blob: faf4d463bbffb9a561beed93bd59282ea7c7092e [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 Molnar0d905bc2009-05-04 19:13:30 +020042#include <linux/perf_counter.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: */
144 spinlock_t rt_runtime_lock;
145 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
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200181 spin_lock_init(&rt_b->rt_runtime_lock);
182
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
203 spin_lock(&rt_b->rt_runtime_lock);
204 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 }
220 spin_unlock(&rt_b->rt_runtime_lock);
221}
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 Heob9bf3122009-06-24 15:13:47 +0900301static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
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
Peter Zijlstra57310a92009-03-09 13:56:21 +0100312#ifdef CONFIG_SMP
313static int root_task_group_empty(void)
314{
315 return list_empty(&root_task_group.children);
316}
317#endif
318
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100319#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100320#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100321# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200322#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100323# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200324#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200325
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800326/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
333 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200334#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800335#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200336
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100337static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100338#endif
339
340/* Default task group.
341 * Every task in system belong to this group at bootup.
342 */
Mike Travis434d53b2008-04-04 18:11:04 -0700343struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200344
345/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200346static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200347{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200348 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200349
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100350#ifdef CONFIG_USER_SCHED
David Howellsc69e8d92008-11-14 10:39:19 +1100351 rcu_read_lock();
352 tg = __task_cred(p)->user->tg;
353 rcu_read_unlock();
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100354#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200357#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100358 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200359#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200360 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200361}
362
363/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100364static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200365{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100366#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100369#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100370
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100371#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100374#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200375}
376
377#else
378
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100379static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Peter Zijlstra83378262008-06-27 13:41:37 +0200380static inline struct task_group *task_group(struct task_struct *p)
381{
382 return NULL;
383}
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200384
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100385#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200386
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200387/* CFS-related fields in a runqueue */
388struct cfs_rq {
389 struct load_weight load;
390 unsigned long nr_running;
391
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200392 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200393 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200394
395 struct rb_root tasks_timeline;
396 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200397
398 struct list_head tasks;
399 struct list_head *balance_iterator;
400
401 /*
402 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200403 * It is set to NULL otherwise (i.e when none are currently running).
404 */
Peter Zijlstra47932412008-11-04 21:25:09 +0100405 struct sched_entity *curr, *next, *last;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200406
Peter Zijlstra5ac5c4d2008-11-10 10:46:32 +0100407 unsigned int nr_spread_over;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200408
Ingo Molnar62160e32007-10-15 17:00:03 +0200409#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200410 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
411
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100412 /*
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
416 *
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
419 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100420 struct list_head leaf_cfs_rq_list;
421 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200422
423#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200424 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200425 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200426 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200427 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200428
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200429 /*
430 * h_load = weight * f(tg)
431 *
432 * Where f(tg) is the recursive weight fraction assigned to
433 * this group.
434 */
435 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200436
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200437 /*
438 * this cpu's part of tg->shares
439 */
440 unsigned long shares;
Peter Zijlstraf1d239f2008-06-27 13:41:38 +0200441
442 /*
443 * load.weight at the time we set shares
444 */
445 unsigned long rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200446#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200447#endif
448};
449
450/* Real-Time classes' related field in a runqueue: */
451struct rt_rq {
452 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100453 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100454#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -0500455 struct {
456 int curr; /* highest queued rt task prio */
Gregory Haskins398a1532009-01-14 09:10:04 -0500457#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500458 int next; /* next highest */
Gregory Haskins398a1532009-01-14 09:10:04 -0500459#endif
Gregory Haskinse864c492008-12-29 09:39:49 -0500460 } highest_prio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100461#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100462#ifdef CONFIG_SMP
Gregory Haskins73fe6aae2008-01-25 21:08:07 +0100463 unsigned long rt_nr_migratory;
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200464 unsigned long rt_nr_total;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100465 int overloaded;
Gregory Haskins917b6272008-12-29 09:39:53 -0500466 struct plist_head pushable_tasks;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100467#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100468 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100469 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200470 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100471 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200472 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100473
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100474#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100475 unsigned long rt_nr_boosted;
476
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100477 struct rq *rq;
478 struct list_head leaf_rt_rq_list;
479 struct task_group *tg;
480 struct sched_rt_entity *rt_se;
481#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200482};
483
Gregory Haskins57d885f2008-01-25 21:08:18 +0100484#ifdef CONFIG_SMP
485
486/*
487 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100490 * exclusive cpuset is created, we also create and attach a new root-domain
491 * object.
492 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100493 */
494struct root_domain {
495 atomic_t refcount;
Rusty Russellc6c49272008-11-25 02:35:05 +1030496 cpumask_var_t span;
497 cpumask_var_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100498
Ingo Molnar0eab9142008-01-25 21:08:19 +0100499 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
502 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030503 cpumask_var_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100504 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200505#ifdef CONFIG_SMP
506 struct cpupri cpupri;
507#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100508};
509
Gregory Haskinsdc938522008-01-25 21:08:26 +0100510/*
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
513 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100514static struct root_domain def_root_domain;
515
516#endif
517
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200518/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 * This is the main, per-CPU runqueue data structure.
520 *
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
524 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700525struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200526 /* runqueue lock: */
527 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528
529 /*
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
532 */
533 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700536#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200537 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700538 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;
Paul Mackerras23a185c2009-02-09 22:42:47 +1100544 u64 nr_migrations_in;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200545
546 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100547 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100548
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200549#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100552#endif
553#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100554 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556
557 /*
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
562 */
563 unsigned long nr_uninterruptible;
564
Ingo Molnar36c8b582006-07-03 00:25:41 -0700565 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800566 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200568
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200569 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200570
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 atomic_t nr_iowait;
572
573#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100574 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 struct sched_domain *sd;
576
Henrik Austada0a522c2009-02-13 20:35:45 +0100577 unsigned char idle_at_tick;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700578 /* For active balancing */
Gregory Haskins3f029d32009-07-29 11:08:47 -0400579 int post_schedule;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 int active_balance;
581 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200582 /* cpu of this runqueue: */
583 int cpu;
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -0400584 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
Peter Zijlstraa8a51d52008-06-27 13:41:26 +0200586 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587
Ingo Molnar36c8b582006-07-03 00:25:41 -0700588 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 struct list_head migration_queue;
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200590
591 u64 rt_avg;
592 u64 age_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593#endif
594
Thomas Gleixnerdce48a82009-04-11 10:43:41 +0200595 /* calc_load related fields */
596 unsigned long calc_load_update;
597 long calc_load_active;
598
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100599#ifdef CONFIG_SCHED_HRTICK
Peter Zijlstra31656512008-07-18 18:01:23 +0200600#ifdef CONFIG_SMP
601 int hrtick_csd_pending;
602 struct call_single_data hrtick_csd;
603#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100604 struct hrtimer hrtick_timer;
605#endif
606
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607#ifdef CONFIG_SCHEDSTATS
608 /* latency stats */
609 struct sched_info rq_sched_info;
Ken Chen9c2c4802008-12-16 23:41:22 -0800610 unsigned long long rq_cpu_time;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612
613 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200614 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
616 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620
621 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200624
625 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200626 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627#endif
628};
629
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700630static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
Peter Zijlstra7d478722009-09-14 19:55:44 +0200632static inline
633void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnardd41f592007-07-09 18:51:59 +0200634{
Peter Zijlstra7d478722009-09-14 19:55:44 +0200635 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
Ingo Molnardd41f592007-07-09 18:51:59 +0200636}
637
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700638static inline int cpu_of(struct rq *rq)
639{
640#ifdef CONFIG_SMP
641 return rq->cpu;
642#else
643 return 0;
644#endif
645}
646
Ingo Molnar20d315d2007-07-09 18:51:58 +0200647/*
Nick Piggin674311d2005-06-25 14:57:27 -0700648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700649 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700650 *
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
653 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700654#define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
657#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658#define this_rq() (&__get_cpu_var(runqueues))
659#define task_rq(p) cpu_rq(task_cpu(p))
660#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
Hitoshi Mitake54d35f22009-06-29 14:44:57 +0900661#define raw_rq() (&__raw_get_cpu_var(runqueues))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
Ingo Molnaraa9c4c02008-12-17 14:10:57 +0100663inline void update_rq_clock(struct rq *rq)
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200664{
665 rq->clock = sched_clock_cpu(cpu_of(rq));
666}
667
Ingo Molnare436d802007-07-19 21:28:35 +0200668/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
670 */
671#ifdef CONFIG_SCHED_DEBUG
672# define const_debug __read_mostly
673#else
674# define const_debug static const
675#endif
676
Ingo Molnar017730c2008-05-12 21:20:52 +0200677/**
678 * runqueue_is_locked
679 *
680 * Returns true if the current cpu runqueue is locked.
681 * This interface allows printk to be called with the runqueue lock
682 * held and know whether or not it is OK to wake up the klogd.
683 */
684int runqueue_is_locked(void)
685{
686 int cpu = get_cpu();
687 struct rq *rq = cpu_rq(cpu);
688 int ret;
689
690 ret = spin_is_locked(&rq->lock);
691 put_cpu();
692 return ret;
693}
694
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200695/*
696 * Debugging: various feature bits
697 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200698
699#define SCHED_FEAT(name, enabled) \
700 __SCHED_FEAT_##name ,
701
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200702enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200703#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200704};
705
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200706#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200707
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200708#define SCHED_FEAT(name, enabled) \
709 (1UL << __SCHED_FEAT_##name) * enabled |
710
711const_debug unsigned int sysctl_sched_features =
712#include "sched_features.h"
713 0;
714
715#undef SCHED_FEAT
716
717#ifdef CONFIG_SCHED_DEBUG
718#define SCHED_FEAT(name, enabled) \
719 #name ,
720
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700721static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200722#include "sched_features.h"
723 NULL
724};
725
726#undef SCHED_FEAT
727
Li Zefan34f3a812008-10-30 15:23:32 +0800728static int sched_feat_show(struct seq_file *m, void *v)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200729{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200730 int i;
731
732 for (i = 0; sched_feat_names[i]; i++) {
Li Zefan34f3a812008-10-30 15:23:32 +0800733 if (!(sysctl_sched_features & (1UL << i)))
734 seq_puts(m, "NO_");
735 seq_printf(m, "%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200736 }
Li Zefan34f3a812008-10-30 15:23:32 +0800737 seq_puts(m, "\n");
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200738
Li Zefan34f3a812008-10-30 15:23:32 +0800739 return 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200740}
741
742static ssize_t
743sched_feat_write(struct file *filp, const char __user *ubuf,
744 size_t cnt, loff_t *ppos)
745{
746 char buf[64];
747 char *cmp = buf;
748 int neg = 0;
749 int i;
750
751 if (cnt > 63)
752 cnt = 63;
753
754 if (copy_from_user(&buf, ubuf, cnt))
755 return -EFAULT;
756
757 buf[cnt] = 0;
758
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200759 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200760 neg = 1;
761 cmp += 3;
762 }
763
764 for (i = 0; sched_feat_names[i]; i++) {
765 int len = strlen(sched_feat_names[i]);
766
767 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
768 if (neg)
769 sysctl_sched_features &= ~(1UL << i);
770 else
771 sysctl_sched_features |= (1UL << i);
772 break;
773 }
774 }
775
776 if (!sched_feat_names[i])
777 return -EINVAL;
778
779 filp->f_pos += cnt;
780
781 return cnt;
782}
783
Li Zefan34f3a812008-10-30 15:23:32 +0800784static int sched_feat_open(struct inode *inode, struct file *filp)
785{
786 return single_open(filp, sched_feat_show, NULL);
787}
788
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200789static struct file_operations sched_feat_fops = {
Li Zefan34f3a812008-10-30 15:23:32 +0800790 .open = sched_feat_open,
791 .write = sched_feat_write,
792 .read = seq_read,
793 .llseek = seq_lseek,
794 .release = single_release,
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200795};
796
797static __init int sched_init_debug(void)
798{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200799 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 &sched_feat_fops);
801
802 return 0;
803}
804late_initcall(sched_init_debug);
805
806#endif
807
808#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200809
810/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100811 * Number of tasks to iterate in a single balance run.
812 * Limited because this is done with IRQs disabled.
813 */
814const_debug unsigned int sysctl_sched_nr_migrate = 32;
815
816/*
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200817 * ratelimit for updating the group shares.
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200818 * default: 0.25ms
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200819 */
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200820unsigned int sysctl_sched_shares_ratelimit = 250000;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200821
822/*
Peter Zijlstraffda12a2008-10-17 19:27:02 +0200823 * Inject some fuzzyness into changing the per-cpu group shares
824 * this avoids remote rq-locks at the expense of fairness.
825 * default: 4
826 */
827unsigned int sysctl_sched_shares_thresh = 4;
828
829/*
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200830 * period over which we average the RT time consumption, measured
831 * in ms.
832 *
833 * default: 1s
834 */
835const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
836
837/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100838 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100839 * default: 1s
840 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100841unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100842
Ingo Molnar6892b752008-02-13 14:02:36 +0100843static __read_mostly int scheduler_running;
844
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100845/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100846 * part of the period that we allow rt tasks to run in us.
847 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100848 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100849int sysctl_sched_rt_runtime = 950000;
850
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200851static inline u64 global_rt_period(void)
852{
853 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
854}
855
856static inline u64 global_rt_runtime(void)
857{
roel kluine26873b2008-07-22 16:51:15 -0400858 if (sysctl_sched_rt_runtime < 0)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200859 return RUNTIME_INF;
860
861 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
862}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700865# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700867#ifndef finish_arch_switch
868# define finish_arch_switch(prev) do { } while (0)
869#endif
870
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100871static inline int task_current(struct rq *rq, struct task_struct *p)
872{
873 return rq->curr == p;
874}
875
Nick Piggin4866cde2005-06-25 14:57:23 -0700876#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700877static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700878{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100879 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700880}
881
Ingo Molnar70b97a72006-07-03 00:25:42 -0700882static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700883{
884}
885
Ingo Molnar70b97a72006-07-03 00:25:42 -0700886static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700887{
Ingo Molnarda04c032005-09-13 11:17:59 +0200888#ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq->lock.owner = current;
891#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700892 /*
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
895 * prev into current:
896 */
897 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
898
Nick Piggin4866cde2005-06-25 14:57:23 -0700899 spin_unlock_irq(&rq->lock);
900}
901
902#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700903static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700904{
905#ifdef CONFIG_SMP
906 return p->oncpu;
907#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100908 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700909#endif
910}
911
Ingo Molnar70b97a72006-07-03 00:25:42 -0700912static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700913{
914#ifdef CONFIG_SMP
915 /*
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
918 * here.
919 */
920 next->oncpu = 1;
921#endif
922#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq->lock);
924#else
925 spin_unlock(&rq->lock);
926#endif
927}
928
Ingo Molnar70b97a72006-07-03 00:25:42 -0700929static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700930{
931#ifdef CONFIG_SMP
932 /*
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
935 * finished.
936 */
937 smp_wmb();
938 prev->oncpu = 0;
939#endif
940#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 local_irq_enable();
942#endif
943}
944#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945
946/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
949 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700950static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700951 __acquires(rq->lock)
952{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200953 for (;;) {
954 struct rq *rq = task_rq(p);
955 spin_lock(&rq->lock);
956 if (likely(rq == task_rq(p)))
957 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700958 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700959 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700960}
961
962/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100964 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 * explicitly disabling preemption.
966 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700967static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 __acquires(rq->lock)
969{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700970 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
Andi Kleen3a5c3592007-10-15 17:00:14 +0200972 for (;;) {
973 local_irq_save(*flags);
974 rq = task_rq(p);
975 spin_lock(&rq->lock);
976 if (likely(rq == task_rq(p)))
977 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980}
981
Oleg Nesterovad474ca2008-11-10 15:39:30 +0100982void task_rq_unlock_wait(struct task_struct *p)
983{
984 struct rq *rq = task_rq(p);
985
986 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
987 spin_unlock_wait(&rq->lock);
988}
989
Alexey Dobriyana9957442007-10-15 17:00:13 +0200990static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700991 __releases(rq->lock)
992{
993 spin_unlock(&rq->lock);
994}
995
Ingo Molnar70b97a72006-07-03 00:25:42 -0700996static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 __releases(rq->lock)
998{
999 spin_unlock_irqrestore(&rq->lock, *flags);
1000}
1001
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -08001003 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001005static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 __acquires(rq->lock)
1007{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001008 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009
1010 local_irq_disable();
1011 rq = this_rq();
1012 spin_lock(&rq->lock);
1013
1014 return rq;
1015}
1016
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001017#ifdef CONFIG_SCHED_HRTICK
1018/*
1019 * Use HR-timers to deliver accurate preemption points.
1020 *
1021 * Its all a bit involved since we cannot program an hrt while holding the
1022 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1023 * reschedule event.
1024 *
1025 * When we get rescheduled we reprogram the hrtick_timer outside of the
1026 * rq->lock.
1027 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001028
1029/*
1030 * Use hrtick when:
1031 * - enabled by features
1032 * - hrtimer is actually high res
1033 */
1034static inline int hrtick_enabled(struct rq *rq)
1035{
1036 if (!sched_feat(HRTICK))
1037 return 0;
Ingo Molnarba420592008-07-20 11:02:06 +02001038 if (!cpu_active(cpu_of(rq)))
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001039 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001040 return hrtimer_is_hres_active(&rq->hrtick_timer);
1041}
1042
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001043static void hrtick_clear(struct rq *rq)
1044{
1045 if (hrtimer_active(&rq->hrtick_timer))
1046 hrtimer_cancel(&rq->hrtick_timer);
1047}
1048
1049/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001050 * High-resolution timer tick.
1051 * Runs from hardirq context with interrupts disabled.
1052 */
1053static enum hrtimer_restart hrtick(struct hrtimer *timer)
1054{
1055 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1056
1057 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1058
1059 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001060 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001061 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1062 spin_unlock(&rq->lock);
1063
1064 return HRTIMER_NORESTART;
1065}
1066
Rabin Vincent95e904c2008-05-11 05:55:33 +05301067#ifdef CONFIG_SMP
Peter Zijlstra31656512008-07-18 18:01:23 +02001068/*
1069 * called from hardirq (IPI) context
1070 */
1071static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001072{
Peter Zijlstra31656512008-07-18 18:01:23 +02001073 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001074
Peter Zijlstra31656512008-07-18 18:01:23 +02001075 spin_lock(&rq->lock);
1076 hrtimer_restart(&rq->hrtick_timer);
1077 rq->hrtick_csd_pending = 0;
1078 spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001079}
1080
Peter Zijlstra31656512008-07-18 18:01:23 +02001081/*
1082 * Called to set the hrtick timer state.
1083 *
1084 * called with rq->lock held and irqs disabled
1085 */
1086static void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001087{
Peter Zijlstra31656512008-07-18 18:01:23 +02001088 struct hrtimer *timer = &rq->hrtick_timer;
1089 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001090
Arjan van de Vencc584b22008-09-01 15:02:30 -07001091 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +02001092
1093 if (rq == this_rq()) {
1094 hrtimer_restart(timer);
1095 } else if (!rq->hrtick_csd_pending) {
Peter Zijlstra6e275632009-02-25 13:59:48 +01001096 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001097 rq->hrtick_csd_pending = 1;
1098 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001099}
1100
1101static int
1102hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1103{
1104 int cpu = (int)(long)hcpu;
1105
1106 switch (action) {
1107 case CPU_UP_CANCELED:
1108 case CPU_UP_CANCELED_FROZEN:
1109 case CPU_DOWN_PREPARE:
1110 case CPU_DOWN_PREPARE_FROZEN:
1111 case CPU_DEAD:
1112 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +02001113 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001114 return NOTIFY_OK;
1115 }
1116
1117 return NOTIFY_DONE;
1118}
1119
Rakib Mullickfa748202008-09-22 14:55:45 -07001120static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001121{
1122 hotcpu_notifier(hotplug_hrtick, 0);
1123}
Peter Zijlstra31656512008-07-18 18:01:23 +02001124#else
1125/*
1126 * Called to set the hrtick timer state.
1127 *
1128 * called with rq->lock held and irqs disabled
1129 */
1130static void hrtick_start(struct rq *rq, u64 delay)
1131{
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +01001132 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +05301133 HRTIMER_MODE_REL_PINNED, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001134}
1135
Andrew Morton006c75f2008-09-22 14:55:46 -07001136static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +02001137{
1138}
Rabin Vincent95e904c2008-05-11 05:55:33 +05301139#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001140
1141static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001142{
Peter Zijlstra31656512008-07-18 18:01:23 +02001143#ifdef CONFIG_SMP
1144 rq->hrtick_csd_pending = 0;
1145
1146 rq->hrtick_csd.flags = 0;
1147 rq->hrtick_csd.func = __hrtick_start;
1148 rq->hrtick_csd.info = rq;
1149#endif
1150
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001151 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1152 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001153}
Andrew Morton006c75f2008-09-22 14:55:46 -07001154#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001155static inline void hrtick_clear(struct rq *rq)
1156{
1157}
1158
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001159static inline void init_rq_hrtick(struct rq *rq)
1160{
1161}
1162
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001163static inline void init_hrtick(void)
1164{
1165}
Andrew Morton006c75f2008-09-22 14:55:46 -07001166#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001167
Ingo Molnar1b9f19c22007-07-09 18:51:59 +02001168/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001169 * resched_task - mark a task 'to be rescheduled now'.
1170 *
1171 * On UP this means the setting of the need_resched flag, on SMP it
1172 * might also involve a cross-CPU call to trigger the scheduler on
1173 * the target CPU.
1174 */
1175#ifdef CONFIG_SMP
1176
1177#ifndef tsk_is_polling
1178#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1179#endif
1180
Peter Zijlstra31656512008-07-18 18:01:23 +02001181static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001182{
1183 int cpu;
1184
1185 assert_spin_locked(&task_rq(p)->lock);
1186
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001187 if (test_tsk_need_resched(p))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001188 return;
1189
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001190 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001191
1192 cpu = task_cpu(p);
1193 if (cpu == smp_processor_id())
1194 return;
1195
1196 /* NEED_RESCHED must be visible before we test polling */
1197 smp_mb();
1198 if (!tsk_is_polling(p))
1199 smp_send_reschedule(cpu);
1200}
1201
1202static void resched_cpu(int cpu)
1203{
1204 struct rq *rq = cpu_rq(cpu);
1205 unsigned long flags;
1206
1207 if (!spin_trylock_irqsave(&rq->lock, flags))
1208 return;
1209 resched_task(cpu_curr(cpu));
1210 spin_unlock_irqrestore(&rq->lock, flags);
1211}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001212
1213#ifdef CONFIG_NO_HZ
1214/*
1215 * When add_timer_on() enqueues a timer into the timer wheel of an
1216 * idle CPU then this timer might expire before the next timer event
1217 * which is scheduled to wake up that CPU. In case of a completely
1218 * idle system the next event might even be infinite time into the
1219 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1220 * leaves the inner idle loop so the newly added timer is taken into
1221 * account when the CPU goes back to idle and evaluates the timer
1222 * wheel for the next timer event.
1223 */
1224void wake_up_idle_cpu(int cpu)
1225{
1226 struct rq *rq = cpu_rq(cpu);
1227
1228 if (cpu == smp_processor_id())
1229 return;
1230
1231 /*
1232 * This is safe, as this function is called with the timer
1233 * wheel base lock of (cpu) held. When the CPU is on the way
1234 * to idle and has not yet set rq->curr to idle then it will
1235 * be serialized on the timer wheel base lock and take the new
1236 * timer into account automatically.
1237 */
1238 if (rq->curr != rq->idle)
1239 return;
1240
1241 /*
1242 * We can set TIF_RESCHED on the idle task of the other CPU
1243 * lockless. The worst case is that the other CPU runs the
1244 * idle task through an additional NOOP schedule()
1245 */
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001246 set_tsk_need_resched(rq->idle);
Thomas Gleixner06d83082008-03-22 09:20:24 +01001247
1248 /* NEED_RESCHED must be visible before we test polling */
1249 smp_mb();
1250 if (!tsk_is_polling(rq->idle))
1251 smp_send_reschedule(cpu);
1252}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001253#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001254
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001255static u64 sched_avg_period(void)
1256{
1257 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1258}
1259
1260static void sched_avg_update(struct rq *rq)
1261{
1262 s64 period = sched_avg_period();
1263
1264 while ((s64)(rq->clock - rq->age_stamp) > period) {
1265 rq->age_stamp += period;
1266 rq->rt_avg /= 2;
1267 }
1268}
1269
1270static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1271{
1272 rq->rt_avg += rt_delta;
1273 sched_avg_update(rq);
1274}
1275
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001276#else /* !CONFIG_SMP */
Peter Zijlstra31656512008-07-18 18:01:23 +02001277static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001278{
1279 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001280 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001281}
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001282
1283static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1284{
1285}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001286#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001287
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001288#if BITS_PER_LONG == 32
1289# define WMULT_CONST (~0UL)
1290#else
1291# define WMULT_CONST (1UL << 32)
1292#endif
1293
1294#define WMULT_SHIFT 32
1295
Ingo Molnar194081e2007-08-09 11:16:51 +02001296/*
1297 * Shift right and round:
1298 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001299#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001300
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001301/*
1302 * delta *= weight / lw
1303 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001304static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001305calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1306 struct load_weight *lw)
1307{
1308 u64 tmp;
1309
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001310 if (!lw->inv_weight) {
1311 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1312 lw->inv_weight = 1;
1313 else
1314 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1315 / (lw->weight+1);
1316 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001317
1318 tmp = (u64)delta_exec * weight;
1319 /*
1320 * Check whether we'd overflow the 64-bit multiplication:
1321 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001322 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001323 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001324 WMULT_SHIFT/2);
1325 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001326 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001327
Ingo Molnarecf691d2007-08-02 17:41:40 +02001328 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001329}
1330
Ingo Molnar10919852007-10-15 17:00:04 +02001331static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001332{
1333 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001334 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001335}
1336
Ingo Molnar10919852007-10-15 17:00:04 +02001337static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001338{
1339 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001340 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001341}
1342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001348 * scaled version of the new time slice allocation that they receive on time
1349 * slice expiry etc.
1350 */
1351
Peter Zijlstracce7ade2009-01-15 14:53:37 +01001352#define WEIGHT_IDLEPRIO 3
1353#define WMULT_IDLEPRIO 1431655765
Ingo Molnardd41f592007-07-09 18:51:59 +02001354
1355/*
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1360 *
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001366 */
1367static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001376};
1377
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001378/*
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 *
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1384 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001385static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001394};
Peter Williams2dd73a42006-06-27 02:54:34 -07001395
Ingo Molnardd41f592007-07-09 18:51:59 +02001396static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1397
1398/*
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1402 */
1403struct rq_iterator {
1404 void *arg;
1405 struct task_struct *(*start)(void *);
1406 struct task_struct *(*next)(void *);
1407};
1408
Peter Williamse1d14842007-10-24 18:23:51 +02001409#ifdef CONFIG_SMP
1410static unsigned long
1411balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 unsigned long max_load_move, struct sched_domain *sd,
1413 enum cpu_idle_type idle, int *all_pinned,
1414 int *this_best_prio, struct rq_iterator *iterator);
1415
1416static int
1417iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 struct sched_domain *sd, enum cpu_idle_type idle,
1419 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001420#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001421
Bharata B Raoef12fef2009-03-31 10:02:22 +05301422/* Time spent by the tasks of the cpu accounting group executing in ... */
1423enum cpuacct_stat_index {
1424 CPUACCT_STAT_USER, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1426
1427 CPUACCT_STAT_NSTATS,
1428};
1429
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001430#ifdef CONFIG_CGROUP_CPUACCT
1431static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
Bharata B Raoef12fef2009-03-31 10:02:22 +05301432static void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001434#else
1435static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
Bharata B Raoef12fef2009-03-31 10:02:22 +05301436static inline void cpuacct_update_stats(struct task_struct *tsk,
1437 enum cpuacct_stat_index idx, cputime_t val) {}
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001438#endif
1439
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001440static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1441{
1442 update_load_add(&rq->load, load);
1443}
1444
1445static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1446{
1447 update_load_sub(&rq->load, load);
1448}
1449
Ingo Molnar7940ca32008-08-19 13:40:47 +02001450#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001451typedef int (*tg_visitor)(struct task_group *, void *);
1452
1453/*
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1456 */
1457static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1458{
1459 struct task_group *parent, *child;
1460 int ret;
1461
1462 rcu_read_lock();
1463 parent = &root_task_group;
1464down:
1465 ret = (*down)(parent, data);
1466 if (ret)
1467 goto out_unlock;
1468 list_for_each_entry_rcu(child, &parent->children, siblings) {
1469 parent = child;
1470 goto down;
1471
1472up:
1473 continue;
1474 }
1475 ret = (*up)(parent, data);
1476 if (ret)
1477 goto out_unlock;
1478
1479 child = parent;
1480 parent = parent->parent;
1481 if (parent)
1482 goto up;
1483out_unlock:
1484 rcu_read_unlock();
1485
1486 return ret;
1487}
1488
1489static int tg_nop(struct task_group *tg, void *data)
1490{
1491 return 0;
1492}
1493#endif
1494
Gregory Haskinse7693a32008-01-25 21:08:09 +01001495#ifdef CONFIG_SMP
Peter Zijlstraf5f08f32009-09-10 13:35:28 +02001496/* Used instead of source_load when we know the type == 0 */
1497static unsigned long weighted_cpuload(const int cpu)
1498{
1499 return cpu_rq(cpu)->load.weight;
1500}
1501
1502/*
1503 * Return a low guess at the load of a migration-source cpu weighted
1504 * according to the scheduling class and "nice" value.
1505 *
1506 * We want to under-estimate the load of migration sources, to
1507 * balance conservatively.
1508 */
1509static unsigned long source_load(int cpu, int type)
1510{
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long total = weighted_cpuload(cpu);
1513
1514 if (type == 0 || !sched_feat(LB_BIAS))
1515 return total;
1516
1517 return min(rq->cpu_load[type-1], total);
1518}
1519
1520/*
1521 * Return a high guess at the load of a migration-target cpu weighted
1522 * according to the scheduling class and "nice" value.
1523 */
1524static unsigned long target_load(int cpu, int type)
1525{
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long total = weighted_cpuload(cpu);
1528
1529 if (type == 0 || !sched_feat(LB_BIAS))
1530 return total;
1531
1532 return max(rq->cpu_load[type-1], total);
1533}
1534
Peter Zijlstraae154be2009-09-10 14:40:57 +02001535static struct sched_group *group_of(int cpu)
1536{
1537 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1538
1539 if (!sd)
1540 return NULL;
1541
1542 return sd->groups;
1543}
1544
1545static unsigned long power_of(int cpu)
1546{
1547 struct sched_group *group = group_of(cpu);
1548
1549 if (!group)
1550 return SCHED_LOAD_SCALE;
1551
1552 return group->cpu_power;
1553}
1554
Gregory Haskinse7693a32008-01-25 21:08:09 +01001555static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001556
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001557static unsigned long cpu_avg_load_per_task(int cpu)
1558{
1559 struct rq *rq = cpu_rq(cpu);
Ingo Molnaraf6d5962008-11-29 20:45:15 +01001560 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001561
Steven Rostedt4cd42622008-11-26 21:04:24 -05001562 if (nr_running)
1563 rq->avg_load_per_task = rq->load.weight / nr_running;
Balbir Singha2d47772008-11-12 16:19:00 +05301564 else
1565 rq->avg_load_per_task = 0;
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001566
1567 return rq->avg_load_per_task;
1568}
1569
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001570#ifdef CONFIG_FAIR_GROUP_SCHED
1571
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001572struct update_shares_data {
1573 unsigned long rq_weight[NR_CPUS];
1574};
1575
1576static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1577
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001578static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1579
1580/*
1581 * Calculate and set the cpu's group shares.
1582 */
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001583static void update_group_shares_cpu(struct task_group *tg, int cpu,
1584 unsigned long sd_shares,
1585 unsigned long sd_rq_weight,
1586 struct update_shares_data *usd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001587{
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001588 unsigned long shares, rq_weight;
Peter Zijlstraa50042782009-07-27 14:04:49 +02001589 int boost = 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001590
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001591 rq_weight = usd->rq_weight[cpu];
Peter Zijlstraa50042782009-07-27 14:04:49 +02001592 if (!rq_weight) {
1593 boost = 1;
1594 rq_weight = NICE_0_LOAD;
1595 }
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001596
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001597 /*
Peter Zijlstraa8af7242009-08-21 13:58:54 +02001598 * \Sum_j shares_j * rq_weight_i
1599 * shares_i = -----------------------------
1600 * \Sum_j rq_weight_j
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001601 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001602 shares = (sd_shares * rq_weight) / sd_rq_weight;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001603 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001604
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001605 if (abs(shares - tg->se[cpu]->load.weight) >
1606 sysctl_sched_shares_thresh) {
1607 struct rq *rq = cpu_rq(cpu);
1608 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001609
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001610 spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001611 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
Peter Zijlstraa50042782009-07-27 14:04:49 +02001612 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001613 __set_se_shares(tg->se[cpu], shares);
1614 spin_unlock_irqrestore(&rq->lock, flags);
1615 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001616}
1617
1618/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001619 * Re-compute the task group their per cpu shares over the given domain.
1620 * This needs to be done in a bottom-up fashion because the rq weight of a
1621 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001622 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001623static int tg_shares_up(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001624{
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001625 unsigned long weight, rq_weight = 0, shares = 0;
1626 struct update_shares_data *usd;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001627 struct sched_domain *sd = data;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001628 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001629 int i;
1630
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001631 if (!tg->se[0])
1632 return 0;
1633
1634 local_irq_save(flags);
1635 usd = &__get_cpu_var(update_shares_data);
1636
Rusty Russell758b2cd2008-11-25 02:35:04 +10301637 for_each_cpu(i, sched_domain_span(sd)) {
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001638 weight = tg->cfs_rq[i]->load.weight;
1639 usd->rq_weight[i] = weight;
1640
Ken Chenec4e0e22008-11-18 22:41:57 -08001641 /*
1642 * If there are currently no tasks on the cpu pretend there
1643 * is one of average load so that when a new task gets to
1644 * run here it will not get delayed by group starvation.
1645 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001646 if (!weight)
1647 weight = NICE_0_LOAD;
1648
Ken Chenec4e0e22008-11-18 22:41:57 -08001649 rq_weight += weight;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001650 shares += tg->cfs_rq[i]->shares;
1651 }
1652
1653 if ((!shares && rq_weight) || shares > tg->shares)
1654 shares = tg->shares;
1655
1656 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1657 shares = tg->shares;
1658
Rusty Russell758b2cd2008-11-25 02:35:04 +10301659 for_each_cpu(i, sched_domain_span(sd))
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001660 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1661
1662 local_irq_restore(flags);
Peter Zijlstraeb755802008-08-19 12:33:05 +02001663
1664 return 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001665}
1666
1667/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001668 * Compute the cpu's hierarchical load factor for each task group.
1669 * This needs to be done in a top-down fashion because the load of a child
1670 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001671 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001672static int tg_load_down(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001673{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001674 unsigned long load;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001675 long cpu = (long)data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001676
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001677 if (!tg->parent) {
1678 load = cpu_rq(cpu)->load.weight;
1679 } else {
1680 load = tg->parent->cfs_rq[cpu]->h_load;
1681 load *= tg->cfs_rq[cpu]->shares;
1682 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1683 }
1684
1685 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001686
Peter Zijlstraeb755802008-08-19 12:33:05 +02001687 return 0;
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001688}
1689
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001690static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001691{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001692 s64 elapsed;
1693 u64 now;
1694
1695 if (root_task_group_empty())
1696 return;
1697
1698 now = cpu_clock(raw_smp_processor_id());
1699 elapsed = now - sd->last_update;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001700
1701 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1702 sd->last_update = now;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001703 walk_tg_tree(tg_nop, tg_shares_up, sd);
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001704 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001705}
1706
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001707static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1708{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001709 if (root_task_group_empty())
1710 return;
1711
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001712 spin_unlock(&rq->lock);
1713 update_shares(sd);
1714 spin_lock(&rq->lock);
1715}
1716
Peter Zijlstraeb755802008-08-19 12:33:05 +02001717static void update_h_load(long cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001718{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001719 if (root_task_group_empty())
1720 return;
1721
Peter Zijlstraeb755802008-08-19 12:33:05 +02001722 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001723}
1724
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001725#else
1726
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001727static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001728{
1729}
1730
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001731static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732{
1733}
1734
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001735#endif
1736
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001737#ifdef CONFIG_PREEMPT
1738
Peter Zijlstrab78bb862009-09-15 14:23:18 +02001739static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1740
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001741/*
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001742 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1743 * way at the expense of forcing extra atomic operations in all
1744 * invocations. This assures that the double_lock is acquired using the
1745 * same underlying policy as the spinlock_t on this architecture, which
1746 * reduces latency compared to the unfair variant below. However, it
1747 * also adds more overhead and therefore may reduce throughput.
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001748 */
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001749static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1750 __releases(this_rq->lock)
1751 __acquires(busiest->lock)
1752 __acquires(this_rq->lock)
1753{
1754 spin_unlock(&this_rq->lock);
1755 double_rq_lock(this_rq, busiest);
1756
1757 return 1;
1758}
1759
1760#else
1761/*
1762 * Unfair double_lock_balance: Optimizes throughput at the expense of
1763 * latency by eliminating extra atomic operations when the locks are
1764 * already in proper order on entry. This favors lower cpu-ids and will
1765 * grant the double lock to lower cpus over higher ids under contention,
1766 * regardless of entry order into the function.
1767 */
1768static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001769 __releases(this_rq->lock)
1770 __acquires(busiest->lock)
1771 __acquires(this_rq->lock)
1772{
1773 int ret = 0;
1774
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001775 if (unlikely(!spin_trylock(&busiest->lock))) {
1776 if (busiest < this_rq) {
1777 spin_unlock(&this_rq->lock);
1778 spin_lock(&busiest->lock);
1779 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1780 ret = 1;
1781 } else
1782 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1783 }
1784 return ret;
1785}
1786
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001787#endif /* CONFIG_PREEMPT */
1788
1789/*
1790 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1791 */
1792static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1793{
1794 if (unlikely(!irqs_disabled())) {
1795 /* printk() doesn't work good under rq->lock */
1796 spin_unlock(&this_rq->lock);
1797 BUG_ON(1);
1798 }
1799
1800 return _double_lock_balance(this_rq, busiest);
1801}
1802
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001803static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1804 __releases(busiest->lock)
1805{
1806 spin_unlock(&busiest->lock);
1807 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1808}
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001809#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001810
1811#ifdef CONFIG_FAIR_GROUP_SCHED
1812static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1813{
Vegard Nossum30432092008-06-27 21:35:50 +02001814#ifdef CONFIG_SMP
Ingo Molnar34e83e82008-06-27 15:42:36 +02001815 cfs_rq->shares = shares;
1816#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001817}
1818#endif
1819
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02001820static void calc_load_account_active(struct rq *this_rq);
1821
Ingo Molnardd41f592007-07-09 18:51:59 +02001822#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001823#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001824#include "sched_fair.c"
1825#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001826#ifdef CONFIG_SCHED_DEBUG
1827# include "sched_debug.c"
1828#endif
1829
1830#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04001831#define for_each_class(class) \
1832 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001833
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001834static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001835{
1836 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001837}
1838
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001839static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001840{
1841 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001842}
1843
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001844static void set_load_weight(struct task_struct *p)
1845{
1846 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001847 p->se.load.weight = prio_to_weight[0] * 2;
1848 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1849 return;
1850 }
1851
1852 /*
1853 * SCHED_IDLE tasks get minimal weight:
1854 */
1855 if (p->policy == SCHED_IDLE) {
1856 p->se.load.weight = WEIGHT_IDLEPRIO;
1857 p->se.load.inv_weight = WMULT_IDLEPRIO;
1858 return;
1859 }
1860
1861 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1862 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001863}
1864
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001865static void update_avg(u64 *avg, u64 sample)
1866{
1867 s64 diff = sample - *avg;
1868 *avg += diff >> 3;
1869}
1870
Ingo Molnar8159f872007-08-09 11:16:49 +02001871static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001872{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001873 if (wakeup)
1874 p->se.start_runtime = p->se.sum_exec_runtime;
1875
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001876 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001877 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001878 p->se.on_rq = 1;
1879}
1880
Ingo Molnar69be72c2007-08-09 11:16:49 +02001881static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001882{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001883 if (sleep) {
1884 if (p->se.last_wakeup) {
1885 update_avg(&p->se.avg_overlap,
1886 p->se.sum_exec_runtime - p->se.last_wakeup);
1887 p->se.last_wakeup = 0;
1888 } else {
1889 update_avg(&p->se.avg_wakeup,
1890 sysctl_sched_wakeup_granularity);
1891 }
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001892 }
1893
Ankita Garg46ac22b2008-07-01 14:30:06 +05301894 sched_info_dequeued(p);
Ingo Molnarf02231e2007-08-09 11:16:48 +02001895 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001896 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001897}
1898
1899/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001900 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001901 */
Ingo Molnar14531182007-07-09 18:51:59 +02001902static inline int __normal_prio(struct task_struct *p)
1903{
Ingo Molnardd41f592007-07-09 18:51:59 +02001904 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001905}
1906
1907/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001908 * Calculate the expected normal priority: i.e. priority
1909 * without taking RT-inheritance into account. Might be
1910 * boosted by interactivity modifiers. Changes upon fork,
1911 * setprio syscalls, and whenever the interactivity
1912 * estimator recalculates.
1913 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001914static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001915{
1916 int prio;
1917
Ingo Molnare05606d2007-07-09 18:51:59 +02001918 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001919 prio = MAX_RT_PRIO-1 - p->rt_priority;
1920 else
1921 prio = __normal_prio(p);
1922 return prio;
1923}
1924
1925/*
1926 * Calculate the current priority, i.e. the priority
1927 * taken into account by the scheduler. This value might
1928 * be boosted by RT tasks, or might be boosted by
1929 * interactivity modifiers. Will be RT if the task got
1930 * RT-boosted. If not then it returns p->normal_prio.
1931 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001932static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001933{
1934 p->normal_prio = normal_prio(p);
1935 /*
1936 * If we are RT tasks or we were boosted to RT priority,
1937 * keep the priority unchanged. Otherwise, update priority
1938 * to the normal priority:
1939 */
1940 if (!rt_prio(p->prio))
1941 return p->normal_prio;
1942 return p->prio;
1943}
1944
1945/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001946 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001948static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001950 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001951 rq->nr_uninterruptible--;
1952
Ingo Molnar8159f872007-08-09 11:16:49 +02001953 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001954 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955}
1956
1957/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 * deactivate_task - remove a task from the runqueue.
1959 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001960static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001962 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001963 rq->nr_uninterruptible++;
1964
Ingo Molnar69be72c2007-08-09 11:16:49 +02001965 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001966 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967}
1968
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969/**
1970 * task_curr - is this task currently executing on a CPU?
1971 * @p: the task in question.
1972 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001973inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974{
1975 return cpu_curr(task_cpu(p)) == p;
1976}
1977
Ingo Molnardd41f592007-07-09 18:51:59 +02001978static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1979{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001980 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001981#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001982 /*
1983 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1984 * successfuly executed on another CPU. We must ensure that updates of
1985 * per-task data have been completed by this moment.
1986 */
1987 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001988 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001989#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001990}
1991
Steven Rostedtcb469842008-01-25 21:08:22 +01001992static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1993 const struct sched_class *prev_class,
1994 int oldprio, int running)
1995{
1996 if (prev_class != p->sched_class) {
1997 if (prev_class->switched_from)
1998 prev_class->switched_from(rq, p, running);
1999 p->sched_class->switched_to(rq, p, running);
2000 } else
2001 p->sched_class->prio_changed(rq, p, oldprio, running);
2002}
2003
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004#ifdef CONFIG_SMP
Ingo Molnarcc367732007-10-15 17:00:18 +02002005/*
2006 * Is this task likely cache-hot:
2007 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002008static int
Ingo Molnarcc367732007-10-15 17:00:18 +02002009task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2010{
2011 s64 delta;
2012
Ingo Molnarf540a602008-03-15 17:10:34 +01002013 /*
2014 * Buddy candidates are cache hot:
2015 */
Peter Zijlstra47932412008-11-04 21:25:09 +01002016 if (sched_feat(CACHE_HOT_BUDDY) &&
2017 (&p->se == cfs_rq_of(&p->se)->next ||
2018 &p->se == cfs_rq_of(&p->se)->last))
Ingo Molnarf540a602008-03-15 17:10:34 +01002019 return 1;
2020
Ingo Molnarcc367732007-10-15 17:00:18 +02002021 if (p->sched_class != &fair_sched_class)
2022 return 0;
2023
Ingo Molnar6bc16652007-10-15 17:00:18 +02002024 if (sysctl_sched_migration_cost == -1)
2025 return 1;
2026 if (sysctl_sched_migration_cost == 0)
2027 return 0;
2028
Ingo Molnarcc367732007-10-15 17:00:18 +02002029 delta = now - p->se.exec_start;
2030
2031 return delta < (s64)sysctl_sched_migration_cost;
2032}
2033
2034
Ingo Molnardd41f592007-07-09 18:51:59 +02002035void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02002036{
Ingo Molnardd41f592007-07-09 18:51:59 +02002037 int old_cpu = task_cpu(p);
2038 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002039 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2040 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02002041 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002042
2043 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002044
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08002045 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01002046
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002047#ifdef CONFIG_SCHEDSTATS
2048 if (p->se.wait_start)
2049 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002050 if (p->se.sleep_start)
2051 p->se.sleep_start -= clock_offset;
2052 if (p->se.block_start)
2053 p->se.block_start -= clock_offset;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002054#endif
Ingo Molnarcc367732007-10-15 17:00:18 +02002055 if (old_cpu != new_cpu) {
Ingo Molnar6c594c22008-12-14 12:34:15 +01002056 p->se.nr_migrations++;
Paul Mackerras23a185c2009-02-09 22:42:47 +11002057 new_rq->nr_migrations_in++;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002058#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002059 if (task_hot(p, old_rq->clock, NULL))
2060 schedstat_inc(p, se.nr_forced2_migrations);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002061#endif
Peter Zijlstrae5289d42009-06-19 13:22:51 +02002062 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2063 1, 1, NULL, 0);
Ingo Molnar6c594c22008-12-14 12:34:15 +01002064 }
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002065 p->se.vruntime -= old_cfsrq->min_vruntime -
2066 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02002067
2068 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02002069}
2070
Ingo Molnar70b97a72006-07-03 00:25:42 -07002071struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002072 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073
Ingo Molnar36c8b582006-07-03 00:25:41 -07002074 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 int dest_cpu;
2076
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002078};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
2080/*
2081 * The task's runqueue lock must be held.
2082 * Returns true if you have to wait for migration thread.
2083 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002084static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002085migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002087 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
2089 /*
2090 * If the task is not on a runqueue (and not running), then
2091 * it is sufficient to simply update the task's cpu field.
2092 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002093 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 set_task_cpu(p, dest_cpu);
2095 return 0;
2096 }
2097
2098 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 req->task = p;
2100 req->dest_cpu = dest_cpu;
2101 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002102
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 return 1;
2104}
2105
2106/*
Markus Metzgera26b89f2009-04-03 16:43:34 +02002107 * wait_task_context_switch - wait for a thread to complete at least one
2108 * context switch.
2109 *
2110 * @p must not be current.
2111 */
2112void wait_task_context_switch(struct task_struct *p)
2113{
2114 unsigned long nvcsw, nivcsw, flags;
2115 int running;
2116 struct rq *rq;
2117
2118 nvcsw = p->nvcsw;
2119 nivcsw = p->nivcsw;
2120 for (;;) {
2121 /*
2122 * The runqueue is assigned before the actual context
2123 * switch. We need to take the runqueue lock.
2124 *
2125 * We could check initially without the lock but it is
2126 * very likely that we need to take the lock in every
2127 * iteration.
2128 */
2129 rq = task_rq_lock(p, &flags);
2130 running = task_running(rq, p);
2131 task_rq_unlock(rq, &flags);
2132
2133 if (likely(!running))
2134 break;
2135 /*
2136 * The switch count is incremented before the actual
2137 * context switch. We thus wait for two switches to be
2138 * sure at least one completed.
2139 */
2140 if ((p->nvcsw - nvcsw) > 1)
2141 break;
2142 if ((p->nivcsw - nivcsw) > 1)
2143 break;
2144
2145 cpu_relax();
2146 }
2147}
2148
2149/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 * wait_task_inactive - wait for a thread to unschedule.
2151 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07002152 * If @match_state is nonzero, it's the @p->state value just checked and
2153 * not expected to change. If it changes, i.e. @p might have woken up,
2154 * then return zero. When we succeed in waiting for @p to be off its CPU,
2155 * we return a positive number (its total switch count). If a second call
2156 * a short while later returns the same number, the caller can be sure that
2157 * @p has remained unscheduled the whole time.
2158 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 * The caller must ensure that the task *will* unschedule sometime soon,
2160 * else this function might spin for a *long* time. This function can't
2161 * be called with interrupts off, or it may introduce deadlock with
2162 * smp_call_function() if an IPI is sent by the same process we are
2163 * waiting to become inactive.
2164 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002165unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166{
2167 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002168 int running, on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002169 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002170 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171
Andi Kleen3a5c3592007-10-15 17:00:14 +02002172 for (;;) {
2173 /*
2174 * We do the initial early heuristics without holding
2175 * any task-queue locks at all. We'll only try to get
2176 * the runqueue lock when things look like they will
2177 * work out!
2178 */
2179 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002180
Andi Kleen3a5c3592007-10-15 17:00:14 +02002181 /*
2182 * If the task is actively running on another CPU
2183 * still, just relax and busy-wait without holding
2184 * any locks.
2185 *
2186 * NOTE! Since we don't hold any locks, it's not
2187 * even sure that "rq" stays as the right runqueue!
2188 * But we don't care, since "task_running()" will
2189 * return false if the runqueue has changed and p
2190 * is actually now running somewhere else!
2191 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002192 while (task_running(rq, p)) {
2193 if (match_state && unlikely(p->state != match_state))
2194 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02002195 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07002196 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002197
Andi Kleen3a5c3592007-10-15 17:00:14 +02002198 /*
2199 * Ok, time to look more closely! We need the rq
2200 * lock now, to be *sure*. If we're wrong, we'll
2201 * just go back and repeat.
2202 */
2203 rq = task_rq_lock(p, &flags);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002204 trace_sched_wait_task(rq, p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02002205 running = task_running(rq, p);
2206 on_rq = p->se.on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002207 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07002208 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07002209 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Andi Kleen3a5c3592007-10-15 17:00:14 +02002210 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002211
Andi Kleen3a5c3592007-10-15 17:00:14 +02002212 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07002213 * If it changed from the expected state, bail out now.
2214 */
2215 if (unlikely(!ncsw))
2216 break;
2217
2218 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02002219 * Was it really running after all now that we
2220 * checked with the proper locks actually held?
2221 *
2222 * Oops. Go back and try again..
2223 */
2224 if (unlikely(running)) {
2225 cpu_relax();
2226 continue;
2227 }
2228
2229 /*
2230 * It's not enough that it's not actively running,
2231 * it must be off the runqueue _entirely_, and not
2232 * preempted!
2233 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00002234 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02002235 * running right now), it's preempted, and we should
2236 * yield - it could be a while.
2237 */
2238 if (unlikely(on_rq)) {
2239 schedule_timeout_uninterruptible(1);
2240 continue;
2241 }
2242
2243 /*
2244 * Ahh, all good. It wasn't running, and it wasn't
2245 * runnable, which means that it will never become
2246 * running in the future either. We're all done!
2247 */
2248 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07002250
2251 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252}
2253
2254/***
2255 * kick_process - kick a running thread to enter/exit the kernel
2256 * @p: the to-be-kicked thread
2257 *
2258 * Cause a process which is running on another CPU to enter
2259 * kernel-mode, without any delay. (to get signals handled.)
2260 *
2261 * NOTE: this function doesnt have to take the runqueue lock,
2262 * because all it wants to ensure is that the remote task enters
2263 * the kernel. If the IPI races and the task has been migrated
2264 * to another CPU then no harm is done and the purpose has been
2265 * achieved as well.
2266 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002267void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268{
2269 int cpu;
2270
2271 preempt_disable();
2272 cpu = task_cpu(p);
2273 if ((cpu != smp_processor_id()) && task_curr(p))
2274 smp_send_reschedule(cpu);
2275 preempt_enable();
2276}
Rusty Russellb43e3522009-06-12 22:27:00 -06002277EXPORT_SYMBOL_GPL(kick_process);
Nick Piggin476d1392005-06-25 14:57:29 -07002278#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279
Thomas Gleixner0793a612008-12-04 20:12:29 +01002280/**
2281 * task_oncpu_function_call - call a function on the cpu on which a task runs
2282 * @p: the task to evaluate
2283 * @func: the function to be called
2284 * @info: the function call argument
2285 *
2286 * Calls the function @func when the task is currently running. This might
2287 * be on the current CPU, which just calls the function directly
2288 */
2289void task_oncpu_function_call(struct task_struct *p,
2290 void (*func) (void *info), void *info)
2291{
2292 int cpu;
2293
2294 preempt_disable();
2295 cpu = task_cpu(p);
2296 if (task_curr(p))
2297 smp_call_function_single(cpu, func, info, 1);
2298 preempt_enable();
2299}
2300
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301/***
2302 * try_to_wake_up - wake up a thread
2303 * @p: the to-be-woken-up thread
2304 * @state: the mask of task states that can be woken
2305 * @sync: do a synchronous wakeup?
2306 *
2307 * Put it on the run-queue if it's not already there. The "current"
2308 * thread is always on the run-queue (except when the actual
2309 * re-schedule is in progress), and as such you're allowed to do
2310 * the simpler "current->state = TASK_RUNNING" to mark yourself
2311 * runnable without the overhead of this.
2312 *
2313 * returns failure only if the task is already active.
2314 */
Peter Zijlstra7d478722009-09-14 19:55:44 +02002315static int try_to_wake_up(struct task_struct *p, unsigned int state,
2316 int wake_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317{
Ingo Molnarcc367732007-10-15 17:00:18 +02002318 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002320 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321
Ingo Molnarb85d0662008-03-16 20:03:22 +01002322 if (!sched_feat(SYNC_WAKEUPS))
Peter Zijlstra7d478722009-09-14 19:55:44 +02002323 wake_flags &= ~WF_SYNC;
Ingo Molnarb85d0662008-03-16 20:03:22 +01002324
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002325 this_cpu = get_cpu();
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002326
Linus Torvalds04e2f172008-02-23 18:05:03 -08002327 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 rq = task_rq_lock(p, &flags);
Mike Galbraith03e89e42008-12-16 08:45:30 +01002329 update_rq_clock(rq);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002330 if (!(p->state & state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 goto out;
2332
Ingo Molnardd41f592007-07-09 18:51:59 +02002333 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 goto out_running;
2335
2336 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002337 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338
2339#ifdef CONFIG_SMP
2340 if (unlikely(task_running(rq, p)))
2341 goto out_activate;
2342
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002343 /*
2344 * In order to handle concurrent wakeups and release the rq->lock
2345 * we put the task in TASK_WAKING state.
Ingo Molnareb240732009-09-16 21:09:13 +02002346 *
2347 * First fix up the nr_uninterruptible count:
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002348 */
Ingo Molnareb240732009-09-16 21:09:13 +02002349 if (task_contributes_to_load(p))
2350 rq->nr_uninterruptible--;
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002351 p->state = TASK_WAKING;
2352 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353
Peter Zijlstra7d478722009-09-14 19:55:44 +02002354 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002355 if (cpu != orig_cpu)
2356 set_task_cpu(p, cpu);
2357
2358 rq = task_rq_lock(p, &flags);
2359 WARN_ON(p->state != TASK_WAKING);
2360 cpu = task_cpu(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
Gregory Haskinse7693a32008-01-25 21:08:09 +01002362#ifdef CONFIG_SCHEDSTATS
2363 schedstat_inc(rq, ttwu_count);
2364 if (cpu == this_cpu)
2365 schedstat_inc(rq, ttwu_local);
2366 else {
2367 struct sched_domain *sd;
2368 for_each_domain(this_cpu, sd) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302369 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Gregory Haskinse7693a32008-01-25 21:08:09 +01002370 schedstat_inc(sd, ttwu_wake_remote);
2371 break;
2372 }
2373 }
2374 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002375#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002376
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377out_activate:
2378#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002379 schedstat_inc(p, se.nr_wakeups);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002380 if (wake_flags & WF_SYNC)
Ingo Molnarcc367732007-10-15 17:00:18 +02002381 schedstat_inc(p, se.nr_wakeups_sync);
2382 if (orig_cpu != cpu)
2383 schedstat_inc(p, se.nr_wakeups_migrate);
2384 if (cpu == this_cpu)
2385 schedstat_inc(p, se.nr_wakeups_local);
2386 else
2387 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnardd41f592007-07-09 18:51:59 +02002388 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 success = 1;
2390
Peter Zijlstra831451a2009-01-14 12:39:18 +01002391 /*
2392 * Only attribute actual wakeups done by this task.
2393 */
2394 if (!in_interrupt()) {
2395 struct sched_entity *se = &current->se;
2396 u64 sample = se->sum_exec_runtime;
2397
2398 if (se->last_wakeup)
2399 sample -= se->last_wakeup;
2400 else
2401 sample -= se->start_runtime;
2402 update_avg(&se->avg_wakeup, sample);
2403
2404 se->last_wakeup = se->sum_exec_runtime;
2405 }
2406
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407out_running:
Peter Zijlstra468a15b2008-12-16 08:07:03 +01002408 trace_sched_wakeup(rq, p, success);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002409 check_preempt_curr(rq, p, wake_flags);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002410
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002412#ifdef CONFIG_SMP
2413 if (p->sched_class->task_wake_up)
2414 p->sched_class->task_wake_up(rq, p);
2415#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416out:
2417 task_rq_unlock(rq, &flags);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002418 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
2420 return success;
2421}
2422
David Howells50fa6102009-04-28 15:01:38 +01002423/**
2424 * wake_up_process - Wake up a specific process
2425 * @p: The process to be woken up.
2426 *
2427 * Attempt to wake up the nominated process and move it to the set of runnable
2428 * processes. Returns 1 if the process was woken up, 0 if it was already
2429 * running.
2430 *
2431 * It may be assumed that this function implies a write memory barrier before
2432 * changing the task state if and only if any tasks are woken up.
2433 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002434int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002436 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438EXPORT_SYMBOL(wake_up_process);
2439
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002440int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441{
2442 return try_to_wake_up(p, state, 0);
2443}
2444
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445/*
2446 * Perform scheduler related setup for a newly forked process p.
2447 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002448 *
2449 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002451static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452{
Ingo Molnardd41f592007-07-09 18:51:59 +02002453 p->se.exec_start = 0;
2454 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002455 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002456 p->se.nr_migrations = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002457 p->se.last_wakeup = 0;
2458 p->se.avg_overlap = 0;
Peter Zijlstra831451a2009-01-14 12:39:18 +01002459 p->se.start_runtime = 0;
2460 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02002461 p->se.avg_running = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002462
2463#ifdef CONFIG_SCHEDSTATS
Lucas De Marchi77935272009-07-09 13:57:20 +02002464 p->se.wait_start = 0;
2465 p->se.wait_max = 0;
2466 p->se.wait_count = 0;
2467 p->se.wait_sum = 0;
2468
2469 p->se.sleep_start = 0;
2470 p->se.sleep_max = 0;
2471 p->se.sum_sleep_runtime = 0;
2472
2473 p->se.block_start = 0;
2474 p->se.block_max = 0;
2475 p->se.exec_max = 0;
2476 p->se.slice_max = 0;
2477
2478 p->se.nr_migrations_cold = 0;
2479 p->se.nr_failed_migrations_affine = 0;
2480 p->se.nr_failed_migrations_running = 0;
2481 p->se.nr_failed_migrations_hot = 0;
2482 p->se.nr_forced_migrations = 0;
2483 p->se.nr_forced2_migrations = 0;
2484
2485 p->se.nr_wakeups = 0;
2486 p->se.nr_wakeups_sync = 0;
2487 p->se.nr_wakeups_migrate = 0;
2488 p->se.nr_wakeups_local = 0;
2489 p->se.nr_wakeups_remote = 0;
2490 p->se.nr_wakeups_affine = 0;
2491 p->se.nr_wakeups_affine_attempts = 0;
2492 p->se.nr_wakeups_passive = 0;
2493 p->se.nr_wakeups_idle = 0;
2494
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002495#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002496
Peter Zijlstrafa717062008-01-25 21:08:27 +01002497 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002498 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002499 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002500
Avi Kivitye107be32007-07-26 13:40:43 +02002501#ifdef CONFIG_PREEMPT_NOTIFIERS
2502 INIT_HLIST_HEAD(&p->preempt_notifiers);
2503#endif
2504
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 /*
2506 * We mark the process as running here, but have not actually
2507 * inserted it onto the runqueue yet. This guarantees that
2508 * nobody will actually run it, and a signal or other external
2509 * event cannot wake it up and insert it on the runqueue either.
2510 */
2511 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002512}
2513
2514/*
2515 * fork()/clone()-time setup:
2516 */
2517void sched_fork(struct task_struct *p, int clone_flags)
2518{
2519 int cpu = get_cpu();
2520
2521 __sched_fork(p);
2522
Ingo Molnarb29739f2006-06-27 02:54:51 -07002523 /*
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002524 * Make sure we do not leak PI boosting priority to the child.
Ingo Molnarb29739f2006-06-27 02:54:51 -07002525 */
2526 p->prio = current->normal_prio;
Lennart Poetteringca94c442009-06-15 17:17:47 +02002527
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002528 /*
2529 * Revert to default priority/policy on fork if requested.
2530 */
2531 if (unlikely(p->sched_reset_on_fork)) {
2532 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2533 p->policy = SCHED_NORMAL;
2534
2535 if (p->normal_prio < DEFAULT_PRIO)
2536 p->prio = DEFAULT_PRIO;
2537
Mike Galbraith6c697bd2009-06-17 10:48:02 +02002538 if (PRIO_TO_NICE(p->static_prio) < 0) {
2539 p->static_prio = NICE_TO_PRIO(0);
2540 set_load_weight(p);
2541 }
2542
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002543 /*
2544 * We don't need the reset flag anymore after the fork. It has
2545 * fulfilled its duty:
2546 */
2547 p->sched_reset_on_fork = 0;
2548 }
Lennart Poetteringca94c442009-06-15 17:17:47 +02002549
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002550 if (!rt_prio(p->prio))
2551 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002552
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002553#ifdef CONFIG_SMP
2554 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2555#endif
2556 set_task_cpu(p, cpu);
2557
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002558#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002559 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002560 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002562#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002563 p->oncpu = 0;
2564#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002566 /* Want to start with kernel preemption disabled. */
Al Viroa1261f542005-11-13 16:06:55 -08002567 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568#endif
Gregory Haskins917b6272008-12-29 09:39:53 -05002569 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2570
Nick Piggin476d1392005-06-25 14:57:29 -07002571 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572}
2573
2574/*
2575 * wake_up_new_task - wake up a newly created task for the first time.
2576 *
2577 * This function will do some initial scheduler statistics housekeeping
2578 * that must be done for every newly created context, then puts the task
2579 * on the runqueue and wakes it.
2580 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002581void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582{
2583 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002584 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585
2586 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002588 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589
2590 p->prio = effective_prio(p);
2591
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002592 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002593 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002596 * Let the scheduling class do new task startup
2597 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002599 p->sched_class->task_new(rq, p);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002600 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002601 }
Ingo Molnarc71dd422008-12-19 01:09:51 +01002602 trace_sched_wakeup_new(rq, p, 1);
Peter Zijlstraa7558e02009-09-14 20:02:34 +02002603 check_preempt_curr(rq, p, WF_FORK);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002604#ifdef CONFIG_SMP
2605 if (p->sched_class->task_wake_up)
2606 p->sched_class->task_wake_up(rq, p);
2607#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002608 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609}
2610
Avi Kivitye107be32007-07-26 13:40:43 +02002611#ifdef CONFIG_PREEMPT_NOTIFIERS
2612
2613/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00002614 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07002615 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002616 */
2617void preempt_notifier_register(struct preempt_notifier *notifier)
2618{
2619 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2620}
2621EXPORT_SYMBOL_GPL(preempt_notifier_register);
2622
2623/**
2624 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002625 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002626 *
2627 * This is safe to call from within a preemption notifier.
2628 */
2629void preempt_notifier_unregister(struct preempt_notifier *notifier)
2630{
2631 hlist_del(&notifier->link);
2632}
2633EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2634
2635static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2636{
2637 struct preempt_notifier *notifier;
2638 struct hlist_node *node;
2639
2640 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2641 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2642}
2643
2644static void
2645fire_sched_out_preempt_notifiers(struct task_struct *curr,
2646 struct task_struct *next)
2647{
2648 struct preempt_notifier *notifier;
2649 struct hlist_node *node;
2650
2651 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2652 notifier->ops->sched_out(notifier, next);
2653}
2654
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002655#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002656
2657static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2658{
2659}
2660
2661static void
2662fire_sched_out_preempt_notifiers(struct task_struct *curr,
2663 struct task_struct *next)
2664{
2665}
2666
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002667#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002668
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002670 * prepare_task_switch - prepare to switch tasks
2671 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002672 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002673 * @next: the task we are going to switch to.
2674 *
2675 * This is called with the rq lock held and interrupts off. It must
2676 * be paired with a subsequent finish_task_switch after the context
2677 * switch.
2678 *
2679 * prepare_task_switch sets up locking and calls architecture specific
2680 * hooks.
2681 */
Avi Kivitye107be32007-07-26 13:40:43 +02002682static inline void
2683prepare_task_switch(struct rq *rq, struct task_struct *prev,
2684 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002685{
Avi Kivitye107be32007-07-26 13:40:43 +02002686 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002687 prepare_lock_switch(rq, next);
2688 prepare_arch_switch(next);
2689}
2690
2691/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002693 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 * @prev: the thread we just switched away from.
2695 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002696 * finish_task_switch must be called after the context switch, paired
2697 * with a prepare_task_switch call before the context switch.
2698 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2699 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 *
2701 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002702 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 * with the lock held can cause deadlocks; see schedule() for
2704 * details.)
2705 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002706static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 __releases(rq->lock)
2708{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002710 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711
2712 rq->prev_mm = NULL;
2713
2714 /*
2715 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002716 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002717 * schedule one last time. The schedule call will never return, and
2718 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002719 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 * still held, otherwise prev could be scheduled on another cpu, die
2721 * there before we look at prev->state, and then the reference would
2722 * be dropped twice.
2723 * Manfred Spraul <manfred@colorfullife.com>
2724 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002725 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002726 finish_arch_switch(prev);
Thomas Gleixner0793a612008-12-04 20:12:29 +01002727 perf_counter_task_sched_in(current, cpu_of(rq));
Nick Piggin4866cde2005-06-25 14:57:23 -07002728 finish_lock_switch(rq, prev);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002729
Avi Kivitye107be32007-07-26 13:40:43 +02002730 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 if (mm)
2732 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002733 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002734 /*
2735 * Remove function-return probe instances associated with this
2736 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002737 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002738 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002740 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741}
2742
Gregory Haskins3f029d32009-07-29 11:08:47 -04002743#ifdef CONFIG_SMP
2744
2745/* assumes rq->lock is held */
2746static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2747{
2748 if (prev->sched_class->pre_schedule)
2749 prev->sched_class->pre_schedule(rq, prev);
2750}
2751
2752/* rq->lock is NOT held, but preemption is disabled */
2753static inline void post_schedule(struct rq *rq)
2754{
2755 if (rq->post_schedule) {
2756 unsigned long flags;
2757
2758 spin_lock_irqsave(&rq->lock, flags);
2759 if (rq->curr->sched_class->post_schedule)
2760 rq->curr->sched_class->post_schedule(rq);
2761 spin_unlock_irqrestore(&rq->lock, flags);
2762
2763 rq->post_schedule = 0;
2764 }
2765}
2766
2767#else
2768
2769static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2770{
2771}
2772
2773static inline void post_schedule(struct rq *rq)
2774{
2775}
2776
2777#endif
2778
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779/**
2780 * schedule_tail - first thing a freshly forked thread must call.
2781 * @prev: the thread we just switched away from.
2782 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002783asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 __releases(rq->lock)
2785{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002786 struct rq *rq = this_rq();
2787
Nick Piggin4866cde2005-06-25 14:57:23 -07002788 finish_task_switch(rq, prev);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002789
Gregory Haskins3f029d32009-07-29 11:08:47 -04002790 /*
2791 * FIXME: do we need to worry about rq being invalidated by the
2792 * task_switch?
2793 */
2794 post_schedule(rq);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002795
Nick Piggin4866cde2005-06-25 14:57:23 -07002796#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2797 /* In this case, finish_task_switch does not reenable preemption */
2798 preempt_enable();
2799#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002801 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802}
2803
2804/*
2805 * context_switch - switch to the new MM and the new
2806 * thread's register state.
2807 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002808static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002809context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002810 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811{
Ingo Molnardd41f592007-07-09 18:51:59 +02002812 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813
Avi Kivitye107be32007-07-26 13:40:43 +02002814 prepare_task_switch(rq, prev, next);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002815 trace_sched_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002816 mm = next->mm;
2817 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002818 /*
2819 * For paravirt, this is coupled with an exit in switch_to to
2820 * combine the page table reload and the switch backend into
2821 * one hypercall.
2822 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08002823 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01002824
Ingo Molnardd41f592007-07-09 18:51:59 +02002825 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 next->active_mm = oldmm;
2827 atomic_inc(&oldmm->mm_count);
2828 enter_lazy_tlb(oldmm, next);
2829 } else
2830 switch_mm(oldmm, mm, next);
2831
Ingo Molnardd41f592007-07-09 18:51:59 +02002832 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 rq->prev_mm = oldmm;
2835 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002836 /*
2837 * Since the runqueue lock will be released by the next
2838 * task (which is an invalid locking op but in the case
2839 * of the scheduler it's an obvious special-case), so we
2840 * do an early lockdep release here:
2841 */
2842#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002843 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002844#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845
2846 /* Here we just switch the register state and the stack. */
2847 switch_to(prev, next, prev);
2848
Ingo Molnardd41f592007-07-09 18:51:59 +02002849 barrier();
2850 /*
2851 * this_rq must be evaluated again because prev may have moved
2852 * CPUs since it called schedule(), thus the 'rq' on its stack
2853 * frame will be invalid.
2854 */
2855 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856}
2857
2858/*
2859 * nr_running, nr_uninterruptible and nr_context_switches:
2860 *
2861 * externally visible scheduler statistics: current number of runnable
2862 * threads, current number of uninterruptible-sleeping threads, total
2863 * number of context switches performed since bootup.
2864 */
2865unsigned long nr_running(void)
2866{
2867 unsigned long i, sum = 0;
2868
2869 for_each_online_cpu(i)
2870 sum += cpu_rq(i)->nr_running;
2871
2872 return sum;
2873}
2874
2875unsigned long nr_uninterruptible(void)
2876{
2877 unsigned long i, sum = 0;
2878
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002879 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 sum += cpu_rq(i)->nr_uninterruptible;
2881
2882 /*
2883 * Since we read the counters lockless, it might be slightly
2884 * inaccurate. Do not allow it to go below zero though:
2885 */
2886 if (unlikely((long)sum < 0))
2887 sum = 0;
2888
2889 return sum;
2890}
2891
2892unsigned long long nr_context_switches(void)
2893{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002894 int i;
2895 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002897 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 sum += cpu_rq(i)->nr_switches;
2899
2900 return sum;
2901}
2902
2903unsigned long nr_iowait(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 += atomic_read(&cpu_rq(i)->nr_iowait);
2909
2910 return sum;
2911}
2912
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002913/* Variables and functions for calc_load */
2914static atomic_long_t calc_load_tasks;
2915static unsigned long calc_load_update;
2916unsigned long avenrun[3];
2917EXPORT_SYMBOL(avenrun);
2918
Thomas Gleixner2d024942009-05-02 20:08:52 +02002919/**
2920 * get_avenrun - get the load average array
2921 * @loads: pointer to dest load array
2922 * @offset: offset to add
2923 * @shift: shift count to shift the result left
2924 *
2925 * These values are estimates at best, so no need for locking.
2926 */
2927void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2928{
2929 loads[0] = (avenrun[0] + offset) << shift;
2930 loads[1] = (avenrun[1] + offset) << shift;
2931 loads[2] = (avenrun[2] + offset) << shift;
2932}
2933
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002934static unsigned long
2935calc_load(unsigned long load, unsigned long exp, unsigned long active)
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002936{
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002937 load *= exp;
2938 load += active * (FIXED_1 - exp);
2939 return load >> FSHIFT;
2940}
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002941
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002942/*
2943 * calc_load - update the avenrun load estimates 10 ticks after the
2944 * CPUs have updated calc_load_tasks.
2945 */
2946void calc_global_load(void)
2947{
2948 unsigned long upd = calc_load_update + 10;
2949 long active;
2950
2951 if (time_before(jiffies, upd))
2952 return;
2953
2954 active = atomic_long_read(&calc_load_tasks);
2955 active = active > 0 ? active * FIXED_1 : 0;
2956
2957 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2958 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2959 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2960
2961 calc_load_update += LOAD_FREQ;
2962}
2963
2964/*
2965 * Either called from update_cpu_load() or from a cpu going idle
2966 */
2967static void calc_load_account_active(struct rq *this_rq)
2968{
2969 long nr_active, delta;
2970
2971 nr_active = this_rq->nr_running;
2972 nr_active += (long) this_rq->nr_uninterruptible;
2973
2974 if (nr_active != this_rq->calc_load_active) {
2975 delta = nr_active - this_rq->calc_load_active;
2976 this_rq->calc_load_active = nr_active;
2977 atomic_long_add(delta, &calc_load_tasks);
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002978 }
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002979}
2980
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981/*
Paul Mackerras23a185c2009-02-09 22:42:47 +11002982 * Externally visible per-cpu scheduler statistics:
Paul Mackerras23a185c2009-02-09 22:42:47 +11002983 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2984 */
Paul Mackerras23a185c2009-02-09 22:42:47 +11002985u64 cpu_nr_migrations(int cpu)
2986{
2987 return cpu_rq(cpu)->nr_migrations_in;
2988}
2989
2990/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002991 * Update rq->cpu_load[] statistics. This function is usually called every
2992 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002993 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002994static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002995{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002996 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002997 int i, scale;
2998
2999 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02003000
3001 /* Update our load: */
3002 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3003 unsigned long old_load, new_load;
3004
3005 /* scale is effectively 1 << i now, and >> i divides by scale */
3006
3007 old_load = this_rq->cpu_load[i];
3008 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02003009 /*
3010 * Round up the averaging division if load is increasing. This
3011 * prevents us from getting stuck on 9 if the load is 10, for
3012 * example.
3013 */
3014 if (new_load > old_load)
3015 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003016 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3017 }
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02003018
3019 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3020 this_rq->calc_load_update += LOAD_FREQ;
3021 calc_load_account_active(this_rq);
3022 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003023}
3024
Ingo Molnardd41f592007-07-09 18:51:59 +02003025#ifdef CONFIG_SMP
3026
Ingo Molnar48f24c42006-07-03 00:25:40 -07003027/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028 * double_rq_lock - safely lock two runqueues
3029 *
3030 * Note this does not disable interrupts like task_rq_lock,
3031 * you need to do so manually before calling.
3032 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003033static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 __acquires(rq1->lock)
3035 __acquires(rq2->lock)
3036{
Kirill Korotaev054b9102006-12-10 02:20:11 -08003037 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 if (rq1 == rq2) {
3039 spin_lock(&rq1->lock);
3040 __acquire(rq2->lock); /* Fake it out ;) */
3041 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003042 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 spin_lock(&rq1->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003044 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 } else {
3046 spin_lock(&rq2->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003047 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048 }
3049 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003050 update_rq_clock(rq1);
3051 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052}
3053
3054/*
3055 * double_rq_unlock - safely unlock two runqueues
3056 *
3057 * Note this does not restore interrupts like task_rq_unlock,
3058 * you need to do so manually after calling.
3059 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003060static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 __releases(rq1->lock)
3062 __releases(rq2->lock)
3063{
3064 spin_unlock(&rq1->lock);
3065 if (rq1 != rq2)
3066 spin_unlock(&rq2->lock);
3067 else
3068 __release(rq2->lock);
3069}
3070
3071/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 * If dest_cpu is allowed for this process, migrate the task to it.
3073 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003074 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 * the cpu_allowed mask is restored.
3076 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003077static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003079 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003081 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082
3083 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e22008-11-25 02:35:14 +10303084 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
Max Krasnyanskye761b772008-07-15 04:43:49 -07003085 || unlikely(!cpu_active(dest_cpu)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086 goto out;
3087
3088 /* force the process onto the specified CPU */
3089 if (migrate_task(p, dest_cpu, &req)) {
3090 /* Need to wait for migration thread (might exit: take ref). */
3091 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003092
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 get_task_struct(mt);
3094 task_rq_unlock(rq, &flags);
3095 wake_up_process(mt);
3096 put_task_struct(mt);
3097 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07003098
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 return;
3100 }
3101out:
3102 task_rq_unlock(rq, &flags);
3103}
3104
3105/*
Nick Piggin476d1392005-06-25 14:57:29 -07003106 * sched_exec - execve() is a valuable balancing opportunity, because at
3107 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 */
3109void sched_exec(void)
3110{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111 int new_cpu, this_cpu = get_cpu();
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02003112 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003114 if (new_cpu != this_cpu)
3115 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116}
3117
3118/*
3119 * pull_task - move a task from a remote runqueue to the local runqueue.
3120 * Both runqueues must be locked.
3121 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003122static void pull_task(struct rq *src_rq, struct task_struct *p,
3123 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003125 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003127 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 /*
3129 * Note that idle threads have a prio of MAX_PRIO, for this test
3130 * to be always true for them.
3131 */
Peter Zijlstra15afe092008-09-20 23:38:02 +02003132 check_preempt_curr(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133}
3134
3135/*
3136 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3137 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08003138static
Ingo Molnar70b97a72006-07-03 00:25:42 -07003139int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003140 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003141 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142{
Luis Henriques708dc512009-03-16 19:59:02 +00003143 int tsk_cache_hot = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 /*
3145 * We do not migrate tasks that are:
3146 * 1) running (obviously), or
3147 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3148 * 3) are cache-hot on their current CPU.
3149 */
Rusty Russell96f874e22008-11-25 02:35:14 +10303150 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003151 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003153 }
Nick Piggin81026792005-06-25 14:57:07 -07003154 *all_pinned = 0;
3155
Ingo Molnarcc367732007-10-15 17:00:18 +02003156 if (task_running(rq, p)) {
3157 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07003158 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003159 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160
Ingo Molnarda84d962007-10-15 17:00:18 +02003161 /*
3162 * Aggressive migration if:
3163 * 1) task is cache cold, or
3164 * 2) too many balance attempts have failed.
3165 */
3166
Luis Henriques708dc512009-03-16 19:59:02 +00003167 tsk_cache_hot = task_hot(p, rq->clock, sd);
3168 if (!tsk_cache_hot ||
3169 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003170#ifdef CONFIG_SCHEDSTATS
Luis Henriques708dc512009-03-16 19:59:02 +00003171 if (tsk_cache_hot) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003172 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02003173 schedstat_inc(p, se.nr_forced_migrations);
3174 }
Ingo Molnarda84d962007-10-15 17:00:18 +02003175#endif
3176 return 1;
3177 }
3178
Luis Henriques708dc512009-03-16 19:59:02 +00003179 if (tsk_cache_hot) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003180 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02003181 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 return 1;
3184}
3185
Peter Williamse1d14842007-10-24 18:23:51 +02003186static unsigned long
3187balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3188 unsigned long max_load_move, struct sched_domain *sd,
3189 enum cpu_idle_type idle, int *all_pinned,
3190 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003191{
Peter Zijlstra051c6762008-06-27 13:41:31 +02003192 int loops = 0, pulled = 0, pinned = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02003193 struct task_struct *p;
3194 long rem_load_move = max_load_move;
3195
Peter Williamse1d14842007-10-24 18:23:51 +02003196 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02003197 goto out;
3198
3199 pinned = 1;
3200
3201 /*
3202 * Start the load-balancing iterator:
3203 */
3204 p = iterator->start(iterator->arg);
3205next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003206 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02003207 goto out;
Peter Zijlstra051c6762008-06-27 13:41:31 +02003208
3209 if ((p->se.load.weight >> 1) > rem_load_move ||
Ingo Molnardd41f592007-07-09 18:51:59 +02003210 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003211 p = iterator->next(iterator->arg);
3212 goto next;
3213 }
3214
3215 pull_task(busiest, p, this_rq, this_cpu);
3216 pulled++;
3217 rem_load_move -= p->se.load.weight;
3218
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003219#ifdef CONFIG_PREEMPT
3220 /*
3221 * NEWIDLE balancing is a source of latency, so preemptible kernels
3222 * will stop after the first task is pulled to minimize the critical
3223 * section.
3224 */
3225 if (idle == CPU_NEWLY_IDLE)
3226 goto out;
3227#endif
3228
Ingo Molnardd41f592007-07-09 18:51:59 +02003229 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003230 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003231 */
Peter Williamse1d14842007-10-24 18:23:51 +02003232 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003233 if (p->prio < *this_best_prio)
3234 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003235 p = iterator->next(iterator->arg);
3236 goto next;
3237 }
3238out:
3239 /*
Peter Williamse1d14842007-10-24 18:23:51 +02003240 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02003241 * so we can safely collect pull_task() stats here rather than
3242 * inside pull_task().
3243 */
3244 schedstat_add(sd, lb_gained[idle], pulled);
3245
3246 if (all_pinned)
3247 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003248
3249 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003250}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003251
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252/*
Peter Williams43010652007-08-09 11:16:46 +02003253 * move_tasks tries to move up to max_load_move weighted load from busiest to
3254 * this_rq, as part of a balancing operation within domain "sd".
3255 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 *
3257 * Called with both runqueues locked.
3258 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003259static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003260 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003261 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003262 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003264 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003265 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003266 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267
Ingo Molnardd41f592007-07-09 18:51:59 +02003268 do {
Peter Williams43010652007-08-09 11:16:46 +02003269 total_load_moved +=
3270 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003271 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003272 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003273 class = class->next;
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003274
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003275#ifdef CONFIG_PREEMPT
3276 /*
3277 * NEWIDLE balancing is a source of latency, so preemptible
3278 * kernels will stop after the first task is pulled to minimize
3279 * the critical section.
3280 */
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003281 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3282 break;
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003283#endif
Peter Williams43010652007-08-09 11:16:46 +02003284 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285
Peter Williams43010652007-08-09 11:16:46 +02003286 return total_load_moved > 0;
3287}
3288
Peter Williamse1d14842007-10-24 18:23:51 +02003289static int
3290iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3291 struct sched_domain *sd, enum cpu_idle_type idle,
3292 struct rq_iterator *iterator)
3293{
3294 struct task_struct *p = iterator->start(iterator->arg);
3295 int pinned = 0;
3296
3297 while (p) {
3298 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3299 pull_task(busiest, p, this_rq, this_cpu);
3300 /*
3301 * Right now, this is only the second place pull_task()
3302 * is called, so we can safely collect pull_task()
3303 * stats here rather than inside pull_task().
3304 */
3305 schedstat_inc(sd, lb_gained[idle]);
3306
3307 return 1;
3308 }
3309 p = iterator->next(iterator->arg);
3310 }
3311
3312 return 0;
3313}
3314
Peter Williams43010652007-08-09 11:16:46 +02003315/*
3316 * move_one_task tries to move exactly one task from busiest to this_rq, as
3317 * part of active balancing operations within "domain".
3318 * Returns 1 if successful and 0 otherwise.
3319 *
3320 * Called with both runqueues locked.
3321 */
3322static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3323 struct sched_domain *sd, enum cpu_idle_type idle)
3324{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003325 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003326
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003327 for_each_class(class) {
Peter Williamse1d14842007-10-24 18:23:51 +02003328 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003329 return 1;
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003330 }
Peter Williams43010652007-08-09 11:16:46 +02003331
3332 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333}
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303334/********** Helpers for find_busiest_group ************************/
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003335/*
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303336 * sd_lb_stats - Structure to store the statistics of a sched_domain
3337 * during load balancing.
3338 */
3339struct sd_lb_stats {
3340 struct sched_group *busiest; /* Busiest group in this sd */
3341 struct sched_group *this; /* Local group in this sd */
3342 unsigned long total_load; /* Total load of all groups in sd */
3343 unsigned long total_pwr; /* Total power of all groups in sd */
3344 unsigned long avg_load; /* Average load across all groups in sd */
3345
3346 /** Statistics of this group */
3347 unsigned long this_load;
3348 unsigned long this_load_per_task;
3349 unsigned long this_nr_running;
3350
3351 /* Statistics of the busiest group */
3352 unsigned long max_load;
3353 unsigned long busiest_load_per_task;
3354 unsigned long busiest_nr_running;
3355
3356 int group_imb; /* Is there imbalance in this sd */
3357#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3358 int power_savings_balance; /* Is powersave balance needed for this sd */
3359 struct sched_group *group_min; /* Least loaded group in sd */
3360 struct sched_group *group_leader; /* Group which relieves group_min */
3361 unsigned long min_load_per_task; /* load_per_task in group_min */
3362 unsigned long leader_nr_running; /* Nr running of group_leader */
3363 unsigned long min_nr_running; /* Nr running of group_min */
3364#endif
3365};
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366
3367/*
Gautham R Shenoy381be782009-03-25 14:43:46 +05303368 * sg_lb_stats - stats of a sched_group required for load_balancing
3369 */
3370struct sg_lb_stats {
3371 unsigned long avg_load; /*Avg load across the CPUs of the group */
3372 unsigned long group_load; /* Total load over the CPUs of the group */
3373 unsigned long sum_nr_running; /* Nr tasks running in the group */
3374 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3375 unsigned long group_capacity;
3376 int group_imb; /* Is there an imbalance in the group ? */
3377};
3378
3379/**
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303380 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3381 * @group: The group whose first cpu is to be returned.
3382 */
3383static inline unsigned int group_first_cpu(struct sched_group *group)
3384{
3385 return cpumask_first(sched_group_cpus(group));
3386}
3387
3388/**
3389 * get_sd_load_idx - Obtain the load index for a given sched domain.
3390 * @sd: The sched_domain whose load_idx is to be obtained.
3391 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3392 */
3393static inline int get_sd_load_idx(struct sched_domain *sd,
3394 enum cpu_idle_type idle)
3395{
3396 int load_idx;
3397
3398 switch (idle) {
3399 case CPU_NOT_IDLE:
3400 load_idx = sd->busy_idx;
3401 break;
3402
3403 case CPU_NEWLY_IDLE:
3404 load_idx = sd->newidle_idx;
3405 break;
3406 default:
3407 load_idx = sd->idle_idx;
3408 break;
3409 }
3410
3411 return load_idx;
3412}
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303413
3414
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303415#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3416/**
3417 * init_sd_power_savings_stats - Initialize power savings statistics for
3418 * the given sched_domain, during load balancing.
3419 *
3420 * @sd: Sched domain whose power-savings statistics are to be initialized.
3421 * @sds: Variable containing the statistics for sd.
3422 * @idle: Idle status of the CPU at which we're performing load-balancing.
3423 */
3424static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3425 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3426{
3427 /*
3428 * Busy processors will not participate in power savings
3429 * balance.
3430 */
3431 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3432 sds->power_savings_balance = 0;
3433 else {
3434 sds->power_savings_balance = 1;
3435 sds->min_nr_running = ULONG_MAX;
3436 sds->leader_nr_running = 0;
3437 }
3438}
3439
3440/**
3441 * update_sd_power_savings_stats - Update the power saving stats for a
3442 * sched_domain while performing load balancing.
3443 *
3444 * @group: sched_group belonging to the sched_domain under consideration.
3445 * @sds: Variable containing the statistics of the sched_domain
3446 * @local_group: Does group contain the CPU for which we're performing
3447 * load balancing ?
3448 * @sgs: Variable containing the statistics of the group.
3449 */
3450static inline void update_sd_power_savings_stats(struct sched_group *group,
3451 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3452{
3453
3454 if (!sds->power_savings_balance)
3455 return;
3456
3457 /*
3458 * If the local group is idle or completely loaded
3459 * no need to do power savings balance at this domain
3460 */
3461 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3462 !sds->this_nr_running))
3463 sds->power_savings_balance = 0;
3464
3465 /*
3466 * If a group is already running at full capacity or idle,
3467 * don't include that group in power savings calculations
3468 */
3469 if (!sds->power_savings_balance ||
3470 sgs->sum_nr_running >= sgs->group_capacity ||
3471 !sgs->sum_nr_running)
3472 return;
3473
3474 /*
3475 * Calculate the group which has the least non-idle load.
3476 * This is the group from where we need to pick up the load
3477 * for saving power
3478 */
3479 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3480 (sgs->sum_nr_running == sds->min_nr_running &&
3481 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3482 sds->group_min = group;
3483 sds->min_nr_running = sgs->sum_nr_running;
3484 sds->min_load_per_task = sgs->sum_weighted_load /
3485 sgs->sum_nr_running;
3486 }
3487
3488 /*
3489 * Calculate the group which is almost near its
3490 * capacity but still has some space to pick up some load
3491 * from other group and save more power
3492 */
Gautham R Shenoyd899a782009-09-02 16:59:10 +05303493 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303494 return;
3495
3496 if (sgs->sum_nr_running > sds->leader_nr_running ||
3497 (sgs->sum_nr_running == sds->leader_nr_running &&
3498 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3499 sds->group_leader = group;
3500 sds->leader_nr_running = sgs->sum_nr_running;
3501 }
3502}
3503
3504/**
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003505 * check_power_save_busiest_group - see if there is potential for some power-savings balance
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303506 * @sds: Variable containing the statistics of the sched_domain
3507 * under consideration.
3508 * @this_cpu: Cpu at which we're currently performing load-balancing.
3509 * @imbalance: Variable to store the imbalance.
3510 *
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003511 * Description:
3512 * Check if we have potential to perform some power-savings balance.
3513 * If yes, set the busiest group to be the least loaded group in the
3514 * sched_domain, so that it's CPUs can be put to idle.
3515 *
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303516 * Returns 1 if there is potential to perform power-savings balance.
3517 * Else returns 0.
3518 */
3519static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3520 int this_cpu, unsigned long *imbalance)
3521{
3522 if (!sds->power_savings_balance)
3523 return 0;
3524
3525 if (sds->this != sds->group_leader ||
3526 sds->group_leader == sds->group_min)
3527 return 0;
3528
3529 *imbalance = sds->min_load_per_task;
3530 sds->busiest = sds->group_min;
3531
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303532 return 1;
3533
3534}
3535#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3536static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3537 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3538{
3539 return;
3540}
3541
3542static inline void update_sd_power_savings_stats(struct sched_group *group,
3543 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3544{
3545 return;
3546}
3547
3548static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3549 int this_cpu, unsigned long *imbalance)
3550{
3551 return 0;
3552}
3553#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3554
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003555
3556unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3557{
3558 return SCHED_LOAD_SCALE;
3559}
3560
3561unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3562{
3563 return default_scale_freq_power(sd, cpu);
3564}
3565
3566unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003567{
3568 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3569 unsigned long smt_gain = sd->smt_gain;
3570
3571 smt_gain /= weight;
3572
3573 return smt_gain;
3574}
3575
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003576unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3577{
3578 return default_scale_smt_power(sd, cpu);
3579}
3580
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003581unsigned long scale_rt_power(int cpu)
3582{
3583 struct rq *rq = cpu_rq(cpu);
3584 u64 total, available;
3585
3586 sched_avg_update(rq);
3587
3588 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3589 available = total - rq->rt_avg;
3590
3591 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3592 total = SCHED_LOAD_SCALE;
3593
3594 total >>= SCHED_LOAD_SHIFT;
3595
3596 return div_u64(available, total);
3597}
3598
Peter Zijlstraab292302009-09-01 10:34:36 +02003599static void update_cpu_power(struct sched_domain *sd, int cpu)
3600{
3601 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3602 unsigned long power = SCHED_LOAD_SCALE;
3603 struct sched_group *sdg = sd->groups;
Peter Zijlstraab292302009-09-01 10:34:36 +02003604
Peter Zijlstra8e6598af2009-09-03 13:20:03 +02003605 if (sched_feat(ARCH_POWER))
3606 power *= arch_scale_freq_power(sd, cpu);
3607 else
3608 power *= default_scale_freq_power(sd, cpu);
3609
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003610 power >>= SCHED_LOAD_SHIFT;
Peter Zijlstraab292302009-09-01 10:34:36 +02003611
3612 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
Peter Zijlstra8e6598af2009-09-03 13:20:03 +02003613 if (sched_feat(ARCH_POWER))
3614 power *= arch_scale_smt_power(sd, cpu);
3615 else
3616 power *= default_scale_smt_power(sd, cpu);
3617
Peter Zijlstraab292302009-09-01 10:34:36 +02003618 power >>= SCHED_LOAD_SHIFT;
3619 }
3620
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003621 power *= scale_rt_power(cpu);
3622 power >>= SCHED_LOAD_SHIFT;
3623
3624 if (!power)
3625 power = 1;
Peter Zijlstraab292302009-09-01 10:34:36 +02003626
Peter Zijlstra18a38852009-09-01 10:34:39 +02003627 sdg->cpu_power = power;
Peter Zijlstraab292302009-09-01 10:34:36 +02003628}
3629
3630static void update_group_power(struct sched_domain *sd, int cpu)
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003631{
3632 struct sched_domain *child = sd->child;
3633 struct sched_group *group, *sdg = sd->groups;
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003634 unsigned long power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003635
3636 if (!child) {
Peter Zijlstraab292302009-09-01 10:34:36 +02003637 update_cpu_power(sd, cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003638 return;
3639 }
3640
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003641 power = 0;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003642
3643 group = child->groups;
3644 do {
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003645 power += group->cpu_power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003646 group = group->next;
3647 } while (group != child->groups);
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003648
3649 sdg->cpu_power = power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003650}
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303651
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303652/**
3653 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3654 * @group: sched_group whose statistics are to be updated.
3655 * @this_cpu: Cpu for which load balance is currently performed.
3656 * @idle: Idle status of this_cpu
3657 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3658 * @sd_idle: Idle status of the sched_domain containing group.
3659 * @local_group: Does group contain this_cpu.
3660 * @cpus: Set of cpus considered for load balancing.
3661 * @balance: Should we balance.
3662 * @sgs: variable to hold the statistics for this group.
3663 */
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003664static inline void update_sg_lb_stats(struct sched_domain *sd,
3665 struct sched_group *group, int this_cpu,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303666 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3667 int local_group, const struct cpumask *cpus,
3668 int *balance, struct sg_lb_stats *sgs)
3669{
3670 unsigned long load, max_cpu_load, min_cpu_load;
3671 int i;
3672 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3673 unsigned long sum_avg_load_per_task;
3674 unsigned long avg_load_per_task;
3675
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003676 if (local_group) {
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303677 balance_cpu = group_first_cpu(group);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003678 if (balance_cpu == this_cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003679 update_group_power(sd, this_cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003680 }
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303681
3682 /* Tally up the load of all CPUs in the group */
3683 sum_avg_load_per_task = avg_load_per_task = 0;
3684 max_cpu_load = 0;
3685 min_cpu_load = ~0UL;
3686
3687 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3688 struct rq *rq = cpu_rq(i);
3689
3690 if (*sd_idle && rq->nr_running)
3691 *sd_idle = 0;
3692
3693 /* Bias balancing toward cpus of our domain */
3694 if (local_group) {
3695 if (idle_cpu(i) && !first_idle_cpu) {
3696 first_idle_cpu = 1;
3697 balance_cpu = i;
3698 }
3699
3700 load = target_load(i, load_idx);
3701 } else {
3702 load = source_load(i, load_idx);
3703 if (load > max_cpu_load)
3704 max_cpu_load = load;
3705 if (min_cpu_load > load)
3706 min_cpu_load = load;
3707 }
3708
3709 sgs->group_load += load;
3710 sgs->sum_nr_running += rq->nr_running;
3711 sgs->sum_weighted_load += weighted_cpuload(i);
3712
3713 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3714 }
3715
3716 /*
3717 * First idle cpu or the first cpu(busiest) in this sched group
3718 * is eligible for doing load balancing at this and above
3719 * domains. In the newly idle case, we will allow all the cpu's
3720 * to do the newly idle load balance.
3721 */
3722 if (idle != CPU_NEWLY_IDLE && local_group &&
3723 balance_cpu != this_cpu && balance) {
3724 *balance = 0;
3725 return;
3726 }
3727
3728 /* Adjust by relative CPU power of the group */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003729 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303730
3731
3732 /*
3733 * Consider the group unbalanced when the imbalance is larger
3734 * than the average weight of two tasks.
3735 *
3736 * APZ: with cgroup the avg task weight can vary wildly and
3737 * might not be a suitable number - should we keep a
3738 * normalized nr_running number somewhere that negates
3739 * the hierarchy?
3740 */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003741 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3742 group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303743
3744 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3745 sgs->group_imb = 1;
3746
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003747 sgs->group_capacity =
Peter Zijlstra18a38852009-09-01 10:34:39 +02003748 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303749}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303751/**
3752 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3753 * @sd: sched_domain whose statistics are to be updated.
3754 * @this_cpu: Cpu for which load balance is currently performed.
3755 * @idle: Idle status of this_cpu
3756 * @sd_idle: Idle status of the sched_domain containing group.
3757 * @cpus: Set of cpus considered for load balancing.
3758 * @balance: Should we balance.
3759 * @sds: variable to hold the statistics for this sched_domain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303761static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3762 enum cpu_idle_type idle, int *sd_idle,
3763 const struct cpumask *cpus, int *balance,
3764 struct sd_lb_stats *sds)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765{
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003766 struct sched_domain *child = sd->child;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303767 struct sched_group *group = sd->groups;
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303768 struct sg_lb_stats sgs;
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003769 int load_idx, prefer_sibling = 0;
3770
3771 if (child && child->flags & SD_PREFER_SIBLING)
3772 prefer_sibling = 1;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303773
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303774 init_sd_power_savings_stats(sd, sds, idle);
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303775 load_idx = get_sd_load_idx(sd, idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776
3777 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003778 int local_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779
Rusty Russell758b2cd2008-11-25 02:35:04 +10303780 local_group = cpumask_test_cpu(this_cpu,
3781 sched_group_cpus(group));
Gautham R Shenoy381be782009-03-25 14:43:46 +05303782 memset(&sgs, 0, sizeof(sgs));
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003783 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303784 local_group, cpus, balance, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003785
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303786 if (local_group && balance && !(*balance))
3787 return;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003788
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303789 sds->total_load += sgs.group_load;
Peter Zijlstra18a38852009-09-01 10:34:39 +02003790 sds->total_pwr += group->cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003792 /*
3793 * In case the child domain prefers tasks go to siblings
3794 * first, lower the group capacity to one so that we'll try
3795 * and move all the excess tasks away.
3796 */
3797 if (prefer_sibling)
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003798 sgs.group_capacity = min(sgs.group_capacity, 1UL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 if (local_group) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303801 sds->this_load = sgs.avg_load;
3802 sds->this = group;
3803 sds->this_nr_running = sgs.sum_nr_running;
3804 sds->this_load_per_task = sgs.sum_weighted_load;
3805 } else if (sgs.avg_load > sds->max_load &&
Gautham R Shenoy381be782009-03-25 14:43:46 +05303806 (sgs.sum_nr_running > sgs.group_capacity ||
3807 sgs.group_imb)) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303808 sds->max_load = sgs.avg_load;
3809 sds->busiest = group;
3810 sds->busiest_nr_running = sgs.sum_nr_running;
3811 sds->busiest_load_per_task = sgs.sum_weighted_load;
3812 sds->group_imb = sgs.group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003814
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303815 update_sd_power_savings_stats(group, sds, local_group, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816 group = group->next;
3817 } while (group != sd->groups);
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303818}
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303819
3820/**
3821 * fix_small_imbalance - Calculate the minor imbalance that exists
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303822 * amongst the groups of a sched_domain, during
3823 * load balancing.
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303824 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3825 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3826 * @imbalance: Variable to store the imbalance.
3827 */
3828static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3829 int this_cpu, unsigned long *imbalance)
3830{
3831 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3832 unsigned int imbn = 2;
3833
3834 if (sds->this_nr_running) {
3835 sds->this_load_per_task /= sds->this_nr_running;
3836 if (sds->busiest_load_per_task >
3837 sds->this_load_per_task)
3838 imbn = 1;
3839 } else
3840 sds->this_load_per_task =
3841 cpu_avg_load_per_task(this_cpu);
3842
3843 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3844 sds->busiest_load_per_task * imbn) {
3845 *imbalance = sds->busiest_load_per_task;
3846 return;
3847 }
3848
3849 /*
3850 * OK, we don't have enough imbalance to justify moving tasks,
3851 * however we may be able to increase total CPU power used by
3852 * moving them.
3853 */
3854
Peter Zijlstra18a38852009-09-01 10:34:39 +02003855 pwr_now += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303856 min(sds->busiest_load_per_task, sds->max_load);
Peter Zijlstra18a38852009-09-01 10:34:39 +02003857 pwr_now += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303858 min(sds->this_load_per_task, sds->this_load);
3859 pwr_now /= SCHED_LOAD_SCALE;
3860
3861 /* Amount of load we'd subtract */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003862 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3863 sds->busiest->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303864 if (sds->max_load > tmp)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003865 pwr_move += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303866 min(sds->busiest_load_per_task, sds->max_load - tmp);
3867
3868 /* Amount of load we'd add */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003869 if (sds->max_load * sds->busiest->cpu_power <
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303870 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003871 tmp = (sds->max_load * sds->busiest->cpu_power) /
3872 sds->this->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303873 else
Peter Zijlstra18a38852009-09-01 10:34:39 +02003874 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3875 sds->this->cpu_power;
3876 pwr_move += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303877 min(sds->this_load_per_task, sds->this_load + tmp);
3878 pwr_move /= SCHED_LOAD_SCALE;
3879
3880 /* Move if we gain throughput */
3881 if (pwr_move > pwr_now)
3882 *imbalance = sds->busiest_load_per_task;
3883}
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303884
3885/**
3886 * calculate_imbalance - Calculate the amount of imbalance present within the
3887 * groups of a given sched_domain during load balance.
3888 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3889 * @this_cpu: Cpu for which currently load balance is being performed.
3890 * @imbalance: The variable to store the imbalance.
3891 */
3892static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3893 unsigned long *imbalance)
3894{
3895 unsigned long max_pull;
3896 /*
3897 * In the presence of smp nice balancing, certain scenarios can have
3898 * max load less than avg load(as we skip the groups at or below
3899 * its cpu_power, while calculating max_load..)
3900 */
3901 if (sds->max_load < sds->avg_load) {
3902 *imbalance = 0;
3903 return fix_small_imbalance(sds, this_cpu, imbalance);
3904 }
3905
3906 /* Don't want to pull so many tasks that a group would go idle */
3907 max_pull = min(sds->max_load - sds->avg_load,
3908 sds->max_load - sds->busiest_load_per_task);
3909
3910 /* How much load to actually move to equalise the imbalance */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003911 *imbalance = min(max_pull * sds->busiest->cpu_power,
3912 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303913 / SCHED_LOAD_SCALE;
3914
3915 /*
3916 * if *imbalance is less than the average load per runnable task
3917 * there is no gaurantee that any tasks will be moved so we'll have
3918 * a think about bumping its value to force at least one task to be
3919 * moved
3920 */
3921 if (*imbalance < sds->busiest_load_per_task)
3922 return fix_small_imbalance(sds, this_cpu, imbalance);
3923
3924}
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303925/******* find_busiest_group() helpers end here *********************/
3926
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303927/**
3928 * find_busiest_group - Returns the busiest group within the sched_domain
3929 * if there is an imbalance. If there isn't an imbalance, and
3930 * the user has opted for power-savings, it returns a group whose
3931 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3932 * such a group exists.
3933 *
3934 * Also calculates the amount of weighted load which should be moved
3935 * to restore balance.
3936 *
3937 * @sd: The sched_domain whose busiest group is to be returned.
3938 * @this_cpu: The cpu for which load balancing is currently being performed.
3939 * @imbalance: Variable which stores amount of weighted load which should
3940 * be moved to restore balance/put a group to idle.
3941 * @idle: The idle status of this_cpu.
3942 * @sd_idle: The idleness of sd
3943 * @cpus: The set of CPUs under consideration for load-balancing.
3944 * @balance: Pointer to a variable indicating if this_cpu
3945 * is the appropriate cpu to perform load balancing at this_level.
3946 *
3947 * Returns: - the busiest group if imbalance exists.
3948 * - If no imbalance and user has opted for power-savings balance,
3949 * return the least loaded group whose CPUs can be
3950 * put to idle by rebalancing its tasks onto our group.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 */
3952static struct sched_group *
3953find_busiest_group(struct sched_domain *sd, int this_cpu,
3954 unsigned long *imbalance, enum cpu_idle_type idle,
3955 int *sd_idle, const struct cpumask *cpus, int *balance)
3956{
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303957 struct sd_lb_stats sds;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303959 memset(&sds, 0, sizeof(sds));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303961 /*
3962 * Compute the various statistics relavent for load balancing at
3963 * this level.
3964 */
3965 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3966 balance, &sds);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303968 /* Cases where imbalance does not exist from POV of this_cpu */
3969 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3970 * at this level.
3971 * 2) There is no busy sibling group to pull from.
3972 * 3) This group is the busiest group.
3973 * 4) This group is more busy than the avg busieness at this
3974 * sched_domain.
3975 * 5) The imbalance is within the specified limit.
3976 * 6) Any rebalance would lead to ping-pong
3977 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303978 if (balance && !(*balance))
3979 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303981 if (!sds.busiest || sds.busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 goto out_balanced;
3983
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303984 if (sds.this_load >= sds.max_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 goto out_balanced;
3986
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303987 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303989 if (sds.this_load >= sds.avg_load)
3990 goto out_balanced;
3991
3992 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 goto out_balanced;
3994
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303995 sds.busiest_load_per_task /= sds.busiest_nr_running;
3996 if (sds.group_imb)
3997 sds.busiest_load_per_task =
3998 min(sds.busiest_load_per_task, sds.avg_load);
Ken Chen908a7c12007-10-17 16:55:11 +02003999
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 /*
4001 * We're trying to get all the cpus to the average_load, so we don't
4002 * want to push ourselves above the average load, nor do we wish to
4003 * reduce the max loaded cpu below the average load, as either of these
4004 * actions would just result in more rebalancing later, and ping-pong
4005 * tasks around. Thus we look for the minimum possible imbalance.
4006 * Negative imbalances (*we* are more loaded than anyone else) will
4007 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004008 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009 * appear as very large values with unsigned longs.
4010 */
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304011 if (sds.max_load <= sds.busiest_load_per_task)
Peter Williams2dd73a42006-06-27 02:54:34 -07004012 goto out_balanced;
4013
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05304014 /* Looks like there is an imbalance. Compute it */
4015 calculate_imbalance(&sds, this_cpu, imbalance);
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304016 return sds.busiest;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017
4018out_balanced:
Gautham R Shenoyc071df12009-03-25 14:44:22 +05304019 /*
4020 * There is no obvious imbalance. But check if we can do some balancing
4021 * to save power.
4022 */
4023 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4024 return sds.busiest;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004025ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 *imbalance = 0;
4027 return NULL;
4028}
4029
4030/*
4031 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4032 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004033static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004034find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Rusty Russell96f874e22008-11-25 02:35:14 +10304035 unsigned long imbalance, const struct cpumask *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004037 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07004038 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 int i;
4040
Rusty Russell758b2cd2008-11-25 02:35:04 +10304041 for_each_cpu(i, sched_group_cpus(group)) {
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004042 unsigned long power = power_of(i);
4043 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
Ingo Molnardd41f592007-07-09 18:51:59 +02004044 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004045
Rusty Russell96f874e22008-11-25 02:35:14 +10304046 if (!cpumask_test_cpu(i, cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004047 continue;
4048
Ingo Molnar48f24c42006-07-03 00:25:40 -07004049 rq = cpu_rq(i);
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004050 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4051 wl /= power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004053 if (capacity && rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07004054 continue;
4055
Ingo Molnardd41f592007-07-09 18:51:59 +02004056 if (wl > max_load) {
4057 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004058 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 }
4060 }
4061
4062 return busiest;
4063}
4064
4065/*
Nick Piggin77391d72005-06-25 14:57:30 -07004066 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4067 * so long as it is large enough.
4068 */
4069#define MAX_PINNED_INTERVAL 512
4070
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304071/* Working cpumask for load_balance and load_balance_newidle. */
4072static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4073
Nick Piggin77391d72005-06-25 14:57:30 -07004074/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4076 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004078static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004079 struct sched_domain *sd, enum cpu_idle_type idle,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304080 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081{
Peter Williams43010652007-08-09 11:16:46 +02004082 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004085 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004086 unsigned long flags;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304087 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004088
Rusty Russell96f874e22008-11-25 02:35:14 +10304089 cpumask_setall(cpus);
Mike Travis7c16ec52008-04-04 18:11:11 -07004090
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004091 /*
4092 * When power savings policy is enabled for the parent domain, idle
4093 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02004094 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004095 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004096 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004097 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004098 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004099 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Ingo Molnar2d723762007-10-15 17:00:12 +02004101 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004103redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004104 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004105 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07004106 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004107
Chen, Kenneth W06066712006-12-10 02:20:35 -08004108 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004109 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004110
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111 if (!group) {
4112 schedstat_inc(sd, lb_nobusyg[idle]);
4113 goto out_balanced;
4114 }
4115
Mike Travis7c16ec52008-04-04 18:11:11 -07004116 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004117 if (!busiest) {
4118 schedstat_inc(sd, lb_nobusyq[idle]);
4119 goto out_balanced;
4120 }
4121
Nick Piggindb935db2005-06-25 14:57:11 -07004122 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123
4124 schedstat_add(sd, lb_imbalance[idle], imbalance);
4125
Peter Williams43010652007-08-09 11:16:46 +02004126 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 if (busiest->nr_running > 1) {
4128 /*
4129 * Attempt to move tasks. If find_busiest_group has found
4130 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02004131 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 * correctly treated as an imbalance.
4133 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004134 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07004135 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02004136 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07004137 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07004138 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004139 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07004140
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004141 /*
4142 * some other cpu did the load balance for us.
4143 */
Peter Williams43010652007-08-09 11:16:46 +02004144 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004145 resched_cpu(this_cpu);
4146
Nick Piggin81026792005-06-25 14:57:07 -07004147 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004148 if (unlikely(all_pinned)) {
Rusty Russell96f874e22008-11-25 02:35:14 +10304149 cpumask_clear_cpu(cpu_of(busiest), cpus);
4150 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004151 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07004152 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004153 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004154 }
Nick Piggin81026792005-06-25 14:57:07 -07004155
Peter Williams43010652007-08-09 11:16:46 +02004156 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004157 schedstat_inc(sd, lb_failed[idle]);
4158 sd->nr_balance_failed++;
4159
4160 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004162 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004163
4164 /* don't kick the migration_thread, if the curr
4165 * task on busiest cpu can't be moved to this_cpu
4166 */
Rusty Russell96f874e22008-11-25 02:35:14 +10304167 if (!cpumask_test_cpu(this_cpu,
4168 &busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004169 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004170 all_pinned = 1;
4171 goto out_one_pinned;
4172 }
4173
Linus Torvalds1da177e2005-04-16 15:20:36 -07004174 if (!busiest->active_balance) {
4175 busiest->active_balance = 1;
4176 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07004177 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004179 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07004180 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 wake_up_process(busiest->migration_thread);
4182
4183 /*
4184 * We've kicked active balancing, reset the failure
4185 * counter.
4186 */
Nick Piggin39507452005-06-25 14:57:09 -07004187 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 }
Nick Piggin81026792005-06-25 14:57:07 -07004189 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 sd->nr_balance_failed = 0;
4191
Nick Piggin81026792005-06-25 14:57:07 -07004192 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193 /* We were unbalanced, so reset the balancing interval */
4194 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07004195 } else {
4196 /*
4197 * If we've begun active balancing, start to back off. This
4198 * case may not be covered by the all_pinned logic if there
4199 * is only 1 task on the busy runqueue (because we don't call
4200 * move_tasks).
4201 */
4202 if (sd->balance_interval < sd->max_interval)
4203 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204 }
4205
Peter Williams43010652007-08-09 11:16:46 +02004206 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004207 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004208 ld_moved = -1;
4209
4210 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211
4212out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 schedstat_inc(sd, lb_balanced[idle]);
4214
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004215 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004216
4217out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07004219 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4220 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 sd->balance_interval *= 2;
4222
Ingo Molnar48f24c42006-07-03 00:25:40 -07004223 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004224 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004225 ld_moved = -1;
4226 else
4227 ld_moved = 0;
4228out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004229 if (ld_moved)
4230 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004231 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232}
4233
4234/*
4235 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4236 * tasks if there is an imbalance.
4237 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004238 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 * this_rq is locked.
4240 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07004241static int
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304242load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243{
4244 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004245 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02004247 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07004248 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004249 int all_pinned = 0;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304250 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004251
Rusty Russell96f874e22008-11-25 02:35:14 +10304252 cpumask_setall(cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07004253
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004254 /*
4255 * When power savings policy is enabled for the parent domain, idle
4256 * sibling can pick up load irrespective of busy siblings. In this case,
4257 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004258 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004259 */
4260 if (sd->flags & SD_SHARE_CPUPOWER &&
4261 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004262 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004263
Ingo Molnar2d723762007-10-15 17:00:12 +02004264 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004265redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004266 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004267 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07004268 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004270 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004271 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 }
4273
Mike Travis7c16ec52008-04-04 18:11:11 -07004274 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07004275 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004276 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004277 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004278 }
4279
Nick Piggindb935db2005-06-25 14:57:11 -07004280 BUG_ON(busiest == this_rq);
4281
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004282 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004283
Peter Williams43010652007-08-09 11:16:46 +02004284 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004285 if (busiest->nr_running > 1) {
4286 /* Attempt to move tasks */
4287 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004288 /* this_rq->clock is already updated */
4289 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02004290 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004291 imbalance, sd, CPU_NEWLY_IDLE,
4292 &all_pinned);
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004293 double_unlock_balance(this_rq, busiest);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004294
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004295 if (unlikely(all_pinned)) {
Rusty Russell96f874e22008-11-25 02:35:14 +10304296 cpumask_clear_cpu(cpu_of(busiest), cpus);
4297 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004298 goto redo;
4299 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004300 }
4301
Peter Williams43010652007-08-09 11:16:46 +02004302 if (!ld_moved) {
Vaidyanathan Srinivasan36dffab2008-12-20 10:06:38 +05304303 int active_balance = 0;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304304
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004305 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004306 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4307 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004308 return -1;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304309
4310 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4311 return -1;
4312
4313 if (sd->nr_balance_failed++ < 2)
4314 return -1;
4315
4316 /*
4317 * The only task running in a non-idle cpu can be moved to this
4318 * cpu in an attempt to completely freeup the other CPU
4319 * package. The same method used to move task in load_balance()
4320 * have been extended for load_balance_newidle() to speedup
4321 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4322 *
4323 * The package power saving logic comes from
4324 * find_busiest_group(). If there are no imbalance, then
4325 * f_b_g() will return NULL. However when sched_mc={1,2} then
4326 * f_b_g() will select a group from which a running task may be
4327 * pulled to this cpu in order to make the other package idle.
4328 * If there is no opportunity to make a package idle and if
4329 * there are no imbalance, then f_b_g() will return NULL and no
4330 * action will be taken in load_balance_newidle().
4331 *
4332 * Under normal task pull operation due to imbalance, there
4333 * will be more than one task in the source run queue and
4334 * move_tasks() will succeed. ld_moved will be true and this
4335 * active balance code will not be triggered.
4336 */
4337
4338 /* Lock busiest in correct order while this_rq is held */
4339 double_lock_balance(this_rq, busiest);
4340
4341 /*
4342 * don't kick the migration_thread, if the curr
4343 * task on busiest cpu can't be moved to this_cpu
4344 */
Mike Travis6ca09df2008-12-31 18:08:45 -08004345 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304346 double_unlock_balance(this_rq, busiest);
4347 all_pinned = 1;
4348 return ld_moved;
4349 }
4350
4351 if (!busiest->active_balance) {
4352 busiest->active_balance = 1;
4353 busiest->push_cpu = this_cpu;
4354 active_balance = 1;
4355 }
4356
4357 double_unlock_balance(this_rq, busiest);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004358 /*
4359 * Should not call ttwu while holding a rq->lock
4360 */
4361 spin_unlock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304362 if (active_balance)
4363 wake_up_process(busiest->migration_thread);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004364 spin_lock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304365
Nick Piggin5969fe02005-09-10 00:26:19 -07004366 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004367 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004369 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02004370 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004371
4372out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004373 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004374 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004375 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004376 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004377 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004378
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004379 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004380}
4381
4382/*
4383 * idle_balance is called by schedule() if this_cpu is about to become
4384 * idle. Attempts to pull tasks from other CPUs.
4385 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004386static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004387{
4388 struct sched_domain *sd;
Vaidyanathan Srinivasanefbe0272008-12-08 20:52:49 +05304389 int pulled_task = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02004390 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004391
4392 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004393 unsigned long interval;
4394
4395 if (!(sd->flags & SD_LOAD_BALANCE))
4396 continue;
4397
4398 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07004399 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07004400 pulled_task = load_balance_newidle(this_cpu, this_rq,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304401 sd);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004402
4403 interval = msecs_to_jiffies(sd->balance_interval);
4404 if (time_after(next_balance, sd->last_balance + interval))
4405 next_balance = sd->last_balance + interval;
4406 if (pulled_task)
4407 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004408 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004409 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004410 /*
4411 * We are going idle. next_balance may be set based on
4412 * a busy processor. So reset next_balance.
4413 */
4414 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02004415 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416}
4417
4418/*
4419 * active_load_balance is run by migration threads. It pushes running tasks
4420 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4421 * running on each physical CPU where possible, and avoids physical /
4422 * logical imbalances.
4423 *
4424 * Called with busiest_rq locked.
4425 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004426static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427{
Nick Piggin39507452005-06-25 14:57:09 -07004428 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004429 struct sched_domain *sd;
4430 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07004431
Ingo Molnar48f24c42006-07-03 00:25:40 -07004432 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07004433 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07004434 return;
4435
4436 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004437
4438 /*
Nick Piggin39507452005-06-25 14:57:09 -07004439 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004440 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07004441 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442 */
Nick Piggin39507452005-06-25 14:57:09 -07004443 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004444
Nick Piggin39507452005-06-25 14:57:09 -07004445 /* move a task from busiest_rq to target_rq */
4446 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004447 update_rq_clock(busiest_rq);
4448 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004449
Nick Piggin39507452005-06-25 14:57:09 -07004450 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004451 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07004452 if ((sd->flags & SD_LOAD_BALANCE) &&
Rusty Russell758b2cd2008-11-25 02:35:04 +10304453 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
Nick Piggin39507452005-06-25 14:57:09 -07004454 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004455 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004456
Ingo Molnar48f24c42006-07-03 00:25:40 -07004457 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004458 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004459
Peter Williams43010652007-08-09 11:16:46 +02004460 if (move_one_task(target_rq, target_cpu, busiest_rq,
4461 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07004462 schedstat_inc(sd, alb_pushed);
4463 else
4464 schedstat_inc(sd, alb_failed);
4465 }
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004466 double_unlock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004467}
4468
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004469#ifdef CONFIG_NO_HZ
4470static struct {
4471 atomic_t load_balancer;
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304472 cpumask_var_t cpu_mask;
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304473 cpumask_var_t ilb_grp_nohz_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004474} nohz ____cacheline_aligned = {
4475 .load_balancer = ATOMIC_INIT(-1),
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004476};
4477
Arun R Bharadwajeea08f32009-04-16 12:16:41 +05304478int get_nohz_load_balancer(void)
4479{
4480 return atomic_read(&nohz.load_balancer);
4481}
4482
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304483#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4484/**
4485 * lowest_flag_domain - Return lowest sched_domain containing flag.
4486 * @cpu: The cpu whose lowest level of sched domain is to
4487 * be returned.
4488 * @flag: The flag to check for the lowest sched_domain
4489 * for the given cpu.
4490 *
4491 * Returns the lowest sched_domain of a cpu which contains the given flag.
4492 */
4493static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4494{
4495 struct sched_domain *sd;
4496
4497 for_each_domain(cpu, sd)
4498 if (sd && (sd->flags & flag))
4499 break;
4500
4501 return sd;
4502}
4503
4504/**
4505 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4506 * @cpu: The cpu whose domains we're iterating over.
4507 * @sd: variable holding the value of the power_savings_sd
4508 * for cpu.
4509 * @flag: The flag to filter the sched_domains to be iterated.
4510 *
4511 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4512 * set, starting from the lowest sched_domain to the highest.
4513 */
4514#define for_each_flag_domain(cpu, sd, flag) \
4515 for (sd = lowest_flag_domain(cpu, flag); \
4516 (sd && (sd->flags & flag)); sd = sd->parent)
4517
4518/**
4519 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4520 * @ilb_group: group to be checked for semi-idleness
4521 *
4522 * Returns: 1 if the group is semi-idle. 0 otherwise.
4523 *
4524 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4525 * and atleast one non-idle CPU. This helper function checks if the given
4526 * sched_group is semi-idle or not.
4527 */
4528static inline int is_semi_idle_group(struct sched_group *ilb_group)
4529{
4530 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4531 sched_group_cpus(ilb_group));
4532
4533 /*
4534 * A sched_group is semi-idle when it has atleast one busy cpu
4535 * and atleast one idle cpu.
4536 */
4537 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4538 return 0;
4539
4540 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4541 return 0;
4542
4543 return 1;
4544}
4545/**
4546 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4547 * @cpu: The cpu which is nominating a new idle_load_balancer.
4548 *
4549 * Returns: Returns the id of the idle load balancer if it exists,
4550 * Else, returns >= nr_cpu_ids.
4551 *
4552 * This algorithm picks the idle load balancer such that it belongs to a
4553 * semi-idle powersavings sched_domain. The idea is to try and avoid
4554 * completely idle packages/cores just for the purpose of idle load balancing
4555 * when there are other idle cpu's which are better suited for that job.
4556 */
4557static int find_new_ilb(int cpu)
4558{
4559 struct sched_domain *sd;
4560 struct sched_group *ilb_group;
4561
4562 /*
4563 * Have idle load balancer selection from semi-idle packages only
4564 * when power-aware load balancing is enabled
4565 */
4566 if (!(sched_smt_power_savings || sched_mc_power_savings))
4567 goto out_done;
4568
4569 /*
4570 * Optimize for the case when we have no idle CPUs or only one
4571 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4572 */
4573 if (cpumask_weight(nohz.cpu_mask) < 2)
4574 goto out_done;
4575
4576 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4577 ilb_group = sd->groups;
4578
4579 do {
4580 if (is_semi_idle_group(ilb_group))
4581 return cpumask_first(nohz.ilb_grp_nohz_mask);
4582
4583 ilb_group = ilb_group->next;
4584
4585 } while (ilb_group != sd->groups);
4586 }
4587
4588out_done:
4589 return cpumask_first(nohz.cpu_mask);
4590}
4591#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4592static inline int find_new_ilb(int call_cpu)
4593{
Gautham R Shenoy6e29ec52009-04-21 08:40:49 +05304594 return cpumask_first(nohz.cpu_mask);
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304595}
4596#endif
4597
Christoph Lameter7835b982006-12-10 02:20:22 -08004598/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004599 * This routine will try to nominate the ilb (idle load balancing)
4600 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4601 * load balancing on behalf of all those cpus. If all the cpus in the system
4602 * go into this tickless mode, then there will be no ilb owner (as there is
4603 * no need for one) and all the cpus will sleep till the next wakeup event
4604 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08004605 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004606 * For the ilb owner, tick is not stopped. And this tick will be used
4607 * for idle load balancing. ilb owner will still be part of
4608 * nohz.cpu_mask..
4609 *
4610 * While stopping the tick, this cpu will become the ilb owner if there
4611 * is no other owner. And will be the owner till that cpu becomes busy
4612 * or if all cpus in the system stop their ticks at which point
4613 * there is no need for ilb owner.
4614 *
4615 * When the ilb owner becomes busy, it nominates another owner, during the
4616 * next busy scheduler_tick()
4617 */
4618int select_nohz_load_balancer(int stop_tick)
4619{
4620 int cpu = smp_processor_id();
4621
4622 if (stop_tick) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004623 cpu_rq(cpu)->in_nohz_recently = 1;
4624
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004625 if (!cpu_active(cpu)) {
4626 if (atomic_read(&nohz.load_balancer) != cpu)
4627 return 0;
4628
4629 /*
4630 * If we are going offline and still the leader,
4631 * give up!
4632 */
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004633 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4634 BUG();
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004635
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004636 return 0;
4637 }
4638
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004639 cpumask_set_cpu(cpu, nohz.cpu_mask);
4640
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004641 /* time for ilb owner also to sleep */
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304642 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004643 if (atomic_read(&nohz.load_balancer) == cpu)
4644 atomic_set(&nohz.load_balancer, -1);
4645 return 0;
4646 }
4647
4648 if (atomic_read(&nohz.load_balancer) == -1) {
4649 /* make me the ilb owner */
4650 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4651 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304652 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4653 int new_ilb;
4654
4655 if (!(sched_smt_power_savings ||
4656 sched_mc_power_savings))
4657 return 1;
4658 /*
4659 * Check to see if there is a more power-efficient
4660 * ilb.
4661 */
4662 new_ilb = find_new_ilb(cpu);
4663 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4664 atomic_set(&nohz.load_balancer, -1);
4665 resched_cpu(new_ilb);
4666 return 0;
4667 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004668 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304669 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004670 } else {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304671 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004672 return 0;
4673
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304674 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004675
4676 if (atomic_read(&nohz.load_balancer) == cpu)
4677 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4678 BUG();
4679 }
4680 return 0;
4681}
4682#endif
4683
4684static DEFINE_SPINLOCK(balancing);
4685
4686/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004687 * It checks each scheduling domain to see if it is due to be balanced,
4688 * and initiates a balancing operation if so.
4689 *
4690 * Balancing parameters are set up in arch_init_sched_domains.
4691 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004692static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08004693{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004694 int balance = 1;
4695 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08004696 unsigned long interval;
4697 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004698 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08004699 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004700 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004701 int need_serialize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004702
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004703 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004704 if (!(sd->flags & SD_LOAD_BALANCE))
4705 continue;
4706
4707 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004708 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004709 interval *= sd->busy_factor;
4710
4711 /* scale ms to jiffies */
4712 interval = msecs_to_jiffies(interval);
4713 if (unlikely(!interval))
4714 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02004715 if (interval > HZ*NR_CPUS/10)
4716 interval = HZ*NR_CPUS/10;
4717
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004718 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004719
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004720 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08004721 if (!spin_trylock(&balancing))
4722 goto out;
4723 }
4724
Christoph Lameterc9819f42006-12-10 02:20:25 -08004725 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304726 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004727 /*
4728 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07004729 * longer idle, or one of our SMT siblings is
4730 * not idle.
4731 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004732 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004733 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004734 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004735 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004736 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08004737 spin_unlock(&balancing);
4738out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02004739 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08004740 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004741 update_next_balance = 1;
4742 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004743
4744 /*
4745 * Stop the load balance at this level. There is another
4746 * CPU in our sched group which is doing load balancing more
4747 * actively.
4748 */
4749 if (!balance)
4750 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004751 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02004752
4753 /*
4754 * next_balance will be updated only when there is a need.
4755 * When the cpu is attached to null domain for ex, it will not be
4756 * updated.
4757 */
4758 if (likely(update_next_balance))
4759 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004760}
4761
4762/*
4763 * run_rebalance_domains is triggered when needed from the scheduler tick.
4764 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4765 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4766 */
4767static void run_rebalance_domains(struct softirq_action *h)
4768{
Ingo Molnardd41f592007-07-09 18:51:59 +02004769 int this_cpu = smp_processor_id();
4770 struct rq *this_rq = cpu_rq(this_cpu);
4771 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4772 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004773
Ingo Molnardd41f592007-07-09 18:51:59 +02004774 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004775
4776#ifdef CONFIG_NO_HZ
4777 /*
4778 * If this cpu is the owner for idle load balancing, then do the
4779 * balancing on behalf of the other idle cpus whose ticks are
4780 * stopped.
4781 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004782 if (this_rq->idle_at_tick &&
4783 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004784 struct rq *rq;
4785 int balance_cpu;
4786
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304787 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4788 if (balance_cpu == this_cpu)
4789 continue;
4790
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004791 /*
4792 * If this cpu gets work to do, stop the load balancing
4793 * work being done for other cpus. Next load
4794 * balancing owner will pick it up.
4795 */
4796 if (need_resched())
4797 break;
4798
Oleg Nesterovde0cf892007-08-12 18:08:19 +02004799 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004800
4801 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004802 if (time_after(this_rq->next_balance, rq->next_balance))
4803 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004804 }
4805 }
4806#endif
4807}
4808
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004809static inline int on_null_domain(int cpu)
4810{
4811 return !rcu_dereference(cpu_rq(cpu)->sd);
4812}
4813
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004814/*
4815 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4816 *
4817 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4818 * idle load balancing owner or decide to stop the periodic load balancing,
4819 * if the whole system is idle.
4820 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004821static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004822{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004823#ifdef CONFIG_NO_HZ
4824 /*
4825 * If we were in the nohz mode recently and busy at the current
4826 * scheduler tick, then check if we need to nominate new idle
4827 * load balancer.
4828 */
4829 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4830 rq->in_nohz_recently = 0;
4831
4832 if (atomic_read(&nohz.load_balancer) == cpu) {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304833 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004834 atomic_set(&nohz.load_balancer, -1);
4835 }
4836
4837 if (atomic_read(&nohz.load_balancer) == -1) {
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304838 int ilb = find_new_ilb(cpu);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004839
Mike Travis434d53b2008-04-04 18:11:04 -07004840 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004841 resched_cpu(ilb);
4842 }
4843 }
4844
4845 /*
4846 * If this cpu is idle and doing idle load balancing for all the
4847 * cpus with ticks stopped, is it time for that to stop?
4848 */
4849 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304850 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004851 resched_cpu(cpu);
4852 return;
4853 }
4854
4855 /*
4856 * If this cpu is idle and the idle load balancing is done by
4857 * someone else, then no need raise the SCHED_SOFTIRQ
4858 */
4859 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304860 cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004861 return;
4862#endif
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004863 /* Don't need to rebalance while attached to NULL domain */
4864 if (time_after_eq(jiffies, rq->next_balance) &&
4865 likely(!on_null_domain(cpu)))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004866 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004867}
Ingo Molnardd41f592007-07-09 18:51:59 +02004868
4869#else /* CONFIG_SMP */
4870
Linus Torvalds1da177e2005-04-16 15:20:36 -07004871/*
4872 * on UP we do not need to balance between CPUs:
4873 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004874static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004875{
4876}
Ingo Molnardd41f592007-07-09 18:51:59 +02004877
Linus Torvalds1da177e2005-04-16 15:20:36 -07004878#endif
4879
Linus Torvalds1da177e2005-04-16 15:20:36 -07004880DEFINE_PER_CPU(struct kernel_stat, kstat);
4881
4882EXPORT_PER_CPU_SYMBOL(kstat);
4883
4884/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004885 * Return any ns on the sched_clock that have not yet been accounted in
Frank Mayharf06febc2008-09-12 09:54:39 -07004886 * @p in case that task is currently running.
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004887 *
4888 * Called with task_rq_lock() held on @rq.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004889 */
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004890static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4891{
4892 u64 ns = 0;
4893
4894 if (task_current(rq, p)) {
4895 update_rq_clock(rq);
4896 ns = rq->clock - p->se.exec_start;
4897 if ((s64)ns < 0)
4898 ns = 0;
4899 }
4900
4901 return ns;
4902}
4903
Frank Mayharbb34d922008-09-12 09:54:39 -07004904unsigned long long task_delta_exec(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004905{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004906 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004907 struct rq *rq;
Frank Mayharbb34d922008-09-12 09:54:39 -07004908 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004909
Ingo Molnar41b86e92007-07-09 18:51:58 +02004910 rq = task_rq_lock(p, &flags);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004911 ns = do_task_delta_exec(p, rq);
4912 task_rq_unlock(rq, &flags);
Ingo Molnar15084872008-09-30 08:28:17 +02004913
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004914 return ns;
4915}
Frank Mayharf06febc2008-09-12 09:54:39 -07004916
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004917/*
4918 * Return accounted runtime for the task.
4919 * In case the task is currently running, return the runtime plus current's
4920 * pending runtime that have not been accounted yet.
4921 */
4922unsigned long long task_sched_runtime(struct task_struct *p)
4923{
4924 unsigned long flags;
4925 struct rq *rq;
4926 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004927
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004928 rq = task_rq_lock(p, &flags);
4929 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4930 task_rq_unlock(rq, &flags);
4931
4932 return ns;
4933}
4934
4935/*
4936 * Return sum_exec_runtime for the thread group.
4937 * In case the task is currently running, return the sum plus current's
4938 * pending runtime that have not been accounted yet.
4939 *
4940 * Note that the thread group might have other running tasks as well,
4941 * so the return value not includes other pending runtime that other
4942 * running tasks might have.
4943 */
4944unsigned long long thread_group_sched_runtime(struct task_struct *p)
4945{
4946 struct task_cputime totals;
4947 unsigned long flags;
4948 struct rq *rq;
4949 u64 ns;
4950
4951 rq = task_rq_lock(p, &flags);
4952 thread_group_cputime(p, &totals);
4953 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004954 task_rq_unlock(rq, &flags);
4955
4956 return ns;
4957}
4958
4959/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004960 * Account user cpu time to a process.
4961 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962 * @cputime: the cpu time spent in user space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004963 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07004964 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004965void account_user_time(struct task_struct *p, cputime_t cputime,
4966 cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004967{
4968 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4969 cputime64_t tmp;
4970
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004971 /* Add user time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004972 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004973 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07004974 account_group_user_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004975
4976 /* Add user time to cpustat. */
4977 tmp = cputime_to_cputime64(cputime);
4978 if (TASK_NICE(p) > 0)
4979 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4980 else
4981 cpustat->user = cputime64_add(cpustat->user, tmp);
Bharata B Raoef12fef2009-03-31 10:02:22 +05304982
4983 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
Jonathan Lim49b5cf32008-07-25 01:48:40 -07004984 /* Account for user time used */
4985 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004986}
4987
4988/*
Laurent Vivier94886b82007-10-15 17:00:19 +02004989 * Account guest cpu time to a process.
4990 * @p: the process that the cpu time gets accounted to
4991 * @cputime: the cpu time spent in virtual machine since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004992 * @cputime_scaled: cputime scaled by cpu frequency
Laurent Vivier94886b82007-10-15 17:00:19 +02004993 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004994static void account_guest_time(struct task_struct *p, cputime_t cputime,
4995 cputime_t cputime_scaled)
Laurent Vivier94886b82007-10-15 17:00:19 +02004996{
4997 cputime64_t tmp;
4998 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4999
5000 tmp = cputime_to_cputime64(cputime);
5001
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005002 /* Add guest time to process. */
Laurent Vivier94886b82007-10-15 17:00:19 +02005003 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005004 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005005 account_group_user_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02005006 p->gtime = cputime_add(p->gtime, cputime);
5007
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005008 /* Add guest time to cpustat. */
Laurent Vivier94886b82007-10-15 17:00:19 +02005009 cpustat->user = cputime64_add(cpustat->user, tmp);
5010 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5011}
5012
5013/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005014 * Account system cpu time to a process.
5015 * @p: the process that the cpu time gets accounted to
5016 * @hardirq_offset: the offset to subtract from hardirq_count()
5017 * @cputime: the cpu time spent in kernel space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005018 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07005019 */
5020void account_system_time(struct task_struct *p, int hardirq_offset,
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005021 cputime_t cputime, cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005022{
5023 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005024 cputime64_t tmp;
5025
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005026 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005027 account_guest_time(p, cputime, cputime_scaled);
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005028 return;
5029 }
Laurent Vivier94886b82007-10-15 17:00:19 +02005030
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005031 /* Add system time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005032 p->stime = cputime_add(p->stime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005033 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005034 account_group_system_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005035
5036 /* Add system time to cpustat. */
5037 tmp = cputime_to_cputime64(cputime);
5038 if (hardirq_count() - hardirq_offset)
5039 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5040 else if (softirq_count())
5041 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005042 else
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005043 cpustat->system = cputime64_add(cpustat->system, tmp);
5044
Bharata B Raoef12fef2009-03-31 10:02:22 +05305045 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5046
Linus Torvalds1da177e2005-04-16 15:20:36 -07005047 /* Account for system time used */
5048 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005049}
5050
5051/*
5052 * Account for involuntary wait time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005053 * @steal: the cpu time spent in involuntary wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005054 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005055void account_steal_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005056{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005058 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5059
5060 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005061}
5062
Christoph Lameter7835b982006-12-10 02:20:22 -08005063/*
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005064 * Account for idle time.
5065 * @cputime: the cpu time spent in idle wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005067void account_idle_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005068{
5069 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005070 cputime64_t cputime64 = cputime_to_cputime64(cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005071 struct rq *rq = this_rq();
5072
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005073 if (atomic_read(&rq->nr_iowait) > 0)
5074 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5075 else
5076 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
Christoph Lameter7835b982006-12-10 02:20:22 -08005077}
5078
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005079#ifndef CONFIG_VIRT_CPU_ACCOUNTING
5080
5081/*
5082 * Account a single tick of cpu time.
5083 * @p: the process that the cpu time gets accounted to
5084 * @user_tick: indicates if the tick is a user or a system tick
5085 */
5086void account_process_tick(struct task_struct *p, int user_tick)
5087{
5088 cputime_t one_jiffy = jiffies_to_cputime(1);
5089 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5090 struct rq *rq = this_rq();
5091
5092 if (user_tick)
5093 account_user_time(p, one_jiffy, one_jiffy_scaled);
Eric Dumazetf5f293a2009-04-29 14:44:49 +02005094 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005095 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5096 one_jiffy_scaled);
5097 else
5098 account_idle_time(one_jiffy);
5099}
5100
5101/*
5102 * Account multiple ticks of steal time.
5103 * @p: the process from which the cpu time has been stolen
5104 * @ticks: number of stolen ticks
5105 */
5106void account_steal_ticks(unsigned long ticks)
5107{
5108 account_steal_time(jiffies_to_cputime(ticks));
5109}
5110
5111/*
5112 * Account multiple ticks of idle time.
5113 * @ticks: number of stolen ticks
5114 */
5115void account_idle_ticks(unsigned long ticks)
5116{
5117 account_idle_time(jiffies_to_cputime(ticks));
5118}
5119
5120#endif
5121
Christoph Lameter7835b982006-12-10 02:20:22 -08005122/*
Balbir Singh49048622008-09-05 18:12:23 +02005123 * Use precise platform statistics if available:
5124 */
5125#ifdef CONFIG_VIRT_CPU_ACCOUNTING
5126cputime_t task_utime(struct task_struct *p)
5127{
5128 return p->utime;
5129}
5130
5131cputime_t task_stime(struct task_struct *p)
5132{
5133 return p->stime;
5134}
5135#else
5136cputime_t task_utime(struct task_struct *p)
5137{
5138 clock_t utime = cputime_to_clock_t(p->utime),
5139 total = utime + cputime_to_clock_t(p->stime);
5140 u64 temp;
5141
5142 /*
5143 * Use CFS's precise accounting:
5144 */
5145 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5146
5147 if (total) {
5148 temp *= utime;
5149 do_div(temp, total);
5150 }
5151 utime = (clock_t)temp;
5152
5153 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5154 return p->prev_utime;
5155}
5156
5157cputime_t task_stime(struct task_struct *p)
5158{
5159 clock_t stime;
5160
5161 /*
5162 * Use CFS's precise accounting. (we subtract utime from
5163 * the total, to make sure the total observed by userspace
5164 * grows monotonically - apps rely on that):
5165 */
5166 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5167 cputime_to_clock_t(task_utime(p));
5168
5169 if (stime >= 0)
5170 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5171
5172 return p->prev_stime;
5173}
5174#endif
5175
5176inline cputime_t task_gtime(struct task_struct *p)
5177{
5178 return p->gtime;
5179}
5180
5181/*
Christoph Lameter7835b982006-12-10 02:20:22 -08005182 * This function gets called by the timer code, with HZ frequency.
5183 * We call it with interrupts disabled.
5184 *
5185 * It also gets called by the fork code, when changing the parent's
5186 * timeslices.
5187 */
5188void scheduler_tick(void)
5189{
Christoph Lameter7835b982006-12-10 02:20:22 -08005190 int cpu = smp_processor_id();
5191 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005192 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005193
5194 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08005195
Ingo Molnardd41f592007-07-09 18:51:59 +02005196 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005197 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02005198 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01005199 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02005200 spin_unlock(&rq->lock);
5201
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02005202 perf_counter_task_tick(curr, cpu);
5203
Christoph Lametere418e1c2006-12-10 02:20:23 -08005204#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02005205 rq->idle_at_tick = idle_cpu(cpu);
5206 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08005207#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005208}
5209
Lai Jiangshan132380a2009-04-02 14:18:25 +08005210notrace unsigned long get_parent_ip(unsigned long addr)
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005211{
5212 if (in_lock_functions(addr)) {
5213 addr = CALLER_ADDR2;
5214 if (in_lock_functions(addr))
5215 addr = CALLER_ADDR3;
5216 }
5217 return addr;
5218}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005219
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05005220#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5221 defined(CONFIG_PREEMPT_TRACER))
5222
Srinivasa Ds43627582008-02-23 15:24:04 -08005223void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005224{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005225#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005226 /*
5227 * Underflow?
5228 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005229 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5230 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005231#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005232 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005233#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005234 /*
5235 * Spinlock count overflowing soon?
5236 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005237 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5238 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005239#endif
5240 if (preempt_count() == val)
5241 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005242}
5243EXPORT_SYMBOL(add_preempt_count);
5244
Srinivasa Ds43627582008-02-23 15:24:04 -08005245void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005246{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005247#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005248 /*
5249 * Underflow?
5250 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01005251 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005252 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005253 /*
5254 * Is the spinlock portion underflowing?
5255 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005256 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5257 !(preempt_count() & PREEMPT_MASK)))
5258 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005259#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005260
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005261 if (preempt_count() == val)
5262 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005263 preempt_count() -= val;
5264}
5265EXPORT_SYMBOL(sub_preempt_count);
5266
5267#endif
5268
5269/*
Ingo Molnardd41f592007-07-09 18:51:59 +02005270 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005271 */
Ingo Molnardd41f592007-07-09 18:51:59 +02005272static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005273{
Satyam Sharma838225b2007-10-24 18:23:50 +02005274 struct pt_regs *regs = get_irq_regs();
5275
5276 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5277 prev->comm, prev->pid, preempt_count());
5278
Ingo Molnardd41f592007-07-09 18:51:59 +02005279 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07005280 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02005281 if (irqs_disabled())
5282 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02005283
5284 if (regs)
5285 show_regs(regs);
5286 else
5287 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02005288}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005289
Ingo Molnardd41f592007-07-09 18:51:59 +02005290/*
5291 * Various schedule()-time debugging checks and statistics:
5292 */
5293static inline void schedule_debug(struct task_struct *prev)
5294{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005295 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005296 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07005297 * schedule() atomically, we ignore that path for now.
5298 * Otherwise, whine if we are scheduling when we should not be.
5299 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02005300 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02005301 __schedule_bug(prev);
5302
Linus Torvalds1da177e2005-04-16 15:20:36 -07005303 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5304
Ingo Molnar2d723762007-10-15 17:00:12 +02005305 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005306#ifdef CONFIG_SCHEDSTATS
5307 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02005308 schedstat_inc(this_rq(), bkl_count);
5309 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005310 }
5311#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005312}
5313
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005314static void put_prev_task(struct rq *rq, struct task_struct *p)
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005315{
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005316 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005317
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005318 update_avg(&p->se.avg_running, runtime);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005319
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005320 if (p->state == TASK_RUNNING) {
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005321 /*
5322 * In order to avoid avg_overlap growing stale when we are
5323 * indeed overlapping and hence not getting put to sleep, grow
5324 * the avg_overlap on preemption.
5325 *
5326 * We use the average preemption runtime because that
5327 * correlates to the amount of cache footprint a task can
5328 * build up.
5329 */
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005330 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5331 update_avg(&p->se.avg_overlap, runtime);
5332 } else {
5333 update_avg(&p->se.avg_running, 0);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005334 }
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005335 p->sched_class->put_prev_task(rq, p);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005336}
5337
Ingo Molnardd41f592007-07-09 18:51:59 +02005338/*
5339 * Pick up the highest-prio task:
5340 */
5341static inline struct task_struct *
Wang Chenb67802e2009-03-02 13:55:26 +08005342pick_next_task(struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02005343{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02005344 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02005345 struct task_struct *p;
5346
5347 /*
5348 * Optimization: we know that if all tasks are in
5349 * the fair class we can call that function directly:
5350 */
5351 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005352 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005353 if (likely(p))
5354 return p;
5355 }
5356
5357 class = sched_class_highest;
5358 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005359 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005360 if (p)
5361 return p;
5362 /*
5363 * Will never be NULL as the idle class always
5364 * returns a non-NULL p:
5365 */
5366 class = class->next;
5367 }
5368}
5369
5370/*
5371 * schedule() is the main scheduler function.
5372 */
Peter Zijlstraff743342009-03-13 12:21:26 +01005373asmlinkage void __sched schedule(void)
Ingo Molnardd41f592007-07-09 18:51:59 +02005374{
5375 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08005376 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02005377 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02005378 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02005379
Peter Zijlstraff743342009-03-13 12:21:26 +01005380need_resched:
5381 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02005382 cpu = smp_processor_id();
5383 rq = cpu_rq(cpu);
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07005384 rcu_sched_qs(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005385 prev = rq->curr;
5386 switch_count = &prev->nivcsw;
5387
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388 release_kernel_lock(prev);
5389need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005390
Ingo Molnardd41f592007-07-09 18:51:59 +02005391 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005392
Peter Zijlstra31656512008-07-18 18:01:23 +02005393 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02005394 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005395
Peter Zijlstra8cd162c2008-10-15 20:37:23 +02005396 spin_lock_irq(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005397 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02005398 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005399
Ingo Molnardd41f592007-07-09 18:51:59 +02005400 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04005401 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02005402 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04005403 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005404 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02005405 switch_count = &prev->nvcsw;
5406 }
5407
Gregory Haskins3f029d32009-07-29 11:08:47 -04005408 pre_schedule(rq, prev);
Steven Rostedtf65eda42008-01-25 21:08:07 +01005409
Ingo Molnardd41f592007-07-09 18:51:59 +02005410 if (unlikely(!rq->nr_running))
5411 idle_balance(cpu, rq);
5412
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005413 put_prev_task(rq, prev);
Wang Chenb67802e2009-03-02 13:55:26 +08005414 next = pick_next_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005415
Linus Torvalds1da177e2005-04-16 15:20:36 -07005416 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01005417 sched_info_switch(prev, next);
Paul Mackerras564c2b22009-05-22 14:27:22 +10005418 perf_counter_task_sched_out(prev, next, cpu);
David Simner673a90a2008-04-29 10:08:59 +01005419
Linus Torvalds1da177e2005-04-16 15:20:36 -07005420 rq->nr_switches++;
5421 rq->curr = next;
5422 ++*switch_count;
5423
Ingo Molnardd41f592007-07-09 18:51:59 +02005424 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005425 /*
5426 * the context switch might have flipped the stack from under
5427 * us, hence refresh the local variables.
5428 */
5429 cpu = smp_processor_id();
5430 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005431 } else
5432 spin_unlock_irq(&rq->lock);
5433
Gregory Haskins3f029d32009-07-29 11:08:47 -04005434 post_schedule(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005435
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005436 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005437 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005438
Linus Torvalds1da177e2005-04-16 15:20:36 -07005439 preempt_enable_no_resched();
Peter Zijlstraff743342009-03-13 12:21:26 +01005440 if (need_resched())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005441 goto need_resched;
5442}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005443EXPORT_SYMBOL(schedule);
5444
Peter Zijlstra0d66bf62009-01-12 14:01:47 +01005445#ifdef CONFIG_SMP
5446/*
5447 * Look out! "owner" is an entirely speculative pointer
5448 * access and not reliable.
5449 */
5450int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5451{
5452 unsigned int cpu;
5453 struct rq *rq;
5454
5455 if (!sched_feat(OWNER_SPIN))
5456 return 0;
5457
5458#ifdef CONFIG_DEBUG_PAGEALLOC
5459 /*
5460 * Need to access the cpu field knowing that
5461 * DEBUG_PAGEALLOC could have unmapped it if
5462 * the mutex owner just released it and exited.
5463 */
5464 if (probe_kernel_address(&owner->cpu, cpu))
5465 goto out;
5466#else
5467 cpu = owner->cpu;
5468#endif
5469
5470 /*
5471 * Even if the access succeeded (likely case),
5472 * the cpu field may no longer be valid.
5473 */
5474 if (cpu >= nr_cpumask_bits)
5475 goto out;
5476
5477 /*
5478 * We need to validate that we can do a
5479 * get_cpu() and that we have the percpu area.
5480 */
5481 if (!cpu_online(cpu))
5482 goto out;
5483
5484 rq = cpu_rq(cpu);
5485
5486 for (;;) {
5487 /*
5488 * Owner changed, break to re-assess state.
5489 */
5490 if (lock->owner != owner)
5491 break;
5492
5493 /*
5494 * Is that owner really running on that cpu?
5495 */
5496 if (task_thread_info(rq->curr) != owner || need_resched())
5497 return 0;
5498
5499 cpu_relax();
5500 }
5501out:
5502 return 1;
5503}
5504#endif
5505
Linus Torvalds1da177e2005-04-16 15:20:36 -07005506#ifdef CONFIG_PREEMPT
5507/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005508 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005509 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07005510 * occur there and call schedule directly.
5511 */
5512asmlinkage void __sched preempt_schedule(void)
5513{
5514 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005515
Linus Torvalds1da177e2005-04-16 15:20:36 -07005516 /*
5517 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005518 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07005519 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07005520 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005521 return;
5522
Andi Kleen3a5c3592007-10-15 17:00:14 +02005523 do {
5524 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005525 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005526 sub_preempt_count(PREEMPT_ACTIVE);
5527
5528 /*
5529 * Check again in case we missed a preemption opportunity
5530 * between schedule and now.
5531 */
5532 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005533 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005534}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005535EXPORT_SYMBOL(preempt_schedule);
5536
5537/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005538 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07005539 * off of irq context.
5540 * Note, that this is called and return with irqs disabled. This will
5541 * protect us against recursive calling from irq.
5542 */
5543asmlinkage void __sched preempt_schedule_irq(void)
5544{
5545 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005546
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005547 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005548 BUG_ON(ti->preempt_count || !irqs_disabled());
5549
Andi Kleen3a5c3592007-10-15 17:00:14 +02005550 do {
5551 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005552 local_irq_enable();
5553 schedule();
5554 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005555 sub_preempt_count(PREEMPT_ACTIVE);
5556
5557 /*
5558 * Check again in case we missed a preemption opportunity
5559 * between schedule and now.
5560 */
5561 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005562 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005563}
5564
5565#endif /* CONFIG_PREEMPT */
5566
Peter Zijlstra63859d42009-09-15 19:14:42 +02005567int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005568 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005569{
Peter Zijlstra63859d42009-09-15 19:14:42 +02005570 return try_to_wake_up(curr->private, mode, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005571}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005572EXPORT_SYMBOL(default_wake_function);
5573
5574/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005575 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5576 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07005577 * number) then we wake all the non-exclusive tasks and one exclusive task.
5578 *
5579 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005580 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07005581 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5582 */
Johannes Weiner78ddb082009-04-14 16:53:05 +02005583static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
Peter Zijlstra63859d42009-09-15 19:14:42 +02005584 int nr_exclusive, int wake_flags, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005585{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005586 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005587
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005588 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005589 unsigned flags = curr->flags;
5590
Peter Zijlstra63859d42009-09-15 19:14:42 +02005591 if (curr->func(curr, mode, wake_flags, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07005592 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005593 break;
5594 }
5595}
5596
5597/**
5598 * __wake_up - wake up threads blocked on a waitqueue.
5599 * @q: the waitqueue
5600 * @mode: which threads
5601 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07005602 * @key: is directly passed to the wakeup function
David Howells50fa6102009-04-28 15:01:38 +01005603 *
5604 * It may be assumed that this function implies a write memory barrier before
5605 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005607void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005608 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005609{
5610 unsigned long flags;
5611
5612 spin_lock_irqsave(&q->lock, flags);
5613 __wake_up_common(q, mode, nr_exclusive, 0, key);
5614 spin_unlock_irqrestore(&q->lock, flags);
5615}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005616EXPORT_SYMBOL(__wake_up);
5617
5618/*
5619 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5620 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005621void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005622{
5623 __wake_up_common(q, mode, 1, 0, NULL);
5624}
5625
Davide Libenzi4ede8162009-03-31 15:24:20 -07005626void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5627{
5628 __wake_up_common(q, mode, 1, 0, key);
5629}
5630
Linus Torvalds1da177e2005-04-16 15:20:36 -07005631/**
Davide Libenzi4ede8162009-03-31 15:24:20 -07005632 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005633 * @q: the waitqueue
5634 * @mode: which threads
5635 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Davide Libenzi4ede8162009-03-31 15:24:20 -07005636 * @key: opaque value to be passed to wakeup targets
Linus Torvalds1da177e2005-04-16 15:20:36 -07005637 *
5638 * The sync wakeup differs that the waker knows that it will schedule
5639 * away soon, so while the target thread will be woken up, it will not
5640 * be migrated to another CPU - ie. the two threads are 'synchronized'
5641 * with each other. This can prevent needless bouncing between CPUs.
5642 *
5643 * On UP it can prevent extra preemption.
David Howells50fa6102009-04-28 15:01:38 +01005644 *
5645 * It may be assumed that this function implies a write memory barrier before
5646 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005647 */
Davide Libenzi4ede8162009-03-31 15:24:20 -07005648void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5649 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005650{
5651 unsigned long flags;
Peter Zijlstra7d478722009-09-14 19:55:44 +02005652 int wake_flags = WF_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005653
5654 if (unlikely(!q))
5655 return;
5656
5657 if (unlikely(!nr_exclusive))
Peter Zijlstra7d478722009-09-14 19:55:44 +02005658 wake_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005659
5660 spin_lock_irqsave(&q->lock, flags);
Peter Zijlstra7d478722009-09-14 19:55:44 +02005661 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005662 spin_unlock_irqrestore(&q->lock, flags);
5663}
Davide Libenzi4ede8162009-03-31 15:24:20 -07005664EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5665
5666/*
5667 * __wake_up_sync - see __wake_up_sync_key()
5668 */
5669void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5670{
5671 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5672}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005673EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5674
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005675/**
5676 * complete: - signals a single thread waiting on this completion
5677 * @x: holds the state of this particular completion
5678 *
5679 * This will wake up a single thread waiting on this completion. Threads will be
5680 * awakened in the same order in which they were queued.
5681 *
5682 * See also complete_all(), wait_for_completion() and related routines.
David Howells50fa6102009-04-28 15:01:38 +01005683 *
5684 * It may be assumed that this function implies a write memory barrier before
5685 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005686 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005687void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005688{
5689 unsigned long flags;
5690
5691 spin_lock_irqsave(&x->wait.lock, flags);
5692 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005693 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005694 spin_unlock_irqrestore(&x->wait.lock, flags);
5695}
5696EXPORT_SYMBOL(complete);
5697
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005698/**
5699 * complete_all: - signals all threads waiting on this completion
5700 * @x: holds the state of this particular completion
5701 *
5702 * This will wake up all threads waiting on this particular completion event.
David Howells50fa6102009-04-28 15:01:38 +01005703 *
5704 * It may be assumed that this function implies a write memory barrier before
5705 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005706 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005707void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005708{
5709 unsigned long flags;
5710
5711 spin_lock_irqsave(&x->wait.lock, flags);
5712 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005713 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005714 spin_unlock_irqrestore(&x->wait.lock, flags);
5715}
5716EXPORT_SYMBOL(complete_all);
5717
Andi Kleen8cbbe862007-10-15 17:00:14 +02005718static inline long __sched
5719do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005720{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005721 if (!x->done) {
5722 DECLARE_WAITQUEUE(wait, current);
5723
5724 wait.flags |= WQ_FLAG_EXCLUSIVE;
5725 __add_wait_queue_tail(&x->wait, &wait);
5726 do {
Oleg Nesterov94d3d822008-08-20 16:54:41 -07005727 if (signal_pending_state(state, current)) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04005728 timeout = -ERESTARTSYS;
5729 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005730 }
5731 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005733 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005734 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005735 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005736 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005737 if (!x->done)
5738 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005739 }
5740 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04005741 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005742}
5743
5744static long __sched
5745wait_for_common(struct completion *x, long timeout, int state)
5746{
5747 might_sleep();
5748
5749 spin_lock_irq(&x->wait.lock);
5750 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005751 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005752 return timeout;
5753}
5754
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005755/**
5756 * wait_for_completion: - waits for completion of a task
5757 * @x: holds the state of this particular completion
5758 *
5759 * This waits to be signaled for completion of a specific task. It is NOT
5760 * interruptible and there is no timeout.
5761 *
5762 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5763 * and interrupt capability. Also see complete().
5764 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005765void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02005766{
5767 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005768}
5769EXPORT_SYMBOL(wait_for_completion);
5770
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005771/**
5772 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5773 * @x: holds the state of this particular completion
5774 * @timeout: timeout value in jiffies
5775 *
5776 * This waits for either a completion of a specific task to be signaled or for a
5777 * specified timeout to expire. The timeout is in jiffies. It is not
5778 * interruptible.
5779 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005780unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005781wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5782{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005783 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005784}
5785EXPORT_SYMBOL(wait_for_completion_timeout);
5786
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005787/**
5788 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5789 * @x: holds the state of this particular completion
5790 *
5791 * This waits for completion of a specific task to be signaled. It is
5792 * interruptible.
5793 */
Andi Kleen8cbbe862007-10-15 17:00:14 +02005794int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005795{
Andi Kleen51e97992007-10-18 21:32:55 +02005796 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5797 if (t == -ERESTARTSYS)
5798 return t;
5799 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005800}
5801EXPORT_SYMBOL(wait_for_completion_interruptible);
5802
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005803/**
5804 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5805 * @x: holds the state of this particular completion
5806 * @timeout: timeout value in jiffies
5807 *
5808 * This waits for either a completion of a specific task to be signaled or for a
5809 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5810 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005811unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005812wait_for_completion_interruptible_timeout(struct completion *x,
5813 unsigned long timeout)
5814{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005815 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005816}
5817EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5818
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005819/**
5820 * wait_for_completion_killable: - waits for completion of a task (killable)
5821 * @x: holds the state of this particular completion
5822 *
5823 * This waits to be signaled for completion of a specific task. It can be
5824 * interrupted by a kill signal.
5825 */
Matthew Wilcox009e5772007-12-06 12:29:54 -05005826int __sched wait_for_completion_killable(struct completion *x)
5827{
5828 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5829 if (t == -ERESTARTSYS)
5830 return t;
5831 return 0;
5832}
5833EXPORT_SYMBOL(wait_for_completion_killable);
5834
Dave Chinnerbe4de352008-08-15 00:40:44 -07005835/**
5836 * try_wait_for_completion - try to decrement a completion without blocking
5837 * @x: completion structure
5838 *
5839 * Returns: 0 if a decrement cannot be done without blocking
5840 * 1 if a decrement succeeded.
5841 *
5842 * If a completion is being used as a counting completion,
5843 * attempt to decrement the counter without blocking. This
5844 * enables us to avoid waiting if the resource the completion
5845 * is protecting is not available.
5846 */
5847bool try_wait_for_completion(struct completion *x)
5848{
5849 int ret = 1;
5850
5851 spin_lock_irq(&x->wait.lock);
5852 if (!x->done)
5853 ret = 0;
5854 else
5855 x->done--;
5856 spin_unlock_irq(&x->wait.lock);
5857 return ret;
5858}
5859EXPORT_SYMBOL(try_wait_for_completion);
5860
5861/**
5862 * completion_done - Test to see if a completion has any waiters
5863 * @x: completion structure
5864 *
5865 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5866 * 1 if there are no waiters.
5867 *
5868 */
5869bool completion_done(struct completion *x)
5870{
5871 int ret = 1;
5872
5873 spin_lock_irq(&x->wait.lock);
5874 if (!x->done)
5875 ret = 0;
5876 spin_unlock_irq(&x->wait.lock);
5877 return ret;
5878}
5879EXPORT_SYMBOL(completion_done);
5880
Andi Kleen8cbbe862007-10-15 17:00:14 +02005881static long __sched
5882sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02005883{
5884 unsigned long flags;
5885 wait_queue_t wait;
5886
5887 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005888
Andi Kleen8cbbe862007-10-15 17:00:14 +02005889 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005890
Andi Kleen8cbbe862007-10-15 17:00:14 +02005891 spin_lock_irqsave(&q->lock, flags);
5892 __add_wait_queue(q, &wait);
5893 spin_unlock(&q->lock);
5894 timeout = schedule_timeout(timeout);
5895 spin_lock_irq(&q->lock);
5896 __remove_wait_queue(q, &wait);
5897 spin_unlock_irqrestore(&q->lock, flags);
5898
5899 return timeout;
5900}
5901
5902void __sched interruptible_sleep_on(wait_queue_head_t *q)
5903{
5904 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005905}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005906EXPORT_SYMBOL(interruptible_sleep_on);
5907
Ingo Molnar0fec1712007-07-09 18:52:01 +02005908long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005909interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005910{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005911 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005912}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005913EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5914
Ingo Molnar0fec1712007-07-09 18:52:01 +02005915void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005916{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005917 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005918}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005919EXPORT_SYMBOL(sleep_on);
5920
Ingo Molnar0fec1712007-07-09 18:52:01 +02005921long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005922{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005923 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005924}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005925EXPORT_SYMBOL(sleep_on_timeout);
5926
Ingo Molnarb29739f2006-06-27 02:54:51 -07005927#ifdef CONFIG_RT_MUTEXES
5928
5929/*
5930 * rt_mutex_setprio - set the current priority of a task
5931 * @p: task
5932 * @prio: prio value (kernel-internal form)
5933 *
5934 * This function changes the 'effective' priority of a task. It does
5935 * not touch ->normal_prio like __setscheduler().
5936 *
5937 * Used by the rt_mutex code to implement priority inheritance logic.
5938 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005939void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07005940{
5941 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02005942 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005943 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01005944 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005945
5946 BUG_ON(prio < 0 || prio > MAX_PRIO);
5947
5948 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005949 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07005950
Andrew Mortond5f9f942007-05-08 20:27:06 -07005951 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02005952 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01005953 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005954 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02005955 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005956 if (running)
5957 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02005958
5959 if (rt_prio(prio))
5960 p->sched_class = &rt_sched_class;
5961 else
5962 p->sched_class = &fair_sched_class;
5963
Ingo Molnarb29739f2006-06-27 02:54:51 -07005964 p->prio = prio;
5965
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005966 if (running)
5967 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005968 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02005969 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01005970
5971 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07005972 }
5973 task_rq_unlock(rq, &flags);
5974}
5975
5976#endif
5977
Ingo Molnar36c8b582006-07-03 00:25:41 -07005978void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005979{
Ingo Molnardd41f592007-07-09 18:51:59 +02005980 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005981 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005982 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005983
5984 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5985 return;
5986 /*
5987 * We have to be careful, if called from sys_setpriority(),
5988 * the task might be in the middle of scheduling on another CPU.
5989 */
5990 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005991 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005992 /*
5993 * The RT priorities are set via sched_setscheduler(), but we still
5994 * allow the 'normal' nice value to be set - but as expected
5995 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02005996 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005997 */
Ingo Molnare05606d2007-07-09 18:51:59 +02005998 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005999 p->static_prio = NICE_TO_PRIO(nice);
6000 goto out_unlock;
6001 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006002 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02006003 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02006004 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006005
Linus Torvalds1da177e2005-04-16 15:20:36 -07006006 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07006007 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006008 old_prio = p->prio;
6009 p->prio = effective_prio(p);
6010 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006011
Ingo Molnardd41f592007-07-09 18:51:59 +02006012 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02006013 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006014 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07006015 * If the task increased its priority or is running and
6016 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006017 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07006018 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006019 resched_task(rq->curr);
6020 }
6021out_unlock:
6022 task_rq_unlock(rq, &flags);
6023}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006024EXPORT_SYMBOL(set_user_nice);
6025
Matt Mackalle43379f2005-05-01 08:59:00 -07006026/*
6027 * can_nice - check if a task can reduce its nice value
6028 * @p: task
6029 * @nice: nice value
6030 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006031int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07006032{
Matt Mackall024f4742005-08-18 11:24:19 -07006033 /* convert nice value [19,-20] to rlimit style value [1,40] */
6034 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006035
Matt Mackalle43379f2005-05-01 08:59:00 -07006036 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6037 capable(CAP_SYS_NICE));
6038}
6039
Linus Torvalds1da177e2005-04-16 15:20:36 -07006040#ifdef __ARCH_WANT_SYS_NICE
6041
6042/*
6043 * sys_nice - change the priority of the current process.
6044 * @increment: priority increment
6045 *
6046 * sys_setpriority is a more generic, but much slower function that
6047 * does similar things.
6048 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006049SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006050{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006051 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006052
6053 /*
6054 * Setpriority might change our priority at the same moment.
6055 * We don't have to worry. Conceptually one call occurs first
6056 * and we have a single winner.
6057 */
Matt Mackalle43379f2005-05-01 08:59:00 -07006058 if (increment < -40)
6059 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006060 if (increment > 40)
6061 increment = 40;
6062
Américo Wang2b8f8362009-02-16 18:54:21 +08006063 nice = TASK_NICE(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006064 if (nice < -20)
6065 nice = -20;
6066 if (nice > 19)
6067 nice = 19;
6068
Matt Mackalle43379f2005-05-01 08:59:00 -07006069 if (increment < 0 && !can_nice(current, nice))
6070 return -EPERM;
6071
Linus Torvalds1da177e2005-04-16 15:20:36 -07006072 retval = security_task_setnice(current, nice);
6073 if (retval)
6074 return retval;
6075
6076 set_user_nice(current, nice);
6077 return 0;
6078}
6079
6080#endif
6081
6082/**
6083 * task_prio - return the priority value of a given task.
6084 * @p: the task in question.
6085 *
6086 * This is the priority value as seen by users in /proc.
6087 * RT tasks are offset by -200. Normal tasks are centered
6088 * around 0, value goes from -16 to +15.
6089 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006090int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006091{
6092 return p->prio - MAX_RT_PRIO;
6093}
6094
6095/**
6096 * task_nice - return the nice value of a given task.
6097 * @p: the task in question.
6098 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006099int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006100{
6101 return TASK_NICE(p);
6102}
Pavel Roskin150d8be2008-03-05 16:56:37 -05006103EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006104
6105/**
6106 * idle_cpu - is a given cpu idle currently?
6107 * @cpu: the processor in question.
6108 */
6109int idle_cpu(int cpu)
6110{
6111 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6112}
6113
Linus Torvalds1da177e2005-04-16 15:20:36 -07006114/**
6115 * idle_task - return the idle task for a given cpu.
6116 * @cpu: the processor in question.
6117 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006118struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006119{
6120 return cpu_rq(cpu)->idle;
6121}
6122
6123/**
6124 * find_process_by_pid - find a process with a matching PID value.
6125 * @pid: the pid in question.
6126 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02006127static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006128{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07006129 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006130}
6131
6132/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02006133static void
6134__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006135{
Ingo Molnardd41f592007-07-09 18:51:59 +02006136 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006137
Linus Torvalds1da177e2005-04-16 15:20:36 -07006138 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02006139 switch (p->policy) {
6140 case SCHED_NORMAL:
6141 case SCHED_BATCH:
6142 case SCHED_IDLE:
6143 p->sched_class = &fair_sched_class;
6144 break;
6145 case SCHED_FIFO:
6146 case SCHED_RR:
6147 p->sched_class = &rt_sched_class;
6148 break;
6149 }
6150
Linus Torvalds1da177e2005-04-16 15:20:36 -07006151 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006152 p->normal_prio = normal_prio(p);
6153 /* we are holding p->pi_lock already */
6154 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07006155 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006156}
6157
David Howellsc69e8d92008-11-14 10:39:19 +11006158/*
6159 * check the target process has a UID that matches the current process's
6160 */
6161static bool check_same_owner(struct task_struct *p)
6162{
6163 const struct cred *cred = current_cred(), *pcred;
6164 bool match;
6165
6166 rcu_read_lock();
6167 pcred = __task_cred(p);
6168 match = (cred->euid == pcred->euid ||
6169 cred->euid == pcred->uid);
6170 rcu_read_unlock();
6171 return match;
6172}
6173
Rusty Russell961ccdd2008-06-23 13:55:38 +10006174static int __sched_setscheduler(struct task_struct *p, int policy,
6175 struct sched_param *param, bool user)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006176{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006177 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006178 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01006179 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006180 struct rq *rq;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006181 int reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006182
Steven Rostedt66e53932006-06-27 02:54:44 -07006183 /* may grab non-irq protected spin_locks */
6184 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07006185recheck:
6186 /* double check policy once rq lock held */
Lennart Poetteringca94c442009-06-15 17:17:47 +02006187 if (policy < 0) {
6188 reset_on_fork = p->sched_reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006189 policy = oldpolicy = p->policy;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006190 } else {
6191 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6192 policy &= ~SCHED_RESET_ON_FORK;
6193
6194 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6195 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6196 policy != SCHED_IDLE)
6197 return -EINVAL;
6198 }
6199
Linus Torvalds1da177e2005-04-16 15:20:36 -07006200 /*
6201 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02006202 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6203 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006204 */
6205 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006206 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04006207 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006208 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02006209 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006210 return -EINVAL;
6211
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006212 /*
6213 * Allow unprivileged RT tasks to decrease priority:
6214 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10006215 if (user && !capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02006216 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006217 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006218
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006219 if (!lock_task_sighand(p, &flags))
6220 return -ESRCH;
6221 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6222 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006223
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006224 /* can't set/change the rt policy */
6225 if (policy != p->policy && !rlim_rtprio)
6226 return -EPERM;
6227
6228 /* can't increase priority */
6229 if (param->sched_priority > p->rt_priority &&
6230 param->sched_priority > rlim_rtprio)
6231 return -EPERM;
6232 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006233 /*
6234 * Like positive nice levels, dont allow tasks to
6235 * move out of SCHED_IDLE either:
6236 */
6237 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6238 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006239
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006240 /* can't change other user's priorities */
David Howellsc69e8d92008-11-14 10:39:19 +11006241 if (!check_same_owner(p))
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006242 return -EPERM;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006243
6244 /* Normal users shall not reset the sched_reset_on_fork flag */
6245 if (p->sched_reset_on_fork && !reset_on_fork)
6246 return -EPERM;
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006247 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006248
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006249 if (user) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006250#ifdef CONFIG_RT_GROUP_SCHED
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006251 /*
6252 * Do not allow realtime tasks into groups that have no runtime
6253 * assigned.
6254 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02006255 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6256 task_group(p)->rt_bandwidth.rt_runtime == 0)
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006257 return -EPERM;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006258#endif
6259
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006260 retval = security_task_setscheduler(p, policy, param);
6261 if (retval)
6262 return retval;
6263 }
6264
Linus Torvalds1da177e2005-04-16 15:20:36 -07006265 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07006266 * make sure no PI-waiters arrive (or leave) while we are
6267 * changing the priority of the task:
6268 */
6269 spin_lock_irqsave(&p->pi_lock, flags);
6270 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07006271 * To be able to change p->policy safely, the apropriate
6272 * runqueue lock must be held.
6273 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07006274 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006275 /* recheck policy now with rq lock held */
6276 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6277 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006278 __task_rq_unlock(rq);
6279 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006280 goto recheck;
6281 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02006282 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006283 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006284 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006285 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006286 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006287 if (running)
6288 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02006289
Lennart Poetteringca94c442009-06-15 17:17:47 +02006290 p->sched_reset_on_fork = reset_on_fork;
6291
Linus Torvalds1da177e2005-04-16 15:20:36 -07006292 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006293 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b532052007-10-15 17:00:08 +02006294
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006295 if (running)
6296 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006297 if (on_rq) {
6298 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006299
6300 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006301 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07006302 __task_rq_unlock(rq);
6303 spin_unlock_irqrestore(&p->pi_lock, flags);
6304
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07006305 rt_mutex_adjust_pi(p);
6306
Linus Torvalds1da177e2005-04-16 15:20:36 -07006307 return 0;
6308}
Rusty Russell961ccdd2008-06-23 13:55:38 +10006309
6310/**
6311 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6312 * @p: the task in question.
6313 * @policy: new policy.
6314 * @param: structure containing the new RT priority.
6315 *
6316 * NOTE that the task may be already dead.
6317 */
6318int sched_setscheduler(struct task_struct *p, int policy,
6319 struct sched_param *param)
6320{
6321 return __sched_setscheduler(p, policy, param, true);
6322}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006323EXPORT_SYMBOL_GPL(sched_setscheduler);
6324
Rusty Russell961ccdd2008-06-23 13:55:38 +10006325/**
6326 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6327 * @p: the task in question.
6328 * @policy: new policy.
6329 * @param: structure containing the new RT priority.
6330 *
6331 * Just like sched_setscheduler, only don't bother checking if the
6332 * current context has permission. For example, this is needed in
6333 * stop_machine(): we create temporary high priority worker threads,
6334 * but our caller might not have that capability.
6335 */
6336int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6337 struct sched_param *param)
6338{
6339 return __sched_setscheduler(p, policy, param, false);
6340}
6341
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006342static int
6343do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006344{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006345 struct sched_param lparam;
6346 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006347 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006348
6349 if (!param || pid < 0)
6350 return -EINVAL;
6351 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6352 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006353
6354 rcu_read_lock();
6355 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006356 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006357 if (p != NULL)
6358 retval = sched_setscheduler(p, policy, &lparam);
6359 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07006360
Linus Torvalds1da177e2005-04-16 15:20:36 -07006361 return retval;
6362}
6363
6364/**
6365 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6366 * @pid: the pid in question.
6367 * @policy: new policy.
6368 * @param: structure containing the new RT priority.
6369 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006370SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6371 struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006372{
Jason Baronc21761f2006-01-18 17:43:03 -08006373 /* negative values for policy are not valid */
6374 if (policy < 0)
6375 return -EINVAL;
6376
Linus Torvalds1da177e2005-04-16 15:20:36 -07006377 return do_sched_setscheduler(pid, policy, param);
6378}
6379
6380/**
6381 * sys_sched_setparam - set/change the RT priority of a thread
6382 * @pid: the pid in question.
6383 * @param: structure containing the new RT priority.
6384 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006385SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006386{
6387 return do_sched_setscheduler(pid, -1, param);
6388}
6389
6390/**
6391 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6392 * @pid: the pid in question.
6393 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006394SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006395{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006396 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006397 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006398
6399 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006400 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006401
6402 retval = -ESRCH;
6403 read_lock(&tasklist_lock);
6404 p = find_process_by_pid(pid);
6405 if (p) {
6406 retval = security_task_getscheduler(p);
6407 if (!retval)
Lennart Poetteringca94c442009-06-15 17:17:47 +02006408 retval = p->policy
6409 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006410 }
6411 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006412 return retval;
6413}
6414
6415/**
Lennart Poetteringca94c442009-06-15 17:17:47 +02006416 * sys_sched_getparam - get the RT priority of a thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07006417 * @pid: the pid in question.
6418 * @param: structure containing the RT priority.
6419 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006420SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006421{
6422 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006423 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006424 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006425
6426 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006427 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006428
6429 read_lock(&tasklist_lock);
6430 p = find_process_by_pid(pid);
6431 retval = -ESRCH;
6432 if (!p)
6433 goto out_unlock;
6434
6435 retval = security_task_getscheduler(p);
6436 if (retval)
6437 goto out_unlock;
6438
6439 lp.sched_priority = p->rt_priority;
6440 read_unlock(&tasklist_lock);
6441
6442 /*
6443 * This one might sleep, we cannot do it with a spinlock held ...
6444 */
6445 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6446
Linus Torvalds1da177e2005-04-16 15:20:36 -07006447 return retval;
6448
6449out_unlock:
6450 read_unlock(&tasklist_lock);
6451 return retval;
6452}
6453
Rusty Russell96f874e22008-11-25 02:35:14 +10306454long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006455{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306456 cpumask_var_t cpus_allowed, new_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006457 struct task_struct *p;
6458 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006459
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006460 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006461 read_lock(&tasklist_lock);
6462
6463 p = find_process_by_pid(pid);
6464 if (!p) {
6465 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006466 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006467 return -ESRCH;
6468 }
6469
6470 /*
6471 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006472 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07006473 * usage count and then drop tasklist_lock.
6474 */
6475 get_task_struct(p);
6476 read_unlock(&tasklist_lock);
6477
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306478 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6479 retval = -ENOMEM;
6480 goto out_put_task;
6481 }
6482 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6483 retval = -ENOMEM;
6484 goto out_free_cpus_allowed;
6485 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006486 retval = -EPERM;
David Howellsc69e8d92008-11-14 10:39:19 +11006487 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006488 goto out_unlock;
6489
David Quigleye7834f82006-06-23 02:03:59 -07006490 retval = security_task_setscheduler(p, 0, NULL);
6491 if (retval)
6492 goto out_unlock;
6493
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306494 cpuset_cpus_allowed(p, cpus_allowed);
6495 cpumask_and(new_mask, in_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006496 again:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306497 retval = set_cpus_allowed_ptr(p, new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006498
Paul Menage8707d8b2007-10-18 23:40:22 -07006499 if (!retval) {
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306500 cpuset_cpus_allowed(p, cpus_allowed);
6501 if (!cpumask_subset(new_mask, cpus_allowed)) {
Paul Menage8707d8b2007-10-18 23:40:22 -07006502 /*
6503 * We must have raced with a concurrent cpuset
6504 * update. Just reset the cpus_allowed to the
6505 * cpuset's cpus_allowed
6506 */
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306507 cpumask_copy(new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006508 goto again;
6509 }
6510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006511out_unlock:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306512 free_cpumask_var(new_mask);
6513out_free_cpus_allowed:
6514 free_cpumask_var(cpus_allowed);
6515out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006517 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006518 return retval;
6519}
6520
6521static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e22008-11-25 02:35:14 +10306522 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006523{
Rusty Russell96f874e22008-11-25 02:35:14 +10306524 if (len < cpumask_size())
6525 cpumask_clear(new_mask);
6526 else if (len > cpumask_size())
6527 len = cpumask_size();
6528
Linus Torvalds1da177e2005-04-16 15:20:36 -07006529 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6530}
6531
6532/**
6533 * sys_sched_setaffinity - set the cpu affinity of a process
6534 * @pid: pid of the process
6535 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6536 * @user_mask_ptr: user-space pointer to the new cpu mask
6537 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006538SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6539 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006540{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306541 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006542 int retval;
6543
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306544 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6545 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006546
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306547 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6548 if (retval == 0)
6549 retval = sched_setaffinity(pid, new_mask);
6550 free_cpumask_var(new_mask);
6551 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006552}
6553
Rusty Russell96f874e22008-11-25 02:35:14 +10306554long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006555{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006556 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006557 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006558
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006559 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006560 read_lock(&tasklist_lock);
6561
6562 retval = -ESRCH;
6563 p = find_process_by_pid(pid);
6564 if (!p)
6565 goto out_unlock;
6566
David Quigleye7834f82006-06-23 02:03:59 -07006567 retval = security_task_getscheduler(p);
6568 if (retval)
6569 goto out_unlock;
6570
Rusty Russell96f874e22008-11-25 02:35:14 +10306571 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006572
6573out_unlock:
6574 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006575 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006576
Ulrich Drepper9531b622007-08-09 11:16:46 +02006577 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006578}
6579
6580/**
6581 * sys_sched_getaffinity - get the cpu affinity of a process
6582 * @pid: pid of the process
6583 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6584 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6585 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006586SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6587 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006588{
6589 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10306590 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006591
Rusty Russellf17c8602008-11-25 02:35:11 +10306592 if (len < cpumask_size())
Linus Torvalds1da177e2005-04-16 15:20:36 -07006593 return -EINVAL;
6594
Rusty Russellf17c8602008-11-25 02:35:11 +10306595 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6596 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006597
Rusty Russellf17c8602008-11-25 02:35:11 +10306598 ret = sched_getaffinity(pid, mask);
6599 if (ret == 0) {
6600 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6601 ret = -EFAULT;
6602 else
6603 ret = cpumask_size();
6604 }
6605 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006606
Rusty Russellf17c8602008-11-25 02:35:11 +10306607 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006608}
6609
6610/**
6611 * sys_sched_yield - yield the current processor to other threads.
6612 *
Ingo Molnardd41f592007-07-09 18:51:59 +02006613 * This function yields the current CPU to other tasks. If there are no
6614 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006615 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006616SYSCALL_DEFINE0(sched_yield)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006617{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006618 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619
Ingo Molnar2d723762007-10-15 17:00:12 +02006620 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02006621 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006622
6623 /*
6624 * Since we are going to call schedule() anyway, there's
6625 * no need to preempt or enable interrupts:
6626 */
6627 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07006628 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006629 _raw_spin_unlock(&rq->lock);
6630 preempt_enable_no_resched();
6631
6632 schedule();
6633
6634 return 0;
6635}
6636
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006637static inline int should_resched(void)
6638{
6639 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6640}
6641
Andrew Mortone7b38402006-06-30 01:56:00 -07006642static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006643{
Frederic Weisbeckere7aaaa62009-07-16 15:44:29 +02006644 add_preempt_count(PREEMPT_ACTIVE);
6645 schedule();
6646 sub_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006647}
6648
Herbert Xu02b67cc32008-01-25 21:08:28 +01006649int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006650{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006651 if (should_resched()) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006652 __cond_resched();
6653 return 1;
6654 }
6655 return 0;
6656}
Herbert Xu02b67cc32008-01-25 21:08:28 +01006657EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006658
6659/*
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006660 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07006661 * call schedule, and on return reacquire the lock.
6662 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006663 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07006664 * operations here to prevent schedule() from being called twice (once via
6665 * spin_unlock(), once by hand).
6666 */
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006667int __cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006668{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006669 int resched = should_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07006670 int ret = 0;
6671
Peter Zijlstraf607c662009-07-20 19:16:29 +02006672 lockdep_assert_held(lock);
6673
Nick Piggin95c354f2008-01-30 13:31:20 +01006674 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006675 spin_unlock(lock);
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006676 if (resched)
Nick Piggin95c354f2008-01-30 13:31:20 +01006677 __cond_resched();
6678 else
6679 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07006680 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006681 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006682 }
Jan Kara6df3cec2005-06-13 15:52:32 -07006683 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006684}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006685EXPORT_SYMBOL(__cond_resched_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006686
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006687int __sched __cond_resched_softirq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006688{
6689 BUG_ON(!in_softirq());
6690
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006691 if (should_resched()) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07006692 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006693 __cond_resched();
6694 local_bh_disable();
6695 return 1;
6696 }
6697 return 0;
6698}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006699EXPORT_SYMBOL(__cond_resched_softirq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006700
Linus Torvalds1da177e2005-04-16 15:20:36 -07006701/**
6702 * yield - yield the current processor to other threads.
6703 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08006704 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07006705 * thread runnable and calls sys_sched_yield().
6706 */
6707void __sched yield(void)
6708{
6709 set_current_state(TASK_RUNNING);
6710 sys_sched_yield();
6711}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006712EXPORT_SYMBOL(yield);
6713
6714/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006715 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07006716 * that process accounting knows that this is a task in IO wait state.
6717 *
6718 * But don't do that if it is a deliberate, throttling IO wait (this task
6719 * has set its backing_dev_info: the queue against which it should throttle)
6720 */
6721void __sched io_schedule(void)
6722{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006723 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006724
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006725 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006726 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006727 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006728 schedule();
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006729 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006730 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006731 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006732}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006733EXPORT_SYMBOL(io_schedule);
6734
6735long __sched io_schedule_timeout(long timeout)
6736{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006737 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006738 long ret;
6739
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006740 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006741 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006742 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006743 ret = schedule_timeout(timeout);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006744 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006745 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006746 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006747 return ret;
6748}
6749
6750/**
6751 * sys_sched_get_priority_max - return maximum RT priority.
6752 * @policy: scheduling class.
6753 *
6754 * this syscall returns the maximum rt_priority that can be used
6755 * by a given scheduling class.
6756 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006757SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006758{
6759 int ret = -EINVAL;
6760
6761 switch (policy) {
6762 case SCHED_FIFO:
6763 case SCHED_RR:
6764 ret = MAX_USER_RT_PRIO-1;
6765 break;
6766 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006767 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006768 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006769 ret = 0;
6770 break;
6771 }
6772 return ret;
6773}
6774
6775/**
6776 * sys_sched_get_priority_min - return minimum RT priority.
6777 * @policy: scheduling class.
6778 *
6779 * this syscall returns the minimum rt_priority that can be used
6780 * by a given scheduling class.
6781 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006782SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006783{
6784 int ret = -EINVAL;
6785
6786 switch (policy) {
6787 case SCHED_FIFO:
6788 case SCHED_RR:
6789 ret = 1;
6790 break;
6791 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006792 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006793 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006794 ret = 0;
6795 }
6796 return ret;
6797}
6798
6799/**
6800 * sys_sched_rr_get_interval - return the default timeslice of a process.
6801 * @pid: pid of the process.
6802 * @interval: userspace pointer to the timeslice value.
6803 *
6804 * this syscall writes the default timeslice value of a given process
6805 * into the user-space timespec buffer. A value of '0' means infinity.
6806 */
Heiko Carstens17da2bd2009-01-14 14:14:10 +01006807SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Heiko Carstens754fe8d2009-01-14 14:14:09 +01006808 struct timespec __user *, interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006809{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006810 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006811 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006812 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006813 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006814
6815 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006816 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006817
6818 retval = -ESRCH;
6819 read_lock(&tasklist_lock);
6820 p = find_process_by_pid(pid);
6821 if (!p)
6822 goto out_unlock;
6823
6824 retval = security_task_getscheduler(p);
6825 if (retval)
6826 goto out_unlock;
6827
Ingo Molnar77034932007-12-04 17:04:39 +01006828 /*
6829 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6830 * tasks that are on an otherwise idle runqueue:
6831 */
6832 time_slice = 0;
6833 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006834 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08006835 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006836 struct sched_entity *se = &p->se;
6837 unsigned long flags;
6838 struct rq *rq;
6839
6840 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01006841 if (rq->cfs.load.weight)
6842 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006843 task_rq_unlock(rq, &flags);
6844 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006845 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006846 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006847 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006848 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006849
Linus Torvalds1da177e2005-04-16 15:20:36 -07006850out_unlock:
6851 read_unlock(&tasklist_lock);
6852 return retval;
6853}
6854
Steven Rostedt7c731e02008-05-12 21:20:41 +02006855static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006856
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006857void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006858{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006859 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006860 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006861
Linus Torvalds1da177e2005-04-16 15:20:36 -07006862 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006863 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07006864 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02006865#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07006866 if (state == TASK_RUNNING)
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006867 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006868 else
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006869 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006870#else
6871 if (state == TASK_RUNNING)
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006872 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006873 else
Ingo Molnarcc4ea7952007-10-18 21:32:56 +02006874 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006875#endif
6876#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05006877 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006878#endif
David Rientjesaa47b7e2009-05-04 01:38:05 -07006879 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6880 task_pid_nr(p), task_pid_nr(p->real_parent),
6881 (unsigned long)task_thread_info(p)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006882
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01006883 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006884}
6885
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006886void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006887{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006888 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006889
Ingo Molnar4bd77322007-07-11 21:21:47 +02006890#if BITS_PER_LONG == 32
6891 printk(KERN_INFO
6892 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006893#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02006894 printk(KERN_INFO
6895 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006896#endif
6897 read_lock(&tasklist_lock);
6898 do_each_thread(g, p) {
6899 /*
6900 * reset the NMI-timeout, listing all files on a slow
6901 * console might take alot of time:
6902 */
6903 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07006904 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006905 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006906 } while_each_thread(g, p);
6907
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07006908 touch_all_softlockup_watchdogs();
6909
Ingo Molnardd41f592007-07-09 18:51:59 +02006910#ifdef CONFIG_SCHED_DEBUG
6911 sysrq_sched_debug_show();
6912#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006913 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006914 /*
6915 * Only show locks if all tasks are dumped:
6916 */
6917 if (state_filter == -1)
6918 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006919}
6920
Ingo Molnar1df21052007-07-09 18:51:58 +02006921void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6922{
Ingo Molnardd41f592007-07-09 18:51:59 +02006923 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02006924}
6925
Ingo Molnarf340c0d2005-06-28 16:40:42 +02006926/**
6927 * init_idle - set up an idle thread for a given CPU
6928 * @idle: task in question
6929 * @cpu: cpu the idle task belongs to
6930 *
6931 * NOTE: this function does not set the idle thread's NEED_RESCHED
6932 * flag, to make booting more robust.
6933 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07006934void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006935{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006936 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006937 unsigned long flags;
6938
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01006939 spin_lock_irqsave(&rq->lock, flags);
6940
Ingo Molnardd41f592007-07-09 18:51:59 +02006941 __sched_fork(idle);
6942 idle->se.exec_start = sched_clock();
6943
Ingo Molnarb29739f2006-06-27 02:54:51 -07006944 idle->prio = idle->normal_prio = MAX_PRIO;
Rusty Russell96f874e22008-11-25 02:35:14 +10306945 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
Ingo Molnardd41f592007-07-09 18:51:59 +02006946 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006947
Linus Torvalds1da177e2005-04-16 15:20:36 -07006948 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07006949#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6950 idle->oncpu = 1;
6951#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006952 spin_unlock_irqrestore(&rq->lock, flags);
6953
6954 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07006955#if defined(CONFIG_PREEMPT)
6956 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6957#else
Al Viroa1261f542005-11-13 16:06:55 -08006958 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07006959#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02006960 /*
6961 * The idle tasks have their own, simple scheduling class:
6962 */
6963 idle->sched_class = &idle_sched_class;
Frederic Weisbeckerfb526072008-11-25 21:07:04 +01006964 ftrace_graph_init_task(idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006965}
6966
6967/*
6968 * In a system that switches off the HZ timer nohz_cpu_mask
6969 * indicates which cpus entered this state. This is used
6970 * in the rcu update to wait only for active cpus. For system
6971 * which do not switch off the HZ timer nohz_cpu_mask should
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10306972 * always be CPU_BITS_NONE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006973 */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10306974cpumask_var_t nohz_cpu_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006975
Ingo Molnar19978ca2007-11-09 22:39:38 +01006976/*
6977 * Increase the granularity value when there are more CPUs,
6978 * because with more CPUs the 'effective latency' as visible
6979 * to users decreases. But the relationship is not linear,
6980 * so pick a second-best guess by going with the log2 of the
6981 * number of CPUs.
6982 *
6983 * This idea comes from the SD scheduler of Con Kolivas:
6984 */
6985static inline void sched_init_granularity(void)
6986{
6987 unsigned int factor = 1 + ilog2(num_online_cpus());
6988 const unsigned long limit = 200000000;
6989
6990 sysctl_sched_min_granularity *= factor;
6991 if (sysctl_sched_min_granularity > limit)
6992 sysctl_sched_min_granularity = limit;
6993
6994 sysctl_sched_latency *= factor;
6995 if (sysctl_sched_latency > limit)
6996 sysctl_sched_latency = limit;
6997
6998 sysctl_sched_wakeup_granularity *= factor;
Peter Zijlstra55cd5342008-08-04 08:54:26 +02006999
7000 sysctl_sched_shares_ratelimit *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01007001}
7002
Linus Torvalds1da177e2005-04-16 15:20:36 -07007003#ifdef CONFIG_SMP
7004/*
7005 * This is how migration works:
7006 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07007007 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07007008 * runqueue and wake up that CPU's migration thread.
7009 * 2) we down() the locked semaphore => thread blocks.
7010 * 3) migration thread wakes up (implicitly it forces the migrated
7011 * thread off the CPU)
7012 * 4) it gets the migration request and checks whether the migrated
7013 * task is still in the wrong runqueue.
7014 * 5) if it's in the wrong runqueue then the migration thread removes
7015 * it and puts it into the right queue.
7016 * 6) migration thread up()s the semaphore.
7017 * 7) we wake up and the migration is done.
7018 */
7019
7020/*
7021 * Change a given task's CPU affinity. Migrate the thread to a
7022 * proper CPU and schedule it away if the CPU it's executing on
7023 * is removed from the allowed bitmask.
7024 *
7025 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007026 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07007027 * call is not atomic; no spinlocks may be held.
7028 */
Rusty Russell96f874e22008-11-25 02:35:14 +10307029int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007030{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007031 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007032 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007033 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007034 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007035
7036 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e22008-11-25 02:35:14 +10307037 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007038 ret = -EINVAL;
7039 goto out;
7040 }
7041
David Rientjes9985b0b2008-06-05 12:57:11 -07007042 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
Rusty Russell96f874e22008-11-25 02:35:14 +10307043 !cpumask_equal(&p->cpus_allowed, new_mask))) {
David Rientjes9985b0b2008-06-05 12:57:11 -07007044 ret = -EINVAL;
7045 goto out;
7046 }
7047
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007048 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007049 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007050 else {
Rusty Russell96f874e22008-11-25 02:35:14 +10307051 cpumask_copy(&p->cpus_allowed, new_mask);
7052 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
Gregory Haskins73fe6aae2008-01-25 21:08:07 +01007053 }
7054
Linus Torvalds1da177e2005-04-16 15:20:36 -07007055 /* Can the task run on the task's current CPU? If so, we're done */
Rusty Russell96f874e22008-11-25 02:35:14 +10307056 if (cpumask_test_cpu(task_cpu(p), new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007057 goto out;
7058
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307059 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007060 /* Need help from migration thread: drop lock and wait. */
Peter Zijlstra693525e2009-07-21 13:56:38 +02007061 struct task_struct *mt = rq->migration_thread;
7062
7063 get_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007064 task_rq_unlock(rq, &flags);
7065 wake_up_process(rq->migration_thread);
Peter Zijlstra693525e2009-07-21 13:56:38 +02007066 put_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007067 wait_for_completion(&req.done);
7068 tlb_migrate_finish(p->mm);
7069 return 0;
7070 }
7071out:
7072 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007073
Linus Torvalds1da177e2005-04-16 15:20:36 -07007074 return ret;
7075}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007076EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007077
7078/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007079 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07007080 * this because either it can't run here any more (set_cpus_allowed()
7081 * away from this CPU, or CPU going down), or because we're
7082 * attempting to rebalance this task on exec (sched_exec).
7083 *
7084 * So we race with normal scheduler movements, but that's OK, as long
7085 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07007086 *
7087 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007088 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07007089static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007090{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007091 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02007092 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007093
Max Krasnyanskye761b772008-07-15 04:43:49 -07007094 if (unlikely(!cpu_active(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07007095 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007096
7097 rq_src = cpu_rq(src_cpu);
7098 rq_dest = cpu_rq(dest_cpu);
7099
7100 double_rq_lock(rq_src, rq_dest);
7101 /* Already moved. */
7102 if (task_cpu(p) != src_cpu)
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007103 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007104 /* Affinity changed (again). */
Rusty Russell96f874e22008-11-25 02:35:14 +10307105 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007106 goto fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007107
Ingo Molnardd41f592007-07-09 18:51:59 +02007108 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007109 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007110 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007111
Linus Torvalds1da177e2005-04-16 15:20:36 -07007112 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007113 if (on_rq) {
7114 activate_task(rq_dest, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02007115 check_preempt_curr(rq_dest, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007116 }
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007117done:
Kirill Korotaevefc30812006-06-27 02:54:32 -07007118 ret = 1;
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007119fail:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007120 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07007121 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007122}
7123
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007124#define RCU_MIGRATION_IDLE 0
7125#define RCU_MIGRATION_NEED_QS 1
7126#define RCU_MIGRATION_GOT_QS 2
7127#define RCU_MIGRATION_MUST_SYNC 3
7128
Linus Torvalds1da177e2005-04-16 15:20:36 -07007129/*
7130 * migration_thread - this is a highprio system thread that performs
7131 * thread migration by bumping thread off CPU then 'pushing' onto
7132 * another runqueue.
7133 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07007134static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007135{
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007136 int badcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007137 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007138 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007139
7140 rq = cpu_rq(cpu);
7141 BUG_ON(rq->migration_thread != current);
7142
7143 set_current_state(TASK_INTERRUPTIBLE);
7144 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007145 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007146 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147
Linus Torvalds1da177e2005-04-16 15:20:36 -07007148 spin_lock_irq(&rq->lock);
7149
7150 if (cpu_is_offline(cpu)) {
7151 spin_unlock_irq(&rq->lock);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007152 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007153 }
7154
7155 if (rq->active_balance) {
7156 active_load_balance(rq, cpu);
7157 rq->active_balance = 0;
7158 }
7159
7160 head = &rq->migration_queue;
7161
7162 if (list_empty(head)) {
7163 spin_unlock_irq(&rq->lock);
7164 schedule();
7165 set_current_state(TASK_INTERRUPTIBLE);
7166 continue;
7167 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07007168 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007169 list_del_init(head->next);
7170
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007171 if (req->task != NULL) {
7172 spin_unlock(&rq->lock);
7173 __migrate_task(req->task, cpu, req->dest_cpu);
7174 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7175 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7176 spin_unlock(&rq->lock);
7177 } else {
7178 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7179 spin_unlock(&rq->lock);
7180 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7181 }
Nick Piggin674311d2005-06-25 14:57:27 -07007182 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007183
7184 complete(&req->done);
7185 }
7186 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007187
Linus Torvalds1da177e2005-04-16 15:20:36 -07007188 return 0;
7189}
7190
7191#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007192
7193static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7194{
7195 int ret;
7196
7197 local_irq_disable();
7198 ret = __migrate_task(p, src_cpu, dest_cpu);
7199 local_irq_enable();
7200 return ret;
7201}
7202
Kirill Korotaev054b9102006-12-10 02:20:11 -08007203/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007204 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08007205 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007206static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007207{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007208 int dest_cpu;
Mike Travis6ca09df2008-12-31 18:08:45 -08007209 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007210
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307211again:
7212 /* Look for allowed, online CPU in same node. */
7213 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7214 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7215 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007216
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307217 /* Any allowed, online CPU? */
7218 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7219 if (dest_cpu < nr_cpu_ids)
7220 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007221
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307222 /* No more Mr. Nice Guy. */
7223 if (dest_cpu >= nr_cpu_ids) {
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307224 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7225 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07007226
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307227 /*
7228 * Don't tell them about moving exiting tasks or
7229 * kernel threads (both mm NULL), since they never
7230 * leave kernel.
7231 */
7232 if (p->mm && printk_ratelimit()) {
7233 printk(KERN_INFO "process %d (%s) no "
7234 "longer affine to cpu%d\n",
7235 task_pid_nr(p), p->comm, dead_cpu);
Andi Kleen3a5c3592007-10-15 17:00:14 +02007236 }
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307237 }
7238
7239move:
7240 /* It can have affinity changed while we were choosing. */
7241 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7242 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007243}
7244
7245/*
7246 * While a dead CPU has no uninterruptible tasks queued at this point,
7247 * it might still have a nonzero ->nr_uninterruptible counter, because
7248 * for performance reasons the counter is not stricly tracking tasks to
7249 * their home CPUs. So we just add the counter to another CPU's counter,
7250 * to keep the global sum constant after CPU-down:
7251 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07007252static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007253{
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307254 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007255 unsigned long flags;
7256
7257 local_irq_save(flags);
7258 double_rq_lock(rq_src, rq_dest);
7259 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7260 rq_src->nr_uninterruptible = 0;
7261 double_rq_unlock(rq_src, rq_dest);
7262 local_irq_restore(flags);
7263}
7264
7265/* Run through task list and migrate tasks from the dead cpu. */
7266static void migrate_live_tasks(int src_cpu)
7267{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007268 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007269
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007270 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007271
Ingo Molnar48f24c42006-07-03 00:25:40 -07007272 do_each_thread(t, p) {
7273 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007274 continue;
7275
Ingo Molnar48f24c42006-07-03 00:25:40 -07007276 if (task_cpu(p) == src_cpu)
7277 move_task_off_dead_cpu(src_cpu, p);
7278 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007279
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007280 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007281}
7282
Ingo Molnardd41f592007-07-09 18:51:59 +02007283/*
7284 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007285 * It does so by boosting its priority to highest possible.
7286 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007287 */
7288void sched_idle_next(void)
7289{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007290 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07007291 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007292 struct task_struct *p = rq->idle;
7293 unsigned long flags;
7294
7295 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007296 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007297
Ingo Molnar48f24c42006-07-03 00:25:40 -07007298 /*
7299 * Strictly not necessary since rest of the CPUs are stopped by now
7300 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007301 */
7302 spin_lock_irqsave(&rq->lock, flags);
7303
Ingo Molnardd41f592007-07-09 18:51:59 +02007304 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007305
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007306 update_rq_clock(rq);
7307 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007308
7309 spin_unlock_irqrestore(&rq->lock, flags);
7310}
7311
Ingo Molnar48f24c42006-07-03 00:25:40 -07007312/*
7313 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07007314 * offline.
7315 */
7316void idle_task_exit(void)
7317{
7318 struct mm_struct *mm = current->active_mm;
7319
7320 BUG_ON(cpu_online(smp_processor_id()));
7321
7322 if (mm != &init_mm)
7323 switch_mm(mm, &init_mm, current);
7324 mmdrop(mm);
7325}
7326
Kirill Korotaev054b9102006-12-10 02:20:11 -08007327/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007328static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007329{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007330 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007331
7332 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07007333 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007334
7335 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07007336 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007337
Ingo Molnar48f24c42006-07-03 00:25:40 -07007338 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007339
7340 /*
7341 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007342 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07007343 * fine.
7344 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007345 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007346 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007347 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007348
Ingo Molnar48f24c42006-07-03 00:25:40 -07007349 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007350}
7351
7352/* release_task() removes task from tasklist, so we won't find dead tasks. */
7353static void migrate_dead_tasks(unsigned int dead_cpu)
7354{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007355 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007356 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007357
Ingo Molnardd41f592007-07-09 18:51:59 +02007358 for ( ; ; ) {
7359 if (!rq->nr_running)
7360 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02007361 update_rq_clock(rq);
Wang Chenb67802e2009-03-02 13:55:26 +08007362 next = pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02007363 if (!next)
7364 break;
Dmitry Adamushko79c53792008-06-29 00:16:56 +02007365 next->sched_class->put_prev_task(rq, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02007366 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02007367
Linus Torvalds1da177e2005-04-16 15:20:36 -07007368 }
7369}
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007370
7371/*
7372 * remove the tasks which were accounted by rq from calc_load_tasks.
7373 */
7374static void calc_global_load_remove(struct rq *rq)
7375{
7376 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
Thomas Gleixnera468d382009-07-17 14:15:46 +02007377 rq->calc_load_active = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007378}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007379#endif /* CONFIG_HOTPLUG_CPU */
7380
Nick Piggine692ab52007-07-26 13:40:43 +02007381#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7382
7383static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007384 {
7385 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007386 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007387 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01007388 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02007389};
7390
7391static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007392 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007393 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007394 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007395 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007396 .child = sd_ctl_dir,
7397 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01007398 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02007399};
7400
7401static struct ctl_table *sd_alloc_ctl_entry(int n)
7402{
7403 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02007404 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02007405
Nick Piggine692ab52007-07-26 13:40:43 +02007406 return entry;
7407}
7408
Milton Miller6382bc92007-10-15 17:00:19 +02007409static void sd_free_ctl_entry(struct ctl_table **tablep)
7410{
Milton Millercd7900762007-10-17 16:55:11 +02007411 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02007412
Milton Millercd7900762007-10-17 16:55:11 +02007413 /*
7414 * In the intermediate directories, both the child directory and
7415 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007416 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02007417 * static strings and all have proc handlers.
7418 */
7419 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02007420 if (entry->child)
7421 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02007422 if (entry->proc_handler == NULL)
7423 kfree(entry->procname);
7424 }
Milton Miller6382bc92007-10-15 17:00:19 +02007425
7426 kfree(*tablep);
7427 *tablep = NULL;
7428}
7429
Nick Piggine692ab52007-07-26 13:40:43 +02007430static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02007431set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02007432 const char *procname, void *data, int maxlen,
7433 mode_t mode, proc_handler *proc_handler)
7434{
Nick Piggine692ab52007-07-26 13:40:43 +02007435 entry->procname = procname;
7436 entry->data = data;
7437 entry->maxlen = maxlen;
7438 entry->mode = mode;
7439 entry->proc_handler = proc_handler;
7440}
7441
7442static struct ctl_table *
7443sd_alloc_ctl_domain_table(struct sched_domain *sd)
7444{
Ingo Molnara5d8c342008-10-09 11:35:51 +02007445 struct ctl_table *table = sd_alloc_ctl_entry(13);
Nick Piggine692ab52007-07-26 13:40:43 +02007446
Milton Millerad1cdc12007-10-15 17:00:19 +02007447 if (table == NULL)
7448 return NULL;
7449
Alexey Dobriyane0361852007-08-09 11:16:46 +02007450 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007451 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007452 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007453 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007454 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007455 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007456 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007457 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007458 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007459 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007460 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007461 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007462 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007463 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007464 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02007465 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007466 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02007467 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007468 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02007469 &sd->cache_nice_tries,
7470 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007471 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02007472 sizeof(int), 0644, proc_dointvec_minmax);
Ingo Molnara5d8c342008-10-09 11:35:51 +02007473 set_table_entry(&table[11], "name", sd->name,
7474 CORENAME_MAX_SIZE, 0444, proc_dostring);
7475 /* &table[12] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02007476
7477 return table;
7478}
7479
Ingo Molnar9a4e7152007-11-28 15:52:56 +01007480static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02007481{
7482 struct ctl_table *entry, *table;
7483 struct sched_domain *sd;
7484 int domain_num = 0, i;
7485 char buf[32];
7486
7487 for_each_domain(cpu, sd)
7488 domain_num++;
7489 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02007490 if (table == NULL)
7491 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02007492
7493 i = 0;
7494 for_each_domain(cpu, sd) {
7495 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007496 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007497 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007498 entry->child = sd_alloc_ctl_domain_table(sd);
7499 entry++;
7500 i++;
7501 }
7502 return table;
7503}
7504
7505static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02007506static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007507{
7508 int i, cpu_num = num_online_cpus();
7509 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7510 char buf[32];
7511
Milton Miller73785472007-10-24 18:23:48 +02007512 WARN_ON(sd_ctl_dir[0].child);
7513 sd_ctl_dir[0].child = entry;
7514
Milton Millerad1cdc12007-10-15 17:00:19 +02007515 if (entry == NULL)
7516 return;
7517
Milton Miller97b6ea72007-10-15 17:00:19 +02007518 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02007519 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007520 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007521 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007522 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02007523 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02007524 }
Milton Miller73785472007-10-24 18:23:48 +02007525
7526 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02007527 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7528}
Milton Miller6382bc92007-10-15 17:00:19 +02007529
Milton Miller73785472007-10-24 18:23:48 +02007530/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02007531static void unregister_sched_domain_sysctl(void)
7532{
Milton Miller73785472007-10-24 18:23:48 +02007533 if (sd_sysctl_header)
7534 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02007535 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007536 if (sd_ctl_dir[0].child)
7537 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02007538}
Nick Piggine692ab52007-07-26 13:40:43 +02007539#else
Milton Miller6382bc92007-10-15 17:00:19 +02007540static void register_sched_domain_sysctl(void)
7541{
7542}
7543static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007544{
7545}
7546#endif
7547
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007548static void set_rq_online(struct rq *rq)
7549{
7550 if (!rq->online) {
7551 const struct sched_class *class;
7552
Rusty Russellc6c49272008-11-25 02:35:05 +10307553 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007554 rq->online = 1;
7555
7556 for_each_class(class) {
7557 if (class->rq_online)
7558 class->rq_online(rq);
7559 }
7560 }
7561}
7562
7563static void set_rq_offline(struct rq *rq)
7564{
7565 if (rq->online) {
7566 const struct sched_class *class;
7567
7568 for_each_class(class) {
7569 if (class->rq_offline)
7570 class->rq_offline(rq);
7571 }
7572
Rusty Russellc6c49272008-11-25 02:35:05 +10307573 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007574 rq->online = 0;
7575 }
7576}
7577
Linus Torvalds1da177e2005-04-16 15:20:36 -07007578/*
7579 * migration_call - callback that gets triggered when a CPU is added.
7580 * Here we can start up the necessary migration thread for the new CPU.
7581 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007582static int __cpuinit
7583migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007584{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007585 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007586 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007587 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007588 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007589
7590 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07007591
Linus Torvalds1da177e2005-04-16 15:20:36 -07007592 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007593 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02007594 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007595 if (IS_ERR(p))
7596 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007597 kthread_bind(p, cpu);
7598 /* Must be high prio: stop_machine expects to yield to it. */
7599 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02007600 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007601 task_rq_unlock(rq, &flags);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007602 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007603 cpu_rq(cpu)->migration_thread = p;
Thomas Gleixnera468d382009-07-17 14:15:46 +02007604 rq->calc_load_update = calc_load_update;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007605 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007606
Linus Torvalds1da177e2005-04-16 15:20:36 -07007607 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007608 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007609 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007610 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007611
7612 /* Update our root-domain */
7613 rq = cpu_rq(cpu);
7614 spin_lock_irqsave(&rq->lock, flags);
7615 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307616 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007617
7618 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007619 }
7620 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007621 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007622
Linus Torvalds1da177e2005-04-16 15:20:36 -07007623#ifdef CONFIG_HOTPLUG_CPU
7624 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007625 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07007626 if (!cpu_rq(cpu)->migration_thread)
7627 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007628 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08007629 kthread_bind(cpu_rq(cpu)->migration_thread,
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307630 cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007631 kthread_stop(cpu_rq(cpu)->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007632 put_task_struct(cpu_rq(cpu)->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007633 cpu_rq(cpu)->migration_thread = NULL;
7634 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007635
Linus Torvalds1da177e2005-04-16 15:20:36 -07007636 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007637 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07007638 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007639 migrate_live_tasks(cpu);
7640 rq = cpu_rq(cpu);
7641 kthread_stop(rq->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007642 put_task_struct(rq->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007643 rq->migration_thread = NULL;
7644 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007645 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02007646 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007647 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007648 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02007649 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7650 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007651 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007652 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07007653 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007654 migrate_nr_uninterruptible(rq);
7655 BUG_ON(rq->nr_running != 0);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007656 calc_global_load_remove(rq);
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007657 /*
7658 * No need to migrate the tasks: it was best-effort if
7659 * they didn't take sched_hotcpu_mutex. Just wake up
7660 * the requestors.
7661 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007662 spin_lock_irq(&rq->lock);
7663 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007664 struct migration_req *req;
7665
Linus Torvalds1da177e2005-04-16 15:20:36 -07007666 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07007667 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007668 list_del_init(&req->list);
Brian King9a2bd242008-12-09 08:47:00 -06007669 spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007670 complete(&req->done);
Brian King9a2bd242008-12-09 08:47:00 -06007671 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007672 }
7673 spin_unlock_irq(&rq->lock);
7674 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007675
Gregory Haskins08f503b2008-03-10 17:59:11 -04007676 case CPU_DYING:
7677 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01007678 /* Update our root-domain */
7679 rq = cpu_rq(cpu);
7680 spin_lock_irqsave(&rq->lock, flags);
7681 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 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007684 }
7685 spin_unlock_irqrestore(&rq->lock, flags);
7686 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007687#endif
7688 }
7689 return NOTIFY_OK;
7690}
7691
Paul Mackerrasf38b0822009-06-02 21:05:16 +10007692/*
7693 * Register at high priority so that task migration (migrate_all_tasks)
7694 * happens before everything else. This has to be lower priority than
7695 * the notifier in the perf_counter subsystem, though.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007696 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07007697static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007698 .notifier_call = migration_call,
7699 .priority = 10
7700};
7701
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007702static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007703{
7704 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07007705 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007706
7707 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07007708 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7709 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007710 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7711 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007712
Thomas Gleixnera004cd42009-07-21 09:54:05 +02007713 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007714}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007715early_initcall(migration_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007716#endif
7717
7718#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07007719
Ingo Molnar3e9830d2007-10-15 17:00:13 +02007720#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007721
Mike Travis7c16ec52008-04-04 18:11:11 -07007722static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
Rusty Russell96f874e22008-11-25 02:35:14 +10307723 struct cpumask *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007724{
7725 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07007726 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007727
Rusty Russell968ea6d2008-12-13 21:55:51 +10307728 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
Rusty Russell96f874e22008-11-25 02:35:14 +10307729 cpumask_clear(groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007730
7731 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7732
7733 if (!(sd->flags & SD_LOAD_BALANCE)) {
7734 printk("does not load-balance\n");
7735 if (sd->parent)
7736 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7737 " has parent");
7738 return -1;
7739 }
7740
Li Zefaneefd7962008-11-04 16:15:37 +08007741 printk(KERN_CONT "span %s level %s\n", str, sd->name);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007742
Rusty Russell758b2cd2008-11-25 02:35:04 +10307743 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007744 printk(KERN_ERR "ERROR: domain->span does not contain "
7745 "CPU%d\n", cpu);
7746 }
Rusty Russell758b2cd2008-11-25 02:35:04 +10307747 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007748 printk(KERN_ERR "ERROR: domain->groups does not contain"
7749 " CPU%d\n", cpu);
7750 }
7751
7752 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7753 do {
7754 if (!group) {
7755 printk("\n");
7756 printk(KERN_ERR "ERROR: group is NULL\n");
7757 break;
7758 }
7759
Peter Zijlstra18a38852009-09-01 10:34:39 +02007760 if (!group->cpu_power) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007761 printk(KERN_CONT "\n");
7762 printk(KERN_ERR "ERROR: domain->cpu_power not "
7763 "set\n");
7764 break;
7765 }
7766
Rusty Russell758b2cd2008-11-25 02:35:04 +10307767 if (!cpumask_weight(sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007768 printk(KERN_CONT "\n");
7769 printk(KERN_ERR "ERROR: empty group\n");
7770 break;
7771 }
7772
Rusty Russell758b2cd2008-11-25 02:35:04 +10307773 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007774 printk(KERN_CONT "\n");
7775 printk(KERN_ERR "ERROR: repeated CPUs\n");
7776 break;
7777 }
7778
Rusty Russell758b2cd2008-11-25 02:35:04 +10307779 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007780
Rusty Russell968ea6d2008-12-13 21:55:51 +10307781 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307782
7783 printk(KERN_CONT " %s", str);
Peter Zijlstra18a38852009-09-01 10:34:39 +02007784 if (group->cpu_power != SCHED_LOAD_SCALE) {
7785 printk(KERN_CONT " (cpu_power = %d)",
7786 group->cpu_power);
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307787 }
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007788
7789 group = group->next;
7790 } while (group != sd->groups);
7791 printk(KERN_CONT "\n");
7792
Rusty Russell758b2cd2008-11-25 02:35:04 +10307793 if (!cpumask_equal(sched_domain_span(sd), groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007794 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7795
Rusty Russell758b2cd2008-11-25 02:35:04 +10307796 if (sd->parent &&
7797 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007798 printk(KERN_ERR "ERROR: parent span is not a superset "
7799 "of domain->span\n");
7800 return 0;
7801}
7802
Linus Torvalds1da177e2005-04-16 15:20:36 -07007803static void sched_domain_debug(struct sched_domain *sd, int cpu)
7804{
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307805 cpumask_var_t groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007806 int level = 0;
7807
Nick Piggin41c7ce92005-06-25 14:57:24 -07007808 if (!sd) {
7809 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7810 return;
7811 }
7812
Linus Torvalds1da177e2005-04-16 15:20:36 -07007813 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7814
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307815 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007816 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7817 return;
7818 }
7819
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007820 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007821 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007822 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007823 level++;
7824 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08007825 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007826 break;
7827 }
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307828 free_cpumask_var(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007829}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007830#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007831# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007832#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007833
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007834static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007835{
Rusty Russell758b2cd2008-11-25 02:35:04 +10307836 if (cpumask_weight(sched_domain_span(sd)) == 1)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007837 return 1;
7838
7839 /* Following flags need at least 2 groups */
7840 if (sd->flags & (SD_LOAD_BALANCE |
7841 SD_BALANCE_NEWIDLE |
7842 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007843 SD_BALANCE_EXEC |
7844 SD_SHARE_CPUPOWER |
7845 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007846 if (sd->groups != sd->groups->next)
7847 return 0;
7848 }
7849
7850 /* Following flags don't use groups */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02007851 if (sd->flags & (SD_WAKE_AFFINE))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007852 return 0;
7853
7854 return 1;
7855}
7856
Ingo Molnar48f24c42006-07-03 00:25:40 -07007857static int
7858sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007859{
7860 unsigned long cflags = sd->flags, pflags = parent->flags;
7861
7862 if (sd_degenerate(parent))
7863 return 1;
7864
Rusty Russell758b2cd2008-11-25 02:35:04 +10307865 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007866 return 0;
7867
Suresh Siddha245af2c2005-06-25 14:57:25 -07007868 /* Flags needing groups don't count if only 1 group in parent */
7869 if (parent->groups == parent->groups->next) {
7870 pflags &= ~(SD_LOAD_BALANCE |
7871 SD_BALANCE_NEWIDLE |
7872 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007873 SD_BALANCE_EXEC |
7874 SD_SHARE_CPUPOWER |
7875 SD_SHARE_PKG_RESOURCES);
Ken Chen54364992008-12-07 18:47:37 -08007876 if (nr_node_ids == 1)
7877 pflags &= ~SD_SERIALIZE;
Suresh Siddha245af2c2005-06-25 14:57:25 -07007878 }
7879 if (~cflags & pflags)
7880 return 0;
7881
7882 return 1;
7883}
7884
Rusty Russellc6c49272008-11-25 02:35:05 +10307885static void free_rootdomain(struct root_domain *rd)
7886{
Rusty Russell68e74562008-11-25 02:35:13 +10307887 cpupri_cleanup(&rd->cpupri);
7888
Rusty Russellc6c49272008-11-25 02:35:05 +10307889 free_cpumask_var(rd->rto_mask);
7890 free_cpumask_var(rd->online);
7891 free_cpumask_var(rd->span);
7892 kfree(rd);
7893}
7894
Gregory Haskins57d885f2008-01-25 21:08:18 +01007895static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7896{
Ingo Molnara0490fa2009-02-12 11:35:40 +01007897 struct root_domain *old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007898 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007899
7900 spin_lock_irqsave(&rq->lock, flags);
7901
7902 if (rq->rd) {
Ingo Molnara0490fa2009-02-12 11:35:40 +01007903 old_rd = rq->rd;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007904
Rusty Russellc6c49272008-11-25 02:35:05 +10307905 if (cpumask_test_cpu(rq->cpu, old_rd->online))
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007906 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007907
Rusty Russellc6c49272008-11-25 02:35:05 +10307908 cpumask_clear_cpu(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007909
Ingo Molnara0490fa2009-02-12 11:35:40 +01007910 /*
7911 * If we dont want to free the old_rt yet then
7912 * set old_rd to NULL to skip the freeing later
7913 * in this function:
7914 */
7915 if (!atomic_dec_and_test(&old_rd->refcount))
7916 old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007917 }
7918
7919 atomic_inc(&rd->refcount);
7920 rq->rd = rd;
7921
Rusty Russellc6c49272008-11-25 02:35:05 +10307922 cpumask_set_cpu(rq->cpu, rd->span);
Gregory Haskins00aec932009-07-30 10:57:23 -04007923 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04007924 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007925
7926 spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnara0490fa2009-02-12 11:35:40 +01007927
7928 if (old_rd)
7929 free_rootdomain(old_rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007930}
7931
Li Zefanfd5e1b52009-06-15 13:34:19 +08007932static int init_rootdomain(struct root_domain *rd, bool bootmem)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007933{
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007934 gfp_t gfp = GFP_KERNEL;
7935
Gregory Haskins57d885f2008-01-25 21:08:18 +01007936 memset(rd, 0, sizeof(*rd));
7937
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007938 if (bootmem)
7939 gfp = GFP_NOWAIT;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02007940
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007941 if (!alloc_cpumask_var(&rd->span, gfp))
Li Zefan0c910d22009-01-06 17:39:06 +08007942 goto out;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007943 if (!alloc_cpumask_var(&rd->online, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10307944 goto free_span;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007945 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10307946 goto free_online;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02007947
Pekka Enberg0fb53022009-06-11 08:41:22 +03007948 if (cpupri_init(&rd->cpupri, bootmem) != 0)
Rusty Russell68e74562008-11-25 02:35:13 +10307949 goto free_rto_mask;
Rusty Russellc6c49272008-11-25 02:35:05 +10307950 return 0;
7951
Rusty Russell68e74562008-11-25 02:35:13 +10307952free_rto_mask:
7953 free_cpumask_var(rd->rto_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10307954free_online:
7955 free_cpumask_var(rd->online);
7956free_span:
7957 free_cpumask_var(rd->span);
Li Zefan0c910d22009-01-06 17:39:06 +08007958out:
Rusty Russellc6c49272008-11-25 02:35:05 +10307959 return -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007960}
7961
7962static void init_defrootdomain(void)
7963{
Rusty Russellc6c49272008-11-25 02:35:05 +10307964 init_rootdomain(&def_root_domain, true);
7965
Gregory Haskins57d885f2008-01-25 21:08:18 +01007966 atomic_set(&def_root_domain.refcount, 1);
7967}
7968
Gregory Haskinsdc938522008-01-25 21:08:26 +01007969static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007970{
7971 struct root_domain *rd;
7972
7973 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7974 if (!rd)
7975 return NULL;
7976
Rusty Russellc6c49272008-11-25 02:35:05 +10307977 if (init_rootdomain(rd, false) != 0) {
7978 kfree(rd);
7979 return NULL;
7980 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01007981
7982 return rd;
7983}
7984
Linus Torvalds1da177e2005-04-16 15:20:36 -07007985/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01007986 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07007987 * hold the hotplug lock.
7988 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01007989static void
7990cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007991{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007992 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07007993 struct sched_domain *tmp;
7994
7995 /* Remove the sched domains which do not contribute to scheduling. */
Li Zefanf29c9b12008-11-06 09:45:16 +08007996 for (tmp = sd; tmp; ) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007997 struct sched_domain *parent = tmp->parent;
7998 if (!parent)
7999 break;
Li Zefanf29c9b12008-11-06 09:45:16 +08008000
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008001 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008002 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008003 if (parent->parent)
8004 parent->parent->child = tmp;
Li Zefanf29c9b12008-11-06 09:45:16 +08008005 } else
8006 tmp = tmp->parent;
Suresh Siddha245af2c2005-06-25 14:57:25 -07008007 }
8008
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008009 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008010 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008011 if (sd)
8012 sd->child = NULL;
8013 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008014
8015 sched_domain_debug(sd, cpu);
8016
Gregory Haskins57d885f2008-01-25 21:08:18 +01008017 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07008018 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008019}
8020
8021/* cpus with isolated domains */
Rusty Russelldcc30a32008-11-25 02:35:12 +10308022static cpumask_var_t cpu_isolated_map;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008023
8024/* Setup the mask of cpus configured for isolated domains */
8025static int __init isolated_cpu_setup(char *str)
8026{
Rusty Russell968ea6d2008-12-13 21:55:51 +10308027 cpulist_parse(str, cpu_isolated_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008028 return 1;
8029}
8030
Ingo Molnar8927f492007-10-15 17:00:13 +02008031__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008032
8033/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008034 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8035 * to a function which identifies what group(along with sched group) a CPU
Rusty Russell96f874e22008-11-25 02:35:14 +10308036 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8037 * (due to the fact that we keep track of groups covered with a struct cpumask).
Linus Torvalds1da177e2005-04-16 15:20:36 -07008038 *
8039 * init_sched_build_groups will build a circular linked list of the groups
8040 * covered by the given span, and will set each group's ->cpumask correctly,
8041 * and ->cpu_power to 0.
8042 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008043static void
Rusty Russell96f874e22008-11-25 02:35:14 +10308044init_sched_build_groups(const struct cpumask *span,
8045 const struct cpumask *cpu_map,
8046 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07008047 struct sched_group **sg,
Rusty Russell96f874e22008-11-25 02:35:14 +10308048 struct cpumask *tmpmask),
8049 struct cpumask *covered, struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008050{
8051 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008052 int i;
8053
Rusty Russell96f874e22008-11-25 02:35:14 +10308054 cpumask_clear(covered);
Mike Travis7c16ec52008-04-04 18:11:11 -07008055
Rusty Russellabcd0832008-11-25 02:35:02 +10308056 for_each_cpu(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008057 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07008058 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008059 int j;
8060
Rusty Russell758b2cd2008-11-25 02:35:04 +10308061 if (cpumask_test_cpu(i, covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008062 continue;
8063
Rusty Russell758b2cd2008-11-25 02:35:04 +10308064 cpumask_clear(sched_group_cpus(sg));
Peter Zijlstra18a38852009-09-01 10:34:39 +02008065 sg->cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008066
Rusty Russellabcd0832008-11-25 02:35:02 +10308067 for_each_cpu(j, span) {
Mike Travis7c16ec52008-04-04 18:11:11 -07008068 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008069 continue;
8070
Rusty Russell96f874e22008-11-25 02:35:14 +10308071 cpumask_set_cpu(j, covered);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308072 cpumask_set_cpu(j, sched_group_cpus(sg));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008073 }
8074 if (!first)
8075 first = sg;
8076 if (last)
8077 last->next = sg;
8078 last = sg;
8079 }
8080 last->next = first;
8081}
8082
John Hawkes9c1cfda2005-09-06 15:18:14 -07008083#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07008084
John Hawkes9c1cfda2005-09-06 15:18:14 -07008085#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08008086
John Hawkes9c1cfda2005-09-06 15:18:14 -07008087/**
8088 * find_next_best_node - find the next node to include in a sched_domain
8089 * @node: node whose sched_domain we're building
8090 * @used_nodes: nodes already in the sched_domain
8091 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008092 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07008093 * finds the closest node not already in the @used_nodes map.
8094 *
8095 * Should use nodemask_t.
8096 */
Mike Travisc5f59f02008-04-04 18:11:10 -07008097static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008098{
8099 int i, n, val, min_val, best_node = 0;
8100
8101 min_val = INT_MAX;
8102
Mike Travis076ac2a2008-05-12 21:21:12 +02008103 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008104 /* Start at @node */
Mike Travis076ac2a2008-05-12 21:21:12 +02008105 n = (node + i) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008106
8107 if (!nr_cpus_node(n))
8108 continue;
8109
8110 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07008111 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008112 continue;
8113
8114 /* Simple min distance search */
8115 val = node_distance(node, n);
8116
8117 if (val < min_val) {
8118 min_val = val;
8119 best_node = n;
8120 }
8121 }
8122
Mike Travisc5f59f02008-04-04 18:11:10 -07008123 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008124 return best_node;
8125}
8126
8127/**
8128 * sched_domain_node_span - get a cpumask for a node's sched_domain
8129 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07008130 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07008131 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008132 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07008133 * should be one that prevents unnecessary balancing, but also spreads tasks
8134 * out optimally.
8135 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308136static void sched_domain_node_span(int node, struct cpumask *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008137{
Mike Travisc5f59f02008-04-04 18:11:10 -07008138 nodemask_t used_nodes;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008139 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008140
Mike Travis6ca09df2008-12-31 18:08:45 -08008141 cpumask_clear(span);
Mike Travisc5f59f02008-04-04 18:11:10 -07008142 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008143
Mike Travis6ca09df2008-12-31 18:08:45 -08008144 cpumask_or(span, span, cpumask_of_node(node));
Mike Travisc5f59f02008-04-04 18:11:10 -07008145 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008146
8147 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07008148 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008149
Mike Travis6ca09df2008-12-31 18:08:45 -08008150 cpumask_or(span, span, cpumask_of_node(next_node));
John Hawkes9c1cfda2005-09-06 15:18:14 -07008151 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008152}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008153#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07008154
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008155int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008156
John Hawkes9c1cfda2005-09-06 15:18:14 -07008157/*
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308158 * The cpus mask in sched_group and sched_domain hangs off the end.
Ingo Molnar4200efd2009-05-19 09:22:19 +02008159 *
8160 * ( See the the comments in include/linux/sched.h:struct sched_group
8161 * and struct sched_domain. )
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308162 */
8163struct static_sched_group {
8164 struct sched_group sg;
8165 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8166};
8167
8168struct static_sched_domain {
8169 struct sched_domain sd;
8170 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8171};
8172
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008173struct s_data {
8174#ifdef CONFIG_NUMA
8175 int sd_allnodes;
8176 cpumask_var_t domainspan;
8177 cpumask_var_t covered;
8178 cpumask_var_t notcovered;
8179#endif
8180 cpumask_var_t nodemask;
8181 cpumask_var_t this_sibling_map;
8182 cpumask_var_t this_core_map;
8183 cpumask_var_t send_covered;
8184 cpumask_var_t tmpmask;
8185 struct sched_group **sched_group_nodes;
8186 struct root_domain *rd;
8187};
8188
Andreas Herrmann2109b992009-08-18 12:53:00 +02008189enum s_alloc {
8190 sa_sched_groups = 0,
8191 sa_rootdomain,
8192 sa_tmpmask,
8193 sa_send_covered,
8194 sa_this_core_map,
8195 sa_this_sibling_map,
8196 sa_nodemask,
8197 sa_sched_group_nodes,
8198#ifdef CONFIG_NUMA
8199 sa_notcovered,
8200 sa_covered,
8201 sa_domainspan,
8202#endif
8203 sa_none,
8204};
8205
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308206/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07008207 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07008208 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008209#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308210static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8211static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008212
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008213static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308214cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8215 struct sched_group **sg, struct cpumask *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008216{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008217 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308218 *sg = &per_cpu(sched_group_cpus, cpu).sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008219 return cpu;
8220}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008221#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008222
Ingo Molnar48f24c42006-07-03 00:25:40 -07008223/*
8224 * multi-core sched-domains:
8225 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008226#ifdef CONFIG_SCHED_MC
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308227static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8228static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008229#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008230
8231#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008232static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308233cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8234 struct sched_group **sg, struct cpumask *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008235{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008236 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07008237
Rusty Russellc69fc562009-03-13 14:49:46 +10308238 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308239 group = cpumask_first(mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008240 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308241 *sg = &per_cpu(sched_group_core, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008242 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008243}
8244#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008245static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308246cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8247 struct sched_group **sg, struct cpumask *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008248{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008249 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308250 *sg = &per_cpu(sched_group_core, cpu).sg;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008251 return cpu;
8252}
8253#endif
8254
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308255static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8256static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008257
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008258static int
Rusty Russell96f874e22008-11-25 02:35:14 +10308259cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8260 struct sched_group **sg, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008261{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008262 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008263#ifdef CONFIG_SCHED_MC
Mike Travis6ca09df2008-12-31 18:08:45 -08008264 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308265 group = cpumask_first(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008266#elif defined(CONFIG_SCHED_SMT)
Rusty Russellc69fc562009-03-13 14:49:46 +10308267 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308268 group = cpumask_first(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008269#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008270 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008271#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008272 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308273 *sg = &per_cpu(sched_group_phys, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008274 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008275}
8276
8277#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07008278/*
8279 * The init_sched_build_groups can't handle what we want to do with node
8280 * groups, so roll our own. Now each node has its own list of groups which
8281 * gets dynamically allocated.
8282 */
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008283static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07008284static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008285
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008286static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308287static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008288
Rusty Russell96f874e22008-11-25 02:35:14 +10308289static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8290 struct sched_group **sg,
8291 struct cpumask *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008292{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008293 int group;
8294
Mike Travis6ca09df2008-12-31 18:08:45 -08008295 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308296 group = cpumask_first(nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008297
8298 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308299 *sg = &per_cpu(sched_group_allnodes, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008300 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008301}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008302
Siddha, Suresh B08069032006-03-27 01:15:23 -08008303static void init_numa_sched_groups_power(struct sched_group *group_head)
8304{
8305 struct sched_group *sg = group_head;
8306 int j;
8307
8308 if (!sg)
8309 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008310 do {
Rusty Russell758b2cd2008-11-25 02:35:04 +10308311 for_each_cpu(j, sched_group_cpus(sg)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008312 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008313
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308314 sd = &per_cpu(phys_domains, j).sd;
Miao Xie13318a72009-04-15 09:59:10 +08008315 if (j != group_first_cpu(sd->groups)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008316 /*
8317 * Only add "power" once for each
8318 * physical package.
8319 */
8320 continue;
8321 }
8322
Peter Zijlstra18a38852009-09-01 10:34:39 +02008323 sg->cpu_power += sd->groups->cpu_power;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008324 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02008325 sg = sg->next;
8326 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08008327}
Andreas Herrmann0601a882009-08-18 13:01:11 +02008328
8329static int build_numa_sched_groups(struct s_data *d,
8330 const struct cpumask *cpu_map, int num)
8331{
8332 struct sched_domain *sd;
8333 struct sched_group *sg, *prev;
8334 int n, j;
8335
8336 cpumask_clear(d->covered);
8337 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8338 if (cpumask_empty(d->nodemask)) {
8339 d->sched_group_nodes[num] = NULL;
8340 goto out;
8341 }
8342
8343 sched_domain_node_span(num, d->domainspan);
8344 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8345
8346 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8347 GFP_KERNEL, num);
8348 if (!sg) {
8349 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8350 num);
8351 return -ENOMEM;
8352 }
8353 d->sched_group_nodes[num] = sg;
8354
8355 for_each_cpu(j, d->nodemask) {
8356 sd = &per_cpu(node_domains, j).sd;
8357 sd->groups = sg;
8358 }
8359
Peter Zijlstra18a38852009-09-01 10:34:39 +02008360 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008361 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8362 sg->next = sg;
8363 cpumask_or(d->covered, d->covered, d->nodemask);
8364
8365 prev = sg;
8366 for (j = 0; j < nr_node_ids; j++) {
8367 n = (num + j) % nr_node_ids;
8368 cpumask_complement(d->notcovered, d->covered);
8369 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8370 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8371 if (cpumask_empty(d->tmpmask))
8372 break;
8373 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8374 if (cpumask_empty(d->tmpmask))
8375 continue;
8376 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8377 GFP_KERNEL, num);
8378 if (!sg) {
8379 printk(KERN_WARNING
8380 "Can not alloc domain group for node %d\n", j);
8381 return -ENOMEM;
8382 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008383 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008384 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8385 sg->next = prev->next;
8386 cpumask_or(d->covered, d->covered, d->tmpmask);
8387 prev->next = sg;
8388 prev = sg;
8389 }
8390out:
8391 return 0;
8392}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008393#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008394
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008395#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008396/* Free memory allocated for various sched_group structures */
Rusty Russell96f874e22008-11-25 02:35:14 +10308397static void free_sched_groups(const struct cpumask *cpu_map,
8398 struct cpumask *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008399{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008400 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008401
Rusty Russellabcd0832008-11-25 02:35:02 +10308402 for_each_cpu(cpu, cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008403 struct sched_group **sched_group_nodes
8404 = sched_group_nodes_bycpu[cpu];
8405
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008406 if (!sched_group_nodes)
8407 continue;
8408
Mike Travis076ac2a2008-05-12 21:21:12 +02008409 for (i = 0; i < nr_node_ids; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008410 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8411
Mike Travis6ca09df2008-12-31 18:08:45 -08008412 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e22008-11-25 02:35:14 +10308413 if (cpumask_empty(nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008414 continue;
8415
8416 if (sg == NULL)
8417 continue;
8418 sg = sg->next;
8419next_sg:
8420 oldsg = sg;
8421 sg = sg->next;
8422 kfree(oldsg);
8423 if (oldsg != sched_group_nodes[i])
8424 goto next_sg;
8425 }
8426 kfree(sched_group_nodes);
8427 sched_group_nodes_bycpu[cpu] = NULL;
8428 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008429}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008430#else /* !CONFIG_NUMA */
Rusty Russell96f874e22008-11-25 02:35:14 +10308431static void free_sched_groups(const struct cpumask *cpu_map,
8432 struct cpumask *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008433{
8434}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008435#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008436
Linus Torvalds1da177e2005-04-16 15:20:36 -07008437/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008438 * Initialize sched groups cpu_power.
8439 *
8440 * cpu_power indicates the capacity of sched group, which is used while
8441 * distributing the load between different sched groups in a sched domain.
8442 * Typically cpu_power for all the groups in a sched domain will be same unless
8443 * there are asymmetries in the topology. If there are asymmetries, group
8444 * having more cpu_power will pickup more load compared to the group having
8445 * less cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008446 */
8447static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8448{
8449 struct sched_domain *child;
8450 struct sched_group *group;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008451 long power;
8452 int weight;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008453
8454 WARN_ON(!sd || !sd->groups);
8455
Miao Xie13318a72009-04-15 09:59:10 +08008456 if (cpu != group_first_cpu(sd->groups))
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008457 return;
8458
8459 child = sd->child;
8460
Peter Zijlstra18a38852009-09-01 10:34:39 +02008461 sd->groups->cpu_power = 0;
Eric Dumazet5517d862007-05-08 00:32:57 -07008462
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008463 if (!child) {
8464 power = SCHED_LOAD_SCALE;
8465 weight = cpumask_weight(sched_domain_span(sd));
8466 /*
8467 * SMT siblings share the power of a single core.
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008468 * Usually multiple threads get a better yield out of
8469 * that one core than a single thread would have,
8470 * reflect that in sd->smt_gain.
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008471 */
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008472 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8473 power *= sd->smt_gain;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008474 power /= weight;
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008475 power >>= SCHED_LOAD_SHIFT;
8476 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008477 sd->groups->cpu_power += power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008478 return;
8479 }
8480
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008481 /*
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008482 * Add cpu_power of each child group to this groups cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008483 */
8484 group = child->groups;
8485 do {
Peter Zijlstra18a38852009-09-01 10:34:39 +02008486 sd->groups->cpu_power += group->cpu_power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008487 group = group->next;
8488 } while (group != child->groups);
8489}
8490
8491/*
Mike Travis7c16ec52008-04-04 18:11:11 -07008492 * Initializers for schedule domains
8493 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8494 */
8495
Ingo Molnara5d8c342008-10-09 11:35:51 +02008496#ifdef CONFIG_SCHED_DEBUG
8497# define SD_INIT_NAME(sd, type) sd->name = #type
8498#else
8499# define SD_INIT_NAME(sd, type) do { } while (0)
8500#endif
8501
Mike Travis7c16ec52008-04-04 18:11:11 -07008502#define SD_INIT(sd, type) sd_init_##type(sd)
Ingo Molnara5d8c342008-10-09 11:35:51 +02008503
Mike Travis7c16ec52008-04-04 18:11:11 -07008504#define SD_INIT_FUNC(type) \
8505static noinline void sd_init_##type(struct sched_domain *sd) \
8506{ \
8507 memset(sd, 0, sizeof(*sd)); \
8508 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008509 sd->level = SD_LV_##type; \
Ingo Molnara5d8c342008-10-09 11:35:51 +02008510 SD_INIT_NAME(sd, type); \
Mike Travis7c16ec52008-04-04 18:11:11 -07008511}
8512
8513SD_INIT_FUNC(CPU)
8514#ifdef CONFIG_NUMA
8515 SD_INIT_FUNC(ALLNODES)
8516 SD_INIT_FUNC(NODE)
8517#endif
8518#ifdef CONFIG_SCHED_SMT
8519 SD_INIT_FUNC(SIBLING)
8520#endif
8521#ifdef CONFIG_SCHED_MC
8522 SD_INIT_FUNC(MC)
8523#endif
8524
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008525static int default_relax_domain_level = -1;
8526
8527static int __init setup_relax_domain_level(char *str)
8528{
Li Zefan30e0e172008-05-13 10:27:17 +08008529 unsigned long val;
8530
8531 val = simple_strtoul(str, NULL, 0);
8532 if (val < SD_LV_MAX)
8533 default_relax_domain_level = val;
8534
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008535 return 1;
8536}
8537__setup("relax_domain_level=", setup_relax_domain_level);
8538
8539static void set_domain_attribute(struct sched_domain *sd,
8540 struct sched_domain_attr *attr)
8541{
8542 int request;
8543
8544 if (!attr || attr->relax_domain_level < 0) {
8545 if (default_relax_domain_level < 0)
8546 return;
8547 else
8548 request = default_relax_domain_level;
8549 } else
8550 request = attr->relax_domain_level;
8551 if (request < sd->level) {
8552 /* turn off idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008553 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008554 } else {
8555 /* turn on idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008556 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008557 }
8558}
8559
Andreas Herrmann2109b992009-08-18 12:53:00 +02008560static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8561 const struct cpumask *cpu_map)
8562{
8563 switch (what) {
8564 case sa_sched_groups:
8565 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8566 d->sched_group_nodes = NULL;
8567 case sa_rootdomain:
8568 free_rootdomain(d->rd); /* fall through */
8569 case sa_tmpmask:
8570 free_cpumask_var(d->tmpmask); /* fall through */
8571 case sa_send_covered:
8572 free_cpumask_var(d->send_covered); /* fall through */
8573 case sa_this_core_map:
8574 free_cpumask_var(d->this_core_map); /* fall through */
8575 case sa_this_sibling_map:
8576 free_cpumask_var(d->this_sibling_map); /* fall through */
8577 case sa_nodemask:
8578 free_cpumask_var(d->nodemask); /* fall through */
8579 case sa_sched_group_nodes:
8580#ifdef CONFIG_NUMA
8581 kfree(d->sched_group_nodes); /* fall through */
8582 case sa_notcovered:
8583 free_cpumask_var(d->notcovered); /* fall through */
8584 case sa_covered:
8585 free_cpumask_var(d->covered); /* fall through */
8586 case sa_domainspan:
8587 free_cpumask_var(d->domainspan); /* fall through */
8588#endif
8589 case sa_none:
8590 break;
8591 }
8592}
8593
8594static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8595 const struct cpumask *cpu_map)
8596{
8597#ifdef CONFIG_NUMA
8598 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8599 return sa_none;
8600 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8601 return sa_domainspan;
8602 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8603 return sa_covered;
8604 /* Allocate the per-node list of sched groups */
8605 d->sched_group_nodes = kcalloc(nr_node_ids,
8606 sizeof(struct sched_group *), GFP_KERNEL);
8607 if (!d->sched_group_nodes) {
8608 printk(KERN_WARNING "Can not alloc sched group node list\n");
8609 return sa_notcovered;
8610 }
8611 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8612#endif
8613 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8614 return sa_sched_group_nodes;
8615 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8616 return sa_nodemask;
8617 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8618 return sa_this_sibling_map;
8619 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8620 return sa_this_core_map;
8621 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8622 return sa_send_covered;
8623 d->rd = alloc_rootdomain();
8624 if (!d->rd) {
8625 printk(KERN_WARNING "Cannot alloc root domain\n");
8626 return sa_tmpmask;
8627 }
8628 return sa_rootdomain;
8629}
8630
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008631static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8632 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8633{
8634 struct sched_domain *sd = NULL;
8635#ifdef CONFIG_NUMA
8636 struct sched_domain *parent;
8637
8638 d->sd_allnodes = 0;
8639 if (cpumask_weight(cpu_map) >
8640 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8641 sd = &per_cpu(allnodes_domains, i).sd;
8642 SD_INIT(sd, ALLNODES);
8643 set_domain_attribute(sd, attr);
8644 cpumask_copy(sched_domain_span(sd), cpu_map);
8645 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8646 d->sd_allnodes = 1;
8647 }
8648 parent = sd;
8649
8650 sd = &per_cpu(node_domains, i).sd;
8651 SD_INIT(sd, NODE);
8652 set_domain_attribute(sd, attr);
8653 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8654 sd->parent = parent;
8655 if (parent)
8656 parent->child = sd;
8657 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8658#endif
8659 return sd;
8660}
8661
Andreas Herrmann87cce662009-08-18 12:54:55 +02008662static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8663 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8664 struct sched_domain *parent, int i)
8665{
8666 struct sched_domain *sd;
8667 sd = &per_cpu(phys_domains, i).sd;
8668 SD_INIT(sd, CPU);
8669 set_domain_attribute(sd, attr);
8670 cpumask_copy(sched_domain_span(sd), d->nodemask);
8671 sd->parent = parent;
8672 if (parent)
8673 parent->child = sd;
8674 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8675 return sd;
8676}
8677
Andreas Herrmann410c4082009-08-18 12:56:14 +02008678static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8679 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8680 struct sched_domain *parent, int i)
8681{
8682 struct sched_domain *sd = parent;
8683#ifdef CONFIG_SCHED_MC
8684 sd = &per_cpu(core_domains, i).sd;
8685 SD_INIT(sd, MC);
8686 set_domain_attribute(sd, attr);
8687 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8688 sd->parent = parent;
8689 parent->child = sd;
8690 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8691#endif
8692 return sd;
8693}
8694
Andreas Herrmannd8173532009-08-18 12:57:03 +02008695static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8696 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8697 struct sched_domain *parent, int i)
8698{
8699 struct sched_domain *sd = parent;
8700#ifdef CONFIG_SCHED_SMT
8701 sd = &per_cpu(cpu_domains, i).sd;
8702 SD_INIT(sd, SIBLING);
8703 set_domain_attribute(sd, attr);
8704 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8705 sd->parent = parent;
8706 parent->child = sd;
8707 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8708#endif
8709 return sd;
8710}
8711
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008712static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8713 const struct cpumask *cpu_map, int cpu)
8714{
8715 switch (l) {
8716#ifdef CONFIG_SCHED_SMT
8717 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8718 cpumask_and(d->this_sibling_map, cpu_map,
8719 topology_thread_cpumask(cpu));
8720 if (cpu == cpumask_first(d->this_sibling_map))
8721 init_sched_build_groups(d->this_sibling_map, cpu_map,
8722 &cpu_to_cpu_group,
8723 d->send_covered, d->tmpmask);
8724 break;
8725#endif
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008726#ifdef CONFIG_SCHED_MC
8727 case SD_LV_MC: /* set up multi-core groups */
8728 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8729 if (cpu == cpumask_first(d->this_core_map))
8730 init_sched_build_groups(d->this_core_map, cpu_map,
8731 &cpu_to_core_group,
8732 d->send_covered, d->tmpmask);
8733 break;
8734#endif
Andreas Herrmann86548092009-08-18 12:59:28 +02008735 case SD_LV_CPU: /* set up physical groups */
8736 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8737 if (!cpumask_empty(d->nodemask))
8738 init_sched_build_groups(d->nodemask, cpu_map,
8739 &cpu_to_phys_group,
8740 d->send_covered, d->tmpmask);
8741 break;
Andreas Herrmannde616e32009-08-18 13:00:13 +02008742#ifdef CONFIG_NUMA
8743 case SD_LV_ALLNODES:
8744 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8745 d->send_covered, d->tmpmask);
8746 break;
8747#endif
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008748 default:
8749 break;
8750 }
8751}
8752
Mike Travis7c16ec52008-04-04 18:11:11 -07008753/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008754 * Build sched domains for a given set of cpus and attach the sched domains
8755 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07008756 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308757static int __build_sched_domains(const struct cpumask *cpu_map,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008758 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008759{
Andreas Herrmann2109b992009-08-18 12:53:00 +02008760 enum s_alloc alloc_state = sa_none;
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008761 struct s_data d;
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008762 struct sched_domain *sd;
Andreas Herrmann2109b992009-08-18 12:53:00 +02008763 int i;
John Hawkesd1b55132005-09-06 15:18:14 -07008764#ifdef CONFIG_NUMA
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008765 d.sd_allnodes = 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308766#endif
8767
Andreas Herrmann2109b992009-08-18 12:53:00 +02008768 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8769 if (alloc_state != sa_rootdomain)
8770 goto error;
8771 alloc_state = sa_sched_groups;
Mike Travis7c16ec52008-04-04 18:11:11 -07008772
Linus Torvalds1da177e2005-04-16 15:20:36 -07008773 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008774 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008775 */
Rusty Russellabcd0832008-11-25 02:35:02 +10308776 for_each_cpu(i, cpu_map) {
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008777 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8778 cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008779
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008780 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
Andreas Herrmann87cce662009-08-18 12:54:55 +02008781 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmann410c4082009-08-18 12:56:14 +02008782 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmannd8173532009-08-18 12:57:03 +02008783 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008784 }
8785
Rusty Russellabcd0832008-11-25 02:35:02 +10308786 for_each_cpu(i, cpu_map) {
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008787 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008788 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008789 }
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008790
Linus Torvalds1da177e2005-04-16 15:20:36 -07008791 /* Set up physical groups */
Andreas Herrmann86548092009-08-18 12:59:28 +02008792 for (i = 0; i < nr_node_ids; i++)
8793 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008794
8795#ifdef CONFIG_NUMA
8796 /* Set up node groups */
Andreas Herrmannde616e32009-08-18 13:00:13 +02008797 if (d.sd_allnodes)
8798 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008799
Andreas Herrmann0601a882009-08-18 13:01:11 +02008800 for (i = 0; i < nr_node_ids; i++)
8801 if (build_numa_sched_groups(&d, cpu_map, i))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008802 goto error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008803#endif
8804
8805 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008806#ifdef CONFIG_SCHED_SMT
Rusty Russellabcd0832008-11-25 02:35:02 +10308807 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008808 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008809 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008810 }
8811#endif
8812#ifdef CONFIG_SCHED_MC
Rusty Russellabcd0832008-11-25 02:35:02 +10308813 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008814 sd = &per_cpu(core_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008815 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008816 }
8817#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008818
Rusty Russellabcd0832008-11-25 02:35:02 +10308819 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008820 sd = &per_cpu(phys_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008821 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008822 }
8823
John Hawkes9c1cfda2005-09-06 15:18:14 -07008824#ifdef CONFIG_NUMA
Mike Travis076ac2a2008-05-12 21:21:12 +02008825 for (i = 0; i < nr_node_ids; i++)
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008826 init_numa_sched_groups_power(d.sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008827
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008828 if (d.sd_allnodes) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008829 struct sched_group *sg;
Siddha, Suresh Bf712c0c72006-07-30 03:02:59 -07008830
Rusty Russell96f874e22008-11-25 02:35:14 +10308831 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008832 d.tmpmask);
Siddha, Suresh Bf712c0c72006-07-30 03:02:59 -07008833 init_numa_sched_groups_power(sg);
8834 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008835#endif
8836
Linus Torvalds1da177e2005-04-16 15:20:36 -07008837 /* Attach the domains */
Rusty Russellabcd0832008-11-25 02:35:02 +10308838 for_each_cpu(i, cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008839#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308840 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008841#elif defined(CONFIG_SCHED_MC)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308842 sd = &per_cpu(core_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008843#else
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308844 sd = &per_cpu(phys_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008845#endif
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008846 cpu_attach_domain(sd, d.rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008847 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008848
Andreas Herrmann2109b992009-08-18 12:53:00 +02008849 d.sched_group_nodes = NULL; /* don't free this we still need it */
8850 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8851 return 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308852
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008853error:
Andreas Herrmann2109b992009-08-18 12:53:00 +02008854 __free_domain_allocs(&d, alloc_state, cpu_map);
8855 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008856}
Paul Jackson029190c2007-10-18 23:40:20 -07008857
Rusty Russell96f874e22008-11-25 02:35:14 +10308858static int build_sched_domains(const struct cpumask *cpu_map)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008859{
8860 return __build_sched_domains(cpu_map, NULL);
8861}
8862
Rusty Russell96f874e22008-11-25 02:35:14 +10308863static struct cpumask *doms_cur; /* current sched domains */
Paul Jackson029190c2007-10-18 23:40:20 -07008864static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02008865static struct sched_domain_attr *dattr_cur;
8866 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07008867
8868/*
8869 * Special case: If a kmalloc of a doms_cur partition (array of
Rusty Russell42128232008-11-25 02:35:12 +10308870 * cpumask) fails, then fallback to a single sched domain,
8871 * as determined by the single cpumask fallback_doms.
Paul Jackson029190c2007-10-18 23:40:20 -07008872 */
Rusty Russell42128232008-11-25 02:35:12 +10308873static cpumask_var_t fallback_doms;
Paul Jackson029190c2007-10-18 23:40:20 -07008874
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008875/*
8876 * arch_update_cpu_topology lets virtualized architectures update the
8877 * cpu core maps. It is supposed to return 1 if the topology changed
8878 * or 0 if it stayed the same.
8879 */
8880int __attribute__((weak)) arch_update_cpu_topology(void)
Heiko Carstens22e52b02008-03-12 18:31:59 +01008881{
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008882 return 0;
Heiko Carstens22e52b02008-03-12 18:31:59 +01008883}
8884
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008885/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008886 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07008887 * For now this just excludes isolated cpus, but could be used to
8888 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008889 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308890static int arch_init_sched_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008891{
Milton Miller73785472007-10-24 18:23:48 +02008892 int err;
8893
Heiko Carstens22e52b02008-03-12 18:31:59 +01008894 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07008895 ndoms_cur = 1;
Rusty Russell96f874e22008-11-25 02:35:14 +10308896 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -07008897 if (!doms_cur)
Rusty Russell42128232008-11-25 02:35:12 +10308898 doms_cur = fallback_doms;
Rusty Russelldcc30a32008-11-25 02:35:12 +10308899 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008900 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02008901 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02008902 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02008903
8904 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008905}
8906
Rusty Russell96f874e22008-11-25 02:35:14 +10308907static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8908 struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008909{
Mike Travis7c16ec52008-04-04 18:11:11 -07008910 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008911}
Linus Torvalds1da177e2005-04-16 15:20:36 -07008912
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008913/*
8914 * Detach sched domains from a group of cpus specified in cpu_map
8915 * These cpus will now be attached to the NULL domain
8916 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308917static void detach_destroy_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008918{
Rusty Russell96f874e22008-11-25 02:35:14 +10308919 /* Save because hotplug lock held. */
8920 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008921 int i;
8922
Rusty Russellabcd0832008-11-25 02:35:02 +10308923 for_each_cpu(i, cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008924 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008925 synchronize_sched();
Rusty Russell96f874e22008-11-25 02:35:14 +10308926 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008927}
8928
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008929/* handle null as "default" */
8930static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8931 struct sched_domain_attr *new, int idx_new)
8932{
8933 struct sched_domain_attr tmp;
8934
8935 /* fast path */
8936 if (!new && !cur)
8937 return 1;
8938
8939 tmp = SD_ATTR_INIT;
8940 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8941 new ? (new + idx_new) : &tmp,
8942 sizeof(struct sched_domain_attr));
8943}
8944
Paul Jackson029190c2007-10-18 23:40:20 -07008945/*
8946 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008947 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07008948 * doms_new[] to the current sched domain partitioning, doms_cur[].
8949 * It destroys each deleted domain and builds each new domain.
8950 *
Rusty Russell96f874e22008-11-25 02:35:14 +10308951 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008952 * The masks don't intersect (don't overlap.) We should setup one
8953 * sched domain for each mask. CPUs not in any of the cpumasks will
8954 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07008955 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8956 * it as it is.
8957 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008958 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8959 * ownership of it and will kfree it when done with it. If the caller
Li Zefan700018e2008-11-18 14:02:03 +08008960 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8961 * ndoms_new == 1, and partition_sched_domains() will fallback to
8962 * the single partition 'fallback_doms', it also forces the domains
8963 * to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07008964 *
Rusty Russell96f874e22008-11-25 02:35:14 +10308965 * If doms_new == NULL it will be replaced with cpu_online_mask.
Li Zefan700018e2008-11-18 14:02:03 +08008966 * ndoms_new == 0 is a special case for destroying existing domains,
8967 * and it will not create the default domain.
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008968 *
Paul Jackson029190c2007-10-18 23:40:20 -07008969 * Call with hotplug lock held
8970 */
Rusty Russell96f874e22008-11-25 02:35:14 +10308971/* FIXME: Change to struct cpumask *doms_new[] */
8972void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008973 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07008974{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008975 int i, j, n;
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008976 int new_topology;
Paul Jackson029190c2007-10-18 23:40:20 -07008977
Heiko Carstens712555e2008-04-28 11:33:07 +02008978 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01008979
Milton Miller73785472007-10-24 18:23:48 +02008980 /* always unregister in case we don't destroy any domains */
8981 unregister_sched_domain_sysctl();
8982
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008983 /* Let architecture update cpu core mappings. */
8984 new_topology = arch_update_cpu_topology();
8985
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008986 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07008987
8988 /* Destroy deleted domains */
8989 for (i = 0; i < ndoms_cur; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008990 for (j = 0; j < n && !new_topology; j++) {
Rusty Russell96f874e22008-11-25 02:35:14 +10308991 if (cpumask_equal(&doms_cur[i], &doms_new[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008992 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07008993 goto match1;
8994 }
8995 /* no match - a current sched domain not in new doms_new[] */
8996 detach_destroy_domains(doms_cur + i);
8997match1:
8998 ;
8999 }
9000
Max Krasnyanskye761b772008-07-15 04:43:49 -07009001 if (doms_new == NULL) {
9002 ndoms_cur = 0;
Rusty Russell42128232008-11-25 02:35:12 +10309003 doms_new = fallback_doms;
Rusty Russelldcc30a32008-11-25 02:35:12 +10309004 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
Li Zefanfaa2f982008-11-04 16:20:23 +08009005 WARN_ON_ONCE(dattr_new);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009006 }
9007
Paul Jackson029190c2007-10-18 23:40:20 -07009008 /* Build new domains */
9009 for (i = 0; i < ndoms_new; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009010 for (j = 0; j < ndoms_cur && !new_topology; j++) {
Rusty Russell96f874e22008-11-25 02:35:14 +10309011 if (cpumask_equal(&doms_new[i], &doms_cur[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009012 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07009013 goto match2;
9014 }
9015 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009016 __build_sched_domains(doms_new + i,
9017 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07009018match2:
9019 ;
9020 }
9021
9022 /* Remember the new sched domains */
Rusty Russell42128232008-11-25 02:35:12 +10309023 if (doms_cur != fallback_doms)
Paul Jackson029190c2007-10-18 23:40:20 -07009024 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009025 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07009026 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009027 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07009028 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02009029
9030 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01009031
Heiko Carstens712555e2008-04-28 11:33:07 +02009032 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07009033}
9034
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009035#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Li Zefanc70f22d2009-01-05 19:07:50 +08009036static void arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009037{
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009038 get_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009039
9040 /* Destroy domains first to force the rebuild */
9041 partition_sched_domains(0, NULL, NULL);
9042
Max Krasnyanskye761b772008-07-15 04:43:49 -07009043 rebuild_sched_domains();
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009044 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009045}
9046
9047static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9048{
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309049 unsigned int level = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009050
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309051 if (sscanf(buf, "%u", &level) != 1)
9052 return -EINVAL;
9053
9054 /*
9055 * level is always be positive so don't check for
9056 * level < POWERSAVINGS_BALANCE_NONE which is 0
9057 * What happens on 0 or 1 byte write,
9058 * need to check for count as well?
9059 */
9060
9061 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009062 return -EINVAL;
9063
9064 if (smt)
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309065 sched_smt_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009066 else
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309067 sched_mc_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009068
Li Zefanc70f22d2009-01-05 19:07:50 +08009069 arch_reinit_sched_domains();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009070
Li Zefanc70f22d2009-01-05 19:07:50 +08009071 return count;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009072}
9073
Adrian Bunk6707de002007-08-12 18:08:19 +02009074#ifdef CONFIG_SCHED_MC
Andi Kleenf718cd42008-07-29 22:33:52 -07009075static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9076 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009077{
9078 return sprintf(page, "%u\n", sched_mc_power_savings);
9079}
Andi Kleenf718cd42008-07-29 22:33:52 -07009080static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
Adrian Bunk6707de002007-08-12 18:08:19 +02009081 const char *buf, size_t count)
9082{
9083 return sched_power_savings_store(buf, count, 0);
9084}
Andi Kleenf718cd42008-07-29 22:33:52 -07009085static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9086 sched_mc_power_savings_show,
9087 sched_mc_power_savings_store);
Adrian Bunk6707de002007-08-12 18:08:19 +02009088#endif
9089
9090#ifdef CONFIG_SCHED_SMT
Andi Kleenf718cd42008-07-29 22:33:52 -07009091static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9092 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009093{
9094 return sprintf(page, "%u\n", sched_smt_power_savings);
9095}
Andi Kleenf718cd42008-07-29 22:33:52 -07009096static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
Adrian Bunk6707de002007-08-12 18:08:19 +02009097 const char *buf, size_t count)
9098{
9099 return sched_power_savings_store(buf, count, 1);
9100}
Andi Kleenf718cd42008-07-29 22:33:52 -07009101static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9102 sched_smt_power_savings_show,
Adrian Bunk6707de002007-08-12 18:08:19 +02009103 sched_smt_power_savings_store);
9104#endif
9105
Li Zefan39aac642009-01-05 19:18:02 +08009106int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009107{
9108 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07009109
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009110#ifdef CONFIG_SCHED_SMT
9111 if (smt_capable())
9112 err = sysfs_create_file(&cls->kset.kobj,
9113 &attr_sched_smt_power_savings.attr);
9114#endif
9115#ifdef CONFIG_SCHED_MC
9116 if (!err && mc_capable())
9117 err = sysfs_create_file(&cls->kset.kobj,
9118 &attr_sched_mc_power_savings.attr);
9119#endif
9120 return err;
9121}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009122#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009123
Max Krasnyanskye761b772008-07-15 04:43:49 -07009124#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009125/*
Max Krasnyanskye761b772008-07-15 04:43:49 -07009126 * Add online and remove offline CPUs from the scheduler domains.
9127 * When cpusets are enabled they take over this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009128 */
9129static int update_sched_domains(struct notifier_block *nfb,
9130 unsigned long action, void *hcpu)
9131{
Max Krasnyanskye761b772008-07-15 04:43:49 -07009132 switch (action) {
9133 case CPU_ONLINE:
9134 case CPU_ONLINE_FROZEN:
9135 case CPU_DEAD:
9136 case CPU_DEAD_FROZEN:
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009137 partition_sched_domains(1, NULL, NULL);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009138 return NOTIFY_OK;
9139
9140 default:
9141 return NOTIFY_DONE;
9142 }
9143}
9144#endif
9145
9146static int update_runtime(struct notifier_block *nfb,
9147 unsigned long action, void *hcpu)
9148{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009149 int cpu = (int)(long)hcpu;
9150
Linus Torvalds1da177e2005-04-16 15:20:36 -07009151 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07009152 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009153 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009154 disable_runtime(cpu_rq(cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07009155 return NOTIFY_OK;
9156
Linus Torvalds1da177e2005-04-16 15:20:36 -07009157 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009158 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07009159 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009160 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009161 enable_runtime(cpu_rq(cpu));
Max Krasnyanskye761b772008-07-15 04:43:49 -07009162 return NOTIFY_OK;
9163
Linus Torvalds1da177e2005-04-16 15:20:36 -07009164 default:
9165 return NOTIFY_DONE;
9166 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009167}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009168
9169void __init sched_init_smp(void)
9170{
Rusty Russelldcc30a32008-11-25 02:35:12 +10309171 cpumask_var_t non_isolated_cpus;
9172
9173 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
Nick Piggin5c1e1762006-10-03 01:14:04 -07009174
Mike Travis434d53b2008-04-04 18:11:04 -07009175#if defined(CONFIG_NUMA)
9176 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9177 GFP_KERNEL);
9178 BUG_ON(sched_group_nodes_bycpu == NULL);
9179#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009180 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02009181 mutex_lock(&sched_domains_mutex);
Rusty Russelldcc30a32008-11-25 02:35:12 +10309182 arch_init_sched_domains(cpu_online_mask);
9183 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9184 if (cpumask_empty(non_isolated_cpus))
9185 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02009186 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009187 put_online_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009188
9189#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009190 /* XXX: Theoretical race here - CPU may be hotplugged now */
9191 hotcpu_notifier(update_sched_domains, 0);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009192#endif
9193
9194 /* RT runtime code needs to handle some hotplug events */
9195 hotcpu_notifier(update_runtime, 0);
9196
Peter Zijlstrab328ca12008-04-29 10:02:46 +02009197 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07009198
9199 /* Move init over to a non-isolated CPU */
Rusty Russelldcc30a32008-11-25 02:35:12 +10309200 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07009201 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01009202 sched_init_granularity();
Rusty Russelldcc30a32008-11-25 02:35:12 +10309203 free_cpumask_var(non_isolated_cpus);
Rusty Russell42128232008-11-25 02:35:12 +10309204
9205 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
Rusty Russell0e3900e2008-11-25 02:35:13 +10309206 init_sched_rt_class();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009207}
9208#else
9209void __init sched_init_smp(void)
9210{
Ingo Molnar19978ca2007-11-09 22:39:38 +01009211 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009212}
9213#endif /* CONFIG_SMP */
9214
Arun R Bharadwajcd1bb942009-04-16 12:15:34 +05309215const_debug unsigned int sysctl_timer_migration = 1;
9216
Linus Torvalds1da177e2005-04-16 15:20:36 -07009217int in_sched_functions(unsigned long addr)
9218{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009219 return in_lock_functions(addr) ||
9220 (addr >= (unsigned long)__sched_text_start
9221 && addr < (unsigned long)__sched_text_end);
9222}
9223
Alexey Dobriyana9957442007-10-15 17:00:13 +02009224static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02009225{
9226 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02009227 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02009228#ifdef CONFIG_FAIR_GROUP_SCHED
9229 cfs_rq->rq = rq;
9230#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02009231 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02009232}
9233
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009234static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9235{
9236 struct rt_prio_array *array;
9237 int i;
9238
9239 array = &rt_rq->active;
9240 for (i = 0; i < MAX_RT_PRIO; i++) {
9241 INIT_LIST_HEAD(array->queue + i);
9242 __clear_bit(i, array->bitmap);
9243 }
9244 /* delimiter for bitsearch: */
9245 __set_bit(MAX_RT_PRIO, array->bitmap);
9246
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009247#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -05009248 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins398a1532009-01-14 09:10:04 -05009249#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -05009250 rt_rq->highest_prio.next = MAX_RT_PRIO;
Peter Zijlstra48d5e252008-01-25 21:08:31 +01009251#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009252#endif
9253#ifdef CONFIG_SMP
9254 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009255 rt_rq->overloaded = 0;
Fabio Checconic20b08e2009-06-15 20:56:38 +02009256 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009257#endif
9258
9259 rt_rq->rt_time = 0;
9260 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009261 rt_rq->rt_runtime = 0;
9262 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009263
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009264#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01009265 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009266 rt_rq->rq = rq;
9267#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009268}
9269
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009270#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009271static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9272 struct sched_entity *se, int cpu, int add,
9273 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009274{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009275 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009276 tg->cfs_rq[cpu] = cfs_rq;
9277 init_cfs_rq(cfs_rq, rq);
9278 cfs_rq->tg = tg;
9279 if (add)
9280 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9281
9282 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009283 /* se could be NULL for init_task_group */
9284 if (!se)
9285 return;
9286
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009287 if (!parent)
9288 se->cfs_rq = &rq->cfs;
9289 else
9290 se->cfs_rq = parent->my_q;
9291
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009292 se->my_q = cfs_rq;
9293 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009294 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009295 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009296}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009297#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009298
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009299#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009300static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9301 struct sched_rt_entity *rt_se, int cpu, int add,
9302 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009303{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009304 struct rq *rq = cpu_rq(cpu);
9305
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009306 tg->rt_rq[cpu] = rt_rq;
9307 init_rt_rq(rt_rq, rq);
9308 rt_rq->tg = tg;
9309 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009310 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009311 if (add)
9312 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9313
9314 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009315 if (!rt_se)
9316 return;
9317
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009318 if (!parent)
9319 rt_se->rt_rq = &rq->rt;
9320 else
9321 rt_se->rt_rq = parent->my_q;
9322
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009323 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009324 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009325 INIT_LIST_HEAD(&rt_se->run_list);
9326}
9327#endif
9328
Linus Torvalds1da177e2005-04-16 15:20:36 -07009329void __init sched_init(void)
9330{
Ingo Molnardd41f592007-07-09 18:51:59 +02009331 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07009332 unsigned long alloc_size = 0, ptr;
9333
9334#ifdef CONFIG_FAIR_GROUP_SCHED
9335 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9336#endif
9337#ifdef CONFIG_RT_GROUP_SCHED
9338 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9339#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009340#ifdef CONFIG_USER_SCHED
9341 alloc_size *= 2;
9342#endif
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309343#ifdef CONFIG_CPUMASK_OFFSTACK
Rusty Russell8c083f02009-03-19 15:22:20 +10309344 alloc_size += num_possible_cpus() * cpumask_size();
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309345#endif
Mike Travis434d53b2008-04-04 18:11:04 -07009346 /*
9347 * As sched_init() is called before page_alloc is setup,
9348 * we use alloc_bootmem().
9349 */
9350 if (alloc_size) {
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03009351 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07009352
9353#ifdef CONFIG_FAIR_GROUP_SCHED
9354 init_task_group.se = (struct sched_entity **)ptr;
9355 ptr += nr_cpu_ids * sizeof(void **);
9356
9357 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9358 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009359
9360#ifdef CONFIG_USER_SCHED
9361 root_task_group.se = (struct sched_entity **)ptr;
9362 ptr += nr_cpu_ids * sizeof(void **);
9363
9364 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9365 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009366#endif /* CONFIG_USER_SCHED */
9367#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009368#ifdef CONFIG_RT_GROUP_SCHED
9369 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9370 ptr += nr_cpu_ids * sizeof(void **);
9371
9372 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009373 ptr += nr_cpu_ids * sizeof(void **);
9374
9375#ifdef CONFIG_USER_SCHED
9376 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9377 ptr += nr_cpu_ids * sizeof(void **);
9378
9379 root_task_group.rt_rq = (struct rt_rq **)ptr;
9380 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009381#endif /* CONFIG_USER_SCHED */
9382#endif /* CONFIG_RT_GROUP_SCHED */
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309383#ifdef CONFIG_CPUMASK_OFFSTACK
9384 for_each_possible_cpu(i) {
9385 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9386 ptr += cpumask_size();
9387 }
9388#endif /* CONFIG_CPUMASK_OFFSTACK */
Mike Travis434d53b2008-04-04 18:11:04 -07009389 }
Ingo Molnardd41f592007-07-09 18:51:59 +02009390
Gregory Haskins57d885f2008-01-25 21:08:18 +01009391#ifdef CONFIG_SMP
9392 init_defrootdomain();
9393#endif
9394
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009395 init_rt_bandwidth(&def_rt_bandwidth,
9396 global_rt_period(), global_rt_runtime());
9397
9398#ifdef CONFIG_RT_GROUP_SCHED
9399 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9400 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009401#ifdef CONFIG_USER_SCHED
9402 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9403 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009404#endif /* CONFIG_USER_SCHED */
9405#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009406
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009407#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009408 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009409 INIT_LIST_HEAD(&init_task_group.children);
9410
9411#ifdef CONFIG_USER_SCHED
9412 INIT_LIST_HEAD(&root_task_group.children);
9413 init_task_group.parent = &root_task_group;
9414 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009415#endif /* CONFIG_USER_SCHED */
9416#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009417
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08009418 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07009419 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009420
9421 rq = cpu_rq(i);
9422 spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07009423 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009424 rq->calc_load_active = 0;
9425 rq->calc_load_update = jiffies + LOAD_FREQ;
Ingo Molnardd41f592007-07-09 18:51:59 +02009426 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009427 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009428#ifdef CONFIG_FAIR_GROUP_SCHED
9429 init_task_group.shares = init_task_group_load;
9430 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009431#ifdef CONFIG_CGROUP_SCHED
9432 /*
9433 * How much cpu bandwidth does init_task_group get?
9434 *
9435 * In case of task-groups formed thr' the cgroup filesystem, it
9436 * gets 100% of the cpu resources in the system. This overall
9437 * system cpu resource is divided among the tasks of
9438 * init_task_group and its child task-groups in a fair manner,
9439 * based on each entity's (task or task-group's) weight
9440 * (se->load.weight).
9441 *
9442 * In other words, if init_task_group has 10 tasks of weight
9443 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9444 * then A0's share of the cpu resource is:
9445 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009446 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02009447 *
9448 * We achieve this by letting init_task_group's tasks sit
9449 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9450 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009451 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009452#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009453 root_task_group.shares = NICE_0_LOAD;
9454 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009455 /*
9456 * In case of task-groups formed thr' the user id of tasks,
9457 * init_task_group represents tasks belonging to root user.
9458 * Hence it forms a sibling of all subsequent groups formed.
9459 * In this case, init_task_group gets only a fraction of overall
9460 * system cpu resource, based on the weight assigned to root
9461 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9462 * by letting tasks of init_task_group sit in a separate cfs_rq
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009463 * (init_tg_cfs_rq) and having one entity represent this group of
Dhaval Giani354d60c2008-04-19 19:44:59 +02009464 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9465 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009466 init_tg_cfs_entry(&init_task_group,
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009467 &per_cpu(init_tg_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009468 &per_cpu(init_sched_entity, i), i, 1,
9469 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009470
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009471#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02009472#endif /* CONFIG_FAIR_GROUP_SCHED */
9473
9474 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009475#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009476 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009477#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009478 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009479#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009480 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009481 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009482 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009483 &per_cpu(init_sched_rt_entity, i), i, 1,
9484 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009485#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009486#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009487
Ingo Molnardd41f592007-07-09 18:51:59 +02009488 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9489 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009490#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07009491 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01009492 rq->rd = NULL;
Gregory Haskins3f029d32009-07-29 11:08:47 -04009493 rq->post_schedule = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009494 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009495 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009496 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07009497 rq->cpu = i;
Gregory Haskins1f11eb6a2008-06-04 15:04:05 -04009498 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009499 rq->migration_thread = NULL;
9500 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01009501 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009502#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009503 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009504 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009505 }
9506
Peter Williams2dd73a42006-06-27 02:54:34 -07009507 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009508
Avi Kivitye107be32007-07-26 13:40:43 +02009509#ifdef CONFIG_PREEMPT_NOTIFIERS
9510 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9511#endif
9512
Christoph Lameterc9819f42006-12-10 02:20:25 -08009513#ifdef CONFIG_SMP
Carlos R. Mafra962cf362008-05-15 11:15:37 -03009514 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
Christoph Lameterc9819f42006-12-10 02:20:25 -08009515#endif
9516
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009517#ifdef CONFIG_RT_MUTEXES
9518 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9519#endif
9520
Linus Torvalds1da177e2005-04-16 15:20:36 -07009521 /*
9522 * The boot idle thread does lazy MMU switching as well:
9523 */
9524 atomic_inc(&init_mm.mm_count);
9525 enter_lazy_tlb(&init_mm, current);
9526
9527 /*
9528 * Make us the idle thread. Technically, schedule() should not be
9529 * called from this thread, however somewhere below it might be,
9530 * but because we are the idle thread, we just pick up running again
9531 * when this runqueue becomes "idle".
9532 */
9533 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009534
9535 calc_load_update = jiffies + LOAD_FREQ;
9536
Ingo Molnardd41f592007-07-09 18:51:59 +02009537 /*
9538 * During early bootup we pretend to be a normal task:
9539 */
9540 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01009541
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309542 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009543 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309544#ifdef CONFIG_SMP
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309545#ifdef CONFIG_NO_HZ
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009546 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9547 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309548#endif
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009549 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309550#endif /* SMP */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309551
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009552 perf_counter_init();
9553
Ingo Molnar6892b752008-02-13 14:02:36 +01009554 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009555}
9556
9557#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009558static inline int preempt_count_equals(int preempt_offset)
9559{
9560 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9561
9562 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9563}
9564
9565void __might_sleep(char *file, int line, int preempt_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009566{
Ingo Molnar48f24c42006-07-03 00:25:40 -07009567#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07009568 static unsigned long prev_jiffy; /* ratelimiting */
9569
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009570 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9571 system_state != SYSTEM_RUNNING || oops_in_progress)
Ingo Molnaraef745f2008-08-28 11:34:43 +02009572 return;
9573 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9574 return;
9575 prev_jiffy = jiffies;
9576
9577 printk(KERN_ERR
9578 "BUG: sleeping function called from invalid context at %s:%d\n",
9579 file, line);
9580 printk(KERN_ERR
9581 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9582 in_atomic(), irqs_disabled(),
9583 current->pid, current->comm);
9584
9585 debug_show_held_locks(current);
9586 if (irqs_disabled())
9587 print_irqtrace_events(current);
9588 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009589#endif
9590}
9591EXPORT_SYMBOL(__might_sleep);
9592#endif
9593
9594#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009595static void normalize_task(struct rq *rq, struct task_struct *p)
9596{
9597 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02009598
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009599 update_rq_clock(rq);
9600 on_rq = p->se.on_rq;
9601 if (on_rq)
9602 deactivate_task(rq, p, 0);
9603 __setscheduler(rq, p, SCHED_NORMAL, 0);
9604 if (on_rq) {
9605 activate_task(rq, p, 0);
9606 resched_task(rq->curr);
9607 }
9608}
9609
Linus Torvalds1da177e2005-04-16 15:20:36 -07009610void normalize_rt_tasks(void)
9611{
Ingo Molnara0f98a12007-06-17 18:37:45 +02009612 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009613 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07009614 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009615
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009616 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009617 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02009618 /*
9619 * Only normalize user tasks:
9620 */
9621 if (!p->mm)
9622 continue;
9623
Ingo Molnardd41f592007-07-09 18:51:59 +02009624 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009625#ifdef CONFIG_SCHEDSTATS
9626 p->se.wait_start = 0;
9627 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009628 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009629#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02009630
9631 if (!rt_task(p)) {
9632 /*
9633 * Renice negative nice level userspace
9634 * tasks back to 0:
9635 */
9636 if (TASK_NICE(p) < 0 && p->mm)
9637 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009638 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02009639 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009640
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009641 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07009642 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009643
Ingo Molnar178be792007-10-15 17:00:18 +02009644 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009645
Ingo Molnarb29739f2006-06-27 02:54:51 -07009646 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009647 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009648 } while_each_thread(g, p);
9649
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009650 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009651}
9652
9653#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07009654
9655#ifdef CONFIG_IA64
9656/*
9657 * These functions are only useful for the IA64 MCA handling.
9658 *
9659 * They can only be called when the whole system has been
9660 * stopped - every CPU needs to be quiescent, and no scheduling
9661 * activity can take place. Using them for anything else would
9662 * be a serious bug, and as a result, they aren't even visible
9663 * under any other configuration.
9664 */
9665
9666/**
9667 * curr_task - return the current task for a given cpu.
9668 * @cpu: the processor in question.
9669 *
9670 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9671 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009672struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009673{
9674 return cpu_curr(cpu);
9675}
9676
9677/**
9678 * set_curr_task - set the current task for a given cpu.
9679 * @cpu: the processor in question.
9680 * @p: the task pointer to set.
9681 *
9682 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009683 * are serviced on a separate stack. It allows the architecture to switch the
9684 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07009685 * must be called with all CPU's synchronized, and interrupts disabled, the
9686 * and caller must save the original value of the current task (see
9687 * curr_task() above) and restore that value before reenabling interrupts and
9688 * re-starting the system.
9689 *
9690 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9691 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009692void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009693{
9694 cpu_curr(cpu) = p;
9695}
9696
9697#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009698
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009699#ifdef CONFIG_FAIR_GROUP_SCHED
9700static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009701{
9702 int i;
9703
9704 for_each_possible_cpu(i) {
9705 if (tg->cfs_rq)
9706 kfree(tg->cfs_rq[i]);
9707 if (tg->se)
9708 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009709 }
9710
9711 kfree(tg->cfs_rq);
9712 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009713}
9714
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009715static
9716int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009717{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009718 struct cfs_rq *cfs_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009719 struct sched_entity *se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009720 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009721 int i;
9722
Mike Travis434d53b2008-04-04 18:11:04 -07009723 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009724 if (!tg->cfs_rq)
9725 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009726 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009727 if (!tg->se)
9728 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009729
9730 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009731
9732 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009733 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009734
Li Zefaneab17222008-10-29 17:03:22 +08009735 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9736 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009737 if (!cfs_rq)
9738 goto err;
9739
Li Zefaneab17222008-10-29 17:03:22 +08009740 se = kzalloc_node(sizeof(struct sched_entity),
9741 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009742 if (!se)
9743 goto err;
9744
Li Zefaneab17222008-10-29 17:03:22 +08009745 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009746 }
9747
9748 return 1;
9749
9750 err:
9751 return 0;
9752}
9753
9754static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9755{
9756 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9757 &cpu_rq(cpu)->leaf_cfs_rq_list);
9758}
9759
9760static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9761{
9762 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9763}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009764#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009765static inline void free_fair_sched_group(struct task_group *tg)
9766{
9767}
9768
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009769static inline
9770int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009771{
9772 return 1;
9773}
9774
9775static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9776{
9777}
9778
9779static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9780{
9781}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009782#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009783
9784#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009785static void free_rt_sched_group(struct task_group *tg)
9786{
9787 int i;
9788
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009789 destroy_rt_bandwidth(&tg->rt_bandwidth);
9790
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009791 for_each_possible_cpu(i) {
9792 if (tg->rt_rq)
9793 kfree(tg->rt_rq[i]);
9794 if (tg->rt_se)
9795 kfree(tg->rt_se[i]);
9796 }
9797
9798 kfree(tg->rt_rq);
9799 kfree(tg->rt_se);
9800}
9801
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009802static
9803int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009804{
9805 struct rt_rq *rt_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009806 struct sched_rt_entity *rt_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009807 struct rq *rq;
9808 int i;
9809
Mike Travis434d53b2008-04-04 18:11:04 -07009810 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009811 if (!tg->rt_rq)
9812 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009813 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009814 if (!tg->rt_se)
9815 goto err;
9816
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009817 init_rt_bandwidth(&tg->rt_bandwidth,
9818 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009819
9820 for_each_possible_cpu(i) {
9821 rq = cpu_rq(i);
9822
Li Zefaneab17222008-10-29 17:03:22 +08009823 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9824 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009825 if (!rt_rq)
9826 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009827
Li Zefaneab17222008-10-29 17:03:22 +08009828 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9829 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009830 if (!rt_se)
9831 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009832
Li Zefaneab17222008-10-29 17:03:22 +08009833 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009834 }
9835
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009836 return 1;
9837
9838 err:
9839 return 0;
9840}
9841
9842static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9843{
9844 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9845 &cpu_rq(cpu)->leaf_rt_rq_list);
9846}
9847
9848static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9849{
9850 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9851}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009852#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009853static inline void free_rt_sched_group(struct task_group *tg)
9854{
9855}
9856
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009857static inline
9858int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009859{
9860 return 1;
9861}
9862
9863static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9864{
9865}
9866
9867static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9868{
9869}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009870#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009871
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009872#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009873static void free_sched_group(struct task_group *tg)
9874{
9875 free_fair_sched_group(tg);
9876 free_rt_sched_group(tg);
9877 kfree(tg);
9878}
9879
9880/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009881struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009882{
9883 struct task_group *tg;
9884 unsigned long flags;
9885 int i;
9886
9887 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9888 if (!tg)
9889 return ERR_PTR(-ENOMEM);
9890
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009891 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009892 goto err;
9893
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009894 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009895 goto err;
9896
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009897 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009898 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009899 register_fair_sched_group(tg, i);
9900 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009901 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009902 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009903
9904 WARN_ON(!parent); /* root should already exist */
9905
9906 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009907 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +08009908 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009909 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009910
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009911 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009912
9913err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009914 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009915 return ERR_PTR(-ENOMEM);
9916}
9917
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009918/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009919static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009920{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009921 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009922 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009923}
9924
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009925/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02009926void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009927{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009928 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009929 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009930
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009931 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009932 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009933 unregister_fair_sched_group(tg, i);
9934 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009935 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009936 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009937 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009938 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009939
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009940 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009941 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009942}
9943
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009944/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02009945 * The caller of this function should have put the task in its new group
9946 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9947 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009948 */
9949void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009950{
9951 int on_rq, running;
9952 unsigned long flags;
9953 struct rq *rq;
9954
9955 rq = task_rq_lock(tsk, &flags);
9956
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009957 update_rq_clock(rq);
9958
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01009959 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009960 on_rq = tsk->se.on_rq;
9961
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009962 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009963 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009964 if (unlikely(running))
9965 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009966
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009967 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009968
Peter Zijlstra810b3812008-02-29 15:21:01 -05009969#ifdef CONFIG_FAIR_GROUP_SCHED
9970 if (tsk->sched_class->moved_group)
9971 tsk->sched_class->moved_group(tsk);
9972#endif
9973
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009974 if (unlikely(running))
9975 tsk->sched_class->set_curr_task(rq);
9976 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02009977 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009978
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009979 task_rq_unlock(rq, &flags);
9980}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009981#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009982
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009983#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009984static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009985{
9986 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009987 int on_rq;
9988
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009989 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009990 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009991 dequeue_entity(cfs_rq, se, 0);
9992
9993 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009994 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009995
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009996 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009997 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009998}
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009999
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010000static void set_se_shares(struct sched_entity *se, unsigned long shares)
10001{
10002 struct cfs_rq *cfs_rq = se->cfs_rq;
10003 struct rq *rq = cfs_rq->rq;
10004 unsigned long flags;
10005
10006 spin_lock_irqsave(&rq->lock, flags);
10007 __set_se_shares(se, shares);
10008 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010009}
10010
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010011static DEFINE_MUTEX(shares_mutex);
10012
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010013int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010014{
10015 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010016 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +010010017
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010018 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010019 * We can't change the weight of the root cgroup.
10020 */
10021 if (!tg->se[0])
10022 return -EINVAL;
10023
Peter Zijlstra18d95a22008-04-19 19:45:00 +020010024 if (shares < MIN_SHARES)
10025 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010026 else if (shares > MAX_SHARES)
10027 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010028
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010029 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010030 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010031 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010032
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010033 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010034 for_each_possible_cpu(i)
10035 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010036 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010037 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010038
10039 /* wait for any ongoing reference to this group to finish */
10040 synchronize_sched();
10041
10042 /*
10043 * Now we are free to modify the group's share on each cpu
10044 * w/o tripping rebalance_share or load_balance_fair.
10045 */
10046 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010047 for_each_possible_cpu(i) {
10048 /*
10049 * force a rebalance
10050 */
10051 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010052 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010053 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010054
10055 /*
10056 * Enable load balance activity on this group, by inserting it back on
10057 * each cpu's rq->leaf_cfs_rq_list.
10058 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010059 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010060 for_each_possible_cpu(i)
10061 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010062 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010063 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010064done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010065 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010066 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010067}
10068
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010069unsigned long sched_group_shares(struct task_group *tg)
10070{
10071 return tg->shares;
10072}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010073#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010074
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010075#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010076/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010077 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010078 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010079static DEFINE_MUTEX(rt_constraints_mutex);
10080
10081static unsigned long to_ratio(u64 period, u64 runtime)
10082{
10083 if (runtime == RUNTIME_INF)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010084 return 1ULL << 20;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010085
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010086 return div64_u64(runtime << 20, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010087}
10088
Dhaval Giani521f1a242008-02-28 15:21:56 +053010089/* Must be called with tasklist_lock held */
10090static inline int tg_has_rt_tasks(struct task_group *tg)
10091{
10092 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010093
Dhaval Giani521f1a242008-02-28 15:21:56 +053010094 do_each_thread(g, p) {
10095 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10096 return 1;
10097 } while_each_thread(g, p);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010098
Dhaval Giani521f1a242008-02-28 15:21:56 +053010099 return 0;
10100}
10101
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010102struct rt_schedulable_data {
10103 struct task_group *tg;
10104 u64 rt_period;
10105 u64 rt_runtime;
10106};
10107
10108static int tg_schedulable(struct task_group *tg, void *data)
10109{
10110 struct rt_schedulable_data *d = data;
10111 struct task_group *child;
10112 unsigned long total, sum = 0;
10113 u64 period, runtime;
10114
10115 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10116 runtime = tg->rt_bandwidth.rt_runtime;
10117
10118 if (tg == d->tg) {
10119 period = d->rt_period;
10120 runtime = d->rt_runtime;
10121 }
10122
Peter Zijlstra98a48262009-01-14 10:56:32 +010010123#ifdef CONFIG_USER_SCHED
10124 if (tg == &root_task_group) {
10125 period = global_rt_period();
10126 runtime = global_rt_runtime();
10127 }
10128#endif
10129
Peter Zijlstra4653f802008-09-23 15:33:44 +020010130 /*
10131 * Cannot have more runtime than the period.
10132 */
10133 if (runtime > period && runtime != RUNTIME_INF)
10134 return -EINVAL;
10135
10136 /*
10137 * Ensure we don't starve existing RT tasks.
10138 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010139 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10140 return -EBUSY;
10141
10142 total = to_ratio(period, runtime);
10143
Peter Zijlstra4653f802008-09-23 15:33:44 +020010144 /*
10145 * Nobody can have more than the global setting allows.
10146 */
10147 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10148 return -EINVAL;
10149
10150 /*
10151 * The sum of our children's runtime should not exceed our own.
10152 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010153 list_for_each_entry_rcu(child, &tg->children, siblings) {
10154 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10155 runtime = child->rt_bandwidth.rt_runtime;
10156
10157 if (child == d->tg) {
10158 period = d->rt_period;
10159 runtime = d->rt_runtime;
10160 }
10161
10162 sum += to_ratio(period, runtime);
10163 }
10164
10165 if (sum > total)
10166 return -EINVAL;
10167
10168 return 0;
10169}
10170
10171static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10172{
10173 struct rt_schedulable_data data = {
10174 .tg = tg,
10175 .rt_period = period,
10176 .rt_runtime = runtime,
10177 };
10178
10179 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10180}
10181
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010182static int tg_set_bandwidth(struct task_group *tg,
10183 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010184{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010185 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010186
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010187 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +053010188 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010189 err = __rt_schedulable(tg, rt_period, rt_runtime);
10190 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +053010191 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010192
10193 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010194 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10195 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010196
10197 for_each_possible_cpu(i) {
10198 struct rt_rq *rt_rq = tg->rt_rq[i];
10199
10200 spin_lock(&rt_rq->rt_runtime_lock);
10201 rt_rq->rt_runtime = rt_runtime;
10202 spin_unlock(&rt_rq->rt_runtime_lock);
10203 }
10204 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010205 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +053010206 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010207 mutex_unlock(&rt_constraints_mutex);
10208
10209 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010210}
10211
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010212int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10213{
10214 u64 rt_runtime, rt_period;
10215
10216 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10217 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10218 if (rt_runtime_us < 0)
10219 rt_runtime = RUNTIME_INF;
10220
10221 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10222}
10223
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010224long sched_group_rt_runtime(struct task_group *tg)
10225{
10226 u64 rt_runtime_us;
10227
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010228 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010229 return -1;
10230
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010231 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010232 do_div(rt_runtime_us, NSEC_PER_USEC);
10233 return rt_runtime_us;
10234}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010235
10236int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10237{
10238 u64 rt_runtime, rt_period;
10239
10240 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10241 rt_runtime = tg->rt_bandwidth.rt_runtime;
10242
Raistlin619b0482008-06-26 18:54:09 +020010243 if (rt_period == 0)
10244 return -EINVAL;
10245
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010246 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10247}
10248
10249long sched_group_rt_period(struct task_group *tg)
10250{
10251 u64 rt_period_us;
10252
10253 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10254 do_div(rt_period_us, NSEC_PER_USEC);
10255 return rt_period_us;
10256}
10257
10258static int sched_rt_global_constraints(void)
10259{
Peter Zijlstra4653f802008-09-23 15:33:44 +020010260 u64 runtime, period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010261 int ret = 0;
10262
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010263 if (sysctl_sched_rt_period <= 0)
10264 return -EINVAL;
10265
Peter Zijlstra4653f802008-09-23 15:33:44 +020010266 runtime = global_rt_runtime();
10267 period = global_rt_period();
10268
10269 /*
10270 * Sanity check on the sysctl variables.
10271 */
10272 if (runtime > period && runtime != RUNTIME_INF)
10273 return -EINVAL;
Peter Zijlstra10b612f2008-06-19 14:22:27 +020010274
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010275 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010276 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +020010277 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010278 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010279 mutex_unlock(&rt_constraints_mutex);
10280
10281 return ret;
10282}
Dhaval Giani54e99122009-02-27 15:13:54 +053010283
10284int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10285{
10286 /* Don't accept realtime tasks when there is no way for them to run */
10287 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10288 return 0;
10289
10290 return 1;
10291}
10292
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010293#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010294static int sched_rt_global_constraints(void)
10295{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010296 unsigned long flags;
10297 int i;
10298
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010299 if (sysctl_sched_rt_period <= 0)
10300 return -EINVAL;
10301
Peter Zijlstra60aa6052009-05-05 17:50:21 +020010302 /*
10303 * There's always some RT tasks in the root group
10304 * -- migration, kstopmachine etc..
10305 */
10306 if (sysctl_sched_rt_runtime == 0)
10307 return -EBUSY;
10308
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010309 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10310 for_each_possible_cpu(i) {
10311 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10312
10313 spin_lock(&rt_rq->rt_runtime_lock);
10314 rt_rq->rt_runtime = global_rt_runtime();
10315 spin_unlock(&rt_rq->rt_runtime_lock);
10316 }
10317 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10318
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010319 return 0;
10320}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010321#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010322
10323int sched_rt_handler(struct ctl_table *table, int write,
10324 struct file *filp, void __user *buffer, size_t *lenp,
10325 loff_t *ppos)
10326{
10327 int ret;
10328 int old_period, old_runtime;
10329 static DEFINE_MUTEX(mutex);
10330
10331 mutex_lock(&mutex);
10332 old_period = sysctl_sched_rt_period;
10333 old_runtime = sysctl_sched_rt_runtime;
10334
10335 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10336
10337 if (!ret && write) {
10338 ret = sched_rt_global_constraints();
10339 if (ret) {
10340 sysctl_sched_rt_period = old_period;
10341 sysctl_sched_rt_runtime = old_runtime;
10342 } else {
10343 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10344 def_rt_bandwidth.rt_period =
10345 ns_to_ktime(global_rt_period());
10346 }
10347 }
10348 mutex_unlock(&mutex);
10349
10350 return ret;
10351}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010352
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010353#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010354
10355/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +020010356static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010357{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010358 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10359 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010360}
10361
10362static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +020010363cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010364{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010365 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010366
Paul Menage2b01dfe2007-10-24 18:23:50 +020010367 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010368 /* This is early initialization for the top cgroup */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010369 return &init_task_group.css;
10370 }
10371
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010372 parent = cgroup_tg(cgrp->parent);
10373 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010374 if (IS_ERR(tg))
10375 return ERR_PTR(-ENOMEM);
10376
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010377 return &tg->css;
10378}
10379
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010380static void
10381cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010382{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010383 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010384
10385 sched_destroy_group(tg);
10386}
10387
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010388static int
10389cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10390 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010391{
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010392#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Giani54e99122009-02-27 15:13:54 +053010393 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010394 return -EINVAL;
10395#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010396 /* We don't support RT-tasks being in separate groups */
10397 if (tsk->sched_class != &fair_sched_class)
10398 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010399#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010400
10401 return 0;
10402}
10403
10404static void
Paul Menage2b01dfe2007-10-24 18:23:50 +020010405cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010406 struct cgroup *old_cont, struct task_struct *tsk)
10407{
10408 sched_move_task(tsk);
10409}
10410
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010411#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -070010412static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +020010413 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010414{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010415 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010416}
10417
Paul Menagef4c753b2008-04-29 00:59:56 -070010418static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010419{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010420 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010421
10422 return (u64) tg->shares;
10423}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010424#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010425
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010426#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -070010427static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -070010428 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010429{
Paul Menage06ecb272008-04-29 01:00:06 -070010430 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010431}
10432
Paul Menage06ecb272008-04-29 01:00:06 -070010433static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010434{
Paul Menage06ecb272008-04-29 01:00:06 -070010435 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010436}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010437
10438static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10439 u64 rt_period_us)
10440{
10441 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10442}
10443
10444static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10445{
10446 return sched_group_rt_period(cgroup_tg(cgrp));
10447}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010448#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010449
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010450static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010451#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010452 {
10453 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -070010454 .read_u64 = cpu_shares_read_u64,
10455 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010456 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010457#endif
10458#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010459 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010460 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -070010461 .read_s64 = cpu_rt_runtime_read,
10462 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010463 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010464 {
10465 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -070010466 .read_u64 = cpu_rt_period_read_uint,
10467 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010468 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010469#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010470};
10471
10472static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10473{
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010474 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010475}
10476
10477struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +010010478 .name = "cpu",
10479 .create = cpu_cgroup_create,
10480 .destroy = cpu_cgroup_destroy,
10481 .can_attach = cpu_cgroup_can_attach,
10482 .attach = cpu_cgroup_attach,
10483 .populate = cpu_cgroup_populate,
10484 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010485 .early_init = 1,
10486};
10487
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010488#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010489
10490#ifdef CONFIG_CGROUP_CPUACCT
10491
10492/*
10493 * CPU accounting code for task groups.
10494 *
10495 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10496 * (balbir@in.ibm.com).
10497 */
10498
Bharata B Rao934352f2008-11-10 20:41:13 +053010499/* track cpu usage of a group of tasks and its child groups */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010500struct cpuacct {
10501 struct cgroup_subsys_state css;
10502 /* cpuusage holds pointer to a u64-type object on every cpu */
10503 u64 *cpuusage;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010504 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
Bharata B Rao934352f2008-11-10 20:41:13 +053010505 struct cpuacct *parent;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010506};
10507
10508struct cgroup_subsys cpuacct_subsys;
10509
10510/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010511static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010512{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010513 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010514 struct cpuacct, css);
10515}
10516
10517/* return cpu accounting group to which this task belongs */
10518static inline struct cpuacct *task_ca(struct task_struct *tsk)
10519{
10520 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10521 struct cpuacct, css);
10522}
10523
10524/* create a new cpu accounting group */
10525static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +053010526 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010527{
10528 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010529 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010530
10531 if (!ca)
Bharata B Raoef12fef2009-03-31 10:02:22 +053010532 goto out;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010533
10534 ca->cpuusage = alloc_percpu(u64);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010535 if (!ca->cpuusage)
10536 goto out_free_ca;
10537
10538 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10539 if (percpu_counter_init(&ca->cpustat[i], 0))
10540 goto out_free_counters;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010541
Bharata B Rao934352f2008-11-10 20:41:13 +053010542 if (cgrp->parent)
10543 ca->parent = cgroup_ca(cgrp->parent);
10544
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010545 return &ca->css;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010546
10547out_free_counters:
10548 while (--i >= 0)
10549 percpu_counter_destroy(&ca->cpustat[i]);
10550 free_percpu(ca->cpuusage);
10551out_free_ca:
10552 kfree(ca);
10553out:
10554 return ERR_PTR(-ENOMEM);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010555}
10556
10557/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010558static void
Dhaval Giani32cd7562008-02-29 10:02:43 +053010559cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010560{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010561 struct cpuacct *ca = cgroup_ca(cgrp);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010562 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010563
Bharata B Raoef12fef2009-03-31 10:02:22 +053010564 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10565 percpu_counter_destroy(&ca->cpustat[i]);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010566 free_percpu(ca->cpuusage);
10567 kfree(ca);
10568}
10569
Ken Chen720f5492008-12-15 22:02:01 -080010570static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10571{
Rusty Russellb36128c2009-02-20 16:29:08 +090010572 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010573 u64 data;
10574
10575#ifndef CONFIG_64BIT
10576 /*
10577 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10578 */
10579 spin_lock_irq(&cpu_rq(cpu)->lock);
10580 data = *cpuusage;
10581 spin_unlock_irq(&cpu_rq(cpu)->lock);
10582#else
10583 data = *cpuusage;
10584#endif
10585
10586 return data;
10587}
10588
10589static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10590{
Rusty Russellb36128c2009-02-20 16:29:08 +090010591 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010592
10593#ifndef CONFIG_64BIT
10594 /*
10595 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10596 */
10597 spin_lock_irq(&cpu_rq(cpu)->lock);
10598 *cpuusage = val;
10599 spin_unlock_irq(&cpu_rq(cpu)->lock);
10600#else
10601 *cpuusage = val;
10602#endif
10603}
10604
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010605/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010606static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010607{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010608 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010609 u64 totalcpuusage = 0;
10610 int i;
10611
Ken Chen720f5492008-12-15 22:02:01 -080010612 for_each_present_cpu(i)
10613 totalcpuusage += cpuacct_cpuusage_read(ca, i);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010614
10615 return totalcpuusage;
10616}
10617
Dhaval Giani0297b802008-02-29 10:02:44 +053010618static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10619 u64 reset)
10620{
10621 struct cpuacct *ca = cgroup_ca(cgrp);
10622 int err = 0;
10623 int i;
10624
10625 if (reset) {
10626 err = -EINVAL;
10627 goto out;
10628 }
10629
Ken Chen720f5492008-12-15 22:02:01 -080010630 for_each_present_cpu(i)
10631 cpuacct_cpuusage_write(ca, i, 0);
Dhaval Giani0297b802008-02-29 10:02:44 +053010632
Dhaval Giani0297b802008-02-29 10:02:44 +053010633out:
10634 return err;
10635}
10636
Ken Chene9515c32008-12-15 22:04:15 -080010637static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10638 struct seq_file *m)
10639{
10640 struct cpuacct *ca = cgroup_ca(cgroup);
10641 u64 percpu;
10642 int i;
10643
10644 for_each_present_cpu(i) {
10645 percpu = cpuacct_cpuusage_read(ca, i);
10646 seq_printf(m, "%llu ", (unsigned long long) percpu);
10647 }
10648 seq_printf(m, "\n");
10649 return 0;
10650}
10651
Bharata B Raoef12fef2009-03-31 10:02:22 +053010652static const char *cpuacct_stat_desc[] = {
10653 [CPUACCT_STAT_USER] = "user",
10654 [CPUACCT_STAT_SYSTEM] = "system",
10655};
10656
10657static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10658 struct cgroup_map_cb *cb)
10659{
10660 struct cpuacct *ca = cgroup_ca(cgrp);
10661 int i;
10662
10663 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10664 s64 val = percpu_counter_read(&ca->cpustat[i]);
10665 val = cputime64_to_clock_t(val);
10666 cb->fill(cb, cpuacct_stat_desc[i], val);
10667 }
10668 return 0;
10669}
10670
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010671static struct cftype files[] = {
10672 {
10673 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -070010674 .read_u64 = cpuusage_read,
10675 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010676 },
Ken Chene9515c32008-12-15 22:04:15 -080010677 {
10678 .name = "usage_percpu",
10679 .read_seq_string = cpuacct_percpu_seq_read,
10680 },
Bharata B Raoef12fef2009-03-31 10:02:22 +053010681 {
10682 .name = "stat",
10683 .read_map = cpuacct_stats_show,
10684 },
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010685};
10686
Dhaval Giani32cd7562008-02-29 10:02:43 +053010687static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010688{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010689 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010690}
10691
10692/*
10693 * charge this task's execution time to its accounting group.
10694 *
10695 * called with rq->lock held.
10696 */
10697static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10698{
10699 struct cpuacct *ca;
Bharata B Rao934352f2008-11-10 20:41:13 +053010700 int cpu;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010701
Li Zefanc40c6f82009-02-26 15:40:15 +080010702 if (unlikely(!cpuacct_subsys.active))
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010703 return;
10704
Bharata B Rao934352f2008-11-10 20:41:13 +053010705 cpu = task_cpu(tsk);
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010706
10707 rcu_read_lock();
10708
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010709 ca = task_ca(tsk);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010710
Bharata B Rao934352f2008-11-10 20:41:13 +053010711 for (; ca; ca = ca->parent) {
Rusty Russellb36128c2009-02-20 16:29:08 +090010712 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010713 *cpuusage += cputime;
10714 }
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010715
10716 rcu_read_unlock();
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010717}
10718
Bharata B Raoef12fef2009-03-31 10:02:22 +053010719/*
10720 * Charge the system/user time to the task's accounting group.
10721 */
10722static void cpuacct_update_stats(struct task_struct *tsk,
10723 enum cpuacct_stat_index idx, cputime_t val)
10724{
10725 struct cpuacct *ca;
10726
10727 if (unlikely(!cpuacct_subsys.active))
10728 return;
10729
10730 rcu_read_lock();
10731 ca = task_ca(tsk);
10732
10733 do {
10734 percpu_counter_add(&ca->cpustat[idx], val);
10735 ca = ca->parent;
10736 } while (ca);
10737 rcu_read_unlock();
10738}
10739
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010740struct cgroup_subsys cpuacct_subsys = {
10741 .name = "cpuacct",
10742 .create = cpuacct_create,
10743 .destroy = cpuacct_destroy,
10744 .populate = cpuacct_populate,
10745 .subsys_id = cpuacct_subsys_id,
10746};
10747#endif /* CONFIG_CGROUP_CPUACCT */
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010748
10749#ifndef CONFIG_SMP
10750
10751int rcu_expedited_torture_stats(char *page)
10752{
10753 return 0;
10754}
10755EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10756
10757void synchronize_sched_expedited(void)
10758{
10759}
10760EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10761
10762#else /* #ifndef CONFIG_SMP */
10763
10764static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10765static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10766
10767#define RCU_EXPEDITED_STATE_POST -2
10768#define RCU_EXPEDITED_STATE_IDLE -1
10769
10770static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10771
10772int rcu_expedited_torture_stats(char *page)
10773{
10774 int cnt = 0;
10775 int cpu;
10776
10777 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10778 for_each_online_cpu(cpu) {
10779 cnt += sprintf(&page[cnt], " %d:%d",
10780 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10781 }
10782 cnt += sprintf(&page[cnt], "\n");
10783 return cnt;
10784}
10785EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10786
10787static long synchronize_sched_expedited_count;
10788
10789/*
10790 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10791 * approach to force grace period to end quickly. This consumes
10792 * significant time on all CPUs, and is thus not recommended for
10793 * any sort of common-case code.
10794 *
10795 * Note that it is illegal to call this function while holding any
10796 * lock that is acquired by a CPU-hotplug notifier. Failing to
10797 * observe this restriction will result in deadlock.
10798 */
10799void synchronize_sched_expedited(void)
10800{
10801 int cpu;
10802 unsigned long flags;
10803 bool need_full_sync = 0;
10804 struct rq *rq;
10805 struct migration_req *req;
10806 long snap;
10807 int trycount = 0;
10808
10809 smp_mb(); /* ensure prior mod happens before capturing snap. */
10810 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10811 get_online_cpus();
10812 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10813 put_online_cpus();
10814 if (trycount++ < 10)
10815 udelay(trycount * num_online_cpus());
10816 else {
10817 synchronize_sched();
10818 return;
10819 }
10820 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10821 smp_mb(); /* ensure test happens before caller kfree */
10822 return;
10823 }
10824 get_online_cpus();
10825 }
10826 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10827 for_each_online_cpu(cpu) {
10828 rq = cpu_rq(cpu);
10829 req = &per_cpu(rcu_migration_req, cpu);
10830 init_completion(&req->done);
10831 req->task = NULL;
10832 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10833 spin_lock_irqsave(&rq->lock, flags);
10834 list_add(&req->list, &rq->migration_queue);
10835 spin_unlock_irqrestore(&rq->lock, flags);
10836 wake_up_process(rq->migration_thread);
10837 }
10838 for_each_online_cpu(cpu) {
10839 rcu_expedited_state = cpu;
10840 req = &per_cpu(rcu_migration_req, cpu);
10841 rq = cpu_rq(cpu);
10842 wait_for_completion(&req->done);
10843 spin_lock_irqsave(&rq->lock, flags);
10844 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10845 need_full_sync = 1;
10846 req->dest_cpu = RCU_MIGRATION_IDLE;
10847 spin_unlock_irqrestore(&rq->lock, flags);
10848 }
10849 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10850 mutex_unlock(&rcu_sched_expedited_mutex);
10851 put_online_cpus();
10852 if (need_full_sync)
10853 synchronize_sched();
10854}
10855EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10856
10857#endif /* #else #ifndef CONFIG_SMP */