type idle,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304109 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110{
Peter Williams43010652007-08-09 11:16:46 +02004111 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004114 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004115 unsigned long flags;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304116 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004117
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01004118 cpumask_copy(cpus, cpu_active_mask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004119
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004120 /*
4121 * When power savings policy is enabled for the parent domain, idle
4122 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02004123 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004124 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004125 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004126 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004127 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004128 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129
Ingo Molnar2d723762007-10-15 17:00:12 +02004130 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004132redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004133 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004134 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07004135 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004136
Chen, Kenneth W06066712006-12-10 02:20:35 -08004137 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004138 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004139
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 if (!group) {
4141 schedstat_inc(sd, lb_nobusyg[idle]);
4142 goto out_balanced;
4143 }
4144
Mike Travis7c16ec52008-04-04 18:11:11 -07004145 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 if (!busiest) {
4147 schedstat_inc(sd, lb_nobusyq[idle]);
4148 goto out_balanced;
4149 }
4150
Nick Piggindb935db2005-06-25 14:57:11 -07004151 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152
4153 schedstat_add(sd, lb_imbalance[idle], imbalance);
4154
Peter Williams43010652007-08-09 11:16:46 +02004155 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 if (busiest->nr_running > 1) {
4157 /*
4158 * Attempt to move tasks. If find_busiest_group has found
4159 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02004160 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 * correctly treated as an imbalance.
4162 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004163 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07004164 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02004165 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07004166 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07004167 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004168 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07004169
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004170 /*
4171 * some other cpu did the load balance for us.
4172 */
Peter Williams43010652007-08-09 11:16:46 +02004173 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004174 resched_cpu(this_cpu);
4175
Nick Piggin81026792005-06-25 14:57:07 -07004176 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004177 if (unlikely(all_pinned)) {
Rusty Russell96f874e22008-11-25 02:35:14 +10304178 cpumask_clear_cpu(cpu_of(busiest), cpus);
4179 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004180 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07004181 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183 }
Nick Piggin81026792005-06-25 14:57:07 -07004184
Peter Williams43010652007-08-09 11:16:46 +02004185 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186 schedstat_inc(sd, lb_failed[idle]);
4187 sd->nr_balance_failed++;
4188
4189 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004191 raw_spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004192
4193 /* don't kick the migration_thread, if the curr
4194 * task on busiest cpu can't be moved to this_cpu
4195 */
Rusty Russell96f874e22008-11-25 02:35:14 +10304196 if (!cpumask_test_cpu(this_cpu,
4197 &busiest->curr->cpus_allowed)) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004198 raw_spin_unlock_irqrestore(&busiest->lock,
4199 flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004200 all_pinned = 1;
4201 goto out_one_pinned;
4202 }
4203
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 if (!busiest->active_balance) {
4205 busiest->active_balance = 1;
4206 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07004207 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 }
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004209 raw_spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07004210 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211 wake_up_process(busiest->migration_thread);
4212
4213 /*
4214 * We've kicked active balancing, reset the failure
4215 * counter.
4216 */
Nick Piggin39507452005-06-25 14:57:09 -07004217 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 }
Nick Piggin81026792005-06-25 14:57:07 -07004219 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220 sd->nr_balance_failed = 0;
4221
Nick Piggin81026792005-06-25 14:57:07 -07004222 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 /* We were unbalanced, so reset the balancing interval */
4224 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07004225 } else {
4226 /*
4227 * If we've begun active balancing, start to back off. This
4228 * case may not be covered by the all_pinned logic if there
4229 * is only 1 task on the busy runqueue (because we don't call
4230 * move_tasks).
4231 */
4232 if (sd->balance_interval < sd->max_interval)
4233 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234 }
4235
Peter Williams43010652007-08-09 11:16:46 +02004236 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004237 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004238 ld_moved = -1;
4239
4240 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241
4242out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243 schedstat_inc(sd, lb_balanced[idle]);
4244
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004245 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004246
4247out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07004249 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4250 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 sd->balance_interval *= 2;
4252
Ingo Molnar48f24c42006-07-03 00:25:40 -07004253 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004254 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004255 ld_moved = -1;
4256 else
4257 ld_moved = 0;
4258out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004259 if (ld_moved)
4260 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004261 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262}
4263
4264/*
4265 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4266 * tasks if there is an imbalance.
4267 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004268 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 * this_rq is locked.
4270 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07004271static int
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304272load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273{
4274 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004275 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004276 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02004277 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07004278 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004279 int all_pinned = 0;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304280 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004281
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01004282 cpumask_copy(cpus, cpu_active_mask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004283
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004284 /*
4285 * When power savings policy is enabled for the parent domain, idle
4286 * sibling can pick up load irrespective of busy siblings. In this case,
4287 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004288 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004289 */
4290 if (sd->flags & SD_SHARE_CPUPOWER &&
4291 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004292 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004293
Ingo Molnar2d723762007-10-15 17:00:12 +02004294 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004295redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004296 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004297 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07004298 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004299 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004300 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004301 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302 }
4303
Mike Travis7c16ec52008-04-04 18:11:11 -07004304 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07004305 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004306 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004307 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 }
4309
Nick Piggindb935db2005-06-25 14:57:11 -07004310 BUG_ON(busiest == this_rq);
4311
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004312 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004313
Peter Williams43010652007-08-09 11:16:46 +02004314 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004315 if (busiest->nr_running > 1) {
4316 /* Attempt to move tasks */
4317 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004318 /* this_rq->clock is already updated */
4319 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02004320 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004321 imbalance, sd, CPU_NEWLY_IDLE,
4322 &all_pinned);
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004323 double_unlock_balance(this_rq, busiest);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004324
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004325 if (unlikely(all_pinned)) {
Rusty Russell96f874e22008-11-25 02:35:14 +10304326 cpumask_clear_cpu(cpu_of(busiest), cpus);
4327 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004328 goto redo;
4329 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004330 }
4331
Peter Williams43010652007-08-09 11:16:46 +02004332 if (!ld_moved) {
Vaidyanathan Srinivasan36dffab2008-12-20 10:06:38 +05304333 int active_balance = 0;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304334
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004335 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004336 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4337 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004338 return -1;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304339
4340 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4341 return -1;
4342
4343 if (sd->nr_balance_failed++ < 2)
4344 return -1;
4345
4346 /*
4347 * The only task running in a non-idle cpu can be moved to this
4348 * cpu in an attempt to completely freeup the other CPU
4349 * package. The same method used to move task in load_balance()
4350 * have been extended for load_balance_newidle() to speedup
4351 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4352 *
4353 * The package power saving logic comes from
4354 * find_busiest_group(). If there are no imbalance, then
4355 * f_b_g() will return NULL. However when sched_mc={1,2} then
4356 * f_b_g() will select a group from which a running task may be
4357 * pulled to this cpu in order to make the other package idle.
4358 * If there is no opportunity to make a package idle and if
4359 * there are no imbalance, then f_b_g() will return NULL and no
4360 * action will be taken in load_balance_newidle().
4361 *
4362 * Under normal task pull operation due to imbalance, there
4363 * will be more than one task in the source run queue and
4364 * move_tasks() will succeed. ld_moved will be true and this
4365 * active balance code will not be triggered.
4366 */
4367
4368 /* Lock busiest in correct order while this_rq is held */
4369 double_lock_balance(this_rq, busiest);
4370
4371 /*
4372 * don't kick the migration_thread, if the curr
4373 * task on busiest cpu can't be moved to this_cpu
4374 */
Mike Travis6ca09df2008-12-31 18:08:45 -08004375 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304376 double_unlock_balance(this_rq, busiest);
4377 all_pinned = 1;
4378 return ld_moved;
4379 }
4380
4381 if (!busiest->active_balance) {
4382 busiest->active_balance = 1;
4383 busiest->push_cpu = this_cpu;
4384 active_balance = 1;
4385 }
4386
4387 double_unlock_balance(this_rq, busiest);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004388 /*
4389 * Should not call ttwu while holding a rq->lock
4390 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004391 raw_spin_unlock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304392 if (active_balance)
4393 wake_up_process(busiest->migration_thread);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01004394 raw_spin_lock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304395
Nick Piggin5969fe02005-09-10 00:26:19 -07004396 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004397 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004398
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004399 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02004400 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004401
4402out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004403 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004404 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004405 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004406 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004407 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004408
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004409 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004410}
4411
4412/*
4413 * idle_balance is called by schedule() if this_cpu is about to become
4414 * idle. Attempts to pull tasks from other CPUs.
4415 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004416static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004417{
4418 struct sched_domain *sd;
Vaidyanathan Srinivasanefbe0272008-12-08 20:52:49 +05304419 int pulled_task = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02004420 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004421
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004422 this_rq->idle_stamp = this_rq->clock;
4423
4424 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4425 return;
4426
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004428 unsigned long interval;
4429
4430 if (!(sd->flags & SD_LOAD_BALANCE))
4431 continue;
4432
4433 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07004434 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07004435 pulled_task = load_balance_newidle(this_cpu, this_rq,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304436 sd);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004437
4438 interval = msecs_to_jiffies(sd->balance_interval);
4439 if (time_after(next_balance, sd->last_balance + interval))
4440 next_balance = sd->last_balance + interval;
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004441 if (pulled_task) {
4442 this_rq->idle_stamp = 0;
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004443 break;
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004444 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004446 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004447 /*
4448 * We are going idle. next_balance may be set based on
4449 * a busy processor. So reset next_balance.
4450 */
4451 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02004452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004453}
4454
4455/*
4456 * active_load_balance is run by migration threads. It pushes running tasks
4457 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4458 * running on each physical CPU where possible, and avoids physical /
4459 * logical imbalances.
4460 *
4461 * Called with busiest_rq locked.
4462 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004463static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004464{
Nick Piggin39507452005-06-25 14:57:09 -07004465 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004466 struct sched_domain *sd;
4467 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07004468
Ingo Molnar48f24c42006-07-03 00:25:40 -07004469 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07004470 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07004471 return;
4472
4473 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004474
4475 /*
Nick Piggin39507452005-06-25 14:57:09 -07004476 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004477 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07004478 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479 */
Nick Piggin39507452005-06-25 14:57:09 -07004480 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481
Nick Piggin39507452005-06-25 14:57:09 -07004482 /* move a task from busiest_rq to target_rq */
4483 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004484 update_rq_clock(busiest_rq);
4485 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004486
Nick Piggin39507452005-06-25 14:57:09 -07004487 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004488 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07004489 if ((sd->flags & SD_LOAD_BALANCE) &&
Rusty Russell758b2cd2008-11-25 02:35:04 +10304490 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
Nick Piggin39507452005-06-25 14:57:09 -07004491 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004492 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493
Ingo Molnar48f24c42006-07-03 00:25:40 -07004494 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004495 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004496
Peter Williams43010652007-08-09 11:16:46 +02004497 if (move_one_task(target_rq, target_cpu, busiest_rq,
4498 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07004499 schedstat_inc(sd, alb_pushed);
4500 else
4501 schedstat_inc(sd, alb_failed);
4502 }
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004503 double_unlock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004504}
4505
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004506#ifdef CONFIG_NO_HZ
4507static struct {
4508 atomic_t load_balancer;
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304509 cpumask_var_t cpu_mask;
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304510 cpumask_var_t ilb_grp_nohz_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004511} nohz ____cacheline_aligned = {
4512 .load_balancer = ATOMIC_INIT(-1),
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004513};
4514
Arun R Bharadwajeea08f32009-04-16 12:16:41 +05304515int get_nohz_load_balancer(void)
4516{
4517 return atomic_read(&nohz.load_balancer);
4518}
4519
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304520#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4521/**
4522 * lowest_flag_domain - Return lowest sched_domain containing flag.
4523 * @cpu: The cpu whose lowest level of sched domain is to
4524 * be returned.
4525 * @flag: The flag to check for the lowest sched_domain
4526 * for the given cpu.
4527 *
4528 * Returns the lowest sched_domain of a cpu which contains the given flag.
4529 */
4530static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4531{
4532 struct sched_domain *sd;
4533
4534 for_each_domain(cpu, sd)
4535 if (sd && (sd->flags & flag))
4536 break;
4537
4538 return sd;
4539}
4540
4541/**
4542 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4543 * @cpu: The cpu whose domains we're iterating over.
4544 * @sd: variable holding the value of the power_savings_sd
4545 * for cpu.
4546 * @flag: The flag to filter the sched_domains to be iterated.
4547 *
4548 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4549 * set, starting from the lowest sched_domain to the highest.
4550 */
4551#define for_each_flag_domain(cpu, sd, flag) \
4552 for (sd = lowest_flag_domain(cpu, flag); \
4553 (sd && (sd->flags & flag)); sd = sd->parent)
4554
4555/**
4556 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4557 * @ilb_group: group to be checked for semi-idleness
4558 *
4559 * Returns: 1 if the group is semi-idle. 0 otherwise.
4560 *
4561 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4562 * and atleast one non-idle CPU. This helper function checks if the given
4563 * sched_group is semi-idle or not.
4564 */
4565static inline int is_semi_idle_group(struct sched_group *ilb_group)
4566{
4567 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4568 sched_group_cpus(ilb_group));
4569
4570 /*
4571 * A sched_group is semi-idle when it has atleast one busy cpu
4572 * and atleast one idle cpu.
4573 */
4574 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4575 return 0;
4576
4577 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4578 return 0;
4579
4580 return 1;
4581}
4582/**
4583 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4584 * @cpu: The cpu which is nominating a new idle_load_balancer.
4585 *
4586 * Returns: Returns the id of the idle load balancer if it exists,
4587 * Else, returns >= nr_cpu_ids.
4588 *
4589 * This algorithm picks the idle load balancer such that it belongs to a
4590 * semi-idle powersavings sched_domain. The idea is to try and avoid
4591 * completely idle packages/cores just for the purpose of idle load balancing
4592 * when there are other idle cpu's which are better suited for that job.
4593 */
4594static int find_new_ilb(int cpu)
4595{
4596 struct sched_domain *sd;
4597 struct sched_group *ilb_group;
4598
4599 /*
4600 * Have idle load balancer selection from semi-idle packages only
4601 * when power-aware load balancing is enabled
4602 */
4603 if (!(sched_smt_power_savings || sched_mc_power_savings))
4604 goto out_done;
4605
4606 /*
4607 * Optimize for the case when we have no idle CPUs or only one
4608 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4609 */
4610 if (cpumask_weight(nohz.cpu_mask) < 2)
4611 goto out_done;
4612
4613 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4614 ilb_group = sd->groups;
4615
4616 do {
4617 if (is_semi_idle_group(ilb_group))
4618 return cpumask_first(nohz.ilb_grp_nohz_mask);
4619
4620 ilb_group = ilb_group->next;
4621
4622 } while (ilb_group != sd->groups);
4623 }
4624
4625out_done:
4626 return cpumask_first(nohz.cpu_mask);
4627}
4628#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4629static inline int find_new_ilb(int call_cpu)
4630{
Gautham R Shenoy6e29ec52009-04-21 08:40:49 +05304631 return cpumask_first(nohz.cpu_mask);
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304632}
4633#endif
4634
Christoph Lameter7835b982006-12-10 02:20:22 -08004635/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004636 * This routine will try to nominate the ilb (idle load balancing)
4637 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4638 * load balancing on behalf of all those cpus. If all the cpus in the system
4639 * go into this tickless mode, then there will be no ilb owner (as there is
4640 * no need for one) and all the cpus will sleep till the next wakeup event
4641 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08004642 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004643 * For the ilb owner, tick is not stopped. And this tick will be used
4644 * for idle load balancing. ilb owner will still be part of
4645 * nohz.cpu_mask..
4646 *
4647 * While stopping the tick, this cpu will become the ilb owner if there
4648 * is no other owner. And will be the owner till that cpu becomes busy
4649 * or if all cpus in the system stop their ticks at which point
4650 * there is no need for ilb owner.
4651 *
4652 * When the ilb owner becomes busy, it nominates another owner, during the
4653 * next busy scheduler_tick()
4654 */
4655int select_nohz_load_balancer(int stop_tick)
4656{
4657 int cpu = smp_processor_id();
4658
4659 if (stop_tick) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004660 cpu_rq(cpu)->in_nohz_recently = 1;
4661
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004662 if (!cpu_active(cpu)) {
4663 if (atomic_read(&nohz.load_balancer) != cpu)
4664 return 0;
4665
4666 /*
4667 * If we are going offline and still the leader,
4668 * give up!
4669 */
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004670 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4671 BUG();
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004672
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004673 return 0;
4674 }
4675
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004676 cpumask_set_cpu(cpu, nohz.cpu_mask);
4677
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004678 /* time for ilb owner also to sleep */
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01004679 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004680 if (atomic_read(&nohz.load_balancer) == cpu)
4681 atomic_set(&nohz.load_balancer, -1);
4682 return 0;
4683 }
4684
4685 if (atomic_read(&nohz.load_balancer) == -1) {
4686 /* make me the ilb owner */
4687 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4688 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304689 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4690 int new_ilb;
4691
4692 if (!(sched_smt_power_savings ||
4693 sched_mc_power_savings))
4694 return 1;
4695 /*
4696 * Check to see if there is a more power-efficient
4697 * ilb.
4698 */
4699 new_ilb = find_new_ilb(cpu);
4700 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4701 atomic_set(&nohz.load_balancer, -1);
4702 resched_cpu(new_ilb);
4703 return 0;
4704 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004705 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304706 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004707 } else {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304708 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004709 return 0;
4710
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304711 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004712
4713 if (atomic_read(&nohz.load_balancer) == cpu)
4714 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4715 BUG();
4716 }
4717 return 0;
4718}
4719#endif
4720
4721static DEFINE_SPINLOCK(balancing);
4722
4723/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004724 * It checks each scheduling domain to see if it is due to be balanced,
4725 * and initiates a balancing operation if so.
4726 *
4727 * Balancing parameters are set up in arch_init_sched_domains.
4728 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004729static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08004730{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004731 int balance = 1;
4732 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08004733 unsigned long interval;
4734 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004735 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08004736 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004737 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004738 int need_serialize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004739
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004740 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004741 if (!(sd->flags & SD_LOAD_BALANCE))
4742 continue;
4743
4744 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004745 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746 interval *= sd->busy_factor;
4747
4748 /* scale ms to jiffies */
4749 interval = msecs_to_jiffies(interval);
4750 if (unlikely(!interval))
4751 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02004752 if (interval > HZ*NR_CPUS/10)
4753 interval = HZ*NR_CPUS/10;
4754
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004755 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004756
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004757 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08004758 if (!spin_trylock(&balancing))
4759 goto out;
4760 }
4761
Christoph Lameterc9819f42006-12-10 02:20:25 -08004762 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304763 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004764 /*
4765 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07004766 * longer idle, or one of our SMT siblings is
4767 * not idle.
4768 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004769 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004771 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004772 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004773 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08004774 spin_unlock(&balancing);
4775out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02004776 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08004777 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004778 update_next_balance = 1;
4779 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004780
4781 /*
4782 * Stop the load balance at this level. There is another
4783 * CPU in our sched group which is doing load balancing more
4784 * actively.
4785 */
4786 if (!balance)
4787 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004788 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02004789
4790 /*
4791 * next_balance will be updated only when there is a need.
4792 * When the cpu is attached to null domain for ex, it will not be
4793 * updated.
4794 */
4795 if (likely(update_next_balance))
4796 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004797}
4798
4799/*
4800 * run_rebalance_domains is triggered when needed from the scheduler tick.
4801 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4802 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4803 */
4804static void run_rebalance_domains(struct softirq_action *h)
4805{
Ingo Molnardd41f592007-07-09 18:51:59 +02004806 int this_cpu = smp_processor_id();
4807 struct rq *this_rq = cpu_rq(this_cpu);
4808 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4809 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004810
Ingo Molnardd41f592007-07-09 18:51:59 +02004811 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004812
4813#ifdef CONFIG_NO_HZ
4814 /*
4815 * If this cpu is the owner for idle load balancing, then do the
4816 * balancing on behalf of the other idle cpus whose ticks are
4817 * stopped.
4818 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004819 if (this_rq->idle_at_tick &&
4820 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004821 struct rq *rq;
4822 int balance_cpu;
4823
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304824 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4825 if (balance_cpu == this_cpu)
4826 continue;
4827
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004828 /*
4829 * If this cpu gets work to do, stop the load balancing
4830 * work being done for other cpus. Next load
4831 * balancing owner will pick it up.
4832 */
4833 if (need_resched())
4834 break;
4835
Oleg Nesterovde0cf892007-08-12 18:08:19 +02004836 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004837
4838 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004839 if (time_after(this_rq->next_balance, rq->next_balance))
4840 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004841 }
4842 }
4843#endif
4844}
4845
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004846static inline int on_null_domain(int cpu)
4847{
4848 return !rcu_dereference(cpu_rq(cpu)->sd);
4849}
4850
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004851/*
4852 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4853 *
4854 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4855 * idle load balancing owner or decide to stop the periodic load balancing,
4856 * if the whole system is idle.
4857 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004858static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004859{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004860#ifdef CONFIG_NO_HZ
4861 /*
4862 * If we were in the nohz mode recently and busy at the current
4863 * scheduler tick, then check if we need to nominate new idle
4864 * load balancer.
4865 */
4866 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4867 rq->in_nohz_recently = 0;
4868
4869 if (atomic_read(&nohz.load_balancer) == cpu) {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304870 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004871 atomic_set(&nohz.load_balancer, -1);
4872 }
4873
4874 if (atomic_read(&nohz.load_balancer) == -1) {
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304875 int ilb = find_new_ilb(cpu);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004876
Mike Travis434d53b2008-04-04 18:11:04 -07004877 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004878 resched_cpu(ilb);
4879 }
4880 }
4881
4882 /*
4883 * If this cpu is idle and doing idle load balancing for all the
4884 * cpus with ticks stopped, is it time for that to stop?
4885 */
4886 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304887 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004888 resched_cpu(cpu);
4889 return;
4890 }
4891
4892 /*
4893 * If this cpu is idle and the idle load balancing is done by
4894 * someone else, then no need raise the SCHED_SOFTIRQ
4895 */
4896 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304897 cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004898 return;
4899#endif
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004900 /* Don't need to rebalance while attached to NULL domain */
4901 if (time_after_eq(jiffies, rq->next_balance) &&
4902 likely(!on_null_domain(cpu)))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004903 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004904}
Ingo Molnardd41f592007-07-09 18:51:59 +02004905
4906#else /* CONFIG_SMP */
4907
Linus Torvalds1da177e2005-04-16 15:20:36 -07004908/*
4909 * on UP we do not need to balance between CPUs:
4910 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004911static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004912{
4913}
Ingo Molnardd41f592007-07-09 18:51:59 +02004914
Linus Torvalds1da177e2005-04-16 15:20:36 -07004915#endif
4916
Linus Torvalds1da177e2005-04-16 15:20:36 -07004917DEFINE_PER_CPU(struct kernel_stat, kstat);
4918
4919EXPORT_PER_CPU_SYMBOL(kstat);
4920
4921/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004922 * Return any ns on the sched_clock that have not yet been accounted in
Frank Mayharf06febc2008-09-12 09:54:39 -07004923 * @p in case that task is currently running.
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004924 *
4925 * Called with task_rq_lock() held on @rq.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004926 */
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004927static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4928{
4929 u64 ns = 0;
4930
4931 if (task_current(rq, p)) {
4932 update_rq_clock(rq);
4933 ns = rq->clock - p->se.exec_start;
4934 if ((s64)ns < 0)
4935 ns = 0;
4936 }
4937
4938 return ns;
4939}
4940
Frank Mayharbb34d922008-09-12 09:54:39 -07004941unsigned long long task_delta_exec(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004942{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004943 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004944 struct rq *rq;
Frank Mayharbb34d922008-09-12 09:54:39 -07004945 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004946
Ingo Molnar41b86e92007-07-09 18:51:58 +02004947 rq = task_rq_lock(p, &flags);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004948 ns = do_task_delta_exec(p, rq);
4949 task_rq_unlock(rq, &flags);
Ingo Molnar15084872008-09-30 08:28:17 +02004950
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004951 return ns;
4952}
Frank Mayharf06febc2008-09-12 09:54:39 -07004953
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004954/*
4955 * Return accounted runtime for the task.
4956 * In case the task is currently running, return the runtime plus current's
4957 * pending runtime that have not been accounted yet.
4958 */
4959unsigned long long task_sched_runtime(struct task_struct *p)
4960{
4961 unsigned long flags;
4962 struct rq *rq;
4963 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004964
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004965 rq = task_rq_lock(p, &flags);
4966 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4967 task_rq_unlock(rq, &flags);
4968
4969 return ns;
4970}
4971
4972/*
4973 * Return sum_exec_runtime for the thread group.
4974 * In case the task is currently running, return the sum plus current's
4975 * pending runtime that have not been accounted yet.
4976 *
4977 * Note that the thread group might have other running tasks as well,
4978 * so the return value not includes other pending runtime that other
4979 * running tasks might have.
4980 */
4981unsigned long long thread_group_sched_runtime(struct task_struct *p)
4982{
4983 struct task_cputime totals;
4984 unsigned long flags;
4985 struct rq *rq;
4986 u64 ns;
4987
4988 rq = task_rq_lock(p, &flags);
4989 thread_group_cputime(p, &totals);
4990 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004991 task_rq_unlock(rq, &flags);
4992
4993 return ns;
4994}
4995
4996/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004997 * Account user cpu time to a process.
4998 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004999 * @cputime: the cpu time spent in user space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005000 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07005001 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005002void account_user_time(struct task_struct *p, cputime_t cputime,
5003 cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005004{
5005 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5006 cputime64_t tmp;
5007
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005008 /* Add user time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005009 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005010 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005011 account_group_user_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005012
5013 /* Add user time to cpustat. */
5014 tmp = cputime_to_cputime64(cputime);
5015 if (TASK_NICE(p) > 0)
5016 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5017 else
5018 cpustat->user = cputime64_add(cpustat->user, tmp);
Bharata B Raoef12fef2009-03-31 10:02:22 +05305019
5020 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
Jonathan Lim49b5cf32008-07-25 01:48:40 -07005021 /* Account for user time used */
5022 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005023}
5024
5025/*
Laurent Vivier94886b82007-10-15 17:00:19 +02005026 * Account guest cpu time to a process.
5027 * @p: the process that the cpu time gets accounted to
5028 * @cputime: the cpu time spent in virtual machine since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005029 * @cputime_scaled: cputime scaled by cpu frequency
Laurent Vivier94886b82007-10-15 17:00:19 +02005030 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005031static void account_guest_time(struct task_struct *p, cputime_t cputime,
5032 cputime_t cputime_scaled)
Laurent Vivier94886b82007-10-15 17:00:19 +02005033{
5034 cputime64_t tmp;
5035 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5036
5037 tmp = cputime_to_cputime64(cputime);
5038
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005039 /* Add guest time to process. */
Laurent Vivier94886b82007-10-15 17:00:19 +02005040 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005041 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005042 account_group_user_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02005043 p->gtime = cputime_add(p->gtime, cputime);
5044
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005045 /* Add guest time to cpustat. */
Ryota Ozakice0e7b22009-10-24 01:20:10 +09005046 if (TASK_NICE(p) > 0) {
5047 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5048 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5049 } else {
5050 cpustat->user = cputime64_add(cpustat->user, tmp);
5051 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5052 }
Laurent Vivier94886b82007-10-15 17:00:19 +02005053}
5054
5055/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005056 * Account system cpu time to a process.
5057 * @p: the process that the cpu time gets accounted to
5058 * @hardirq_offset: the offset to subtract from hardirq_count()
5059 * @cputime: the cpu time spent in kernel space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005060 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07005061 */
5062void account_system_time(struct task_struct *p, int hardirq_offset,
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005063 cputime_t cputime, cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005064{
5065 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066 cputime64_t tmp;
5067
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005068 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005069 account_guest_time(p, cputime, cputime_scaled);
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005070 return;
5071 }
Laurent Vivier94886b82007-10-15 17:00:19 +02005072
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005073 /* Add system time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005074 p->stime = cputime_add(p->stime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005075 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005076 account_group_system_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005077
5078 /* Add system time to cpustat. */
5079 tmp = cputime_to_cputime64(cputime);
5080 if (hardirq_count() - hardirq_offset)
5081 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5082 else if (softirq_count())
5083 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005084 else
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005085 cpustat->system = cputime64_add(cpustat->system, tmp);
5086
Bharata B Raoef12fef2009-03-31 10:02:22 +05305087 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5088
Linus Torvalds1da177e2005-04-16 15:20:36 -07005089 /* Account for system time used */
5090 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005091}
5092
5093/*
5094 * Account for involuntary wait time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095 * @steal: the cpu time spent in involuntary wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005096 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005097void account_steal_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005098{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005099 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005100 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5101
5102 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005103}
5104
Christoph Lameter7835b982006-12-10 02:20:22 -08005105/*
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005106 * Account for idle time.
5107 * @cputime: the cpu time spent in idle wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005108 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005109void account_idle_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005110{
5111 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005112 cputime64_t cputime64 = cputime_to_cputime64(cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005113 struct rq *rq = this_rq();
5114
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005115 if (atomic_read(&rq->nr_iowait) > 0)
5116 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5117 else
5118 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
Christoph Lameter7835b982006-12-10 02:20:22 -08005119}
5120
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005121#ifndef CONFIG_VIRT_CPU_ACCOUNTING
5122
5123/*
5124 * Account a single tick of cpu time.
5125 * @p: the process that the cpu time gets accounted to
5126 * @user_tick: indicates if the tick is a user or a system tick
5127 */
5128void account_process_tick(struct task_struct *p, int user_tick)
5129{
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005130 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005131 struct rq *rq = this_rq();
5132
5133 if (user_tick)
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005134 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
Eric Dumazetf5f293a2009-04-29 14:44:49 +02005135 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005136 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005137 one_jiffy_scaled);
5138 else
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005139 account_idle_time(cputime_one_jiffy);
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005140}
5141
5142/*
5143 * Account multiple ticks of steal time.
5144 * @p: the process from which the cpu time has been stolen
5145 * @ticks: number of stolen ticks
5146 */
5147void account_steal_ticks(unsigned long ticks)
5148{
5149 account_steal_time(jiffies_to_cputime(ticks));
5150}
5151
5152/*
5153 * Account multiple ticks of idle time.
5154 * @ticks: number of stolen ticks
5155 */
5156void account_idle_ticks(unsigned long ticks)
5157{
5158 account_idle_time(jiffies_to_cputime(ticks));
5159}
5160
5161#endif
5162
Christoph Lameter7835b982006-12-10 02:20:22 -08005163/*
Balbir Singh49048622008-09-05 18:12:23 +02005164 * Use precise platform statistics if available:
5165 */
5166#ifdef CONFIG_VIRT_CPU_ACCOUNTING
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005167void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005168{
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005169 *ut = p->utime;
5170 *st = p->stime;
Balbir Singh49048622008-09-05 18:12:23 +02005171}
5172
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005173void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005174{
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005175 struct task_cputime cputime;
5176
5177 thread_group_cputime(p, &cputime);
5178
5179 *ut = cputime.utime;
5180 *st = cputime.stime;
Balbir Singh49048622008-09-05 18:12:23 +02005181}
5182#else
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005183
5184#ifndef nsecs_to_cputime
Hidetoshi Setob7b20df92009-11-26 14:49:27 +09005185# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005186#endif
5187
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005188void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005189{
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005190 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
Balbir Singh49048622008-09-05 18:12:23 +02005191
5192 /*
5193 * Use CFS's precise accounting:
5194 */
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005195 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
Balbir Singh49048622008-09-05 18:12:23 +02005196
5197 if (total) {
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005198 u64 temp;
Balbir Singh49048622008-09-05 18:12:23 +02005199
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005200 temp = (u64)(rtime * utime);
Balbir Singh49048622008-09-05 18:12:23 +02005201 do_div(temp, total);
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005202 utime = (cputime_t)temp;
5203 } else
5204 utime = rtime;
Balbir Singh49048622008-09-05 18:12:23 +02005205
5206 /*
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005207 * Compare with previous values, to keep monotonicity:
Balbir Singh49048622008-09-05 18:12:23 +02005208 */
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005209 p->prev_utime = max(p->prev_utime, utime);
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005210 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
Balbir Singh49048622008-09-05 18:12:23 +02005211
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005212 *ut = p->prev_utime;
5213 *st = p->prev_stime;
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005214}
Balbir Singh49048622008-09-05 18:12:23 +02005215
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005216/*
5217 * Must be called with siglock held.
5218 */
5219void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5220{
5221 struct signal_struct *sig = p->signal;
5222 struct task_cputime cputime;
5223 cputime_t rtime, utime, total;
5224
5225 thread_group_cputime(p, &cputime);
5226
5227 total = cputime_add(cputime.utime, cputime.stime);
5228 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5229
5230 if (total) {
5231 u64 temp;
5232
5233 temp = (u64)(rtime * cputime.utime);
5234 do_div(temp, total);
5235 utime = (cputime_t)temp;
5236 } else
5237 utime = rtime;
5238
5239 sig->prev_utime = max(sig->prev_utime, utime);
5240 sig->prev_stime = max(sig->prev_stime,
5241 cputime_sub(rtime, sig->prev_utime));
5242
5243 *ut = sig->prev_utime;
5244 *st = sig->prev_stime;
Balbir Singh49048622008-09-05 18:12:23 +02005245}
5246#endif
5247
Balbir Singh49048622008-09-05 18:12:23 +02005248/*
Christoph Lameter7835b982006-12-10 02:20:22 -08005249 * This function gets called by the timer code, with HZ frequency.
5250 * We call it with interrupts disabled.
5251 *
5252 * It also gets called by the fork code, when changing the parent's
5253 * timeslices.
5254 */
5255void scheduler_tick(void)
5256{
Christoph Lameter7835b982006-12-10 02:20:22 -08005257 int cpu = smp_processor_id();
5258 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005259 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005260
5261 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08005262
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005263 raw_spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005264 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02005265 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01005266 curr->sched_class->task_tick(rq, curr, 0);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005267 raw_spin_unlock(&rq->lock);
Ingo Molnardd41f592007-07-09 18:51:59 +02005268
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005269 perf_event_task_tick(curr, cpu);
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02005270
Christoph Lametere418e1c2006-12-10 02:20:23 -08005271#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02005272 rq->idle_at_tick = idle_cpu(cpu);
5273 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08005274#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005275}
5276
Lai Jiangshan132380a2009-04-02 14:18:25 +08005277notrace unsigned long get_parent_ip(unsigned long addr)
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005278{
5279 if (in_lock_functions(addr)) {
5280 addr = CALLER_ADDR2;
5281 if (in_lock_functions(addr))
5282 addr = CALLER_ADDR3;
5283 }
5284 return addr;
5285}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005286
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05005287#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5288 defined(CONFIG_PREEMPT_TRACER))
5289
Srinivasa Ds43627582008-02-23 15:24:04 -08005290void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005291{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005292#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293 /*
5294 * Underflow?
5295 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005296 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5297 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005298#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005299 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005300#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005301 /*
5302 * Spinlock count overflowing soon?
5303 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005304 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5305 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005306#endif
5307 if (preempt_count() == val)
5308 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309}
5310EXPORT_SYMBOL(add_preempt_count);
5311
Srinivasa Ds43627582008-02-23 15:24:04 -08005312void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005314#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005315 /*
5316 * Underflow?
5317 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01005318 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005319 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005320 /*
5321 * Is the spinlock portion underflowing?
5322 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005323 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5324 !(preempt_count() & PREEMPT_MASK)))
5325 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005326#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005327
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005328 if (preempt_count() == val)
5329 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005330 preempt_count() -= val;
5331}
5332EXPORT_SYMBOL(sub_preempt_count);
5333
5334#endif
5335
5336/*
Ingo Molnardd41f592007-07-09 18:51:59 +02005337 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338 */
Ingo Molnardd41f592007-07-09 18:51:59 +02005339static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005340{
Satyam Sharma838225b2007-10-24 18:23:50 +02005341 struct pt_regs *regs = get_irq_regs();
5342
5343 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5344 prev->comm, prev->pid, preempt_count());
5345
Ingo Molnardd41f592007-07-09 18:51:59 +02005346 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07005347 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02005348 if (irqs_disabled())
5349 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02005350
5351 if (regs)
5352 show_regs(regs);
5353 else
5354 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02005355}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356
Ingo Molnardd41f592007-07-09 18:51:59 +02005357/*
5358 * Various schedule()-time debugging checks and statistics:
5359 */
5360static inline void schedule_debug(struct task_struct *prev)
5361{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005362 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005363 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07005364 * schedule() atomically, we ignore that path for now.
5365 * Otherwise, whine if we are scheduling when we should not be.
5366 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02005367 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02005368 __schedule_bug(prev);
5369
Linus Torvalds1da177e2005-04-16 15:20:36 -07005370 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5371
Ingo Molnar2d723762007-10-15 17:00:12 +02005372 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005373#ifdef CONFIG_SCHEDSTATS
5374 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02005375 schedstat_inc(this_rq(), bkl_count);
5376 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005377 }
5378#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005379}
5380
Peter Zijlstra6cecd082009-11-30 13:00:37 +01005381static void put_prev_task(struct rq *rq, struct task_struct *prev)
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005382{
Peter Zijlstra6cecd082009-11-30 13:00:37 +01005383 if (prev->state == TASK_RUNNING) {
5384 u64 runtime = prev->se.sum_exec_runtime;
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005385
Peter Zijlstra6cecd082009-11-30 13:00:37 +01005386 runtime -= prev->se.prev_sum_exec_runtime;
5387 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005388
5389 /*
5390 * In order to avoid avg_overlap growing stale when we are
5391 * indeed overlapping and hence not getting put to sleep, grow
5392 * the avg_overlap on preemption.
5393 *
5394 * We use the average preemption runtime because that
5395 * correlates to the amount of cache footprint a task can
5396 * build up.
5397 */
Peter Zijlstra6cecd082009-11-30 13:00:37 +01005398 update_avg(&prev->se.avg_overlap, runtime);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005399 }
Peter Zijlstra6cecd082009-11-30 13:00:37 +01005400 prev->sched_class->put_prev_task(rq, prev);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005401}
5402
Ingo Molnardd41f592007-07-09 18:51:59 +02005403/*
5404 * Pick up the highest-prio task:
5405 */
5406static inline struct task_struct *
Wang Chenb67802e2009-03-02 13:55:26 +08005407pick_next_task(struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02005408{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02005409 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02005410 struct task_struct *p;
5411
5412 /*
5413 * Optimization: we know that if all tasks are in
5414 * the fair class we can call that function directly:
5415 */
5416 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005417 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005418 if (likely(p))
5419 return p;
5420 }
5421
5422 class = sched_class_highest;
5423 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005424 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005425 if (p)
5426 return p;
5427 /*
5428 * Will never be NULL as the idle class always
5429 * returns a non-NULL p:
5430 */
5431 class = class->next;
5432 }
5433}
5434
5435/*
5436 * schedule() is the main scheduler function.
5437 */
Peter Zijlstraff743342009-03-13 12:21:26 +01005438asmlinkage void __sched schedule(void)
Ingo Molnardd41f592007-07-09 18:51:59 +02005439{
5440 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08005441 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02005442 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02005443 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02005444
Peter Zijlstraff743342009-03-13 12:21:26 +01005445need_resched:
5446 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02005447 cpu = smp_processor_id();
5448 rq = cpu_rq(cpu);
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07005449 rcu_sched_qs(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005450 prev = rq->curr;
5451 switch_count = &prev->nivcsw;
5452
Linus Torvalds1da177e2005-04-16 15:20:36 -07005453 release_kernel_lock(prev);
5454need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005455
Ingo Molnardd41f592007-07-09 18:51:59 +02005456 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005457
Peter Zijlstra31656512008-07-18 18:01:23 +02005458 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02005459 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005460
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005461 raw_spin_lock_irq(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005462 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02005463 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005464
Ingo Molnardd41f592007-07-09 18:51:59 +02005465 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04005466 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02005467 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04005468 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005469 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02005470 switch_count = &prev->nvcsw;
5471 }
5472
Gregory Haskins3f029d32009-07-29 11:08:47 -04005473 pre_schedule(rq, prev);
Steven Rostedtf65eda42008-01-25 21:08:07 +01005474
Ingo Molnardd41f592007-07-09 18:51:59 +02005475 if (unlikely(!rq->nr_running))
5476 idle_balance(cpu, rq);
5477
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005478 put_prev_task(rq, prev);
Wang Chenb67802e2009-03-02 13:55:26 +08005479 next = pick_next_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01005482 sched_info_switch(prev, next);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005483 perf_event_task_sched_out(prev, next, cpu);
David Simner673a90a2008-04-29 10:08:59 +01005484
Linus Torvalds1da177e2005-04-16 15:20:36 -07005485 rq->nr_switches++;
5486 rq->curr = next;
5487 ++*switch_count;
5488
Ingo Molnardd41f592007-07-09 18:51:59 +02005489 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005490 /*
5491 * the context switch might have flipped the stack from under
5492 * us, hence refresh the local variables.
5493 */
5494 cpu = smp_processor_id();
5495 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005496 } else
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005497 raw_spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005498
Gregory Haskins3f029d32009-07-29 11:08:47 -04005499 post_schedule(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005501 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005502 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005503
Linus Torvalds1da177e2005-04-16 15:20:36 -07005504 preempt_enable_no_resched();
Peter Zijlstraff743342009-03-13 12:21:26 +01005505 if (need_resched())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005506 goto need_resched;
5507}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005508EXPORT_SYMBOL(schedule);
5509
Frederic Weisbeckerc08f7822009-12-02 20:49:17 +01005510#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
Peter Zijlstra0d66bf62009-01-12 14:01:47 +01005511/*
5512 * Look out! "owner" is an entirely speculative pointer
5513 * access and not reliable.
5514 */
5515int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5516{
5517 unsigned int cpu;
5518 struct rq *rq;
5519
5520 if (!sched_feat(OWNER_SPIN))
5521 return 0;
5522
5523#ifdef CONFIG_DEBUG_PAGEALLOC
5524 /*
5525 * Need to access the cpu field knowing that
5526 * DEBUG_PAGEALLOC could have unmapped it if
5527 * the mutex owner just released it and exited.
5528 */
5529 if (probe_kernel_address(&owner->cpu, cpu))
5530 goto out;
5531#else
5532 cpu = owner->cpu;
5533#endif
5534
5535 /*
5536 * Even if the access succeeded (likely case),
5537 * the cpu field may no longer be valid.
5538 */
5539 if (cpu >= nr_cpumask_bits)
5540 goto out;
5541
5542 /*
5543 * We need to validate that we can do a
5544 * get_cpu() and that we have the percpu area.
5545 */
5546 if (!cpu_online(cpu))
5547 goto out;
5548
5549 rq = cpu_rq(cpu);
5550
5551 for (;;) {
5552 /*
5553 * Owner changed, break to re-assess state.
5554 */
5555 if (lock->owner != owner)
5556 break;
5557
5558 /*
5559 * Is that owner really running on that cpu?
5560 */
5561 if (task_thread_info(rq->curr) != owner || need_resched())
5562 return 0;
5563
5564 cpu_relax();
5565 }
5566out:
5567 return 1;
5568}
5569#endif
5570
Linus Torvalds1da177e2005-04-16 15:20:36 -07005571#ifdef CONFIG_PREEMPT
5572/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005573 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005574 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07005575 * occur there and call schedule directly.
5576 */
5577asmlinkage void __sched preempt_schedule(void)
5578{
5579 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005580
Linus Torvalds1da177e2005-04-16 15:20:36 -07005581 /*
5582 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005583 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07005585 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005586 return;
5587
Andi Kleen3a5c3592007-10-15 17:00:14 +02005588 do {
5589 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005590 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005591 sub_preempt_count(PREEMPT_ACTIVE);
5592
5593 /*
5594 * Check again in case we missed a preemption opportunity
5595 * between schedule and now.
5596 */
5597 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005598 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005599}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005600EXPORT_SYMBOL(preempt_schedule);
5601
5602/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005603 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604 * off of irq context.
5605 * Note, that this is called and return with irqs disabled. This will
5606 * protect us against recursive calling from irq.
5607 */
5608asmlinkage void __sched preempt_schedule_irq(void)
5609{
5610 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005611
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005612 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005613 BUG_ON(ti->preempt_count || !irqs_disabled());
5614
Andi Kleen3a5c3592007-10-15 17:00:14 +02005615 do {
5616 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005617 local_irq_enable();
5618 schedule();
5619 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005620 sub_preempt_count(PREEMPT_ACTIVE);
5621
5622 /*
5623 * Check again in case we missed a preemption opportunity
5624 * between schedule and now.
5625 */
5626 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005627 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005628}
5629
5630#endif /* CONFIG_PREEMPT */
5631
Peter Zijlstra63859d42009-09-15 19:14:42 +02005632int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005633 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005634{
Peter Zijlstra63859d42009-09-15 19:14:42 +02005635 return try_to_wake_up(curr->private, mode, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005636}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005637EXPORT_SYMBOL(default_wake_function);
5638
5639/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005640 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5641 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642 * number) then we wake all the non-exclusive tasks and one exclusive task.
5643 *
5644 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005645 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07005646 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5647 */
Johannes Weiner78ddb082009-04-14 16:53:05 +02005648static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
Peter Zijlstra63859d42009-09-15 19:14:42 +02005649 int nr_exclusive, int wake_flags, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005650{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005651 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005652
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005653 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005654 unsigned flags = curr->flags;
5655
Peter Zijlstra63859d42009-09-15 19:14:42 +02005656 if (curr->func(curr, mode, wake_flags, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07005657 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005658 break;
5659 }
5660}
5661
5662/**
5663 * __wake_up - wake up threads blocked on a waitqueue.
5664 * @q: the waitqueue
5665 * @mode: which threads
5666 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07005667 * @key: is directly passed to the wakeup function
David Howells50fa6102009-04-28 15:01:38 +01005668 *
5669 * It may be assumed that this function implies a write memory barrier before
5670 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005671 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005672void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005673 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005674{
5675 unsigned long flags;
5676
5677 spin_lock_irqsave(&q->lock, flags);
5678 __wake_up_common(q, mode, nr_exclusive, 0, key);
5679 spin_unlock_irqrestore(&q->lock, flags);
5680}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005681EXPORT_SYMBOL(__wake_up);
5682
5683/*
5684 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5685 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005686void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687{
5688 __wake_up_common(q, mode, 1, 0, NULL);
5689}
5690
Davide Libenzi4ede8162009-03-31 15:24:20 -07005691void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5692{
5693 __wake_up_common(q, mode, 1, 0, key);
5694}
5695
Linus Torvalds1da177e2005-04-16 15:20:36 -07005696/**
Davide Libenzi4ede8162009-03-31 15:24:20 -07005697 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005698 * @q: the waitqueue
5699 * @mode: which threads
5700 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Davide Libenzi4ede8162009-03-31 15:24:20 -07005701 * @key: opaque value to be passed to wakeup targets
Linus Torvalds1da177e2005-04-16 15:20:36 -07005702 *
5703 * The sync wakeup differs that the waker knows that it will schedule
5704 * away soon, so while the target thread will be woken up, it will not
5705 * be migrated to another CPU - ie. the two threads are 'synchronized'
5706 * with each other. This can prevent needless bouncing between CPUs.
5707 *
5708 * On UP it can prevent extra preemption.
David Howells50fa6102009-04-28 15:01:38 +01005709 *
5710 * It may be assumed that this function implies a write memory barrier before
5711 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712 */
Davide Libenzi4ede8162009-03-31 15:24:20 -07005713void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5714 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005715{
5716 unsigned long flags;
Peter Zijlstra7d478722009-09-14 19:55:44 +02005717 int wake_flags = WF_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718
5719 if (unlikely(!q))
5720 return;
5721
5722 if (unlikely(!nr_exclusive))
Peter Zijlstra7d478722009-09-14 19:55:44 +02005723 wake_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005724
5725 spin_lock_irqsave(&q->lock, flags);
Peter Zijlstra7d478722009-09-14 19:55:44 +02005726 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005727 spin_unlock_irqrestore(&q->lock, flags);
5728}
Davide Libenzi4ede8162009-03-31 15:24:20 -07005729EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5730
5731/*
5732 * __wake_up_sync - see __wake_up_sync_key()
5733 */
5734void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5735{
5736 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5737}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005738EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5739
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005740/**
5741 * complete: - signals a single thread waiting on this completion
5742 * @x: holds the state of this particular completion
5743 *
5744 * This will wake up a single thread waiting on this completion. Threads will be
5745 * awakened in the same order in which they were queued.
5746 *
5747 * See also complete_all(), wait_for_completion() and related routines.
David Howells50fa6102009-04-28 15:01:38 +01005748 *
5749 * It may be assumed that this function implies a write memory barrier before
5750 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005751 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005752void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005753{
5754 unsigned long flags;
5755
5756 spin_lock_irqsave(&x->wait.lock, flags);
5757 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005758 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005759 spin_unlock_irqrestore(&x->wait.lock, flags);
5760}
5761EXPORT_SYMBOL(complete);
5762
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005763/**
5764 * complete_all: - signals all threads waiting on this completion
5765 * @x: holds the state of this particular completion
5766 *
5767 * This will wake up all threads waiting on this particular completion event.
David Howells50fa6102009-04-28 15:01:38 +01005768 *
5769 * It may be assumed that this function implies a write memory barrier before
5770 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005771 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005772void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005773{
5774 unsigned long flags;
5775
5776 spin_lock_irqsave(&x->wait.lock, flags);
5777 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005778 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005779 spin_unlock_irqrestore(&x->wait.lock, flags);
5780}
5781EXPORT_SYMBOL(complete_all);
5782
Andi Kleen8cbbe862007-10-15 17:00:14 +02005783static inline long __sched
5784do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005785{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005786 if (!x->done) {
5787 DECLARE_WAITQUEUE(wait, current);
5788
5789 wait.flags |= WQ_FLAG_EXCLUSIVE;
5790 __add_wait_queue_tail(&x->wait, &wait);
5791 do {
Oleg Nesterov94d3d822008-08-20 16:54:41 -07005792 if (signal_pending_state(state, current)) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04005793 timeout = -ERESTARTSYS;
5794 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005795 }
5796 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005797 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005798 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005799 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005800 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005801 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005802 if (!x->done)
5803 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005804 }
5805 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04005806 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005807}
5808
5809static long __sched
5810wait_for_common(struct completion *x, long timeout, int state)
5811{
5812 might_sleep();
5813
5814 spin_lock_irq(&x->wait.lock);
5815 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005816 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005817 return timeout;
5818}
5819
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005820/**
5821 * wait_for_completion: - waits for completion of a task
5822 * @x: holds the state of this particular completion
5823 *
5824 * This waits to be signaled for completion of a specific task. It is NOT
5825 * interruptible and there is no timeout.
5826 *
5827 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5828 * and interrupt capability. Also see complete().
5829 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005830void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02005831{
5832 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005833}
5834EXPORT_SYMBOL(wait_for_completion);
5835
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005836/**
5837 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5838 * @x: holds the state of this particular completion
5839 * @timeout: timeout value in jiffies
5840 *
5841 * This waits for either a completion of a specific task to be signaled or for a
5842 * specified timeout to expire. The timeout is in jiffies. It is not
5843 * interruptible.
5844 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005845unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5847{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005848 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005849}
5850EXPORT_SYMBOL(wait_for_completion_timeout);
5851
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005852/**
5853 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5854 * @x: holds the state of this particular completion
5855 *
5856 * This waits for completion of a specific task to be signaled. It is
5857 * interruptible.
5858 */
Andi Kleen8cbbe862007-10-15 17:00:14 +02005859int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005860{
Andi Kleen51e97992007-10-18 21:32:55 +02005861 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5862 if (t == -ERESTARTSYS)
5863 return t;
5864 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005865}
5866EXPORT_SYMBOL(wait_for_completion_interruptible);
5867
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005868/**
5869 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5870 * @x: holds the state of this particular completion
5871 * @timeout: timeout value in jiffies
5872 *
5873 * This waits for either a completion of a specific task to be signaled or for a
5874 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5875 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005876unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005877wait_for_completion_interruptible_timeout(struct completion *x,
5878 unsigned long timeout)
5879{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005880 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005881}
5882EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5883
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005884/**
5885 * wait_for_completion_killable: - waits for completion of a task (killable)
5886 * @x: holds the state of this particular completion
5887 *
5888 * This waits to be signaled for completion of a specific task. It can be
5889 * interrupted by a kill signal.
5890 */
Matthew Wilcox009e5772007-12-06 12:29:54 -05005891int __sched wait_for_completion_killable(struct completion *x)
5892{
5893 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5894 if (t == -ERESTARTSYS)
5895 return t;
5896 return 0;
5897}
5898EXPORT_SYMBOL(wait_for_completion_killable);
5899
Dave Chinnerbe4de352008-08-15 00:40:44 -07005900/**
5901 * try_wait_for_completion - try to decrement a completion without blocking
5902 * @x: completion structure
5903 *
5904 * Returns: 0 if a decrement cannot be done without blocking
5905 * 1 if a decrement succeeded.
5906 *
5907 * If a completion is being used as a counting completion,
5908 * attempt to decrement the counter without blocking. This
5909 * enables us to avoid waiting if the resource the completion
5910 * is protecting is not available.
5911 */
5912bool try_wait_for_completion(struct completion *x)
5913{
5914 int ret = 1;
5915
5916 spin_lock_irq(&x->wait.lock);
5917 if (!x->done)
5918 ret = 0;
5919 else
5920 x->done--;
5921 spin_unlock_irq(&x->wait.lock);
5922 return ret;
5923}
5924EXPORT_SYMBOL(try_wait_for_completion);
5925
5926/**
5927 * completion_done - Test to see if a completion has any waiters
5928 * @x: completion structure
5929 *
5930 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5931 * 1 if there are no waiters.
5932 *
5933 */
5934bool completion_done(struct completion *x)
5935{
5936 int ret = 1;
5937
5938 spin_lock_irq(&x->wait.lock);
5939 if (!x->done)
5940 ret = 0;
5941 spin_unlock_irq(&x->wait.lock);
5942 return ret;
5943}
5944EXPORT_SYMBOL(completion_done);
5945
Andi Kleen8cbbe862007-10-15 17:00:14 +02005946static long __sched
5947sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02005948{
5949 unsigned long flags;
5950 wait_queue_t wait;
5951
5952 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005953
Andi Kleen8cbbe862007-10-15 17:00:14 +02005954 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005955
Andi Kleen8cbbe862007-10-15 17:00:14 +02005956 spin_lock_irqsave(&q->lock, flags);
5957 __add_wait_queue(q, &wait);
5958 spin_unlock(&q->lock);
5959 timeout = schedule_timeout(timeout);
5960 spin_lock_irq(&q->lock);
5961 __remove_wait_queue(q, &wait);
5962 spin_unlock_irqrestore(&q->lock, flags);
5963
5964 return timeout;
5965}
5966
5967void __sched interruptible_sleep_on(wait_queue_head_t *q)
5968{
5969 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005970}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005971EXPORT_SYMBOL(interruptible_sleep_on);
5972
Ingo Molnar0fec1712007-07-09 18:52:01 +02005973long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005974interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005975{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005976 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005977}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005978EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5979
Ingo Molnar0fec1712007-07-09 18:52:01 +02005980void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005981{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005982 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005983}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005984EXPORT_SYMBOL(sleep_on);
5985
Ingo Molnar0fec1712007-07-09 18:52:01 +02005986long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005987{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005988 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005989}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005990EXPORT_SYMBOL(sleep_on_timeout);
5991
Ingo Molnarb29739f2006-06-27 02:54:51 -07005992#ifdef CONFIG_RT_MUTEXES
5993
5994/*
5995 * rt_mutex_setprio - set the current priority of a task
5996 * @p: task
5997 * @prio: prio value (kernel-internal form)
5998 *
5999 * This function changes the 'effective' priority of a task. It does
6000 * not touch ->normal_prio like __setscheduler().
6001 *
6002 * Used by the rt_mutex code to implement priority inheritance logic.
6003 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006004void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07006005{
6006 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006007 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006008 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01006009 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006010
6011 BUG_ON(prio < 0 || prio > MAX_PRIO);
6012
6013 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006014 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006015
Andrew Mortond5f9f942007-05-08 20:27:06 -07006016 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006017 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006018 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006019 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02006020 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006021 if (running)
6022 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02006023
6024 if (rt_prio(prio))
6025 p->sched_class = &rt_sched_class;
6026 else
6027 p->sched_class = &fair_sched_class;
6028
Ingo Molnarb29739f2006-06-27 02:54:51 -07006029 p->prio = prio;
6030
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006031 if (running)
6032 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006033 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02006034 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006035
6036 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006037 }
6038 task_rq_unlock(rq, &flags);
6039}
6040
6041#endif
6042
Ingo Molnar36c8b582006-07-03 00:25:41 -07006043void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006044{
Ingo Molnardd41f592007-07-09 18:51:59 +02006045 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006046 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006047 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006048
6049 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6050 return;
6051 /*
6052 * We have to be careful, if called from sys_setpriority(),
6053 * the task might be in the middle of scheduling on another CPU.
6054 */
6055 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006056 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006057 /*
6058 * The RT priorities are set via sched_setscheduler(), but we still
6059 * allow the 'normal' nice value to be set - but as expected
6060 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02006061 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006062 */
Ingo Molnare05606d2007-07-09 18:51:59 +02006063 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006064 p->static_prio = NICE_TO_PRIO(nice);
6065 goto out_unlock;
6066 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006067 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02006068 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02006069 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006070
Linus Torvalds1da177e2005-04-16 15:20:36 -07006071 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07006072 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006073 old_prio = p->prio;
6074 p->prio = effective_prio(p);
6075 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006076
Ingo Molnardd41f592007-07-09 18:51:59 +02006077 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02006078 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006079 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07006080 * If the task increased its priority or is running and
6081 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006082 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07006083 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006084 resched_task(rq->curr);
6085 }
6086out_unlock:
6087 task_rq_unlock(rq, &flags);
6088}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006089EXPORT_SYMBOL(set_user_nice);
6090
Matt Mackalle43379f2005-05-01 08:59:00 -07006091/*
6092 * can_nice - check if a task can reduce its nice value
6093 * @p: task
6094 * @nice: nice value
6095 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006096int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07006097{
Matt Mackall024f4742005-08-18 11:24:19 -07006098 /* convert nice value [19,-20] to rlimit style value [1,40] */
6099 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006100
Matt Mackalle43379f2005-05-01 08:59:00 -07006101 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6102 capable(CAP_SYS_NICE));
6103}
6104
Linus Torvalds1da177e2005-04-16 15:20:36 -07006105#ifdef __ARCH_WANT_SYS_NICE
6106
6107/*
6108 * sys_nice - change the priority of the current process.
6109 * @increment: priority increment
6110 *
6111 * sys_setpriority is a more generic, but much slower function that
6112 * does similar things.
6113 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006114SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006115{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006116 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006117
6118 /*
6119 * Setpriority might change our priority at the same moment.
6120 * We don't have to worry. Conceptually one call occurs first
6121 * and we have a single winner.
6122 */
Matt Mackalle43379f2005-05-01 08:59:00 -07006123 if (increment < -40)
6124 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006125 if (increment > 40)
6126 increment = 40;
6127
Américo Wang2b8f8362009-02-16 18:54:21 +08006128 nice = TASK_NICE(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006129 if (nice < -20)
6130 nice = -20;
6131 if (nice > 19)
6132 nice = 19;
6133
Matt Mackalle43379f2005-05-01 08:59:00 -07006134 if (increment < 0 && !can_nice(current, nice))
6135 return -EPERM;
6136
Linus Torvalds1da177e2005-04-16 15:20:36 -07006137 retval = security_task_setnice(current, nice);
6138 if (retval)
6139 return retval;
6140
6141 set_user_nice(current, nice);
6142 return 0;
6143}
6144
6145#endif
6146
6147/**
6148 * task_prio - return the priority value of a given task.
6149 * @p: the task in question.
6150 *
6151 * This is the priority value as seen by users in /proc.
6152 * RT tasks are offset by -200. Normal tasks are centered
6153 * around 0, value goes from -16 to +15.
6154 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006155int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006156{
6157 return p->prio - MAX_RT_PRIO;
6158}
6159
6160/**
6161 * task_nice - return the nice value of a given task.
6162 * @p: the task in question.
6163 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006164int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006165{
6166 return TASK_NICE(p);
6167}
Pavel Roskin150d8be2008-03-05 16:56:37 -05006168EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006169
6170/**
6171 * idle_cpu - is a given cpu idle currently?
6172 * @cpu: the processor in question.
6173 */
6174int idle_cpu(int cpu)
6175{
6176 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6177}
6178
Linus Torvalds1da177e2005-04-16 15:20:36 -07006179/**
6180 * idle_task - return the idle task for a given cpu.
6181 * @cpu: the processor in question.
6182 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006183struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006184{
6185 return cpu_rq(cpu)->idle;
6186}
6187
6188/**
6189 * find_process_by_pid - find a process with a matching PID value.
6190 * @pid: the pid in question.
6191 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02006192static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006193{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07006194 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006195}
6196
6197/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02006198static void
6199__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006200{
Ingo Molnardd41f592007-07-09 18:51:59 +02006201 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006202
Linus Torvalds1da177e2005-04-16 15:20:36 -07006203 p->policy = policy;
6204 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006205 p->normal_prio = normal_prio(p);
6206 /* we are holding p->pi_lock already */
6207 p->prio = rt_mutex_getprio(p);
Peter Zijlstraffd44db2009-11-10 20:12:01 +01006208 if (rt_prio(p->prio))
6209 p->sched_class = &rt_sched_class;
6210 else
6211 p->sched_class = &fair_sched_class;
Peter Williams2dd73a42006-06-27 02:54:34 -07006212 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006213}
6214
David Howellsc69e8d92008-11-14 10:39:19 +11006215/*
6216 * check the target process has a UID that matches the current process's
6217 */
6218static bool check_same_owner(struct task_struct *p)
6219{
6220 const struct cred *cred = current_cred(), *pcred;
6221 bool match;
6222
6223 rcu_read_lock();
6224 pcred = __task_cred(p);
6225 match = (cred->euid == pcred->euid ||
6226 cred->euid == pcred->uid);
6227 rcu_read_unlock();
6228 return match;
6229}
6230
Rusty Russell961ccdd2008-06-23 13:55:38 +10006231static int __sched_setscheduler(struct task_struct *p, int policy,
6232 struct sched_param *param, bool user)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006233{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006234 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006235 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01006236 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006237 struct rq *rq;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006238 int reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006239
Steven Rostedt66e53932006-06-27 02:54:44 -07006240 /* may grab non-irq protected spin_locks */
6241 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07006242recheck:
6243 /* double check policy once rq lock held */
Lennart Poetteringca94c442009-06-15 17:17:47 +02006244 if (policy < 0) {
6245 reset_on_fork = p->sched_reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006246 policy = oldpolicy = p->policy;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006247 } else {
6248 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6249 policy &= ~SCHED_RESET_ON_FORK;
6250
6251 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6252 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6253 policy != SCHED_IDLE)
6254 return -EINVAL;
6255 }
6256
Linus Torvalds1da177e2005-04-16 15:20:36 -07006257 /*
6258 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02006259 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6260 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006261 */
6262 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006263 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04006264 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006265 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02006266 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006267 return -EINVAL;
6268
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006269 /*
6270 * Allow unprivileged RT tasks to decrease priority:
6271 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10006272 if (user && !capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02006273 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006274 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006275
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006276 if (!lock_task_sighand(p, &flags))
6277 return -ESRCH;
6278 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6279 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006280
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006281 /* can't set/change the rt policy */
6282 if (policy != p->policy && !rlim_rtprio)
6283 return -EPERM;
6284
6285 /* can't increase priority */
6286 if (param->sched_priority > p->rt_priority &&
6287 param->sched_priority > rlim_rtprio)
6288 return -EPERM;
6289 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006290 /*
6291 * Like positive nice levels, dont allow tasks to
6292 * move out of SCHED_IDLE either:
6293 */
6294 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6295 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006296
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006297 /* can't change other user's priorities */
David Howellsc69e8d92008-11-14 10:39:19 +11006298 if (!check_same_owner(p))
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006299 return -EPERM;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006300
6301 /* Normal users shall not reset the sched_reset_on_fork flag */
6302 if (p->sched_reset_on_fork && !reset_on_fork)
6303 return -EPERM;
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006304 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006305
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006306 if (user) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006307#ifdef CONFIG_RT_GROUP_SCHED
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006308 /*
6309 * Do not allow realtime tasks into groups that have no runtime
6310 * assigned.
6311 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02006312 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6313 task_group(p)->rt_bandwidth.rt_runtime == 0)
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006314 return -EPERM;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006315#endif
6316
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006317 retval = security_task_setscheduler(p, policy, param);
6318 if (retval)
6319 return retval;
6320 }
6321
Linus Torvalds1da177e2005-04-16 15:20:36 -07006322 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07006323 * make sure no PI-waiters arrive (or leave) while we are
6324 * changing the priority of the task:
6325 */
Thomas Gleixner1d615482009-11-17 14:54:03 +01006326 raw_spin_lock_irqsave(&p->pi_lock, flags);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006327 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07006328 * To be able to change p->policy safely, the apropriate
6329 * runqueue lock must be held.
6330 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07006331 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006332 /* recheck policy now with rq lock held */
6333 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6334 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006335 __task_rq_unlock(rq);
Thomas Gleixner1d615482009-11-17 14:54:03 +01006336 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006337 goto recheck;
6338 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02006339 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006340 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006341 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006342 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006343 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006344 if (running)
6345 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02006346
Lennart Poetteringca94c442009-06-15 17:17:47 +02006347 p->sched_reset_on_fork = reset_on_fork;
6348
Linus Torvalds1da177e2005-04-16 15:20:36 -07006349 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006350 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02006351
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006352 if (running)
6353 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006354 if (on_rq) {
6355 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006356
6357 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006358 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07006359 __task_rq_unlock(rq);
Thomas Gleixner1d615482009-11-17 14:54:03 +01006360 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006361
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07006362 rt_mutex_adjust_pi(p);
6363
Linus Torvalds1da177e2005-04-16 15:20:36 -07006364 return 0;
6365}
Rusty Russell961ccdd2008-06-23 13:55:38 +10006366
6367/**
6368 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6369 * @p: the task in question.
6370 * @policy: new policy.
6371 * @param: structure containing the new RT priority.
6372 *
6373 * NOTE that the task may be already dead.
6374 */
6375int sched_setscheduler(struct task_struct *p, int policy,
6376 struct sched_param *param)
6377{
6378 return __sched_setscheduler(p, policy, param, true);
6379}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006380EXPORT_SYMBOL_GPL(sched_setscheduler);
6381
Rusty Russell961ccdd2008-06-23 13:55:38 +10006382/**
6383 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6384 * @p: the task in question.
6385 * @policy: new policy.
6386 * @param: structure containing the new RT priority.
6387 *
6388 * Just like sched_setscheduler, only don't bother checking if the
6389 * current context has permission. For example, this is needed in
6390 * stop_machine(): we create temporary high priority worker threads,
6391 * but our caller might not have that capability.
6392 */
6393int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6394 struct sched_param *param)
6395{
6396 return __sched_setscheduler(p, policy, param, false);
6397}
6398
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006399static int
6400do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006401{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006402 struct sched_param lparam;
6403 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006404 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006405
6406 if (!param || pid < 0)
6407 return -EINVAL;
6408 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6409 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006410
6411 rcu_read_lock();
6412 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006413 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006414 if (p != NULL)
6415 retval = sched_setscheduler(p, policy, &lparam);
6416 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07006417
Linus Torvalds1da177e2005-04-16 15:20:36 -07006418 return retval;
6419}
6420
6421/**
6422 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6423 * @pid: the pid in question.
6424 * @policy: new policy.
6425 * @param: structure containing the new RT priority.
6426 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006427SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6428 struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006429{
Jason Baronc21761f2006-01-18 17:43:03 -08006430 /* negative values for policy are not valid */
6431 if (policy < 0)
6432 return -EINVAL;
6433
Linus Torvalds1da177e2005-04-16 15:20:36 -07006434 return do_sched_setscheduler(pid, policy, param);
6435}
6436
6437/**
6438 * sys_sched_setparam - set/change the RT priority of a thread
6439 * @pid: the pid in question.
6440 * @param: structure containing the new RT priority.
6441 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006442SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006443{
6444 return do_sched_setscheduler(pid, -1, param);
6445}
6446
6447/**
6448 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6449 * @pid: the pid in question.
6450 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006451SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006452{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006453 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006454 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006455
6456 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006457 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006458
6459 retval = -ESRCH;
6460 read_lock(&tasklist_lock);
6461 p = find_process_by_pid(pid);
6462 if (p) {
6463 retval = security_task_getscheduler(p);
6464 if (!retval)
Lennart Poetteringca94c442009-06-15 17:17:47 +02006465 retval = p->policy
6466 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006467 }
6468 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006469 return retval;
6470}
6471
6472/**
Lennart Poetteringca94c442009-06-15 17:17:47 +02006473 * sys_sched_getparam - get the RT priority of a thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07006474 * @pid: the pid in question.
6475 * @param: structure containing the RT priority.
6476 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006477SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006478{
6479 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006480 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006481 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006482
6483 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006484 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006485
6486 read_lock(&tasklist_lock);
6487 p = find_process_by_pid(pid);
6488 retval = -ESRCH;
6489 if (!p)
6490 goto out_unlock;
6491
6492 retval = security_task_getscheduler(p);
6493 if (retval)
6494 goto out_unlock;
6495
6496 lp.sched_priority = p->rt_priority;
6497 read_unlock(&tasklist_lock);
6498
6499 /*
6500 * This one might sleep, we cannot do it with a spinlock held ...
6501 */
6502 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6503
Linus Torvalds1da177e2005-04-16 15:20:36 -07006504 return retval;
6505
6506out_unlock:
6507 read_unlock(&tasklist_lock);
6508 return retval;
6509}
6510
Rusty Russell96f874e22008-11-25 02:35:14 +10306511long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006512{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306513 cpumask_var_t cpus_allowed, new_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006514 struct task_struct *p;
6515 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006517 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006518 read_lock(&tasklist_lock);
6519
6520 p = find_process_by_pid(pid);
6521 if (!p) {
6522 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006523 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006524 return -ESRCH;
6525 }
6526
6527 /*
6528 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006529 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07006530 * usage count and then drop tasklist_lock.
6531 */
6532 get_task_struct(p);
6533 read_unlock(&tasklist_lock);
6534
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306535 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6536 retval = -ENOMEM;
6537 goto out_put_task;
6538 }
6539 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6540 retval = -ENOMEM;
6541 goto out_free_cpus_allowed;
6542 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006543 retval = -EPERM;
David Howellsc69e8d92008-11-14 10:39:19 +11006544 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006545 goto out_unlock;
6546
David Quigleye7834f82006-06-23 02:03:59 -07006547 retval = security_task_setscheduler(p, 0, NULL);
6548 if (retval)
6549 goto out_unlock;
6550
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306551 cpuset_cpus_allowed(p, cpus_allowed);
6552 cpumask_and(new_mask, in_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006553 again:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306554 retval = set_cpus_allowed_ptr(p, new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006555
Paul Menage8707d8b2007-10-18 23:40:22 -07006556 if (!retval) {
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306557 cpuset_cpus_allowed(p, cpus_allowed);
6558 if (!cpumask_subset(new_mask, cpus_allowed)) {
Paul Menage8707d8b2007-10-18 23:40:22 -07006559 /*
6560 * We must have raced with a concurrent cpuset
6561 * update. Just reset the cpus_allowed to the
6562 * cpuset's cpus_allowed
6563 */
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306564 cpumask_copy(new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006565 goto again;
6566 }
6567 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006568out_unlock:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306569 free_cpumask_var(new_mask);
6570out_free_cpus_allowed:
6571 free_cpumask_var(cpus_allowed);
6572out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006573 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006574 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006575 return retval;
6576}
6577
6578static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e22008-11-25 02:35:14 +10306579 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006580{
Rusty Russell96f874e22008-11-25 02:35:14 +10306581 if (len < cpumask_size())
6582 cpumask_clear(new_mask);
6583 else if (len > cpumask_size())
6584 len = cpumask_size();
6585
Linus Torvalds1da177e2005-04-16 15:20:36 -07006586 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6587}
6588
6589/**
6590 * sys_sched_setaffinity - set the cpu affinity of a process
6591 * @pid: pid of the process
6592 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6593 * @user_mask_ptr: user-space pointer to the new cpu mask
6594 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006595SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6596 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006597{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306598 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006599 int retval;
6600
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306601 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6602 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006603
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306604 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6605 if (retval == 0)
6606 retval = sched_setaffinity(pid, new_mask);
6607 free_cpumask_var(new_mask);
6608 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006609}
6610
Rusty Russell96f874e22008-11-25 02:35:14 +10306611long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006612{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006613 struct task_struct *p;
Thomas Gleixner31605682009-12-08 20:24:16 +00006614 unsigned long flags;
6615 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006616 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006617
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006618 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619 read_lock(&tasklist_lock);
6620
6621 retval = -ESRCH;
6622 p = find_process_by_pid(pid);
6623 if (!p)
6624 goto out_unlock;
6625
David Quigleye7834f82006-06-23 02:03:59 -07006626 retval = security_task_getscheduler(p);
6627 if (retval)
6628 goto out_unlock;
6629
Thomas Gleixner31605682009-12-08 20:24:16 +00006630 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e22008-11-25 02:35:14 +10306631 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
Thomas Gleixner31605682009-12-08 20:24:16 +00006632 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006633
6634out_unlock:
6635 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006636 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006637
Ulrich Drepper9531b622007-08-09 11:16:46 +02006638 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006639}
6640
6641/**
6642 * sys_sched_getaffinity - get the cpu affinity of a process
6643 * @pid: pid of the process
6644 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6645 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6646 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006647SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6648 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006649{
6650 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10306651 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006652
Rusty Russellf17c8602008-11-25 02:35:11 +10306653 if (len < cpumask_size())
Linus Torvalds1da177e2005-04-16 15:20:36 -07006654 return -EINVAL;
6655
Rusty Russellf17c8602008-11-25 02:35:11 +10306656 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6657 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006658
Rusty Russellf17c8602008-11-25 02:35:11 +10306659 ret = sched_getaffinity(pid, mask);
6660 if (ret == 0) {
6661 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6662 ret = -EFAULT;
6663 else
6664 ret = cpumask_size();
6665 }
6666 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006667
Rusty Russellf17c8602008-11-25 02:35:11 +10306668 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006669}
6670
6671/**
6672 * sys_sched_yield - yield the current processor to other threads.
6673 *
Ingo Molnardd41f592007-07-09 18:51:59 +02006674 * This function yields the current CPU to other tasks. If there are no
6675 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006676 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006677SYSCALL_DEFINE0(sched_yield)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006678{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006679 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006680
Ingo Molnar2d723762007-10-15 17:00:12 +02006681 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02006682 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006683
6684 /*
6685 * Since we are going to call schedule() anyway, there's
6686 * no need to preempt or enable interrupts:
6687 */
6688 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07006689 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Thomas Gleixner9828ea92009-12-03 20:55:53 +01006690 do_raw_spin_unlock(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006691 preempt_enable_no_resched();
6692
6693 schedule();
6694
6695 return 0;
6696}
6697
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006698static inline int should_resched(void)
6699{
6700 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6701}
6702
Andrew Mortone7b38402006-06-30 01:56:00 -07006703static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006704{
Frederic Weisbeckere7aaaa62009-07-16 15:44:29 +02006705 add_preempt_count(PREEMPT_ACTIVE);
6706 schedule();
6707 sub_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006708}
6709
Herbert Xu02b67cc32008-01-25 21:08:28 +01006710int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006711{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006712 if (should_resched()) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006713 __cond_resched();
6714 return 1;
6715 }
6716 return 0;
6717}
Herbert Xu02b67cc32008-01-25 21:08:28 +01006718EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006719
6720/*
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006721 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07006722 * call schedule, and on return reacquire the lock.
6723 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006724 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07006725 * operations here to prevent schedule() from being called twice (once via
6726 * spin_unlock(), once by hand).
6727 */
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006728int __cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006729{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006730 int resched = should_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07006731 int ret = 0;
6732
Peter Zijlstraf607c662009-07-20 19:16:29 +02006733 lockdep_assert_held(lock);
6734
Nick Piggin95c354f2008-01-30 13:31:20 +01006735 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006736 spin_unlock(lock);
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006737 if (resched)
Nick Piggin95c354f2008-01-30 13:31:20 +01006738 __cond_resched();
6739 else
6740 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07006741 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006742 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006743 }
Jan Kara6df3cec2005-06-13 15:52:32 -07006744 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006745}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006746EXPORT_SYMBOL(__cond_resched_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006747
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006748int __sched __cond_resched_softirq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006749{
6750 BUG_ON(!in_softirq());
6751
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006752 if (should_resched()) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07006753 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006754 __cond_resched();
6755 local_bh_disable();
6756 return 1;
6757 }
6758 return 0;
6759}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006760EXPORT_SYMBOL(__cond_resched_softirq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006761
Linus Torvalds1da177e2005-04-16 15:20:36 -07006762/**
6763 * yield - yield the current processor to other threads.
6764 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08006765 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07006766 * thread runnable and calls sys_sched_yield().
6767 */
6768void __sched yield(void)
6769{
6770 set_current_state(TASK_RUNNING);
6771 sys_sched_yield();
6772}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006773EXPORT_SYMBOL(yield);
6774
6775/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006776 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07006777 * that process accounting knows that this is a task in IO wait state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006778 */
6779void __sched io_schedule(void)
6780{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006781 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006782
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006783 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006784 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006785 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006786 schedule();
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006787 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006788 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006789 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006790}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006791EXPORT_SYMBOL(io_schedule);
6792
6793long __sched io_schedule_timeout(long timeout)
6794{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006795 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006796 long ret;
6797
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006798 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006799 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006800 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006801 ret = schedule_timeout(timeout);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006802 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006803 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006804 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006805 return ret;
6806}
6807
6808/**
6809 * sys_sched_get_priority_max - return maximum RT priority.
6810 * @policy: scheduling class.
6811 *
6812 * this syscall returns the maximum rt_priority that can be used
6813 * by a given scheduling class.
6814 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006815SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006816{
6817 int ret = -EINVAL;
6818
6819 switch (policy) {
6820 case SCHED_FIFO:
6821 case SCHED_RR:
6822 ret = MAX_USER_RT_PRIO-1;
6823 break;
6824 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006825 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006826 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006827 ret = 0;
6828 break;
6829 }
6830 return ret;
6831}
6832
6833/**
6834 * sys_sched_get_priority_min - return minimum RT priority.
6835 * @policy: scheduling class.
6836 *
6837 * this syscall returns the minimum rt_priority that can be used
6838 * by a given scheduling class.
6839 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006840SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006841{
6842 int ret = -EINVAL;
6843
6844 switch (policy) {
6845 case SCHED_FIFO:
6846 case SCHED_RR:
6847 ret = 1;
6848 break;
6849 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006850 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006851 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006852 ret = 0;
6853 }
6854 return ret;
6855}
6856
6857/**
6858 * sys_sched_rr_get_interval - return the default timeslice of a process.
6859 * @pid: pid of the process.
6860 * @interval: userspace pointer to the timeslice value.
6861 *
6862 * this syscall writes the default timeslice value of a given process
6863 * into the user-space timespec buffer. A value of '0' means infinity.
6864 */
Heiko Carstens17da2bd2009-01-14 14:14:10 +01006865SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Heiko Carstens754fe8d2009-01-14 14:14:09 +01006866 struct timespec __user *, interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006867{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006868 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006869 unsigned int time_slice;
Thomas Gleixnerdba091b2009-12-09 09:32:03 +01006870 unsigned long flags;
6871 struct rq *rq;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006872 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006873 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006874
6875 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006876 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006877
6878 retval = -ESRCH;
6879 read_lock(&tasklist_lock);
6880 p = find_process_by_pid(pid);
6881 if (!p)
6882 goto out_unlock;
6883
6884 retval = security_task_getscheduler(p);
6885 if (retval)
6886 goto out_unlock;
6887
Thomas Gleixnerdba091b2009-12-09 09:32:03 +01006888 rq = task_rq_lock(p, &flags);
6889 time_slice = p->sched_class->get_rr_interval(rq, p);
6890 task_rq_unlock(rq, &flags);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006891
Linus Torvalds1da177e2005-04-16 15:20:36 -07006892 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006893 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006894 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006895 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006896
Linus Torvalds1da177e2005-04-16 15:20:36 -07006897out_unlock:
6898 read_unlock(&tasklist_lock);
6899 return retval;
6900}
6901
Steven Rostedt7c731e02008-05-12 21:20:41 +02006902static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006903
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006904void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006905{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006906 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006907 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006908
Linus Torvalds1da177e2005-04-16 15:20:36 -07006909 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006910 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07006911 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02006912#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07006913 if (state == TASK_RUNNING)
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006914 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006915 else
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006916 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006917#else
6918 if (state == TASK_RUNNING)
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006919 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006920 else
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006921 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006922#endif
6923#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05006924 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006925#endif
David Rientjesaa47b7e2009-05-04 01:38:05 -07006926 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6927 task_pid_nr(p), task_pid_nr(p->real_parent),
6928 (unsigned long)task_thread_info(p)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006929
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01006930 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006931}
6932
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006933void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006934{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006935 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006936
Ingo Molnar4bd77322007-07-11 21:21:47 +02006937#if BITS_PER_LONG == 32
6938 printk(KERN_INFO
6939 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006940#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02006941 printk(KERN_INFO
6942 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006943#endif
6944 read_lock(&tasklist_lock);
6945 do_each_thread(g, p) {
6946 /*
6947 * reset the NMI-timeout, listing all files on a slow
6948 * console might take alot of time:
6949 */
6950 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07006951 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006952 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006953 } while_each_thread(g, p);
6954
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07006955 touch_all_softlockup_watchdogs();
6956
Ingo Molnardd41f592007-07-09 18:51:59 +02006957#ifdef CONFIG_SCHED_DEBUG
6958 sysrq_sched_debug_show();
6959#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006960 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006961 /*
6962 * Only show locks if all tasks are dumped:
6963 */
Shmulik Ladkani93335a22009-11-25 15:23:41 +02006964 if (!state_filter)
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006965 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006966}
6967
Ingo Molnar1df21052007-07-09 18:51:58 +02006968void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6969{
Ingo Molnardd41f592007-07-09 18:51:59 +02006970 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02006971}
6972
Ingo Molnarf340c0d2005-06-28 16:40:42 +02006973/**
6974 * init_idle - set up an idle thread for a given CPU
6975 * @idle: task in question
6976 * @cpu: cpu the idle task belongs to
6977 *
6978 * NOTE: this function does not set the idle thread's NEED_RESCHED
6979 * flag, to make booting more robust.
6980 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07006981void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006982{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006983 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006984 unsigned long flags;
6985
Thomas Gleixner05fa7852009-11-17 14:28:38 +01006986 raw_spin_lock_irqsave(&rq->lock, flags);
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01006987
Ingo Molnardd41f592007-07-09 18:51:59 +02006988 __sched_fork(idle);
6989 idle->se.exec_start = sched_clock();
6990
Rusty Russell96f874e22008-11-25 02:35:14 +10306991 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
Ingo Molnardd41f592007-07-09 18:51:59 +02006992 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006993
Linus Torvalds1da177e2005-04-16 15:20:36 -07006994 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07006995#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6996 idle->oncpu = 1;
6997#endif
Thomas Gleixner05fa7852009-11-17 14:28:38 +01006998 raw_spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006999
7000 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07007001#if defined(CONFIG_PREEMPT)
7002 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7003#else
Al Viroa1261f542005-11-13 16:06:55 -08007004 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07007005#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02007006 /*
7007 * The idle tasks have their own, simple scheduling class:
7008 */
7009 idle->sched_class = &idle_sched_class;
Frederic Weisbeckerfb526072008-11-25 21:07:04 +01007010 ftrace_graph_init_task(idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007011}
7012
7013/*
7014 * In a system that switches off the HZ timer nohz_cpu_mask
7015 * indicates which cpus entered this state. This is used
7016 * in the rcu update to wait only for active cpus. For system
7017 * which do not switch off the HZ timer nohz_cpu_mask should
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10307018 * always be CPU_BITS_NONE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007019 */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10307020cpumask_var_t nohz_cpu_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007021
Ingo Molnar19978ca2007-11-09 22:39:38 +01007022/*
7023 * Increase the granularity value when there are more CPUs,
7024 * because with more CPUs the 'effective latency' as visible
7025 * to users decreases. But the relationship is not linear,
7026 * so pick a second-best guess by going with the log2 of the
7027 * number of CPUs.
7028 *
7029 * This idea comes from the SD scheduler of Con Kolivas:
7030 */
Christian Ehrhardtacb4a842009-11-30 12:16:48 +01007031static int get_update_sysctl_factor(void)
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01007032{
Mike Galbraith4ca3ef72009-12-10 09:25:53 +01007033 unsigned int cpus = min_t(int, num_online_cpus(), 8);
Christian Ehrhardt1983a922009-11-30 12:16:47 +01007034 unsigned int factor;
7035
7036 switch (sysctl_sched_tunable_scaling) {
7037 case SCHED_TUNABLESCALING_NONE:
7038 factor = 1;
7039 break;
7040 case SCHED_TUNABLESCALING_LINEAR:
7041 factor = cpus;
7042 break;
7043 case SCHED_TUNABLESCALING_LOG:
7044 default:
7045 factor = 1 + ilog2(cpus);
7046 break;
7047 }
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01007048
Christian Ehrhardtacb4a842009-11-30 12:16:48 +01007049 return factor;
7050}
7051
7052static void update_sysctl(void)
7053{
7054 unsigned int factor = get_update_sysctl_factor();
7055
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01007056#define SET_SYSCTL(name) \
7057 (sysctl_##name = (factor) * normalized_sysctl_##name)
7058 SET_SYSCTL(sched_min_granularity);
7059 SET_SYSCTL(sched_latency);
7060 SET_SYSCTL(sched_wakeup_granularity);
7061 SET_SYSCTL(sched_shares_ratelimit);
7062#undef SET_SYSCTL
7063}
7064
Ingo Molnar19978ca2007-11-09 22:39:38 +01007065static inline void sched_init_granularity(void)
7066{
Christian Ehrhardt0bcdcf22009-11-30 12:16:46 +01007067 update_sysctl();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007068}
7069
Linus Torvalds1da177e2005-04-16 15:20:36 -07007070#ifdef CONFIG_SMP
7071/*
7072 * This is how migration works:
7073 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07007074 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07007075 * runqueue and wake up that CPU's migration thread.
7076 * 2) we down() the locked semaphore => thread blocks.
7077 * 3) migration thread wakes up (implicitly it forces the migrated
7078 * thread off the CPU)
7079 * 4) it gets the migration request and checks whether the migrated
7080 * task is still in the wrong runqueue.
7081 * 5) if it's in the wrong runqueue then the migration thread removes
7082 * it and puts it into the right queue.
7083 * 6) migration thread up()s the semaphore.
7084 * 7) we wake up and the migration is done.
7085 */
7086
7087/*
7088 * Change a given task's CPU affinity. Migrate the thread to a
7089 * proper CPU and schedule it away if the CPU it's executing on
7090 * is removed from the allowed bitmask.
7091 *
7092 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007093 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07007094 * call is not atomic; no spinlocks may be held.
7095 */
Rusty Russell96f874e22008-11-25 02:35:14 +10307096int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007097{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007098 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007099 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007100 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007101 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007102
7103 rq = task_rq_lock(p, &flags);
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007104 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007105 ret = -EINVAL;
7106 goto out;
7107 }
7108
David Rientjes9985b0b2008-06-05 12:57:11 -07007109 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
Rusty Russell96f874e22008-11-25 02:35:14 +10307110 !cpumask_equal(&p->cpus_allowed, new_mask))) {
David Rientjes9985b0b2008-06-05 12:57:11 -07007111 ret = -EINVAL;
7112 goto out;
7113 }
7114
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007115 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007116 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007117 else {
Rusty Russell96f874e22008-11-25 02:35:14 +10307118 cpumask_copy(&p->cpus_allowed, new_mask);
7119 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007120 }
7121
Linus Torvalds1da177e2005-04-16 15:20:36 -07007122 /* Can the task run on the task's current CPU? If so, we're done */
Rusty Russell96f874e22008-11-25 02:35:14 +10307123 if (cpumask_test_cpu(task_cpu(p), new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007124 goto out;
7125
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007126 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007127 /* Need help from migration thread: drop lock and wait. */
Peter Zijlstra693525e2009-07-21 13:56:38 +02007128 struct task_struct *mt = rq->migration_thread;
7129
7130 get_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007131 task_rq_unlock(rq, &flags);
7132 wake_up_process(rq->migration_thread);
Peter Zijlstra693525e2009-07-21 13:56:38 +02007133 put_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007134 wait_for_completion(&req.done);
7135 tlb_migrate_finish(p->mm);
7136 return 0;
7137 }
7138out:
7139 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007140
Linus Torvalds1da177e2005-04-16 15:20:36 -07007141 return ret;
7142}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007143EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007144
7145/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007146 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147 * this because either it can't run here any more (set_cpus_allowed()
7148 * away from this CPU, or CPU going down), or because we're
7149 * attempting to rebalance this task on exec (sched_exec).
7150 *
7151 * So we race with normal scheduler movements, but that's OK, as long
7152 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07007153 *
7154 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007155 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07007156static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007157{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007158 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02007159 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007160
Max Krasnyanskye761b772008-07-15 04:43:49 -07007161 if (unlikely(!cpu_active(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07007162 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007163
7164 rq_src = cpu_rq(src_cpu);
7165 rq_dest = cpu_rq(dest_cpu);
7166
7167 double_rq_lock(rq_src, rq_dest);
7168 /* Already moved. */
7169 if (task_cpu(p) != src_cpu)
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007170 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007171 /* Affinity changed (again). */
Rusty Russell96f874e22008-11-25 02:35:14 +10307172 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007173 goto fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007174
Ingo Molnardd41f592007-07-09 18:51:59 +02007175 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007176 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007177 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007178
Linus Torvalds1da177e2005-04-16 15:20:36 -07007179 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007180 if (on_rq) {
7181 activate_task(rq_dest, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02007182 check_preempt_curr(rq_dest, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007183 }
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007184done:
Kirill Korotaevefc30812006-06-27 02:54:32 -07007185 ret = 1;
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007186fail:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007187 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07007188 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007189}
7190
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007191#define RCU_MIGRATION_IDLE 0
7192#define RCU_MIGRATION_NEED_QS 1
7193#define RCU_MIGRATION_GOT_QS 2
7194#define RCU_MIGRATION_MUST_SYNC 3
7195
Linus Torvalds1da177e2005-04-16 15:20:36 -07007196/*
7197 * migration_thread - this is a highprio system thread that performs
7198 * thread migration by bumping thread off CPU then 'pushing' onto
7199 * another runqueue.
7200 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07007201static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007202{
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007203 int badcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007204 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007205 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007206
7207 rq = cpu_rq(cpu);
7208 BUG_ON(rq->migration_thread != current);
7209
7210 set_current_state(TASK_INTERRUPTIBLE);
7211 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007212 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007213 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007214
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007215 raw_spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007216
7217 if (cpu_is_offline(cpu)) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007218 raw_spin_unlock_irq(&rq->lock);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007219 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007220 }
7221
7222 if (rq->active_balance) {
7223 active_load_balance(rq, cpu);
7224 rq->active_balance = 0;
7225 }
7226
7227 head = &rq->migration_queue;
7228
7229 if (list_empty(head)) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007230 raw_spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007231 schedule();
7232 set_current_state(TASK_INTERRUPTIBLE);
7233 continue;
7234 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07007235 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007236 list_del_init(head->next);
7237
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007238 if (req->task != NULL) {
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007239 raw_spin_unlock(&rq->lock);
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007240 __migrate_task(req->task, cpu, req->dest_cpu);
7241 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7242 req->dest_cpu = RCU_MIGRATION_GOT_QS;
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007243 raw_spin_unlock(&rq->lock);
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007244 } else {
7245 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007246 raw_spin_unlock(&rq->lock);
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007247 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7248 }
Nick Piggin674311d2005-06-25 14:57:27 -07007249 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007250
7251 complete(&req->done);
7252 }
7253 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007254
Linus Torvalds1da177e2005-04-16 15:20:36 -07007255 return 0;
7256}
7257
7258#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007259
7260static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7261{
7262 int ret;
7263
7264 local_irq_disable();
7265 ret = __migrate_task(p, src_cpu, dest_cpu);
7266 local_irq_enable();
7267 return ret;
7268}
7269
Kirill Korotaev054b9102006-12-10 02:20:11 -08007270/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007271 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08007272 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007273static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007274{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007275 int dest_cpu;
Mike Travis6ca09df2008-12-31 18:08:45 -08007276 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007277
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307278again:
7279 /* Look for allowed, online CPU in same node. */
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007280 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307281 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7282 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007283
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307284 /* Any allowed, online CPU? */
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007285 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307286 if (dest_cpu < nr_cpu_ids)
7287 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007288
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307289 /* No more Mr. Nice Guy. */
7290 if (dest_cpu >= nr_cpu_ids) {
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307291 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007292 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07007293
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307294 /*
7295 * Don't tell them about moving exiting tasks or
7296 * kernel threads (both mm NULL), since they never
7297 * leave kernel.
7298 */
7299 if (p->mm && printk_ratelimit()) {
7300 printk(KERN_INFO "process %d (%s) no "
7301 "longer affine to cpu%d\n",
7302 task_pid_nr(p), p->comm, dead_cpu);
Andi Kleen3a5c3592007-10-15 17:00:14 +02007303 }
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307304 }
7305
7306move:
7307 /* It can have affinity changed while we were choosing. */
7308 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7309 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007310}
7311
7312/*
7313 * While a dead CPU has no uninterruptible tasks queued at this point,
7314 * it might still have a nonzero ->nr_uninterruptible counter, because
7315 * for performance reasons the counter is not stricly tracking tasks to
7316 * their home CPUs. So we just add the counter to another CPU's counter,
7317 * to keep the global sum constant after CPU-down:
7318 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07007319static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007320{
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007321 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007322 unsigned long flags;
7323
7324 local_irq_save(flags);
7325 double_rq_lock(rq_src, rq_dest);
7326 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7327 rq_src->nr_uninterruptible = 0;
7328 double_rq_unlock(rq_src, rq_dest);
7329 local_irq_restore(flags);
7330}
7331
7332/* Run through task list and migrate tasks from the dead cpu. */
7333static void migrate_live_tasks(int src_cpu)
7334{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007335 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007336
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007337 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007338
Ingo Molnar48f24c42006-07-03 00:25:40 -07007339 do_each_thread(t, p) {
7340 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007341 continue;
7342
Ingo Molnar48f24c42006-07-03 00:25:40 -07007343 if (task_cpu(p) == src_cpu)
7344 move_task_off_dead_cpu(src_cpu, p);
7345 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007346
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007347 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007348}
7349
Ingo Molnardd41f592007-07-09 18:51:59 +02007350/*
7351 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007352 * It does so by boosting its priority to highest possible.
7353 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007354 */
7355void sched_idle_next(void)
7356{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007357 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07007358 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007359 struct task_struct *p = rq->idle;
7360 unsigned long flags;
7361
7362 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007363 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007364
Ingo Molnar48f24c42006-07-03 00:25:40 -07007365 /*
7366 * Strictly not necessary since rest of the CPUs are stopped by now
7367 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007368 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007369 raw_spin_lock_irqsave(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007370
Ingo Molnardd41f592007-07-09 18:51:59 +02007371 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007372
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007373 update_rq_clock(rq);
7374 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007375
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007376 raw_spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007377}
7378
Ingo Molnar48f24c42006-07-03 00:25:40 -07007379/*
7380 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07007381 * offline.
7382 */
7383void idle_task_exit(void)
7384{
7385 struct mm_struct *mm = current->active_mm;
7386
7387 BUG_ON(cpu_online(smp_processor_id()));
7388
7389 if (mm != &init_mm)
7390 switch_mm(mm, &init_mm, current);
7391 mmdrop(mm);
7392}
7393
Kirill Korotaev054b9102006-12-10 02:20:11 -08007394/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007395static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007396{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007397 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007398
7399 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07007400 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007401
7402 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07007403 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007404
Ingo Molnar48f24c42006-07-03 00:25:40 -07007405 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007406
7407 /*
7408 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007409 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07007410 * fine.
7411 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007412 raw_spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007413 move_task_off_dead_cpu(dead_cpu, p);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007414 raw_spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007415
Ingo Molnar48f24c42006-07-03 00:25:40 -07007416 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007417}
7418
7419/* release_task() removes task from tasklist, so we won't find dead tasks. */
7420static void migrate_dead_tasks(unsigned int dead_cpu)
7421{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007422 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007423 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007424
Ingo Molnardd41f592007-07-09 18:51:59 +02007425 for ( ; ; ) {
7426 if (!rq->nr_running)
7427 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02007428 update_rq_clock(rq);
Wang Chenb67802e2009-03-02 13:55:26 +08007429 next = pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02007430 if (!next)
7431 break;
Dmitry Adamushko79c53792008-06-29 00:16:56 +02007432 next->sched_class->put_prev_task(rq, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02007433 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02007434
Linus Torvalds1da177e2005-04-16 15:20:36 -07007435 }
7436}
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007437
7438/*
7439 * remove the tasks which were accounted by rq from calc_load_tasks.
7440 */
7441static void calc_global_load_remove(struct rq *rq)
7442{
7443 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
Thomas Gleixnera468d382009-07-17 14:15:46 +02007444 rq->calc_load_active = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007445}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007446#endif /* CONFIG_HOTPLUG_CPU */
7447
Nick Piggine692ab52007-07-26 13:40:43 +02007448#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7449
7450static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007451 {
7452 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007453 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007454 },
Eric W. Biederman56992302009-11-05 15:38:40 -08007455 {}
Nick Piggine692ab52007-07-26 13:40:43 +02007456};
7457
7458static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007459 {
7460 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007461 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007462 .child = sd_ctl_dir,
7463 },
Eric W. Biederman56992302009-11-05 15:38:40 -08007464 {}
Nick Piggine692ab52007-07-26 13:40:43 +02007465};
7466
7467static struct ctl_table *sd_alloc_ctl_entry(int n)
7468{
7469 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02007470 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02007471
Nick Piggine692ab52007-07-26 13:40:43 +02007472 return entry;
7473}
7474
Milton Miller6382bc92007-10-15 17:00:19 +02007475static void sd_free_ctl_entry(struct ctl_table **tablep)
7476{
Milton Millercd7900762007-10-17 16:55:11 +02007477 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02007478
Milton Millercd7900762007-10-17 16:55:11 +02007479 /*
7480 * In the intermediate directories, both the child directory and
7481 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007482 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02007483 * static strings and all have proc handlers.
7484 */
7485 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02007486 if (entry->child)
7487 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02007488 if (entry->proc_handler == NULL)
7489 kfree(entry->procname);
7490 }
Milton Miller6382bc92007-10-15 17:00:19 +02007491
7492 kfree(*tablep);
7493 *tablep = NULL;
7494}
7495
Nick Piggine692ab52007-07-26 13:40:43 +02007496static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02007497set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02007498 const char *procname, void *data, int maxlen,
7499 mode_t mode, proc_handler *proc_handler)
7500{
Nick Piggine692ab52007-07-26 13:40:43 +02007501 entry->procname = procname;
7502 entry->data = data;
7503 entry->maxlen = maxlen;
7504 entry->mode = mode;
7505 entry->proc_handler = proc_handler;
7506}
7507
7508static struct ctl_table *
7509sd_alloc_ctl_domain_table(struct sched_domain *sd)
7510{
Ingo Molnara5d8c342008-10-09 11:35:51 +02007511 struct ctl_table *table = sd_alloc_ctl_entry(13);
Nick Piggine692ab52007-07-26 13:40:43 +02007512
Milton Millerad1cdc12007-10-15 17:00:19 +02007513 if (table == NULL)
7514 return NULL;
7515
Alexey Dobriyane0361852007-08-09 11:16:46 +02007516 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007517 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007518 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007519 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007520 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007521 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007522 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007523 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007524 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007525 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007526 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007527 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007528 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007529 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007530 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02007531 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007532 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02007533 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007534 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02007535 &sd->cache_nice_tries,
7536 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007537 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02007538 sizeof(int), 0644, proc_dointvec_minmax);
Ingo Molnara5d8c342008-10-09 11:35:51 +02007539 set_table_entry(&table[11], "name", sd->name,
7540 CORENAME_MAX_SIZE, 0444, proc_dostring);
7541 /* &table[12] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02007542
7543 return table;
7544}
7545
Ingo Molnar9a4e7152007-11-28 15:52:56 +01007546static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02007547{
7548 struct ctl_table *entry, *table;
7549 struct sched_domain *sd;
7550 int domain_num = 0, i;
7551 char buf[32];
7552
7553 for_each_domain(cpu, sd)
7554 domain_num++;
7555 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02007556 if (table == NULL)
7557 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02007558
7559 i = 0;
7560 for_each_domain(cpu, sd) {
7561 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007562 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007563 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007564 entry->child = sd_alloc_ctl_domain_table(sd);
7565 entry++;
7566 i++;
7567 }
7568 return table;
7569}
7570
7571static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02007572static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007573{
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007574 int i, cpu_num = num_possible_cpus();
Nick Piggine692ab52007-07-26 13:40:43 +02007575 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7576 char buf[32];
7577
Milton Miller73785472007-10-24 18:23:48 +02007578 WARN_ON(sd_ctl_dir[0].child);
7579 sd_ctl_dir[0].child = entry;
7580
Milton Millerad1cdc12007-10-15 17:00:19 +02007581 if (entry == NULL)
7582 return;
7583
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007584 for_each_possible_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02007585 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007586 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007587 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007588 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02007589 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02007590 }
Milton Miller73785472007-10-24 18:23:48 +02007591
7592 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02007593 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7594}
Milton Miller6382bc92007-10-15 17:00:19 +02007595
Milton Miller73785472007-10-24 18:23:48 +02007596/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02007597static void unregister_sched_domain_sysctl(void)
7598{
Milton Miller73785472007-10-24 18:23:48 +02007599 if (sd_sysctl_header)
7600 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02007601 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007602 if (sd_ctl_dir[0].child)
7603 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02007604}
Nick Piggine692ab52007-07-26 13:40:43 +02007605#else
Milton Miller6382bc92007-10-15 17:00:19 +02007606static void register_sched_domain_sysctl(void)
7607{
7608}
7609static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007610{
7611}
7612#endif
7613
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007614static void set_rq_online(struct rq *rq)
7615{
7616 if (!rq->online) {
7617 const struct sched_class *class;
7618
Rusty Russellc6c49272008-11-25 02:35:05 +10307619 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007620 rq->online = 1;
7621
7622 for_each_class(class) {
7623 if (class->rq_online)
7624 class->rq_online(rq);
7625 }
7626 }
7627}
7628
7629static void set_rq_offline(struct rq *rq)
7630{
7631 if (rq->online) {
7632 const struct sched_class *class;
7633
7634 for_each_class(class) {
7635 if (class->rq_offline)
7636 class->rq_offline(rq);
7637 }
7638
Rusty Russellc6c49272008-11-25 02:35:05 +10307639 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007640 rq->online = 0;
7641 }
7642}
7643
Linus Torvalds1da177e2005-04-16 15:20:36 -07007644/*
7645 * migration_call - callback that gets triggered when a CPU is added.
7646 * Here we can start up the necessary migration thread for the new CPU.
7647 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007648static int __cpuinit
7649migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007650{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007651 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007652 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007653 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007654 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007655
7656 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07007657
Linus Torvalds1da177e2005-04-16 15:20:36 -07007658 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007659 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02007660 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007661 if (IS_ERR(p))
7662 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007663 kthread_bind(p, cpu);
7664 /* Must be high prio: stop_machine expects to yield to it. */
7665 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02007666 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007667 task_rq_unlock(rq, &flags);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007668 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007669 cpu_rq(cpu)->migration_thread = p;
Thomas Gleixnera468d382009-07-17 14:15:46 +02007670 rq->calc_load_update = calc_load_update;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007671 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007672
Linus Torvalds1da177e2005-04-16 15:20:36 -07007673 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007674 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007675 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007676 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007677
7678 /* Update our root-domain */
7679 rq = cpu_rq(cpu);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007680 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007681 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307682 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007683
7684 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007685 }
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007686 raw_spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007687 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007688
Linus Torvalds1da177e2005-04-16 15:20:36 -07007689#ifdef CONFIG_HOTPLUG_CPU
7690 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007691 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07007692 if (!cpu_rq(cpu)->migration_thread)
7693 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007694 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08007695 kthread_bind(cpu_rq(cpu)->migration_thread,
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307696 cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007697 kthread_stop(cpu_rq(cpu)->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007698 put_task_struct(cpu_rq(cpu)->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007699 cpu_rq(cpu)->migration_thread = NULL;
7700 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007701
Linus Torvalds1da177e2005-04-16 15:20:36 -07007702 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007703 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07007704 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007705 migrate_live_tasks(cpu);
7706 rq = cpu_rq(cpu);
7707 kthread_stop(rq->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007708 put_task_struct(rq->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007709 rq->migration_thread = NULL;
7710 /* Idle task back to normal (off runqueue, low prio) */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007711 raw_spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02007712 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007713 deactivate_task(rq, rq->idle, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02007714 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7715 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007716 migrate_dead_tasks(cpu);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007717 raw_spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07007718 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007719 migrate_nr_uninterruptible(rq);
7720 BUG_ON(rq->nr_running != 0);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007721 calc_global_load_remove(rq);
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007722 /*
7723 * No need to migrate the tasks: it was best-effort if
7724 * they didn't take sched_hotcpu_mutex. Just wake up
7725 * the requestors.
7726 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007727 raw_spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007728 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007729 struct migration_req *req;
7730
Linus Torvalds1da177e2005-04-16 15:20:36 -07007731 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07007732 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007733 list_del_init(&req->list);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007734 raw_spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007735 complete(&req->done);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007736 raw_spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007737 }
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007738 raw_spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007739 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007740
Gregory Haskins08f503b2008-03-10 17:59:11 -04007741 case CPU_DYING:
7742 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01007743 /* Update our root-domain */
7744 rq = cpu_rq(cpu);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007745 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007746 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307747 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007748 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007749 }
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007750 raw_spin_unlock_irqrestore(&rq->lock, flags);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007751 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007752#endif
7753 }
7754 return NOTIFY_OK;
7755}
7756
Paul Mackerrasf38b0822009-06-02 21:05:16 +10007757/*
7758 * Register at high priority so that task migration (migrate_all_tasks)
7759 * happens before everything else. This has to be lower priority than
Ingo Molnarcdd6c482009-09-21 12:02:48 +02007760 * the notifier in the perf_event subsystem, though.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007761 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07007762static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007763 .notifier_call = migration_call,
7764 .priority = 10
7765};
7766
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007767static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007768{
7769 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07007770 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007771
7772 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07007773 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7774 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007775 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7776 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007777
Thomas Gleixnera004cd42009-07-21 09:54:05 +02007778 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007779}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007780early_initcall(migration_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007781#endif
7782
7783#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07007784
Ingo Molnar3e9830d2007-10-15 17:00:13 +02007785#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007786
Mike Travisf6630112009-11-17 18:22:15 -06007787static __read_mostly int sched_domain_debug_enabled;
7788
7789static int __init sched_domain_debug_setup(char *str)
7790{
7791 sched_domain_debug_enabled = 1;
7792
7793 return 0;
7794}
7795early_param("sched_debug", sched_domain_debug_setup);
7796
Mike Travis7c16ec52008-04-04 18:11:11 -07007797static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
Rusty Russell96f874e22008-11-25 02:35:14 +10307798 struct cpumask *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007799{
7800 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07007801 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007802
Rusty Russell968ea6d2008-12-13 21:55:51 +10307803 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
Rusty Russell96f874e22008-11-25 02:35:14 +10307804 cpumask_clear(groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007805
7806 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7807
7808 if (!(sd->flags & SD_LOAD_BALANCE)) {
7809 printk("does not load-balance\n");
7810 if (sd->parent)
7811 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7812 " has parent");
7813 return -1;
7814 }
7815
Li Zefaneefd7962008-11-04 16:15:37 +08007816 printk(KERN_CONT "span %s level %s\n", str, sd->name);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007817
Rusty Russell758b2cd2008-11-25 02:35:04 +10307818 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007819 printk(KERN_ERR "ERROR: domain->span does not contain "
7820 "CPU%d\n", cpu);
7821 }
Rusty Russell758b2cd2008-11-25 02:35:04 +10307822 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007823 printk(KERN_ERR "ERROR: domain->groups does not contain"
7824 " CPU%d\n", cpu);
7825 }
7826
7827 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7828 do {
7829 if (!group) {
7830 printk("\n");
7831 printk(KERN_ERR "ERROR: group is NULL\n");
7832 break;
7833 }
7834
Peter Zijlstra18a38852009-09-01 10:34:39 +02007835 if (!group->cpu_power) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007836 printk(KERN_CONT "\n");
7837 printk(KERN_ERR "ERROR: domain->cpu_power not "
7838 "set\n");
7839 break;
7840 }
7841
Rusty Russell758b2cd2008-11-25 02:35:04 +10307842 if (!cpumask_weight(sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007843 printk(KERN_CONT "\n");
7844 printk(KERN_ERR "ERROR: empty group\n");
7845 break;
7846 }
7847
Rusty Russell758b2cd2008-11-25 02:35:04 +10307848 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007849 printk(KERN_CONT "\n");
7850 printk(KERN_ERR "ERROR: repeated CPUs\n");
7851 break;
7852 }
7853
Rusty Russell758b2cd2008-11-25 02:35:04 +10307854 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007855
Rusty Russell968ea6d2008-12-13 21:55:51 +10307856 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307857
7858 printk(KERN_CONT " %s", str);
Peter Zijlstra18a38852009-09-01 10:34:39 +02007859 if (group->cpu_power != SCHED_LOAD_SCALE) {
7860 printk(KERN_CONT " (cpu_power = %d)",
7861 group->cpu_power);
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307862 }
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007863
7864 group = group->next;
7865 } while (group != sd->groups);
7866 printk(KERN_CONT "\n");
7867
Rusty Russell758b2cd2008-11-25 02:35:04 +10307868 if (!cpumask_equal(sched_domain_span(sd), groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007869 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7870
Rusty Russell758b2cd2008-11-25 02:35:04 +10307871 if (sd->parent &&
7872 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007873 printk(KERN_ERR "ERROR: parent span is not a superset "
7874 "of domain->span\n");
7875 return 0;
7876}
7877
Linus Torvalds1da177e2005-04-16 15:20:36 -07007878static void sched_domain_debug(struct sched_domain *sd, int cpu)
7879{
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307880 cpumask_var_t groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007881 int level = 0;
7882
Mike Travisf6630112009-11-17 18:22:15 -06007883 if (!sched_domain_debug_enabled)
7884 return;
7885
Nick Piggin41c7ce92005-06-25 14:57:24 -07007886 if (!sd) {
7887 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7888 return;
7889 }
7890
Linus Torvalds1da177e2005-04-16 15:20:36 -07007891 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7892
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307893 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007894 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7895 return;
7896 }
7897
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007898 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007899 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007900 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007901 level++;
7902 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08007903 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007904 break;
7905 }
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307906 free_cpumask_var(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007907}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007908#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007909# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007910#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007911
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007912static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007913{
Rusty Russell758b2cd2008-11-25 02:35:04 +10307914 if (cpumask_weight(sched_domain_span(sd)) == 1)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007915 return 1;
7916
7917 /* Following flags need at least 2 groups */
7918 if (sd->flags & (SD_LOAD_BALANCE |
7919 SD_BALANCE_NEWIDLE |
7920 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007921 SD_BALANCE_EXEC |
7922 SD_SHARE_CPUPOWER |
7923 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007924 if (sd->groups != sd->groups->next)
7925 return 0;
7926 }
7927
7928 /* Following flags don't use groups */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02007929 if (sd->flags & (SD_WAKE_AFFINE))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007930 return 0;
7931
7932 return 1;
7933}
7934
Ingo Molnar48f24c42006-07-03 00:25:40 -07007935static int
7936sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007937{
7938 unsigned long cflags = sd->flags, pflags = parent->flags;
7939
7940 if (sd_degenerate(parent))
7941 return 1;
7942
Rusty Russell758b2cd2008-11-25 02:35:04 +10307943 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007944 return 0;
7945
Suresh Siddha245af2c2005-06-25 14:57:25 -07007946 /* Flags needing groups don't count if only 1 group in parent */
7947 if (parent->groups == parent->groups->next) {
7948 pflags &= ~(SD_LOAD_BALANCE |
7949 SD_BALANCE_NEWIDLE |
7950 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007951 SD_BALANCE_EXEC |
7952 SD_SHARE_CPUPOWER |
7953 SD_SHARE_PKG_RESOURCES);
Ken Chen54364992008-12-07 18:47:37 -08007954 if (nr_node_ids == 1)
7955 pflags &= ~SD_SERIALIZE;
Suresh Siddha245af2c2005-06-25 14:57:25 -07007956 }
7957 if (~cflags & pflags)
7958 return 0;
7959
7960 return 1;
7961}
7962
Rusty Russellc6c49272008-11-25 02:35:05 +10307963static void free_rootdomain(struct root_domain *rd)
7964{
Peter Zijlstra047106a2009-11-16 10:28:09 +01007965 synchronize_sched();
7966
Rusty Russell68e74562008-11-25 02:35:13 +10307967 cpupri_cleanup(&rd->cpupri);
7968
Rusty Russellc6c49272008-11-25 02:35:05 +10307969 free_cpumask_var(rd->rto_mask);
7970 free_cpumask_var(rd->online);
7971 free_cpumask_var(rd->span);
7972 kfree(rd);
7973}
7974
Gregory Haskins57d885f2008-01-25 21:08:18 +01007975static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7976{
Ingo Molnara0490fa2009-02-12 11:35:40 +01007977 struct root_domain *old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007978 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007979
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007980 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007981
7982 if (rq->rd) {
Ingo Molnara0490fa2009-02-12 11:35:40 +01007983 old_rd = rq->rd;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007984
Rusty Russellc6c49272008-11-25 02:35:05 +10307985 if (cpumask_test_cpu(rq->cpu, old_rd->online))
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007986 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007987
Rusty Russellc6c49272008-11-25 02:35:05 +10307988 cpumask_clear_cpu(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007989
Ingo Molnara0490fa2009-02-12 11:35:40 +01007990 /*
7991 * If we dont want to free the old_rt yet then
7992 * set old_rd to NULL to skip the freeing later
7993 * in this function:
7994 */
7995 if (!atomic_dec_and_test(&old_rd->refcount))
7996 old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007997 }
7998
7999 atomic_inc(&rd->refcount);
8000 rq->rd = rd;
8001
Rusty Russellc6c49272008-11-25 02:35:05 +10308002 cpumask_set_cpu(rq->cpu, rd->span);
Gregory Haskins00aec932009-07-30 10:57:23 -04008003 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04008004 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01008005
Thomas Gleixner05fa7852009-11-17 14:28:38 +01008006 raw_spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnara0490fa2009-02-12 11:35:40 +01008007
8008 if (old_rd)
8009 free_rootdomain(old_rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01008010}
8011
Li Zefanfd5e1b52009-06-15 13:34:19 +08008012static int init_rootdomain(struct root_domain *rd, bool bootmem)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008013{
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008014 gfp_t gfp = GFP_KERNEL;
8015
Gregory Haskins57d885f2008-01-25 21:08:18 +01008016 memset(rd, 0, sizeof(*rd));
8017
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008018 if (bootmem)
8019 gfp = GFP_NOWAIT;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02008020
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008021 if (!alloc_cpumask_var(&rd->span, gfp))
Li Zefan0c910d22009-01-06 17:39:06 +08008022 goto out;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008023 if (!alloc_cpumask_var(&rd->online, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10308024 goto free_span;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008025 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10308026 goto free_online;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02008027
Pekka Enberg0fb53022009-06-11 08:41:22 +03008028 if (cpupri_init(&rd->cpupri, bootmem) != 0)
Rusty Russell68e74562008-11-25 02:35:13 +10308029 goto free_rto_mask;
Rusty Russellc6c49272008-11-25 02:35:05 +10308030 return 0;
8031
Rusty Russell68e74562008-11-25 02:35:13 +10308032free_rto_mask:
8033 free_cpumask_var(rd->rto_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10308034free_online:
8035 free_cpumask_var(rd->online);
8036free_span:
8037 free_cpumask_var(rd->span);
Li Zefan0c910d22009-01-06 17:39:06 +08008038out:
Rusty Russellc6c49272008-11-25 02:35:05 +10308039 return -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008040}
8041
8042static void init_defrootdomain(void)
8043{
Rusty Russellc6c49272008-11-25 02:35:05 +10308044 init_rootdomain(&def_root_domain, true);
8045
Gregory Haskins57d885f2008-01-25 21:08:18 +01008046 atomic_set(&def_root_domain.refcount, 1);
8047}
8048
Gregory Haskinsdc938522008-01-25 21:08:26 +01008049static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008050{
8051 struct root_domain *rd;
8052
8053 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8054 if (!rd)
8055 return NULL;
8056
Rusty Russellc6c49272008-11-25 02:35:05 +10308057 if (init_rootdomain(rd, false) != 0) {
8058 kfree(rd);
8059 return NULL;
8060 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01008061
8062 return rd;
8063}
8064
Linus Torvalds1da177e2005-04-16 15:20:36 -07008065/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01008066 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07008067 * hold the hotplug lock.
8068 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01008069static void
8070cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008071{
Ingo Molnar70b97a72006-07-03 00:25:42 -07008072 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07008073 struct sched_domain *tmp;
8074
8075 /* Remove the sched domains which do not contribute to scheduling. */
Li Zefanf29c9b12008-11-06 09:45:16 +08008076 for (tmp = sd; tmp; ) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008077 struct sched_domain *parent = tmp->parent;
8078 if (!parent)
8079 break;
Li Zefanf29c9b12008-11-06 09:45:16 +08008080
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008081 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008082 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008083 if (parent->parent)
8084 parent->parent->child = tmp;
Li Zefanf29c9b12008-11-06 09:45:16 +08008085 } else
8086 tmp = tmp->parent;
Suresh Siddha245af2c2005-06-25 14:57:25 -07008087 }
8088
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008089 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008090 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008091 if (sd)
8092 sd->child = NULL;
8093 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008094
8095 sched_domain_debug(sd, cpu);
8096
Gregory Haskins57d885f2008-01-25 21:08:18 +01008097 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07008098 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008099}
8100
8101/* cpus with isolated domains */
Rusty Russelldcc30a32008-11-25 02:35:12 +10308102static cpumask_var_t cpu_isolated_map;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008103
8104/* Setup the mask of cpus configured for isolated domains */
8105static int __init isolated_cpu_setup(char *str)
8106{
Rusty Russellbdddd292009-12-02 14:09:16 +10308107 alloc_bootmem_cpumask_var(&cpu_isolated_map);
Rusty Russell968ea6d2008-12-13 21:55:51 +10308108 cpulist_parse(str, cpu_isolated_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008109 return 1;
8110}
8111
Ingo Molnar8927f492007-10-15 17:00:13 +02008112__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008113
8114/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008115 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8116 * to a function which identifies what group(along with sched group) a CPU
Rusty Russell96f874e22008-11-25 02:35:14 +10308117 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8118 * (due to the fact that we keep track of groups covered with a struct cpumask).
Linus Torvalds1da177e2005-04-16 15:20:36 -07008119 *
8120 * init_sched_build_groups will build a circular linked list of the groups
8121 * covered by the given span, and will set each group's ->cpumask correctly,
8122 * and ->cpu_power to 0.
8123 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008124static void
Rusty Russell96f874e22008-11-25 02:35:14 +10308125init_sched_build_groups(const struct cpumask *span,
8126 const struct cpumask *cpu_map,
8127 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07008128 struct sched_group **sg,
Rusty Russell96f874e22008-11-25 02:35:14 +10308129 struct cpumask *tmpmask),
8130 struct cpumask *covered, struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008131{
8132 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008133 int i;
8134
Rusty Russell96f874e22008-11-25 02:35:14 +10308135 cpumask_clear(covered);
Mike Travis7c16ec52008-04-04 18:11:11 -07008136
Rusty Russellabcd0832008-11-25 02:35:02 +10308137 for_each_cpu(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008138 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07008139 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008140 int j;
8141
Rusty Russell758b2cd2008-11-25 02:35:04 +10308142 if (cpumask_test_cpu(i, covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008143 continue;
8144
Rusty Russell758b2cd2008-11-25 02:35:04 +10308145 cpumask_clear(sched_group_cpus(sg));
Peter Zijlstra18a38852009-09-01 10:34:39 +02008146 sg->cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008147
Rusty Russellabcd0832008-11-25 02:35:02 +10308148 for_each_cpu(j, span) {
Mike Travis7c16ec52008-04-04 18:11:11 -07008149 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008150 continue;
8151
Rusty Russell96f874e22008-11-25 02:35:14 +10308152 cpumask_set_cpu(j, covered);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308153 cpumask_set_cpu(j, sched_group_cpus(sg));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008154 }
8155 if (!first)
8156 first = sg;
8157 if (last)
8158 last->next = sg;
8159 last = sg;
8160 }
8161 last->next = first;
8162}
8163
John Hawkes9c1cfda2005-09-06 15:18:14 -07008164#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07008165
John Hawkes9c1cfda2005-09-06 15:18:14 -07008166#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08008167
John Hawkes9c1cfda2005-09-06 15:18:14 -07008168/**
8169 * find_next_best_node - find the next node to include in a sched_domain
8170 * @node: node whose sched_domain we're building
8171 * @used_nodes: nodes already in the sched_domain
8172 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008173 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07008174 * finds the closest node not already in the @used_nodes map.
8175 *
8176 * Should use nodemask_t.
8177 */
Mike Travisc5f59f02008-04-04 18:11:10 -07008178static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008179{
8180 int i, n, val, min_val, best_node = 0;
8181
8182 min_val = INT_MAX;
8183
Mike Travis076ac2a2008-05-12 21:21:12 +02008184 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008185 /* Start at @node */
Mike Travis076ac2a2008-05-12 21:21:12 +02008186 n = (node + i) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008187
8188 if (!nr_cpus_node(n))
8189 continue;
8190
8191 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07008192 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008193 continue;
8194
8195 /* Simple min distance search */
8196 val = node_distance(node, n);
8197
8198 if (val < min_val) {
8199 min_val = val;
8200 best_node = n;
8201 }
8202 }
8203
Mike Travisc5f59f02008-04-04 18:11:10 -07008204 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008205 return best_node;
8206}
8207
8208/**
8209 * sched_domain_node_span - get a cpumask for a node's sched_domain
8210 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07008211 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07008212 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008213 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07008214 * should be one that prevents unnecessary balancing, but also spreads tasks
8215 * out optimally.
8216 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308217static void sched_domain_node_span(int node, struct cpumask *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008218{
Mike Travisc5f59f02008-04-04 18:11:10 -07008219 nodemask_t used_nodes;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008220 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008221
Mike Travis6ca09df2008-12-31 18:08:45 -08008222 cpumask_clear(span);
Mike Travisc5f59f02008-04-04 18:11:10 -07008223 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008224
Mike Travis6ca09df2008-12-31 18:08:45 -08008225 cpumask_or(span, span, cpumask_of_node(node));
Mike Travisc5f59f02008-04-04 18:11:10 -07008226 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008227
8228 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07008229 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008230
Mike Travis6ca09df2008-12-31 18:08:45 -08008231 cpumask_or(span, span, cpumask_of_node(next_node));
John Hawkes9c1cfda2005-09-06 15:18:14 -07008232 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008233}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008234#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07008235
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008236int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008237
John Hawkes9c1cfda2005-09-06 15:18:14 -07008238/*
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308239 * The cpus mask in sched_group and sched_domain hangs off the end.
Ingo Molnar4200efd2009-05-19 09:22:19 +02008240 *
8241 * ( See the the comments in include/linux/sched.h:struct sched_group
8242 * and struct sched_domain. )
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308243 */
8244struct static_sched_group {
8245 struct sched_group sg;
8246 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8247};
8248
8249struct static_sched_domain {
8250 struct sched_domain sd;
8251 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8252};
8253
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008254struct s_data {
8255#ifdef CONFIG_NUMA
8256 int sd_allnodes;
8257 cpumask_var_t domainspan;
8258 cpumask_var_t covered;
8259 cpumask_var_t notcovered;
8260#endif
8261 cpumask_var_t nodemask;
8262 cpumask_var_t this_sibling_map;
8263 cpumask_var_t this_core_map;
8264 cpumask_var_t send_covered;
8265 cpumask_var_t tmpmask;
8266 struct sched_group **sched_group_nodes;
8267 struct root_domain *rd;
8268};
8269
Andreas Herrmann2109b992009-08-18 12:53:00 +02008270enum s_alloc {
8271 sa_sched_groups = 0,
8272 sa_rootdomain,
8273 sa_tmpmask,
8274 sa_send_covered,
8275 sa_this_core_map,
8276 sa_this_sibling_map,
8277 sa_nodemask,
8278 sa_sched_group_nodes,
8279#ifdef CONFIG_NUMA
8280 sa_notcovered,
8281 sa_covered,
8282 sa_domainspan,
8283#endif
8284 sa_none,
8285};
8286
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308287/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07008288 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07008289 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008290#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308291static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
Tejun Heo1871e522009-10-29 22:34:13 +09008292static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008293
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008294static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308295cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8296 struct sched_group **sg, struct cpumask *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008297{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008298 if (sg)
Tejun Heo1871e522009-10-29 22:34:13 +09008299 *sg = &per_cpu(sched_groups, cpu).sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008300 return cpu;
8301}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008302#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008303
Ingo Molnar48f24c42006-07-03 00:25:40 -07008304/*
8305 * multi-core sched-domains:
8306 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008307#ifdef CONFIG_SCHED_MC
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308308static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8309static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008310#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008311
8312#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008313static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308314cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8315 struct sched_group **sg, struct cpumask *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008316{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008317 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07008318
Rusty Russellc69fc562009-03-13 14:49:46 +10308319 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308320 group = cpumask_first(mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008321 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308322 *sg = &per_cpu(sched_group_core, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008323 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008324}
8325#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008326static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308327cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8328 struct sched_group **sg, struct cpumask *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008329{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008330 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308331 *sg = &per_cpu(sched_group_core, cpu).sg;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008332 return cpu;
8333}
8334#endif
8335
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308336static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8337static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008338
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008339static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308340cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8341 struct sched_group **sg, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008342{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008343 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008344#ifdef CONFIG_SCHED_MC
Mike Travis6ca09df2008-12-31 18:08:45 -08008345 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308346 group = cpumask_first(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008347#elif defined(CONFIG_SCHED_SMT)
Rusty Russellc69fc562009-03-13 14:49:46 +10308348 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308349 group = cpumask_first(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008350#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008351 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008352#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008353 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308354 *sg = &per_cpu(sched_group_phys, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008355 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008356}
8357
8358#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07008359/*
8360 * The init_sched_build_groups can't handle what we want to do with node
8361 * groups, so roll our own. Now each node has its own list of groups which
8362 * gets dynamically allocated.
8363 */
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008364static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07008365static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008366
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008367static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308368static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008369
Rusty Russell96f874e22008-11-25 02:35:14 +10308370static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8371 struct sched_group **sg,
8372 struct cpumask *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008373{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008374 int group;
8375
Mike Travis6ca09df2008-12-31 18:08:45 -08008376 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308377 group = cpumask_first(nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008378
8379 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308380 *sg = &per_cpu(sched_group_allnodes, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008381 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008382}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008383
Siddha, Suresh B08069032006-03-27 01:15:23 -08008384static void init_numa_sched_groups_power(struct sched_group *group_head)
8385{
8386 struct sched_group *sg = group_head;
8387 int j;
8388
8389 if (!sg)
8390 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008391 do {
Rusty Russell758b2cd2008-11-25 02:35:04 +10308392 for_each_cpu(j, sched_group_cpus(sg)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008393 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008394
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308395 sd = &per_cpu(phys_domains, j).sd;
Miao Xie13318a72009-04-15 09:59:10 +08008396 if (j != group_first_cpu(sd->groups)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008397 /*
8398 * Only add "power" once for each
8399 * physical package.
8400 */
8401 continue;
8402 }
8403
Peter Zijlstra18a38852009-09-01 10:34:39 +02008404 sg->cpu_power += sd->groups->cpu_power;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008405 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02008406 sg = sg->next;
8407 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08008408}
Andreas Herrmann0601a882009-08-18 13:01:11 +02008409
8410static int build_numa_sched_groups(struct s_data *d,
8411 const struct cpumask *cpu_map, int num)
8412{
8413 struct sched_domain *sd;
8414 struct sched_group *sg, *prev;
8415 int n, j;
8416
8417 cpumask_clear(d->covered);
8418 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8419 if (cpumask_empty(d->nodemask)) {
8420 d->sched_group_nodes[num] = NULL;
8421 goto out;
8422 }
8423
8424 sched_domain_node_span(num, d->domainspan);
8425 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8426
8427 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8428 GFP_KERNEL, num);
8429 if (!sg) {
8430 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8431 num);
8432 return -ENOMEM;
8433 }
8434 d->sched_group_nodes[num] = sg;
8435
8436 for_each_cpu(j, d->nodemask) {
8437 sd = &per_cpu(node_domains, j).sd;
8438 sd->groups = sg;
8439 }
8440
Peter Zijlstra18a38852009-09-01 10:34:39 +02008441 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008442 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8443 sg->next = sg;
8444 cpumask_or(d->covered, d->covered, d->nodemask);
8445
8446 prev = sg;
8447 for (j = 0; j < nr_node_ids; j++) {
8448 n = (num + j) % nr_node_ids;
8449 cpumask_complement(d->notcovered, d->covered);
8450 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8451 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8452 if (cpumask_empty(d->tmpmask))
8453 break;
8454 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8455 if (cpumask_empty(d->tmpmask))
8456 continue;
8457 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8458 GFP_KERNEL, num);
8459 if (!sg) {
8460 printk(KERN_WARNING
8461 "Can not alloc domain group for node %d\n", j);
8462 return -ENOMEM;
8463 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008464 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008465 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8466 sg->next = prev->next;
8467 cpumask_or(d->covered, d->covered, d->tmpmask);
8468 prev->next = sg;
8469 prev = sg;
8470 }
8471out:
8472 return 0;
8473}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008474#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008475
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008476#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008477/* Free memory allocated for various sched_group structures */
Rusty Russell96f874e22008-11-25 02:35:14 +10308478static void free_sched_groups(const struct cpumask *cpu_map,
8479 struct cpumask *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008480{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008481 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008482
Rusty Russellabcd0832008-11-25 02:35:02 +10308483 for_each_cpu(cpu, cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008484 struct sched_group **sched_group_nodes
8485 = sched_group_nodes_bycpu[cpu];
8486
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008487 if (!sched_group_nodes)
8488 continue;
8489
Mike Travis076ac2a2008-05-12 21:21:12 +02008490 for (i = 0; i < nr_node_ids; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008491 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8492
Mike Travis6ca09df2008-12-31 18:08:45 -08008493 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308494 if (cpumask_empty(nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008495 continue;
8496
8497 if (sg == NULL)
8498 continue;
8499 sg = sg->next;
8500next_sg:
8501 oldsg = sg;
8502 sg = sg->next;
8503 kfree(oldsg);
8504 if (oldsg != sched_group_nodes[i])
8505 goto next_sg;
8506 }
8507 kfree(sched_group_nodes);
8508 sched_group_nodes_bycpu[cpu] = NULL;
8509 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008510}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008511#else /* !CONFIG_NUMA */
Rusty Russell96f874e22008-11-25 02:35:14 +10308512static void free_sched_groups(const struct cpumask *cpu_map,
8513 struct cpumask *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008514{
8515}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008516#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008517
Linus Torvalds1da177e2005-04-16 15:20:36 -07008518/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008519 * Initialize sched groups cpu_power.
8520 *
8521 * cpu_power indicates the capacity of sched group, which is used while
8522 * distributing the load between different sched groups in a sched domain.
8523 * Typically cpu_power for all the groups in a sched domain will be same unless
8524 * there are asymmetries in the topology. If there are asymmetries, group
8525 * having more cpu_power will pickup more load compared to the group having
8526 * less cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008527 */
8528static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8529{
8530 struct sched_domain *child;
8531 struct sched_group *group;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008532 long power;
8533 int weight;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008534
8535 WARN_ON(!sd || !sd->groups);
8536
Miao Xie13318a72009-04-15 09:59:10 +08008537 if (cpu != group_first_cpu(sd->groups))
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008538 return;
8539
8540 child = sd->child;
8541
Peter Zijlstra18a38852009-09-01 10:34:39 +02008542 sd->groups->cpu_power = 0;
Eric Dumazet5517d862007-05-08 00:32:57 -07008543
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008544 if (!child) {
8545 power = SCHED_LOAD_SCALE;
8546 weight = cpumask_weight(sched_domain_span(sd));
8547 /*
8548 * SMT siblings share the power of a single core.
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008549 * Usually multiple threads get a better yield out of
8550 * that one core than a single thread would have,
8551 * reflect that in sd->smt_gain.
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008552 */
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008553 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8554 power *= sd->smt_gain;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008555 power /= weight;
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008556 power >>= SCHED_LOAD_SHIFT;
8557 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008558 sd->groups->cpu_power += power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008559 return;
8560 }
8561
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008562 /*
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008563 * Add cpu_power of each child group to this groups cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008564 */
8565 group = child->groups;
8566 do {
Peter Zijlstra18a38852009-09-01 10:34:39 +02008567 sd->groups->cpu_power += group->cpu_power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008568 group = group->next;
8569 } while (group != child->groups);
8570}
8571
8572/*
Mike Travis7c16ec52008-04-04 18:11:11 -07008573 * Initializers for schedule domains
8574 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8575 */
8576
Ingo Molnara5d8c342008-10-09 11:35:51 +02008577#ifdef CONFIG_SCHED_DEBUG
8578# define SD_INIT_NAME(sd, type) sd->name = #type
8579#else
8580# define SD_INIT_NAME(sd, type) do { } while (0)
8581#endif
8582
Mike Travis7c16ec52008-04-04 18:11:11 -07008583#define SD_INIT(sd, type) sd_init_##type(sd)
Ingo Molnara5d8c342008-10-09 11:35:51 +02008584
Mike Travis7c16ec52008-04-04 18:11:11 -07008585#define SD_INIT_FUNC(type) \
8586static noinline void sd_init_##type(struct sched_domain *sd) \
8587{ \
8588 memset(sd, 0, sizeof(*sd)); \
8589 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008590 sd->level = SD_LV_##type; \
Ingo Molnara5d8c342008-10-09 11:35:51 +02008591 SD_INIT_NAME(sd, type); \
Mike Travis7c16ec52008-04-04 18:11:11 -07008592}
8593
8594SD_INIT_FUNC(CPU)
8595#ifdef CONFIG_NUMA
8596 SD_INIT_FUNC(ALLNODES)
8597 SD_INIT_FUNC(NODE)
8598#endif
8599#ifdef CONFIG_SCHED_SMT
8600 SD_INIT_FUNC(SIBLING)
8601#endif
8602#ifdef CONFIG_SCHED_MC
8603 SD_INIT_FUNC(MC)
8604#endif
8605
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008606static int default_relax_domain_level = -1;
8607
8608static int __init setup_relax_domain_level(char *str)
8609{
Li Zefan30e0e172008-05-13 10:27:17 +08008610 unsigned long val;
8611
8612 val = simple_strtoul(str, NULL, 0);
8613 if (val < SD_LV_MAX)
8614 default_relax_domain_level = val;
8615
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008616 return 1;
8617}
8618__setup("relax_domain_level=", setup_relax_domain_level);
8619
8620static void set_domain_attribute(struct sched_domain *sd,
8621 struct sched_domain_attr *attr)
8622{
8623 int request;
8624
8625 if (!attr || attr->relax_domain_level < 0) {
8626 if (default_relax_domain_level < 0)
8627 return;
8628 else
8629 request = default_relax_domain_level;
8630 } else
8631 request = attr->relax_domain_level;
8632 if (request < sd->level) {
8633 /* turn off idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008634 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008635 } else {
8636 /* turn on idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008637 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008638 }
8639}
8640
Andreas Herrmann2109b992009-08-18 12:53:00 +02008641static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8642 const struct cpumask *cpu_map)
8643{
8644 switch (what) {
8645 case sa_sched_groups:
8646 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8647 d->sched_group_nodes = NULL;
8648 case sa_rootdomain:
8649 free_rootdomain(d->rd); /* fall through */
8650 case sa_tmpmask:
8651 free_cpumask_var(d->tmpmask); /* fall through */
8652 case sa_send_covered:
8653 free_cpumask_var(d->send_covered); /* fall through */
8654 case sa_this_core_map:
8655 free_cpumask_var(d->this_core_map); /* fall through */
8656 case sa_this_sibling_map:
8657 free_cpumask_var(d->this_sibling_map); /* fall through */
8658 case sa_nodemask:
8659 free_cpumask_var(d->nodemask); /* fall through */
8660 case sa_sched_group_nodes:
8661#ifdef CONFIG_NUMA
8662 kfree(d->sched_group_nodes); /* fall through */
8663 case sa_notcovered:
8664 free_cpumask_var(d->notcovered); /* fall through */
8665 case sa_covered:
8666 free_cpumask_var(d->covered); /* fall through */
8667 case sa_domainspan:
8668 free_cpumask_var(d->domainspan); /* fall through */
8669#endif
8670 case sa_none:
8671 break;
8672 }
8673}
8674
8675static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8676 const struct cpumask *cpu_map)
8677{
8678#ifdef CONFIG_NUMA
8679 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8680 return sa_none;
8681 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8682 return sa_domainspan;
8683 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8684 return sa_covered;
8685 /* Allocate the per-node list of sched groups */
8686 d->sched_group_nodes = kcalloc(nr_node_ids,
8687 sizeof(struct sched_group *), GFP_KERNEL);
8688 if (!d->sched_group_nodes) {
8689 printk(KERN_WARNING "Can not alloc sched group node list\n");
8690 return sa_notcovered;
8691 }
8692 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8693#endif
8694 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8695 return sa_sched_group_nodes;
8696 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8697 return sa_nodemask;
8698 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8699 return sa_this_sibling_map;
8700 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8701 return sa_this_core_map;
8702 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8703 return sa_send_covered;
8704 d->rd = alloc_rootdomain();
8705 if (!d->rd) {
8706 printk(KERN_WARNING "Cannot alloc root domain\n");
8707 return sa_tmpmask;
8708 }
8709 return sa_rootdomain;
8710}
8711
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008712static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8713 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8714{
8715 struct sched_domain *sd = NULL;
8716#ifdef CONFIG_NUMA
8717 struct sched_domain *parent;
8718
8719 d->sd_allnodes = 0;
8720 if (cpumask_weight(cpu_map) >
8721 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8722 sd = &per_cpu(allnodes_domains, i).sd;
8723 SD_INIT(sd, ALLNODES);
8724 set_domain_attribute(sd, attr);
8725 cpumask_copy(sched_domain_span(sd), cpu_map);
8726 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8727 d->sd_allnodes = 1;
8728 }
8729 parent = sd;
8730
8731 sd = &per_cpu(node_domains, i).sd;
8732 SD_INIT(sd, NODE);
8733 set_domain_attribute(sd, attr);
8734 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8735 sd->parent = parent;
8736 if (parent)
8737 parent->child = sd;
8738 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8739#endif
8740 return sd;
8741}
8742
Andreas Herrmann87cce662009-08-18 12:54:55 +02008743static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8744 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8745 struct sched_domain *parent, int i)
8746{
8747 struct sched_domain *sd;
8748 sd = &per_cpu(phys_domains, i).sd;
8749 SD_INIT(sd, CPU);
8750 set_domain_attribute(sd, attr);
8751 cpumask_copy(sched_domain_span(sd), d->nodemask);
8752 sd->parent = parent;
8753 if (parent)
8754 parent->child = sd;
8755 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8756 return sd;
8757}
8758
Andreas Herrmann410c4082009-08-18 12:56:14 +02008759static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8760 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8761 struct sched_domain *parent, int i)
8762{
8763 struct sched_domain *sd = parent;
8764#ifdef CONFIG_SCHED_MC
8765 sd = &per_cpu(core_domains, i).sd;
8766 SD_INIT(sd, MC);
8767 set_domain_attribute(sd, attr);
8768 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8769 sd->parent = parent;
8770 parent->child = sd;
8771 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8772#endif
8773 return sd;
8774}
8775
Andreas Herrmannd8173532009-08-18 12:57:03 +02008776static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8777 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8778 struct sched_domain *parent, int i)
8779{
8780 struct sched_domain *sd = parent;
8781#ifdef CONFIG_SCHED_SMT
8782 sd = &per_cpu(cpu_domains, i).sd;
8783 SD_INIT(sd, SIBLING);
8784 set_domain_attribute(sd, attr);
8785 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8786 sd->parent = parent;
8787 parent->child = sd;
8788 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8789#endif
8790 return sd;
8791}
8792
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008793static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8794 const struct cpumask *cpu_map, int cpu)
8795{
8796 switch (l) {
8797#ifdef CONFIG_SCHED_SMT
8798 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8799 cpumask_and(d->this_sibling_map, cpu_map,
8800 topology_thread_cpumask(cpu));
8801 if (cpu == cpumask_first(d->this_sibling_map))
8802 init_sched_build_groups(d->this_sibling_map, cpu_map,
8803 &cpu_to_cpu_group,
8804 d->send_covered, d->tmpmask);
8805 break;
8806#endif
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008807#ifdef CONFIG_SCHED_MC
8808 case SD_LV_MC: /* set up multi-core groups */
8809 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8810 if (cpu == cpumask_first(d->this_core_map))
8811 init_sched_build_groups(d->this_core_map, cpu_map,
8812 &cpu_to_core_group,
8813 d->send_covered, d->tmpmask);
8814 break;
8815#endif
Andreas Herrmann86548092009-08-18 12:59:28 +02008816 case SD_LV_CPU: /* set up physical groups */
8817 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8818 if (!cpumask_empty(d->nodemask))
8819 init_sched_build_groups(d->nodemask, cpu_map,
8820 &cpu_to_phys_group,
8821 d->send_covered, d->tmpmask);
8822 break;
Andreas Herrmannde616e32009-08-18 13:00:13 +02008823#ifdef CONFIG_NUMA
8824 case SD_LV_ALLNODES:
8825 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8826 d->send_covered, d->tmpmask);
8827 break;
8828#endif
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008829 default:
8830 break;
8831 }
8832}
8833
Mike Travis7c16ec52008-04-04 18:11:11 -07008834/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008835 * Build sched domains for a given set of cpus and attach the sched domains
8836 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07008837 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308838static int __build_sched_domains(const struct cpumask *cpu_map,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008839 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008840{
Andreas Herrmann2109b992009-08-18 12:53:00 +02008841 enum s_alloc alloc_state = sa_none;
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008842 struct s_data d;
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008843 struct sched_domain *sd;
Andreas Herrmann2109b992009-08-18 12:53:00 +02008844 int i;
John Hawkesd1b55132005-09-06 15:18:14 -07008845#ifdef CONFIG_NUMA
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008846 d.sd_allnodes = 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308847#endif
8848
Andreas Herrmann2109b992009-08-18 12:53:00 +02008849 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8850 if (alloc_state != sa_rootdomain)
8851 goto error;
8852 alloc_state = sa_sched_groups;
Mike Travis7c16ec52008-04-04 18:11:11 -07008853
Linus Torvalds1da177e2005-04-16 15:20:36 -07008854 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008855 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008856 */
Rusty Russellabcd0832008-11-25 02:35:02 +10308857 for_each_cpu(i, cpu_map) {
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008858 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8859 cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008860
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008861 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
Andreas Herrmann87cce662009-08-18 12:54:55 +02008862 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmann410c4082009-08-18 12:56:14 +02008863 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmannd8173532009-08-18 12:57:03 +02008864 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008865 }
8866
Rusty Russellabcd0832008-11-25 02:35:02 +10308867 for_each_cpu(i, cpu_map) {
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008868 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008869 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008870 }
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008871
Linus Torvalds1da177e2005-04-16 15:20:36 -07008872 /* Set up physical groups */
Andreas Herrmann86548092009-08-18 12:59:28 +02008873 for (i = 0; i < nr_node_ids; i++)
8874 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008875
8876#ifdef CONFIG_NUMA
8877 /* Set up node groups */
Andreas Herrmannde616e32009-08-18 13:00:13 +02008878 if (d.sd_allnodes)
8879 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008880
Andreas Herrmann0601a882009-08-18 13:01:11 +02008881 for (i = 0; i < nr_node_ids; i++)
8882 if (build_numa_sched_groups(&d, cpu_map, i))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008883 goto error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008884#endif
8885
8886 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008887#ifdef CONFIG_SCHED_SMT
Rusty Russellabcd0832008-11-25 02:35:02 +10308888 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008889 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008890 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008891 }
8892#endif
8893#ifdef CONFIG_SCHED_MC
Rusty Russellabcd0832008-11-25 02:35:02 +10308894 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008895 sd = &per_cpu(core_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008896 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008897 }
8898#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008899
Rusty Russellabcd0832008-11-25 02:35:02 +10308900 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008901 sd = &per_cpu(phys_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008902 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008903 }
8904
John Hawkes9c1cfda2005-09-06 15:18:14 -07008905#ifdef CONFIG_NUMA
Mike Travis076ac2a2008-05-12 21:21:12 +02008906 for (i = 0; i < nr_node_ids; i++)
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008907 init_numa_sched_groups_power(d.sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008908
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008909 if (d.sd_allnodes) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008910 struct sched_group *sg;
Siddha, Suresh Bf712c0c72006-07-30 03:02:59 -07008911
Rusty Russell96f874e22008-11-25 02:35:14 +10308912 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008913 d.tmpmask);
Siddha, Suresh Bf712c0c72006-07-30 03:02:59 -07008914 init_numa_sched_groups_power(sg);
8915 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008916#endif
8917
Linus Torvalds1da177e2005-04-16 15:20:36 -07008918 /* Attach the domains */
Rusty Russellabcd0832008-11-25 02:35:02 +10308919 for_each_cpu(i, cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008920#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308921 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008922#elif defined(CONFIG_SCHED_MC)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308923 sd = &per_cpu(core_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008924#else
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308925 sd = &per_cpu(phys_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008926#endif
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008927 cpu_attach_domain(sd, d.rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008928 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008929
Andreas Herrmann2109b992009-08-18 12:53:00 +02008930 d.sched_group_nodes = NULL; /* don't free this we still need it */
8931 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8932 return 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308933
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008934error:
Andreas Herrmann2109b992009-08-18 12:53:00 +02008935 __free_domain_allocs(&d, alloc_state, cpu_map);
8936 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008937}
Paul Jackson029190c2007-10-18 23:40:20 -07008938
Rusty Russell96f874e22008-11-25 02:35:14 +10308939static int build_sched_domains(const struct cpumask *cpu_map)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008940{
8941 return __build_sched_domains(cpu_map, NULL);
8942}
8943
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308944static cpumask_var_t *doms_cur; /* current sched domains */
Paul Jackson029190c2007-10-18 23:40:20 -07008945static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02008946static struct sched_domain_attr *dattr_cur;
8947 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07008948
8949/*
8950 * Special case: If a kmalloc of a doms_cur partition (array of
Rusty Russell42128232008-11-25 02:35:12 +10308951 * cpumask) fails, then fallback to a single sched domain,
8952 * as determined by the single cpumask fallback_doms.
Paul Jackson029190c2007-10-18 23:40:20 -07008953 */
Rusty Russell42128232008-11-25 02:35:12 +10308954static cpumask_var_t fallback_doms;
Paul Jackson029190c2007-10-18 23:40:20 -07008955
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008956/*
8957 * arch_update_cpu_topology lets virtualized architectures update the
8958 * cpu core maps. It is supposed to return 1 if the topology changed
8959 * or 0 if it stayed the same.
8960 */
8961int __attribute__((weak)) arch_update_cpu_topology(void)
Heiko Carstens22e52b02008-03-12 18:31:59 +01008962{
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008963 return 0;
Heiko Carstens22e52b02008-03-12 18:31:59 +01008964}
8965
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308966cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8967{
8968 int i;
8969 cpumask_var_t *doms;
8970
8971 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8972 if (!doms)
8973 return NULL;
8974 for (i = 0; i < ndoms; i++) {
8975 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8976 free_sched_domains(doms, i);
8977 return NULL;
8978 }
8979 }
8980 return doms;
8981}
8982
8983void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8984{
8985 unsigned int i;
8986 for (i = 0; i < ndoms; i++)
8987 free_cpumask_var(doms[i]);
8988 kfree(doms);
8989}
8990
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008991/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008992 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07008993 * For now this just excludes isolated cpus, but could be used to
8994 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008995 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308996static int arch_init_sched_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008997{
Milton Miller73785472007-10-24 18:23:48 +02008998 int err;
8999
Heiko Carstens22e52b02008-03-12 18:31:59 +01009000 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07009001 ndoms_cur = 1;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309002 doms_cur = alloc_sched_domains(ndoms_cur);
Paul Jackson029190c2007-10-18 23:40:20 -07009003 if (!doms_cur)
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309004 doms_cur = &fallback_doms;
9005 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009006 dattr_cur = NULL;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309007 err = build_sched_domains(doms_cur[0]);
Milton Miller6382bc92007-10-15 17:00:19 +02009008 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02009009
9010 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009011}
9012
Rusty Russell96f874e22008-11-25 02:35:14 +10309013static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9014 struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009015{
Mike Travis7c16ec52008-04-04 18:11:11 -07009016 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07009017}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009018
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009019/*
9020 * Detach sched domains from a group of cpus specified in cpu_map
9021 * These cpus will now be attached to the NULL domain
9022 */
Rusty Russell96f874e22008-11-25 02:35:14 +10309023static void detach_destroy_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009024{
Rusty Russell96f874e22008-11-25 02:35:14 +10309025 /* Save because hotplug lock held. */
9026 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009027 int i;
9028
Rusty Russellabcd0832008-11-25 02:35:02 +10309029 for_each_cpu(i, cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01009030 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009031 synchronize_sched();
Rusty Russell96f874e22008-11-25 02:35:14 +10309032 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009033}
9034
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009035/* handle null as "default" */
9036static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9037 struct sched_domain_attr *new, int idx_new)
9038{
9039 struct sched_domain_attr tmp;
9040
9041 /* fast path */
9042 if (!new && !cur)
9043 return 1;
9044
9045 tmp = SD_ATTR_INIT;
9046 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9047 new ? (new + idx_new) : &tmp,
9048 sizeof(struct sched_domain_attr));
9049}
9050
Paul Jackson029190c2007-10-18 23:40:20 -07009051/*
9052 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009053 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07009054 * doms_new[] to the current sched domain partitioning, doms_cur[].
9055 * It destroys each deleted domain and builds each new domain.
9056 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309057 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009058 * The masks don't intersect (don't overlap.) We should setup one
9059 * sched domain for each mask. CPUs not in any of the cpumasks will
9060 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07009061 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9062 * it as it is.
9063 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309064 * The passed in 'doms_new' should be allocated using
9065 * alloc_sched_domains. This routine takes ownership of it and will
9066 * free_sched_domains it when done with it. If the caller failed the
9067 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9068 * and partition_sched_domains() will fallback to the single partition
9069 * 'fallback_doms', it also forces the domains to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07009070 *
Rusty Russell96f874e22008-11-25 02:35:14 +10309071 * If doms_new == NULL it will be replaced with cpu_online_mask.
Li Zefan700018e2008-11-18 14:02:03 +08009072 * ndoms_new == 0 is a special case for destroying existing domains,
9073 * and it will not create the default domain.
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009074 *
Paul Jackson029190c2007-10-18 23:40:20 -07009075 * Call with hotplug lock held
9076 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309077void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009078 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07009079{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009080 int i, j, n;
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009081 int new_topology;
Paul Jackson029190c2007-10-18 23:40:20 -07009082
Heiko Carstens712555e2008-04-28 11:33:07 +02009083 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01009084
Milton Miller73785472007-10-24 18:23:48 +02009085 /* always unregister in case we don't destroy any domains */
9086 unregister_sched_domain_sysctl();
9087
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009088 /* Let architecture update cpu core mappings. */
9089 new_topology = arch_update_cpu_topology();
9090
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009091 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07009092
9093 /* Destroy deleted domains */
9094 for (i = 0; i < ndoms_cur; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009095 for (j = 0; j < n && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309096 if (cpumask_equal(doms_cur[i], doms_new[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009097 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07009098 goto match1;
9099 }
9100 /* no match - a current sched domain not in new doms_new[] */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309101 detach_destroy_domains(doms_cur[i]);
Paul Jackson029190c2007-10-18 23:40:20 -07009102match1:
9103 ;
9104 }
9105
Max Krasnyanskye761b772008-07-15 04:43:49 -07009106 if (doms_new == NULL) {
9107 ndoms_cur = 0;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309108 doms_new = &fallback_doms;
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01009109 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
Li Zefanfaa2f982008-11-04 16:20:23 +08009110 WARN_ON_ONCE(dattr_new);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009111 }
9112
Paul Jackson029190c2007-10-18 23:40:20 -07009113 /* Build new domains */
9114 for (i = 0; i < ndoms_new; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009115 for (j = 0; j < ndoms_cur && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309116 if (cpumask_equal(doms_new[i], doms_cur[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009117 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07009118 goto match2;
9119 }
9120 /* no match - add a new doms_new */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309121 __build_sched_domains(doms_new[i],
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009122 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07009123match2:
9124 ;
9125 }
9126
9127 /* Remember the new sched domains */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309128 if (doms_cur != &fallback_doms)
9129 free_sched_domains(doms_cur, ndoms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009130 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07009131 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009132 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07009133 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02009134
9135 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01009136
Heiko Carstens712555e2008-04-28 11:33:07 +02009137 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07009138}
9139
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009140#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Li Zefanc70f22d2009-01-05 19:07:50 +08009141static void arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009142{
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009143 get_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009144
9145 /* Destroy domains first to force the rebuild */
9146 partition_sched_domains(0, NULL, NULL);
9147
Max Krasnyanskye761b772008-07-15 04:43:49 -07009148 rebuild_sched_domains();
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009149 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009150}
9151
9152static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9153{
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309154 unsigned int level = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009155
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309156 if (sscanf(buf, "%u", &level) != 1)
9157 return -EINVAL;
9158
9159 /*
9160 * level is always be positive so don't check for
9161 * level < POWERSAVINGS_BALANCE_NONE which is 0
9162 * What happens on 0 or 1 byte write,
9163 * need to check for count as well?
9164 */
9165
9166 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009167 return -EINVAL;
9168
9169 if (smt)
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309170 sched_smt_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009171 else
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309172 sched_mc_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009173
Li Zefanc70f22d2009-01-05 19:07:50 +08009174 arch_reinit_sched_domains();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009175
Li Zefanc70f22d2009-01-05 19:07:50 +08009176 return count;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009177}
9178
Adrian Bunk6707de002007-08-12 18:08:19 +02009179#ifdef CONFIG_SCHED_MC
Andi Kleenf718cd42008-07-29 22:33:52 -07009180static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9181 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009182{
9183 return sprintf(page, "%u\n", sched_mc_power_savings);
9184}
Andi Kleenf718cd42008-07-29 22:33:52 -07009185static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
Adrian Bunk6707de002007-08-12 18:08:19 +02009186 const char *buf, size_t count)
9187{
9188 return sched_power_savings_store(buf, count, 0);
9189}
Andi Kleenf718cd42008-07-29 22:33:52 -07009190static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9191 sched_mc_power_savings_show,
9192 sched_mc_power_savings_store);
Adrian Bunk6707de002007-08-12 18:08:19 +02009193#endif
9194
9195#ifdef CONFIG_SCHED_SMT
Andi Kleenf718cd42008-07-29 22:33:52 -07009196static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9197 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009198{
9199 return sprintf(page, "%u\n", sched_smt_power_savings);
9200}
Andi Kleenf718cd42008-07-29 22:33:52 -07009201static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
Adrian Bunk6707de002007-08-12 18:08:19 +02009202 const char *buf, size_t count)
9203{
9204 return sched_power_savings_store(buf, count, 1);
9205}
Andi Kleenf718cd42008-07-29 22:33:52 -07009206static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9207 sched_smt_power_savings_show,
Adrian Bunk6707de002007-08-12 18:08:19 +02009208 sched_smt_power_savings_store);
9209#endif
9210
Li Zefan39aac642009-01-05 19:18:02 +08009211int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009212{
9213 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07009214
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009215#ifdef CONFIG_SCHED_SMT
9216 if (smt_capable())
9217 err = sysfs_create_file(&cls->kset.kobj,
9218 &attr_sched_smt_power_savings.attr);
9219#endif
9220#ifdef CONFIG_SCHED_MC
9221 if (!err && mc_capable())
9222 err = sysfs_create_file(&cls->kset.kobj,
9223 &attr_sched_mc_power_savings.attr);
9224#endif
9225 return err;
9226}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009227#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009228
Max Krasnyanskye761b772008-07-15 04:43:49 -07009229#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009230/*
Max Krasnyanskye761b772008-07-15 04:43:49 -07009231 * Add online and remove offline CPUs from the scheduler domains.
9232 * When cpusets are enabled they take over this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009233 */
9234static int update_sched_domains(struct notifier_block *nfb,
9235 unsigned long action, void *hcpu)
9236{
Max Krasnyanskye761b772008-07-15 04:43:49 -07009237 switch (action) {
9238 case CPU_ONLINE:
9239 case CPU_ONLINE_FROZEN:
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01009240 case CPU_DOWN_PREPARE:
9241 case CPU_DOWN_PREPARE_FROZEN:
9242 case CPU_DOWN_FAILED:
9243 case CPU_DOWN_FAILED_FROZEN:
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009244 partition_sched_domains(1, NULL, NULL);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009245 return NOTIFY_OK;
9246
9247 default:
9248 return NOTIFY_DONE;
9249 }
9250}
9251#endif
9252
9253static int update_runtime(struct notifier_block *nfb,
9254 unsigned long action, void *hcpu)
9255{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009256 int cpu = (int)(long)hcpu;
9257
Linus Torvalds1da177e2005-04-16 15:20:36 -07009258 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07009259 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009260 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009261 disable_runtime(cpu_rq(cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07009262 return NOTIFY_OK;
9263
Linus Torvalds1da177e2005-04-16 15:20:36 -07009264 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009265 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07009266 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009267 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009268 enable_runtime(cpu_rq(cpu));
Max Krasnyanskye761b772008-07-15 04:43:49 -07009269 return NOTIFY_OK;
9270
Linus Torvalds1da177e2005-04-16 15:20:36 -07009271 default:
9272 return NOTIFY_DONE;
9273 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009274}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009275
9276void __init sched_init_smp(void)
9277{
Rusty Russelldcc30a32008-11-25 02:35:12 +10309278 cpumask_var_t non_isolated_cpus;
9279
9280 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
Yong Zhangcb5fd132009-09-14 20:20:16 +08009281 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
Nick Piggin5c1e1762006-10-03 01:14:04 -07009282
Mike Travis434d53b2008-04-04 18:11:04 -07009283#if defined(CONFIG_NUMA)
9284 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9285 GFP_KERNEL);
9286 BUG_ON(sched_group_nodes_bycpu == NULL);
9287#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009288 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02009289 mutex_lock(&sched_domains_mutex);
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01009290 arch_init_sched_domains(cpu_active_mask);
Rusty Russelldcc30a32008-11-25 02:35:12 +10309291 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9292 if (cpumask_empty(non_isolated_cpus))
9293 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02009294 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009295 put_online_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009296
9297#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009298 /* XXX: Theoretical race here - CPU may be hotplugged now */
9299 hotcpu_notifier(update_sched_domains, 0);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009300#endif
9301
9302 /* RT runtime code needs to handle some hotplug events */
9303 hotcpu_notifier(update_runtime, 0);
9304
Peter Zijlstrab328ca12008-04-29 10:02:46 +02009305 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07009306
9307 /* Move init over to a non-isolated CPU */
Rusty Russelldcc30a32008-11-25 02:35:12 +10309308 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07009309 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01009310 sched_init_granularity();
Rusty Russelldcc30a32008-11-25 02:35:12 +10309311 free_cpumask_var(non_isolated_cpus);
Rusty Russell42128232008-11-25 02:35:12 +10309312
Rusty Russell0e3900e2008-11-25 02:35:13 +10309313 init_sched_rt_class();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009314}
9315#else
9316void __init sched_init_smp(void)
9317{
Ingo Molnar19978ca2007-11-09 22:39:38 +01009318 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009319}
9320#endif /* CONFIG_SMP */
9321
Arun R Bharadwajcd1bb942009-04-16 12:15:34 +05309322const_debug unsigned int sysctl_timer_migration = 1;
9323
Linus Torvalds1da177e2005-04-16 15:20:36 -07009324int in_sched_functions(unsigned long addr)
9325{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009326 return in_lock_functions(addr) ||
9327 (addr >= (unsigned long)__sched_text_start
9328 && addr < (unsigned long)__sched_text_end);
9329}
9330
Alexey Dobriyana9957442007-10-15 17:00:13 +02009331static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02009332{
9333 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02009334 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02009335#ifdef CONFIG_FAIR_GROUP_SCHED
9336 cfs_rq->rq = rq;
9337#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02009338 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02009339}
9340
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009341static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9342{
9343 struct rt_prio_array *array;
9344 int i;
9345
9346 array = &rt_rq->active;
9347 for (i = 0; i < MAX_RT_PRIO; i++) {
9348 INIT_LIST_HEAD(array->queue + i);
9349 __clear_bit(i, array->bitmap);
9350 }
9351 /* delimiter for bitsearch: */
9352 __set_bit(MAX_RT_PRIO, array->bitmap);
9353
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009354#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -05009355 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins398a1532009-01-14 09:10:04 -05009356#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -05009357 rt_rq->highest_prio.next = MAX_RT_PRIO;
Peter Zijlstra48d5e252008-01-25 21:08:31 +01009358#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009359#endif
9360#ifdef CONFIG_SMP
9361 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009362 rt_rq->overloaded = 0;
Thomas Gleixner05fa7852009-11-17 14:28:38 +01009363 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009364#endif
9365
9366 rt_rq->rt_time = 0;
9367 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009368 rt_rq->rt_runtime = 0;
Thomas Gleixner0986b112009-11-17 15:32:06 +01009369 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009370
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009371#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01009372 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009373 rt_rq->rq = rq;
9374#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009375}
9376
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009377#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009378static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9379 struct sched_entity *se, int cpu, int add,
9380 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009381{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009382 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009383 tg->cfs_rq[cpu] = cfs_rq;
9384 init_cfs_rq(cfs_rq, rq);
9385 cfs_rq->tg = tg;
9386 if (add)
9387 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9388
9389 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009390 /* se could be NULL for init_task_group */
9391 if (!se)
9392 return;
9393
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009394 if (!parent)
9395 se->cfs_rq = &rq->cfs;
9396 else
9397 se->cfs_rq = parent->my_q;
9398
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009399 se->my_q = cfs_rq;
9400 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009401 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009402 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009403}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009404#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009405
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009406#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009407static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9408 struct sched_rt_entity *rt_se, int cpu, int add,
9409 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009410{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009411 struct rq *rq = cpu_rq(cpu);
9412
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009413 tg->rt_rq[cpu] = rt_rq;
9414 init_rt_rq(rt_rq, rq);
9415 rt_rq->tg = tg;
9416 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009417 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009418 if (add)
9419 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9420
9421 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009422 if (!rt_se)
9423 return;
9424
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009425 if (!parent)
9426 rt_se->rt_rq = &rq->rt;
9427 else
9428 rt_se->rt_rq = parent->my_q;
9429
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009430 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009431 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009432 INIT_LIST_HEAD(&rt_se->run_list);
9433}
9434#endif
9435
Linus Torvalds1da177e2005-04-16 15:20:36 -07009436void __init sched_init(void)
9437{
Ingo Molnardd41f592007-07-09 18:51:59 +02009438 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07009439 unsigned long alloc_size = 0, ptr;
9440
9441#ifdef CONFIG_FAIR_GROUP_SCHED
9442 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9443#endif
9444#ifdef CONFIG_RT_GROUP_SCHED
9445 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9446#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009447#ifdef CONFIG_USER_SCHED
9448 alloc_size *= 2;
9449#endif
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309450#ifdef CONFIG_CPUMASK_OFFSTACK
Rusty Russell8c083f02009-03-19 15:22:20 +10309451 alloc_size += num_possible_cpus() * cpumask_size();
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309452#endif
Mike Travis434d53b2008-04-04 18:11:04 -07009453 if (alloc_size) {
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03009454 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07009455
9456#ifdef CONFIG_FAIR_GROUP_SCHED
9457 init_task_group.se = (struct sched_entity **)ptr;
9458 ptr += nr_cpu_ids * sizeof(void **);
9459
9460 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9461 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009462
9463#ifdef CONFIG_USER_SCHED
9464 root_task_group.se = (struct sched_entity **)ptr;
9465 ptr += nr_cpu_ids * sizeof(void **);
9466
9467 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9468 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009469#endif /* CONFIG_USER_SCHED */
9470#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009471#ifdef CONFIG_RT_GROUP_SCHED
9472 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9473 ptr += nr_cpu_ids * sizeof(void **);
9474
9475 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009476 ptr += nr_cpu_ids * sizeof(void **);
9477
9478#ifdef CONFIG_USER_SCHED
9479 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9480 ptr += nr_cpu_ids * sizeof(void **);
9481
9482 root_task_group.rt_rq = (struct rt_rq **)ptr;
9483 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009484#endif /* CONFIG_USER_SCHED */
9485#endif /* CONFIG_RT_GROUP_SCHED */
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309486#ifdef CONFIG_CPUMASK_OFFSTACK
9487 for_each_possible_cpu(i) {
9488 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9489 ptr += cpumask_size();
9490 }
9491#endif /* CONFIG_CPUMASK_OFFSTACK */
Mike Travis434d53b2008-04-04 18:11:04 -07009492 }
Ingo Molnardd41f592007-07-09 18:51:59 +02009493
Gregory Haskins57d885f2008-01-25 21:08:18 +01009494#ifdef CONFIG_SMP
9495 init_defrootdomain();
9496#endif
9497
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009498 init_rt_bandwidth(&def_rt_bandwidth,
9499 global_rt_period(), global_rt_runtime());
9500
9501#ifdef CONFIG_RT_GROUP_SCHED
9502 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9503 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009504#ifdef CONFIG_USER_SCHED
9505 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9506 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009507#endif /* CONFIG_USER_SCHED */
9508#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009509
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009510#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009511 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009512 INIT_LIST_HEAD(&init_task_group.children);
9513
9514#ifdef CONFIG_USER_SCHED
9515 INIT_LIST_HEAD(&root_task_group.children);
9516 init_task_group.parent = &root_task_group;
9517 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009518#endif /* CONFIG_USER_SCHED */
9519#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009520
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09009521#if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9522 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9523 __alignof__(unsigned long));
9524#endif
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08009525 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07009526 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009527
9528 rq = cpu_rq(i);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01009529 raw_spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07009530 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009531 rq->calc_load_active = 0;
9532 rq->calc_load_update = jiffies + LOAD_FREQ;
Ingo Molnardd41f592007-07-09 18:51:59 +02009533 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009534 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009535#ifdef CONFIG_FAIR_GROUP_SCHED
9536 init_task_group.shares = init_task_group_load;
9537 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009538#ifdef CONFIG_CGROUP_SCHED
9539 /*
9540 * How much cpu bandwidth does init_task_group get?
9541 *
9542 * In case of task-groups formed thr' the cgroup filesystem, it
9543 * gets 100% of the cpu resources in the system. This overall
9544 * system cpu resource is divided among the tasks of
9545 * init_task_group and its child task-groups in a fair manner,
9546 * based on each entity's (task or task-group's) weight
9547 * (se->load.weight).
9548 *
9549 * In other words, if init_task_group has 10 tasks of weight
9550 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9551 * then A0's share of the cpu resource is:
9552 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009553 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02009554 *
9555 * We achieve this by letting init_task_group's tasks sit
9556 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9557 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009558 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009559#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009560 root_task_group.shares = NICE_0_LOAD;
9561 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009562 /*
9563 * In case of task-groups formed thr' the user id of tasks,
9564 * init_task_group represents tasks belonging to root user.
9565 * Hence it forms a sibling of all subsequent groups formed.
9566 * In this case, init_task_group gets only a fraction of overall
9567 * system cpu resource, based on the weight assigned to root
9568 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9569 * by letting tasks of init_task_group sit in a separate cfs_rq
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009570 * (init_tg_cfs_rq) and having one entity represent this group of
Dhaval Giani354d60c2008-04-19 19:44:59 +02009571 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9572 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009573 init_tg_cfs_entry(&init_task_group,
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009574 &per_cpu(init_tg_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009575 &per_cpu(init_sched_entity, i), i, 1,
9576 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009577
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009578#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02009579#endif /* CONFIG_FAIR_GROUP_SCHED */
9580
9581 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009582#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009583 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009584#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009585 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009586#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009587 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009588 init_tg_rt_entry(&init_task_group,
Tejun Heo1871e522009-10-29 22:34:13 +09009589 &per_cpu(init_rt_rq_var, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009590 &per_cpu(init_sched_rt_entity, i), i, 1,
9591 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009592#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009593#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009594
Ingo Molnardd41f592007-07-09 18:51:59 +02009595 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9596 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009597#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07009598 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01009599 rq->rd = NULL;
Gregory Haskins3f029d32009-07-29 11:08:47 -04009600 rq->post_schedule = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009601 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009602 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009603 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07009604 rq->cpu = i;
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009605 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009606 rq->migration_thread = NULL;
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01009607 rq->idle_stamp = 0;
9608 rq->avg_idle = 2*sysctl_sched_migration_cost;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009609 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01009610 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009611#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009612 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009613 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009614 }
9615
Peter Williams2dd73a42006-06-27 02:54:34 -07009616 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009617
Avi Kivitye107be32007-07-26 13:40:43 +02009618#ifdef CONFIG_PREEMPT_NOTIFIERS
9619 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9620#endif
9621
Christoph Lameterc9819f42006-12-10 02:20:25 -08009622#ifdef CONFIG_SMP
Carlos R. Mafra962cf362008-05-15 11:15:37 -03009623 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
Christoph Lameterc9819f42006-12-10 02:20:25 -08009624#endif
9625
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009626#ifdef CONFIG_RT_MUTEXES
Thomas Gleixner1d615482009-11-17 14:54:03 +01009627 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009628#endif
9629
Linus Torvalds1da177e2005-04-16 15:20:36 -07009630 /*
9631 * The boot idle thread does lazy MMU switching as well:
9632 */
9633 atomic_inc(&init_mm.mm_count);
9634 enter_lazy_tlb(&init_mm, current);
9635
9636 /*
9637 * Make us the idle thread. Technically, schedule() should not be
9638 * called from this thread, however somewhere below it might be,
9639 * but because we are the idle thread, we just pick up running again
9640 * when this runqueue becomes "idle".
9641 */
9642 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009643
9644 calc_load_update = jiffies + LOAD_FREQ;
9645
Ingo Molnardd41f592007-07-09 18:51:59 +02009646 /*
9647 * During early bootup we pretend to be a normal task:
9648 */
9649 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01009650
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309651 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
Rusty Russell49557e62009-11-02 20:37:20 +10309652 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309653#ifdef CONFIG_SMP
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309654#ifdef CONFIG_NO_HZ
Rusty Russell49557e62009-11-02 20:37:20 +10309655 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009656 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309657#endif
Rusty Russellbdddd292009-12-02 14:09:16 +10309658 /* May be allocated at isolcpus cmdline parse time */
9659 if (cpu_isolated_map == NULL)
9660 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309661#endif /* SMP */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309662
Ingo Molnarcdd6c482009-09-21 12:02:48 +02009663 perf_event_init();
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009664
Ingo Molnar6892b752008-02-13 14:02:36 +01009665 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009666}
9667
9668#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009669static inline int preempt_count_equals(int preempt_offset)
9670{
9671 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9672
9673 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9674}
9675
9676void __might_sleep(char *file, int line, int preempt_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009677{
Ingo Molnar48f24c42006-07-03 00:25:40 -07009678#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07009679 static unsigned long prev_jiffy; /* ratelimiting */
9680
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009681 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9682 system_state != SYSTEM_RUNNING || oops_in_progress)
Ingo Molnaraef745f2008-08-28 11:34:43 +02009683 return;
9684 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9685 return;
9686 prev_jiffy = jiffies;
9687
9688 printk(KERN_ERR
9689 "BUG: sleeping function called from invalid context at %s:%d\n",
9690 file, line);
9691 printk(KERN_ERR
9692 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9693 in_atomic(), irqs_disabled(),
9694 current->pid, current->comm);
9695
9696 debug_show_held_locks(current);
9697 if (irqs_disabled())
9698 print_irqtrace_events(current);
9699 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009700#endif
9701}
9702EXPORT_SYMBOL(__might_sleep);
9703#endif
9704
9705#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009706static void normalize_task(struct rq *rq, struct task_struct *p)
9707{
9708 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02009709
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009710 update_rq_clock(rq);
9711 on_rq = p->se.on_rq;
9712 if (on_rq)
9713 deactivate_task(rq, p, 0);
9714 __setscheduler(rq, p, SCHED_NORMAL, 0);
9715 if (on_rq) {
9716 activate_task(rq, p, 0);
9717 resched_task(rq->curr);
9718 }
9719}
9720
Linus Torvalds1da177e2005-04-16 15:20:36 -07009721void normalize_rt_tasks(void)
9722{
Ingo Molnara0f98a12007-06-17 18:37:45 +02009723 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009724 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07009725 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009726
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009727 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009728 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02009729 /*
9730 * Only normalize user tasks:
9731 */
9732 if (!p->mm)
9733 continue;
9734
Ingo Molnardd41f592007-07-09 18:51:59 +02009735 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009736#ifdef CONFIG_SCHEDSTATS
9737 p->se.wait_start = 0;
9738 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009739 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009740#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02009741
9742 if (!rt_task(p)) {
9743 /*
9744 * Renice negative nice level userspace
9745 * tasks back to 0:
9746 */
9747 if (TASK_NICE(p) < 0 && p->mm)
9748 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009749 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02009750 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009751
Thomas Gleixner1d615482009-11-17 14:54:03 +01009752 raw_spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07009753 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009754
Ingo Molnar178be792007-10-15 17:00:18 +02009755 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009756
Ingo Molnarb29739f2006-06-27 02:54:51 -07009757 __task_rq_unlock(rq);
Thomas Gleixner1d615482009-11-17 14:54:03 +01009758 raw_spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009759 } while_each_thread(g, p);
9760
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009761 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009762}
9763
9764#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07009765
9766#ifdef CONFIG_IA64
9767/*
9768 * These functions are only useful for the IA64 MCA handling.
9769 *
9770 * They can only be called when the whole system has been
9771 * stopped - every CPU needs to be quiescent, and no scheduling
9772 * activity can take place. Using them for anything else would
9773 * be a serious bug, and as a result, they aren't even visible
9774 * under any other configuration.
9775 */
9776
9777/**
9778 * curr_task - return the current task for a given cpu.
9779 * @cpu: the processor in question.
9780 *
9781 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9782 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009783struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009784{
9785 return cpu_curr(cpu);
9786}
9787
9788/**
9789 * set_curr_task - set the current task for a given cpu.
9790 * @cpu: the processor in question.
9791 * @p: the task pointer to set.
9792 *
9793 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009794 * are serviced on a separate stack. It allows the architecture to switch the
9795 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07009796 * must be called with all CPU's synchronized, and interrupts disabled, the
9797 * and caller must save the original value of the current task (see
9798 * curr_task() above) and restore that value before reenabling interrupts and
9799 * re-starting the system.
9800 *
9801 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9802 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009803void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009804{
9805 cpu_curr(cpu) = p;
9806}
9807
9808#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009809
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009810#ifdef CONFIG_FAIR_GROUP_SCHED
9811static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009812{
9813 int i;
9814
9815 for_each_possible_cpu(i) {
9816 if (tg->cfs_rq)
9817 kfree(tg->cfs_rq[i]);
9818 if (tg->se)
9819 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009820 }
9821
9822 kfree(tg->cfs_rq);
9823 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009824}
9825
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009826static
9827int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009828{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009829 struct cfs_rq *cfs_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009830 struct sched_entity *se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009831 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009832 int i;
9833
Mike Travis434d53b2008-04-04 18:11:04 -07009834 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009835 if (!tg->cfs_rq)
9836 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009837 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009838 if (!tg->se)
9839 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009840
9841 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009842
9843 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009844 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009845
Li Zefaneab17222008-10-29 17:03:22 +08009846 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9847 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009848 if (!cfs_rq)
9849 goto err;
9850
Li Zefaneab17222008-10-29 17:03:22 +08009851 se = kzalloc_node(sizeof(struct sched_entity),
9852 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009853 if (!se)
Phil Carmodydfc12eb2009-12-10 14:29:37 +02009854 goto err_free_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009855
Li Zefaneab17222008-10-29 17:03:22 +08009856 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009857 }
9858
9859 return 1;
9860
Phil Carmodydfc12eb2009-12-10 14:29:37 +02009861 err_free_rq:
9862 kfree(cfs_rq);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009863 err:
9864 return 0;
9865}
9866
9867static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9868{
9869 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9870 &cpu_rq(cpu)->leaf_cfs_rq_list);
9871}
9872
9873static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9874{
9875 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9876}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009877#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009878static inline void free_fair_sched_group(struct task_group *tg)
9879{
9880}
9881
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009882static inline
9883int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009884{
9885 return 1;
9886}
9887
9888static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9889{
9890}
9891
9892static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9893{
9894}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009895#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009896
9897#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009898static void free_rt_sched_group(struct task_group *tg)
9899{
9900 int i;
9901
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009902 destroy_rt_bandwidth(&tg->rt_bandwidth);
9903
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009904 for_each_possible_cpu(i) {
9905 if (tg->rt_rq)
9906 kfree(tg->rt_rq[i]);
9907 if (tg->rt_se)
9908 kfree(tg->rt_se[i]);
9909 }
9910
9911 kfree(tg->rt_rq);
9912 kfree(tg->rt_se);
9913}
9914
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009915static
9916int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009917{
9918 struct rt_rq *rt_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009919 struct sched_rt_entity *rt_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009920 struct rq *rq;
9921 int i;
9922
Mike Travis434d53b2008-04-04 18:11:04 -07009923 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009924 if (!tg->rt_rq)
9925 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009926 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009927 if (!tg->rt_se)
9928 goto err;
9929
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009930 init_rt_bandwidth(&tg->rt_bandwidth,
9931 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009932
9933 for_each_possible_cpu(i) {
9934 rq = cpu_rq(i);
9935
Li Zefaneab17222008-10-29 17:03:22 +08009936 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9937 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009938 if (!rt_rq)
9939 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009940
Li Zefaneab17222008-10-29 17:03:22 +08009941 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9942 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009943 if (!rt_se)
Phil Carmodydfc12eb2009-12-10 14:29:37 +02009944 goto err_free_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009945
Li Zefaneab17222008-10-29 17:03:22 +08009946 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009947 }
9948
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009949 return 1;
9950
Phil Carmodydfc12eb2009-12-10 14:29:37 +02009951 err_free_rq:
9952 kfree(rt_rq);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009953 err:
9954 return 0;
9955}
9956
9957static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9958{
9959 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9960 &cpu_rq(cpu)->leaf_rt_rq_list);
9961}
9962
9963static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9964{
9965 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9966}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009967#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009968static inline void free_rt_sched_group(struct task_group *tg)
9969{
9970}
9971
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009972static inline
9973int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009974{
9975 return 1;
9976}
9977
9978static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9979{
9980}
9981
9982static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9983{
9984}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009985#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009986
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009987#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009988static void free_sched_group(struct task_group *tg)
9989{
9990 free_fair_sched_group(tg);
9991 free_rt_sched_group(tg);
9992 kfree(tg);
9993}
9994
9995/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009996struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009997{
9998 struct task_group *tg;
9999 unsigned long flags;
10000 int i;
10001
10002 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10003 if (!tg)
10004 return ERR_PTR(-ENOMEM);
10005
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010006 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010007 goto err;
10008
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010009 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010010 goto err;
10011
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010012 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010013 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010014 register_fair_sched_group(tg, i);
10015 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010016 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010017 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010018
10019 WARN_ON(!parent); /* root should already exist */
10020
10021 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010022 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +080010023 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010024 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010025
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010026 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010027
10028err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010029 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010030 return ERR_PTR(-ENOMEM);
10031}
10032
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010033/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010034static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010035{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010036 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010037 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010038}
10039
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010040/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010041void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010042{
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010043 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010044 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010045
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010046 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010047 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010048 unregister_fair_sched_group(tg, i);
10049 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010050 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010051 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010052 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010053 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010054
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010055 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010056 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010057}
10058
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010059/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +020010060 * The caller of this function should have put the task in its new group
10061 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10062 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010063 */
10064void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010065{
10066 int on_rq, running;
10067 unsigned long flags;
10068 struct rq *rq;
10069
10070 rq = task_rq_lock(tsk, &flags);
10071
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010072 update_rq_clock(rq);
10073
Dmitry Adamushko051a1d12007-12-18 15:21:13 +010010074 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010075 on_rq = tsk->se.on_rq;
10076
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010077 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010078 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010079 if (unlikely(running))
10080 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010081
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010082 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010083
Peter Zijlstra810b3812008-02-29 15:21:01 -050010084#ifdef CONFIG_FAIR_GROUP_SCHED
10085 if (tsk->sched_class->moved_group)
10086 tsk->sched_class->moved_group(tsk);
10087#endif
10088
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010089 if (unlikely(running))
10090 tsk->sched_class->set_curr_task(rq);
10091 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +020010092 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010093
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010094 task_rq_unlock(rq, &flags);
10095}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010096#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010097
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010098#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010099static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010100{
10101 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010102 int on_rq;
10103
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010104 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010105 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010106 dequeue_entity(cfs_rq, se, 0);
10107
10108 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +020010109 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010110
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010111 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010112 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010113}
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010114
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010115static void set_se_shares(struct sched_entity *se, unsigned long shares)
10116{
10117 struct cfs_rq *cfs_rq = se->cfs_rq;
10118 struct rq *rq = cfs_rq->rq;
10119 unsigned long flags;
10120
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010121 raw_spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010122 __set_se_shares(se, shares);
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010123 raw_spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010124}
10125
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010126static DEFINE_MUTEX(shares_mutex);
10127
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010128int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010129{
10130 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010131 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +010010132
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010133 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010134 * We can't change the weight of the root cgroup.
10135 */
10136 if (!tg->se[0])
10137 return -EINVAL;
10138
Peter Zijlstra18d95a22008-04-19 19:45:00 +020010139 if (shares < MIN_SHARES)
10140 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010141 else if (shares > MAX_SHARES)
10142 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010143
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010144 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010145 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010146 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010147
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010148 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010149 for_each_possible_cpu(i)
10150 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010151 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010152 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010153
10154 /* wait for any ongoing reference to this group to finish */
10155 synchronize_sched();
10156
10157 /*
10158 * Now we are free to modify the group's share on each cpu
10159 * w/o tripping rebalance_share or load_balance_fair.
10160 */
10161 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010162 for_each_possible_cpu(i) {
10163 /*
10164 * force a rebalance
10165 */
10166 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010167 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010168 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010169
10170 /*
10171 * Enable load balance activity on this group, by inserting it back on
10172 * each cpu's rq->leaf_cfs_rq_list.
10173 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010174 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010175 for_each_possible_cpu(i)
10176 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010177 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010178 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010179done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010180 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010181 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010182}
10183
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010184unsigned long sched_group_shares(struct task_group *tg)
10185{
10186 return tg->shares;
10187}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010188#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010189
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010190#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010191/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010192 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010193 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010194static DEFINE_MUTEX(rt_constraints_mutex);
10195
10196static unsigned long to_ratio(u64 period, u64 runtime)
10197{
10198 if (runtime == RUNTIME_INF)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010199 return 1ULL << 20;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010200
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010201 return div64_u64(runtime << 20, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010202}
10203
Dhaval Giani521f1a242008-02-28 15:21:56 +053010204/* Must be called with tasklist_lock held */
10205static inline int tg_has_rt_tasks(struct task_group *tg)
10206{
10207 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010208
Dhaval Giani521f1a242008-02-28 15:21:56 +053010209 do_each_thread(g, p) {
10210 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10211 return 1;
10212 } while_each_thread(g, p);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010213
Dhaval Giani521f1a242008-02-28 15:21:56 +053010214 return 0;
10215}
10216
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010217struct rt_schedulable_data {
10218 struct task_group *tg;
10219 u64 rt_period;
10220 u64 rt_runtime;
10221};
10222
10223static int tg_schedulable(struct task_group *tg, void *data)
10224{
10225 struct rt_schedulable_data *d = data;
10226 struct task_group *child;
10227 unsigned long total, sum = 0;
10228 u64 period, runtime;
10229
10230 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10231 runtime = tg->rt_bandwidth.rt_runtime;
10232
10233 if (tg == d->tg) {
10234 period = d->rt_period;
10235 runtime = d->rt_runtime;
10236 }
10237
Peter Zijlstra98a48262009-01-14 10:56:32 +010010238#ifdef CONFIG_USER_SCHED
10239 if (tg == &root_task_group) {
10240 period = global_rt_period();
10241 runtime = global_rt_runtime();
10242 }
10243#endif
10244
Peter Zijlstra4653f802008-09-23 15:33:44 +020010245 /*
10246 * Cannot have more runtime than the period.
10247 */
10248 if (runtime > period && runtime != RUNTIME_INF)
10249 return -EINVAL;
10250
10251 /*
10252 * Ensure we don't starve existing RT tasks.
10253 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010254 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10255 return -EBUSY;
10256
10257 total = to_ratio(period, runtime);
10258
Peter Zijlstra4653f802008-09-23 15:33:44 +020010259 /*
10260 * Nobody can have more than the global setting allows.
10261 */
10262 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10263 return -EINVAL;
10264
10265 /*
10266 * The sum of our children's runtime should not exceed our own.
10267 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010268 list_for_each_entry_rcu(child, &tg->children, siblings) {
10269 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10270 runtime = child->rt_bandwidth.rt_runtime;
10271
10272 if (child == d->tg) {
10273 period = d->rt_period;
10274 runtime = d->rt_runtime;
10275 }
10276
10277 sum += to_ratio(period, runtime);
10278 }
10279
10280 if (sum > total)
10281 return -EINVAL;
10282
10283 return 0;
10284}
10285
10286static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10287{
10288 struct rt_schedulable_data data = {
10289 .tg = tg,
10290 .rt_period = period,
10291 .rt_runtime = runtime,
10292 };
10293
10294 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10295}
10296
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010297static int tg_set_bandwidth(struct task_group *tg,
10298 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010299{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010300 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010301
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010302 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +053010303 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010304 err = __rt_schedulable(tg, rt_period, rt_runtime);
10305 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +053010306 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010307
Thomas Gleixner0986b112009-11-17 15:32:06 +010010308 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010309 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10310 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010311
10312 for_each_possible_cpu(i) {
10313 struct rt_rq *rt_rq = tg->rt_rq[i];
10314
Thomas Gleixner0986b112009-11-17 15:32:06 +010010315 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010316 rt_rq->rt_runtime = rt_runtime;
Thomas Gleixner0986b112009-11-17 15:32:06 +010010317 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010318 }
Thomas Gleixner0986b112009-11-17 15:32:06 +010010319 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010320 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +053010321 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010322 mutex_unlock(&rt_constraints_mutex);
10323
10324 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010325}
10326
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010327int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10328{
10329 u64 rt_runtime, rt_period;
10330
10331 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10332 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10333 if (rt_runtime_us < 0)
10334 rt_runtime = RUNTIME_INF;
10335
10336 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10337}
10338
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010339long sched_group_rt_runtime(struct task_group *tg)
10340{
10341 u64 rt_runtime_us;
10342
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010343 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010344 return -1;
10345
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010346 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010347 do_div(rt_runtime_us, NSEC_PER_USEC);
10348 return rt_runtime_us;
10349}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010350
10351int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10352{
10353 u64 rt_runtime, rt_period;
10354
10355 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10356 rt_runtime = tg->rt_bandwidth.rt_runtime;
10357
Raistlin619b0482008-06-26 18:54:09 +020010358 if (rt_period == 0)
10359 return -EINVAL;
10360
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010361 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10362}
10363
10364long sched_group_rt_period(struct task_group *tg)
10365{
10366 u64 rt_period_us;
10367
10368 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10369 do_div(rt_period_us, NSEC_PER_USEC);
10370 return rt_period_us;
10371}
10372
10373static int sched_rt_global_constraints(void)
10374{
Peter Zijlstra4653f802008-09-23 15:33:44 +020010375 u64 runtime, period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010376 int ret = 0;
10377
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010378 if (sysctl_sched_rt_period <= 0)
10379 return -EINVAL;
10380
Peter Zijlstra4653f802008-09-23 15:33:44 +020010381 runtime = global_rt_runtime();
10382 period = global_rt_period();
10383
10384 /*
10385 * Sanity check on the sysctl variables.
10386 */
10387 if (runtime > period && runtime != RUNTIME_INF)
10388 return -EINVAL;
Peter Zijlstra10b612f2008-06-19 14:22:27 +020010389
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010390 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010391 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +020010392 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010393 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010394 mutex_unlock(&rt_constraints_mutex);
10395
10396 return ret;
10397}
Dhaval Giani54e99122009-02-27 15:13:54 +053010398
10399int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10400{
10401 /* Don't accept realtime tasks when there is no way for them to run */
10402 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10403 return 0;
10404
10405 return 1;
10406}
10407
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010408#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010409static int sched_rt_global_constraints(void)
10410{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010411 unsigned long flags;
10412 int i;
10413
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010414 if (sysctl_sched_rt_period <= 0)
10415 return -EINVAL;
10416
Peter Zijlstra60aa6052009-05-05 17:50:21 +020010417 /*
10418 * There's always some RT tasks in the root group
10419 * -- migration, kstopmachine etc..
10420 */
10421 if (sysctl_sched_rt_runtime == 0)
10422 return -EBUSY;
10423
Thomas Gleixner0986b112009-11-17 15:32:06 +010010424 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010425 for_each_possible_cpu(i) {
10426 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10427
Thomas Gleixner0986b112009-11-17 15:32:06 +010010428 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010429 rt_rq->rt_runtime = global_rt_runtime();
Thomas Gleixner0986b112009-11-17 15:32:06 +010010430 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010431 }
Thomas Gleixner0986b112009-11-17 15:32:06 +010010432 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010433
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010434 return 0;
10435}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010436#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010437
10438int sched_rt_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070010439 void __user *buffer, size_t *lenp,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010440 loff_t *ppos)
10441{
10442 int ret;
10443 int old_period, old_runtime;
10444 static DEFINE_MUTEX(mutex);
10445
10446 mutex_lock(&mutex);
10447 old_period = sysctl_sched_rt_period;
10448 old_runtime = sysctl_sched_rt_runtime;
10449
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070010450 ret = proc_dointvec(table, write, buffer, lenp, ppos);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010451
10452 if (!ret && write) {
10453 ret = sched_rt_global_constraints();
10454 if (ret) {
10455 sysctl_sched_rt_period = old_period;
10456 sysctl_sched_rt_runtime = old_runtime;
10457 } else {
10458 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10459 def_rt_bandwidth.rt_period =
10460 ns_to_ktime(global_rt_period());
10461 }
10462 }
10463 mutex_unlock(&mutex);
10464
10465 return ret;
10466}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010467
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010468#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010469
10470/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +020010471static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010472{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010473 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10474 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010475}
10476
10477static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +020010478cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010479{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010480 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010481
Paul Menage2b01dfe2007-10-24 18:23:50 +020010482 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010483 /* This is early initialization for the top cgroup */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010484 return &init_task_group.css;
10485 }
10486
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010487 parent = cgroup_tg(cgrp->parent);
10488 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010489 if (IS_ERR(tg))
10490 return ERR_PTR(-ENOMEM);
10491
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010492 return &tg->css;
10493}
10494
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010495static void
10496cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010497{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010498 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010499
10500 sched_destroy_group(tg);
10501}
10502
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010503static int
Ben Blumbe367d02009-09-23 15:56:31 -070010504cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010505{
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010506#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Giani54e99122009-02-27 15:13:54 +053010507 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010508 return -EINVAL;
10509#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010510 /* We don't support RT-tasks being in separate groups */
10511 if (tsk->sched_class != &fair_sched_class)
10512 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010513#endif
Ben Blumbe367d02009-09-23 15:56:31 -070010514 return 0;
10515}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010516
Ben Blumbe367d02009-09-23 15:56:31 -070010517static int
10518cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10519 struct task_struct *tsk, bool threadgroup)
10520{
10521 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10522 if (retval)
10523 return retval;
10524 if (threadgroup) {
10525 struct task_struct *c;
10526 rcu_read_lock();
10527 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10528 retval = cpu_cgroup_can_attach_task(cgrp, c);
10529 if (retval) {
10530 rcu_read_unlock();
10531 return retval;
10532 }
10533 }
10534 rcu_read_unlock();
10535 }
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010536 return 0;
10537}
10538
10539static void
Paul Menage2b01dfe2007-10-24 18:23:50 +020010540cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Ben Blumbe367d02009-09-23 15:56:31 -070010541 struct cgroup *old_cont, struct task_struct *tsk,
10542 bool threadgroup)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010543{
10544 sched_move_task(tsk);
Ben Blumbe367d02009-09-23 15:56:31 -070010545 if (threadgroup) {
10546 struct task_struct *c;
10547 rcu_read_lock();
10548 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10549 sched_move_task(c);
10550 }
10551 rcu_read_unlock();
10552 }
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010553}
10554
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010555#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -070010556static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +020010557 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010558{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010559 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010560}
10561
Paul Menagef4c753b2008-04-29 00:59:56 -070010562static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010563{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010564 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010565
10566 return (u64) tg->shares;
10567}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010568#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010569
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010570#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -070010571static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -070010572 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010573{
Paul Menage06ecb272008-04-29 01:00:06 -070010574 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010575}
10576
Paul Menage06ecb272008-04-29 01:00:06 -070010577static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010578{
Paul Menage06ecb272008-04-29 01:00:06 -070010579 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010580}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010581
10582static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10583 u64 rt_period_us)
10584{
10585 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10586}
10587
10588static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10589{
10590 return sched_group_rt_period(cgroup_tg(cgrp));
10591}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010592#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010593
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010594static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010595#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010596 {
10597 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -070010598 .read_u64 = cpu_shares_read_u64,
10599 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010600 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010601#endif
10602#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010603 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010604 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -070010605 .read_s64 = cpu_rt_runtime_read,
10606 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010607 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010608 {
10609 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -070010610 .read_u64 = cpu_rt_period_read_uint,
10611 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010612 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010613#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010614};
10615
10616static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10617{
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010618 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010619}
10620
10621struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +010010622 .name = "cpu",
10623 .create = cpu_cgroup_create,
10624 .destroy = cpu_cgroup_destroy,
10625 .can_attach = cpu_cgroup_can_attach,
10626 .attach = cpu_cgroup_attach,
10627 .populate = cpu_cgroup_populate,
10628 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010629 .early_init = 1,
10630};
10631
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010632#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010633
10634#ifdef CONFIG_CGROUP_CPUACCT
10635
10636/*
10637 * CPU accounting code for task groups.
10638 *
10639 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10640 * (balbir@in.ibm.com).
10641 */
10642
Bharata B Rao934352f2008-11-10 20:41:13 +053010643/* track cpu usage of a group of tasks and its child groups */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010644struct cpuacct {
10645 struct cgroup_subsys_state css;
10646 /* cpuusage holds pointer to a u64-type object on every cpu */
10647 u64 *cpuusage;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010648 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
Bharata B Rao934352f2008-11-10 20:41:13 +053010649 struct cpuacct *parent;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010650};
10651
10652struct cgroup_subsys cpuacct_subsys;
10653
10654/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010655static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010656{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010657 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010658 struct cpuacct, css);
10659}
10660
10661/* return cpu accounting group to which this task belongs */
10662static inline struct cpuacct *task_ca(struct task_struct *tsk)
10663{
10664 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10665 struct cpuacct, css);
10666}
10667
10668/* create a new cpu accounting group */
10669static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +053010670 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010671{
10672 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010673 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010674
10675 if (!ca)
Bharata B Raoef12fef2009-03-31 10:02:22 +053010676 goto out;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010677
10678 ca->cpuusage = alloc_percpu(u64);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010679 if (!ca->cpuusage)
10680 goto out_free_ca;
10681
10682 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10683 if (percpu_counter_init(&ca->cpustat[i], 0))
10684 goto out_free_counters;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010685
Bharata B Rao934352f2008-11-10 20:41:13 +053010686 if (cgrp->parent)
10687 ca->parent = cgroup_ca(cgrp->parent);
10688
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010689 return &ca->css;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010690
10691out_free_counters:
10692 while (--i >= 0)
10693 percpu_counter_destroy(&ca->cpustat[i]);
10694 free_percpu(ca->cpuusage);
10695out_free_ca:
10696 kfree(ca);
10697out:
10698 return ERR_PTR(-ENOMEM);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010699}
10700
10701/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010702static void
Dhaval Giani32cd7562008-02-29 10:02:43 +053010703cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010704{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010705 struct cpuacct *ca = cgroup_ca(cgrp);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010706 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010707
Bharata B Raoef12fef2009-03-31 10:02:22 +053010708 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10709 percpu_counter_destroy(&ca->cpustat[i]);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010710 free_percpu(ca->cpuusage);
10711 kfree(ca);
10712}
10713
Ken Chen720f5492008-12-15 22:02:01 -080010714static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10715{
Rusty Russellb36128c2009-02-20 16:29:08 +090010716 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010717 u64 data;
10718
10719#ifndef CONFIG_64BIT
10720 /*
10721 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10722 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010723 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
Ken Chen720f5492008-12-15 22:02:01 -080010724 data = *cpuusage;
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010725 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
Ken Chen720f5492008-12-15 22:02:01 -080010726#else
10727 data = *cpuusage;
10728#endif
10729
10730 return data;
10731}
10732
10733static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10734{
Rusty Russellb36128c2009-02-20 16:29:08 +090010735 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010736
10737#ifndef CONFIG_64BIT
10738 /*
10739 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10740 */
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010741 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
Ken Chen720f5492008-12-15 22:02:01 -080010742 *cpuusage = val;
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010743 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
Ken Chen720f5492008-12-15 22:02:01 -080010744#else
10745 *cpuusage = val;
10746#endif
10747}
10748
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010749/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010750static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010751{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010752 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010753 u64 totalcpuusage = 0;
10754 int i;
10755
Ken Chen720f5492008-12-15 22:02:01 -080010756 for_each_present_cpu(i)
10757 totalcpuusage += cpuacct_cpuusage_read(ca, i);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010758
10759 return totalcpuusage;
10760}
10761
Dhaval Giani0297b802008-02-29 10:02:44 +053010762static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10763 u64 reset)
10764{
10765 struct cpuacct *ca = cgroup_ca(cgrp);
10766 int err = 0;
10767 int i;
10768
10769 if (reset) {
10770 err = -EINVAL;
10771 goto out;
10772 }
10773
Ken Chen720f5492008-12-15 22:02:01 -080010774 for_each_present_cpu(i)
10775 cpuacct_cpuusage_write(ca, i, 0);
Dhaval Giani0297b802008-02-29 10:02:44 +053010776
Dhaval Giani0297b802008-02-29 10:02:44 +053010777out:
10778 return err;
10779}
10780
Ken Chene9515c32008-12-15 22:04:15 -080010781static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10782 struct seq_file *m)
10783{
10784 struct cpuacct *ca = cgroup_ca(cgroup);
10785 u64 percpu;
10786 int i;
10787
10788 for_each_present_cpu(i) {
10789 percpu = cpuacct_cpuusage_read(ca, i);
10790 seq_printf(m, "%llu ", (unsigned long long) percpu);
10791 }
10792 seq_printf(m, "\n");
10793 return 0;
10794}
10795
Bharata B Raoef12fef2009-03-31 10:02:22 +053010796static const char *cpuacct_stat_desc[] = {
10797 [CPUACCT_STAT_USER] = "user",
10798 [CPUACCT_STAT_SYSTEM] = "system",
10799};
10800
10801static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10802 struct cgroup_map_cb *cb)
10803{
10804 struct cpuacct *ca = cgroup_ca(cgrp);
10805 int i;
10806
10807 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10808 s64 val = percpu_counter_read(&ca->cpustat[i]);
10809 val = cputime64_to_clock_t(val);
10810 cb->fill(cb, cpuacct_stat_desc[i], val);
10811 }
10812 return 0;
10813}
10814
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010815static struct cftype files[] = {
10816 {
10817 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -070010818 .read_u64 = cpuusage_read,
10819 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010820 },
Ken Chene9515c32008-12-15 22:04:15 -080010821 {
10822 .name = "usage_percpu",
10823 .read_seq_string = cpuacct_percpu_seq_read,
10824 },
Bharata B Raoef12fef2009-03-31 10:02:22 +053010825 {
10826 .name = "stat",
10827 .read_map = cpuacct_stats_show,
10828 },
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010829};
10830
Dhaval Giani32cd7562008-02-29 10:02:43 +053010831static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010832{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010833 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010834}
10835
10836/*
10837 * charge this task's execution time to its accounting group.
10838 *
10839 * called with rq->lock held.
10840 */
10841static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10842{
10843 struct cpuacct *ca;
Bharata B Rao934352f2008-11-10 20:41:13 +053010844 int cpu;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010845
Li Zefanc40c6f82009-02-26 15:40:15 +080010846 if (unlikely(!cpuacct_subsys.active))
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010847 return;
10848
Bharata B Rao934352f2008-11-10 20:41:13 +053010849 cpu = task_cpu(tsk);
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010850
10851 rcu_read_lock();
10852
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010853 ca = task_ca(tsk);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010854
Bharata B Rao934352f2008-11-10 20:41:13 +053010855 for (; ca; ca = ca->parent) {
Rusty Russellb36128c2009-02-20 16:29:08 +090010856 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010857 *cpuusage += cputime;
10858 }
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010859
10860 rcu_read_unlock();
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010861}
10862
Bharata B Raoef12fef2009-03-31 10:02:22 +053010863/*
10864 * Charge the system/user time to the task's accounting group.
10865 */
10866static void cpuacct_update_stats(struct task_struct *tsk,
10867 enum cpuacct_stat_index idx, cputime_t val)
10868{
10869 struct cpuacct *ca;
10870
10871 if (unlikely(!cpuacct_subsys.active))
10872 return;
10873
10874 rcu_read_lock();
10875 ca = task_ca(tsk);
10876
10877 do {
10878 percpu_counter_add(&ca->cpustat[idx], val);
10879 ca = ca->parent;
10880 } while (ca);
10881 rcu_read_unlock();
10882}
10883
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010884struct cgroup_subsys cpuacct_subsys = {
10885 .name = "cpuacct",
10886 .create = cpuacct_create,
10887 .destroy = cpuacct_destroy,
10888 .populate = cpuacct_populate,
10889 .subsys_id = cpuacct_subsys_id,
10890};
10891#endif /* CONFIG_CGROUP_CPUACCT */
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010892
10893#ifndef CONFIG_SMP
10894
10895int rcu_expedited_torture_stats(char *page)
10896{
10897 return 0;
10898}
10899EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10900
10901void synchronize_sched_expedited(void)
10902{
10903}
10904EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10905
10906#else /* #ifndef CONFIG_SMP */
10907
10908static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10909static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10910
10911#define RCU_EXPEDITED_STATE_POST -2
10912#define RCU_EXPEDITED_STATE_IDLE -1
10913
10914static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10915
10916int rcu_expedited_torture_stats(char *page)
10917{
10918 int cnt = 0;
10919 int cpu;
10920
10921 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10922 for_each_online_cpu(cpu) {
10923 cnt += sprintf(&page[cnt], " %d:%d",
10924 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10925 }
10926 cnt += sprintf(&page[cnt], "\n");
10927 return cnt;
10928}
10929EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10930
10931static long synchronize_sched_expedited_count;
10932
10933/*
10934 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10935 * approach to force grace period to end quickly. This consumes
10936 * significant time on all CPUs, and is thus not recommended for
10937 * any sort of common-case code.
10938 *
10939 * Note that it is illegal to call this function while holding any
10940 * lock that is acquired by a CPU-hotplug notifier. Failing to
10941 * observe this restriction will result in deadlock.
10942 */
10943void synchronize_sched_expedited(void)
10944{
10945 int cpu;
10946 unsigned long flags;
10947 bool need_full_sync = 0;
10948 struct rq *rq;
10949 struct migration_req *req;
10950 long snap;
10951 int trycount = 0;
10952
10953 smp_mb(); /* ensure prior mod happens before capturing snap. */
10954 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10955 get_online_cpus();
10956 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10957 put_online_cpus();
10958 if (trycount++ < 10)
10959 udelay(trycount * num_online_cpus());
10960 else {
10961 synchronize_sched();
10962 return;
10963 }
10964 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10965 smp_mb(); /* ensure test happens before caller kfree */
10966 return;
10967 }
10968 get_online_cpus();
10969 }
10970 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10971 for_each_online_cpu(cpu) {
10972 rq = cpu_rq(cpu);
10973 req = &per_cpu(rcu_migration_req, cpu);
10974 init_completion(&req->done);
10975 req->task = NULL;
10976 req->dest_cpu = RCU_MIGRATION_NEED_QS;
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010977 raw_spin_lock_irqsave(&rq->lock, flags);
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010978 list_add(&req->list, &rq->migration_queue);
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010979 raw_spin_unlock_irqrestore(&rq->lock, flags);
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010980 wake_up_process(rq->migration_thread);
10981 }
10982 for_each_online_cpu(cpu) {
10983 rcu_expedited_state = cpu;
10984 req = &per_cpu(rcu_migration_req, cpu);
10985 rq = cpu_rq(cpu);
10986 wait_for_completion(&req->done);
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010987 raw_spin_lock_irqsave(&rq->lock, flags);
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010988 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10989 need_full_sync = 1;
10990 req->dest_cpu = RCU_MIGRATION_IDLE;
Thomas Gleixner05fa7852009-11-17 14:28:38 +010010991 raw_spin_unlock_irqrestore(&rq->lock, flags);
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010992 }
10993 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
Paul E. McKenney956539b2009-11-10 13:37:20 -080010994 synchronize_sched_expedited_count++;
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010995 mutex_unlock(&rcu_sched_expedited_mutex);
10996 put_online_cpus();
10997 if (need_full_sync)
10998 synchronize_sched();
10999}
11000EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11001
11002#endif /* #else #ifndef CONFIG_SMP */