sched: reduce stack size requirements in kernel/sched.c
[linux-2.6-block.git] / kernel / sched.c
CommitLineData
1da177e4
LT
1/*
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
c31f2e8a
IM
19 * 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
b9131769
IM
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
1da177e4
LT
27 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
dff06c15 33#include <linux/uaccess.h>
1da177e4
LT
34#include <linux/highmem.h>
35#include <linux/smp_lock.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
c59ede7b 38#include <linux/capability.h>
1da177e4
LT
39#include <linux/completion.h>
40#include <linux/kernel_stat.h>
9a11b49a 41#include <linux/debug_locks.h>
1da177e4
LT
42#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
7dfb7103 45#include <linux/freezer.h>
198e2f18 46#include <linux/vmalloc.h>
1da177e4
LT
47#include <linux/blkdev.h>
48#include <linux/delay.h>
b488893a 49#include <linux/pid_namespace.h>
1da177e4
LT
50#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/kthread.h>
b5aadf7f 58#include <linux/proc_fs.h>
1da177e4 59#include <linux/seq_file.h>
e692ab53 60#include <linux/sysctl.h>
1da177e4
LT
61#include <linux/syscalls.h>
62#include <linux/times.h>
8f0ab514 63#include <linux/tsacct_kern.h>
c6fd91f0 64#include <linux/kprobes.h>
0ff92245 65#include <linux/delayacct.h>
5517d86b 66#include <linux/reciprocal_div.h>
dff06c15 67#include <linux/unistd.h>
f5ff8422 68#include <linux/pagemap.h>
8f4d37ec 69#include <linux/hrtimer.h>
30914a58 70#include <linux/tick.h>
434d53b0 71#include <linux/bootmem.h>
f00b45c1
PZ
72#include <linux/debugfs.h>
73#include <linux/ctype.h>
6cd8a4bb 74#include <linux/ftrace.h>
0a16b607 75#include <trace/sched.h>
1da177e4 76
5517d86b 77#include <asm/tlb.h>
838225b4 78#include <asm/irq_regs.h>
1da177e4 79
6e0534f2
GH
80#include "sched_cpupri.h"
81
1da177e4
LT
82/*
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 * and back.
86 */
87#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90
91/*
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
95 */
96#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99
100/*
d7876a08 101 * Helpers for converting nanosecond timing to jiffy resolution
1da177e4 102 */
d6322faf 103#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
1da177e4 104
6aa645ea
IM
105#define NICE_0_LOAD SCHED_LOAD_SCALE
106#define NICE_0_SHIFT SCHED_LOAD_SHIFT
107
1da177e4
LT
108/*
109 * These are the 'tuning knobs' of the scheduler:
110 *
a4ec24b4 111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
1da177e4
LT
112 * Timeslices get refilled after they expire.
113 */
1da177e4 114#define DEF_TIMESLICE (100 * HZ / 1000)
2dd73a4f 115
d0b27fa7
PZ
116/*
117 * single value that denotes runtime == period, ie unlimited time.
118 */
119#define RUNTIME_INF ((u64)~0ULL)
120
7e066fb8
MD
121DEFINE_TRACE(sched_wait_task);
122DEFINE_TRACE(sched_wakeup);
123DEFINE_TRACE(sched_wakeup_new);
124DEFINE_TRACE(sched_switch);
125DEFINE_TRACE(sched_migrate_task);
126
5517d86b
ED
127#ifdef CONFIG_SMP
128/*
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 */
132static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133{
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
135}
136
137/*
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
140 */
141static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142{
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
145}
146#endif
147
e05606d3
IM
148static inline int rt_policy(int policy)
149{
3f33a7ce 150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
e05606d3
IM
151 return 1;
152 return 0;
153}
154
155static inline int task_has_rt_policy(struct task_struct *p)
156{
157 return rt_policy(p->policy);
158}
159
1da177e4 160/*
6aa645ea 161 * This is the priority-queue data structure of the RT scheduling class:
1da177e4 162 */
6aa645ea
IM
163struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
166};
167
d0b27fa7 168struct rt_bandwidth {
ea736ed5
IM
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
171 ktime_t rt_period;
172 u64 rt_runtime;
173 struct hrtimer rt_period_timer;
d0b27fa7
PZ
174};
175
176static struct rt_bandwidth def_rt_bandwidth;
177
178static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179
180static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181{
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
184 ktime_t now;
185 int overrun;
186 int idle = 0;
187
188 for (;;) {
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
191
192 if (!overrun)
193 break;
194
195 idle = do_sched_rt_period_timer(rt_b, overrun);
196 }
197
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
199}
200
201static
202void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203{
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
206
ac086bc2
PZ
207 spin_lock_init(&rt_b->rt_runtime_lock);
208
d0b27fa7
PZ
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
ccc7dadf 212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
d0b27fa7
PZ
213}
214
c8bfff6d
KH
215static inline int rt_bandwidth_enabled(void)
216{
217 return sysctl_sched_rt_runtime >= 0;
d0b27fa7
PZ
218}
219
220static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
221{
222 ktime_t now;
223
0b148fa0 224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
d0b27fa7
PZ
225 return;
226
227 if (hrtimer_active(&rt_b->rt_period_timer))
228 return;
229
230 spin_lock(&rt_b->rt_runtime_lock);
231 for (;;) {
232 if (hrtimer_active(&rt_b->rt_period_timer))
233 break;
234
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
cc584b21
AV
237 hrtimer_start_expires(&rt_b->rt_period_timer,
238 HRTIMER_MODE_ABS);
d0b27fa7
PZ
239 }
240 spin_unlock(&rt_b->rt_runtime_lock);
241}
242
243#ifdef CONFIG_RT_GROUP_SCHED
244static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
245{
246 hrtimer_cancel(&rt_b->rt_period_timer);
247}
248#endif
249
712555ee
HC
250/*
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
253 */
254static DEFINE_MUTEX(sched_domains_mutex);
255
052f1dc7 256#ifdef CONFIG_GROUP_SCHED
29f59db3 257
68318b8e
SV
258#include <linux/cgroup.h>
259
29f59db3
SV
260struct cfs_rq;
261
6f505b16
PZ
262static LIST_HEAD(task_groups);
263
29f59db3 264/* task group related information */
4cf86d77 265struct task_group {
052f1dc7 266#ifdef CONFIG_CGROUP_SCHED
68318b8e
SV
267 struct cgroup_subsys_state css;
268#endif
052f1dc7
PZ
269
270#ifdef CONFIG_FAIR_GROUP_SCHED
29f59db3
SV
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity **se;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq **cfs_rq;
275 unsigned long shares;
052f1dc7
PZ
276#endif
277
278#ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity **rt_se;
280 struct rt_rq **rt_rq;
281
d0b27fa7 282 struct rt_bandwidth rt_bandwidth;
052f1dc7 283#endif
6b2d7700 284
ae8393e5 285 struct rcu_head rcu;
6f505b16 286 struct list_head list;
f473aa5e
PZ
287
288 struct task_group *parent;
289 struct list_head siblings;
290 struct list_head children;
29f59db3
SV
291};
292
354d60c2 293#ifdef CONFIG_USER_SCHED
eff766a6
PZ
294
295/*
296 * Root task group.
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
299 */
300struct task_group root_task_group;
301
052f1dc7 302#ifdef CONFIG_FAIR_GROUP_SCHED
29f59db3
SV
303/* Default task group's sched entity on each cpu */
304static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
305/* Default task group's cfs_rq on each cpu */
306static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
6d6bc0ad 307#endif /* CONFIG_FAIR_GROUP_SCHED */
052f1dc7
PZ
308
309#ifdef CONFIG_RT_GROUP_SCHED
310static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
311static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
6d6bc0ad 312#endif /* CONFIG_RT_GROUP_SCHED */
9a7e0b18 313#else /* !CONFIG_USER_SCHED */
eff766a6 314#define root_task_group init_task_group
9a7e0b18 315#endif /* CONFIG_USER_SCHED */
6f505b16 316
8ed36996 317/* task_group_lock serializes add/remove of task groups and also changes to
ec2c507f
SV
318 * a task group's cpu shares.
319 */
8ed36996 320static DEFINE_SPINLOCK(task_group_lock);
ec2c507f 321
052f1dc7 322#ifdef CONFIG_FAIR_GROUP_SCHED
052f1dc7
PZ
323#ifdef CONFIG_USER_SCHED
324# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
6d6bc0ad 325#else /* !CONFIG_USER_SCHED */
052f1dc7 326# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
6d6bc0ad 327#endif /* CONFIG_USER_SCHED */
052f1dc7 328
cb4ad1ff 329/*
2e084786
LJ
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
cb4ad1ff
MX
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
336 */
18d95a28 337#define MIN_SHARES 2
2e084786 338#define MAX_SHARES (1UL << 18)
18d95a28 339
052f1dc7
PZ
340static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341#endif
342
29f59db3 343/* Default task group.
3a252015 344 * Every task in system belong to this group at bootup.
29f59db3 345 */
434d53b0 346struct task_group init_task_group;
29f59db3
SV
347
348/* return group to which a task belongs */
4cf86d77 349static inline struct task_group *task_group(struct task_struct *p)
29f59db3 350{
4cf86d77 351 struct task_group *tg;
9b5b7751 352
052f1dc7 353#ifdef CONFIG_USER_SCHED
24e377a8 354 tg = p->user->tg;
052f1dc7 355#elif defined(CONFIG_CGROUP_SCHED)
68318b8e
SV
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
24e377a8 358#else
41a2d6cf 359 tg = &init_task_group;
24e377a8 360#endif
9b5b7751 361 return tg;
29f59db3
SV
362}
363
364/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
6f505b16 365static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
29f59db3 366{
052f1dc7 367#ifdef CONFIG_FAIR_GROUP_SCHED
ce96b5ac
DA
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
052f1dc7 370#endif
6f505b16 371
052f1dc7 372#ifdef CONFIG_RT_GROUP_SCHED
6f505b16
PZ
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
052f1dc7 375#endif
29f59db3
SV
376}
377
378#else
379
6f505b16 380static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
83378269
PZ
381static inline struct task_group *task_group(struct task_struct *p)
382{
383 return NULL;
384}
29f59db3 385
052f1dc7 386#endif /* CONFIG_GROUP_SCHED */
29f59db3 387
6aa645ea
IM
388/* CFS-related fields in a runqueue */
389struct cfs_rq {
390 struct load_weight load;
391 unsigned long nr_running;
392
6aa645ea 393 u64 exec_clock;
e9acbff6 394 u64 min_vruntime;
6aa645ea
IM
395
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
4a55bd5e
PZ
398
399 struct list_head tasks;
400 struct list_head *balance_iterator;
401
402 /*
403 * 'curr' points to currently running entity on this cfs_rq.
6aa645ea
IM
404 * It is set to NULL otherwise (i.e when none are currently running).
405 */
4793241b 406 struct sched_entity *curr, *next, *last;
ddc97297 407
5ac5c4d6 408 unsigned int nr_spread_over;
ddc97297 409
62160e3f 410#ifdef CONFIG_FAIR_GROUP_SCHED
6aa645ea
IM
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
412
41a2d6cf
IM
413 /*
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
6aa645ea
IM
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
417 *
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
420 */
41a2d6cf
IM
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
c09595f6
PZ
423
424#ifdef CONFIG_SMP
c09595f6 425 /*
c8cba857 426 * the part of load.weight contributed by tasks
c09595f6 427 */
c8cba857 428 unsigned long task_weight;
c09595f6 429
c8cba857
PZ
430 /*
431 * h_load = weight * f(tg)
432 *
433 * Where f(tg) is the recursive weight fraction assigned to
434 * this group.
435 */
436 unsigned long h_load;
c09595f6 437
c8cba857
PZ
438 /*
439 * this cpu's part of tg->shares
440 */
441 unsigned long shares;
f1d239f7
PZ
442
443 /*
444 * load.weight at the time we set shares
445 */
446 unsigned long rq_weight;
c09595f6 447#endif
6aa645ea
IM
448#endif
449};
1da177e4 450
6aa645ea
IM
451/* Real-Time classes' related field in a runqueue: */
452struct rt_rq {
453 struct rt_prio_array active;
63489e45 454 unsigned long rt_nr_running;
052f1dc7 455#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
6f505b16
PZ
456 int highest_prio; /* highest queued rt task prio */
457#endif
fa85ae24 458#ifdef CONFIG_SMP
73fe6aae 459 unsigned long rt_nr_migratory;
a22d7fc1 460 int overloaded;
fa85ae24 461#endif
6f505b16 462 int rt_throttled;
fa85ae24 463 u64 rt_time;
ac086bc2 464 u64 rt_runtime;
ea736ed5 465 /* Nests inside the rq lock: */
ac086bc2 466 spinlock_t rt_runtime_lock;
6f505b16 467
052f1dc7 468#ifdef CONFIG_RT_GROUP_SCHED
23b0fdfc
PZ
469 unsigned long rt_nr_boosted;
470
6f505b16
PZ
471 struct rq *rq;
472 struct list_head leaf_rt_rq_list;
473 struct task_group *tg;
474 struct sched_rt_entity *rt_se;
475#endif
6aa645ea
IM
476};
477
57d885fe
GH
478#ifdef CONFIG_SMP
479
480/*
481 * We add the notion of a root-domain which will be used to define per-domain
0eab9146
IM
482 * variables. Each exclusive cpuset essentially defines an island domain by
483 * fully partitioning the member cpus from any other cpuset. Whenever a new
57d885fe
GH
484 * exclusive cpuset is created, we also create and attach a new root-domain
485 * object.
486 *
57d885fe
GH
487 */
488struct root_domain {
489 atomic_t refcount;
490 cpumask_t span;
491 cpumask_t online;
637f5085 492
0eab9146 493 /*
637f5085
GH
494 * The "RT overload" flag: it gets set if a CPU has more than
495 * one runnable RT task.
496 */
497 cpumask_t rto_mask;
0eab9146 498 atomic_t rto_count;
6e0534f2
GH
499#ifdef CONFIG_SMP
500 struct cpupri cpupri;
501#endif
57d885fe
GH
502};
503
dc938520
GH
504/*
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
507 */
57d885fe
GH
508static struct root_domain def_root_domain;
509
510#endif
511
1da177e4
LT
512/*
513 * This is the main, per-CPU runqueue data structure.
514 *
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
518 */
70b97a7f 519struct rq {
d8016491
IM
520 /* runqueue lock: */
521 spinlock_t lock;
1da177e4
LT
522
523 /*
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
526 */
527 unsigned long nr_running;
6aa645ea
IM
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
bdecea3a 530 unsigned char idle_at_tick;
46cb4b7c 531#ifdef CONFIG_NO_HZ
15934a37 532 unsigned long last_tick_seen;
46cb4b7c
SS
533 unsigned char in_nohz_recently;
534#endif
d8016491
IM
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load;
6aa645ea
IM
537 unsigned long nr_load_updates;
538 u64 nr_switches;
539
540 struct cfs_rq cfs;
6f505b16 541 struct rt_rq rt;
6f505b16 542
6aa645ea 543#ifdef CONFIG_FAIR_GROUP_SCHED
d8016491
IM
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list;
052f1dc7
PZ
546#endif
547#ifdef CONFIG_RT_GROUP_SCHED
6f505b16 548 struct list_head leaf_rt_rq_list;
1da177e4 549#endif
1da177e4
LT
550
551 /*
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
556 */
557 unsigned long nr_uninterruptible;
558
36c8b586 559 struct task_struct *curr, *idle;
c9819f45 560 unsigned long next_balance;
1da177e4 561 struct mm_struct *prev_mm;
6aa645ea 562
3e51f33f 563 u64 clock;
6aa645ea 564
1da177e4
LT
565 atomic_t nr_iowait;
566
567#ifdef CONFIG_SMP
0eab9146 568 struct root_domain *rd;
1da177e4
LT
569 struct sched_domain *sd;
570
571 /* For active balancing */
572 int active_balance;
573 int push_cpu;
d8016491
IM
574 /* cpu of this runqueue: */
575 int cpu;
1f11eb6a 576 int online;
1da177e4 577
a8a51d5e 578 unsigned long avg_load_per_task;
1da177e4 579
36c8b586 580 struct task_struct *migration_thread;
1da177e4
LT
581 struct list_head migration_queue;
582#endif
583
8f4d37ec 584#ifdef CONFIG_SCHED_HRTICK
31656519
PZ
585#ifdef CONFIG_SMP
586 int hrtick_csd_pending;
587 struct call_single_data hrtick_csd;
588#endif
8f4d37ec
PZ
589 struct hrtimer hrtick_timer;
590#endif
591
1da177e4
LT
592#ifdef CONFIG_SCHEDSTATS
593 /* latency stats */
594 struct sched_info rq_sched_info;
595
596 /* sys_sched_yield() stats */
480b9434
KC
597 unsigned int yld_exp_empty;
598 unsigned int yld_act_empty;
599 unsigned int yld_both_empty;
600 unsigned int yld_count;
1da177e4
LT
601
602 /* schedule() stats */
480b9434
KC
603 unsigned int sched_switch;
604 unsigned int sched_count;
605 unsigned int sched_goidle;
1da177e4
LT
606
607 /* try_to_wake_up() stats */
480b9434
KC
608 unsigned int ttwu_count;
609 unsigned int ttwu_local;
b8efb561
IM
610
611 /* BKL stats */
480b9434 612 unsigned int bkl_count;
1da177e4
LT
613#endif
614};
615
f34e3b61 616static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1da177e4 617
15afe09b 618static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
dd41f596 619{
15afe09b 620 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
dd41f596
IM
621}
622
0a2966b4
CL
623static inline int cpu_of(struct rq *rq)
624{
625#ifdef CONFIG_SMP
626 return rq->cpu;
627#else
628 return 0;
629#endif
630}
631
674311d5
NP
632/*
633 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1a20ff27 634 * See detach_destroy_domains: synchronize_sched for details.
674311d5
NP
635 *
636 * The domain tree of any CPU may only be accessed from within
637 * preempt-disabled sections.
638 */
48f24c4d
IM
639#define for_each_domain(cpu, __sd) \
640 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
1da177e4
LT
641
642#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
643#define this_rq() (&__get_cpu_var(runqueues))
644#define task_rq(p) cpu_rq(task_cpu(p))
645#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
646
3e51f33f
PZ
647static inline void update_rq_clock(struct rq *rq)
648{
649 rq->clock = sched_clock_cpu(cpu_of(rq));
650}
651
bf5c91ba
IM
652/*
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
654 */
655#ifdef CONFIG_SCHED_DEBUG
656# define const_debug __read_mostly
657#else
658# define const_debug static const
659#endif
660
017730c1
IM
661/**
662 * runqueue_is_locked
663 *
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
667 */
668int runqueue_is_locked(void)
669{
670 int cpu = get_cpu();
671 struct rq *rq = cpu_rq(cpu);
672 int ret;
673
674 ret = spin_is_locked(&rq->lock);
675 put_cpu();
676 return ret;
677}
678
bf5c91ba
IM
679/*
680 * Debugging: various feature bits
681 */
f00b45c1
PZ
682
683#define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
685
bf5c91ba 686enum {
f00b45c1 687#include "sched_features.h"
bf5c91ba
IM
688};
689
f00b45c1
PZ
690#undef SCHED_FEAT
691
692#define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
694
bf5c91ba 695const_debug unsigned int sysctl_sched_features =
f00b45c1
PZ
696#include "sched_features.h"
697 0;
698
699#undef SCHED_FEAT
700
701#ifdef CONFIG_SCHED_DEBUG
702#define SCHED_FEAT(name, enabled) \
703 #name ,
704
983ed7a6 705static __read_mostly char *sched_feat_names[] = {
f00b45c1
PZ
706#include "sched_features.h"
707 NULL
708};
709
710#undef SCHED_FEAT
711
34f3a814 712static int sched_feat_show(struct seq_file *m, void *v)
f00b45c1 713{
f00b45c1
PZ
714 int i;
715
716 for (i = 0; sched_feat_names[i]; i++) {
34f3a814
LZ
717 if (!(sysctl_sched_features & (1UL << i)))
718 seq_puts(m, "NO_");
719 seq_printf(m, "%s ", sched_feat_names[i]);
f00b45c1 720 }
34f3a814 721 seq_puts(m, "\n");
f00b45c1 722
34f3a814 723 return 0;
f00b45c1
PZ
724}
725
726static ssize_t
727sched_feat_write(struct file *filp, const char __user *ubuf,
728 size_t cnt, loff_t *ppos)
729{
730 char buf[64];
731 char *cmp = buf;
732 int neg = 0;
733 int i;
734
735 if (cnt > 63)
736 cnt = 63;
737
738 if (copy_from_user(&buf, ubuf, cnt))
739 return -EFAULT;
740
741 buf[cnt] = 0;
742
c24b7c52 743 if (strncmp(buf, "NO_", 3) == 0) {
f00b45c1
PZ
744 neg = 1;
745 cmp += 3;
746 }
747
748 for (i = 0; sched_feat_names[i]; i++) {
749 int len = strlen(sched_feat_names[i]);
750
751 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
752 if (neg)
753 sysctl_sched_features &= ~(1UL << i);
754 else
755 sysctl_sched_features |= (1UL << i);
756 break;
757 }
758 }
759
760 if (!sched_feat_names[i])
761 return -EINVAL;
762
763 filp->f_pos += cnt;
764
765 return cnt;
766}
767
34f3a814
LZ
768static int sched_feat_open(struct inode *inode, struct file *filp)
769{
770 return single_open(filp, sched_feat_show, NULL);
771}
772
f00b45c1 773static struct file_operations sched_feat_fops = {
34f3a814
LZ
774 .open = sched_feat_open,
775 .write = sched_feat_write,
776 .read = seq_read,
777 .llseek = seq_lseek,
778 .release = single_release,
f00b45c1
PZ
779};
780
781static __init int sched_init_debug(void)
782{
f00b45c1
PZ
783 debugfs_create_file("sched_features", 0644, NULL, NULL,
784 &sched_feat_fops);
785
786 return 0;
787}
788late_initcall(sched_init_debug);
789
790#endif
791
792#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
bf5c91ba 793
b82d9fdd
PZ
794/*
795 * Number of tasks to iterate in a single balance run.
796 * Limited because this is done with IRQs disabled.
797 */
798const_debug unsigned int sysctl_sched_nr_migrate = 32;
799
2398f2c6
PZ
800/*
801 * ratelimit for updating the group shares.
55cd5340 802 * default: 0.25ms
2398f2c6 803 */
55cd5340 804unsigned int sysctl_sched_shares_ratelimit = 250000;
2398f2c6 805
ffda12a1
PZ
806/*
807 * Inject some fuzzyness into changing the per-cpu group shares
808 * this avoids remote rq-locks at the expense of fairness.
809 * default: 4
810 */
811unsigned int sysctl_sched_shares_thresh = 4;
812
fa85ae24 813/*
9f0c1e56 814 * period over which we measure -rt task cpu usage in us.
fa85ae24
PZ
815 * default: 1s
816 */
9f0c1e56 817unsigned int sysctl_sched_rt_period = 1000000;
fa85ae24 818
6892b75e
IM
819static __read_mostly int scheduler_running;
820
9f0c1e56
PZ
821/*
822 * part of the period that we allow rt tasks to run in us.
823 * default: 0.95s
824 */
825int sysctl_sched_rt_runtime = 950000;
fa85ae24 826
d0b27fa7
PZ
827static inline u64 global_rt_period(void)
828{
829 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
830}
831
832static inline u64 global_rt_runtime(void)
833{
e26873bb 834 if (sysctl_sched_rt_runtime < 0)
d0b27fa7
PZ
835 return RUNTIME_INF;
836
837 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
838}
fa85ae24 839
1da177e4 840#ifndef prepare_arch_switch
4866cde0
NP
841# define prepare_arch_switch(next) do { } while (0)
842#endif
843#ifndef finish_arch_switch
844# define finish_arch_switch(prev) do { } while (0)
845#endif
846
051a1d1a
DA
847static inline int task_current(struct rq *rq, struct task_struct *p)
848{
849 return rq->curr == p;
850}
851
4866cde0 852#ifndef __ARCH_WANT_UNLOCKED_CTXSW
70b97a7f 853static inline int task_running(struct rq *rq, struct task_struct *p)
4866cde0 854{
051a1d1a 855 return task_current(rq, p);
4866cde0
NP
856}
857
70b97a7f 858static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
4866cde0
NP
859{
860}
861
70b97a7f 862static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
4866cde0 863{
da04c035
IM
864#ifdef CONFIG_DEBUG_SPINLOCK
865 /* this is a valid case when another task releases the spinlock */
866 rq->lock.owner = current;
867#endif
8a25d5de
IM
868 /*
869 * If we are tracking spinlock dependencies then we have to
870 * fix up the runqueue lock - which gets 'carried over' from
871 * prev into current:
872 */
873 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
874
4866cde0
NP
875 spin_unlock_irq(&rq->lock);
876}
877
878#else /* __ARCH_WANT_UNLOCKED_CTXSW */
70b97a7f 879static inline int task_running(struct rq *rq, struct task_struct *p)
4866cde0
NP
880{
881#ifdef CONFIG_SMP
882 return p->oncpu;
883#else
051a1d1a 884 return task_current(rq, p);
4866cde0
NP
885#endif
886}
887
70b97a7f 888static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
4866cde0
NP
889{
890#ifdef CONFIG_SMP
891 /*
892 * We can optimise this out completely for !SMP, because the
893 * SMP rebalancing from interrupt is the only thing that cares
894 * here.
895 */
896 next->oncpu = 1;
897#endif
898#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 spin_unlock_irq(&rq->lock);
900#else
901 spin_unlock(&rq->lock);
902#endif
903}
904
70b97a7f 905static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
4866cde0
NP
906{
907#ifdef CONFIG_SMP
908 /*
909 * After ->oncpu is cleared, the task can be moved to a different CPU.
910 * We must ensure this doesn't happen until the switch is completely
911 * finished.
912 */
913 smp_wmb();
914 prev->oncpu = 0;
915#endif
916#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 local_irq_enable();
1da177e4 918#endif
4866cde0
NP
919}
920#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1da177e4 921
b29739f9
IM
922/*
923 * __task_rq_lock - lock the runqueue a given task resides on.
924 * Must be called interrupts disabled.
925 */
70b97a7f 926static inline struct rq *__task_rq_lock(struct task_struct *p)
b29739f9
IM
927 __acquires(rq->lock)
928{
3a5c359a
AK
929 for (;;) {
930 struct rq *rq = task_rq(p);
931 spin_lock(&rq->lock);
932 if (likely(rq == task_rq(p)))
933 return rq;
b29739f9 934 spin_unlock(&rq->lock);
b29739f9 935 }
b29739f9
IM
936}
937
1da177e4
LT
938/*
939 * task_rq_lock - lock the runqueue a given task resides on and disable
41a2d6cf 940 * interrupts. Note the ordering: we can safely lookup the task_rq without
1da177e4
LT
941 * explicitly disabling preemption.
942 */
70b97a7f 943static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1da177e4
LT
944 __acquires(rq->lock)
945{
70b97a7f 946 struct rq *rq;
1da177e4 947
3a5c359a
AK
948 for (;;) {
949 local_irq_save(*flags);
950 rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
953 return rq;
1da177e4 954 spin_unlock_irqrestore(&rq->lock, *flags);
1da177e4 955 }
1da177e4
LT
956}
957
ad474cac
ON
958void task_rq_unlock_wait(struct task_struct *p)
959{
960 struct rq *rq = task_rq(p);
961
962 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
963 spin_unlock_wait(&rq->lock);
964}
965
a9957449 966static void __task_rq_unlock(struct rq *rq)
b29739f9
IM
967 __releases(rq->lock)
968{
969 spin_unlock(&rq->lock);
970}
971
70b97a7f 972static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1da177e4
LT
973 __releases(rq->lock)
974{
975 spin_unlock_irqrestore(&rq->lock, *flags);
976}
977
1da177e4 978/*
cc2a73b5 979 * this_rq_lock - lock this runqueue and disable interrupts.
1da177e4 980 */
a9957449 981static struct rq *this_rq_lock(void)
1da177e4
LT
982 __acquires(rq->lock)
983{
70b97a7f 984 struct rq *rq;
1da177e4
LT
985
986 local_irq_disable();
987 rq = this_rq();
988 spin_lock(&rq->lock);
989
990 return rq;
991}
992
8f4d37ec
PZ
993#ifdef CONFIG_SCHED_HRTICK
994/*
995 * Use HR-timers to deliver accurate preemption points.
996 *
997 * Its all a bit involved since we cannot program an hrt while holding the
998 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
999 * reschedule event.
1000 *
1001 * When we get rescheduled we reprogram the hrtick_timer outside of the
1002 * rq->lock.
1003 */
8f4d37ec
PZ
1004
1005/*
1006 * Use hrtick when:
1007 * - enabled by features
1008 * - hrtimer is actually high res
1009 */
1010static inline int hrtick_enabled(struct rq *rq)
1011{
1012 if (!sched_feat(HRTICK))
1013 return 0;
ba42059f 1014 if (!cpu_active(cpu_of(rq)))
b328ca18 1015 return 0;
8f4d37ec
PZ
1016 return hrtimer_is_hres_active(&rq->hrtick_timer);
1017}
1018
8f4d37ec
PZ
1019static void hrtick_clear(struct rq *rq)
1020{
1021 if (hrtimer_active(&rq->hrtick_timer))
1022 hrtimer_cancel(&rq->hrtick_timer);
1023}
1024
8f4d37ec
PZ
1025/*
1026 * High-resolution timer tick.
1027 * Runs from hardirq context with interrupts disabled.
1028 */
1029static enum hrtimer_restart hrtick(struct hrtimer *timer)
1030{
1031 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1032
1033 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1034
1035 spin_lock(&rq->lock);
3e51f33f 1036 update_rq_clock(rq);
8f4d37ec
PZ
1037 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1038 spin_unlock(&rq->lock);
1039
1040 return HRTIMER_NORESTART;
1041}
1042
95e904c7 1043#ifdef CONFIG_SMP
31656519
PZ
1044/*
1045 * called from hardirq (IPI) context
1046 */
1047static void __hrtick_start(void *arg)
b328ca18 1048{
31656519 1049 struct rq *rq = arg;
b328ca18 1050
31656519
PZ
1051 spin_lock(&rq->lock);
1052 hrtimer_restart(&rq->hrtick_timer);
1053 rq->hrtick_csd_pending = 0;
1054 spin_unlock(&rq->lock);
b328ca18
PZ
1055}
1056
31656519
PZ
1057/*
1058 * Called to set the hrtick timer state.
1059 *
1060 * called with rq->lock held and irqs disabled
1061 */
1062static void hrtick_start(struct rq *rq, u64 delay)
b328ca18 1063{
31656519
PZ
1064 struct hrtimer *timer = &rq->hrtick_timer;
1065 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
b328ca18 1066
cc584b21 1067 hrtimer_set_expires(timer, time);
31656519
PZ
1068
1069 if (rq == this_rq()) {
1070 hrtimer_restart(timer);
1071 } else if (!rq->hrtick_csd_pending) {
1072 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1073 rq->hrtick_csd_pending = 1;
1074 }
b328ca18
PZ
1075}
1076
1077static int
1078hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1079{
1080 int cpu = (int)(long)hcpu;
1081
1082 switch (action) {
1083 case CPU_UP_CANCELED:
1084 case CPU_UP_CANCELED_FROZEN:
1085 case CPU_DOWN_PREPARE:
1086 case CPU_DOWN_PREPARE_FROZEN:
1087 case CPU_DEAD:
1088 case CPU_DEAD_FROZEN:
31656519 1089 hrtick_clear(cpu_rq(cpu));
b328ca18
PZ
1090 return NOTIFY_OK;
1091 }
1092
1093 return NOTIFY_DONE;
1094}
1095
fa748203 1096static __init void init_hrtick(void)
b328ca18
PZ
1097{
1098 hotcpu_notifier(hotplug_hrtick, 0);
1099}
31656519
PZ
1100#else
1101/*
1102 * Called to set the hrtick timer state.
1103 *
1104 * called with rq->lock held and irqs disabled
1105 */
1106static void hrtick_start(struct rq *rq, u64 delay)
1107{
1108 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1109}
b328ca18 1110
006c75f1 1111static inline void init_hrtick(void)
8f4d37ec 1112{
8f4d37ec 1113}
31656519 1114#endif /* CONFIG_SMP */
8f4d37ec 1115
31656519 1116static void init_rq_hrtick(struct rq *rq)
8f4d37ec 1117{
31656519
PZ
1118#ifdef CONFIG_SMP
1119 rq->hrtick_csd_pending = 0;
8f4d37ec 1120
31656519
PZ
1121 rq->hrtick_csd.flags = 0;
1122 rq->hrtick_csd.func = __hrtick_start;
1123 rq->hrtick_csd.info = rq;
1124#endif
8f4d37ec 1125
31656519
PZ
1126 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1127 rq->hrtick_timer.function = hrtick;
ccc7dadf 1128 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
8f4d37ec 1129}
006c75f1 1130#else /* CONFIG_SCHED_HRTICK */
8f4d37ec
PZ
1131static inline void hrtick_clear(struct rq *rq)
1132{
1133}
1134
8f4d37ec
PZ
1135static inline void init_rq_hrtick(struct rq *rq)
1136{
1137}
1138
b328ca18
PZ
1139static inline void init_hrtick(void)
1140{
1141}
006c75f1 1142#endif /* CONFIG_SCHED_HRTICK */
8f4d37ec 1143
c24d20db
IM
1144/*
1145 * resched_task - mark a task 'to be rescheduled now'.
1146 *
1147 * On UP this means the setting of the need_resched flag, on SMP it
1148 * might also involve a cross-CPU call to trigger the scheduler on
1149 * the target CPU.
1150 */
1151#ifdef CONFIG_SMP
1152
1153#ifndef tsk_is_polling
1154#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1155#endif
1156
31656519 1157static void resched_task(struct task_struct *p)
c24d20db
IM
1158{
1159 int cpu;
1160
1161 assert_spin_locked(&task_rq(p)->lock);
1162
31656519 1163 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
c24d20db
IM
1164 return;
1165
31656519 1166 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
c24d20db
IM
1167
1168 cpu = task_cpu(p);
1169 if (cpu == smp_processor_id())
1170 return;
1171
1172 /* NEED_RESCHED must be visible before we test polling */
1173 smp_mb();
1174 if (!tsk_is_polling(p))
1175 smp_send_reschedule(cpu);
1176}
1177
1178static void resched_cpu(int cpu)
1179{
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long flags;
1182
1183 if (!spin_trylock_irqsave(&rq->lock, flags))
1184 return;
1185 resched_task(cpu_curr(cpu));
1186 spin_unlock_irqrestore(&rq->lock, flags);
1187}
06d8308c
TG
1188
1189#ifdef CONFIG_NO_HZ
1190/*
1191 * When add_timer_on() enqueues a timer into the timer wheel of an
1192 * idle CPU then this timer might expire before the next timer event
1193 * which is scheduled to wake up that CPU. In case of a completely
1194 * idle system the next event might even be infinite time into the
1195 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1196 * leaves the inner idle loop so the newly added timer is taken into
1197 * account when the CPU goes back to idle and evaluates the timer
1198 * wheel for the next timer event.
1199 */
1200void wake_up_idle_cpu(int cpu)
1201{
1202 struct rq *rq = cpu_rq(cpu);
1203
1204 if (cpu == smp_processor_id())
1205 return;
1206
1207 /*
1208 * This is safe, as this function is called with the timer
1209 * wheel base lock of (cpu) held. When the CPU is on the way
1210 * to idle and has not yet set rq->curr to idle then it will
1211 * be serialized on the timer wheel base lock and take the new
1212 * timer into account automatically.
1213 */
1214 if (rq->curr != rq->idle)
1215 return;
1216
1217 /*
1218 * We can set TIF_RESCHED on the idle task of the other CPU
1219 * lockless. The worst case is that the other CPU runs the
1220 * idle task through an additional NOOP schedule()
1221 */
1222 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1223
1224 /* NEED_RESCHED must be visible before we test polling */
1225 smp_mb();
1226 if (!tsk_is_polling(rq->idle))
1227 smp_send_reschedule(cpu);
1228}
6d6bc0ad 1229#endif /* CONFIG_NO_HZ */
06d8308c 1230
6d6bc0ad 1231#else /* !CONFIG_SMP */
31656519 1232static void resched_task(struct task_struct *p)
c24d20db
IM
1233{
1234 assert_spin_locked(&task_rq(p)->lock);
31656519 1235 set_tsk_need_resched(p);
c24d20db 1236}
6d6bc0ad 1237#endif /* CONFIG_SMP */
c24d20db 1238
45bf76df
IM
1239#if BITS_PER_LONG == 32
1240# define WMULT_CONST (~0UL)
1241#else
1242# define WMULT_CONST (1UL << 32)
1243#endif
1244
1245#define WMULT_SHIFT 32
1246
194081eb
IM
1247/*
1248 * Shift right and round:
1249 */
cf2ab469 1250#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
194081eb 1251
a7be37ac
PZ
1252/*
1253 * delta *= weight / lw
1254 */
cb1c4fc9 1255static unsigned long
45bf76df
IM
1256calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1257 struct load_weight *lw)
1258{
1259 u64 tmp;
1260
7a232e03
LJ
1261 if (!lw->inv_weight) {
1262 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1263 lw->inv_weight = 1;
1264 else
1265 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1266 / (lw->weight+1);
1267 }
45bf76df
IM
1268
1269 tmp = (u64)delta_exec * weight;
1270 /*
1271 * Check whether we'd overflow the 64-bit multiplication:
1272 */
194081eb 1273 if (unlikely(tmp > WMULT_CONST))
cf2ab469 1274 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
194081eb
IM
1275 WMULT_SHIFT/2);
1276 else
cf2ab469 1277 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
45bf76df 1278
ecf691da 1279 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
45bf76df
IM
1280}
1281
1091985b 1282static inline void update_load_add(struct load_weight *lw, unsigned long inc)
45bf76df
IM
1283{
1284 lw->weight += inc;
e89996ae 1285 lw->inv_weight = 0;
45bf76df
IM
1286}
1287
1091985b 1288static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
45bf76df
IM
1289{
1290 lw->weight -= dec;
e89996ae 1291 lw->inv_weight = 0;
45bf76df
IM
1292}
1293
2dd73a4f
PW
1294/*
1295 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1296 * of tasks with abnormal "nice" values across CPUs the contribution that
1297 * each task makes to its run queue's load is weighted according to its
41a2d6cf 1298 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2dd73a4f
PW
1299 * scaled version of the new time slice allocation that they receive on time
1300 * slice expiry etc.
1301 */
1302
dd41f596
IM
1303#define WEIGHT_IDLEPRIO 2
1304#define WMULT_IDLEPRIO (1 << 31)
1305
1306/*
1307 * Nice levels are multiplicative, with a gentle 10% change for every
1308 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1309 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1310 * that remained on nice 0.
1311 *
1312 * The "10% effect" is relative and cumulative: from _any_ nice level,
1313 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
f9153ee6
IM
1314 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1315 * If a task goes up by ~10% and another task goes down by ~10% then
1316 * the relative distance between them is ~25%.)
dd41f596
IM
1317 */
1318static const int prio_to_weight[40] = {
254753dc
IM
1319 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1320 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1321 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1322 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1323 /* 0 */ 1024, 820, 655, 526, 423,
1324 /* 5 */ 335, 272, 215, 172, 137,
1325 /* 10 */ 110, 87, 70, 56, 45,
1326 /* 15 */ 36, 29, 23, 18, 15,
dd41f596
IM
1327};
1328
5714d2de
IM
1329/*
1330 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1331 *
1332 * In cases where the weight does not change often, we can use the
1333 * precalculated inverse to speed up arithmetics by turning divisions
1334 * into multiplications:
1335 */
dd41f596 1336static const u32 prio_to_wmult[40] = {
254753dc
IM
1337 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1338 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1339 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1340 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1341 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1342 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1343 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1344 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
dd41f596 1345};
2dd73a4f 1346
dd41f596
IM
1347static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1348
1349/*
1350 * runqueue iterator, to support SMP load-balancing between different
1351 * scheduling classes, without having to expose their internal data
1352 * structures to the load-balancing proper:
1353 */
1354struct rq_iterator {
1355 void *arg;
1356 struct task_struct *(*start)(void *);
1357 struct task_struct *(*next)(void *);
1358};
1359
e1d1484f
PW
1360#ifdef CONFIG_SMP
1361static unsigned long
1362balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363 unsigned long max_load_move, struct sched_domain *sd,
1364 enum cpu_idle_type idle, int *all_pinned,
1365 int *this_best_prio, struct rq_iterator *iterator);
1366
1367static int
1368iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1369 struct sched_domain *sd, enum cpu_idle_type idle,
1370 struct rq_iterator *iterator);
e1d1484f 1371#endif
dd41f596 1372
d842de87
SV
1373#ifdef CONFIG_CGROUP_CPUACCT
1374static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1375#else
1376static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1377#endif
1378
18d95a28
PZ
1379static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1380{
1381 update_load_add(&rq->load, load);
1382}
1383
1384static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1385{
1386 update_load_sub(&rq->load, load);
1387}
1388
7940ca36 1389#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
eb755805 1390typedef int (*tg_visitor)(struct task_group *, void *);
c09595f6
PZ
1391
1392/*
1393 * Iterate the full tree, calling @down when first entering a node and @up when
1394 * leaving it for the final time.
1395 */
eb755805 1396static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
c09595f6
PZ
1397{
1398 struct task_group *parent, *child;
eb755805 1399 int ret;
c09595f6
PZ
1400
1401 rcu_read_lock();
1402 parent = &root_task_group;
1403down:
eb755805
PZ
1404 ret = (*down)(parent, data);
1405 if (ret)
1406 goto out_unlock;
c09595f6
PZ
1407 list_for_each_entry_rcu(child, &parent->children, siblings) {
1408 parent = child;
1409 goto down;
1410
1411up:
1412 continue;
1413 }
eb755805
PZ
1414 ret = (*up)(parent, data);
1415 if (ret)
1416 goto out_unlock;
c09595f6
PZ
1417
1418 child = parent;
1419 parent = parent->parent;
1420 if (parent)
1421 goto up;
eb755805 1422out_unlock:
c09595f6 1423 rcu_read_unlock();
eb755805
PZ
1424
1425 return ret;
c09595f6
PZ
1426}
1427
eb755805
PZ
1428static int tg_nop(struct task_group *tg, void *data)
1429{
1430 return 0;
c09595f6 1431}
eb755805
PZ
1432#endif
1433
1434#ifdef CONFIG_SMP
1435static unsigned long source_load(int cpu, int type);
1436static unsigned long target_load(int cpu, int type);
1437static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1438
1439static unsigned long cpu_avg_load_per_task(int cpu)
1440{
1441 struct rq *rq = cpu_rq(cpu);
1442
1443 if (rq->nr_running)
1444 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
a2d47777
BS
1445 else
1446 rq->avg_load_per_task = 0;
eb755805
PZ
1447
1448 return rq->avg_load_per_task;
1449}
1450
1451#ifdef CONFIG_FAIR_GROUP_SCHED
c09595f6 1452
c09595f6
PZ
1453static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1454
1455/*
1456 * Calculate and set the cpu's group shares.
1457 */
1458static void
ffda12a1
PZ
1459update_group_shares_cpu(struct task_group *tg, int cpu,
1460 unsigned long sd_shares, unsigned long sd_rq_weight)
18d95a28 1461{
c09595f6
PZ
1462 unsigned long shares;
1463 unsigned long rq_weight;
1464
c8cba857 1465 if (!tg->se[cpu])
c09595f6
PZ
1466 return;
1467
ec4e0e2f 1468 rq_weight = tg->cfs_rq[cpu]->rq_weight;
c8cba857 1469
c09595f6
PZ
1470 /*
1471 * \Sum shares * rq_weight
1472 * shares = -----------------------
1473 * \Sum rq_weight
1474 *
1475 */
ec4e0e2f 1476 shares = (sd_shares * rq_weight) / sd_rq_weight;
ffda12a1 1477 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
c09595f6 1478
ffda12a1
PZ
1479 if (abs(shares - tg->se[cpu]->load.weight) >
1480 sysctl_sched_shares_thresh) {
1481 struct rq *rq = cpu_rq(cpu);
1482 unsigned long flags;
c09595f6 1483
ffda12a1 1484 spin_lock_irqsave(&rq->lock, flags);
ec4e0e2f 1485 tg->cfs_rq[cpu]->shares = shares;
c09595f6 1486
ffda12a1
PZ
1487 __set_se_shares(tg->se[cpu], shares);
1488 spin_unlock_irqrestore(&rq->lock, flags);
1489 }
18d95a28 1490}
c09595f6
PZ
1491
1492/*
c8cba857
PZ
1493 * Re-compute the task group their per cpu shares over the given domain.
1494 * This needs to be done in a bottom-up fashion because the rq weight of a
1495 * parent group depends on the shares of its child groups.
c09595f6 1496 */
eb755805 1497static int tg_shares_up(struct task_group *tg, void *data)
c09595f6 1498{
ec4e0e2f 1499 unsigned long weight, rq_weight = 0;
c8cba857 1500 unsigned long shares = 0;
eb755805 1501 struct sched_domain *sd = data;
c8cba857 1502 int i;
c09595f6 1503
c8cba857 1504 for_each_cpu_mask(i, sd->span) {
ec4e0e2f
KC
1505 /*
1506 * If there are currently no tasks on the cpu pretend there
1507 * is one of average load so that when a new task gets to
1508 * run here it will not get delayed by group starvation.
1509 */
1510 weight = tg->cfs_rq[i]->load.weight;
1511 if (!weight)
1512 weight = NICE_0_LOAD;
1513
1514 tg->cfs_rq[i]->rq_weight = weight;
1515 rq_weight += weight;
c8cba857 1516 shares += tg->cfs_rq[i]->shares;
c09595f6 1517 }
c09595f6 1518
c8cba857
PZ
1519 if ((!shares && rq_weight) || shares > tg->shares)
1520 shares = tg->shares;
1521
1522 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1523 shares = tg->shares;
c09595f6 1524
ffda12a1
PZ
1525 for_each_cpu_mask(i, sd->span)
1526 update_group_shares_cpu(tg, i, shares, rq_weight);
eb755805
PZ
1527
1528 return 0;
c09595f6
PZ
1529}
1530
1531/*
c8cba857
PZ
1532 * Compute the cpu's hierarchical load factor for each task group.
1533 * This needs to be done in a top-down fashion because the load of a child
1534 * group is a fraction of its parents load.
c09595f6 1535 */
eb755805 1536static int tg_load_down(struct task_group *tg, void *data)
c09595f6 1537{
c8cba857 1538 unsigned long load;
eb755805 1539 long cpu = (long)data;
c09595f6 1540
c8cba857
PZ
1541 if (!tg->parent) {
1542 load = cpu_rq(cpu)->load.weight;
1543 } else {
1544 load = tg->parent->cfs_rq[cpu]->h_load;
1545 load *= tg->cfs_rq[cpu]->shares;
1546 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1547 }
c09595f6 1548
c8cba857 1549 tg->cfs_rq[cpu]->h_load = load;
c09595f6 1550
eb755805 1551 return 0;
c09595f6
PZ
1552}
1553
c8cba857 1554static void update_shares(struct sched_domain *sd)
4d8d595d 1555{
2398f2c6
PZ
1556 u64 now = cpu_clock(raw_smp_processor_id());
1557 s64 elapsed = now - sd->last_update;
1558
1559 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1560 sd->last_update = now;
eb755805 1561 walk_tg_tree(tg_nop, tg_shares_up, sd);
2398f2c6 1562 }
4d8d595d
PZ
1563}
1564
3e5459b4
PZ
1565static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1566{
1567 spin_unlock(&rq->lock);
1568 update_shares(sd);
1569 spin_lock(&rq->lock);
1570}
1571
eb755805 1572static void update_h_load(long cpu)
c09595f6 1573{
eb755805 1574 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
c09595f6
PZ
1575}
1576
c09595f6
PZ
1577#else
1578
c8cba857 1579static inline void update_shares(struct sched_domain *sd)
4d8d595d
PZ
1580{
1581}
1582
3e5459b4
PZ
1583static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1584{
1585}
1586
18d95a28
PZ
1587#endif
1588
18d95a28
PZ
1589#endif
1590
30432094 1591#ifdef CONFIG_FAIR_GROUP_SCHED
34e83e85
IM
1592static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1593{
30432094 1594#ifdef CONFIG_SMP
34e83e85
IM
1595 cfs_rq->shares = shares;
1596#endif
1597}
30432094 1598#endif
e7693a36 1599
dd41f596 1600#include "sched_stats.h"
dd41f596 1601#include "sched_idletask.c"
5522d5d5
IM
1602#include "sched_fair.c"
1603#include "sched_rt.c"
dd41f596
IM
1604#ifdef CONFIG_SCHED_DEBUG
1605# include "sched_debug.c"
1606#endif
1607
1608#define sched_class_highest (&rt_sched_class)
1f11eb6a
GH
1609#define for_each_class(class) \
1610 for (class = sched_class_highest; class; class = class->next)
dd41f596 1611
c09595f6 1612static void inc_nr_running(struct rq *rq)
9c217245
IM
1613{
1614 rq->nr_running++;
9c217245
IM
1615}
1616
c09595f6 1617static void dec_nr_running(struct rq *rq)
9c217245
IM
1618{
1619 rq->nr_running--;
9c217245
IM
1620}
1621
45bf76df
IM
1622static void set_load_weight(struct task_struct *p)
1623{
1624 if (task_has_rt_policy(p)) {
dd41f596
IM
1625 p->se.load.weight = prio_to_weight[0] * 2;
1626 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1627 return;
1628 }
45bf76df 1629
dd41f596
IM
1630 /*
1631 * SCHED_IDLE tasks get minimal weight:
1632 */
1633 if (p->policy == SCHED_IDLE) {
1634 p->se.load.weight = WEIGHT_IDLEPRIO;
1635 p->se.load.inv_weight = WMULT_IDLEPRIO;
1636 return;
1637 }
71f8bd46 1638
dd41f596
IM
1639 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1640 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
71f8bd46
IM
1641}
1642
2087a1ad
GH
1643static void update_avg(u64 *avg, u64 sample)
1644{
1645 s64 diff = sample - *avg;
1646 *avg += diff >> 3;
1647}
1648
8159f87e 1649static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
71f8bd46 1650{
dd41f596 1651 sched_info_queued(p);
fd390f6a 1652 p->sched_class->enqueue_task(rq, p, wakeup);
dd41f596 1653 p->se.on_rq = 1;
71f8bd46
IM
1654}
1655
69be72c1 1656static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
71f8bd46 1657{
2087a1ad
GH
1658 if (sleep && p->se.last_wakeup) {
1659 update_avg(&p->se.avg_overlap,
1660 p->se.sum_exec_runtime - p->se.last_wakeup);
1661 p->se.last_wakeup = 0;
1662 }
1663
46ac22ba 1664 sched_info_dequeued(p);
f02231e5 1665 p->sched_class->dequeue_task(rq, p, sleep);
dd41f596 1666 p->se.on_rq = 0;
71f8bd46
IM
1667}
1668
14531189 1669/*
dd41f596 1670 * __normal_prio - return the priority that is based on the static prio
14531189 1671 */
14531189
IM
1672static inline int __normal_prio(struct task_struct *p)
1673{
dd41f596 1674 return p->static_prio;
14531189
IM
1675}
1676
b29739f9
IM
1677/*
1678 * Calculate the expected normal priority: i.e. priority
1679 * without taking RT-inheritance into account. Might be
1680 * boosted by interactivity modifiers. Changes upon fork,
1681 * setprio syscalls, and whenever the interactivity
1682 * estimator recalculates.
1683 */
36c8b586 1684static inline int normal_prio(struct task_struct *p)
b29739f9
IM
1685{
1686 int prio;
1687
e05606d3 1688 if (task_has_rt_policy(p))
b29739f9
IM
1689 prio = MAX_RT_PRIO-1 - p->rt_priority;
1690 else
1691 prio = __normal_prio(p);
1692 return prio;
1693}
1694
1695/*
1696 * Calculate the current priority, i.e. the priority
1697 * taken into account by the scheduler. This value might
1698 * be boosted by RT tasks, or might be boosted by
1699 * interactivity modifiers. Will be RT if the task got
1700 * RT-boosted. If not then it returns p->normal_prio.
1701 */
36c8b586 1702static int effective_prio(struct task_struct *p)
b29739f9
IM
1703{
1704 p->normal_prio = normal_prio(p);
1705 /*
1706 * If we are RT tasks or we were boosted to RT priority,
1707 * keep the priority unchanged. Otherwise, update priority
1708 * to the normal priority:
1709 */
1710 if (!rt_prio(p->prio))
1711 return p->normal_prio;
1712 return p->prio;
1713}
1714
1da177e4 1715/*
dd41f596 1716 * activate_task - move a task to the runqueue.
1da177e4 1717 */
dd41f596 1718static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1da177e4 1719{
d9514f6c 1720 if (task_contributes_to_load(p))
dd41f596 1721 rq->nr_uninterruptible--;
1da177e4 1722
8159f87e 1723 enqueue_task(rq, p, wakeup);
c09595f6 1724 inc_nr_running(rq);
1da177e4
LT
1725}
1726
1da177e4
LT
1727/*
1728 * deactivate_task - remove a task from the runqueue.
1729 */
2e1cb74a 1730static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1da177e4 1731{
d9514f6c 1732 if (task_contributes_to_load(p))
dd41f596
IM
1733 rq->nr_uninterruptible++;
1734
69be72c1 1735 dequeue_task(rq, p, sleep);
c09595f6 1736 dec_nr_running(rq);
1da177e4
LT
1737}
1738
1da177e4
LT
1739/**
1740 * task_curr - is this task currently executing on a CPU?
1741 * @p: the task in question.
1742 */
36c8b586 1743inline int task_curr(const struct task_struct *p)
1da177e4
LT
1744{
1745 return cpu_curr(task_cpu(p)) == p;
1746}
1747
dd41f596
IM
1748static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1749{
6f505b16 1750 set_task_rq(p, cpu);
dd41f596 1751#ifdef CONFIG_SMP
ce96b5ac
DA
1752 /*
1753 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1754 * successfuly executed on another CPU. We must ensure that updates of
1755 * per-task data have been completed by this moment.
1756 */
1757 smp_wmb();
dd41f596 1758 task_thread_info(p)->cpu = cpu;
dd41f596 1759#endif
2dd73a4f
PW
1760}
1761
cb469845
SR
1762static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1763 const struct sched_class *prev_class,
1764 int oldprio, int running)
1765{
1766 if (prev_class != p->sched_class) {
1767 if (prev_class->switched_from)
1768 prev_class->switched_from(rq, p, running);
1769 p->sched_class->switched_to(rq, p, running);
1770 } else
1771 p->sched_class->prio_changed(rq, p, oldprio, running);
1772}
1773
1da177e4 1774#ifdef CONFIG_SMP
c65cc870 1775
e958b360
TG
1776/* Used instead of source_load when we know the type == 0 */
1777static unsigned long weighted_cpuload(const int cpu)
1778{
1779 return cpu_rq(cpu)->load.weight;
1780}
1781
cc367732
IM
1782/*
1783 * Is this task likely cache-hot:
1784 */
e7693a36 1785static int
cc367732
IM
1786task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1787{
1788 s64 delta;
1789
f540a608
IM
1790 /*
1791 * Buddy candidates are cache hot:
1792 */
4793241b
PZ
1793 if (sched_feat(CACHE_HOT_BUDDY) &&
1794 (&p->se == cfs_rq_of(&p->se)->next ||
1795 &p->se == cfs_rq_of(&p->se)->last))
f540a608
IM
1796 return 1;
1797
cc367732
IM
1798 if (p->sched_class != &fair_sched_class)
1799 return 0;
1800
6bc1665b
IM
1801 if (sysctl_sched_migration_cost == -1)
1802 return 1;
1803 if (sysctl_sched_migration_cost == 0)
1804 return 0;
1805
cc367732
IM
1806 delta = now - p->se.exec_start;
1807
1808 return delta < (s64)sysctl_sched_migration_cost;
1809}
1810
1811
dd41f596 1812void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
c65cc870 1813{
dd41f596
IM
1814 int old_cpu = task_cpu(p);
1815 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2830cf8c
SV
1816 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1817 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
bbdba7c0 1818 u64 clock_offset;
dd41f596
IM
1819
1820 clock_offset = old_rq->clock - new_rq->clock;
6cfb0d5d
IM
1821
1822#ifdef CONFIG_SCHEDSTATS
1823 if (p->se.wait_start)
1824 p->se.wait_start -= clock_offset;
dd41f596
IM
1825 if (p->se.sleep_start)
1826 p->se.sleep_start -= clock_offset;
1827 if (p->se.block_start)
1828 p->se.block_start -= clock_offset;
cc367732
IM
1829 if (old_cpu != new_cpu) {
1830 schedstat_inc(p, se.nr_migrations);
1831 if (task_hot(p, old_rq->clock, NULL))
1832 schedstat_inc(p, se.nr_forced2_migrations);
1833 }
6cfb0d5d 1834#endif
2830cf8c
SV
1835 p->se.vruntime -= old_cfsrq->min_vruntime -
1836 new_cfsrq->min_vruntime;
dd41f596
IM
1837
1838 __set_task_cpu(p, new_cpu);
c65cc870
IM
1839}
1840
70b97a7f 1841struct migration_req {
1da177e4 1842 struct list_head list;
1da177e4 1843
36c8b586 1844 struct task_struct *task;
1da177e4
LT
1845 int dest_cpu;
1846
1da177e4 1847 struct completion done;
70b97a7f 1848};
1da177e4
LT
1849
1850/*
1851 * The task's runqueue lock must be held.
1852 * Returns true if you have to wait for migration thread.
1853 */
36c8b586 1854static int
70b97a7f 1855migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1da177e4 1856{
70b97a7f 1857 struct rq *rq = task_rq(p);
1da177e4
LT
1858
1859 /*
1860 * If the task is not on a runqueue (and not running), then
1861 * it is sufficient to simply update the task's cpu field.
1862 */
dd41f596 1863 if (!p->se.on_rq && !task_running(rq, p)) {
1da177e4
LT
1864 set_task_cpu(p, dest_cpu);
1865 return 0;
1866 }
1867
1868 init_completion(&req->done);
1da177e4
LT
1869 req->task = p;
1870 req->dest_cpu = dest_cpu;
1871 list_add(&req->list, &rq->migration_queue);
48f24c4d 1872
1da177e4
LT
1873 return 1;
1874}
1875
1876/*
1877 * wait_task_inactive - wait for a thread to unschedule.
1878 *
85ba2d86
RM
1879 * If @match_state is nonzero, it's the @p->state value just checked and
1880 * not expected to change. If it changes, i.e. @p might have woken up,
1881 * then return zero. When we succeed in waiting for @p to be off its CPU,
1882 * we return a positive number (its total switch count). If a second call
1883 * a short while later returns the same number, the caller can be sure that
1884 * @p has remained unscheduled the whole time.
1885 *
1da177e4
LT
1886 * The caller must ensure that the task *will* unschedule sometime soon,
1887 * else this function might spin for a *long* time. This function can't
1888 * be called with interrupts off, or it may introduce deadlock with
1889 * smp_call_function() if an IPI is sent by the same process we are
1890 * waiting to become inactive.
1891 */
85ba2d86 1892unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1da177e4
LT
1893{
1894 unsigned long flags;
dd41f596 1895 int running, on_rq;
85ba2d86 1896 unsigned long ncsw;
70b97a7f 1897 struct rq *rq;
1da177e4 1898
3a5c359a
AK
1899 for (;;) {
1900 /*
1901 * We do the initial early heuristics without holding
1902 * any task-queue locks at all. We'll only try to get
1903 * the runqueue lock when things look like they will
1904 * work out!
1905 */
1906 rq = task_rq(p);
fa490cfd 1907
3a5c359a
AK
1908 /*
1909 * If the task is actively running on another CPU
1910 * still, just relax and busy-wait without holding
1911 * any locks.
1912 *
1913 * NOTE! Since we don't hold any locks, it's not
1914 * even sure that "rq" stays as the right runqueue!
1915 * But we don't care, since "task_running()" will
1916 * return false if the runqueue has changed and p
1917 * is actually now running somewhere else!
1918 */
85ba2d86
RM
1919 while (task_running(rq, p)) {
1920 if (match_state && unlikely(p->state != match_state))
1921 return 0;
3a5c359a 1922 cpu_relax();
85ba2d86 1923 }
fa490cfd 1924
3a5c359a
AK
1925 /*
1926 * Ok, time to look more closely! We need the rq
1927 * lock now, to be *sure*. If we're wrong, we'll
1928 * just go back and repeat.
1929 */
1930 rq = task_rq_lock(p, &flags);
0a16b607 1931 trace_sched_wait_task(rq, p);
3a5c359a
AK
1932 running = task_running(rq, p);
1933 on_rq = p->se.on_rq;
85ba2d86 1934 ncsw = 0;
f31e11d8 1935 if (!match_state || p->state == match_state)
93dcf55f 1936 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
3a5c359a 1937 task_rq_unlock(rq, &flags);
fa490cfd 1938
85ba2d86
RM
1939 /*
1940 * If it changed from the expected state, bail out now.
1941 */
1942 if (unlikely(!ncsw))
1943 break;
1944
3a5c359a
AK
1945 /*
1946 * Was it really running after all now that we
1947 * checked with the proper locks actually held?
1948 *
1949 * Oops. Go back and try again..
1950 */
1951 if (unlikely(running)) {
1952 cpu_relax();
1953 continue;
1954 }
fa490cfd 1955
3a5c359a
AK
1956 /*
1957 * It's not enough that it's not actively running,
1958 * it must be off the runqueue _entirely_, and not
1959 * preempted!
1960 *
1961 * So if it wa still runnable (but just not actively
1962 * running right now), it's preempted, and we should
1963 * yield - it could be a while.
1964 */
1965 if (unlikely(on_rq)) {
1966 schedule_timeout_uninterruptible(1);
1967 continue;
1968 }
fa490cfd 1969
3a5c359a
AK
1970 /*
1971 * Ahh, all good. It wasn't running, and it wasn't
1972 * runnable, which means that it will never become
1973 * running in the future either. We're all done!
1974 */
1975 break;
1976 }
85ba2d86
RM
1977
1978 return ncsw;
1da177e4
LT
1979}
1980
1981/***
1982 * kick_process - kick a running thread to enter/exit the kernel
1983 * @p: the to-be-kicked thread
1984 *
1985 * Cause a process which is running on another CPU to enter
1986 * kernel-mode, without any delay. (to get signals handled.)
1987 *
1988 * NOTE: this function doesnt have to take the runqueue lock,
1989 * because all it wants to ensure is that the remote task enters
1990 * the kernel. If the IPI races and the task has been migrated
1991 * to another CPU then no harm is done and the purpose has been
1992 * achieved as well.
1993 */
36c8b586 1994void kick_process(struct task_struct *p)
1da177e4
LT
1995{
1996 int cpu;
1997
1998 preempt_disable();
1999 cpu = task_cpu(p);
2000 if ((cpu != smp_processor_id()) && task_curr(p))
2001 smp_send_reschedule(cpu);
2002 preempt_enable();
2003}
2004
2005/*
2dd73a4f
PW
2006 * Return a low guess at the load of a migration-source cpu weighted
2007 * according to the scheduling class and "nice" value.
1da177e4
LT
2008 *
2009 * We want to under-estimate the load of migration sources, to
2010 * balance conservatively.
2011 */
a9957449 2012static unsigned long source_load(int cpu, int type)
1da177e4 2013{
70b97a7f 2014 struct rq *rq = cpu_rq(cpu);
dd41f596 2015 unsigned long total = weighted_cpuload(cpu);
2dd73a4f 2016
93b75217 2017 if (type == 0 || !sched_feat(LB_BIAS))
dd41f596 2018 return total;
b910472d 2019
dd41f596 2020 return min(rq->cpu_load[type-1], total);
1da177e4
LT
2021}
2022
2023/*
2dd73a4f
PW
2024 * Return a high guess at the load of a migration-target cpu weighted
2025 * according to the scheduling class and "nice" value.
1da177e4 2026 */
a9957449 2027static unsigned long target_load(int cpu, int type)
1da177e4 2028{
70b97a7f 2029 struct rq *rq = cpu_rq(cpu);
dd41f596 2030 unsigned long total = weighted_cpuload(cpu);
2dd73a4f 2031
93b75217 2032 if (type == 0 || !sched_feat(LB_BIAS))
dd41f596 2033 return total;
3b0bd9bc 2034
dd41f596 2035 return max(rq->cpu_load[type-1], total);
2dd73a4f
PW
2036}
2037
147cbb4b
NP
2038/*
2039 * find_idlest_group finds and returns the least busy CPU group within the
2040 * domain.
2041 */
2042static struct sched_group *
2043find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2044{
2045 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2046 unsigned long min_load = ULONG_MAX, this_load = 0;
2047 int load_idx = sd->forkexec_idx;
2048 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2049
2050 do {
2051 unsigned long load, avg_load;
2052 int local_group;
2053 int i;
2054
da5a5522
BD
2055 /* Skip over this group if it has no CPUs allowed */
2056 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
3a5c359a 2057 continue;
da5a5522 2058
147cbb4b 2059 local_group = cpu_isset(this_cpu, group->cpumask);
147cbb4b
NP
2060
2061 /* Tally up the load of all CPUs in the group */
2062 avg_load = 0;
2063
363ab6f1 2064 for_each_cpu_mask_nr(i, group->cpumask) {
147cbb4b
NP
2065 /* Bias balancing toward cpus of our domain */
2066 if (local_group)
2067 load = source_load(i, load_idx);
2068 else
2069 load = target_load(i, load_idx);
2070
2071 avg_load += load;
2072 }
2073
2074 /* Adjust by relative CPU power of the group */
5517d86b
ED
2075 avg_load = sg_div_cpu_power(group,
2076 avg_load * SCHED_LOAD_SCALE);
147cbb4b
NP
2077
2078 if (local_group) {
2079 this_load = avg_load;
2080 this = group;
2081 } else if (avg_load < min_load) {
2082 min_load = avg_load;
2083 idlest = group;
2084 }
3a5c359a 2085 } while (group = group->next, group != sd->groups);
147cbb4b
NP
2086
2087 if (!idlest || 100*this_load < imbalance*min_load)
2088 return NULL;
2089 return idlest;
2090}
2091
2092/*
0feaece9 2093 * find_idlest_cpu - find the idlest cpu among the cpus in group.
147cbb4b 2094 */
95cdf3b7 2095static int
7c16ec58
MT
2096find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2097 cpumask_t *tmp)
147cbb4b
NP
2098{
2099 unsigned long load, min_load = ULONG_MAX;
2100 int idlest = -1;
2101 int i;
2102
da5a5522 2103 /* Traverse only the allowed CPUs */
7c16ec58 2104 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
da5a5522 2105
363ab6f1 2106 for_each_cpu_mask_nr(i, *tmp) {
2dd73a4f 2107 load = weighted_cpuload(i);
147cbb4b
NP
2108
2109 if (load < min_load || (load == min_load && i == this_cpu)) {
2110 min_load = load;
2111 idlest = i;
2112 }
2113 }
2114
2115 return idlest;
2116}
2117
476d139c
NP
2118/*
2119 * sched_balance_self: balance the current task (running on cpu) in domains
2120 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2121 * SD_BALANCE_EXEC.
2122 *
2123 * Balance, ie. select the least loaded group.
2124 *
2125 * Returns the target CPU number, or the same CPU if no balancing is needed.
2126 *
2127 * preempt must be disabled.
2128 */
2129static int sched_balance_self(int cpu, int flag)
2130{
2131 struct task_struct *t = current;
2132 struct sched_domain *tmp, *sd = NULL;
147cbb4b 2133
c96d145e 2134 for_each_domain(cpu, tmp) {
9761eea8
IM
2135 /*
2136 * If power savings logic is enabled for a domain, stop there.
2137 */
5c45bf27
SS
2138 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2139 break;
476d139c
NP
2140 if (tmp->flags & flag)
2141 sd = tmp;
c96d145e 2142 }
476d139c 2143
039a1c41
PZ
2144 if (sd)
2145 update_shares(sd);
2146
476d139c 2147 while (sd) {
7c16ec58 2148 cpumask_t span, tmpmask;
476d139c 2149 struct sched_group *group;
1a848870
SS
2150 int new_cpu, weight;
2151
2152 if (!(sd->flags & flag)) {
2153 sd = sd->child;
2154 continue;
2155 }
476d139c
NP
2156
2157 span = sd->span;
2158 group = find_idlest_group(sd, t, cpu);
1a848870
SS
2159 if (!group) {
2160 sd = sd->child;
2161 continue;
2162 }
476d139c 2163
7c16ec58 2164 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
1a848870
SS
2165 if (new_cpu == -1 || new_cpu == cpu) {
2166 /* Now try balancing at a lower domain level of cpu */
2167 sd = sd->child;
2168 continue;
2169 }
476d139c 2170
1a848870 2171 /* Now try balancing at a lower domain level of new_cpu */
476d139c 2172 cpu = new_cpu;
476d139c
NP
2173 sd = NULL;
2174 weight = cpus_weight(span);
2175 for_each_domain(cpu, tmp) {
2176 if (weight <= cpus_weight(tmp->span))
2177 break;
2178 if (tmp->flags & flag)
2179 sd = tmp;
2180 }
2181 /* while loop will break here if sd == NULL */
2182 }
2183
2184 return cpu;
2185}
2186
2187#endif /* CONFIG_SMP */
1da177e4 2188
1da177e4
LT
2189/***
2190 * try_to_wake_up - wake up a thread
2191 * @p: the to-be-woken-up thread
2192 * @state: the mask of task states that can be woken
2193 * @sync: do a synchronous wakeup?
2194 *
2195 * Put it on the run-queue if it's not already there. The "current"
2196 * thread is always on the run-queue (except when the actual
2197 * re-schedule is in progress), and as such you're allowed to do
2198 * the simpler "current->state = TASK_RUNNING" to mark yourself
2199 * runnable without the overhead of this.
2200 *
2201 * returns failure only if the task is already active.
2202 */
36c8b586 2203static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1da177e4 2204{
cc367732 2205 int cpu, orig_cpu, this_cpu, success = 0;
1da177e4
LT
2206 unsigned long flags;
2207 long old_state;
70b97a7f 2208 struct rq *rq;
1da177e4 2209
b85d0667
IM
2210 if (!sched_feat(SYNC_WAKEUPS))
2211 sync = 0;
2212
2398f2c6
PZ
2213#ifdef CONFIG_SMP
2214 if (sched_feat(LB_WAKEUP_UPDATE)) {
2215 struct sched_domain *sd;
2216
2217 this_cpu = raw_smp_processor_id();
2218 cpu = task_cpu(p);
2219
2220 for_each_domain(this_cpu, sd) {
2221 if (cpu_isset(cpu, sd->span)) {
2222 update_shares(sd);
2223 break;
2224 }
2225 }
2226 }
2227#endif
2228
04e2f174 2229 smp_wmb();
1da177e4
LT
2230 rq = task_rq_lock(p, &flags);
2231 old_state = p->state;
2232 if (!(old_state & state))
2233 goto out;
2234
dd41f596 2235 if (p->se.on_rq)
1da177e4
LT
2236 goto out_running;
2237
2238 cpu = task_cpu(p);
cc367732 2239 orig_cpu = cpu;
1da177e4
LT
2240 this_cpu = smp_processor_id();
2241
2242#ifdef CONFIG_SMP
2243 if (unlikely(task_running(rq, p)))
2244 goto out_activate;
2245
5d2f5a61
DA
2246 cpu = p->sched_class->select_task_rq(p, sync);
2247 if (cpu != orig_cpu) {
2248 set_task_cpu(p, cpu);
1da177e4
LT
2249 task_rq_unlock(rq, &flags);
2250 /* might preempt at this point */
2251 rq = task_rq_lock(p, &flags);
2252 old_state = p->state;
2253 if (!(old_state & state))
2254 goto out;
dd41f596 2255 if (p->se.on_rq)
1da177e4
LT
2256 goto out_running;
2257
2258 this_cpu = smp_processor_id();
2259 cpu = task_cpu(p);
2260 }
2261
e7693a36
GH
2262#ifdef CONFIG_SCHEDSTATS
2263 schedstat_inc(rq, ttwu_count);
2264 if (cpu == this_cpu)
2265 schedstat_inc(rq, ttwu_local);
2266 else {
2267 struct sched_domain *sd;
2268 for_each_domain(this_cpu, sd) {
2269 if (cpu_isset(cpu, sd->span)) {
2270 schedstat_inc(sd, ttwu_wake_remote);
2271 break;
2272 }
2273 }
2274 }
6d6bc0ad 2275#endif /* CONFIG_SCHEDSTATS */
e7693a36 2276
1da177e4
LT
2277out_activate:
2278#endif /* CONFIG_SMP */
cc367732
IM
2279 schedstat_inc(p, se.nr_wakeups);
2280 if (sync)
2281 schedstat_inc(p, se.nr_wakeups_sync);
2282 if (orig_cpu != cpu)
2283 schedstat_inc(p, se.nr_wakeups_migrate);
2284 if (cpu == this_cpu)
2285 schedstat_inc(p, se.nr_wakeups_local);
2286 else
2287 schedstat_inc(p, se.nr_wakeups_remote);
2daa3577 2288 update_rq_clock(rq);
dd41f596 2289 activate_task(rq, p, 1);
1da177e4
LT
2290 success = 1;
2291
2292out_running:
0a16b607 2293 trace_sched_wakeup(rq, p);
15afe09b 2294 check_preempt_curr(rq, p, sync);
4ae7d5ce 2295
1da177e4 2296 p->state = TASK_RUNNING;
9a897c5a
SR
2297#ifdef CONFIG_SMP
2298 if (p->sched_class->task_wake_up)
2299 p->sched_class->task_wake_up(rq, p);
2300#endif
1da177e4 2301out:
2087a1ad
GH
2302 current->se.last_wakeup = current->se.sum_exec_runtime;
2303
1da177e4
LT
2304 task_rq_unlock(rq, &flags);
2305
2306 return success;
2307}
2308
7ad5b3a5 2309int wake_up_process(struct task_struct *p)
1da177e4 2310{
d9514f6c 2311 return try_to_wake_up(p, TASK_ALL, 0);
1da177e4 2312}
1da177e4
LT
2313EXPORT_SYMBOL(wake_up_process);
2314
7ad5b3a5 2315int wake_up_state(struct task_struct *p, unsigned int state)
1da177e4
LT
2316{
2317 return try_to_wake_up(p, state, 0);
2318}
2319
1da177e4
LT
2320/*
2321 * Perform scheduler related setup for a newly forked process p.
2322 * p is forked by current.
dd41f596
IM
2323 *
2324 * __sched_fork() is basic setup used by init_idle() too:
2325 */
2326static void __sched_fork(struct task_struct *p)
2327{
dd41f596
IM
2328 p->se.exec_start = 0;
2329 p->se.sum_exec_runtime = 0;
f6cf891c 2330 p->se.prev_sum_exec_runtime = 0;
4ae7d5ce
IM
2331 p->se.last_wakeup = 0;
2332 p->se.avg_overlap = 0;
6cfb0d5d
IM
2333
2334#ifdef CONFIG_SCHEDSTATS
2335 p->se.wait_start = 0;
dd41f596
IM
2336 p->se.sum_sleep_runtime = 0;
2337 p->se.sleep_start = 0;
dd41f596
IM
2338 p->se.block_start = 0;
2339 p->se.sleep_max = 0;
2340 p->se.block_max = 0;
2341 p->se.exec_max = 0;
eba1ed4b 2342 p->se.slice_max = 0;
dd41f596 2343 p->se.wait_max = 0;
6cfb0d5d 2344#endif
476d139c 2345
fa717060 2346 INIT_LIST_HEAD(&p->rt.run_list);
dd41f596 2347 p->se.on_rq = 0;
4a55bd5e 2348 INIT_LIST_HEAD(&p->se.group_node);
476d139c 2349
e107be36
AK
2350#ifdef CONFIG_PREEMPT_NOTIFIERS
2351 INIT_HLIST_HEAD(&p->preempt_notifiers);
2352#endif
2353
1da177e4
LT
2354 /*
2355 * We mark the process as running here, but have not actually
2356 * inserted it onto the runqueue yet. This guarantees that
2357 * nobody will actually run it, and a signal or other external
2358 * event cannot wake it up and insert it on the runqueue either.
2359 */
2360 p->state = TASK_RUNNING;
dd41f596
IM
2361}
2362
2363/*
2364 * fork()/clone()-time setup:
2365 */
2366void sched_fork(struct task_struct *p, int clone_flags)
2367{
2368 int cpu = get_cpu();
2369
2370 __sched_fork(p);
2371
2372#ifdef CONFIG_SMP
2373 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2374#endif
02e4bac2 2375 set_task_cpu(p, cpu);
b29739f9
IM
2376
2377 /*
2378 * Make sure we do not leak PI boosting priority to the child:
2379 */
2380 p->prio = current->normal_prio;
2ddbf952
HS
2381 if (!rt_prio(p->prio))
2382 p->sched_class = &fair_sched_class;
b29739f9 2383
52f17b6c 2384#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
dd41f596 2385 if (likely(sched_info_on()))
52f17b6c 2386 memset(&p->sched_info, 0, sizeof(p->sched_info));
1da177e4 2387#endif
d6077cb8 2388#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4866cde0
NP
2389 p->oncpu = 0;
2390#endif
1da177e4 2391#ifdef CONFIG_PREEMPT
4866cde0 2392 /* Want to start with kernel preemption disabled. */
a1261f54 2393 task_thread_info(p)->preempt_count = 1;
1da177e4 2394#endif
476d139c 2395 put_cpu();
1da177e4
LT
2396}
2397
2398/*
2399 * wake_up_new_task - wake up a newly created task for the first time.
2400 *
2401 * This function will do some initial scheduler statistics housekeeping
2402 * that must be done for every newly created context, then puts the task
2403 * on the runqueue and wakes it.
2404 */
7ad5b3a5 2405void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1da177e4
LT
2406{
2407 unsigned long flags;
dd41f596 2408 struct rq *rq;
1da177e4
LT
2409
2410 rq = task_rq_lock(p, &flags);
147cbb4b 2411 BUG_ON(p->state != TASK_RUNNING);
a8e504d2 2412 update_rq_clock(rq);
1da177e4
LT
2413
2414 p->prio = effective_prio(p);
2415
b9dca1e0 2416 if (!p->sched_class->task_new || !current->se.on_rq) {
dd41f596 2417 activate_task(rq, p, 0);
1da177e4 2418 } else {
1da177e4 2419 /*
dd41f596
IM
2420 * Let the scheduling class do new task startup
2421 * management (if any):
1da177e4 2422 */
ee0827d8 2423 p->sched_class->task_new(rq, p);
c09595f6 2424 inc_nr_running(rq);
1da177e4 2425 }
0a16b607 2426 trace_sched_wakeup_new(rq, p);
15afe09b 2427 check_preempt_curr(rq, p, 0);
9a897c5a
SR
2428#ifdef CONFIG_SMP
2429 if (p->sched_class->task_wake_up)
2430 p->sched_class->task_wake_up(rq, p);
2431#endif
dd41f596 2432 task_rq_unlock(rq, &flags);
1da177e4
LT
2433}
2434
e107be36
AK
2435#ifdef CONFIG_PREEMPT_NOTIFIERS
2436
2437/**
421cee29
RD
2438 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2439 * @notifier: notifier struct to register
e107be36
AK
2440 */
2441void preempt_notifier_register(struct preempt_notifier *notifier)
2442{
2443 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2444}
2445EXPORT_SYMBOL_GPL(preempt_notifier_register);
2446
2447/**
2448 * preempt_notifier_unregister - no longer interested in preemption notifications
421cee29 2449 * @notifier: notifier struct to unregister
e107be36
AK
2450 *
2451 * This is safe to call from within a preemption notifier.
2452 */
2453void preempt_notifier_unregister(struct preempt_notifier *notifier)
2454{
2455 hlist_del(&notifier->link);
2456}
2457EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2458
2459static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2460{
2461 struct preempt_notifier *notifier;
2462 struct hlist_node *node;
2463
2464 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2465 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2466}
2467
2468static void
2469fire_sched_out_preempt_notifiers(struct task_struct *curr,
2470 struct task_struct *next)
2471{
2472 struct preempt_notifier *notifier;
2473 struct hlist_node *node;
2474
2475 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2476 notifier->ops->sched_out(notifier, next);
2477}
2478
6d6bc0ad 2479#else /* !CONFIG_PREEMPT_NOTIFIERS */
e107be36
AK
2480
2481static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2482{
2483}
2484
2485static void
2486fire_sched_out_preempt_notifiers(struct task_struct *curr,
2487 struct task_struct *next)
2488{
2489}
2490
6d6bc0ad 2491#endif /* CONFIG_PREEMPT_NOTIFIERS */
e107be36 2492
4866cde0
NP
2493/**
2494 * prepare_task_switch - prepare to switch tasks
2495 * @rq: the runqueue preparing to switch
421cee29 2496 * @prev: the current task that is being switched out
4866cde0
NP
2497 * @next: the task we are going to switch to.
2498 *
2499 * This is called with the rq lock held and interrupts off. It must
2500 * be paired with a subsequent finish_task_switch after the context
2501 * switch.
2502 *
2503 * prepare_task_switch sets up locking and calls architecture specific
2504 * hooks.
2505 */
e107be36
AK
2506static inline void
2507prepare_task_switch(struct rq *rq, struct task_struct *prev,
2508 struct task_struct *next)
4866cde0 2509{
e107be36 2510 fire_sched_out_preempt_notifiers(prev, next);
4866cde0
NP
2511 prepare_lock_switch(rq, next);
2512 prepare_arch_switch(next);
2513}
2514
1da177e4
LT
2515/**
2516 * finish_task_switch - clean up after a task-switch
344babaa 2517 * @rq: runqueue associated with task-switch
1da177e4
LT
2518 * @prev: the thread we just switched away from.
2519 *
4866cde0
NP
2520 * finish_task_switch must be called after the context switch, paired
2521 * with a prepare_task_switch call before the context switch.
2522 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2523 * and do any other architecture-specific cleanup actions.
1da177e4
LT
2524 *
2525 * Note that we may have delayed dropping an mm in context_switch(). If
41a2d6cf 2526 * so, we finish that here outside of the runqueue lock. (Doing it
1da177e4
LT
2527 * with the lock held can cause deadlocks; see schedule() for
2528 * details.)
2529 */
a9957449 2530static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1da177e4
LT
2531 __releases(rq->lock)
2532{
1da177e4 2533 struct mm_struct *mm = rq->prev_mm;
55a101f8 2534 long prev_state;
1da177e4
LT
2535
2536 rq->prev_mm = NULL;
2537
2538 /*
2539 * A task struct has one reference for the use as "current".
c394cc9f 2540 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
55a101f8
ON
2541 * schedule one last time. The schedule call will never return, and
2542 * the scheduled task must drop that reference.
c394cc9f 2543 * The test for TASK_DEAD must occur while the runqueue locks are
1da177e4
LT
2544 * still held, otherwise prev could be scheduled on another cpu, die
2545 * there before we look at prev->state, and then the reference would
2546 * be dropped twice.
2547 * Manfred Spraul <manfred@colorfullife.com>
2548 */
55a101f8 2549 prev_state = prev->state;
4866cde0
NP
2550 finish_arch_switch(prev);
2551 finish_lock_switch(rq, prev);
9a897c5a
SR
2552#ifdef CONFIG_SMP
2553 if (current->sched_class->post_schedule)
2554 current->sched_class->post_schedule(rq);
2555#endif
e8fa1362 2556
e107be36 2557 fire_sched_in_preempt_notifiers(current);
1da177e4
LT
2558 if (mm)
2559 mmdrop(mm);
c394cc9f 2560 if (unlikely(prev_state == TASK_DEAD)) {
c6fd91f0 2561 /*
2562 * Remove function-return probe instances associated with this
2563 * task and put them back on the free list.
9761eea8 2564 */
c6fd91f0 2565 kprobe_flush_task(prev);
1da177e4 2566 put_task_struct(prev);
c6fd91f0 2567 }
1da177e4
LT
2568}
2569
2570/**
2571 * schedule_tail - first thing a freshly forked thread must call.
2572 * @prev: the thread we just switched away from.
2573 */
36c8b586 2574asmlinkage void schedule_tail(struct task_struct *prev)
1da177e4
LT
2575 __releases(rq->lock)
2576{
70b97a7f
IM
2577 struct rq *rq = this_rq();
2578
4866cde0
NP
2579 finish_task_switch(rq, prev);
2580#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2581 /* In this case, finish_task_switch does not reenable preemption */
2582 preempt_enable();
2583#endif
1da177e4 2584 if (current->set_child_tid)
b488893a 2585 put_user(task_pid_vnr(current), current->set_child_tid);
1da177e4
LT
2586}
2587
2588/*
2589 * context_switch - switch to the new MM and the new
2590 * thread's register state.
2591 */
dd41f596 2592static inline void
70b97a7f 2593context_switch(struct rq *rq, struct task_struct *prev,
36c8b586 2594 struct task_struct *next)
1da177e4 2595{
dd41f596 2596 struct mm_struct *mm, *oldmm;
1da177e4 2597
e107be36 2598 prepare_task_switch(rq, prev, next);
0a16b607 2599 trace_sched_switch(rq, prev, next);
dd41f596
IM
2600 mm = next->mm;
2601 oldmm = prev->active_mm;
9226d125
ZA
2602 /*
2603 * For paravirt, this is coupled with an exit in switch_to to
2604 * combine the page table reload and the switch backend into
2605 * one hypercall.
2606 */
2607 arch_enter_lazy_cpu_mode();
2608
dd41f596 2609 if (unlikely(!mm)) {
1da177e4
LT
2610 next->active_mm = oldmm;
2611 atomic_inc(&oldmm->mm_count);
2612 enter_lazy_tlb(oldmm, next);
2613 } else
2614 switch_mm(oldmm, mm, next);
2615
dd41f596 2616 if (unlikely(!prev->mm)) {
1da177e4 2617 prev->active_mm = NULL;
1da177e4
LT
2618 rq->prev_mm = oldmm;
2619 }
3a5f5e48
IM
2620 /*
2621 * Since the runqueue lock will be released by the next
2622 * task (which is an invalid locking op but in the case
2623 * of the scheduler it's an obvious special-case), so we
2624 * do an early lockdep release here:
2625 */
2626#ifndef __ARCH_WANT_UNLOCKED_CTXSW
8a25d5de 2627 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3a5f5e48 2628#endif
1da177e4
LT
2629
2630 /* Here we just switch the register state and the stack. */
2631 switch_to(prev, next, prev);
2632
dd41f596
IM
2633 barrier();
2634 /*
2635 * this_rq must be evaluated again because prev may have moved
2636 * CPUs since it called schedule(), thus the 'rq' on its stack
2637 * frame will be invalid.
2638 */
2639 finish_task_switch(this_rq(), prev);
1da177e4
LT
2640}
2641
2642/*
2643 * nr_running, nr_uninterruptible and nr_context_switches:
2644 *
2645 * externally visible scheduler statistics: current number of runnable
2646 * threads, current number of uninterruptible-sleeping threads, total
2647 * number of context switches performed since bootup.
2648 */
2649unsigned long nr_running(void)
2650{
2651 unsigned long i, sum = 0;
2652
2653 for_each_online_cpu(i)
2654 sum += cpu_rq(i)->nr_running;
2655
2656 return sum;
2657}
2658
2659unsigned long nr_uninterruptible(void)
2660{
2661 unsigned long i, sum = 0;
2662
0a945022 2663 for_each_possible_cpu(i)
1da177e4
LT
2664 sum += cpu_rq(i)->nr_uninterruptible;
2665
2666 /*
2667 * Since we read the counters lockless, it might be slightly
2668 * inaccurate. Do not allow it to go below zero though:
2669 */
2670 if (unlikely((long)sum < 0))
2671 sum = 0;
2672
2673 return sum;
2674}
2675
2676unsigned long long nr_context_switches(void)
2677{
cc94abfc
SR
2678 int i;
2679 unsigned long long sum = 0;
1da177e4 2680
0a945022 2681 for_each_possible_cpu(i)
1da177e4
LT
2682 sum += cpu_rq(i)->nr_switches;
2683
2684 return sum;
2685}
2686
2687unsigned long nr_iowait(void)
2688{
2689 unsigned long i, sum = 0;
2690
0a945022 2691 for_each_possible_cpu(i)
1da177e4
LT
2692 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2693
2694 return sum;
2695}
2696
db1b1fef
JS
2697unsigned long nr_active(void)
2698{
2699 unsigned long i, running = 0, uninterruptible = 0;
2700
2701 for_each_online_cpu(i) {
2702 running += cpu_rq(i)->nr_running;
2703 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2704 }
2705
2706 if (unlikely((long)uninterruptible < 0))
2707 uninterruptible = 0;
2708
2709 return running + uninterruptible;
2710}
2711
48f24c4d 2712/*
dd41f596
IM
2713 * Update rq->cpu_load[] statistics. This function is usually called every
2714 * scheduler tick (TICK_NSEC).
48f24c4d 2715 */
dd41f596 2716static void update_cpu_load(struct rq *this_rq)
48f24c4d 2717{
495eca49 2718 unsigned long this_load = this_rq->load.weight;
dd41f596
IM
2719 int i, scale;
2720
2721 this_rq->nr_load_updates++;
dd41f596
IM
2722
2723 /* Update our load: */
2724 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2725 unsigned long old_load, new_load;
2726
2727 /* scale is effectively 1 << i now, and >> i divides by scale */
2728
2729 old_load = this_rq->cpu_load[i];
2730 new_load = this_load;
a25707f3
IM
2731 /*
2732 * Round up the averaging division if load is increasing. This
2733 * prevents us from getting stuck on 9 if the load is 10, for
2734 * example.
2735 */
2736 if (new_load > old_load)
2737 new_load += scale-1;
dd41f596
IM
2738 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2739 }
48f24c4d
IM
2740}
2741
dd41f596
IM
2742#ifdef CONFIG_SMP
2743
1da177e4
LT
2744/*
2745 * double_rq_lock - safely lock two runqueues
2746 *
2747 * Note this does not disable interrupts like task_rq_lock,
2748 * you need to do so manually before calling.
2749 */
70b97a7f 2750static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1da177e4
LT
2751 __acquires(rq1->lock)
2752 __acquires(rq2->lock)
2753{
054b9108 2754 BUG_ON(!irqs_disabled());
1da177e4
LT
2755 if (rq1 == rq2) {
2756 spin_lock(&rq1->lock);
2757 __acquire(rq2->lock); /* Fake it out ;) */
2758 } else {
c96d145e 2759 if (rq1 < rq2) {
1da177e4 2760 spin_lock(&rq1->lock);
5e710e37 2761 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1da177e4
LT
2762 } else {
2763 spin_lock(&rq2->lock);
5e710e37 2764 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1da177e4
LT
2765 }
2766 }
6e82a3be
IM
2767 update_rq_clock(rq1);
2768 update_rq_clock(rq2);
1da177e4
LT
2769}
2770
2771/*
2772 * double_rq_unlock - safely unlock two runqueues
2773 *
2774 * Note this does not restore interrupts like task_rq_unlock,
2775 * you need to do so manually after calling.
2776 */
70b97a7f 2777static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1da177e4
LT
2778 __releases(rq1->lock)
2779 __releases(rq2->lock)
2780{
2781 spin_unlock(&rq1->lock);
2782 if (rq1 != rq2)
2783 spin_unlock(&rq2->lock);
2784 else
2785 __release(rq2->lock);
2786}
2787
2788/*
2789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2790 */
e8fa1362 2791static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1da177e4
LT
2792 __releases(this_rq->lock)
2793 __acquires(busiest->lock)
2794 __acquires(this_rq->lock)
2795{
e8fa1362
SR
2796 int ret = 0;
2797
054b9108
KK
2798 if (unlikely(!irqs_disabled())) {
2799 /* printk() doesn't work good under rq->lock */
2800 spin_unlock(&this_rq->lock);
2801 BUG_ON(1);
2802 }
1da177e4 2803 if (unlikely(!spin_trylock(&busiest->lock))) {
c96d145e 2804 if (busiest < this_rq) {
1da177e4
LT
2805 spin_unlock(&this_rq->lock);
2806 spin_lock(&busiest->lock);
5e710e37 2807 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
e8fa1362 2808 ret = 1;
1da177e4 2809 } else
5e710e37 2810 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1da177e4 2811 }
e8fa1362 2812 return ret;
1da177e4
LT
2813}
2814
cf7f8690 2815static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1b12bbc7
PZ
2816 __releases(busiest->lock)
2817{
2818 spin_unlock(&busiest->lock);
2819 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2820}
2821
1da177e4
LT
2822/*
2823 * If dest_cpu is allowed for this process, migrate the task to it.
2824 * This is accomplished by forcing the cpu_allowed mask to only
41a2d6cf 2825 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1da177e4
LT
2826 * the cpu_allowed mask is restored.
2827 */
36c8b586 2828static void sched_migrate_task(struct task_struct *p, int dest_cpu)
1da177e4 2829{
70b97a7f 2830 struct migration_req req;
1da177e4 2831 unsigned long flags;
70b97a7f 2832 struct rq *rq;
1da177e4
LT
2833
2834 rq = task_rq_lock(p, &flags);
2835 if (!cpu_isset(dest_cpu, p->cpus_allowed)
e761b772 2836 || unlikely(!cpu_active(dest_cpu)))
1da177e4
LT
2837 goto out;
2838
0a16b607 2839 trace_sched_migrate_task(rq, p, dest_cpu);
1da177e4
LT
2840 /* force the process onto the specified CPU */
2841 if (migrate_task(p, dest_cpu, &req)) {
2842 /* Need to wait for migration thread (might exit: take ref). */
2843 struct task_struct *mt = rq->migration_thread;
36c8b586 2844
1da177e4
LT
2845 get_task_struct(mt);
2846 task_rq_unlock(rq, &flags);
2847 wake_up_process(mt);
2848 put_task_struct(mt);
2849 wait_for_completion(&req.done);
36c8b586 2850
1da177e4
LT
2851 return;
2852 }
2853out:
2854 task_rq_unlock(rq, &flags);
2855}
2856
2857/*
476d139c
NP
2858 * sched_exec - execve() is a valuable balancing opportunity, because at
2859 * this point the task has the smallest effective memory and cache footprint.
1da177e4
LT
2860 */
2861void sched_exec(void)
2862{
1da177e4 2863 int new_cpu, this_cpu = get_cpu();
476d139c 2864 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1da177e4 2865 put_cpu();
476d139c
NP
2866 if (new_cpu != this_cpu)
2867 sched_migrate_task(current, new_cpu);
1da177e4
LT
2868}
2869
2870/*
2871 * pull_task - move a task from a remote runqueue to the local runqueue.
2872 * Both runqueues must be locked.
2873 */
dd41f596
IM
2874static void pull_task(struct rq *src_rq, struct task_struct *p,
2875 struct rq *this_rq, int this_cpu)
1da177e4 2876{
2e1cb74a 2877 deactivate_task(src_rq, p, 0);
1da177e4 2878 set_task_cpu(p, this_cpu);
dd41f596 2879 activate_task(this_rq, p, 0);
1da177e4
LT
2880 /*
2881 * Note that idle threads have a prio of MAX_PRIO, for this test
2882 * to be always true for them.
2883 */
15afe09b 2884 check_preempt_curr(this_rq, p, 0);
1da177e4
LT
2885}
2886
2887/*
2888 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2889 */
858119e1 2890static
70b97a7f 2891int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
d15bcfdb 2892 struct sched_domain *sd, enum cpu_idle_type idle,
95cdf3b7 2893 int *all_pinned)
1da177e4
LT
2894{
2895 /*
2896 * We do not migrate tasks that are:
2897 * 1) running (obviously), or
2898 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2899 * 3) are cache-hot on their current CPU.
2900 */
cc367732
IM
2901 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2902 schedstat_inc(p, se.nr_failed_migrations_affine);
1da177e4 2903 return 0;
cc367732 2904 }
81026794
NP
2905 *all_pinned = 0;
2906
cc367732
IM
2907 if (task_running(rq, p)) {
2908 schedstat_inc(p, se.nr_failed_migrations_running);
81026794 2909 return 0;
cc367732 2910 }
1da177e4 2911
da84d961
IM
2912 /*
2913 * Aggressive migration if:
2914 * 1) task is cache cold, or
2915 * 2) too many balance attempts have failed.
2916 */
2917
6bc1665b
IM
2918 if (!task_hot(p, rq->clock, sd) ||
2919 sd->nr_balance_failed > sd->cache_nice_tries) {
da84d961 2920#ifdef CONFIG_SCHEDSTATS
cc367732 2921 if (task_hot(p, rq->clock, sd)) {
da84d961 2922 schedstat_inc(sd, lb_hot_gained[idle]);
cc367732
IM
2923 schedstat_inc(p, se.nr_forced_migrations);
2924 }
da84d961
IM
2925#endif
2926 return 1;
2927 }
2928
cc367732
IM
2929 if (task_hot(p, rq->clock, sd)) {
2930 schedstat_inc(p, se.nr_failed_migrations_hot);
da84d961 2931 return 0;
cc367732 2932 }
1da177e4
LT
2933 return 1;
2934}
2935
e1d1484f
PW
2936static unsigned long
2937balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2938 unsigned long max_load_move, struct sched_domain *sd,
2939 enum cpu_idle_type idle, int *all_pinned,
2940 int *this_best_prio, struct rq_iterator *iterator)
1da177e4 2941{
051c6764 2942 int loops = 0, pulled = 0, pinned = 0;
dd41f596
IM
2943 struct task_struct *p;
2944 long rem_load_move = max_load_move;
1da177e4 2945
e1d1484f 2946 if (max_load_move == 0)
1da177e4
LT
2947 goto out;
2948
81026794
NP
2949 pinned = 1;
2950
1da177e4 2951 /*
dd41f596 2952 * Start the load-balancing iterator:
1da177e4 2953 */
dd41f596
IM
2954 p = iterator->start(iterator->arg);
2955next:
b82d9fdd 2956 if (!p || loops++ > sysctl_sched_nr_migrate)
1da177e4 2957 goto out;
051c6764
PZ
2958
2959 if ((p->se.load.weight >> 1) > rem_load_move ||
dd41f596 2960 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
dd41f596
IM
2961 p = iterator->next(iterator->arg);
2962 goto next;
1da177e4
LT
2963 }
2964
dd41f596 2965 pull_task(busiest, p, this_rq, this_cpu);
1da177e4 2966 pulled++;
dd41f596 2967 rem_load_move -= p->se.load.weight;
1da177e4 2968
2dd73a4f 2969 /*
b82d9fdd 2970 * We only want to steal up to the prescribed amount of weighted load.
2dd73a4f 2971 */
e1d1484f 2972 if (rem_load_move > 0) {
a4ac01c3
PW
2973 if (p->prio < *this_best_prio)
2974 *this_best_prio = p->prio;
dd41f596
IM
2975 p = iterator->next(iterator->arg);
2976 goto next;
1da177e4
LT
2977 }
2978out:
2979 /*
e1d1484f 2980 * Right now, this is one of only two places pull_task() is called,
1da177e4
LT
2981 * so we can safely collect pull_task() stats here rather than
2982 * inside pull_task().
2983 */
2984 schedstat_add(sd, lb_gained[idle], pulled);
81026794
NP
2985
2986 if (all_pinned)
2987 *all_pinned = pinned;
e1d1484f
PW
2988
2989 return max_load_move - rem_load_move;
1da177e4
LT
2990}
2991
dd41f596 2992/*
43010659
PW
2993 * move_tasks tries to move up to max_load_move weighted load from busiest to
2994 * this_rq, as part of a balancing operation within domain "sd".
2995 * Returns 1 if successful and 0 otherwise.
dd41f596
IM
2996 *
2997 * Called with both runqueues locked.
2998 */
2999static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
43010659 3000 unsigned long max_load_move,
dd41f596
IM
3001 struct sched_domain *sd, enum cpu_idle_type idle,
3002 int *all_pinned)
3003{
5522d5d5 3004 const struct sched_class *class = sched_class_highest;
43010659 3005 unsigned long total_load_moved = 0;
a4ac01c3 3006 int this_best_prio = this_rq->curr->prio;
dd41f596
IM
3007
3008 do {
43010659
PW
3009 total_load_moved +=
3010 class->load_balance(this_rq, this_cpu, busiest,
e1d1484f 3011 max_load_move - total_load_moved,
a4ac01c3 3012 sd, idle, all_pinned, &this_best_prio);
dd41f596 3013 class = class->next;
c4acb2c0
GH
3014
3015 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3016 break;
3017
43010659 3018 } while (class && max_load_move > total_load_moved);
dd41f596 3019
43010659
PW
3020 return total_load_moved > 0;
3021}
3022
e1d1484f
PW
3023static int
3024iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3025 struct sched_domain *sd, enum cpu_idle_type idle,
3026 struct rq_iterator *iterator)
3027{
3028 struct task_struct *p = iterator->start(iterator->arg);
3029 int pinned = 0;
3030
3031 while (p) {
3032 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3033 pull_task(busiest, p, this_rq, this_cpu);
3034 /*
3035 * Right now, this is only the second place pull_task()
3036 * is called, so we can safely collect pull_task()
3037 * stats here rather than inside pull_task().
3038 */
3039 schedstat_inc(sd, lb_gained[idle]);
3040
3041 return 1;
3042 }
3043 p = iterator->next(iterator->arg);
3044 }
3045
3046 return 0;
3047}
3048
43010659
PW
3049/*
3050 * move_one_task tries to move exactly one task from busiest to this_rq, as
3051 * part of active balancing operations within "domain".
3052 * Returns 1 if successful and 0 otherwise.
3053 *
3054 * Called with both runqueues locked.
3055 */
3056static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3057 struct sched_domain *sd, enum cpu_idle_type idle)
3058{
5522d5d5 3059 const struct sched_class *class;
43010659
PW
3060
3061 for (class = sched_class_highest; class; class = class->next)
e1d1484f 3062 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
43010659
PW
3063 return 1;
3064
3065 return 0;
dd41f596
IM
3066}
3067
1da177e4
LT
3068/*
3069 * find_busiest_group finds and returns the busiest CPU group within the
48f24c4d
IM
3070 * domain. It calculates and returns the amount of weighted load which
3071 * should be moved to restore balance via the imbalance parameter.
1da177e4
LT
3072 */
3073static struct sched_group *
3074find_busiest_group(struct sched_domain *sd, int this_cpu,
dd41f596 3075 unsigned long *imbalance, enum cpu_idle_type idle,
7c16ec58 3076 int *sd_idle, const cpumask_t *cpus, int *balance)
1da177e4
LT
3077{
3078 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3079 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
0c117f1b 3080 unsigned long max_pull;
2dd73a4f
PW
3081 unsigned long busiest_load_per_task, busiest_nr_running;
3082 unsigned long this_load_per_task, this_nr_running;
908a7c1b 3083 int load_idx, group_imb = 0;
5c45bf27
SS
3084#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3085 int power_savings_balance = 1;
3086 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3087 unsigned long min_nr_running = ULONG_MAX;
3088 struct sched_group *group_min = NULL, *group_leader = NULL;
3089#endif
1da177e4
LT
3090
3091 max_load = this_load = total_load = total_pwr = 0;
2dd73a4f
PW
3092 busiest_load_per_task = busiest_nr_running = 0;
3093 this_load_per_task = this_nr_running = 0;
408ed066 3094
d15bcfdb 3095 if (idle == CPU_NOT_IDLE)
7897986b 3096 load_idx = sd->busy_idx;
d15bcfdb 3097 else if (idle == CPU_NEWLY_IDLE)
7897986b
NP
3098 load_idx = sd->newidle_idx;
3099 else
3100 load_idx = sd->idle_idx;
1da177e4
LT
3101
3102 do {
908a7c1b 3103 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
1da177e4
LT
3104 int local_group;
3105 int i;
908a7c1b 3106 int __group_imb = 0;
783609c6 3107 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2dd73a4f 3108 unsigned long sum_nr_running, sum_weighted_load;
408ed066
PZ
3109 unsigned long sum_avg_load_per_task;
3110 unsigned long avg_load_per_task;
1da177e4
LT
3111
3112 local_group = cpu_isset(this_cpu, group->cpumask);
3113
783609c6
SS
3114 if (local_group)
3115 balance_cpu = first_cpu(group->cpumask);
3116
1da177e4 3117 /* Tally up the load of all CPUs in the group */
2dd73a4f 3118 sum_weighted_load = sum_nr_running = avg_load = 0;
408ed066
PZ
3119 sum_avg_load_per_task = avg_load_per_task = 0;
3120
908a7c1b
KC
3121 max_cpu_load = 0;
3122 min_cpu_load = ~0UL;
1da177e4 3123
363ab6f1 3124 for_each_cpu_mask_nr(i, group->cpumask) {
0a2966b4
CL
3125 struct rq *rq;
3126
3127 if (!cpu_isset(i, *cpus))
3128 continue;
3129
3130 rq = cpu_rq(i);
2dd73a4f 3131
9439aab8 3132 if (*sd_idle && rq->nr_running)
5969fe06
NP
3133 *sd_idle = 0;
3134
1da177e4 3135 /* Bias balancing toward cpus of our domain */
783609c6
SS
3136 if (local_group) {
3137 if (idle_cpu(i) && !first_idle_cpu) {
3138 first_idle_cpu = 1;
3139 balance_cpu = i;
3140 }
3141
a2000572 3142 load = target_load(i, load_idx);
908a7c1b 3143 } else {
a2000572 3144 load = source_load(i, load_idx);
908a7c1b
KC
3145 if (load > max_cpu_load)
3146 max_cpu_load = load;
3147 if (min_cpu_load > load)
3148 min_cpu_load = load;
3149 }
1da177e4
LT
3150
3151 avg_load += load;
2dd73a4f 3152 sum_nr_running += rq->nr_running;
dd41f596 3153 sum_weighted_load += weighted_cpuload(i);
408ed066
PZ
3154
3155 sum_avg_load_per_task += cpu_avg_load_per_task(i);
1da177e4
LT
3156 }
3157
783609c6
SS
3158 /*
3159 * First idle cpu or the first cpu(busiest) in this sched group
3160 * is eligible for doing load balancing at this and above
9439aab8
SS
3161 * domains. In the newly idle case, we will allow all the cpu's
3162 * to do the newly idle load balance.
783609c6 3163 */
9439aab8
SS
3164 if (idle != CPU_NEWLY_IDLE && local_group &&
3165 balance_cpu != this_cpu && balance) {
783609c6
SS
3166 *balance = 0;
3167 goto ret;
3168 }
3169
1da177e4 3170 total_load += avg_load;
5517d86b 3171 total_pwr += group->__cpu_power;
1da177e4
LT
3172
3173 /* Adjust by relative CPU power of the group */
5517d86b
ED
3174 avg_load = sg_div_cpu_power(group,
3175 avg_load * SCHED_LOAD_SCALE);
1da177e4 3176
408ed066
PZ
3177
3178 /*
3179 * Consider the group unbalanced when the imbalance is larger
3180 * than the average weight of two tasks.
3181 *
3182 * APZ: with cgroup the avg task weight can vary wildly and
3183 * might not be a suitable number - should we keep a
3184 * normalized nr_running number somewhere that negates
3185 * the hierarchy?
3186 */
3187 avg_load_per_task = sg_div_cpu_power(group,
3188 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3189
3190 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
908a7c1b
KC
3191 __group_imb = 1;
3192
5517d86b 3193 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
5c45bf27 3194
1da177e4
LT
3195 if (local_group) {
3196 this_load = avg_load;
3197 this = group;
2dd73a4f
PW
3198 this_nr_running = sum_nr_running;
3199 this_load_per_task = sum_weighted_load;
3200 } else if (avg_load > max_load &&
908a7c1b 3201 (sum_nr_running > group_capacity || __group_imb)) {
1da177e4
LT
3202 max_load = avg_load;
3203 busiest = group;
2dd73a4f
PW
3204 busiest_nr_running = sum_nr_running;
3205 busiest_load_per_task = sum_weighted_load;
908a7c1b 3206 group_imb = __group_imb;
1da177e4 3207 }
5c45bf27
SS
3208
3209#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3210 /*
3211 * Busy processors will not participate in power savings
3212 * balance.
3213 */
dd41f596
IM
3214 if (idle == CPU_NOT_IDLE ||
3215 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3216 goto group_next;
5c45bf27
SS
3217
3218 /*
3219 * If the local group is idle or completely loaded
3220 * no need to do power savings balance at this domain
3221 */
3222 if (local_group && (this_nr_running >= group_capacity ||
3223 !this_nr_running))
3224 power_savings_balance = 0;
3225
dd41f596 3226 /*
5c45bf27
SS
3227 * If a group is already running at full capacity or idle,
3228 * don't include that group in power savings calculations
dd41f596
IM
3229 */
3230 if (!power_savings_balance || sum_nr_running >= group_capacity
5c45bf27 3231 || !sum_nr_running)
dd41f596 3232 goto group_next;
5c45bf27 3233
dd41f596 3234 /*
5c45bf27 3235 * Calculate the group which has the least non-idle load.
dd41f596
IM
3236 * This is the group from where we need to pick up the load
3237 * for saving power
3238 */
3239 if ((sum_nr_running < min_nr_running) ||
3240 (sum_nr_running == min_nr_running &&
5c45bf27
SS
3241 first_cpu(group->cpumask) <
3242 first_cpu(group_min->cpumask))) {
dd41f596
IM
3243 group_min = group;
3244 min_nr_running = sum_nr_running;
5c45bf27
SS
3245 min_load_per_task = sum_weighted_load /
3246 sum_nr_running;
dd41f596 3247 }
5c45bf27 3248
dd41f596 3249 /*
5c45bf27 3250 * Calculate the group which is almost near its
dd41f596
IM
3251 * capacity but still has some space to pick up some load
3252 * from other group and save more power
3253 */
3254 if (sum_nr_running <= group_capacity - 1) {
3255 if (sum_nr_running > leader_nr_running ||
3256 (sum_nr_running == leader_nr_running &&
3257 first_cpu(group->cpumask) >
3258 first_cpu(group_leader->cpumask))) {
3259 group_leader = group;
3260 leader_nr_running = sum_nr_running;
3261 }
48f24c4d 3262 }
5c45bf27
SS
3263group_next:
3264#endif
1da177e4
LT
3265 group = group->next;
3266 } while (group != sd->groups);
3267
2dd73a4f 3268 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
1da177e4
LT
3269 goto out_balanced;
3270
3271 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3272
3273 if (this_load >= avg_load ||
3274 100*max_load <= sd->imbalance_pct*this_load)
3275 goto out_balanced;
3276
2dd73a4f 3277 busiest_load_per_task /= busiest_nr_running;
908a7c1b
KC
3278 if (group_imb)
3279 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3280
1da177e4
LT
3281 /*
3282 * We're trying to get all the cpus to the average_load, so we don't
3283 * want to push ourselves above the average load, nor do we wish to
3284 * reduce the max loaded cpu below the average load, as either of these
3285 * actions would just result in more rebalancing later, and ping-pong
3286 * tasks around. Thus we look for the minimum possible imbalance.
3287 * Negative imbalances (*we* are more loaded than anyone else) will
3288 * be counted as no imbalance for these purposes -- we can't fix that
41a2d6cf 3289 * by pulling tasks to us. Be careful of negative numbers as they'll
1da177e4
LT
3290 * appear as very large values with unsigned longs.
3291 */
2dd73a4f
PW
3292 if (max_load <= busiest_load_per_task)
3293 goto out_balanced;
3294
3295 /*
3296 * In the presence of smp nice balancing, certain scenarios can have
3297 * max load less than avg load(as we skip the groups at or below
3298 * its cpu_power, while calculating max_load..)
3299 */
3300 if (max_load < avg_load) {
3301 *imbalance = 0;
3302 goto small_imbalance;
3303 }
0c117f1b
SS
3304
3305 /* Don't want to pull so many tasks that a group would go idle */
2dd73a4f 3306 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
0c117f1b 3307
1da177e4 3308 /* How much load to actually move to equalise the imbalance */
5517d86b
ED
3309 *imbalance = min(max_pull * busiest->__cpu_power,
3310 (avg_load - this_load) * this->__cpu_power)
1da177e4
LT
3311 / SCHED_LOAD_SCALE;
3312
2dd73a4f
PW
3313 /*
3314 * if *imbalance is less than the average load per runnable task
3315 * there is no gaurantee that any tasks will be moved so we'll have
3316 * a think about bumping its value to force at least one task to be
3317 * moved
3318 */
7fd0d2dd 3319 if (*imbalance < busiest_load_per_task) {
48f24c4d 3320 unsigned long tmp, pwr_now, pwr_move;
2dd73a4f
PW
3321 unsigned int imbn;
3322
3323small_imbalance:
3324 pwr_move = pwr_now = 0;
3325 imbn = 2;
3326 if (this_nr_running) {
3327 this_load_per_task /= this_nr_running;
3328 if (busiest_load_per_task > this_load_per_task)
3329 imbn = 1;
3330 } else
408ed066 3331 this_load_per_task = cpu_avg_load_per_task(this_cpu);
1da177e4 3332
01c8c57d 3333 if (max_load - this_load + busiest_load_per_task >=
dd41f596 3334 busiest_load_per_task * imbn) {
2dd73a4f 3335 *imbalance = busiest_load_per_task;
1da177e4
LT
3336 return busiest;
3337 }
3338
3339 /*
3340 * OK, we don't have enough imbalance to justify moving tasks,
3341 * however we may be able to increase total CPU power used by
3342 * moving them.
3343 */
3344
5517d86b
ED
3345 pwr_now += busiest->__cpu_power *
3346 min(busiest_load_per_task, max_load);
3347 pwr_now += this->__cpu_power *
3348 min(this_load_per_task, this_load);
1da177e4
LT
3349 pwr_now /= SCHED_LOAD_SCALE;
3350
3351 /* Amount of load we'd subtract */
5517d86b
ED
3352 tmp = sg_div_cpu_power(busiest,
3353 busiest_load_per_task * SCHED_LOAD_SCALE);
1da177e4 3354 if (max_load > tmp)
5517d86b 3355 pwr_move += busiest->__cpu_power *
2dd73a4f 3356 min(busiest_load_per_task, max_load - tmp);
1da177e4
LT
3357
3358 /* Amount of load we'd add */
5517d86b 3359 if (max_load * busiest->__cpu_power <
33859f7f 3360 busiest_load_per_task * SCHED_LOAD_SCALE)
5517d86b
ED
3361 tmp = sg_div_cpu_power(this,
3362 max_load * busiest->__cpu_power);
1da177e4 3363 else
5517d86b
ED
3364 tmp = sg_div_cpu_power(this,
3365 busiest_load_per_task * SCHED_LOAD_SCALE);
3366 pwr_move += this->__cpu_power *
3367 min(this_load_per_task, this_load + tmp);
1da177e4
LT
3368 pwr_move /= SCHED_LOAD_SCALE;
3369
3370 /* Move if we gain throughput */
7fd0d2dd
SS
3371 if (pwr_move > pwr_now)
3372 *imbalance = busiest_load_per_task;
1da177e4
LT
3373 }
3374
1da177e4
LT
3375 return busiest;
3376
3377out_balanced:
5c45bf27 3378#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
d15bcfdb 3379 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
5c45bf27 3380 goto ret;
1da177e4 3381
5c45bf27
SS
3382 if (this == group_leader && group_leader != group_min) {
3383 *imbalance = min_load_per_task;
3384 return group_min;
3385 }
5c45bf27 3386#endif
783609c6 3387ret:
1da177e4
LT
3388 *imbalance = 0;
3389 return NULL;
3390}
3391
3392/*
3393 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3394 */
70b97a7f 3395static struct rq *
d15bcfdb 3396find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
7c16ec58 3397 unsigned long imbalance, const cpumask_t *cpus)
1da177e4 3398{
70b97a7f 3399 struct rq *busiest = NULL, *rq;
2dd73a4f 3400 unsigned long max_load = 0;
1da177e4
LT
3401 int i;
3402
363ab6f1 3403 for_each_cpu_mask_nr(i, group->cpumask) {
dd41f596 3404 unsigned long wl;
0a2966b4
CL
3405
3406 if (!cpu_isset(i, *cpus))
3407 continue;
3408
48f24c4d 3409 rq = cpu_rq(i);
dd41f596 3410 wl = weighted_cpuload(i);
2dd73a4f 3411
dd41f596 3412 if (rq->nr_running == 1 && wl > imbalance)
2dd73a4f 3413 continue;
1da177e4 3414
dd41f596
IM
3415 if (wl > max_load) {
3416 max_load = wl;
48f24c4d 3417 busiest = rq;
1da177e4
LT
3418 }
3419 }
3420
3421 return busiest;
3422}
3423
77391d71
NP
3424/*
3425 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3426 * so long as it is large enough.
3427 */
3428#define MAX_PINNED_INTERVAL 512
3429
1da177e4
LT
3430/*
3431 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3432 * tasks if there is an imbalance.
1da177e4 3433 */
70b97a7f 3434static int load_balance(int this_cpu, struct rq *this_rq,
d15bcfdb 3435 struct sched_domain *sd, enum cpu_idle_type idle,
7c16ec58 3436 int *balance, cpumask_t *cpus)
1da177e4 3437{
43010659 3438 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
1da177e4 3439 struct sched_group *group;
1da177e4 3440 unsigned long imbalance;
70b97a7f 3441 struct rq *busiest;
fe2eea3f 3442 unsigned long flags;
5969fe06 3443
7c16ec58
MT
3444 cpus_setall(*cpus);
3445
89c4710e
SS
3446 /*
3447 * When power savings policy is enabled for the parent domain, idle
3448 * sibling can pick up load irrespective of busy siblings. In this case,
dd41f596 3449 * let the state of idle sibling percolate up as CPU_IDLE, instead of
d15bcfdb 3450 * portraying it as CPU_NOT_IDLE.
89c4710e 3451 */
d15bcfdb 3452 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 3453 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 3454 sd_idle = 1;
1da177e4 3455
2d72376b 3456 schedstat_inc(sd, lb_count[idle]);
1da177e4 3457
0a2966b4 3458redo:
c8cba857 3459 update_shares(sd);
0a2966b4 3460 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
7c16ec58 3461 cpus, balance);
783609c6 3462
06066714 3463 if (*balance == 0)
783609c6 3464 goto out_balanced;
783609c6 3465
1da177e4
LT
3466 if (!group) {
3467 schedstat_inc(sd, lb_nobusyg[idle]);
3468 goto out_balanced;
3469 }
3470
7c16ec58 3471 busiest = find_busiest_queue(group, idle, imbalance, cpus);
1da177e4
LT
3472 if (!busiest) {
3473 schedstat_inc(sd, lb_nobusyq[idle]);
3474 goto out_balanced;
3475 }
3476
db935dbd 3477 BUG_ON(busiest == this_rq);
1da177e4
LT
3478
3479 schedstat_add(sd, lb_imbalance[idle], imbalance);
3480
43010659 3481 ld_moved = 0;
1da177e4
LT
3482 if (busiest->nr_running > 1) {
3483 /*
3484 * Attempt to move tasks. If find_busiest_group has found
3485 * an imbalance but busiest->nr_running <= 1, the group is
43010659 3486 * still unbalanced. ld_moved simply stays zero, so it is
1da177e4
LT
3487 * correctly treated as an imbalance.
3488 */
fe2eea3f 3489 local_irq_save(flags);
e17224bf 3490 double_rq_lock(this_rq, busiest);
43010659 3491 ld_moved = move_tasks(this_rq, this_cpu, busiest,
48f24c4d 3492 imbalance, sd, idle, &all_pinned);
e17224bf 3493 double_rq_unlock(this_rq, busiest);
fe2eea3f 3494 local_irq_restore(flags);
81026794 3495
46cb4b7c
SS
3496 /*
3497 * some other cpu did the load balance for us.
3498 */
43010659 3499 if (ld_moved && this_cpu != smp_processor_id())
46cb4b7c
SS
3500 resched_cpu(this_cpu);
3501
81026794 3502 /* All tasks on this runqueue were pinned by CPU affinity */
0a2966b4 3503 if (unlikely(all_pinned)) {
7c16ec58
MT
3504 cpu_clear(cpu_of(busiest), *cpus);
3505 if (!cpus_empty(*cpus))
0a2966b4 3506 goto redo;
81026794 3507 goto out_balanced;
0a2966b4 3508 }
1da177e4 3509 }
81026794 3510
43010659 3511 if (!ld_moved) {
1da177e4
LT
3512 schedstat_inc(sd, lb_failed[idle]);
3513 sd->nr_balance_failed++;
3514
3515 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
1da177e4 3516
fe2eea3f 3517 spin_lock_irqsave(&busiest->lock, flags);
fa3b6ddc
SS
3518
3519 /* don't kick the migration_thread, if the curr
3520 * task on busiest cpu can't be moved to this_cpu
3521 */
3522 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
fe2eea3f 3523 spin_unlock_irqrestore(&busiest->lock, flags);
fa3b6ddc
SS
3524 all_pinned = 1;
3525 goto out_one_pinned;
3526 }
3527
1da177e4
LT
3528 if (!busiest->active_balance) {
3529 busiest->active_balance = 1;
3530 busiest->push_cpu = this_cpu;
81026794 3531 active_balance = 1;
1da177e4 3532 }
fe2eea3f 3533 spin_unlock_irqrestore(&busiest->lock, flags);
81026794 3534 if (active_balance)
1da177e4
LT
3535 wake_up_process(busiest->migration_thread);
3536
3537 /*
3538 * We've kicked active balancing, reset the failure
3539 * counter.
3540 */
39507451 3541 sd->nr_balance_failed = sd->cache_nice_tries+1;
1da177e4 3542 }
81026794 3543 } else
1da177e4
LT
3544 sd->nr_balance_failed = 0;
3545
81026794 3546 if (likely(!active_balance)) {
1da177e4
LT
3547 /* We were unbalanced, so reset the balancing interval */
3548 sd->balance_interval = sd->min_interval;
81026794
NP
3549 } else {
3550 /*
3551 * If we've begun active balancing, start to back off. This
3552 * case may not be covered by the all_pinned logic if there
3553 * is only 1 task on the busy runqueue (because we don't call
3554 * move_tasks).
3555 */
3556 if (sd->balance_interval < sd->max_interval)
3557 sd->balance_interval *= 2;
1da177e4
LT
3558 }
3559
43010659 3560 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 3561 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
c09595f6
PZ
3562 ld_moved = -1;
3563
3564 goto out;
1da177e4
LT
3565
3566out_balanced:
1da177e4
LT
3567 schedstat_inc(sd, lb_balanced[idle]);
3568
16cfb1c0 3569 sd->nr_balance_failed = 0;
fa3b6ddc
SS
3570
3571out_one_pinned:
1da177e4 3572 /* tune up the balancing interval */
77391d71
NP
3573 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3574 (sd->balance_interval < sd->max_interval))
1da177e4
LT
3575 sd->balance_interval *= 2;
3576
48f24c4d 3577 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 3578 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
c09595f6
PZ
3579 ld_moved = -1;
3580 else
3581 ld_moved = 0;
3582out:
c8cba857
PZ
3583 if (ld_moved)
3584 update_shares(sd);
c09595f6 3585 return ld_moved;
1da177e4
LT
3586}
3587
3588/*
3589 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3590 * tasks if there is an imbalance.
3591 *
d15bcfdb 3592 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
1da177e4
LT
3593 * this_rq is locked.
3594 */
48f24c4d 3595static int
7c16ec58
MT
3596load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3597 cpumask_t *cpus)
1da177e4
LT
3598{
3599 struct sched_group *group;
70b97a7f 3600 struct rq *busiest = NULL;
1da177e4 3601 unsigned long imbalance;
43010659 3602 int ld_moved = 0;
5969fe06 3603 int sd_idle = 0;
969bb4e4 3604 int all_pinned = 0;
7c16ec58
MT
3605
3606 cpus_setall(*cpus);
5969fe06 3607
89c4710e
SS
3608 /*
3609 * When power savings policy is enabled for the parent domain, idle
3610 * sibling can pick up load irrespective of busy siblings. In this case,
3611 * let the state of idle sibling percolate up as IDLE, instead of
d15bcfdb 3612 * portraying it as CPU_NOT_IDLE.
89c4710e
SS
3613 */
3614 if (sd->flags & SD_SHARE_CPUPOWER &&
3615 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 3616 sd_idle = 1;
1da177e4 3617
2d72376b 3618 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
0a2966b4 3619redo:
3e5459b4 3620 update_shares_locked(this_rq, sd);
d15bcfdb 3621 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
7c16ec58 3622 &sd_idle, cpus, NULL);
1da177e4 3623 if (!group) {
d15bcfdb 3624 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
16cfb1c0 3625 goto out_balanced;
1da177e4
LT
3626 }
3627
7c16ec58 3628 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
db935dbd 3629 if (!busiest) {
d15bcfdb 3630 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
16cfb1c0 3631 goto out_balanced;
1da177e4
LT
3632 }
3633
db935dbd
NP
3634 BUG_ON(busiest == this_rq);
3635
d15bcfdb 3636 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
d6d5cfaf 3637
43010659 3638 ld_moved = 0;
d6d5cfaf
NP
3639 if (busiest->nr_running > 1) {
3640 /* Attempt to move tasks */
3641 double_lock_balance(this_rq, busiest);
6e82a3be
IM
3642 /* this_rq->clock is already updated */
3643 update_rq_clock(busiest);
43010659 3644 ld_moved = move_tasks(this_rq, this_cpu, busiest,
969bb4e4
SS
3645 imbalance, sd, CPU_NEWLY_IDLE,
3646 &all_pinned);
1b12bbc7 3647 double_unlock_balance(this_rq, busiest);
0a2966b4 3648
969bb4e4 3649 if (unlikely(all_pinned)) {
7c16ec58
MT
3650 cpu_clear(cpu_of(busiest), *cpus);
3651 if (!cpus_empty(*cpus))
0a2966b4
CL
3652 goto redo;
3653 }
d6d5cfaf
NP
3654 }
3655
43010659 3656 if (!ld_moved) {
d15bcfdb 3657 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
89c4710e
SS
3658 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3659 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06
NP
3660 return -1;
3661 } else
16cfb1c0 3662 sd->nr_balance_failed = 0;
1da177e4 3663
3e5459b4 3664 update_shares_locked(this_rq, sd);
43010659 3665 return ld_moved;
16cfb1c0
NP
3666
3667out_balanced:
d15bcfdb 3668 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
48f24c4d 3669 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 3670 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 3671 return -1;
16cfb1c0 3672 sd->nr_balance_failed = 0;
48f24c4d 3673
16cfb1c0 3674 return 0;
1da177e4
LT
3675}
3676
3677/*
3678 * idle_balance is called by schedule() if this_cpu is about to become
3679 * idle. Attempts to pull tasks from other CPUs.
3680 */
70b97a7f 3681static void idle_balance(int this_cpu, struct rq *this_rq)
1da177e4
LT
3682{
3683 struct sched_domain *sd;
dd41f596
IM
3684 int pulled_task = -1;
3685 unsigned long next_balance = jiffies + HZ;
7c16ec58 3686 cpumask_t tmpmask;
1da177e4
LT
3687
3688 for_each_domain(this_cpu, sd) {
92c4ca5c
CL
3689 unsigned long interval;
3690
3691 if (!(sd->flags & SD_LOAD_BALANCE))
3692 continue;
3693
3694 if (sd->flags & SD_BALANCE_NEWIDLE)
48f24c4d 3695 /* If we've pulled tasks over stop searching: */
7c16ec58
MT
3696 pulled_task = load_balance_newidle(this_cpu, this_rq,
3697 sd, &tmpmask);
92c4ca5c
CL
3698
3699 interval = msecs_to_jiffies(sd->balance_interval);
3700 if (time_after(next_balance, sd->last_balance + interval))
3701 next_balance = sd->last_balance + interval;
3702 if (pulled_task)
3703 break;
1da177e4 3704 }
dd41f596 3705 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
1bd77f2d
CL
3706 /*
3707 * We are going idle. next_balance may be set based on
3708 * a busy processor. So reset next_balance.
3709 */
3710 this_rq->next_balance = next_balance;
dd41f596 3711 }
1da177e4
LT
3712}
3713
3714/*
3715 * active_load_balance is run by migration threads. It pushes running tasks
3716 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3717 * running on each physical CPU where possible, and avoids physical /
3718 * logical imbalances.
3719 *
3720 * Called with busiest_rq locked.
3721 */
70b97a7f 3722static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
1da177e4 3723{
39507451 3724 int target_cpu = busiest_rq->push_cpu;
70b97a7f
IM
3725 struct sched_domain *sd;
3726 struct rq *target_rq;
39507451 3727
48f24c4d 3728 /* Is there any task to move? */
39507451 3729 if (busiest_rq->nr_running <= 1)
39507451
NP
3730 return;
3731
3732 target_rq = cpu_rq(target_cpu);
1da177e4
LT
3733
3734 /*
39507451 3735 * This condition is "impossible", if it occurs
41a2d6cf 3736 * we need to fix it. Originally reported by
39507451 3737 * Bjorn Helgaas on a 128-cpu setup.
1da177e4 3738 */
39507451 3739 BUG_ON(busiest_rq == target_rq);
1da177e4 3740
39507451
NP
3741 /* move a task from busiest_rq to target_rq */
3742 double_lock_balance(busiest_rq, target_rq);
6e82a3be
IM
3743 update_rq_clock(busiest_rq);
3744 update_rq_clock(target_rq);
39507451
NP
3745
3746 /* Search for an sd spanning us and the target CPU. */
c96d145e 3747 for_each_domain(target_cpu, sd) {
39507451 3748 if ((sd->flags & SD_LOAD_BALANCE) &&
48f24c4d 3749 cpu_isset(busiest_cpu, sd->span))
39507451 3750 break;
c96d145e 3751 }
39507451 3752
48f24c4d 3753 if (likely(sd)) {
2d72376b 3754 schedstat_inc(sd, alb_count);
39507451 3755
43010659
PW
3756 if (move_one_task(target_rq, target_cpu, busiest_rq,
3757 sd, CPU_IDLE))
48f24c4d
IM
3758 schedstat_inc(sd, alb_pushed);
3759 else
3760 schedstat_inc(sd, alb_failed);
3761 }
1b12bbc7 3762 double_unlock_balance(busiest_rq, target_rq);
1da177e4
LT
3763}
3764
46cb4b7c
SS
3765#ifdef CONFIG_NO_HZ
3766static struct {
3767 atomic_t load_balancer;
41a2d6cf 3768 cpumask_t cpu_mask;
46cb4b7c
SS
3769} nohz ____cacheline_aligned = {
3770 .load_balancer = ATOMIC_INIT(-1),
3771 .cpu_mask = CPU_MASK_NONE,
3772};
3773
7835b98b 3774/*
46cb4b7c
SS
3775 * This routine will try to nominate the ilb (idle load balancing)
3776 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3777 * load balancing on behalf of all those cpus. If all the cpus in the system
3778 * go into this tickless mode, then there will be no ilb owner (as there is
3779 * no need for one) and all the cpus will sleep till the next wakeup event
3780 * arrives...
3781 *
3782 * For the ilb owner, tick is not stopped. And this tick will be used
3783 * for idle load balancing. ilb owner will still be part of
3784 * nohz.cpu_mask..
7835b98b 3785 *
46cb4b7c
SS
3786 * While stopping the tick, this cpu will become the ilb owner if there
3787 * is no other owner. And will be the owner till that cpu becomes busy
3788 * or if all cpus in the system stop their ticks at which point
3789 * there is no need for ilb owner.
3790 *
3791 * When the ilb owner becomes busy, it nominates another owner, during the
3792 * next busy scheduler_tick()
3793 */
3794int select_nohz_load_balancer(int stop_tick)
3795{
3796 int cpu = smp_processor_id();
3797
3798 if (stop_tick) {
3799 cpu_set(cpu, nohz.cpu_mask);
3800 cpu_rq(cpu)->in_nohz_recently = 1;
3801
3802 /*
3803 * If we are going offline and still the leader, give up!
3804 */
e761b772 3805 if (!cpu_active(cpu) &&
46cb4b7c
SS
3806 atomic_read(&nohz.load_balancer) == cpu) {
3807 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3808 BUG();
3809 return 0;
3810 }
3811
3812 /* time for ilb owner also to sleep */
3813 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3814 if (atomic_read(&nohz.load_balancer) == cpu)
3815 atomic_set(&nohz.load_balancer, -1);
3816 return 0;
3817 }
3818
3819 if (atomic_read(&nohz.load_balancer) == -1) {
3820 /* make me the ilb owner */
3821 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3822 return 1;
3823 } else if (atomic_read(&nohz.load_balancer) == cpu)
3824 return 1;
3825 } else {
3826 if (!cpu_isset(cpu, nohz.cpu_mask))
3827 return 0;
3828
3829 cpu_clear(cpu, nohz.cpu_mask);
3830
3831 if (atomic_read(&nohz.load_balancer) == cpu)
3832 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3833 BUG();
3834 }
3835 return 0;
3836}
3837#endif
3838
3839static DEFINE_SPINLOCK(balancing);
3840
3841/*
7835b98b
CL
3842 * It checks each scheduling domain to see if it is due to be balanced,
3843 * and initiates a balancing operation if so.
3844 *
3845 * Balancing parameters are set up in arch_init_sched_domains.
3846 */
a9957449 3847static void rebalance_domains(int cpu, enum cpu_idle_type idle)
7835b98b 3848{
46cb4b7c
SS
3849 int balance = 1;
3850 struct rq *rq = cpu_rq(cpu);
7835b98b
CL
3851 unsigned long interval;
3852 struct sched_domain *sd;
46cb4b7c 3853 /* Earliest time when we have to do rebalance again */
c9819f45 3854 unsigned long next_balance = jiffies + 60*HZ;
f549da84 3855 int update_next_balance = 0;
d07355f5 3856 int need_serialize;
7c16ec58 3857 cpumask_t tmp;
1da177e4 3858
46cb4b7c 3859 for_each_domain(cpu, sd) {
1da177e4
LT
3860 if (!(sd->flags & SD_LOAD_BALANCE))
3861 continue;
3862
3863 interval = sd->balance_interval;
d15bcfdb 3864 if (idle != CPU_IDLE)
1da177e4
LT
3865 interval *= sd->busy_factor;
3866
3867 /* scale ms to jiffies */
3868 interval = msecs_to_jiffies(interval);
3869 if (unlikely(!interval))
3870 interval = 1;
dd41f596
IM
3871 if (interval > HZ*NR_CPUS/10)
3872 interval = HZ*NR_CPUS/10;
3873
d07355f5 3874 need_serialize = sd->flags & SD_SERIALIZE;
1da177e4 3875
d07355f5 3876 if (need_serialize) {
08c183f3
CL
3877 if (!spin_trylock(&balancing))
3878 goto out;
3879 }
3880
c9819f45 3881 if (time_after_eq(jiffies, sd->last_balance + interval)) {
7c16ec58 3882 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
fa3b6ddc
SS
3883 /*
3884 * We've pulled tasks over so either we're no
5969fe06
NP
3885 * longer idle, or one of our SMT siblings is
3886 * not idle.
3887 */
d15bcfdb 3888 idle = CPU_NOT_IDLE;
1da177e4 3889 }
1bd77f2d 3890 sd->last_balance = jiffies;
1da177e4 3891 }
d07355f5 3892 if (need_serialize)
08c183f3
CL
3893 spin_unlock(&balancing);
3894out:
f549da84 3895 if (time_after(next_balance, sd->last_balance + interval)) {
c9819f45 3896 next_balance = sd->last_balance + interval;
f549da84
SS
3897 update_next_balance = 1;
3898 }
783609c6
SS
3899
3900 /*
3901 * Stop the load balance at this level. There is another
3902 * CPU in our sched group which is doing load balancing more
3903 * actively.
3904 */
3905 if (!balance)
3906 break;
1da177e4 3907 }
f549da84
SS
3908
3909 /*
3910 * next_balance will be updated only when there is a need.
3911 * When the cpu is attached to null domain for ex, it will not be
3912 * updated.
3913 */
3914 if (likely(update_next_balance))
3915 rq->next_balance = next_balance;
46cb4b7c
SS
3916}
3917
3918/*
3919 * run_rebalance_domains is triggered when needed from the scheduler tick.
3920 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3921 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3922 */
3923static void run_rebalance_domains(struct softirq_action *h)
3924{
dd41f596
IM
3925 int this_cpu = smp_processor_id();
3926 struct rq *this_rq = cpu_rq(this_cpu);
3927 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3928 CPU_IDLE : CPU_NOT_IDLE;
46cb4b7c 3929
dd41f596 3930 rebalance_domains(this_cpu, idle);
46cb4b7c
SS
3931
3932#ifdef CONFIG_NO_HZ
3933 /*
3934 * If this cpu is the owner for idle load balancing, then do the
3935 * balancing on behalf of the other idle cpus whose ticks are
3936 * stopped.
3937 */
dd41f596
IM
3938 if (this_rq->idle_at_tick &&
3939 atomic_read(&nohz.load_balancer) == this_cpu) {
46cb4b7c
SS
3940 cpumask_t cpus = nohz.cpu_mask;
3941 struct rq *rq;
3942 int balance_cpu;
3943
dd41f596 3944 cpu_clear(this_cpu, cpus);
363ab6f1 3945 for_each_cpu_mask_nr(balance_cpu, cpus) {
46cb4b7c
SS
3946 /*
3947 * If this cpu gets work to do, stop the load balancing
3948 * work being done for other cpus. Next load
3949 * balancing owner will pick it up.
3950 */
3951 if (need_resched())
3952 break;
3953
de0cf899 3954 rebalance_domains(balance_cpu, CPU_IDLE);
46cb4b7c
SS
3955
3956 rq = cpu_rq(balance_cpu);
dd41f596
IM
3957 if (time_after(this_rq->next_balance, rq->next_balance))
3958 this_rq->next_balance = rq->next_balance;
46cb4b7c
SS
3959 }
3960 }
3961#endif
3962}
3963
3964/*
3965 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3966 *
3967 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3968 * idle load balancing owner or decide to stop the periodic load balancing,
3969 * if the whole system is idle.
3970 */
dd41f596 3971static inline void trigger_load_balance(struct rq *rq, int cpu)
46cb4b7c 3972{
46cb4b7c
SS
3973#ifdef CONFIG_NO_HZ
3974 /*
3975 * If we were in the nohz mode recently and busy at the current
3976 * scheduler tick, then check if we need to nominate new idle
3977 * load balancer.
3978 */
3979 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3980 rq->in_nohz_recently = 0;
3981
3982 if (atomic_read(&nohz.load_balancer) == cpu) {
3983 cpu_clear(cpu, nohz.cpu_mask);
3984 atomic_set(&nohz.load_balancer, -1);
3985 }
3986
3987 if (atomic_read(&nohz.load_balancer) == -1) {
3988 /*
3989 * simple selection for now: Nominate the
3990 * first cpu in the nohz list to be the next
3991 * ilb owner.
3992 *
3993 * TBD: Traverse the sched domains and nominate
3994 * the nearest cpu in the nohz.cpu_mask.
3995 */
3996 int ilb = first_cpu(nohz.cpu_mask);
3997
434d53b0 3998 if (ilb < nr_cpu_ids)
46cb4b7c
SS
3999 resched_cpu(ilb);
4000 }
4001 }
4002
4003 /*
4004 * If this cpu is idle and doing idle load balancing for all the
4005 * cpus with ticks stopped, is it time for that to stop?
4006 */
4007 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4008 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4009 resched_cpu(cpu);
4010 return;
4011 }
4012
4013 /*
4014 * If this cpu is idle and the idle load balancing is done by
4015 * someone else, then no need raise the SCHED_SOFTIRQ
4016 */
4017 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4018 cpu_isset(cpu, nohz.cpu_mask))
4019 return;
4020#endif
4021 if (time_after_eq(jiffies, rq->next_balance))
4022 raise_softirq(SCHED_SOFTIRQ);
1da177e4 4023}
dd41f596
IM
4024
4025#else /* CONFIG_SMP */
4026
1da177e4
LT
4027/*
4028 * on UP we do not need to balance between CPUs:
4029 */
70b97a7f 4030static inline void idle_balance(int cpu, struct rq *rq)
1da177e4
LT
4031{
4032}
dd41f596 4033
1da177e4
LT
4034#endif
4035
1da177e4
LT
4036DEFINE_PER_CPU(struct kernel_stat, kstat);
4037
4038EXPORT_PER_CPU_SYMBOL(kstat);
4039
4040/*
f06febc9
FM
4041 * Return any ns on the sched_clock that have not yet been banked in
4042 * @p in case that task is currently running.
1da177e4 4043 */
bb34d92f 4044unsigned long long task_delta_exec(struct task_struct *p)
1da177e4 4045{
1da177e4 4046 unsigned long flags;
41b86e9c 4047 struct rq *rq;
bb34d92f 4048 u64 ns = 0;
48f24c4d 4049
41b86e9c 4050 rq = task_rq_lock(p, &flags);
1508487e 4051
051a1d1a 4052 if (task_current(rq, p)) {
f06febc9
FM
4053 u64 delta_exec;
4054
a8e504d2
IM
4055 update_rq_clock(rq);
4056 delta_exec = rq->clock - p->se.exec_start;
41b86e9c 4057 if ((s64)delta_exec > 0)
bb34d92f 4058 ns = delta_exec;
41b86e9c 4059 }
48f24c4d 4060
41b86e9c 4061 task_rq_unlock(rq, &flags);
48f24c4d 4062
1da177e4
LT
4063 return ns;
4064}
4065
1da177e4
LT
4066/*
4067 * Account user cpu time to a process.
4068 * @p: the process that the cpu time gets accounted to
1da177e4
LT
4069 * @cputime: the cpu time spent in user space since the last update
4070 */
4071void account_user_time(struct task_struct *p, cputime_t cputime)
4072{
4073 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4074 cputime64_t tmp;
4075
4076 p->utime = cputime_add(p->utime, cputime);
f06febc9 4077 account_group_user_time(p, cputime);
1da177e4
LT
4078
4079 /* Add user time to cpustat. */
4080 tmp = cputime_to_cputime64(cputime);
4081 if (TASK_NICE(p) > 0)
4082 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4083 else
4084 cpustat->user = cputime64_add(cpustat->user, tmp);
49b5cf34
JL
4085 /* Account for user time used */
4086 acct_update_integrals(p);
1da177e4
LT
4087}
4088
94886b84
LV
4089/*
4090 * Account guest cpu time to a process.
4091 * @p: the process that the cpu time gets accounted to
4092 * @cputime: the cpu time spent in virtual machine since the last update
4093 */
f7402e03 4094static void account_guest_time(struct task_struct *p, cputime_t cputime)
94886b84
LV
4095{
4096 cputime64_t tmp;
4097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4098
4099 tmp = cputime_to_cputime64(cputime);
4100
4101 p->utime = cputime_add(p->utime, cputime);
f06febc9 4102 account_group_user_time(p, cputime);
94886b84
LV
4103 p->gtime = cputime_add(p->gtime, cputime);
4104
4105 cpustat->user = cputime64_add(cpustat->user, tmp);
4106 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4107}
4108
c66f08be
MN
4109/*
4110 * Account scaled user cpu time to a process.
4111 * @p: the process that the cpu time gets accounted to
4112 * @cputime: the cpu time spent in user space since the last update
4113 */
4114void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4115{
4116 p->utimescaled = cputime_add(p->utimescaled, cputime);
4117}
4118
1da177e4
LT
4119/*
4120 * Account system cpu time to a process.
4121 * @p: the process that the cpu time gets accounted to
4122 * @hardirq_offset: the offset to subtract from hardirq_count()
4123 * @cputime: the cpu time spent in kernel space since the last update
4124 */
4125void account_system_time(struct task_struct *p, int hardirq_offset,
4126 cputime_t cputime)
4127{
4128 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
70b97a7f 4129 struct rq *rq = this_rq();
1da177e4
LT
4130 cputime64_t tmp;
4131
983ed7a6
HH
4132 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4133 account_guest_time(p, cputime);
4134 return;
4135 }
94886b84 4136
1da177e4 4137 p->stime = cputime_add(p->stime, cputime);
f06febc9 4138 account_group_system_time(p, cputime);
1da177e4
LT
4139
4140 /* Add system time to cpustat. */
4141 tmp = cputime_to_cputime64(cputime);
4142 if (hardirq_count() - hardirq_offset)
4143 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4144 else if (softirq_count())
4145 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
cfb52856 4146 else if (p != rq->idle)
1da177e4 4147 cpustat->system = cputime64_add(cpustat->system, tmp);
cfb52856 4148 else if (atomic_read(&rq->nr_iowait) > 0)
1da177e4
LT
4149 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4150 else
4151 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4152 /* Account for system time used */
4153 acct_update_integrals(p);
1da177e4
LT
4154}
4155
c66f08be
MN
4156/*
4157 * Account scaled system cpu time to a process.
4158 * @p: the process that the cpu time gets accounted to
4159 * @hardirq_offset: the offset to subtract from hardirq_count()
4160 * @cputime: the cpu time spent in kernel space since the last update
4161 */
4162void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4163{
4164 p->stimescaled = cputime_add(p->stimescaled, cputime);
4165}
4166
1da177e4
LT
4167/*
4168 * Account for involuntary wait time.
4169 * @p: the process from which the cpu time has been stolen
4170 * @steal: the cpu time spent in involuntary wait
4171 */
4172void account_steal_time(struct task_struct *p, cputime_t steal)
4173{
4174 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4175 cputime64_t tmp = cputime_to_cputime64(steal);
70b97a7f 4176 struct rq *rq = this_rq();
1da177e4
LT
4177
4178 if (p == rq->idle) {
4179 p->stime = cputime_add(p->stime, steal);
4180 if (atomic_read(&rq->nr_iowait) > 0)
4181 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4182 else
4183 cpustat->idle = cputime64_add(cpustat->idle, tmp);
cfb52856 4184 } else
1da177e4
LT
4185 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4186}
4187
49048622
BS
4188/*
4189 * Use precise platform statistics if available:
4190 */
4191#ifdef CONFIG_VIRT_CPU_ACCOUNTING
4192cputime_t task_utime(struct task_struct *p)
4193{
4194 return p->utime;
4195}
4196
4197cputime_t task_stime(struct task_struct *p)
4198{
4199 return p->stime;
4200}
4201#else
4202cputime_t task_utime(struct task_struct *p)
4203{
4204 clock_t utime = cputime_to_clock_t(p->utime),
4205 total = utime + cputime_to_clock_t(p->stime);
4206 u64 temp;
4207
4208 /*
4209 * Use CFS's precise accounting:
4210 */
4211 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4212
4213 if (total) {
4214 temp *= utime;
4215 do_div(temp, total);
4216 }
4217 utime = (clock_t)temp;
4218
4219 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4220 return p->prev_utime;
4221}
4222
4223cputime_t task_stime(struct task_struct *p)
4224{
4225 clock_t stime;
4226
4227 /*
4228 * Use CFS's precise accounting. (we subtract utime from
4229 * the total, to make sure the total observed by userspace
4230 * grows monotonically - apps rely on that):
4231 */
4232 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4233 cputime_to_clock_t(task_utime(p));
4234
4235 if (stime >= 0)
4236 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4237
4238 return p->prev_stime;
4239}
4240#endif
4241
4242inline cputime_t task_gtime(struct task_struct *p)
4243{
4244 return p->gtime;
4245}
4246
7835b98b
CL
4247/*
4248 * This function gets called by the timer code, with HZ frequency.
4249 * We call it with interrupts disabled.
4250 *
4251 * It also gets called by the fork code, when changing the parent's
4252 * timeslices.
4253 */
4254void scheduler_tick(void)
4255{
7835b98b
CL
4256 int cpu = smp_processor_id();
4257 struct rq *rq = cpu_rq(cpu);
dd41f596 4258 struct task_struct *curr = rq->curr;
3e51f33f
PZ
4259
4260 sched_clock_tick();
dd41f596
IM
4261
4262 spin_lock(&rq->lock);
3e51f33f 4263 update_rq_clock(rq);
f1a438d8 4264 update_cpu_load(rq);
fa85ae24 4265 curr->sched_class->task_tick(rq, curr, 0);
dd41f596 4266 spin_unlock(&rq->lock);
7835b98b 4267
e418e1c2 4268#ifdef CONFIG_SMP
dd41f596
IM
4269 rq->idle_at_tick = idle_cpu(cpu);
4270 trigger_load_balance(rq, cpu);
e418e1c2 4271#endif
1da177e4
LT
4272}
4273
6cd8a4bb
SR
4274#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4275 defined(CONFIG_PREEMPT_TRACER))
4276
4277static inline unsigned long get_parent_ip(unsigned long addr)
4278{
4279 if (in_lock_functions(addr)) {
4280 addr = CALLER_ADDR2;
4281 if (in_lock_functions(addr))
4282 addr = CALLER_ADDR3;
4283 }
4284 return addr;
4285}
1da177e4 4286
43627582 4287void __kprobes add_preempt_count(int val)
1da177e4 4288{
6cd8a4bb 4289#ifdef CONFIG_DEBUG_PREEMPT
1da177e4
LT
4290 /*
4291 * Underflow?
4292 */
9a11b49a
IM
4293 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4294 return;
6cd8a4bb 4295#endif
1da177e4 4296 preempt_count() += val;
6cd8a4bb 4297#ifdef CONFIG_DEBUG_PREEMPT
1da177e4
LT
4298 /*
4299 * Spinlock count overflowing soon?
4300 */
33859f7f
MOS
4301 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4302 PREEMPT_MASK - 10);
6cd8a4bb
SR
4303#endif
4304 if (preempt_count() == val)
4305 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
1da177e4
LT
4306}
4307EXPORT_SYMBOL(add_preempt_count);
4308
43627582 4309void __kprobes sub_preempt_count(int val)
1da177e4 4310{
6cd8a4bb 4311#ifdef CONFIG_DEBUG_PREEMPT
1da177e4
LT
4312 /*
4313 * Underflow?
4314 */
7317d7b8 4315 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
9a11b49a 4316 return;
1da177e4
LT
4317 /*
4318 * Is the spinlock portion underflowing?
4319 */
9a11b49a
IM
4320 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4321 !(preempt_count() & PREEMPT_MASK)))
4322 return;
6cd8a4bb 4323#endif
9a11b49a 4324
6cd8a4bb
SR
4325 if (preempt_count() == val)
4326 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
1da177e4
LT
4327 preempt_count() -= val;
4328}
4329EXPORT_SYMBOL(sub_preempt_count);
4330
4331#endif
4332
4333/*
dd41f596 4334 * Print scheduling while atomic bug:
1da177e4 4335 */
dd41f596 4336static noinline void __schedule_bug(struct task_struct *prev)
1da177e4 4337{
838225b4
SS
4338 struct pt_regs *regs = get_irq_regs();
4339
4340 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4341 prev->comm, prev->pid, preempt_count());
4342
dd41f596 4343 debug_show_held_locks(prev);
e21f5b15 4344 print_modules();
dd41f596
IM
4345 if (irqs_disabled())
4346 print_irqtrace_events(prev);
838225b4
SS
4347
4348 if (regs)
4349 show_regs(regs);
4350 else
4351 dump_stack();
dd41f596 4352}
1da177e4 4353
dd41f596
IM
4354/*
4355 * Various schedule()-time debugging checks and statistics:
4356 */
4357static inline void schedule_debug(struct task_struct *prev)
4358{
1da177e4 4359 /*
41a2d6cf 4360 * Test if we are atomic. Since do_exit() needs to call into
1da177e4
LT
4361 * schedule() atomically, we ignore that path for now.
4362 * Otherwise, whine if we are scheduling when we should not be.
4363 */
3f33a7ce 4364 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
dd41f596
IM
4365 __schedule_bug(prev);
4366
1da177e4
LT
4367 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4368
2d72376b 4369 schedstat_inc(this_rq(), sched_count);
b8efb561
IM
4370#ifdef CONFIG_SCHEDSTATS
4371 if (unlikely(prev->lock_depth >= 0)) {
2d72376b
IM
4372 schedstat_inc(this_rq(), bkl_count);
4373 schedstat_inc(prev, sched_info.bkl_count);
b8efb561
IM
4374 }
4375#endif
dd41f596
IM
4376}
4377
4378/*
4379 * Pick up the highest-prio task:
4380 */
4381static inline struct task_struct *
ff95f3df 4382pick_next_task(struct rq *rq, struct task_struct *prev)
dd41f596 4383{
5522d5d5 4384 const struct sched_class *class;
dd41f596 4385 struct task_struct *p;
1da177e4
LT
4386
4387 /*
dd41f596
IM
4388 * Optimization: we know that if all tasks are in
4389 * the fair class we can call that function directly:
1da177e4 4390 */
dd41f596 4391 if (likely(rq->nr_running == rq->cfs.nr_running)) {
fb8d4724 4392 p = fair_sched_class.pick_next_task(rq);
dd41f596
IM
4393 if (likely(p))
4394 return p;
1da177e4
LT
4395 }
4396
dd41f596
IM
4397 class = sched_class_highest;
4398 for ( ; ; ) {
fb8d4724 4399 p = class->pick_next_task(rq);
dd41f596
IM
4400 if (p)
4401 return p;
4402 /*
4403 * Will never be NULL as the idle class always
4404 * returns a non-NULL p:
4405 */
4406 class = class->next;
4407 }
4408}
1da177e4 4409
dd41f596
IM
4410/*
4411 * schedule() is the main scheduler function.
4412 */
4413asmlinkage void __sched schedule(void)
4414{
4415 struct task_struct *prev, *next;
67ca7bde 4416 unsigned long *switch_count;
dd41f596 4417 struct rq *rq;
31656519 4418 int cpu;
dd41f596
IM
4419
4420need_resched:
4421 preempt_disable();
4422 cpu = smp_processor_id();
4423 rq = cpu_rq(cpu);
4424 rcu_qsctr_inc(cpu);
4425 prev = rq->curr;
4426 switch_count = &prev->nivcsw;
4427
4428 release_kernel_lock(prev);
4429need_resched_nonpreemptible:
4430
4431 schedule_debug(prev);
1da177e4 4432
31656519 4433 if (sched_feat(HRTICK))
f333fdc9 4434 hrtick_clear(rq);
8f4d37ec 4435
8cd162ce 4436 spin_lock_irq(&rq->lock);
3e51f33f 4437 update_rq_clock(rq);
1e819950 4438 clear_tsk_need_resched(prev);
1da177e4 4439
1da177e4 4440 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
16882c1e 4441 if (unlikely(signal_pending_state(prev->state, prev)))
1da177e4 4442 prev->state = TASK_RUNNING;
16882c1e 4443 else
2e1cb74a 4444 deactivate_task(rq, prev, 1);
dd41f596 4445 switch_count = &prev->nvcsw;
1da177e4
LT
4446 }
4447
9a897c5a
SR
4448#ifdef CONFIG_SMP
4449 if (prev->sched_class->pre_schedule)
4450 prev->sched_class->pre_schedule(rq, prev);
4451#endif
f65eda4f 4452
dd41f596 4453 if (unlikely(!rq->nr_running))
1da177e4 4454 idle_balance(cpu, rq);
1da177e4 4455
31ee529c 4456 prev->sched_class->put_prev_task(rq, prev);
ff95f3df 4457 next = pick_next_task(rq, prev);
1da177e4 4458
1da177e4 4459 if (likely(prev != next)) {
673a90a1
DS
4460 sched_info_switch(prev, next);
4461
1da177e4
LT
4462 rq->nr_switches++;
4463 rq->curr = next;
4464 ++*switch_count;
4465
dd41f596 4466 context_switch(rq, prev, next); /* unlocks the rq */
8f4d37ec
PZ
4467 /*
4468 * the context switch might have flipped the stack from under
4469 * us, hence refresh the local variables.
4470 */
4471 cpu = smp_processor_id();
4472 rq = cpu_rq(cpu);
1da177e4
LT
4473 } else
4474 spin_unlock_irq(&rq->lock);
4475
8f4d37ec 4476 if (unlikely(reacquire_kernel_lock(current) < 0))
1da177e4 4477 goto need_resched_nonpreemptible;
8f4d37ec 4478
1da177e4
LT
4479 preempt_enable_no_resched();
4480 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4481 goto need_resched;
4482}
1da177e4
LT
4483EXPORT_SYMBOL(schedule);
4484
4485#ifdef CONFIG_PREEMPT
4486/*
2ed6e34f 4487 * this is the entry point to schedule() from in-kernel preemption
41a2d6cf 4488 * off of preempt_enable. Kernel preemptions off return from interrupt
1da177e4
LT
4489 * occur there and call schedule directly.
4490 */
4491asmlinkage void __sched preempt_schedule(void)
4492{
4493 struct thread_info *ti = current_thread_info();
6478d880 4494
1da177e4
LT
4495 /*
4496 * If there is a non-zero preempt_count or interrupts are disabled,
41a2d6cf 4497 * we do not want to preempt the current task. Just return..
1da177e4 4498 */
beed33a8 4499 if (likely(ti->preempt_count || irqs_disabled()))
1da177e4
LT
4500 return;
4501
3a5c359a
AK
4502 do {
4503 add_preempt_count(PREEMPT_ACTIVE);
3a5c359a 4504 schedule();
3a5c359a 4505 sub_preempt_count(PREEMPT_ACTIVE);
1da177e4 4506
3a5c359a
AK
4507 /*
4508 * Check again in case we missed a preemption opportunity
4509 * between schedule and now.
4510 */
4511 barrier();
4512 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
1da177e4 4513}
1da177e4
LT
4514EXPORT_SYMBOL(preempt_schedule);
4515
4516/*
2ed6e34f 4517 * this is the entry point to schedule() from kernel preemption
1da177e4
LT
4518 * off of irq context.
4519 * Note, that this is called and return with irqs disabled. This will
4520 * protect us against recursive calling from irq.
4521 */
4522asmlinkage void __sched preempt_schedule_irq(void)
4523{
4524 struct thread_info *ti = current_thread_info();
6478d880 4525
2ed6e34f 4526 /* Catch callers which need to be fixed */
1da177e4
LT
4527 BUG_ON(ti->preempt_count || !irqs_disabled());
4528
3a5c359a
AK
4529 do {
4530 add_preempt_count(PREEMPT_ACTIVE);
3a5c359a
AK
4531 local_irq_enable();
4532 schedule();
4533 local_irq_disable();
3a5c359a 4534 sub_preempt_count(PREEMPT_ACTIVE);
1da177e4 4535
3a5c359a
AK
4536 /*
4537 * Check again in case we missed a preemption opportunity
4538 * between schedule and now.
4539 */
4540 barrier();
4541 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
1da177e4
LT
4542}
4543
4544#endif /* CONFIG_PREEMPT */
4545
95cdf3b7
IM
4546int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4547 void *key)
1da177e4 4548{
48f24c4d 4549 return try_to_wake_up(curr->private, mode, sync);
1da177e4 4550}
1da177e4
LT
4551EXPORT_SYMBOL(default_wake_function);
4552
4553/*
41a2d6cf
IM
4554 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4555 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
1da177e4
LT
4556 * number) then we wake all the non-exclusive tasks and one exclusive task.
4557 *
4558 * There are circumstances in which we can try to wake a task which has already
41a2d6cf 4559 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
1da177e4
LT
4560 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4561 */
4562static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4563 int nr_exclusive, int sync, void *key)
4564{
2e45874c 4565 wait_queue_t *curr, *next;
1da177e4 4566
2e45874c 4567 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
48f24c4d
IM
4568 unsigned flags = curr->flags;
4569
1da177e4 4570 if (curr->func(curr, mode, sync, key) &&
48f24c4d 4571 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
1da177e4
LT
4572 break;
4573 }
4574}
4575
4576/**
4577 * __wake_up - wake up threads blocked on a waitqueue.
4578 * @q: the waitqueue
4579 * @mode: which threads
4580 * @nr_exclusive: how many wake-one or wake-many threads to wake up
67be2dd1 4581 * @key: is directly passed to the wakeup function
1da177e4 4582 */
7ad5b3a5 4583void __wake_up(wait_queue_head_t *q, unsigned int mode,
95cdf3b7 4584 int nr_exclusive, void *key)
1da177e4
LT
4585{
4586 unsigned long flags;
4587
4588 spin_lock_irqsave(&q->lock, flags);
4589 __wake_up_common(q, mode, nr_exclusive, 0, key);
4590 spin_unlock_irqrestore(&q->lock, flags);
4591}
1da177e4
LT
4592EXPORT_SYMBOL(__wake_up);
4593
4594/*
4595 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4596 */
7ad5b3a5 4597void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
1da177e4
LT
4598{
4599 __wake_up_common(q, mode, 1, 0, NULL);
4600}
4601
4602/**
67be2dd1 4603 * __wake_up_sync - wake up threads blocked on a waitqueue.
1da177e4
LT
4604 * @q: the waitqueue
4605 * @mode: which threads
4606 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4607 *
4608 * The sync wakeup differs that the waker knows that it will schedule
4609 * away soon, so while the target thread will be woken up, it will not
4610 * be migrated to another CPU - ie. the two threads are 'synchronized'
4611 * with each other. This can prevent needless bouncing between CPUs.
4612 *
4613 * On UP it can prevent extra preemption.
4614 */
7ad5b3a5 4615void
95cdf3b7 4616__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
1da177e4
LT
4617{
4618 unsigned long flags;
4619 int sync = 1;
4620
4621 if (unlikely(!q))
4622 return;
4623
4624 if (unlikely(!nr_exclusive))
4625 sync = 0;
4626
4627 spin_lock_irqsave(&q->lock, flags);
4628 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4629 spin_unlock_irqrestore(&q->lock, flags);
4630}
4631EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4632
65eb3dc6
KD
4633/**
4634 * complete: - signals a single thread waiting on this completion
4635 * @x: holds the state of this particular completion
4636 *
4637 * This will wake up a single thread waiting on this completion. Threads will be
4638 * awakened in the same order in which they were queued.
4639 *
4640 * See also complete_all(), wait_for_completion() and related routines.
4641 */
b15136e9 4642void complete(struct completion *x)
1da177e4
LT
4643{
4644 unsigned long flags;
4645
4646 spin_lock_irqsave(&x->wait.lock, flags);
4647 x->done++;
d9514f6c 4648 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
1da177e4
LT
4649 spin_unlock_irqrestore(&x->wait.lock, flags);
4650}
4651EXPORT_SYMBOL(complete);
4652
65eb3dc6
KD
4653/**
4654 * complete_all: - signals all threads waiting on this completion
4655 * @x: holds the state of this particular completion
4656 *
4657 * This will wake up all threads waiting on this particular completion event.
4658 */
b15136e9 4659void complete_all(struct completion *x)
1da177e4
LT
4660{
4661 unsigned long flags;
4662
4663 spin_lock_irqsave(&x->wait.lock, flags);
4664 x->done += UINT_MAX/2;
d9514f6c 4665 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
1da177e4
LT
4666 spin_unlock_irqrestore(&x->wait.lock, flags);
4667}
4668EXPORT_SYMBOL(complete_all);
4669
8cbbe86d
AK
4670static inline long __sched
4671do_wait_for_common(struct completion *x, long timeout, int state)
1da177e4 4672{
1da177e4
LT
4673 if (!x->done) {
4674 DECLARE_WAITQUEUE(wait, current);
4675
4676 wait.flags |= WQ_FLAG_EXCLUSIVE;
4677 __add_wait_queue_tail(&x->wait, &wait);
4678 do {
94d3d824 4679 if (signal_pending_state(state, current)) {
ea71a546
ON
4680 timeout = -ERESTARTSYS;
4681 break;
8cbbe86d
AK
4682 }
4683 __set_current_state(state);
1da177e4
LT
4684 spin_unlock_irq(&x->wait.lock);
4685 timeout = schedule_timeout(timeout);
4686 spin_lock_irq(&x->wait.lock);
ea71a546 4687 } while (!x->done && timeout);
1da177e4 4688 __remove_wait_queue(&x->wait, &wait);
ea71a546
ON
4689 if (!x->done)
4690 return timeout;
1da177e4
LT
4691 }
4692 x->done--;
ea71a546 4693 return timeout ?: 1;
1da177e4 4694}
1da177e4 4695
8cbbe86d
AK
4696static long __sched
4697wait_for_common(struct completion *x, long timeout, int state)
1da177e4 4698{
1da177e4
LT
4699 might_sleep();
4700
4701 spin_lock_irq(&x->wait.lock);
8cbbe86d 4702 timeout = do_wait_for_common(x, timeout, state);
1da177e4 4703 spin_unlock_irq(&x->wait.lock);
8cbbe86d
AK
4704 return timeout;
4705}
1da177e4 4706
65eb3dc6
KD
4707/**
4708 * wait_for_completion: - waits for completion of a task
4709 * @x: holds the state of this particular completion
4710 *
4711 * This waits to be signaled for completion of a specific task. It is NOT
4712 * interruptible and there is no timeout.
4713 *
4714 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4715 * and interrupt capability. Also see complete().
4716 */
b15136e9 4717void __sched wait_for_completion(struct completion *x)
8cbbe86d
AK
4718{
4719 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
1da177e4 4720}
8cbbe86d 4721EXPORT_SYMBOL(wait_for_completion);
1da177e4 4722
65eb3dc6
KD
4723/**
4724 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4725 * @x: holds the state of this particular completion
4726 * @timeout: timeout value in jiffies
4727 *
4728 * This waits for either a completion of a specific task to be signaled or for a
4729 * specified timeout to expire. The timeout is in jiffies. It is not
4730 * interruptible.
4731 */
b15136e9 4732unsigned long __sched
8cbbe86d 4733wait_for_completion_timeout(struct completion *x, unsigned long timeout)
1da177e4 4734{
8cbbe86d 4735 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
1da177e4 4736}
8cbbe86d 4737EXPORT_SYMBOL(wait_for_completion_timeout);
1da177e4 4738
65eb3dc6
KD
4739/**
4740 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4741 * @x: holds the state of this particular completion
4742 *
4743 * This waits for completion of a specific task to be signaled. It is
4744 * interruptible.
4745 */
8cbbe86d 4746int __sched wait_for_completion_interruptible(struct completion *x)
0fec171c 4747{
51e97990
AK
4748 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4749 if (t == -ERESTARTSYS)
4750 return t;
4751 return 0;
0fec171c 4752}
8cbbe86d 4753EXPORT_SYMBOL(wait_for_completion_interruptible);
1da177e4 4754
65eb3dc6
KD
4755/**
4756 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4757 * @x: holds the state of this particular completion
4758 * @timeout: timeout value in jiffies
4759 *
4760 * This waits for either a completion of a specific task to be signaled or for a
4761 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4762 */
b15136e9 4763unsigned long __sched
8cbbe86d
AK
4764wait_for_completion_interruptible_timeout(struct completion *x,
4765 unsigned long timeout)
0fec171c 4766{
8cbbe86d 4767 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
0fec171c 4768}
8cbbe86d 4769EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
1da177e4 4770
65eb3dc6
KD
4771/**
4772 * wait_for_completion_killable: - waits for completion of a task (killable)
4773 * @x: holds the state of this particular completion
4774 *
4775 * This waits to be signaled for completion of a specific task. It can be
4776 * interrupted by a kill signal.
4777 */
009e577e
MW
4778int __sched wait_for_completion_killable(struct completion *x)
4779{
4780 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4781 if (t == -ERESTARTSYS)
4782 return t;
4783 return 0;
4784}
4785EXPORT_SYMBOL(wait_for_completion_killable);
4786
be4de352
DC
4787/**
4788 * try_wait_for_completion - try to decrement a completion without blocking
4789 * @x: completion structure
4790 *
4791 * Returns: 0 if a decrement cannot be done without blocking
4792 * 1 if a decrement succeeded.
4793 *
4794 * If a completion is being used as a counting completion,
4795 * attempt to decrement the counter without blocking. This
4796 * enables us to avoid waiting if the resource the completion
4797 * is protecting is not available.
4798 */
4799bool try_wait_for_completion(struct completion *x)
4800{
4801 int ret = 1;
4802
4803 spin_lock_irq(&x->wait.lock);
4804 if (!x->done)
4805 ret = 0;
4806 else
4807 x->done--;
4808 spin_unlock_irq(&x->wait.lock);
4809 return ret;
4810}
4811EXPORT_SYMBOL(try_wait_for_completion);
4812
4813/**
4814 * completion_done - Test to see if a completion has any waiters
4815 * @x: completion structure
4816 *
4817 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4818 * 1 if there are no waiters.
4819 *
4820 */
4821bool completion_done(struct completion *x)
4822{
4823 int ret = 1;
4824
4825 spin_lock_irq(&x->wait.lock);
4826 if (!x->done)
4827 ret = 0;
4828 spin_unlock_irq(&x->wait.lock);
4829 return ret;
4830}
4831EXPORT_SYMBOL(completion_done);
4832
8cbbe86d
AK
4833static long __sched
4834sleep_on_common(wait_queue_head_t *q, int state, long timeout)
1da177e4 4835{
0fec171c
IM
4836 unsigned long flags;
4837 wait_queue_t wait;
4838
4839 init_waitqueue_entry(&wait, current);
1da177e4 4840
8cbbe86d 4841 __set_current_state(state);
1da177e4 4842
8cbbe86d
AK
4843 spin_lock_irqsave(&q->lock, flags);
4844 __add_wait_queue(q, &wait);
4845 spin_unlock(&q->lock);
4846 timeout = schedule_timeout(timeout);
4847 spin_lock_irq(&q->lock);
4848 __remove_wait_queue(q, &wait);
4849 spin_unlock_irqrestore(&q->lock, flags);
4850
4851 return timeout;
4852}
4853
4854void __sched interruptible_sleep_on(wait_queue_head_t *q)
4855{
4856 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
1da177e4 4857}
1da177e4
LT
4858EXPORT_SYMBOL(interruptible_sleep_on);
4859
0fec171c 4860long __sched
95cdf3b7 4861interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
1da177e4 4862{
8cbbe86d 4863 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
1da177e4 4864}
1da177e4
LT
4865EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4866
0fec171c 4867void __sched sleep_on(wait_queue_head_t *q)
1da177e4 4868{
8cbbe86d 4869 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
1da177e4 4870}
1da177e4
LT
4871EXPORT_SYMBOL(sleep_on);
4872
0fec171c 4873long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
1da177e4 4874{
8cbbe86d 4875 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
1da177e4 4876}
1da177e4
LT
4877EXPORT_SYMBOL(sleep_on_timeout);
4878
b29739f9
IM
4879#ifdef CONFIG_RT_MUTEXES
4880
4881/*
4882 * rt_mutex_setprio - set the current priority of a task
4883 * @p: task
4884 * @prio: prio value (kernel-internal form)
4885 *
4886 * This function changes the 'effective' priority of a task. It does
4887 * not touch ->normal_prio like __setscheduler().
4888 *
4889 * Used by the rt_mutex code to implement priority inheritance logic.
4890 */
36c8b586 4891void rt_mutex_setprio(struct task_struct *p, int prio)
b29739f9
IM
4892{
4893 unsigned long flags;
83b699ed 4894 int oldprio, on_rq, running;
70b97a7f 4895 struct rq *rq;
cb469845 4896 const struct sched_class *prev_class = p->sched_class;
b29739f9
IM
4897
4898 BUG_ON(prio < 0 || prio > MAX_PRIO);
4899
4900 rq = task_rq_lock(p, &flags);
a8e504d2 4901 update_rq_clock(rq);
b29739f9 4902
d5f9f942 4903 oldprio = p->prio;
dd41f596 4904 on_rq = p->se.on_rq;
051a1d1a 4905 running = task_current(rq, p);
0e1f3483 4906 if (on_rq)
69be72c1 4907 dequeue_task(rq, p, 0);
0e1f3483
HS
4908 if (running)
4909 p->sched_class->put_prev_task(rq, p);
dd41f596
IM
4910
4911 if (rt_prio(prio))
4912 p->sched_class = &rt_sched_class;
4913 else
4914 p->sched_class = &fair_sched_class;
4915
b29739f9
IM
4916 p->prio = prio;
4917
0e1f3483
HS
4918 if (running)
4919 p->sched_class->set_curr_task(rq);
dd41f596 4920 if (on_rq) {
8159f87e 4921 enqueue_task(rq, p, 0);
cb469845
SR
4922
4923 check_class_changed(rq, p, prev_class, oldprio, running);
b29739f9
IM
4924 }
4925 task_rq_unlock(rq, &flags);
4926}
4927
4928#endif
4929
36c8b586 4930void set_user_nice(struct task_struct *p, long nice)
1da177e4 4931{
dd41f596 4932 int old_prio, delta, on_rq;
1da177e4 4933 unsigned long flags;
70b97a7f 4934 struct rq *rq;
1da177e4
LT
4935
4936 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4937 return;
4938 /*
4939 * We have to be careful, if called from sys_setpriority(),
4940 * the task might be in the middle of scheduling on another CPU.
4941 */
4942 rq = task_rq_lock(p, &flags);
a8e504d2 4943 update_rq_clock(rq);
1da177e4
LT
4944 /*
4945 * The RT priorities are set via sched_setscheduler(), but we still
4946 * allow the 'normal' nice value to be set - but as expected
4947 * it wont have any effect on scheduling until the task is
dd41f596 4948 * SCHED_FIFO/SCHED_RR:
1da177e4 4949 */
e05606d3 4950 if (task_has_rt_policy(p)) {
1da177e4
LT
4951 p->static_prio = NICE_TO_PRIO(nice);
4952 goto out_unlock;
4953 }
dd41f596 4954 on_rq = p->se.on_rq;
c09595f6 4955 if (on_rq)
69be72c1 4956 dequeue_task(rq, p, 0);
1da177e4 4957
1da177e4 4958 p->static_prio = NICE_TO_PRIO(nice);
2dd73a4f 4959 set_load_weight(p);
b29739f9
IM
4960 old_prio = p->prio;
4961 p->prio = effective_prio(p);
4962 delta = p->prio - old_prio;
1da177e4 4963
dd41f596 4964 if (on_rq) {
8159f87e 4965 enqueue_task(rq, p, 0);
1da177e4 4966 /*
d5f9f942
AM
4967 * If the task increased its priority or is running and
4968 * lowered its priority, then reschedule its CPU:
1da177e4 4969 */
d5f9f942 4970 if (delta < 0 || (delta > 0 && task_running(rq, p)))
1da177e4
LT
4971 resched_task(rq->curr);
4972 }
4973out_unlock:
4974 task_rq_unlock(rq, &flags);
4975}
1da177e4
LT
4976EXPORT_SYMBOL(set_user_nice);
4977
e43379f1
MM
4978/*
4979 * can_nice - check if a task can reduce its nice value
4980 * @p: task
4981 * @nice: nice value
4982 */
36c8b586 4983int can_nice(const struct task_struct *p, const int nice)
e43379f1 4984{
024f4747
MM
4985 /* convert nice value [19,-20] to rlimit style value [1,40] */
4986 int nice_rlim = 20 - nice;
48f24c4d 4987
e43379f1
MM
4988 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4989 capable(CAP_SYS_NICE));
4990}
4991
1da177e4
LT
4992#ifdef __ARCH_WANT_SYS_NICE
4993
4994/*
4995 * sys_nice - change the priority of the current process.
4996 * @increment: priority increment
4997 *
4998 * sys_setpriority is a more generic, but much slower function that
4999 * does similar things.
5000 */
5001asmlinkage long sys_nice(int increment)
5002{
48f24c4d 5003 long nice, retval;
1da177e4
LT
5004
5005 /*
5006 * Setpriority might change our priority at the same moment.
5007 * We don't have to worry. Conceptually one call occurs first
5008 * and we have a single winner.
5009 */
e43379f1
MM
5010 if (increment < -40)
5011 increment = -40;
1da177e4
LT
5012 if (increment > 40)
5013 increment = 40;
5014
5015 nice = PRIO_TO_NICE(current->static_prio) + increment;
5016 if (nice < -20)
5017 nice = -20;
5018 if (nice > 19)
5019 nice = 19;
5020
e43379f1
MM
5021 if (increment < 0 && !can_nice(current, nice))
5022 return -EPERM;
5023
1da177e4
LT
5024 retval = security_task_setnice(current, nice);
5025 if (retval)
5026 return retval;
5027
5028 set_user_nice(current, nice);
5029 return 0;
5030}
5031
5032#endif
5033
5034/**
5035 * task_prio - return the priority value of a given task.
5036 * @p: the task in question.
5037 *
5038 * This is the priority value as seen by users in /proc.
5039 * RT tasks are offset by -200. Normal tasks are centered
5040 * around 0, value goes from -16 to +15.
5041 */
36c8b586 5042int task_prio(const struct task_struct *p)
1da177e4
LT
5043{
5044 return p->prio - MAX_RT_PRIO;
5045}
5046
5047/**
5048 * task_nice - return the nice value of a given task.
5049 * @p: the task in question.
5050 */
36c8b586 5051int task_nice(const struct task_struct *p)
1da177e4
LT
5052{
5053 return TASK_NICE(p);
5054}
150d8bed 5055EXPORT_SYMBOL(task_nice);
1da177e4
LT
5056
5057/**
5058 * idle_cpu - is a given cpu idle currently?
5059 * @cpu: the processor in question.
5060 */
5061int idle_cpu(int cpu)
5062{
5063 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5064}
5065
1da177e4
LT
5066/**
5067 * idle_task - return the idle task for a given cpu.
5068 * @cpu: the processor in question.
5069 */
36c8b586 5070struct task_struct *idle_task(int cpu)
1da177e4
LT
5071{
5072 return cpu_rq(cpu)->idle;
5073}
5074
5075/**
5076 * find_process_by_pid - find a process with a matching PID value.
5077 * @pid: the pid in question.
5078 */
a9957449 5079static struct task_struct *find_process_by_pid(pid_t pid)
1da177e4 5080{
228ebcbe 5081 return pid ? find_task_by_vpid(pid) : current;
1da177e4
LT
5082}
5083
5084/* Actually do priority change: must hold rq lock. */
dd41f596
IM
5085static void
5086__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
1da177e4 5087{
dd41f596 5088 BUG_ON(p->se.on_rq);
48f24c4d 5089
1da177e4 5090 p->policy = policy;
dd41f596
IM
5091 switch (p->policy) {
5092 case SCHED_NORMAL:
5093 case SCHED_BATCH:
5094 case SCHED_IDLE:
5095 p->sched_class = &fair_sched_class;
5096 break;
5097 case SCHED_FIFO:
5098 case SCHED_RR:
5099 p->sched_class = &rt_sched_class;
5100 break;
5101 }
5102
1da177e4 5103 p->rt_priority = prio;
b29739f9
IM
5104 p->normal_prio = normal_prio(p);
5105 /* we are holding p->pi_lock already */
5106 p->prio = rt_mutex_getprio(p);
2dd73a4f 5107 set_load_weight(p);
1da177e4
LT
5108}
5109
961ccddd
RR
5110static int __sched_setscheduler(struct task_struct *p, int policy,
5111 struct sched_param *param, bool user)
1da177e4 5112{
83b699ed 5113 int retval, oldprio, oldpolicy = -1, on_rq, running;
1da177e4 5114 unsigned long flags;
cb469845 5115 const struct sched_class *prev_class = p->sched_class;
70b97a7f 5116 struct rq *rq;
1da177e4 5117
66e5393a
SR
5118 /* may grab non-irq protected spin_locks */
5119 BUG_ON(in_interrupt());
1da177e4
LT
5120recheck:
5121 /* double check policy once rq lock held */
5122 if (policy < 0)
5123 policy = oldpolicy = p->policy;
5124 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
dd41f596
IM
5125 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5126 policy != SCHED_IDLE)
b0a9499c 5127 return -EINVAL;
1da177e4
LT
5128 /*
5129 * Valid priorities for SCHED_FIFO and SCHED_RR are
dd41f596
IM
5130 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5131 * SCHED_BATCH and SCHED_IDLE is 0.
1da177e4
LT
5132 */
5133 if (param->sched_priority < 0 ||
95cdf3b7 5134 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
d46523ea 5135 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
1da177e4 5136 return -EINVAL;
e05606d3 5137 if (rt_policy(policy) != (param->sched_priority != 0))
1da177e4
LT
5138 return -EINVAL;
5139
37e4ab3f
OC
5140 /*
5141 * Allow unprivileged RT tasks to decrease priority:
5142 */
961ccddd 5143 if (user && !capable(CAP_SYS_NICE)) {
e05606d3 5144 if (rt_policy(policy)) {
8dc3e909 5145 unsigned long rlim_rtprio;
8dc3e909
ON
5146
5147 if (!lock_task_sighand(p, &flags))
5148 return -ESRCH;
5149 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5150 unlock_task_sighand(p, &flags);
5151
5152 /* can't set/change the rt policy */
5153 if (policy != p->policy && !rlim_rtprio)
5154 return -EPERM;
5155
5156 /* can't increase priority */
5157 if (param->sched_priority > p->rt_priority &&
5158 param->sched_priority > rlim_rtprio)
5159 return -EPERM;
5160 }
dd41f596
IM
5161 /*
5162 * Like positive nice levels, dont allow tasks to
5163 * move out of SCHED_IDLE either:
5164 */
5165 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5166 return -EPERM;
5fe1d75f 5167
37e4ab3f
OC
5168 /* can't change other user's priorities */
5169 if ((current->euid != p->euid) &&
5170 (current->euid != p->uid))
5171 return -EPERM;
5172 }
1da177e4 5173
725aad24 5174 if (user) {
b68aa230 5175#ifdef CONFIG_RT_GROUP_SCHED
725aad24
JF
5176 /*
5177 * Do not allow realtime tasks into groups that have no runtime
5178 * assigned.
5179 */
9a7e0b18
PZ
5180 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5181 task_group(p)->rt_bandwidth.rt_runtime == 0)
725aad24 5182 return -EPERM;
b68aa230
PZ
5183#endif
5184
725aad24
JF
5185 retval = security_task_setscheduler(p, policy, param);
5186 if (retval)
5187 return retval;
5188 }
5189
b29739f9
IM
5190 /*
5191 * make sure no PI-waiters arrive (or leave) while we are
5192 * changing the priority of the task:
5193 */
5194 spin_lock_irqsave(&p->pi_lock, flags);
1da177e4
LT
5195 /*
5196 * To be able to change p->policy safely, the apropriate
5197 * runqueue lock must be held.
5198 */
b29739f9 5199 rq = __task_rq_lock(p);
1da177e4
LT
5200 /* recheck policy now with rq lock held */
5201 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5202 policy = oldpolicy = -1;
b29739f9
IM
5203 __task_rq_unlock(rq);
5204 spin_unlock_irqrestore(&p->pi_lock, flags);
1da177e4
LT
5205 goto recheck;
5206 }
2daa3577 5207 update_rq_clock(rq);
dd41f596 5208 on_rq = p->se.on_rq;
051a1d1a 5209 running = task_current(rq, p);
0e1f3483 5210 if (on_rq)
2e1cb74a 5211 deactivate_task(rq, p, 0);
0e1f3483
HS
5212 if (running)
5213 p->sched_class->put_prev_task(rq, p);
f6b53205 5214
1da177e4 5215 oldprio = p->prio;
dd41f596 5216 __setscheduler(rq, p, policy, param->sched_priority);
f6b53205 5217
0e1f3483
HS
5218 if (running)
5219 p->sched_class->set_curr_task(rq);
dd41f596
IM
5220 if (on_rq) {
5221 activate_task(rq, p, 0);
cb469845
SR
5222
5223 check_class_changed(rq, p, prev_class, oldprio, running);
1da177e4 5224 }
b29739f9
IM
5225 __task_rq_unlock(rq);
5226 spin_unlock_irqrestore(&p->pi_lock, flags);
5227
95e02ca9
TG
5228 rt_mutex_adjust_pi(p);
5229
1da177e4
LT
5230 return 0;
5231}
961ccddd
RR
5232
5233/**
5234 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5235 * @p: the task in question.
5236 * @policy: new policy.
5237 * @param: structure containing the new RT priority.
5238 *
5239 * NOTE that the task may be already dead.
5240 */
5241int sched_setscheduler(struct task_struct *p, int policy,
5242 struct sched_param *param)
5243{
5244 return __sched_setscheduler(p, policy, param, true);
5245}
1da177e4
LT
5246EXPORT_SYMBOL_GPL(sched_setscheduler);
5247
961ccddd
RR
5248/**
5249 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5250 * @p: the task in question.
5251 * @policy: new policy.
5252 * @param: structure containing the new RT priority.
5253 *
5254 * Just like sched_setscheduler, only don't bother checking if the
5255 * current context has permission. For example, this is needed in
5256 * stop_machine(): we create temporary high priority worker threads,
5257 * but our caller might not have that capability.
5258 */
5259int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5260 struct sched_param *param)
5261{
5262 return __sched_setscheduler(p, policy, param, false);
5263}
5264
95cdf3b7
IM
5265static int
5266do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
1da177e4 5267{
1da177e4
LT
5268 struct sched_param lparam;
5269 struct task_struct *p;
36c8b586 5270 int retval;
1da177e4
LT
5271
5272 if (!param || pid < 0)
5273 return -EINVAL;
5274 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5275 return -EFAULT;
5fe1d75f
ON
5276
5277 rcu_read_lock();
5278 retval = -ESRCH;
1da177e4 5279 p = find_process_by_pid(pid);
5fe1d75f
ON
5280 if (p != NULL)
5281 retval = sched_setscheduler(p, policy, &lparam);
5282 rcu_read_unlock();
36c8b586 5283
1da177e4
LT
5284 return retval;
5285}
5286
5287/**
5288 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5289 * @pid: the pid in question.
5290 * @policy: new policy.
5291 * @param: structure containing the new RT priority.
5292 */
41a2d6cf
IM
5293asmlinkage long
5294sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
1da177e4 5295{
c21761f1
JB
5296 /* negative values for policy are not valid */
5297 if (policy < 0)
5298 return -EINVAL;
5299
1da177e4
LT
5300 return do_sched_setscheduler(pid, policy, param);
5301}
5302
5303/**
5304 * sys_sched_setparam - set/change the RT priority of a thread
5305 * @pid: the pid in question.
5306 * @param: structure containing the new RT priority.
5307 */
5308asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5309{
5310 return do_sched_setscheduler(pid, -1, param);
5311}
5312
5313/**
5314 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5315 * @pid: the pid in question.
5316 */
5317asmlinkage long sys_sched_getscheduler(pid_t pid)
5318{
36c8b586 5319 struct task_struct *p;
3a5c359a 5320 int retval;
1da177e4
LT
5321
5322 if (pid < 0)
3a5c359a 5323 return -EINVAL;
1da177e4
LT
5324
5325 retval = -ESRCH;
5326 read_lock(&tasklist_lock);
5327 p = find_process_by_pid(pid);
5328 if (p) {
5329 retval = security_task_getscheduler(p);
5330 if (!retval)
5331 retval = p->policy;
5332 }
5333 read_unlock(&tasklist_lock);
1da177e4
LT
5334 return retval;
5335}
5336
5337/**
5338 * sys_sched_getscheduler - get the RT priority of a thread
5339 * @pid: the pid in question.
5340 * @param: structure containing the RT priority.
5341 */
5342asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5343{
5344 struct sched_param lp;
36c8b586 5345 struct task_struct *p;
3a5c359a 5346 int retval;
1da177e4
LT
5347
5348 if (!param || pid < 0)
3a5c359a 5349 return -EINVAL;
1da177e4
LT
5350
5351 read_lock(&tasklist_lock);
5352 p = find_process_by_pid(pid);
5353 retval = -ESRCH;
5354 if (!p)
5355 goto out_unlock;
5356
5357 retval = security_task_getscheduler(p);
5358 if (retval)
5359 goto out_unlock;
5360
5361 lp.sched_priority = p->rt_priority;
5362 read_unlock(&tasklist_lock);
5363
5364 /*
5365 * This one might sleep, we cannot do it with a spinlock held ...
5366 */
5367 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5368
1da177e4
LT
5369 return retval;
5370
5371out_unlock:
5372 read_unlock(&tasklist_lock);
5373 return retval;
5374}
5375
b53e921b 5376long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
1da177e4 5377{
1da177e4 5378 cpumask_t cpus_allowed;
b53e921b 5379 cpumask_t new_mask = *in_mask;
36c8b586
IM
5380 struct task_struct *p;
5381 int retval;
1da177e4 5382
95402b38 5383 get_online_cpus();
1da177e4
LT
5384 read_lock(&tasklist_lock);
5385
5386 p = find_process_by_pid(pid);
5387 if (!p) {
5388 read_unlock(&tasklist_lock);
95402b38 5389 put_online_cpus();
1da177e4
LT
5390 return -ESRCH;
5391 }
5392
5393 /*
5394 * It is not safe to call set_cpus_allowed with the
41a2d6cf 5395 * tasklist_lock held. We will bump the task_struct's
1da177e4
LT
5396 * usage count and then drop tasklist_lock.
5397 */
5398 get_task_struct(p);
5399 read_unlock(&tasklist_lock);
5400
5401 retval = -EPERM;
5402 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5403 !capable(CAP_SYS_NICE))
5404 goto out_unlock;
5405
e7834f8f
DQ
5406 retval = security_task_setscheduler(p, 0, NULL);
5407 if (retval)
5408 goto out_unlock;
5409
f9a86fcb 5410 cpuset_cpus_allowed(p, &cpus_allowed);
1da177e4 5411 cpus_and(new_mask, new_mask, cpus_allowed);
8707d8b8 5412 again:
7c16ec58 5413 retval = set_cpus_allowed_ptr(p, &new_mask);
1da177e4 5414
8707d8b8 5415 if (!retval) {
f9a86fcb 5416 cpuset_cpus_allowed(p, &cpus_allowed);
8707d8b8
PM
5417 if (!cpus_subset(new_mask, cpus_allowed)) {
5418 /*
5419 * We must have raced with a concurrent cpuset
5420 * update. Just reset the cpus_allowed to the
5421 * cpuset's cpus_allowed
5422 */
5423 new_mask = cpus_allowed;
5424 goto again;
5425 }
5426 }
1da177e4
LT
5427out_unlock:
5428 put_task_struct(p);
95402b38 5429 put_online_cpus();
1da177e4
LT
5430 return retval;
5431}
5432
5433static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5434 cpumask_t *new_mask)
5435{
5436 if (len < sizeof(cpumask_t)) {
5437 memset(new_mask, 0, sizeof(cpumask_t));
5438 } else if (len > sizeof(cpumask_t)) {
5439 len = sizeof(cpumask_t);
5440 }
5441 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5442}
5443
5444/**
5445 * sys_sched_setaffinity - set the cpu affinity of a process
5446 * @pid: pid of the process
5447 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5448 * @user_mask_ptr: user-space pointer to the new cpu mask
5449 */
5450asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5451 unsigned long __user *user_mask_ptr)
5452{
5453 cpumask_t new_mask;
5454 int retval;
5455
5456 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5457 if (retval)
5458 return retval;
5459
b53e921b 5460 return sched_setaffinity(pid, &new_mask);
1da177e4
LT
5461}
5462
1da177e4
LT
5463long sched_getaffinity(pid_t pid, cpumask_t *mask)
5464{
36c8b586 5465 struct task_struct *p;
1da177e4 5466 int retval;
1da177e4 5467
95402b38 5468 get_online_cpus();
1da177e4
LT
5469 read_lock(&tasklist_lock);
5470
5471 retval = -ESRCH;
5472 p = find_process_by_pid(pid);
5473 if (!p)
5474 goto out_unlock;
5475
e7834f8f
DQ
5476 retval = security_task_getscheduler(p);
5477 if (retval)
5478 goto out_unlock;
5479
2f7016d9 5480 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
1da177e4
LT
5481
5482out_unlock:
5483 read_unlock(&tasklist_lock);
95402b38 5484 put_online_cpus();
1da177e4 5485
9531b62f 5486 return retval;
1da177e4
LT
5487}
5488
5489/**
5490 * sys_sched_getaffinity - get the cpu affinity of a process
5491 * @pid: pid of the process
5492 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5493 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5494 */
5495asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5496 unsigned long __user *user_mask_ptr)
5497{
5498 int ret;
5499 cpumask_t mask;
5500
5501 if (len < sizeof(cpumask_t))
5502 return -EINVAL;
5503
5504 ret = sched_getaffinity(pid, &mask);
5505 if (ret < 0)
5506 return ret;
5507
5508 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5509 return -EFAULT;
5510
5511 return sizeof(cpumask_t);
5512}
5513
5514/**
5515 * sys_sched_yield - yield the current processor to other threads.
5516 *
dd41f596
IM
5517 * This function yields the current CPU to other tasks. If there are no
5518 * other threads running on this CPU then this function will return.
1da177e4
LT
5519 */
5520asmlinkage long sys_sched_yield(void)
5521{
70b97a7f 5522 struct rq *rq = this_rq_lock();
1da177e4 5523
2d72376b 5524 schedstat_inc(rq, yld_count);
4530d7ab 5525 current->sched_class->yield_task(rq);
1da177e4
LT
5526
5527 /*
5528 * Since we are going to call schedule() anyway, there's
5529 * no need to preempt or enable interrupts:
5530 */
5531 __release(rq->lock);
8a25d5de 5532 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1da177e4
LT
5533 _raw_spin_unlock(&rq->lock);
5534 preempt_enable_no_resched();
5535
5536 schedule();
5537
5538 return 0;
5539}
5540
e7b38404 5541static void __cond_resched(void)
1da177e4 5542{
8e0a43d8
IM
5543#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5544 __might_sleep(__FILE__, __LINE__);
5545#endif
5bbcfd90
IM
5546 /*
5547 * The BKS might be reacquired before we have dropped
5548 * PREEMPT_ACTIVE, which could trigger a second
5549 * cond_resched() call.
5550 */
1da177e4
LT
5551 do {
5552 add_preempt_count(PREEMPT_ACTIVE);
5553 schedule();
5554 sub_preempt_count(PREEMPT_ACTIVE);
5555 } while (need_resched());
5556}
5557
02b67cc3 5558int __sched _cond_resched(void)
1da177e4 5559{
9414232f
IM
5560 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5561 system_state == SYSTEM_RUNNING) {
1da177e4
LT
5562 __cond_resched();
5563 return 1;
5564 }
5565 return 0;
5566}
02b67cc3 5567EXPORT_SYMBOL(_cond_resched);
1da177e4
LT
5568
5569/*
5570 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5571 * call schedule, and on return reacquire the lock.
5572 *
41a2d6cf 5573 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
1da177e4
LT
5574 * operations here to prevent schedule() from being called twice (once via
5575 * spin_unlock(), once by hand).
5576 */
95cdf3b7 5577int cond_resched_lock(spinlock_t *lock)
1da177e4 5578{
95c354fe 5579 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6df3cecb
JK
5580 int ret = 0;
5581
95c354fe 5582 if (spin_needbreak(lock) || resched) {
1da177e4 5583 spin_unlock(lock);
95c354fe
NP
5584 if (resched && need_resched())
5585 __cond_resched();
5586 else
5587 cpu_relax();
6df3cecb 5588 ret = 1;
1da177e4 5589 spin_lock(lock);
1da177e4 5590 }
6df3cecb 5591 return ret;
1da177e4 5592}
1da177e4
LT
5593EXPORT_SYMBOL(cond_resched_lock);
5594
5595int __sched cond_resched_softirq(void)
5596{
5597 BUG_ON(!in_softirq());
5598
9414232f 5599 if (need_resched() && system_state == SYSTEM_RUNNING) {
98d82567 5600 local_bh_enable();
1da177e4
LT
5601 __cond_resched();
5602 local_bh_disable();
5603 return 1;
5604 }
5605 return 0;
5606}
1da177e4
LT
5607EXPORT_SYMBOL(cond_resched_softirq);
5608
1da177e4
LT
5609/**
5610 * yield - yield the current processor to other threads.
5611 *
72fd4a35 5612 * This is a shortcut for kernel-space yielding - it marks the
1da177e4
LT
5613 * thread runnable and calls sys_sched_yield().
5614 */
5615void __sched yield(void)
5616{
5617 set_current_state(TASK_RUNNING);
5618 sys_sched_yield();
5619}
1da177e4
LT
5620EXPORT_SYMBOL(yield);
5621
5622/*
41a2d6cf 5623 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
1da177e4
LT
5624 * that process accounting knows that this is a task in IO wait state.
5625 *
5626 * But don't do that if it is a deliberate, throttling IO wait (this task
5627 * has set its backing_dev_info: the queue against which it should throttle)
5628 */
5629void __sched io_schedule(void)
5630{
70b97a7f 5631 struct rq *rq = &__raw_get_cpu_var(runqueues);
1da177e4 5632
0ff92245 5633 delayacct_blkio_start();
1da177e4
LT
5634 atomic_inc(&rq->nr_iowait);
5635 schedule();
5636 atomic_dec(&rq->nr_iowait);
0ff92245 5637 delayacct_blkio_end();
1da177e4 5638}
1da177e4
LT
5639EXPORT_SYMBOL(io_schedule);
5640
5641long __sched io_schedule_timeout(long timeout)
5642{
70b97a7f 5643 struct rq *rq = &__raw_get_cpu_var(runqueues);
1da177e4
LT
5644 long ret;
5645
0ff92245 5646 delayacct_blkio_start();
1da177e4
LT
5647 atomic_inc(&rq->nr_iowait);
5648 ret = schedule_timeout(timeout);
5649 atomic_dec(&rq->nr_iowait);
0ff92245 5650 delayacct_blkio_end();
1da177e4
LT
5651 return ret;
5652}
5653
5654/**
5655 * sys_sched_get_priority_max - return maximum RT priority.
5656 * @policy: scheduling class.
5657 *
5658 * this syscall returns the maximum rt_priority that can be used
5659 * by a given scheduling class.
5660 */
5661asmlinkage long sys_sched_get_priority_max(int policy)
5662{
5663 int ret = -EINVAL;
5664
5665 switch (policy) {
5666 case SCHED_FIFO:
5667 case SCHED_RR:
5668 ret = MAX_USER_RT_PRIO-1;
5669 break;
5670 case SCHED_NORMAL:
b0a9499c 5671 case SCHED_BATCH:
dd41f596 5672 case SCHED_IDLE:
1da177e4
LT
5673 ret = 0;
5674 break;
5675 }
5676 return ret;
5677}
5678
5679/**
5680 * sys_sched_get_priority_min - return minimum RT priority.
5681 * @policy: scheduling class.
5682 *
5683 * this syscall returns the minimum rt_priority that can be used
5684 * by a given scheduling class.
5685 */
5686asmlinkage long sys_sched_get_priority_min(int policy)
5687{
5688 int ret = -EINVAL;
5689
5690 switch (policy) {
5691 case SCHED_FIFO:
5692 case SCHED_RR:
5693 ret = 1;
5694 break;
5695 case SCHED_NORMAL:
b0a9499c 5696 case SCHED_BATCH:
dd41f596 5697 case SCHED_IDLE:
1da177e4
LT
5698 ret = 0;
5699 }
5700 return ret;
5701}
5702
5703/**
5704 * sys_sched_rr_get_interval - return the default timeslice of a process.
5705 * @pid: pid of the process.
5706 * @interval: userspace pointer to the timeslice value.
5707 *
5708 * this syscall writes the default timeslice value of a given process
5709 * into the user-space timespec buffer. A value of '0' means infinity.
5710 */
5711asmlinkage
5712long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5713{
36c8b586 5714 struct task_struct *p;
a4ec24b4 5715 unsigned int time_slice;
3a5c359a 5716 int retval;
1da177e4 5717 struct timespec t;
1da177e4
LT
5718
5719 if (pid < 0)
3a5c359a 5720 return -EINVAL;
1da177e4
LT
5721
5722 retval = -ESRCH;
5723 read_lock(&tasklist_lock);
5724 p = find_process_by_pid(pid);
5725 if (!p)
5726 goto out_unlock;
5727
5728 retval = security_task_getscheduler(p);
5729 if (retval)
5730 goto out_unlock;
5731
77034937
IM
5732 /*
5733 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5734 * tasks that are on an otherwise idle runqueue:
5735 */
5736 time_slice = 0;
5737 if (p->policy == SCHED_RR) {
a4ec24b4 5738 time_slice = DEF_TIMESLICE;
1868f958 5739 } else if (p->policy != SCHED_FIFO) {
a4ec24b4
DA
5740 struct sched_entity *se = &p->se;
5741 unsigned long flags;
5742 struct rq *rq;
5743
5744 rq = task_rq_lock(p, &flags);
77034937
IM
5745 if (rq->cfs.load.weight)
5746 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
a4ec24b4
DA
5747 task_rq_unlock(rq, &flags);
5748 }
1da177e4 5749 read_unlock(&tasklist_lock);
a4ec24b4 5750 jiffies_to_timespec(time_slice, &t);
1da177e4 5751 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
1da177e4 5752 return retval;
3a5c359a 5753
1da177e4
LT
5754out_unlock:
5755 read_unlock(&tasklist_lock);
5756 return retval;
5757}
5758
7c731e0a 5759static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
36c8b586 5760
82a1fcb9 5761void sched_show_task(struct task_struct *p)
1da177e4 5762{
1da177e4 5763 unsigned long free = 0;
36c8b586 5764 unsigned state;
1da177e4 5765
1da177e4 5766 state = p->state ? __ffs(p->state) + 1 : 0;
cc4ea795 5767 printk(KERN_INFO "%-13.13s %c", p->comm,
2ed6e34f 5768 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4bd77321 5769#if BITS_PER_LONG == 32
1da177e4 5770 if (state == TASK_RUNNING)
cc4ea795 5771 printk(KERN_CONT " running ");
1da177e4 5772 else
cc4ea795 5773 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
1da177e4
LT
5774#else
5775 if (state == TASK_RUNNING)
cc4ea795 5776 printk(KERN_CONT " running task ");
1da177e4 5777 else
cc4ea795 5778 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
1da177e4
LT
5779#endif
5780#ifdef CONFIG_DEBUG_STACK_USAGE
5781 {
10ebffde 5782 unsigned long *n = end_of_stack(p);
1da177e4
LT
5783 while (!*n)
5784 n++;
10ebffde 5785 free = (unsigned long)n - (unsigned long)end_of_stack(p);
1da177e4
LT
5786 }
5787#endif
ba25f9dc 5788 printk(KERN_CONT "%5lu %5d %6d\n", free,
fcfd50af 5789 task_pid_nr(p), task_pid_nr(p->real_parent));
1da177e4 5790
5fb5e6de 5791 show_stack(p, NULL);
1da177e4
LT
5792}
5793
e59e2ae2 5794void show_state_filter(unsigned long state_filter)
1da177e4 5795{
36c8b586 5796 struct task_struct *g, *p;
1da177e4 5797
4bd77321
IM
5798#if BITS_PER_LONG == 32
5799 printk(KERN_INFO
5800 " task PC stack pid father\n");
1da177e4 5801#else
4bd77321
IM
5802 printk(KERN_INFO
5803 " task PC stack pid father\n");
1da177e4
LT
5804#endif
5805 read_lock(&tasklist_lock);
5806 do_each_thread(g, p) {
5807 /*
5808 * reset the NMI-timeout, listing all files on a slow
5809 * console might take alot of time:
5810 */
5811 touch_nmi_watchdog();
39bc89fd 5812 if (!state_filter || (p->state & state_filter))
82a1fcb9 5813 sched_show_task(p);
1da177e4
LT
5814 } while_each_thread(g, p);
5815
04c9167f
JF
5816 touch_all_softlockup_watchdogs();
5817
dd41f596
IM
5818#ifdef CONFIG_SCHED_DEBUG
5819 sysrq_sched_debug_show();
5820#endif
1da177e4 5821 read_unlock(&tasklist_lock);
e59e2ae2
IM
5822 /*
5823 * Only show locks if all tasks are dumped:
5824 */
5825 if (state_filter == -1)
5826 debug_show_all_locks();
1da177e4
LT
5827}
5828
1df21055
IM
5829void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5830{
dd41f596 5831 idle->sched_class = &idle_sched_class;
1df21055
IM
5832}
5833
f340c0d1
IM
5834/**
5835 * init_idle - set up an idle thread for a given CPU
5836 * @idle: task in question
5837 * @cpu: cpu the idle task belongs to
5838 *
5839 * NOTE: this function does not set the idle thread's NEED_RESCHED
5840 * flag, to make booting more robust.
5841 */
5c1e1767 5842void __cpuinit init_idle(struct task_struct *idle, int cpu)
1da177e4 5843{
70b97a7f 5844 struct rq *rq = cpu_rq(cpu);
1da177e4
LT
5845 unsigned long flags;
5846
5cbd54ef
IM
5847 spin_lock_irqsave(&rq->lock, flags);
5848
dd41f596
IM
5849 __sched_fork(idle);
5850 idle->se.exec_start = sched_clock();
5851
b29739f9 5852 idle->prio = idle->normal_prio = MAX_PRIO;
1da177e4 5853 idle->cpus_allowed = cpumask_of_cpu(cpu);
dd41f596 5854 __set_task_cpu(idle, cpu);
1da177e4 5855
1da177e4 5856 rq->curr = rq->idle = idle;
4866cde0
NP
5857#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5858 idle->oncpu = 1;
5859#endif
1da177e4
LT
5860 spin_unlock_irqrestore(&rq->lock, flags);
5861
5862 /* Set the preempt count _outside_ the spinlocks! */
8e3e076c
LT
5863#if defined(CONFIG_PREEMPT)
5864 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5865#else
a1261f54 5866 task_thread_info(idle)->preempt_count = 0;
8e3e076c 5867#endif
dd41f596
IM
5868 /*
5869 * The idle tasks have their own, simple scheduling class:
5870 */
5871 idle->sched_class = &idle_sched_class;
f201ae23 5872 ftrace_retfunc_init_task(idle);
1da177e4
LT
5873}
5874
5875/*
5876 * In a system that switches off the HZ timer nohz_cpu_mask
5877 * indicates which cpus entered this state. This is used
5878 * in the rcu update to wait only for active cpus. For system
5879 * which do not switch off the HZ timer nohz_cpu_mask should
5880 * always be CPU_MASK_NONE.
5881 */
5882cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5883
19978ca6
IM
5884/*
5885 * Increase the granularity value when there are more CPUs,
5886 * because with more CPUs the 'effective latency' as visible
5887 * to users decreases. But the relationship is not linear,
5888 * so pick a second-best guess by going with the log2 of the
5889 * number of CPUs.
5890 *
5891 * This idea comes from the SD scheduler of Con Kolivas:
5892 */
5893static inline void sched_init_granularity(void)
5894{
5895 unsigned int factor = 1 + ilog2(num_online_cpus());
5896 const unsigned long limit = 200000000;
5897
5898 sysctl_sched_min_granularity *= factor;
5899 if (sysctl_sched_min_granularity > limit)
5900 sysctl_sched_min_granularity = limit;
5901
5902 sysctl_sched_latency *= factor;
5903 if (sysctl_sched_latency > limit)
5904 sysctl_sched_latency = limit;
5905
5906 sysctl_sched_wakeup_granularity *= factor;
55cd5340
PZ
5907
5908 sysctl_sched_shares_ratelimit *= factor;
19978ca6
IM
5909}
5910
1da177e4
LT
5911#ifdef CONFIG_SMP
5912/*
5913 * This is how migration works:
5914 *
70b97a7f 5915 * 1) we queue a struct migration_req structure in the source CPU's
1da177e4
LT
5916 * runqueue and wake up that CPU's migration thread.
5917 * 2) we down() the locked semaphore => thread blocks.
5918 * 3) migration thread wakes up (implicitly it forces the migrated
5919 * thread off the CPU)
5920 * 4) it gets the migration request and checks whether the migrated
5921 * task is still in the wrong runqueue.
5922 * 5) if it's in the wrong runqueue then the migration thread removes
5923 * it and puts it into the right queue.
5924 * 6) migration thread up()s the semaphore.
5925 * 7) we wake up and the migration is done.
5926 */
5927
5928/*
5929 * Change a given task's CPU affinity. Migrate the thread to a
5930 * proper CPU and schedule it away if the CPU it's executing on
5931 * is removed from the allowed bitmask.
5932 *
5933 * NOTE: the caller must have a valid reference to the task, the
41a2d6cf 5934 * task must not exit() & deallocate itself prematurely. The
1da177e4
LT
5935 * call is not atomic; no spinlocks may be held.
5936 */
cd8ba7cd 5937int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
1da177e4 5938{
70b97a7f 5939 struct migration_req req;
1da177e4 5940 unsigned long flags;
70b97a7f 5941 struct rq *rq;
48f24c4d 5942 int ret = 0;
1da177e4
LT
5943
5944 rq = task_rq_lock(p, &flags);
cd8ba7cd 5945 if (!cpus_intersects(*new_mask, cpu_online_map)) {
1da177e4
LT
5946 ret = -EINVAL;
5947 goto out;
5948 }
5949
9985b0ba
DR
5950 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5951 !cpus_equal(p->cpus_allowed, *new_mask))) {
5952 ret = -EINVAL;
5953 goto out;
5954 }
5955
73fe6aae 5956 if (p->sched_class->set_cpus_allowed)
cd8ba7cd 5957 p->sched_class->set_cpus_allowed(p, new_mask);
73fe6aae 5958 else {
cd8ba7cd
MT
5959 p->cpus_allowed = *new_mask;
5960 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
73fe6aae
GH
5961 }
5962
1da177e4 5963 /* Can the task run on the task's current CPU? If so, we're done */
cd8ba7cd 5964 if (cpu_isset(task_cpu(p), *new_mask))
1da177e4
LT
5965 goto out;
5966
cd8ba7cd 5967 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
1da177e4
LT
5968 /* Need help from migration thread: drop lock and wait. */
5969 task_rq_unlock(rq, &flags);
5970 wake_up_process(rq->migration_thread);
5971 wait_for_completion(&req.done);
5972 tlb_migrate_finish(p->mm);
5973 return 0;
5974 }
5975out:
5976 task_rq_unlock(rq, &flags);
48f24c4d 5977
1da177e4
LT
5978 return ret;
5979}
cd8ba7cd 5980EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1da177e4
LT
5981
5982/*
41a2d6cf 5983 * Move (not current) task off this cpu, onto dest cpu. We're doing
1da177e4
LT
5984 * this because either it can't run here any more (set_cpus_allowed()
5985 * away from this CPU, or CPU going down), or because we're
5986 * attempting to rebalance this task on exec (sched_exec).
5987 *
5988 * So we race with normal scheduler movements, but that's OK, as long
5989 * as the task is no longer on this CPU.
efc30814
KK
5990 *
5991 * Returns non-zero if task was successfully migrated.
1da177e4 5992 */
efc30814 5993static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
1da177e4 5994{
70b97a7f 5995 struct rq *rq_dest, *rq_src;
dd41f596 5996 int ret = 0, on_rq;
1da177e4 5997
e761b772 5998 if (unlikely(!cpu_active(dest_cpu)))
efc30814 5999 return ret;
1da177e4
LT
6000
6001 rq_src = cpu_rq(src_cpu);
6002 rq_dest = cpu_rq(dest_cpu);
6003
6004 double_rq_lock(rq_src, rq_dest);
6005 /* Already moved. */
6006 if (task_cpu(p) != src_cpu)
b1e38734 6007 goto done;
1da177e4
LT
6008 /* Affinity changed (again). */
6009 if (!cpu_isset(dest_cpu, p->cpus_allowed))
b1e38734 6010 goto fail;
1da177e4 6011
dd41f596 6012 on_rq = p->se.on_rq;
6e82a3be 6013 if (on_rq)
2e1cb74a 6014 deactivate_task(rq_src, p, 0);
6e82a3be 6015
1da177e4 6016 set_task_cpu(p, dest_cpu);
dd41f596
IM
6017 if (on_rq) {
6018 activate_task(rq_dest, p, 0);
15afe09b 6019 check_preempt_curr(rq_dest, p, 0);
1da177e4 6020 }
b1e38734 6021done:
efc30814 6022 ret = 1;
b1e38734 6023fail:
1da177e4 6024 double_rq_unlock(rq_src, rq_dest);
efc30814 6025 return ret;
1da177e4
LT
6026}
6027
6028/*
6029 * migration_thread - this is a highprio system thread that performs
6030 * thread migration by bumping thread off CPU then 'pushing' onto
6031 * another runqueue.
6032 */
95cdf3b7 6033static int migration_thread(void *data)
1da177e4 6034{
1da177e4 6035 int cpu = (long)data;
70b97a7f 6036 struct rq *rq;
1da177e4
LT
6037
6038 rq = cpu_rq(cpu);
6039 BUG_ON(rq->migration_thread != current);
6040
6041 set_current_state(TASK_INTERRUPTIBLE);
6042 while (!kthread_should_stop()) {
70b97a7f 6043 struct migration_req *req;
1da177e4 6044 struct list_head *head;
1da177e4 6045
1da177e4
LT
6046 spin_lock_irq(&rq->lock);
6047
6048 if (cpu_is_offline(cpu)) {
6049 spin_unlock_irq(&rq->lock);
6050 goto wait_to_die;
6051 }
6052
6053 if (rq->active_balance) {
6054 active_load_balance(rq, cpu);
6055 rq->active_balance = 0;
6056 }
6057
6058 head = &rq->migration_queue;
6059
6060 if (list_empty(head)) {
6061 spin_unlock_irq(&rq->lock);
6062 schedule();
6063 set_current_state(TASK_INTERRUPTIBLE);
6064 continue;
6065 }
70b97a7f 6066 req = list_entry(head->next, struct migration_req, list);
1da177e4
LT
6067 list_del_init(head->next);
6068
674311d5
NP
6069 spin_unlock(&rq->lock);
6070 __migrate_task(req->task, cpu, req->dest_cpu);
6071 local_irq_enable();
1da177e4
LT
6072
6073 complete(&req->done);
6074 }
6075 __set_current_state(TASK_RUNNING);
6076 return 0;
6077
6078wait_to_die:
6079 /* Wait for kthread_stop */
6080 set_current_state(TASK_INTERRUPTIBLE);
6081 while (!kthread_should_stop()) {
6082 schedule();
6083 set_current_state(TASK_INTERRUPTIBLE);
6084 }
6085 __set_current_state(TASK_RUNNING);
6086 return 0;
6087}
6088
6089#ifdef CONFIG_HOTPLUG_CPU
f7b4cddc
ON
6090
6091static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6092{
6093 int ret;
6094
6095 local_irq_disable();
6096 ret = __migrate_task(p, src_cpu, dest_cpu);
6097 local_irq_enable();
6098 return ret;
6099}
6100
054b9108 6101/*
3a4fa0a2 6102 * Figure out where task on dead CPU should go, use force if necessary.
054b9108 6103 */
48f24c4d 6104static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
1da177e4 6105{
efc30814 6106 unsigned long flags;
1da177e4 6107 cpumask_t mask;
70b97a7f
IM
6108 struct rq *rq;
6109 int dest_cpu;
1da177e4 6110
3a5c359a
AK
6111 do {
6112 /* On same node? */
ea6f18ed
MT
6113 node_to_cpumask_ptr(pnodemask, cpu_to_node(dead_cpu));
6114
6115 cpus_and(mask, *pnodemask, p->cpus_allowed);
3a5c359a
AK
6116 dest_cpu = any_online_cpu(mask);
6117
6118 /* On any allowed CPU? */
434d53b0 6119 if (dest_cpu >= nr_cpu_ids)
3a5c359a
AK
6120 dest_cpu = any_online_cpu(p->cpus_allowed);
6121
6122 /* No more Mr. Nice Guy. */
434d53b0 6123 if (dest_cpu >= nr_cpu_ids) {
f9a86fcb
MT
6124 cpumask_t cpus_allowed;
6125
6126 cpuset_cpus_allowed_locked(p, &cpus_allowed);
470fd646
CW
6127 /*
6128 * Try to stay on the same cpuset, where the
6129 * current cpuset may be a subset of all cpus.
6130 * The cpuset_cpus_allowed_locked() variant of
41a2d6cf 6131 * cpuset_cpus_allowed() will not block. It must be
470fd646
CW
6132 * called within calls to cpuset_lock/cpuset_unlock.
6133 */
3a5c359a 6134 rq = task_rq_lock(p, &flags);
470fd646 6135 p->cpus_allowed = cpus_allowed;
3a5c359a
AK
6136 dest_cpu = any_online_cpu(p->cpus_allowed);
6137 task_rq_unlock(rq, &flags);
1da177e4 6138
3a5c359a
AK
6139 /*
6140 * Don't tell them about moving exiting tasks or
6141 * kernel threads (both mm NULL), since they never
6142 * leave kernel.
6143 */
41a2d6cf 6144 if (p->mm && printk_ratelimit()) {
3a5c359a
AK
6145 printk(KERN_INFO "process %d (%s) no "
6146 "longer affine to cpu%d\n",
41a2d6cf
IM
6147 task_pid_nr(p), p->comm, dead_cpu);
6148 }
3a5c359a 6149 }
f7b4cddc 6150 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
1da177e4
LT
6151}
6152
6153/*
6154 * While a dead CPU has no uninterruptible tasks queued at this point,
6155 * it might still have a nonzero ->nr_uninterruptible counter, because
6156 * for performance reasons the counter is not stricly tracking tasks to
6157 * their home CPUs. So we just add the counter to another CPU's counter,
6158 * to keep the global sum constant after CPU-down:
6159 */
70b97a7f 6160static void migrate_nr_uninterruptible(struct rq *rq_src)
1da177e4 6161{
7c16ec58 6162 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
1da177e4
LT
6163 unsigned long flags;
6164
6165 local_irq_save(flags);
6166 double_rq_lock(rq_src, rq_dest);
6167 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6168 rq_src->nr_uninterruptible = 0;
6169 double_rq_unlock(rq_src, rq_dest);
6170 local_irq_restore(flags);
6171}
6172
6173/* Run through task list and migrate tasks from the dead cpu. */
6174static void migrate_live_tasks(int src_cpu)
6175{
48f24c4d 6176 struct task_struct *p, *t;
1da177e4 6177
f7b4cddc 6178 read_lock(&tasklist_lock);
1da177e4 6179
48f24c4d
IM
6180 do_each_thread(t, p) {
6181 if (p == current)
1da177e4
LT
6182 continue;
6183
48f24c4d
IM
6184 if (task_cpu(p) == src_cpu)
6185 move_task_off_dead_cpu(src_cpu, p);
6186 } while_each_thread(t, p);
1da177e4 6187
f7b4cddc 6188 read_unlock(&tasklist_lock);
1da177e4
LT
6189}
6190
dd41f596
IM
6191/*
6192 * Schedules idle task to be the next runnable task on current CPU.
94bc9a7b
DA
6193 * It does so by boosting its priority to highest possible.
6194 * Used by CPU offline code.
1da177e4
LT
6195 */
6196void sched_idle_next(void)
6197{
48f24c4d 6198 int this_cpu = smp_processor_id();
70b97a7f 6199 struct rq *rq = cpu_rq(this_cpu);
1da177e4
LT
6200 struct task_struct *p = rq->idle;
6201 unsigned long flags;
6202
6203 /* cpu has to be offline */
48f24c4d 6204 BUG_ON(cpu_online(this_cpu));
1da177e4 6205
48f24c4d
IM
6206 /*
6207 * Strictly not necessary since rest of the CPUs are stopped by now
6208 * and interrupts disabled on the current cpu.
1da177e4
LT
6209 */
6210 spin_lock_irqsave(&rq->lock, flags);
6211
dd41f596 6212 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
48f24c4d 6213
94bc9a7b
DA
6214 update_rq_clock(rq);
6215 activate_task(rq, p, 0);
1da177e4
LT
6216
6217 spin_unlock_irqrestore(&rq->lock, flags);
6218}
6219
48f24c4d
IM
6220/*
6221 * Ensures that the idle task is using init_mm right before its cpu goes
1da177e4
LT
6222 * offline.
6223 */
6224void idle_task_exit(void)
6225{
6226 struct mm_struct *mm = current->active_mm;
6227
6228 BUG_ON(cpu_online(smp_processor_id()));
6229
6230 if (mm != &init_mm)
6231 switch_mm(mm, &init_mm, current);
6232 mmdrop(mm);
6233}
6234
054b9108 6235/* called under rq->lock with disabled interrupts */
36c8b586 6236static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
1da177e4 6237{
70b97a7f 6238 struct rq *rq = cpu_rq(dead_cpu);
1da177e4
LT
6239
6240 /* Must be exiting, otherwise would be on tasklist. */
270f722d 6241 BUG_ON(!p->exit_state);
1da177e4
LT
6242
6243 /* Cannot have done final schedule yet: would have vanished. */
c394cc9f 6244 BUG_ON(p->state == TASK_DEAD);
1da177e4 6245
48f24c4d 6246 get_task_struct(p);
1da177e4
LT
6247
6248 /*
6249 * Drop lock around migration; if someone else moves it,
41a2d6cf 6250 * that's OK. No task can be added to this CPU, so iteration is
1da177e4
LT
6251 * fine.
6252 */
f7b4cddc 6253 spin_unlock_irq(&rq->lock);
48f24c4d 6254 move_task_off_dead_cpu(dead_cpu, p);
f7b4cddc 6255 spin_lock_irq(&rq->lock);
1da177e4 6256
48f24c4d 6257 put_task_struct(p);
1da177e4
LT
6258}
6259
6260/* release_task() removes task from tasklist, so we won't find dead tasks. */
6261static void migrate_dead_tasks(unsigned int dead_cpu)
6262{
70b97a7f 6263 struct rq *rq = cpu_rq(dead_cpu);
dd41f596 6264 struct task_struct *next;
48f24c4d 6265
dd41f596
IM
6266 for ( ; ; ) {
6267 if (!rq->nr_running)
6268 break;
a8e504d2 6269 update_rq_clock(rq);
ff95f3df 6270 next = pick_next_task(rq, rq->curr);
dd41f596
IM
6271 if (!next)
6272 break;
79c53799 6273 next->sched_class->put_prev_task(rq, next);
dd41f596 6274 migrate_dead(dead_cpu, next);
e692ab53 6275
1da177e4
LT
6276 }
6277}
6278#endif /* CONFIG_HOTPLUG_CPU */
6279
e692ab53
NP
6280#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6281
6282static struct ctl_table sd_ctl_dir[] = {
e0361851
AD
6283 {
6284 .procname = "sched_domain",
c57baf1e 6285 .mode = 0555,
e0361851 6286 },
38605cae 6287 {0, },
e692ab53
NP
6288};
6289
6290static struct ctl_table sd_ctl_root[] = {
e0361851 6291 {
c57baf1e 6292 .ctl_name = CTL_KERN,
e0361851 6293 .procname = "kernel",
c57baf1e 6294 .mode = 0555,
e0361851
AD
6295 .child = sd_ctl_dir,
6296 },
38605cae 6297 {0, },
e692ab53
NP
6298};
6299
6300static struct ctl_table *sd_alloc_ctl_entry(int n)
6301{
6302 struct ctl_table *entry =
5cf9f062 6303 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
e692ab53 6304
e692ab53
NP
6305 return entry;
6306}
6307
6382bc90
MM
6308static void sd_free_ctl_entry(struct ctl_table **tablep)
6309{
cd790076 6310 struct ctl_table *entry;
6382bc90 6311
cd790076
MM
6312 /*
6313 * In the intermediate directories, both the child directory and
6314 * procname are dynamically allocated and could fail but the mode
41a2d6cf 6315 * will always be set. In the lowest directory the names are
cd790076
MM
6316 * static strings and all have proc handlers.
6317 */
6318 for (entry = *tablep; entry->mode; entry++) {
6382bc90
MM
6319 if (entry->child)
6320 sd_free_ctl_entry(&entry->child);
cd790076
MM
6321 if (entry->proc_handler == NULL)
6322 kfree(entry->procname);
6323 }
6382bc90
MM
6324
6325 kfree(*tablep);
6326 *tablep = NULL;
6327}
6328
e692ab53 6329static void
e0361851 6330set_table_entry(struct ctl_table *entry,
e692ab53
NP
6331 const char *procname, void *data, int maxlen,
6332 mode_t mode, proc_handler *proc_handler)
6333{
e692ab53
NP
6334 entry->procname = procname;
6335 entry->data = data;
6336 entry->maxlen = maxlen;
6337 entry->mode = mode;
6338 entry->proc_handler = proc_handler;
6339}
6340
6341static struct ctl_table *
6342sd_alloc_ctl_domain_table(struct sched_domain *sd)
6343{
a5d8c348 6344 struct ctl_table *table = sd_alloc_ctl_entry(13);
e692ab53 6345
ad1cdc1d
MM
6346 if (table == NULL)
6347 return NULL;
6348
e0361851 6349 set_table_entry(&table[0], "min_interval", &sd->min_interval,
e692ab53 6350 sizeof(long), 0644, proc_doulongvec_minmax);
e0361851 6351 set_table_entry(&table[1], "max_interval", &sd->max_interval,
e692ab53 6352 sizeof(long), 0644, proc_doulongvec_minmax);
e0361851 6353 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
e692ab53 6354 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6355 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
e692ab53 6356 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6357 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
e692ab53 6358 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6359 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
e692ab53 6360 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6361 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
e692ab53 6362 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6363 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
e692ab53 6364 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 6365 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
e692ab53 6366 sizeof(int), 0644, proc_dointvec_minmax);
ace8b3d6 6367 set_table_entry(&table[9], "cache_nice_tries",
e692ab53
NP
6368 &sd->cache_nice_tries,
6369 sizeof(int), 0644, proc_dointvec_minmax);
ace8b3d6 6370 set_table_entry(&table[10], "flags", &sd->flags,
e692ab53 6371 sizeof(int), 0644, proc_dointvec_minmax);
a5d8c348
IM
6372 set_table_entry(&table[11], "name", sd->name,
6373 CORENAME_MAX_SIZE, 0444, proc_dostring);
6374 /* &table[12] is terminator */
e692ab53
NP
6375
6376 return table;
6377}
6378
9a4e7159 6379static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
e692ab53
NP
6380{
6381 struct ctl_table *entry, *table;
6382 struct sched_domain *sd;
6383 int domain_num = 0, i;
6384 char buf[32];
6385
6386 for_each_domain(cpu, sd)
6387 domain_num++;
6388 entry = table = sd_alloc_ctl_entry(domain_num + 1);
ad1cdc1d
MM
6389 if (table == NULL)
6390 return NULL;
e692ab53
NP
6391
6392 i = 0;
6393 for_each_domain(cpu, sd) {
6394 snprintf(buf, 32, "domain%d", i);
e692ab53 6395 entry->procname = kstrdup(buf, GFP_KERNEL);
c57baf1e 6396 entry->mode = 0555;
e692ab53
NP
6397 entry->child = sd_alloc_ctl_domain_table(sd);
6398 entry++;
6399 i++;
6400 }
6401 return table;
6402}
6403
6404static struct ctl_table_header *sd_sysctl_header;
6382bc90 6405static void register_sched_domain_sysctl(void)
e692ab53
NP
6406{
6407 int i, cpu_num = num_online_cpus();
6408 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6409 char buf[32];
6410
7378547f
MM
6411 WARN_ON(sd_ctl_dir[0].child);
6412 sd_ctl_dir[0].child = entry;
6413
ad1cdc1d
MM
6414 if (entry == NULL)
6415 return;
6416
97b6ea7b 6417 for_each_online_cpu(i) {
e692ab53 6418 snprintf(buf, 32, "cpu%d", i);
e692ab53 6419 entry->procname = kstrdup(buf, GFP_KERNEL);
c57baf1e 6420 entry->mode = 0555;
e692ab53 6421 entry->child = sd_alloc_ctl_cpu_table(i);
97b6ea7b 6422 entry++;
e692ab53 6423 }
7378547f
MM
6424
6425 WARN_ON(sd_sysctl_header);
e692ab53
NP
6426 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6427}
6382bc90 6428
7378547f 6429/* may be called multiple times per register */
6382bc90
MM
6430static void unregister_sched_domain_sysctl(void)
6431{
7378547f
MM
6432 if (sd_sysctl_header)
6433 unregister_sysctl_table(sd_sysctl_header);
6382bc90 6434 sd_sysctl_header = NULL;
7378547f
MM
6435 if (sd_ctl_dir[0].child)
6436 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6382bc90 6437}
e692ab53 6438#else
6382bc90
MM
6439static void register_sched_domain_sysctl(void)
6440{
6441}
6442static void unregister_sched_domain_sysctl(void)
e692ab53
NP
6443{
6444}
6445#endif
6446
1f11eb6a
GH
6447static void set_rq_online(struct rq *rq)
6448{
6449 if (!rq->online) {
6450 const struct sched_class *class;
6451
6452 cpu_set(rq->cpu, rq->rd->online);
6453 rq->online = 1;
6454
6455 for_each_class(class) {
6456 if (class->rq_online)
6457 class->rq_online(rq);
6458 }
6459 }
6460}
6461
6462static void set_rq_offline(struct rq *rq)
6463{
6464 if (rq->online) {
6465 const struct sched_class *class;
6466
6467 for_each_class(class) {
6468 if (class->rq_offline)
6469 class->rq_offline(rq);
6470 }
6471
6472 cpu_clear(rq->cpu, rq->rd->online);
6473 rq->online = 0;
6474 }
6475}
6476
1da177e4
LT
6477/*
6478 * migration_call - callback that gets triggered when a CPU is added.
6479 * Here we can start up the necessary migration thread for the new CPU.
6480 */
48f24c4d
IM
6481static int __cpuinit
6482migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
1da177e4 6483{
1da177e4 6484 struct task_struct *p;
48f24c4d 6485 int cpu = (long)hcpu;
1da177e4 6486 unsigned long flags;
70b97a7f 6487 struct rq *rq;
1da177e4
LT
6488
6489 switch (action) {
5be9361c 6490
1da177e4 6491 case CPU_UP_PREPARE:
8bb78442 6492 case CPU_UP_PREPARE_FROZEN:
dd41f596 6493 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
1da177e4
LT
6494 if (IS_ERR(p))
6495 return NOTIFY_BAD;
1da177e4
LT
6496 kthread_bind(p, cpu);
6497 /* Must be high prio: stop_machine expects to yield to it. */
6498 rq = task_rq_lock(p, &flags);
dd41f596 6499 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
1da177e4
LT
6500 task_rq_unlock(rq, &flags);
6501 cpu_rq(cpu)->migration_thread = p;
6502 break;
48f24c4d 6503
1da177e4 6504 case CPU_ONLINE:
8bb78442 6505 case CPU_ONLINE_FROZEN:
3a4fa0a2 6506 /* Strictly unnecessary, as first user will wake it. */
1da177e4 6507 wake_up_process(cpu_rq(cpu)->migration_thread);
1f94ef59
GH
6508
6509 /* Update our root-domain */
6510 rq = cpu_rq(cpu);
6511 spin_lock_irqsave(&rq->lock, flags);
6512 if (rq->rd) {
6513 BUG_ON(!cpu_isset(cpu, rq->rd->span));
1f11eb6a
GH
6514
6515 set_rq_online(rq);
1f94ef59
GH
6516 }
6517 spin_unlock_irqrestore(&rq->lock, flags);
1da177e4 6518 break;
48f24c4d 6519
1da177e4
LT
6520#ifdef CONFIG_HOTPLUG_CPU
6521 case CPU_UP_CANCELED:
8bb78442 6522 case CPU_UP_CANCELED_FROZEN:
fc75cdfa
HC
6523 if (!cpu_rq(cpu)->migration_thread)
6524 break;
41a2d6cf 6525 /* Unbind it from offline cpu so it can run. Fall thru. */
a4c4af7c
HC
6526 kthread_bind(cpu_rq(cpu)->migration_thread,
6527 any_online_cpu(cpu_online_map));
1da177e4
LT
6528 kthread_stop(cpu_rq(cpu)->migration_thread);
6529 cpu_rq(cpu)->migration_thread = NULL;
6530 break;
48f24c4d 6531
1da177e4 6532 case CPU_DEAD:
8bb78442 6533 case CPU_DEAD_FROZEN:
470fd646 6534 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
1da177e4
LT
6535 migrate_live_tasks(cpu);
6536 rq = cpu_rq(cpu);
6537 kthread_stop(rq->migration_thread);
6538 rq->migration_thread = NULL;
6539 /* Idle task back to normal (off runqueue, low prio) */
d2da272a 6540 spin_lock_irq(&rq->lock);
a8e504d2 6541 update_rq_clock(rq);
2e1cb74a 6542 deactivate_task(rq, rq->idle, 0);
1da177e4 6543 rq->idle->static_prio = MAX_PRIO;
dd41f596
IM
6544 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6545 rq->idle->sched_class = &idle_sched_class;
1da177e4 6546 migrate_dead_tasks(cpu);
d2da272a 6547 spin_unlock_irq(&rq->lock);
470fd646 6548 cpuset_unlock();
1da177e4
LT
6549 migrate_nr_uninterruptible(rq);
6550 BUG_ON(rq->nr_running != 0);
6551
41a2d6cf
IM
6552 /*
6553 * No need to migrate the tasks: it was best-effort if
6554 * they didn't take sched_hotcpu_mutex. Just wake up
6555 * the requestors.
6556 */
1da177e4
LT
6557 spin_lock_irq(&rq->lock);
6558 while (!list_empty(&rq->migration_queue)) {
70b97a7f
IM
6559 struct migration_req *req;
6560
1da177e4 6561 req = list_entry(rq->migration_queue.next,
70b97a7f 6562 struct migration_req, list);
1da177e4
LT
6563 list_del_init(&req->list);
6564 complete(&req->done);
6565 }
6566 spin_unlock_irq(&rq->lock);
6567 break;
57d885fe 6568
08f503b0
GH
6569 case CPU_DYING:
6570 case CPU_DYING_FROZEN:
57d885fe
GH
6571 /* Update our root-domain */
6572 rq = cpu_rq(cpu);
6573 spin_lock_irqsave(&rq->lock, flags);
6574 if (rq->rd) {
6575 BUG_ON(!cpu_isset(cpu, rq->rd->span));
1f11eb6a 6576 set_rq_offline(rq);
57d885fe
GH
6577 }
6578 spin_unlock_irqrestore(&rq->lock, flags);
6579 break;
1da177e4
LT
6580#endif
6581 }
6582 return NOTIFY_OK;
6583}
6584
6585/* Register at highest priority so that task migration (migrate_all_tasks)
6586 * happens before everything else.
6587 */
26c2143b 6588static struct notifier_block __cpuinitdata migration_notifier = {
1da177e4
LT
6589 .notifier_call = migration_call,
6590 .priority = 10
6591};
6592
7babe8db 6593static int __init migration_init(void)
1da177e4
LT
6594{
6595 void *cpu = (void *)(long)smp_processor_id();
07dccf33 6596 int err;
48f24c4d
IM
6597
6598 /* Start one for the boot CPU: */
07dccf33
AM
6599 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6600 BUG_ON(err == NOTIFY_BAD);
1da177e4
LT
6601 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6602 register_cpu_notifier(&migration_notifier);
7babe8db
EGM
6603
6604 return err;
1da177e4 6605}
7babe8db 6606early_initcall(migration_init);
1da177e4
LT
6607#endif
6608
6609#ifdef CONFIG_SMP
476f3534 6610
3e9830dc 6611#ifdef CONFIG_SCHED_DEBUG
4dcf6aff 6612
7c16ec58
MT
6613static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6614 cpumask_t *groupmask)
1da177e4 6615{
4dcf6aff 6616 struct sched_group *group = sd->groups;
434d53b0 6617 char str[256];
1da177e4 6618
434d53b0 6619 cpulist_scnprintf(str, sizeof(str), sd->span);
7c16ec58 6620 cpus_clear(*groupmask);
4dcf6aff
IM
6621
6622 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6623
6624 if (!(sd->flags & SD_LOAD_BALANCE)) {
6625 printk("does not load-balance\n");
6626 if (sd->parent)
6627 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6628 " has parent");
6629 return -1;
41c7ce9a
NP
6630 }
6631
eefd796a 6632 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4dcf6aff
IM
6633
6634 if (!cpu_isset(cpu, sd->span)) {
6635 printk(KERN_ERR "ERROR: domain->span does not contain "
6636 "CPU%d\n", cpu);
6637 }
6638 if (!cpu_isset(cpu, group->cpumask)) {
6639 printk(KERN_ERR "ERROR: domain->groups does not contain"
6640 " CPU%d\n", cpu);
6641 }
1da177e4 6642
4dcf6aff 6643 printk(KERN_DEBUG "%*s groups:", level + 1, "");
1da177e4 6644 do {
4dcf6aff
IM
6645 if (!group) {
6646 printk("\n");
6647 printk(KERN_ERR "ERROR: group is NULL\n");
1da177e4
LT
6648 break;
6649 }
6650
4dcf6aff
IM
6651 if (!group->__cpu_power) {
6652 printk(KERN_CONT "\n");
6653 printk(KERN_ERR "ERROR: domain->cpu_power not "
6654 "set\n");
6655 break;
6656 }
1da177e4 6657
4dcf6aff
IM
6658 if (!cpus_weight(group->cpumask)) {
6659 printk(KERN_CONT "\n");
6660 printk(KERN_ERR "ERROR: empty group\n");
6661 break;
6662 }
1da177e4 6663
7c16ec58 6664 if (cpus_intersects(*groupmask, group->cpumask)) {
4dcf6aff
IM
6665 printk(KERN_CONT "\n");
6666 printk(KERN_ERR "ERROR: repeated CPUs\n");
6667 break;
6668 }
1da177e4 6669
7c16ec58 6670 cpus_or(*groupmask, *groupmask, group->cpumask);
1da177e4 6671
434d53b0 6672 cpulist_scnprintf(str, sizeof(str), group->cpumask);
4dcf6aff 6673 printk(KERN_CONT " %s", str);
1da177e4 6674
4dcf6aff
IM
6675 group = group->next;
6676 } while (group != sd->groups);
6677 printk(KERN_CONT "\n");
1da177e4 6678
7c16ec58 6679 if (!cpus_equal(sd->span, *groupmask))
4dcf6aff 6680 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
1da177e4 6681
7c16ec58 6682 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
4dcf6aff
IM
6683 printk(KERN_ERR "ERROR: parent span is not a superset "
6684 "of domain->span\n");
6685 return 0;
6686}
1da177e4 6687
4dcf6aff
IM
6688static void sched_domain_debug(struct sched_domain *sd, int cpu)
6689{
7c16ec58 6690 cpumask_t *groupmask;
4dcf6aff 6691 int level = 0;
1da177e4 6692
4dcf6aff
IM
6693 if (!sd) {
6694 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6695 return;
6696 }
1da177e4 6697
4dcf6aff
IM
6698 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6699
7c16ec58
MT
6700 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6701 if (!groupmask) {
6702 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6703 return;
6704 }
6705
4dcf6aff 6706 for (;;) {
7c16ec58 6707 if (sched_domain_debug_one(sd, cpu, level, groupmask))
4dcf6aff 6708 break;
1da177e4
LT
6709 level++;
6710 sd = sd->parent;
33859f7f 6711 if (!sd)
4dcf6aff
IM
6712 break;
6713 }
7c16ec58 6714 kfree(groupmask);
1da177e4 6715}
6d6bc0ad 6716#else /* !CONFIG_SCHED_DEBUG */
48f24c4d 6717# define sched_domain_debug(sd, cpu) do { } while (0)
6d6bc0ad 6718#endif /* CONFIG_SCHED_DEBUG */
1da177e4 6719
1a20ff27 6720static int sd_degenerate(struct sched_domain *sd)
245af2c7
SS
6721{
6722 if (cpus_weight(sd->span) == 1)
6723 return 1;
6724
6725 /* Following flags need at least 2 groups */
6726 if (sd->flags & (SD_LOAD_BALANCE |
6727 SD_BALANCE_NEWIDLE |
6728 SD_BALANCE_FORK |
89c4710e
SS
6729 SD_BALANCE_EXEC |
6730 SD_SHARE_CPUPOWER |
6731 SD_SHARE_PKG_RESOURCES)) {
245af2c7
SS
6732 if (sd->groups != sd->groups->next)
6733 return 0;
6734 }
6735
6736 /* Following flags don't use groups */
6737 if (sd->flags & (SD_WAKE_IDLE |
6738 SD_WAKE_AFFINE |
6739 SD_WAKE_BALANCE))
6740 return 0;
6741
6742 return 1;
6743}
6744
48f24c4d
IM
6745static int
6746sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
245af2c7
SS
6747{
6748 unsigned long cflags = sd->flags, pflags = parent->flags;
6749
6750 if (sd_degenerate(parent))
6751 return 1;
6752
6753 if (!cpus_equal(sd->span, parent->span))
6754 return 0;
6755
6756 /* Does parent contain flags not in child? */
6757 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6758 if (cflags & SD_WAKE_AFFINE)
6759 pflags &= ~SD_WAKE_BALANCE;
6760 /* Flags needing groups don't count if only 1 group in parent */
6761 if (parent->groups == parent->groups->next) {
6762 pflags &= ~(SD_LOAD_BALANCE |
6763 SD_BALANCE_NEWIDLE |
6764 SD_BALANCE_FORK |
89c4710e
SS
6765 SD_BALANCE_EXEC |
6766 SD_SHARE_CPUPOWER |
6767 SD_SHARE_PKG_RESOURCES);
245af2c7
SS
6768 }
6769 if (~cflags & pflags)
6770 return 0;
6771
6772 return 1;
6773}
6774
57d885fe
GH
6775static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6776{
6777 unsigned long flags;
57d885fe
GH
6778
6779 spin_lock_irqsave(&rq->lock, flags);
6780
6781 if (rq->rd) {
6782 struct root_domain *old_rd = rq->rd;
6783
1f11eb6a
GH
6784 if (cpu_isset(rq->cpu, old_rd->online))
6785 set_rq_offline(rq);
57d885fe 6786
dc938520 6787 cpu_clear(rq->cpu, old_rd->span);
dc938520 6788
57d885fe
GH
6789 if (atomic_dec_and_test(&old_rd->refcount))
6790 kfree(old_rd);
6791 }
6792
6793 atomic_inc(&rd->refcount);
6794 rq->rd = rd;
6795
dc938520 6796 cpu_set(rq->cpu, rd->span);
1f94ef59 6797 if (cpu_isset(rq->cpu, cpu_online_map))
1f11eb6a 6798 set_rq_online(rq);
57d885fe
GH
6799
6800 spin_unlock_irqrestore(&rq->lock, flags);
6801}
6802
dc938520 6803static void init_rootdomain(struct root_domain *rd)
57d885fe
GH
6804{
6805 memset(rd, 0, sizeof(*rd));
6806
dc938520
GH
6807 cpus_clear(rd->span);
6808 cpus_clear(rd->online);
6e0534f2
GH
6809
6810 cpupri_init(&rd->cpupri);
57d885fe
GH
6811}
6812
6813static void init_defrootdomain(void)
6814{
dc938520 6815 init_rootdomain(&def_root_domain);
57d885fe
GH
6816 atomic_set(&def_root_domain.refcount, 1);
6817}
6818
dc938520 6819static struct root_domain *alloc_rootdomain(void)
57d885fe
GH
6820{
6821 struct root_domain *rd;
6822
6823 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6824 if (!rd)
6825 return NULL;
6826
dc938520 6827 init_rootdomain(rd);
57d885fe
GH
6828
6829 return rd;
6830}
6831
1da177e4 6832/*
0eab9146 6833 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
1da177e4
LT
6834 * hold the hotplug lock.
6835 */
0eab9146
IM
6836static void
6837cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
1da177e4 6838{
70b97a7f 6839 struct rq *rq = cpu_rq(cpu);
245af2c7
SS
6840 struct sched_domain *tmp;
6841
6842 /* Remove the sched domains which do not contribute to scheduling. */
f29c9b1c 6843 for (tmp = sd; tmp; ) {
245af2c7
SS
6844 struct sched_domain *parent = tmp->parent;
6845 if (!parent)
6846 break;
f29c9b1c 6847
1a848870 6848 if (sd_parent_degenerate(tmp, parent)) {
245af2c7 6849 tmp->parent = parent->parent;
1a848870
SS
6850 if (parent->parent)
6851 parent->parent->child = tmp;
f29c9b1c
LZ
6852 } else
6853 tmp = tmp->parent;
245af2c7
SS
6854 }
6855
1a848870 6856 if (sd && sd_degenerate(sd)) {
245af2c7 6857 sd = sd->parent;
1a848870
SS
6858 if (sd)
6859 sd->child = NULL;
6860 }
1da177e4
LT
6861
6862 sched_domain_debug(sd, cpu);
6863
57d885fe 6864 rq_attach_root(rq, rd);
674311d5 6865 rcu_assign_pointer(rq->sd, sd);
1da177e4
LT
6866}
6867
6868/* cpus with isolated domains */
67af63a6 6869static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
1da177e4
LT
6870
6871/* Setup the mask of cpus configured for isolated domains */
6872static int __init isolated_cpu_setup(char *str)
6873{
13b40c1e
MT
6874 static int __initdata ints[NR_CPUS];
6875 int i;
1da177e4
LT
6876
6877 str = get_options(str, ARRAY_SIZE(ints), ints);
6878 cpus_clear(cpu_isolated_map);
6879 for (i = 1; i <= ints[0]; i++)
6880 if (ints[i] < NR_CPUS)
6881 cpu_set(ints[i], cpu_isolated_map);
6882 return 1;
6883}
6884
8927f494 6885__setup("isolcpus=", isolated_cpu_setup);
1da177e4
LT
6886
6887/*
6711cab4
SS
6888 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6889 * to a function which identifies what group(along with sched group) a CPU
6890 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6891 * (due to the fact that we keep track of groups covered with a cpumask_t).
1da177e4
LT
6892 *
6893 * init_sched_build_groups will build a circular linked list of the groups
6894 * covered by the given span, and will set each group's ->cpumask correctly,
6895 * and ->cpu_power to 0.
6896 */
a616058b 6897static void
7c16ec58 6898init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6711cab4 6899 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
7c16ec58
MT
6900 struct sched_group **sg,
6901 cpumask_t *tmpmask),
6902 cpumask_t *covered, cpumask_t *tmpmask)
1da177e4
LT
6903{
6904 struct sched_group *first = NULL, *last = NULL;
1da177e4
LT
6905 int i;
6906
7c16ec58
MT
6907 cpus_clear(*covered);
6908
363ab6f1 6909 for_each_cpu_mask_nr(i, *span) {
6711cab4 6910 struct sched_group *sg;
7c16ec58 6911 int group = group_fn(i, cpu_map, &sg, tmpmask);
1da177e4
LT
6912 int j;
6913
7c16ec58 6914 if (cpu_isset(i, *covered))
1da177e4
LT
6915 continue;
6916
7c16ec58 6917 cpus_clear(sg->cpumask);
5517d86b 6918 sg->__cpu_power = 0;
1da177e4 6919
363ab6f1 6920 for_each_cpu_mask_nr(j, *span) {
7c16ec58 6921 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
1da177e4
LT
6922 continue;
6923
7c16ec58 6924 cpu_set(j, *covered);
1da177e4
LT
6925 cpu_set(j, sg->cpumask);
6926 }
6927 if (!first)
6928 first = sg;
6929 if (last)
6930 last->next = sg;
6931 last = sg;
6932 }
6933 last->next = first;
6934}
6935
9c1cfda2 6936#define SD_NODES_PER_DOMAIN 16
1da177e4 6937
9c1cfda2 6938#ifdef CONFIG_NUMA
198e2f18 6939
9c1cfda2
JH
6940/**
6941 * find_next_best_node - find the next node to include in a sched_domain
6942 * @node: node whose sched_domain we're building
6943 * @used_nodes: nodes already in the sched_domain
6944 *
41a2d6cf 6945 * Find the next node to include in a given scheduling domain. Simply
9c1cfda2
JH
6946 * finds the closest node not already in the @used_nodes map.
6947 *
6948 * Should use nodemask_t.
6949 */
c5f59f08 6950static int find_next_best_node(int node, nodemask_t *used_nodes)
9c1cfda2
JH
6951{
6952 int i, n, val, min_val, best_node = 0;
6953
6954 min_val = INT_MAX;
6955
076ac2af 6956 for (i = 0; i < nr_node_ids; i++) {
9c1cfda2 6957 /* Start at @node */
076ac2af 6958 n = (node + i) % nr_node_ids;
9c1cfda2
JH
6959
6960 if (!nr_cpus_node(n))
6961 continue;
6962
6963 /* Skip already used nodes */
c5f59f08 6964 if (node_isset(n, *used_nodes))
9c1cfda2
JH
6965 continue;
6966
6967 /* Simple min distance search */
6968 val = node_distance(node, n);
6969
6970 if (val < min_val) {
6971 min_val = val;
6972 best_node = n;
6973 }
6974 }
6975
c5f59f08 6976 node_set(best_node, *used_nodes);
9c1cfda2
JH
6977 return best_node;
6978}
6979
6980/**
6981 * sched_domain_node_span - get a cpumask for a node's sched_domain
6982 * @node: node whose cpumask we're constructing
73486722 6983 * @span: resulting cpumask
9c1cfda2 6984 *
41a2d6cf 6985 * Given a node, construct a good cpumask for its sched_domain to span. It
9c1cfda2
JH
6986 * should be one that prevents unnecessary balancing, but also spreads tasks
6987 * out optimally.
6988 */
4bdbaad3 6989static void sched_domain_node_span(int node, cpumask_t *span)
9c1cfda2 6990{
c5f59f08 6991 nodemask_t used_nodes;
c5f59f08 6992 node_to_cpumask_ptr(nodemask, node);
48f24c4d 6993 int i;
9c1cfda2 6994
4bdbaad3 6995 cpus_clear(*span);
c5f59f08 6996 nodes_clear(used_nodes);
9c1cfda2 6997
4bdbaad3 6998 cpus_or(*span, *span, *nodemask);
c5f59f08 6999 node_set(node, used_nodes);
9c1cfda2
JH
7000
7001 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
c5f59f08 7002 int next_node = find_next_best_node(node, &used_nodes);
48f24c4d 7003
c5f59f08 7004 node_to_cpumask_ptr_next(nodemask, next_node);
4bdbaad3 7005 cpus_or(*span, *span, *nodemask);
9c1cfda2 7006 }
9c1cfda2 7007}
6d6bc0ad 7008#endif /* CONFIG_NUMA */
9c1cfda2 7009
5c45bf27 7010int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
48f24c4d 7011
9c1cfda2 7012/*
48f24c4d 7013 * SMT sched-domains:
9c1cfda2 7014 */
1da177e4
LT
7015#ifdef CONFIG_SCHED_SMT
7016static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6711cab4 7017static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
48f24c4d 7018
41a2d6cf 7019static int
7c16ec58
MT
7020cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7021 cpumask_t *unused)
1da177e4 7022{
6711cab4
SS
7023 if (sg)
7024 *sg = &per_cpu(sched_group_cpus, cpu);
1da177e4
LT
7025 return cpu;
7026}
6d6bc0ad 7027#endif /* CONFIG_SCHED_SMT */
1da177e4 7028
48f24c4d
IM
7029/*
7030 * multi-core sched-domains:
7031 */
1e9f28fa
SS
7032#ifdef CONFIG_SCHED_MC
7033static DEFINE_PER_CPU(struct sched_domain, core_domains);
6711cab4 7034static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6d6bc0ad 7035#endif /* CONFIG_SCHED_MC */
1e9f28fa
SS
7036
7037#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
41a2d6cf 7038static int
7c16ec58
MT
7039cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7040 cpumask_t *mask)
1e9f28fa 7041{
6711cab4 7042 int group;
7c16ec58
MT
7043
7044 *mask = per_cpu(cpu_sibling_map, cpu);
7045 cpus_and(*mask, *mask, *cpu_map);
7046 group = first_cpu(*mask);
6711cab4
SS
7047 if (sg)
7048 *sg = &per_cpu(sched_group_core, group);
7049 return group;
1e9f28fa
SS
7050}
7051#elif defined(CONFIG_SCHED_MC)
41a2d6cf 7052static int
7c16ec58
MT
7053cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7054 cpumask_t *unused)
1e9f28fa 7055{
6711cab4
SS
7056 if (sg)
7057 *sg = &per_cpu(sched_group_core, cpu);
1e9f28fa
SS
7058 return cpu;
7059}
7060#endif
7061
1da177e4 7062static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6711cab4 7063static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
48f24c4d 7064
41a2d6cf 7065static int
7c16ec58
MT
7066cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7067 cpumask_t *mask)
1da177e4 7068{
6711cab4 7069 int group;
48f24c4d 7070#ifdef CONFIG_SCHED_MC
7c16ec58
MT
7071 *mask = cpu_coregroup_map(cpu);
7072 cpus_and(*mask, *mask, *cpu_map);
7073 group = first_cpu(*mask);
1e9f28fa 7074#elif defined(CONFIG_SCHED_SMT)
7c16ec58
MT
7075 *mask = per_cpu(cpu_sibling_map, cpu);
7076 cpus_and(*mask, *mask, *cpu_map);
7077 group = first_cpu(*mask);
1da177e4 7078#else
6711cab4 7079 group = cpu;
1da177e4 7080#endif
6711cab4
SS
7081 if (sg)
7082 *sg = &per_cpu(sched_group_phys, group);
7083 return group;
1da177e4
LT
7084}
7085
7086#ifdef CONFIG_NUMA
1da177e4 7087/*
9c1cfda2
JH
7088 * The init_sched_build_groups can't handle what we want to do with node
7089 * groups, so roll our own. Now each node has its own list of groups which
7090 * gets dynamically allocated.
1da177e4 7091 */
9c1cfda2 7092static DEFINE_PER_CPU(struct sched_domain, node_domains);
434d53b0 7093static struct sched_group ***sched_group_nodes_bycpu;
1da177e4 7094
9c1cfda2 7095static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6711cab4 7096static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
9c1cfda2 7097
6711cab4 7098static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7c16ec58 7099 struct sched_group **sg, cpumask_t *nodemask)
9c1cfda2 7100{
6711cab4 7101 int group;
ea6f18ed 7102 node_to_cpumask_ptr(pnodemask, cpu_to_node(cpu));
6711cab4 7103
ea6f18ed 7104 cpus_and(*nodemask, *pnodemask, *cpu_map);
7c16ec58 7105 group = first_cpu(*nodemask);
6711cab4
SS
7106
7107 if (sg)
7108 *sg = &per_cpu(sched_group_allnodes, group);
7109 return group;
1da177e4 7110}
6711cab4 7111
08069033
SS
7112static void init_numa_sched_groups_power(struct sched_group *group_head)
7113{
7114 struct sched_group *sg = group_head;
7115 int j;
7116
7117 if (!sg)
7118 return;
3a5c359a 7119 do {
363ab6f1 7120 for_each_cpu_mask_nr(j, sg->cpumask) {
3a5c359a 7121 struct sched_domain *sd;
08069033 7122
3a5c359a
AK
7123 sd = &per_cpu(phys_domains, j);
7124 if (j != first_cpu(sd->groups->cpumask)) {
7125 /*
7126 * Only add "power" once for each
7127 * physical package.
7128 */
7129 continue;
7130 }
08069033 7131
3a5c359a
AK
7132 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7133 }
7134 sg = sg->next;
7135 } while (sg != group_head);
08069033 7136}
6d6bc0ad 7137#endif /* CONFIG_NUMA */
1da177e4 7138
a616058b 7139#ifdef CONFIG_NUMA
51888ca2 7140/* Free memory allocated for various sched_group structures */
7c16ec58 7141static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
51888ca2 7142{
a616058b 7143 int cpu, i;
51888ca2 7144
363ab6f1 7145 for_each_cpu_mask_nr(cpu, *cpu_map) {
51888ca2
SV
7146 struct sched_group **sched_group_nodes
7147 = sched_group_nodes_bycpu[cpu];
7148
51888ca2
SV
7149 if (!sched_group_nodes)
7150 continue;
7151
076ac2af 7152 for (i = 0; i < nr_node_ids; i++) {
51888ca2 7153 struct sched_group *oldsg, *sg = sched_group_nodes[i];
ea6f18ed 7154 node_to_cpumask_ptr(pnodemask, i);
51888ca2 7155
ea6f18ed 7156 cpus_and(*nodemask, *pnodemask, *cpu_map);
7c16ec58 7157 if (cpus_empty(*nodemask))
51888ca2
SV
7158 continue;
7159
7160 if (sg == NULL)
7161 continue;
7162 sg = sg->next;
7163next_sg:
7164 oldsg = sg;
7165 sg = sg->next;
7166 kfree(oldsg);
7167 if (oldsg != sched_group_nodes[i])
7168 goto next_sg;
7169 }
7170 kfree(sched_group_nodes);
7171 sched_group_nodes_bycpu[cpu] = NULL;
7172 }
51888ca2 7173}
6d6bc0ad 7174#else /* !CONFIG_NUMA */
7c16ec58 7175static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
a616058b
SS
7176{
7177}
6d6bc0ad 7178#endif /* CONFIG_NUMA */
51888ca2 7179
89c4710e
SS
7180/*
7181 * Initialize sched groups cpu_power.
7182 *
7183 * cpu_power indicates the capacity of sched group, which is used while
7184 * distributing the load between different sched groups in a sched domain.
7185 * Typically cpu_power for all the groups in a sched domain will be same unless
7186 * there are asymmetries in the topology. If there are asymmetries, group
7187 * having more cpu_power will pickup more load compared to the group having
7188 * less cpu_power.
7189 *
7190 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7191 * the maximum number of tasks a group can handle in the presence of other idle
7192 * or lightly loaded groups in the same sched domain.
7193 */
7194static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7195{
7196 struct sched_domain *child;
7197 struct sched_group *group;
7198
7199 WARN_ON(!sd || !sd->groups);
7200
7201 if (cpu != first_cpu(sd->groups->cpumask))
7202 return;
7203
7204 child = sd->child;
7205
5517d86b
ED
7206 sd->groups->__cpu_power = 0;
7207
89c4710e
SS
7208 /*
7209 * For perf policy, if the groups in child domain share resources
7210 * (for example cores sharing some portions of the cache hierarchy
7211 * or SMT), then set this domain groups cpu_power such that each group
7212 * can handle only one task, when there are other idle groups in the
7213 * same sched domain.
7214 */
7215 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7216 (child->flags &
7217 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5517d86b 7218 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
89c4710e
SS
7219 return;
7220 }
7221
89c4710e
SS
7222 /*
7223 * add cpu_power of each child group to this groups cpu_power
7224 */
7225 group = child->groups;
7226 do {
5517d86b 7227 sg_inc_cpu_power(sd->groups, group->__cpu_power);
89c4710e
SS
7228 group = group->next;
7229 } while (group != child->groups);
7230}
7231
7c16ec58
MT
7232/*
7233 * Initializers for schedule domains
7234 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7235 */
7236
a5d8c348
IM
7237#ifdef CONFIG_SCHED_DEBUG
7238# define SD_INIT_NAME(sd, type) sd->name = #type
7239#else
7240# define SD_INIT_NAME(sd, type) do { } while (0)
7241#endif
7242
7c16ec58 7243#define SD_INIT(sd, type) sd_init_##type(sd)
a5d8c348 7244
7c16ec58
MT
7245#define SD_INIT_FUNC(type) \
7246static noinline void sd_init_##type(struct sched_domain *sd) \
7247{ \
7248 memset(sd, 0, sizeof(*sd)); \
7249 *sd = SD_##type##_INIT; \
1d3504fc 7250 sd->level = SD_LV_##type; \
a5d8c348 7251 SD_INIT_NAME(sd, type); \
7c16ec58
MT
7252}
7253
7254SD_INIT_FUNC(CPU)
7255#ifdef CONFIG_NUMA
7256 SD_INIT_FUNC(ALLNODES)
7257 SD_INIT_FUNC(NODE)
7258#endif
7259#ifdef CONFIG_SCHED_SMT
7260 SD_INIT_FUNC(SIBLING)
7261#endif
7262#ifdef CONFIG_SCHED_MC
7263 SD_INIT_FUNC(MC)
7264#endif
7265
7266/*
7267 * To minimize stack usage kmalloc room for cpumasks and share the
7268 * space as the usage in build_sched_domains() dictates. Used only
7269 * if the amount of space is significant.
7270 */
7271struct allmasks {
7272 cpumask_t tmpmask; /* make this one first */
7273 union {
7274 cpumask_t nodemask;
7275 cpumask_t this_sibling_map;
7276 cpumask_t this_core_map;
7277 };
7278 cpumask_t send_covered;
7279
7280#ifdef CONFIG_NUMA
7281 cpumask_t domainspan;
7282 cpumask_t covered;
7283 cpumask_t notcovered;
7284#endif
7285};
7286
7287#if NR_CPUS > 128
6d21cd62
LZ
7288#define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7289static inline void sched_cpumask_alloc(struct allmasks **masks)
7290{
7291 *masks = kmalloc(sizeof(**masks), GFP_KERNEL);
7292}
7293static inline void sched_cpumask_free(struct allmasks *masks)
7294{
7295 kfree(masks);
7296}
7c16ec58 7297#else
6d21cd62
LZ
7298#define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7299static inline void sched_cpumask_alloc(struct allmasks **masks)
7300{ }
7301static inline void sched_cpumask_free(struct allmasks *masks)
7302{ }
7c16ec58
MT
7303#endif
7304
7305#define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7306 ((unsigned long)(a) + offsetof(struct allmasks, v))
7307
1d3504fc
HS
7308static int default_relax_domain_level = -1;
7309
7310static int __init setup_relax_domain_level(char *str)
7311{
30e0e178
LZ
7312 unsigned long val;
7313
7314 val = simple_strtoul(str, NULL, 0);
7315 if (val < SD_LV_MAX)
7316 default_relax_domain_level = val;
7317
1d3504fc
HS
7318 return 1;
7319}
7320__setup("relax_domain_level=", setup_relax_domain_level);
7321
7322static void set_domain_attribute(struct sched_domain *sd,
7323 struct sched_domain_attr *attr)
7324{
7325 int request;
7326
7327 if (!attr || attr->relax_domain_level < 0) {
7328 if (default_relax_domain_level < 0)
7329 return;
7330 else
7331 request = default_relax_domain_level;
7332 } else
7333 request = attr->relax_domain_level;
7334 if (request < sd->level) {
7335 /* turn off idle balance on this domain */
7336 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7337 } else {
7338 /* turn on idle balance on this domain */
7339 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7340 }
7341}
7342
1da177e4 7343/*
1a20ff27
DG
7344 * Build sched domains for a given set of cpus and attach the sched domains
7345 * to the individual cpus
1da177e4 7346 */
1d3504fc
HS
7347static int __build_sched_domains(const cpumask_t *cpu_map,
7348 struct sched_domain_attr *attr)
1da177e4
LT
7349{
7350 int i;
57d885fe 7351 struct root_domain *rd;
7c16ec58
MT
7352 SCHED_CPUMASK_DECLARE(allmasks);
7353 cpumask_t *tmpmask;
d1b55138
JH
7354#ifdef CONFIG_NUMA
7355 struct sched_group **sched_group_nodes = NULL;
6711cab4 7356 int sd_allnodes = 0;
d1b55138
JH
7357
7358 /*
7359 * Allocate the per-node list of sched groups
7360 */
076ac2af 7361 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
41a2d6cf 7362 GFP_KERNEL);
d1b55138
JH
7363 if (!sched_group_nodes) {
7364 printk(KERN_WARNING "Can not alloc sched group node list\n");
51888ca2 7365 return -ENOMEM;
d1b55138 7366 }
d1b55138 7367#endif
1da177e4 7368
dc938520 7369 rd = alloc_rootdomain();
57d885fe
GH
7370 if (!rd) {
7371 printk(KERN_WARNING "Cannot alloc root domain\n");
7c16ec58
MT
7372#ifdef CONFIG_NUMA
7373 kfree(sched_group_nodes);
7374#endif
57d885fe
GH
7375 return -ENOMEM;
7376 }
7377
7c16ec58 7378 /* get space for all scratch cpumask variables */
6d21cd62 7379 sched_cpumask_alloc(&allmasks);
7c16ec58
MT
7380 if (!allmasks) {
7381 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7382 kfree(rd);
7383#ifdef CONFIG_NUMA
7384 kfree(sched_group_nodes);
7385#endif
7386 return -ENOMEM;
7387 }
6d21cd62 7388
7c16ec58
MT
7389 tmpmask = (cpumask_t *)allmasks;
7390
7391
7392#ifdef CONFIG_NUMA
7393 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7394#endif
7395
1da177e4 7396 /*
1a20ff27 7397 * Set up domains for cpus specified by the cpu_map.
1da177e4 7398 */
363ab6f1 7399 for_each_cpu_mask_nr(i, *cpu_map) {
1da177e4 7400 struct sched_domain *sd = NULL, *p;
7c16ec58 7401 SCHED_CPUMASK_VAR(nodemask, allmasks);
1da177e4 7402
7c16ec58
MT
7403 *nodemask = node_to_cpumask(cpu_to_node(i));
7404 cpus_and(*nodemask, *nodemask, *cpu_map);
1da177e4
LT
7405
7406#ifdef CONFIG_NUMA
dd41f596 7407 if (cpus_weight(*cpu_map) >
7c16ec58 7408 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
9c1cfda2 7409 sd = &per_cpu(allnodes_domains, i);
7c16ec58 7410 SD_INIT(sd, ALLNODES);
1d3504fc 7411 set_domain_attribute(sd, attr);
9c1cfda2 7412 sd->span = *cpu_map;
7c16ec58 7413 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
9c1cfda2 7414 p = sd;
6711cab4 7415 sd_allnodes = 1;
9c1cfda2
JH
7416 } else
7417 p = NULL;
7418
1da177e4 7419 sd = &per_cpu(node_domains, i);
7c16ec58 7420 SD_INIT(sd, NODE);
1d3504fc 7421 set_domain_attribute(sd, attr);
4bdbaad3 7422 sched_domain_node_span(cpu_to_node(i), &sd->span);
9c1cfda2 7423 sd->parent = p;
1a848870
SS
7424 if (p)
7425 p->child = sd;
9c1cfda2 7426 cpus_and(sd->span, sd->span, *cpu_map);
1da177e4
LT
7427#endif
7428
7429 p = sd;
7430 sd = &per_cpu(phys_domains, i);
7c16ec58 7431 SD_INIT(sd, CPU);
1d3504fc 7432 set_domain_attribute(sd, attr);
7c16ec58 7433 sd->span = *nodemask;
1da177e4 7434 sd->parent = p;
1a848870
SS
7435 if (p)
7436 p->child = sd;
7c16ec58 7437 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
1da177e4 7438
1e9f28fa
SS
7439#ifdef CONFIG_SCHED_MC
7440 p = sd;
7441 sd = &per_cpu(core_domains, i);
7c16ec58 7442 SD_INIT(sd, MC);
1d3504fc 7443 set_domain_attribute(sd, attr);
1e9f28fa
SS
7444 sd->span = cpu_coregroup_map(i);
7445 cpus_and(sd->span, sd->span, *cpu_map);
7446 sd->parent = p;
1a848870 7447 p->child = sd;
7c16ec58 7448 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
1e9f28fa
SS
7449#endif
7450
1da177e4
LT
7451#ifdef CONFIG_SCHED_SMT
7452 p = sd;
7453 sd = &per_cpu(cpu_domains, i);
7c16ec58 7454 SD_INIT(sd, SIBLING);
1d3504fc 7455 set_domain_attribute(sd, attr);
d5a7430d 7456 sd->span = per_cpu(cpu_sibling_map, i);
1a20ff27 7457 cpus_and(sd->span, sd->span, *cpu_map);
1da177e4 7458 sd->parent = p;
1a848870 7459 p->child = sd;
7c16ec58 7460 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
1da177e4
LT
7461#endif
7462 }
7463
7464#ifdef CONFIG_SCHED_SMT
7465 /* Set up CPU (sibling) groups */
363ab6f1 7466 for_each_cpu_mask_nr(i, *cpu_map) {
7c16ec58
MT
7467 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7468 SCHED_CPUMASK_VAR(send_covered, allmasks);
7469
7470 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7471 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7472 if (i != first_cpu(*this_sibling_map))
1da177e4
LT
7473 continue;
7474
dd41f596 7475 init_sched_build_groups(this_sibling_map, cpu_map,
7c16ec58
MT
7476 &cpu_to_cpu_group,
7477 send_covered, tmpmask);
1da177e4
LT
7478 }
7479#endif
7480
1e9f28fa
SS
7481#ifdef CONFIG_SCHED_MC
7482 /* Set up multi-core groups */
363ab6f1 7483 for_each_cpu_mask_nr(i, *cpu_map) {
7c16ec58
MT
7484 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7485 SCHED_CPUMASK_VAR(send_covered, allmasks);
7486
7487 *this_core_map = cpu_coregroup_map(i);
7488 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7489 if (i != first_cpu(*this_core_map))
1e9f28fa 7490 continue;
7c16ec58 7491
dd41f596 7492 init_sched_build_groups(this_core_map, cpu_map,
7c16ec58
MT
7493 &cpu_to_core_group,
7494 send_covered, tmpmask);
1e9f28fa
SS
7495 }
7496#endif
7497
1da177e4 7498 /* Set up physical groups */
076ac2af 7499 for (i = 0; i < nr_node_ids; i++) {
7c16ec58
MT
7500 SCHED_CPUMASK_VAR(nodemask, allmasks);
7501 SCHED_CPUMASK_VAR(send_covered, allmasks);
1da177e4 7502
7c16ec58
MT
7503 *nodemask = node_to_cpumask(i);
7504 cpus_and(*nodemask, *nodemask, *cpu_map);
7505 if (cpus_empty(*nodemask))
1da177e4
LT
7506 continue;
7507
7c16ec58
MT
7508 init_sched_build_groups(nodemask, cpu_map,
7509 &cpu_to_phys_group,
7510 send_covered, tmpmask);
1da177e4
LT
7511 }
7512
7513#ifdef CONFIG_NUMA
7514 /* Set up node groups */
7c16ec58
MT
7515 if (sd_allnodes) {
7516 SCHED_CPUMASK_VAR(send_covered, allmasks);
7517
7518 init_sched_build_groups(cpu_map, cpu_map,
7519 &cpu_to_allnodes_group,
7520 send_covered, tmpmask);
7521 }
9c1cfda2 7522
076ac2af 7523 for (i = 0; i < nr_node_ids; i++) {
9c1cfda2
JH
7524 /* Set up node groups */
7525 struct sched_group *sg, *prev;
7c16ec58
MT
7526 SCHED_CPUMASK_VAR(nodemask, allmasks);
7527 SCHED_CPUMASK_VAR(domainspan, allmasks);
7528 SCHED_CPUMASK_VAR(covered, allmasks);
9c1cfda2
JH
7529 int j;
7530
7c16ec58
MT
7531 *nodemask = node_to_cpumask(i);
7532 cpus_clear(*covered);
7533
7534 cpus_and(*nodemask, *nodemask, *cpu_map);
7535 if (cpus_empty(*nodemask)) {
d1b55138 7536 sched_group_nodes[i] = NULL;
9c1cfda2 7537 continue;
d1b55138 7538 }
9c1cfda2 7539
4bdbaad3 7540 sched_domain_node_span(i, domainspan);
7c16ec58 7541 cpus_and(*domainspan, *domainspan, *cpu_map);
9c1cfda2 7542
15f0b676 7543 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
51888ca2
SV
7544 if (!sg) {
7545 printk(KERN_WARNING "Can not alloc domain group for "
7546 "node %d\n", i);
7547 goto error;
7548 }
9c1cfda2 7549 sched_group_nodes[i] = sg;
363ab6f1 7550 for_each_cpu_mask_nr(j, *nodemask) {
9c1cfda2 7551 struct sched_domain *sd;
9761eea8 7552
9c1cfda2
JH
7553 sd = &per_cpu(node_domains, j);
7554 sd->groups = sg;
9c1cfda2 7555 }
5517d86b 7556 sg->__cpu_power = 0;
7c16ec58 7557 sg->cpumask = *nodemask;
51888ca2 7558 sg->next = sg;
7c16ec58 7559 cpus_or(*covered, *covered, *nodemask);
9c1cfda2
JH
7560 prev = sg;
7561
076ac2af 7562 for (j = 0; j < nr_node_ids; j++) {
7c16ec58 7563 SCHED_CPUMASK_VAR(notcovered, allmasks);
076ac2af 7564 int n = (i + j) % nr_node_ids;
c5f59f08 7565 node_to_cpumask_ptr(pnodemask, n);
9c1cfda2 7566
7c16ec58
MT
7567 cpus_complement(*notcovered, *covered);
7568 cpus_and(*tmpmask, *notcovered, *cpu_map);
7569 cpus_and(*tmpmask, *tmpmask, *domainspan);
7570 if (cpus_empty(*tmpmask))
9c1cfda2
JH
7571 break;
7572
7c16ec58
MT
7573 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7574 if (cpus_empty(*tmpmask))
9c1cfda2
JH
7575 continue;
7576
15f0b676
SV
7577 sg = kmalloc_node(sizeof(struct sched_group),
7578 GFP_KERNEL, i);
9c1cfda2
JH
7579 if (!sg) {
7580 printk(KERN_WARNING
7581 "Can not alloc domain group for node %d\n", j);
51888ca2 7582 goto error;
9c1cfda2 7583 }
5517d86b 7584 sg->__cpu_power = 0;
7c16ec58 7585 sg->cpumask = *tmpmask;
51888ca2 7586 sg->next = prev->next;
7c16ec58 7587 cpus_or(*covered, *covered, *tmpmask);
9c1cfda2
JH
7588 prev->next = sg;
7589 prev = sg;
7590 }
9c1cfda2 7591 }
1da177e4
LT
7592#endif
7593
7594 /* Calculate CPU power for physical packages and nodes */
5c45bf27 7595#ifdef CONFIG_SCHED_SMT
363ab6f1 7596 for_each_cpu_mask_nr(i, *cpu_map) {
dd41f596
IM
7597 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7598
89c4710e 7599 init_sched_groups_power(i, sd);
5c45bf27 7600 }
1da177e4 7601#endif
1e9f28fa 7602#ifdef CONFIG_SCHED_MC
363ab6f1 7603 for_each_cpu_mask_nr(i, *cpu_map) {
dd41f596
IM
7604 struct sched_domain *sd = &per_cpu(core_domains, i);
7605
89c4710e 7606 init_sched_groups_power(i, sd);
5c45bf27
SS
7607 }
7608#endif
1e9f28fa 7609
363ab6f1 7610 for_each_cpu_mask_nr(i, *cpu_map) {
dd41f596
IM
7611 struct sched_domain *sd = &per_cpu(phys_domains, i);
7612
89c4710e 7613 init_sched_groups_power(i, sd);
1da177e4
LT
7614 }
7615
9c1cfda2 7616#ifdef CONFIG_NUMA
076ac2af 7617 for (i = 0; i < nr_node_ids; i++)
08069033 7618 init_numa_sched_groups_power(sched_group_nodes[i]);
9c1cfda2 7619
6711cab4
SS
7620 if (sd_allnodes) {
7621 struct sched_group *sg;
f712c0c7 7622
7c16ec58
MT
7623 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7624 tmpmask);
f712c0c7
SS
7625 init_numa_sched_groups_power(sg);
7626 }
9c1cfda2
JH
7627#endif
7628
1da177e4 7629 /* Attach the domains */
363ab6f1 7630 for_each_cpu_mask_nr(i, *cpu_map) {
1da177e4
LT
7631 struct sched_domain *sd;
7632#ifdef CONFIG_SCHED_SMT
7633 sd = &per_cpu(cpu_domains, i);
1e9f28fa
SS
7634#elif defined(CONFIG_SCHED_MC)
7635 sd = &per_cpu(core_domains, i);
1da177e4
LT
7636#else
7637 sd = &per_cpu(phys_domains, i);
7638#endif
57d885fe 7639 cpu_attach_domain(sd, rd, i);
1da177e4 7640 }
51888ca2 7641
6d21cd62 7642 sched_cpumask_free(allmasks);
51888ca2
SV
7643 return 0;
7644
a616058b 7645#ifdef CONFIG_NUMA
51888ca2 7646error:
7c16ec58 7647 free_sched_groups(cpu_map, tmpmask);
6d21cd62 7648 sched_cpumask_free(allmasks);
ca3273f9 7649 kfree(rd);
51888ca2 7650 return -ENOMEM;
a616058b 7651#endif
1da177e4 7652}
029190c5 7653
1d3504fc
HS
7654static int build_sched_domains(const cpumask_t *cpu_map)
7655{
7656 return __build_sched_domains(cpu_map, NULL);
7657}
7658
029190c5
PJ
7659static cpumask_t *doms_cur; /* current sched domains */
7660static int ndoms_cur; /* number of sched domains in 'doms_cur' */
4285f594
IM
7661static struct sched_domain_attr *dattr_cur;
7662 /* attribues of custom domains in 'doms_cur' */
029190c5
PJ
7663
7664/*
7665 * Special case: If a kmalloc of a doms_cur partition (array of
7666 * cpumask_t) fails, then fallback to a single sched domain,
7667 * as determined by the single cpumask_t fallback_doms.
7668 */
7669static cpumask_t fallback_doms;
7670
22e52b07
HC
7671void __attribute__((weak)) arch_update_cpu_topology(void)
7672{
7673}
7674
1a20ff27 7675/*
41a2d6cf 7676 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
029190c5
PJ
7677 * For now this just excludes isolated cpus, but could be used to
7678 * exclude other special cases in the future.
1a20ff27 7679 */
51888ca2 7680static int arch_init_sched_domains(const cpumask_t *cpu_map)
1a20ff27 7681{
7378547f
MM
7682 int err;
7683
22e52b07 7684 arch_update_cpu_topology();
029190c5
PJ
7685 ndoms_cur = 1;
7686 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7687 if (!doms_cur)
7688 doms_cur = &fallback_doms;
7689 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
1d3504fc 7690 dattr_cur = NULL;
7378547f 7691 err = build_sched_domains(doms_cur);
6382bc90 7692 register_sched_domain_sysctl();
7378547f
MM
7693
7694 return err;
1a20ff27
DG
7695}
7696
7c16ec58
MT
7697static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7698 cpumask_t *tmpmask)
1da177e4 7699{
7c16ec58 7700 free_sched_groups(cpu_map, tmpmask);
9c1cfda2 7701}
1da177e4 7702
1a20ff27
DG
7703/*
7704 * Detach sched domains from a group of cpus specified in cpu_map
7705 * These cpus will now be attached to the NULL domain
7706 */
858119e1 7707static void detach_destroy_domains(const cpumask_t *cpu_map)
1a20ff27 7708{
7c16ec58 7709 cpumask_t tmpmask;
1a20ff27
DG
7710 int i;
7711
363ab6f1 7712 for_each_cpu_mask_nr(i, *cpu_map)
57d885fe 7713 cpu_attach_domain(NULL, &def_root_domain, i);
1a20ff27 7714 synchronize_sched();
7c16ec58 7715 arch_destroy_sched_domains(cpu_map, &tmpmask);
1a20ff27
DG
7716}
7717
1d3504fc
HS
7718/* handle null as "default" */
7719static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7720 struct sched_domain_attr *new, int idx_new)
7721{
7722 struct sched_domain_attr tmp;
7723
7724 /* fast path */
7725 if (!new && !cur)
7726 return 1;
7727
7728 tmp = SD_ATTR_INIT;
7729 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7730 new ? (new + idx_new) : &tmp,
7731 sizeof(struct sched_domain_attr));
7732}
7733
029190c5
PJ
7734/*
7735 * Partition sched domains as specified by the 'ndoms_new'
41a2d6cf 7736 * cpumasks in the array doms_new[] of cpumasks. This compares
029190c5
PJ
7737 * doms_new[] to the current sched domain partitioning, doms_cur[].
7738 * It destroys each deleted domain and builds each new domain.
7739 *
7740 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
41a2d6cf
IM
7741 * The masks don't intersect (don't overlap.) We should setup one
7742 * sched domain for each mask. CPUs not in any of the cpumasks will
7743 * not be load balanced. If the same cpumask appears both in the
029190c5
PJ
7744 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7745 * it as it is.
7746 *
41a2d6cf
IM
7747 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7748 * ownership of it and will kfree it when done with it. If the caller
700018e0
LZ
7749 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7750 * ndoms_new == 1, and partition_sched_domains() will fallback to
7751 * the single partition 'fallback_doms', it also forces the domains
7752 * to be rebuilt.
029190c5 7753 *
700018e0
LZ
7754 * If doms_new == NULL it will be replaced with cpu_online_map.
7755 * ndoms_new == 0 is a special case for destroying existing domains,
7756 * and it will not create the default domain.
dfb512ec 7757 *
029190c5
PJ
7758 * Call with hotplug lock held
7759 */
1d3504fc
HS
7760void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7761 struct sched_domain_attr *dattr_new)
029190c5 7762{
dfb512ec 7763 int i, j, n;
029190c5 7764
712555ee 7765 mutex_lock(&sched_domains_mutex);
a1835615 7766
7378547f
MM
7767 /* always unregister in case we don't destroy any domains */
7768 unregister_sched_domain_sysctl();
7769
dfb512ec 7770 n = doms_new ? ndoms_new : 0;
029190c5
PJ
7771
7772 /* Destroy deleted domains */
7773 for (i = 0; i < ndoms_cur; i++) {
dfb512ec 7774 for (j = 0; j < n; j++) {
1d3504fc
HS
7775 if (cpus_equal(doms_cur[i], doms_new[j])
7776 && dattrs_equal(dattr_cur, i, dattr_new, j))
029190c5
PJ
7777 goto match1;
7778 }
7779 /* no match - a current sched domain not in new doms_new[] */
7780 detach_destroy_domains(doms_cur + i);
7781match1:
7782 ;
7783 }
7784
e761b772
MK
7785 if (doms_new == NULL) {
7786 ndoms_cur = 0;
e761b772
MK
7787 doms_new = &fallback_doms;
7788 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
faa2f98f 7789 WARN_ON_ONCE(dattr_new);
e761b772
MK
7790 }
7791
029190c5
PJ
7792 /* Build new domains */
7793 for (i = 0; i < ndoms_new; i++) {
7794 for (j = 0; j < ndoms_cur; j++) {
1d3504fc
HS
7795 if (cpus_equal(doms_new[i], doms_cur[j])
7796 && dattrs_equal(dattr_new, i, dattr_cur, j))
029190c5
PJ
7797 goto match2;
7798 }
7799 /* no match - add a new doms_new */
1d3504fc
HS
7800 __build_sched_domains(doms_new + i,
7801 dattr_new ? dattr_new + i : NULL);
029190c5
PJ
7802match2:
7803 ;
7804 }
7805
7806 /* Remember the new sched domains */
7807 if (doms_cur != &fallback_doms)
7808 kfree(doms_cur);
1d3504fc 7809 kfree(dattr_cur); /* kfree(NULL) is safe */
029190c5 7810 doms_cur = doms_new;
1d3504fc 7811 dattr_cur = dattr_new;
029190c5 7812 ndoms_cur = ndoms_new;
7378547f
MM
7813
7814 register_sched_domain_sysctl();
a1835615 7815
712555ee 7816 mutex_unlock(&sched_domains_mutex);
029190c5
PJ
7817}
7818
5c45bf27 7819#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9aefd0ab 7820int arch_reinit_sched_domains(void)
5c45bf27 7821{
95402b38 7822 get_online_cpus();
dfb512ec
MK
7823
7824 /* Destroy domains first to force the rebuild */
7825 partition_sched_domains(0, NULL, NULL);
7826
e761b772 7827 rebuild_sched_domains();
95402b38 7828 put_online_cpus();
dfb512ec 7829
e761b772 7830 return 0;
5c45bf27
SS
7831}
7832
7833static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7834{
7835 int ret;
7836
7837 if (buf[0] != '0' && buf[0] != '1')
7838 return -EINVAL;
7839
7840 if (smt)
7841 sched_smt_power_savings = (buf[0] == '1');
7842 else
7843 sched_mc_power_savings = (buf[0] == '1');
7844
7845 ret = arch_reinit_sched_domains();
7846
7847 return ret ? ret : count;
7848}
7849
5c45bf27 7850#ifdef CONFIG_SCHED_MC
f718cd4a
AK
7851static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7852 char *page)
5c45bf27
SS
7853{
7854 return sprintf(page, "%u\n", sched_mc_power_savings);
7855}
f718cd4a 7856static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
48f24c4d 7857 const char *buf, size_t count)
5c45bf27
SS
7858{
7859 return sched_power_savings_store(buf, count, 0);
7860}
f718cd4a
AK
7861static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7862 sched_mc_power_savings_show,
7863 sched_mc_power_savings_store);
5c45bf27
SS
7864#endif
7865
7866#ifdef CONFIG_SCHED_SMT
f718cd4a
AK
7867static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7868 char *page)
5c45bf27
SS
7869{
7870 return sprintf(page, "%u\n", sched_smt_power_savings);
7871}
f718cd4a 7872static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
48f24c4d 7873 const char *buf, size_t count)
5c45bf27
SS
7874{
7875 return sched_power_savings_store(buf, count, 1);
7876}
f718cd4a
AK
7877static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7878 sched_smt_power_savings_show,
6707de00
AB
7879 sched_smt_power_savings_store);
7880#endif
7881
7882int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7883{
7884 int err = 0;
7885
7886#ifdef CONFIG_SCHED_SMT
7887 if (smt_capable())
7888 err = sysfs_create_file(&cls->kset.kobj,
7889 &attr_sched_smt_power_savings.attr);
7890#endif
7891#ifdef CONFIG_SCHED_MC
7892 if (!err && mc_capable())
7893 err = sysfs_create_file(&cls->kset.kobj,
7894 &attr_sched_mc_power_savings.attr);
7895#endif
7896 return err;
7897}
6d6bc0ad 7898#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
5c45bf27 7899
e761b772 7900#ifndef CONFIG_CPUSETS
1da177e4 7901/*
e761b772
MK
7902 * Add online and remove offline CPUs from the scheduler domains.
7903 * When cpusets are enabled they take over this function.
1da177e4
LT
7904 */
7905static int update_sched_domains(struct notifier_block *nfb,
7906 unsigned long action, void *hcpu)
e761b772
MK
7907{
7908 switch (action) {
7909 case CPU_ONLINE:
7910 case CPU_ONLINE_FROZEN:
7911 case CPU_DEAD:
7912 case CPU_DEAD_FROZEN:
dfb512ec 7913 partition_sched_domains(1, NULL, NULL);
e761b772
MK
7914 return NOTIFY_OK;
7915
7916 default:
7917 return NOTIFY_DONE;
7918 }
7919}
7920#endif
7921
7922static int update_runtime(struct notifier_block *nfb,
7923 unsigned long action, void *hcpu)
1da177e4 7924{
7def2be1
PZ
7925 int cpu = (int)(long)hcpu;
7926
1da177e4 7927 switch (action) {
1da177e4 7928 case CPU_DOWN_PREPARE:
8bb78442 7929 case CPU_DOWN_PREPARE_FROZEN:
7def2be1 7930 disable_runtime(cpu_rq(cpu));
1da177e4
LT
7931 return NOTIFY_OK;
7932
1da177e4 7933 case CPU_DOWN_FAILED:
8bb78442 7934 case CPU_DOWN_FAILED_FROZEN:
1da177e4 7935 case CPU_ONLINE:
8bb78442 7936 case CPU_ONLINE_FROZEN:
7def2be1 7937 enable_runtime(cpu_rq(cpu));
e761b772
MK
7938 return NOTIFY_OK;
7939
1da177e4
LT
7940 default:
7941 return NOTIFY_DONE;
7942 }
1da177e4 7943}
1da177e4
LT
7944
7945void __init sched_init_smp(void)
7946{
5c1e1767
NP
7947 cpumask_t non_isolated_cpus;
7948
434d53b0
MT
7949#if defined(CONFIG_NUMA)
7950 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7951 GFP_KERNEL);
7952 BUG_ON(sched_group_nodes_bycpu == NULL);
7953#endif
95402b38 7954 get_online_cpus();
712555ee 7955 mutex_lock(&sched_domains_mutex);
1a20ff27 7956 arch_init_sched_domains(&cpu_online_map);
e5e5673f 7957 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
5c1e1767
NP
7958 if (cpus_empty(non_isolated_cpus))
7959 cpu_set(smp_processor_id(), non_isolated_cpus);
712555ee 7960 mutex_unlock(&sched_domains_mutex);
95402b38 7961 put_online_cpus();
e761b772
MK
7962
7963#ifndef CONFIG_CPUSETS
1da177e4
LT
7964 /* XXX: Theoretical race here - CPU may be hotplugged now */
7965 hotcpu_notifier(update_sched_domains, 0);
e761b772
MK
7966#endif
7967
7968 /* RT runtime code needs to handle some hotplug events */
7969 hotcpu_notifier(update_runtime, 0);
7970
b328ca18 7971 init_hrtick();
5c1e1767
NP
7972
7973 /* Move init over to a non-isolated CPU */
7c16ec58 7974 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
5c1e1767 7975 BUG();
19978ca6 7976 sched_init_granularity();
1da177e4
LT
7977}
7978#else
7979void __init sched_init_smp(void)
7980{
19978ca6 7981 sched_init_granularity();
1da177e4
LT
7982}
7983#endif /* CONFIG_SMP */
7984
7985int in_sched_functions(unsigned long addr)
7986{
1da177e4
LT
7987 return in_lock_functions(addr) ||
7988 (addr >= (unsigned long)__sched_text_start
7989 && addr < (unsigned long)__sched_text_end);
7990}
7991
a9957449 7992static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
dd41f596
IM
7993{
7994 cfs_rq->tasks_timeline = RB_ROOT;
4a55bd5e 7995 INIT_LIST_HEAD(&cfs_rq->tasks);
dd41f596
IM
7996#ifdef CONFIG_FAIR_GROUP_SCHED
7997 cfs_rq->rq = rq;
7998#endif
67e9fb2a 7999 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
dd41f596
IM
8000}
8001
fa85ae24
PZ
8002static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8003{
8004 struct rt_prio_array *array;
8005 int i;
8006
8007 array = &rt_rq->active;
8008 for (i = 0; i < MAX_RT_PRIO; i++) {
8009 INIT_LIST_HEAD(array->queue + i);
8010 __clear_bit(i, array->bitmap);
8011 }
8012 /* delimiter for bitsearch: */
8013 __set_bit(MAX_RT_PRIO, array->bitmap);
8014
052f1dc7 8015#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
48d5e258
PZ
8016 rt_rq->highest_prio = MAX_RT_PRIO;
8017#endif
fa85ae24
PZ
8018#ifdef CONFIG_SMP
8019 rt_rq->rt_nr_migratory = 0;
fa85ae24
PZ
8020 rt_rq->overloaded = 0;
8021#endif
8022
8023 rt_rq->rt_time = 0;
8024 rt_rq->rt_throttled = 0;
ac086bc2
PZ
8025 rt_rq->rt_runtime = 0;
8026 spin_lock_init(&rt_rq->rt_runtime_lock);
6f505b16 8027
052f1dc7 8028#ifdef CONFIG_RT_GROUP_SCHED
23b0fdfc 8029 rt_rq->rt_nr_boosted = 0;
6f505b16
PZ
8030 rt_rq->rq = rq;
8031#endif
fa85ae24
PZ
8032}
8033
6f505b16 8034#ifdef CONFIG_FAIR_GROUP_SCHED
ec7dc8ac
DG
8035static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8036 struct sched_entity *se, int cpu, int add,
8037 struct sched_entity *parent)
6f505b16 8038{
ec7dc8ac 8039 struct rq *rq = cpu_rq(cpu);
6f505b16
PZ
8040 tg->cfs_rq[cpu] = cfs_rq;
8041 init_cfs_rq(cfs_rq, rq);
8042 cfs_rq->tg = tg;
8043 if (add)
8044 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8045
8046 tg->se[cpu] = se;
354d60c2
DG
8047 /* se could be NULL for init_task_group */
8048 if (!se)
8049 return;
8050
ec7dc8ac
DG
8051 if (!parent)
8052 se->cfs_rq = &rq->cfs;
8053 else
8054 se->cfs_rq = parent->my_q;
8055
6f505b16
PZ
8056 se->my_q = cfs_rq;
8057 se->load.weight = tg->shares;
e05510d0 8058 se->load.inv_weight = 0;
ec7dc8ac 8059 se->parent = parent;
6f505b16 8060}
052f1dc7 8061#endif
6f505b16 8062
052f1dc7 8063#ifdef CONFIG_RT_GROUP_SCHED
ec7dc8ac
DG
8064static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8065 struct sched_rt_entity *rt_se, int cpu, int add,
8066 struct sched_rt_entity *parent)
6f505b16 8067{
ec7dc8ac
DG
8068 struct rq *rq = cpu_rq(cpu);
8069
6f505b16
PZ
8070 tg->rt_rq[cpu] = rt_rq;
8071 init_rt_rq(rt_rq, rq);
8072 rt_rq->tg = tg;
8073 rt_rq->rt_se = rt_se;
ac086bc2 8074 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
6f505b16
PZ
8075 if (add)
8076 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8077
8078 tg->rt_se[cpu] = rt_se;
354d60c2
DG
8079 if (!rt_se)
8080 return;
8081
ec7dc8ac
DG
8082 if (!parent)
8083 rt_se->rt_rq = &rq->rt;
8084 else
8085 rt_se->rt_rq = parent->my_q;
8086
6f505b16 8087 rt_se->my_q = rt_rq;
ec7dc8ac 8088 rt_se->parent = parent;
6f505b16
PZ
8089 INIT_LIST_HEAD(&rt_se->run_list);
8090}
8091#endif
8092
1da177e4
LT
8093void __init sched_init(void)
8094{
dd41f596 8095 int i, j;
434d53b0
MT
8096 unsigned long alloc_size = 0, ptr;
8097
8098#ifdef CONFIG_FAIR_GROUP_SCHED
8099 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8100#endif
8101#ifdef CONFIG_RT_GROUP_SCHED
8102 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
eff766a6
PZ
8103#endif
8104#ifdef CONFIG_USER_SCHED
8105 alloc_size *= 2;
434d53b0
MT
8106#endif
8107 /*
8108 * As sched_init() is called before page_alloc is setup,
8109 * we use alloc_bootmem().
8110 */
8111 if (alloc_size) {
5a9d3225 8112 ptr = (unsigned long)alloc_bootmem(alloc_size);
434d53b0
MT
8113
8114#ifdef CONFIG_FAIR_GROUP_SCHED
8115 init_task_group.se = (struct sched_entity **)ptr;
8116 ptr += nr_cpu_ids * sizeof(void **);
8117
8118 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8119 ptr += nr_cpu_ids * sizeof(void **);
eff766a6
PZ
8120
8121#ifdef CONFIG_USER_SCHED
8122 root_task_group.se = (struct sched_entity **)ptr;
8123 ptr += nr_cpu_ids * sizeof(void **);
8124
8125 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8126 ptr += nr_cpu_ids * sizeof(void **);
6d6bc0ad
DG
8127#endif /* CONFIG_USER_SCHED */
8128#endif /* CONFIG_FAIR_GROUP_SCHED */
434d53b0
MT
8129#ifdef CONFIG_RT_GROUP_SCHED
8130 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8131 ptr += nr_cpu_ids * sizeof(void **);
8132
8133 init_task_group.rt_rq = (struct rt_rq **)ptr;
eff766a6
PZ
8134 ptr += nr_cpu_ids * sizeof(void **);
8135
8136#ifdef CONFIG_USER_SCHED
8137 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8138 ptr += nr_cpu_ids * sizeof(void **);
8139
8140 root_task_group.rt_rq = (struct rt_rq **)ptr;
8141 ptr += nr_cpu_ids * sizeof(void **);
6d6bc0ad
DG
8142#endif /* CONFIG_USER_SCHED */
8143#endif /* CONFIG_RT_GROUP_SCHED */
434d53b0 8144 }
dd41f596 8145
57d885fe
GH
8146#ifdef CONFIG_SMP
8147 init_defrootdomain();
8148#endif
8149
d0b27fa7
PZ
8150 init_rt_bandwidth(&def_rt_bandwidth,
8151 global_rt_period(), global_rt_runtime());
8152
8153#ifdef CONFIG_RT_GROUP_SCHED
8154 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8155 global_rt_period(), global_rt_runtime());
eff766a6
PZ
8156#ifdef CONFIG_USER_SCHED
8157 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8158 global_rt_period(), RUNTIME_INF);
6d6bc0ad
DG
8159#endif /* CONFIG_USER_SCHED */
8160#endif /* CONFIG_RT_GROUP_SCHED */
d0b27fa7 8161
052f1dc7 8162#ifdef CONFIG_GROUP_SCHED
6f505b16 8163 list_add(&init_task_group.list, &task_groups);
f473aa5e
PZ
8164 INIT_LIST_HEAD(&init_task_group.children);
8165
8166#ifdef CONFIG_USER_SCHED
8167 INIT_LIST_HEAD(&root_task_group.children);
8168 init_task_group.parent = &root_task_group;
8169 list_add(&init_task_group.siblings, &root_task_group.children);
6d6bc0ad
DG
8170#endif /* CONFIG_USER_SCHED */
8171#endif /* CONFIG_GROUP_SCHED */
6f505b16 8172
0a945022 8173 for_each_possible_cpu(i) {
70b97a7f 8174 struct rq *rq;
1da177e4
LT
8175
8176 rq = cpu_rq(i);
8177 spin_lock_init(&rq->lock);
7897986b 8178 rq->nr_running = 0;
dd41f596 8179 init_cfs_rq(&rq->cfs, rq);
6f505b16 8180 init_rt_rq(&rq->rt, rq);
dd41f596 8181#ifdef CONFIG_FAIR_GROUP_SCHED
4cf86d77 8182 init_task_group.shares = init_task_group_load;
6f505b16 8183 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
354d60c2
DG
8184#ifdef CONFIG_CGROUP_SCHED
8185 /*
8186 * How much cpu bandwidth does init_task_group get?
8187 *
8188 * In case of task-groups formed thr' the cgroup filesystem, it
8189 * gets 100% of the cpu resources in the system. This overall
8190 * system cpu resource is divided among the tasks of
8191 * init_task_group and its child task-groups in a fair manner,
8192 * based on each entity's (task or task-group's) weight
8193 * (se->load.weight).
8194 *
8195 * In other words, if init_task_group has 10 tasks of weight
8196 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8197 * then A0's share of the cpu resource is:
8198 *
8199 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8200 *
8201 * We achieve this by letting init_task_group's tasks sit
8202 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8203 */
ec7dc8ac 8204 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
354d60c2 8205#elif defined CONFIG_USER_SCHED
eff766a6
PZ
8206 root_task_group.shares = NICE_0_LOAD;
8207 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
354d60c2
DG
8208 /*
8209 * In case of task-groups formed thr' the user id of tasks,
8210 * init_task_group represents tasks belonging to root user.
8211 * Hence it forms a sibling of all subsequent groups formed.
8212 * In this case, init_task_group gets only a fraction of overall
8213 * system cpu resource, based on the weight assigned to root
8214 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8215 * by letting tasks of init_task_group sit in a separate cfs_rq
8216 * (init_cfs_rq) and having one entity represent this group of
8217 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8218 */
ec7dc8ac 8219 init_tg_cfs_entry(&init_task_group,
6f505b16 8220 &per_cpu(init_cfs_rq, i),
eff766a6
PZ
8221 &per_cpu(init_sched_entity, i), i, 1,
8222 root_task_group.se[i]);
6f505b16 8223
052f1dc7 8224#endif
354d60c2
DG
8225#endif /* CONFIG_FAIR_GROUP_SCHED */
8226
8227 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
052f1dc7 8228#ifdef CONFIG_RT_GROUP_SCHED
6f505b16 8229 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
354d60c2 8230#ifdef CONFIG_CGROUP_SCHED
ec7dc8ac 8231 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
354d60c2 8232#elif defined CONFIG_USER_SCHED
eff766a6 8233 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
ec7dc8ac 8234 init_tg_rt_entry(&init_task_group,
6f505b16 8235 &per_cpu(init_rt_rq, i),
eff766a6
PZ
8236 &per_cpu(init_sched_rt_entity, i), i, 1,
8237 root_task_group.rt_se[i]);
354d60c2 8238#endif
dd41f596 8239#endif
1da177e4 8240
dd41f596
IM
8241 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8242 rq->cpu_load[j] = 0;
1da177e4 8243#ifdef CONFIG_SMP
41c7ce9a 8244 rq->sd = NULL;
57d885fe 8245 rq->rd = NULL;
1da177e4 8246 rq->active_balance = 0;
dd41f596 8247 rq->next_balance = jiffies;
1da177e4 8248 rq->push_cpu = 0;
0a2966b4 8249 rq->cpu = i;
1f11eb6a 8250 rq->online = 0;
1da177e4
LT
8251 rq->migration_thread = NULL;
8252 INIT_LIST_HEAD(&rq->migration_queue);
dc938520 8253 rq_attach_root(rq, &def_root_domain);
1da177e4 8254#endif
8f4d37ec 8255 init_rq_hrtick(rq);
1da177e4 8256 atomic_set(&rq->nr_iowait, 0);
1da177e4
LT
8257 }
8258
2dd73a4f 8259 set_load_weight(&init_task);
b50f60ce 8260
e107be36
AK
8261#ifdef CONFIG_PREEMPT_NOTIFIERS
8262 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8263#endif
8264
c9819f45 8265#ifdef CONFIG_SMP
962cf36c 8266 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
c9819f45
CL
8267#endif
8268
b50f60ce
HC
8269#ifdef CONFIG_RT_MUTEXES
8270 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8271#endif
8272
1da177e4
LT
8273 /*
8274 * The boot idle thread does lazy MMU switching as well:
8275 */
8276 atomic_inc(&init_mm.mm_count);
8277 enter_lazy_tlb(&init_mm, current);
8278
8279 /*
8280 * Make us the idle thread. Technically, schedule() should not be
8281 * called from this thread, however somewhere below it might be,
8282 * but because we are the idle thread, we just pick up running again
8283 * when this runqueue becomes "idle".
8284 */
8285 init_idle(current, smp_processor_id());
dd41f596
IM
8286 /*
8287 * During early bootup we pretend to be a normal task:
8288 */
8289 current->sched_class = &fair_sched_class;
6892b75e
IM
8290
8291 scheduler_running = 1;
1da177e4
LT
8292}
8293
8294#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8295void __might_sleep(char *file, int line)
8296{
48f24c4d 8297#ifdef in_atomic
1da177e4
LT
8298 static unsigned long prev_jiffy; /* ratelimiting */
8299
aef745fc
IM
8300 if ((!in_atomic() && !irqs_disabled()) ||
8301 system_state != SYSTEM_RUNNING || oops_in_progress)
8302 return;
8303 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8304 return;
8305 prev_jiffy = jiffies;
8306
8307 printk(KERN_ERR
8308 "BUG: sleeping function called from invalid context at %s:%d\n",
8309 file, line);
8310 printk(KERN_ERR
8311 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8312 in_atomic(), irqs_disabled(),
8313 current->pid, current->comm);
8314
8315 debug_show_held_locks(current);
8316 if (irqs_disabled())
8317 print_irqtrace_events(current);
8318 dump_stack();
1da177e4
LT
8319#endif
8320}
8321EXPORT_SYMBOL(__might_sleep);
8322#endif
8323
8324#ifdef CONFIG_MAGIC_SYSRQ
3a5e4dc1
AK
8325static void normalize_task(struct rq *rq, struct task_struct *p)
8326{
8327 int on_rq;
3e51f33f 8328
3a5e4dc1
AK
8329 update_rq_clock(rq);
8330 on_rq = p->se.on_rq;
8331 if (on_rq)
8332 deactivate_task(rq, p, 0);
8333 __setscheduler(rq, p, SCHED_NORMAL, 0);
8334 if (on_rq) {
8335 activate_task(rq, p, 0);
8336 resched_task(rq->curr);
8337 }
8338}
8339
1da177e4
LT
8340void normalize_rt_tasks(void)
8341{
a0f98a1c 8342 struct task_struct *g, *p;
1da177e4 8343 unsigned long flags;
70b97a7f 8344 struct rq *rq;
1da177e4 8345
4cf5d77a 8346 read_lock_irqsave(&tasklist_lock, flags);
a0f98a1c 8347 do_each_thread(g, p) {
178be793
IM
8348 /*
8349 * Only normalize user tasks:
8350 */
8351 if (!p->mm)
8352 continue;
8353
6cfb0d5d 8354 p->se.exec_start = 0;
6cfb0d5d 8355#ifdef CONFIG_SCHEDSTATS
dd41f596 8356 p->se.wait_start = 0;
dd41f596 8357 p->se.sleep_start = 0;
dd41f596 8358 p->se.block_start = 0;
6cfb0d5d 8359#endif
dd41f596
IM
8360
8361 if (!rt_task(p)) {
8362 /*
8363 * Renice negative nice level userspace
8364 * tasks back to 0:
8365 */
8366 if (TASK_NICE(p) < 0 && p->mm)
8367 set_user_nice(p, 0);
1da177e4 8368 continue;
dd41f596 8369 }
1da177e4 8370
4cf5d77a 8371 spin_lock(&p->pi_lock);
b29739f9 8372 rq = __task_rq_lock(p);
1da177e4 8373
178be793 8374 normalize_task(rq, p);
3a5e4dc1 8375
b29739f9 8376 __task_rq_unlock(rq);
4cf5d77a 8377 spin_unlock(&p->pi_lock);
a0f98a1c
IM
8378 } while_each_thread(g, p);
8379
4cf5d77a 8380 read_unlock_irqrestore(&tasklist_lock, flags);
1da177e4
LT
8381}
8382
8383#endif /* CONFIG_MAGIC_SYSRQ */
1df5c10a
LT
8384
8385#ifdef CONFIG_IA64
8386/*
8387 * These functions are only useful for the IA64 MCA handling.
8388 *
8389 * They can only be called when the whole system has been
8390 * stopped - every CPU needs to be quiescent, and no scheduling
8391 * activity can take place. Using them for anything else would
8392 * be a serious bug, and as a result, they aren't even visible
8393 * under any other configuration.
8394 */
8395
8396/**
8397 * curr_task - return the current task for a given cpu.
8398 * @cpu: the processor in question.
8399 *
8400 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8401 */
36c8b586 8402struct task_struct *curr_task(int cpu)
1df5c10a
LT
8403{
8404 return cpu_curr(cpu);
8405}
8406
8407/**
8408 * set_curr_task - set the current task for a given cpu.
8409 * @cpu: the processor in question.
8410 * @p: the task pointer to set.
8411 *
8412 * Description: This function must only be used when non-maskable interrupts
41a2d6cf
IM
8413 * are serviced on a separate stack. It allows the architecture to switch the
8414 * notion of the current task on a cpu in a non-blocking manner. This function
1df5c10a
LT
8415 * must be called with all CPU's synchronized, and interrupts disabled, the
8416 * and caller must save the original value of the current task (see
8417 * curr_task() above) and restore that value before reenabling interrupts and
8418 * re-starting the system.
8419 *
8420 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8421 */
36c8b586 8422void set_curr_task(int cpu, struct task_struct *p)
1df5c10a
LT
8423{
8424 cpu_curr(cpu) = p;
8425}
8426
8427#endif
29f59db3 8428
bccbe08a
PZ
8429#ifdef CONFIG_FAIR_GROUP_SCHED
8430static void free_fair_sched_group(struct task_group *tg)
6f505b16
PZ
8431{
8432 int i;
8433
8434 for_each_possible_cpu(i) {
8435 if (tg->cfs_rq)
8436 kfree(tg->cfs_rq[i]);
8437 if (tg->se)
8438 kfree(tg->se[i]);
6f505b16
PZ
8439 }
8440
8441 kfree(tg->cfs_rq);
8442 kfree(tg->se);
6f505b16
PZ
8443}
8444
ec7dc8ac
DG
8445static
8446int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
29f59db3 8447{
29f59db3 8448 struct cfs_rq *cfs_rq;
eab17229 8449 struct sched_entity *se;
9b5b7751 8450 struct rq *rq;
29f59db3
SV
8451 int i;
8452
434d53b0 8453 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
29f59db3
SV
8454 if (!tg->cfs_rq)
8455 goto err;
434d53b0 8456 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
29f59db3
SV
8457 if (!tg->se)
8458 goto err;
052f1dc7
PZ
8459
8460 tg->shares = NICE_0_LOAD;
29f59db3
SV
8461
8462 for_each_possible_cpu(i) {
9b5b7751 8463 rq = cpu_rq(i);
29f59db3 8464
eab17229
LZ
8465 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8466 GFP_KERNEL, cpu_to_node(i));
29f59db3
SV
8467 if (!cfs_rq)
8468 goto err;
8469
eab17229
LZ
8470 se = kzalloc_node(sizeof(struct sched_entity),
8471 GFP_KERNEL, cpu_to_node(i));
29f59db3
SV
8472 if (!se)
8473 goto err;
8474
eab17229 8475 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
bccbe08a
PZ
8476 }
8477
8478 return 1;
8479
8480 err:
8481 return 0;
8482}
8483
8484static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8485{
8486 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8487 &cpu_rq(cpu)->leaf_cfs_rq_list);
8488}
8489
8490static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8491{
8492 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8493}
6d6bc0ad 8494#else /* !CONFG_FAIR_GROUP_SCHED */
bccbe08a
PZ
8495static inline void free_fair_sched_group(struct task_group *tg)
8496{
8497}
8498
ec7dc8ac
DG
8499static inline
8500int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
bccbe08a
PZ
8501{
8502 return 1;
8503}
8504
8505static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8506{
8507}
8508
8509static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8510{
8511}
6d6bc0ad 8512#endif /* CONFIG_FAIR_GROUP_SCHED */
052f1dc7
PZ
8513
8514#ifdef CONFIG_RT_GROUP_SCHED
bccbe08a
PZ
8515static void free_rt_sched_group(struct task_group *tg)
8516{
8517 int i;
8518
d0b27fa7
PZ
8519 destroy_rt_bandwidth(&tg->rt_bandwidth);
8520
bccbe08a
PZ
8521 for_each_possible_cpu(i) {
8522 if (tg->rt_rq)
8523 kfree(tg->rt_rq[i]);
8524 if (tg->rt_se)
8525 kfree(tg->rt_se[i]);
8526 }
8527
8528 kfree(tg->rt_rq);
8529 kfree(tg->rt_se);
8530}
8531
ec7dc8ac
DG
8532static
8533int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
bccbe08a
PZ
8534{
8535 struct rt_rq *rt_rq;
eab17229 8536 struct sched_rt_entity *rt_se;
bccbe08a
PZ
8537 struct rq *rq;
8538 int i;
8539
434d53b0 8540 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
bccbe08a
PZ
8541 if (!tg->rt_rq)
8542 goto err;
434d53b0 8543 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
bccbe08a
PZ
8544 if (!tg->rt_se)
8545 goto err;
8546
d0b27fa7
PZ
8547 init_rt_bandwidth(&tg->rt_bandwidth,
8548 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
bccbe08a
PZ
8549
8550 for_each_possible_cpu(i) {
8551 rq = cpu_rq(i);
8552
eab17229
LZ
8553 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8554 GFP_KERNEL, cpu_to_node(i));
6f505b16
PZ
8555 if (!rt_rq)
8556 goto err;
29f59db3 8557
eab17229
LZ
8558 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8559 GFP_KERNEL, cpu_to_node(i));
6f505b16
PZ
8560 if (!rt_se)
8561 goto err;
29f59db3 8562
eab17229 8563 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
29f59db3
SV
8564 }
8565
bccbe08a
PZ
8566 return 1;
8567
8568 err:
8569 return 0;
8570}
8571
8572static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8573{
8574 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8575 &cpu_rq(cpu)->leaf_rt_rq_list);
8576}
8577
8578static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8579{
8580 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8581}
6d6bc0ad 8582#else /* !CONFIG_RT_GROUP_SCHED */
bccbe08a
PZ
8583static inline void free_rt_sched_group(struct task_group *tg)
8584{
8585}
8586
ec7dc8ac
DG
8587static inline
8588int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
bccbe08a
PZ
8589{
8590 return 1;
8591}
8592
8593static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8594{
8595}
8596
8597static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8598{
8599}
6d6bc0ad 8600#endif /* CONFIG_RT_GROUP_SCHED */
bccbe08a 8601
d0b27fa7 8602#ifdef CONFIG_GROUP_SCHED
bccbe08a
PZ
8603static void free_sched_group(struct task_group *tg)
8604{
8605 free_fair_sched_group(tg);
8606 free_rt_sched_group(tg);
8607 kfree(tg);
8608}
8609
8610/* allocate runqueue etc for a new task group */
ec7dc8ac 8611struct task_group *sched_create_group(struct task_group *parent)
bccbe08a
PZ
8612{
8613 struct task_group *tg;
8614 unsigned long flags;
8615 int i;
8616
8617 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8618 if (!tg)
8619 return ERR_PTR(-ENOMEM);
8620
ec7dc8ac 8621 if (!alloc_fair_sched_group(tg, parent))
bccbe08a
PZ
8622 goto err;
8623
ec7dc8ac 8624 if (!alloc_rt_sched_group(tg, parent))
bccbe08a
PZ
8625 goto err;
8626
8ed36996 8627 spin_lock_irqsave(&task_group_lock, flags);
9b5b7751 8628 for_each_possible_cpu(i) {
bccbe08a
PZ
8629 register_fair_sched_group(tg, i);
8630 register_rt_sched_group(tg, i);
9b5b7751 8631 }
6f505b16 8632 list_add_rcu(&tg->list, &task_groups);
f473aa5e
PZ
8633
8634 WARN_ON(!parent); /* root should already exist */
8635
8636 tg->parent = parent;
f473aa5e 8637 INIT_LIST_HEAD(&tg->children);
09f2724a 8638 list_add_rcu(&tg->siblings, &parent->children);
8ed36996 8639 spin_unlock_irqrestore(&task_group_lock, flags);
29f59db3 8640
9b5b7751 8641 return tg;
29f59db3
SV
8642
8643err:
6f505b16 8644 free_sched_group(tg);
29f59db3
SV
8645 return ERR_PTR(-ENOMEM);
8646}
8647
9b5b7751 8648/* rcu callback to free various structures associated with a task group */
6f505b16 8649static void free_sched_group_rcu(struct rcu_head *rhp)
29f59db3 8650{
29f59db3 8651 /* now it should be safe to free those cfs_rqs */
6f505b16 8652 free_sched_group(container_of(rhp, struct task_group, rcu));
29f59db3
SV
8653}
8654
9b5b7751 8655/* Destroy runqueue etc associated with a task group */
4cf86d77 8656void sched_destroy_group(struct task_group *tg)
29f59db3 8657{
8ed36996 8658 unsigned long flags;
9b5b7751 8659 int i;
29f59db3 8660
8ed36996 8661 spin_lock_irqsave(&task_group_lock, flags);
9b5b7751 8662 for_each_possible_cpu(i) {
bccbe08a
PZ
8663 unregister_fair_sched_group(tg, i);
8664 unregister_rt_sched_group(tg, i);
9b5b7751 8665 }
6f505b16 8666 list_del_rcu(&tg->list);
f473aa5e 8667 list_del_rcu(&tg->siblings);
8ed36996 8668 spin_unlock_irqrestore(&task_group_lock, flags);
9b5b7751 8669
9b5b7751 8670 /* wait for possible concurrent references to cfs_rqs complete */
6f505b16 8671 call_rcu(&tg->rcu, free_sched_group_rcu);
29f59db3
SV
8672}
8673
9b5b7751 8674/* change task's runqueue when it moves between groups.
3a252015
IM
8675 * The caller of this function should have put the task in its new group
8676 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8677 * reflect its new group.
9b5b7751
SV
8678 */
8679void sched_move_task(struct task_struct *tsk)
29f59db3
SV
8680{
8681 int on_rq, running;
8682 unsigned long flags;
8683 struct rq *rq;
8684
8685 rq = task_rq_lock(tsk, &flags);
8686
29f59db3
SV
8687 update_rq_clock(rq);
8688
051a1d1a 8689 running = task_current(rq, tsk);
29f59db3
SV
8690 on_rq = tsk->se.on_rq;
8691
0e1f3483 8692 if (on_rq)
29f59db3 8693 dequeue_task(rq, tsk, 0);
0e1f3483
HS
8694 if (unlikely(running))
8695 tsk->sched_class->put_prev_task(rq, tsk);
29f59db3 8696
6f505b16 8697 set_task_rq(tsk, task_cpu(tsk));
29f59db3 8698
810b3817
PZ
8699#ifdef CONFIG_FAIR_GROUP_SCHED
8700 if (tsk->sched_class->moved_group)
8701 tsk->sched_class->moved_group(tsk);
8702#endif
8703
0e1f3483
HS
8704 if (unlikely(running))
8705 tsk->sched_class->set_curr_task(rq);
8706 if (on_rq)
7074badb 8707 enqueue_task(rq, tsk, 0);
29f59db3 8708
29f59db3
SV
8709 task_rq_unlock(rq, &flags);
8710}
6d6bc0ad 8711#endif /* CONFIG_GROUP_SCHED */
29f59db3 8712
052f1dc7 8713#ifdef CONFIG_FAIR_GROUP_SCHED
c09595f6 8714static void __set_se_shares(struct sched_entity *se, unsigned long shares)
29f59db3
SV
8715{
8716 struct cfs_rq *cfs_rq = se->cfs_rq;
29f59db3
SV
8717 int on_rq;
8718
29f59db3 8719 on_rq = se->on_rq;
62fb1851 8720 if (on_rq)
29f59db3
SV
8721 dequeue_entity(cfs_rq, se, 0);
8722
8723 se->load.weight = shares;
e05510d0 8724 se->load.inv_weight = 0;
29f59db3 8725
62fb1851 8726 if (on_rq)
29f59db3 8727 enqueue_entity(cfs_rq, se, 0);
c09595f6 8728}
62fb1851 8729
c09595f6
PZ
8730static void set_se_shares(struct sched_entity *se, unsigned long shares)
8731{
8732 struct cfs_rq *cfs_rq = se->cfs_rq;
8733 struct rq *rq = cfs_rq->rq;
8734 unsigned long flags;
8735
8736 spin_lock_irqsave(&rq->lock, flags);
8737 __set_se_shares(se, shares);
8738 spin_unlock_irqrestore(&rq->lock, flags);
29f59db3
SV
8739}
8740
8ed36996
PZ
8741static DEFINE_MUTEX(shares_mutex);
8742
4cf86d77 8743int sched_group_set_shares(struct task_group *tg, unsigned long shares)
29f59db3
SV
8744{
8745 int i;
8ed36996 8746 unsigned long flags;
c61935fd 8747
ec7dc8ac
DG
8748 /*
8749 * We can't change the weight of the root cgroup.
8750 */
8751 if (!tg->se[0])
8752 return -EINVAL;
8753
18d95a28
PZ
8754 if (shares < MIN_SHARES)
8755 shares = MIN_SHARES;
cb4ad1ff
MX
8756 else if (shares > MAX_SHARES)
8757 shares = MAX_SHARES;
62fb1851 8758
8ed36996 8759 mutex_lock(&shares_mutex);
9b5b7751 8760 if (tg->shares == shares)
5cb350ba 8761 goto done;
29f59db3 8762
8ed36996 8763 spin_lock_irqsave(&task_group_lock, flags);
bccbe08a
PZ
8764 for_each_possible_cpu(i)
8765 unregister_fair_sched_group(tg, i);
f473aa5e 8766 list_del_rcu(&tg->siblings);
8ed36996 8767 spin_unlock_irqrestore(&task_group_lock, flags);
6b2d7700
SV
8768
8769 /* wait for any ongoing reference to this group to finish */
8770 synchronize_sched();
8771
8772 /*
8773 * Now we are free to modify the group's share on each cpu
8774 * w/o tripping rebalance_share or load_balance_fair.
8775 */
9b5b7751 8776 tg->shares = shares;
c09595f6
PZ
8777 for_each_possible_cpu(i) {
8778 /*
8779 * force a rebalance
8780 */
8781 cfs_rq_set_shares(tg->cfs_rq[i], 0);
cb4ad1ff 8782 set_se_shares(tg->se[i], shares);
c09595f6 8783 }
29f59db3 8784
6b2d7700
SV
8785 /*
8786 * Enable load balance activity on this group, by inserting it back on
8787 * each cpu's rq->leaf_cfs_rq_list.
8788 */
8ed36996 8789 spin_lock_irqsave(&task_group_lock, flags);
bccbe08a
PZ
8790 for_each_possible_cpu(i)
8791 register_fair_sched_group(tg, i);
f473aa5e 8792 list_add_rcu(&tg->siblings, &tg->parent->children);
8ed36996 8793 spin_unlock_irqrestore(&task_group_lock, flags);
5cb350ba 8794done:
8ed36996 8795 mutex_unlock(&shares_mutex);
9b5b7751 8796 return 0;
29f59db3
SV
8797}
8798
5cb350ba
DG
8799unsigned long sched_group_shares(struct task_group *tg)
8800{
8801 return tg->shares;
8802}
052f1dc7 8803#endif
5cb350ba 8804
052f1dc7 8805#ifdef CONFIG_RT_GROUP_SCHED
6f505b16 8806/*
9f0c1e56 8807 * Ensure that the real time constraints are schedulable.
6f505b16 8808 */
9f0c1e56
PZ
8809static DEFINE_MUTEX(rt_constraints_mutex);
8810
8811static unsigned long to_ratio(u64 period, u64 runtime)
8812{
8813 if (runtime == RUNTIME_INF)
9a7e0b18 8814 return 1ULL << 20;
9f0c1e56 8815
9a7e0b18 8816 return div64_u64(runtime << 20, period);
9f0c1e56
PZ
8817}
8818
9a7e0b18
PZ
8819/* Must be called with tasklist_lock held */
8820static inline int tg_has_rt_tasks(struct task_group *tg)
b40b2e8e 8821{
9a7e0b18 8822 struct task_struct *g, *p;
b40b2e8e 8823
9a7e0b18
PZ
8824 do_each_thread(g, p) {
8825 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8826 return 1;
8827 } while_each_thread(g, p);
b40b2e8e 8828
9a7e0b18
PZ
8829 return 0;
8830}
b40b2e8e 8831
9a7e0b18
PZ
8832struct rt_schedulable_data {
8833 struct task_group *tg;
8834 u64 rt_period;
8835 u64 rt_runtime;
8836};
b40b2e8e 8837
9a7e0b18
PZ
8838static int tg_schedulable(struct task_group *tg, void *data)
8839{
8840 struct rt_schedulable_data *d = data;
8841 struct task_group *child;
8842 unsigned long total, sum = 0;
8843 u64 period, runtime;
b40b2e8e 8844
9a7e0b18
PZ
8845 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8846 runtime = tg->rt_bandwidth.rt_runtime;
b40b2e8e 8847
9a7e0b18
PZ
8848 if (tg == d->tg) {
8849 period = d->rt_period;
8850 runtime = d->rt_runtime;
b40b2e8e 8851 }
b40b2e8e 8852
4653f803
PZ
8853 /*
8854 * Cannot have more runtime than the period.
8855 */
8856 if (runtime > period && runtime != RUNTIME_INF)
8857 return -EINVAL;
6f505b16 8858
4653f803
PZ
8859 /*
8860 * Ensure we don't starve existing RT tasks.
8861 */
9a7e0b18
PZ
8862 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8863 return -EBUSY;
6f505b16 8864
9a7e0b18 8865 total = to_ratio(period, runtime);
6f505b16 8866
4653f803
PZ
8867 /*
8868 * Nobody can have more than the global setting allows.
8869 */
8870 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8871 return -EINVAL;
6f505b16 8872
4653f803
PZ
8873 /*
8874 * The sum of our children's runtime should not exceed our own.
8875 */
9a7e0b18
PZ
8876 list_for_each_entry_rcu(child, &tg->children, siblings) {
8877 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8878 runtime = child->rt_bandwidth.rt_runtime;
6f505b16 8879
9a7e0b18
PZ
8880 if (child == d->tg) {
8881 period = d->rt_period;
8882 runtime = d->rt_runtime;
8883 }
6f505b16 8884
9a7e0b18 8885 sum += to_ratio(period, runtime);
9f0c1e56 8886 }
6f505b16 8887
9a7e0b18
PZ
8888 if (sum > total)
8889 return -EINVAL;
8890
8891 return 0;
6f505b16
PZ
8892}
8893
9a7e0b18 8894static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
521f1a24 8895{
9a7e0b18
PZ
8896 struct rt_schedulable_data data = {
8897 .tg = tg,
8898 .rt_period = period,
8899 .rt_runtime = runtime,
8900 };
8901
8902 return walk_tg_tree(tg_schedulable, tg_nop, &data);
521f1a24
DG
8903}
8904
d0b27fa7
PZ
8905static int tg_set_bandwidth(struct task_group *tg,
8906 u64 rt_period, u64 rt_runtime)
6f505b16 8907{
ac086bc2 8908 int i, err = 0;
9f0c1e56 8909
9f0c1e56 8910 mutex_lock(&rt_constraints_mutex);
521f1a24 8911 read_lock(&tasklist_lock);
9a7e0b18
PZ
8912 err = __rt_schedulable(tg, rt_period, rt_runtime);
8913 if (err)
9f0c1e56 8914 goto unlock;
ac086bc2
PZ
8915
8916 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
d0b27fa7
PZ
8917 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8918 tg->rt_bandwidth.rt_runtime = rt_runtime;
ac086bc2
PZ
8919
8920 for_each_possible_cpu(i) {
8921 struct rt_rq *rt_rq = tg->rt_rq[i];
8922
8923 spin_lock(&rt_rq->rt_runtime_lock);
8924 rt_rq->rt_runtime = rt_runtime;
8925 spin_unlock(&rt_rq->rt_runtime_lock);
8926 }
8927 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9f0c1e56 8928 unlock:
521f1a24 8929 read_unlock(&tasklist_lock);
9f0c1e56
PZ
8930 mutex_unlock(&rt_constraints_mutex);
8931
8932 return err;
6f505b16
PZ
8933}
8934
d0b27fa7
PZ
8935int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8936{
8937 u64 rt_runtime, rt_period;
8938
8939 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8940 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8941 if (rt_runtime_us < 0)
8942 rt_runtime = RUNTIME_INF;
8943
8944 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8945}
8946
9f0c1e56
PZ
8947long sched_group_rt_runtime(struct task_group *tg)
8948{
8949 u64 rt_runtime_us;
8950
d0b27fa7 8951 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9f0c1e56
PZ
8952 return -1;
8953
d0b27fa7 8954 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9f0c1e56
PZ
8955 do_div(rt_runtime_us, NSEC_PER_USEC);
8956 return rt_runtime_us;
8957}
d0b27fa7
PZ
8958
8959int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8960{
8961 u64 rt_runtime, rt_period;
8962
8963 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8964 rt_runtime = tg->rt_bandwidth.rt_runtime;
8965
619b0488
R
8966 if (rt_period == 0)
8967 return -EINVAL;
8968
d0b27fa7
PZ
8969 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8970}
8971
8972long sched_group_rt_period(struct task_group *tg)
8973{
8974 u64 rt_period_us;
8975
8976 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8977 do_div(rt_period_us, NSEC_PER_USEC);
8978 return rt_period_us;
8979}
8980
8981static int sched_rt_global_constraints(void)
8982{
4653f803 8983 u64 runtime, period;
d0b27fa7
PZ
8984 int ret = 0;
8985
ec5d4989
HS
8986 if (sysctl_sched_rt_period <= 0)
8987 return -EINVAL;
8988
4653f803
PZ
8989 runtime = global_rt_runtime();
8990 period = global_rt_period();
8991
8992 /*
8993 * Sanity check on the sysctl variables.
8994 */
8995 if (runtime > period && runtime != RUNTIME_INF)
8996 return -EINVAL;
10b612f4 8997
d0b27fa7 8998 mutex_lock(&rt_constraints_mutex);
9a7e0b18 8999 read_lock(&tasklist_lock);
4653f803 9000 ret = __rt_schedulable(NULL, 0, 0);
9a7e0b18 9001 read_unlock(&tasklist_lock);
d0b27fa7
PZ
9002 mutex_unlock(&rt_constraints_mutex);
9003
9004 return ret;
9005}
6d6bc0ad 9006#else /* !CONFIG_RT_GROUP_SCHED */
d0b27fa7
PZ
9007static int sched_rt_global_constraints(void)
9008{
ac086bc2
PZ
9009 unsigned long flags;
9010 int i;
9011
ec5d4989
HS
9012 if (sysctl_sched_rt_period <= 0)
9013 return -EINVAL;
9014
ac086bc2
PZ
9015 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9016 for_each_possible_cpu(i) {
9017 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9018
9019 spin_lock(&rt_rq->rt_runtime_lock);
9020 rt_rq->rt_runtime = global_rt_runtime();
9021 spin_unlock(&rt_rq->rt_runtime_lock);
9022 }
9023 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9024
d0b27fa7
PZ
9025 return 0;
9026}
6d6bc0ad 9027#endif /* CONFIG_RT_GROUP_SCHED */
d0b27fa7
PZ
9028
9029int sched_rt_handler(struct ctl_table *table, int write,
9030 struct file *filp, void __user *buffer, size_t *lenp,
9031 loff_t *ppos)
9032{
9033 int ret;
9034 int old_period, old_runtime;
9035 static DEFINE_MUTEX(mutex);
9036
9037 mutex_lock(&mutex);
9038 old_period = sysctl_sched_rt_period;
9039 old_runtime = sysctl_sched_rt_runtime;
9040
9041 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9042
9043 if (!ret && write) {
9044 ret = sched_rt_global_constraints();
9045 if (ret) {
9046 sysctl_sched_rt_period = old_period;
9047 sysctl_sched_rt_runtime = old_runtime;
9048 } else {
9049 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9050 def_rt_bandwidth.rt_period =
9051 ns_to_ktime(global_rt_period());
9052 }
9053 }
9054 mutex_unlock(&mutex);
9055
9056 return ret;
9057}
68318b8e 9058
052f1dc7 9059#ifdef CONFIG_CGROUP_SCHED
68318b8e
SV
9060
9061/* return corresponding task_group object of a cgroup */
2b01dfe3 9062static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
68318b8e 9063{
2b01dfe3
PM
9064 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9065 struct task_group, css);
68318b8e
SV
9066}
9067
9068static struct cgroup_subsys_state *
2b01dfe3 9069cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
68318b8e 9070{
ec7dc8ac 9071 struct task_group *tg, *parent;
68318b8e 9072
2b01dfe3 9073 if (!cgrp->parent) {
68318b8e 9074 /* This is early initialization for the top cgroup */
68318b8e
SV
9075 return &init_task_group.css;
9076 }
9077
ec7dc8ac
DG
9078 parent = cgroup_tg(cgrp->parent);
9079 tg = sched_create_group(parent);
68318b8e
SV
9080 if (IS_ERR(tg))
9081 return ERR_PTR(-ENOMEM);
9082
68318b8e
SV
9083 return &tg->css;
9084}
9085
41a2d6cf
IM
9086static void
9087cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
68318b8e 9088{
2b01dfe3 9089 struct task_group *tg = cgroup_tg(cgrp);
68318b8e
SV
9090
9091 sched_destroy_group(tg);
9092}
9093
41a2d6cf
IM
9094static int
9095cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9096 struct task_struct *tsk)
68318b8e 9097{
b68aa230
PZ
9098#ifdef CONFIG_RT_GROUP_SCHED
9099 /* Don't accept realtime tasks when there is no way for them to run */
d0b27fa7 9100 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
b68aa230
PZ
9101 return -EINVAL;
9102#else
68318b8e
SV
9103 /* We don't support RT-tasks being in separate groups */
9104 if (tsk->sched_class != &fair_sched_class)
9105 return -EINVAL;
b68aa230 9106#endif
68318b8e
SV
9107
9108 return 0;
9109}
9110
9111static void
2b01dfe3 9112cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
68318b8e
SV
9113 struct cgroup *old_cont, struct task_struct *tsk)
9114{
9115 sched_move_task(tsk);
9116}
9117
052f1dc7 9118#ifdef CONFIG_FAIR_GROUP_SCHED
f4c753b7 9119static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
2b01dfe3 9120 u64 shareval)
68318b8e 9121{
2b01dfe3 9122 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
68318b8e
SV
9123}
9124
f4c753b7 9125static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
68318b8e 9126{
2b01dfe3 9127 struct task_group *tg = cgroup_tg(cgrp);
68318b8e
SV
9128
9129 return (u64) tg->shares;
9130}
6d6bc0ad 9131#endif /* CONFIG_FAIR_GROUP_SCHED */
68318b8e 9132
052f1dc7 9133#ifdef CONFIG_RT_GROUP_SCHED
0c70814c 9134static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
06ecb27c 9135 s64 val)
6f505b16 9136{
06ecb27c 9137 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
6f505b16
PZ
9138}
9139
06ecb27c 9140static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
6f505b16 9141{
06ecb27c 9142 return sched_group_rt_runtime(cgroup_tg(cgrp));
6f505b16 9143}
d0b27fa7
PZ
9144
9145static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9146 u64 rt_period_us)
9147{
9148 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9149}
9150
9151static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9152{
9153 return sched_group_rt_period(cgroup_tg(cgrp));
9154}
6d6bc0ad 9155#endif /* CONFIG_RT_GROUP_SCHED */
6f505b16 9156
fe5c7cc2 9157static struct cftype cpu_files[] = {
052f1dc7 9158#ifdef CONFIG_FAIR_GROUP_SCHED
fe5c7cc2
PM
9159 {
9160 .name = "shares",
f4c753b7
PM
9161 .read_u64 = cpu_shares_read_u64,
9162 .write_u64 = cpu_shares_write_u64,
fe5c7cc2 9163 },
052f1dc7
PZ
9164#endif
9165#ifdef CONFIG_RT_GROUP_SCHED
6f505b16 9166 {
9f0c1e56 9167 .name = "rt_runtime_us",
06ecb27c
PM
9168 .read_s64 = cpu_rt_runtime_read,
9169 .write_s64 = cpu_rt_runtime_write,
6f505b16 9170 },
d0b27fa7
PZ
9171 {
9172 .name = "rt_period_us",
f4c753b7
PM
9173 .read_u64 = cpu_rt_period_read_uint,
9174 .write_u64 = cpu_rt_period_write_uint,
d0b27fa7 9175 },
052f1dc7 9176#endif
68318b8e
SV
9177};
9178
9179static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9180{
fe5c7cc2 9181 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
68318b8e
SV
9182}
9183
9184struct cgroup_subsys cpu_cgroup_subsys = {
38605cae
IM
9185 .name = "cpu",
9186 .create = cpu_cgroup_create,
9187 .destroy = cpu_cgroup_destroy,
9188 .can_attach = cpu_cgroup_can_attach,
9189 .attach = cpu_cgroup_attach,
9190 .populate = cpu_cgroup_populate,
9191 .subsys_id = cpu_cgroup_subsys_id,
68318b8e
SV
9192 .early_init = 1,
9193};
9194
052f1dc7 9195#endif /* CONFIG_CGROUP_SCHED */
d842de87
SV
9196
9197#ifdef CONFIG_CGROUP_CPUACCT
9198
9199/*
9200 * CPU accounting code for task groups.
9201 *
9202 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9203 * (balbir@in.ibm.com).
9204 */
9205
934352f2 9206/* track cpu usage of a group of tasks and its child groups */
d842de87
SV
9207struct cpuacct {
9208 struct cgroup_subsys_state css;
9209 /* cpuusage holds pointer to a u64-type object on every cpu */
9210 u64 *cpuusage;
934352f2 9211 struct cpuacct *parent;
d842de87
SV
9212};
9213
9214struct cgroup_subsys cpuacct_subsys;
9215
9216/* return cpu accounting group corresponding to this container */
32cd756a 9217static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
d842de87 9218{
32cd756a 9219 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
d842de87
SV
9220 struct cpuacct, css);
9221}
9222
9223/* return cpu accounting group to which this task belongs */
9224static inline struct cpuacct *task_ca(struct task_struct *tsk)
9225{
9226 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9227 struct cpuacct, css);
9228}
9229
9230/* create a new cpu accounting group */
9231static struct cgroup_subsys_state *cpuacct_create(
32cd756a 9232 struct cgroup_subsys *ss, struct cgroup *cgrp)
d842de87
SV
9233{
9234 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9235
9236 if (!ca)
9237 return ERR_PTR(-ENOMEM);
9238
9239 ca->cpuusage = alloc_percpu(u64);
9240 if (!ca->cpuusage) {
9241 kfree(ca);
9242 return ERR_PTR(-ENOMEM);
9243 }
9244
934352f2
BR
9245 if (cgrp->parent)
9246 ca->parent = cgroup_ca(cgrp->parent);
9247
d842de87
SV
9248 return &ca->css;
9249}
9250
9251/* destroy an existing cpu accounting group */
41a2d6cf 9252static void
32cd756a 9253cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
d842de87 9254{
32cd756a 9255 struct cpuacct *ca = cgroup_ca(cgrp);
d842de87
SV
9256
9257 free_percpu(ca->cpuusage);
9258 kfree(ca);
9259}
9260
9261/* return total cpu usage (in nanoseconds) of a group */
32cd756a 9262static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
d842de87 9263{
32cd756a 9264 struct cpuacct *ca = cgroup_ca(cgrp);
d842de87
SV
9265 u64 totalcpuusage = 0;
9266 int i;
9267
9268 for_each_possible_cpu(i) {
9269 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9270
9271 /*
9272 * Take rq->lock to make 64-bit addition safe on 32-bit
9273 * platforms.
9274 */
9275 spin_lock_irq(&cpu_rq(i)->lock);
9276 totalcpuusage += *cpuusage;
9277 spin_unlock_irq(&cpu_rq(i)->lock);
9278 }
9279
9280 return totalcpuusage;
9281}
9282
0297b803
DG
9283static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9284 u64 reset)
9285{
9286 struct cpuacct *ca = cgroup_ca(cgrp);
9287 int err = 0;
9288 int i;
9289
9290 if (reset) {
9291 err = -EINVAL;
9292 goto out;
9293 }
9294
9295 for_each_possible_cpu(i) {
9296 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9297
9298 spin_lock_irq(&cpu_rq(i)->lock);
9299 *cpuusage = 0;
9300 spin_unlock_irq(&cpu_rq(i)->lock);
9301 }
9302out:
9303 return err;
9304}
9305
d842de87
SV
9306static struct cftype files[] = {
9307 {
9308 .name = "usage",
f4c753b7
PM
9309 .read_u64 = cpuusage_read,
9310 .write_u64 = cpuusage_write,
d842de87
SV
9311 },
9312};
9313
32cd756a 9314static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
d842de87 9315{
32cd756a 9316 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
d842de87
SV
9317}
9318
9319/*
9320 * charge this task's execution time to its accounting group.
9321 *
9322 * called with rq->lock held.
9323 */
9324static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9325{
9326 struct cpuacct *ca;
934352f2 9327 int cpu;
d842de87
SV
9328
9329 if (!cpuacct_subsys.active)
9330 return;
9331
934352f2 9332 cpu = task_cpu(tsk);
d842de87 9333 ca = task_ca(tsk);
d842de87 9334
934352f2
BR
9335 for (; ca; ca = ca->parent) {
9336 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
d842de87
SV
9337 *cpuusage += cputime;
9338 }
9339}
9340
9341struct cgroup_subsys cpuacct_subsys = {
9342 .name = "cpuacct",
9343 .create = cpuacct_create,
9344 .destroy = cpuacct_destroy,
9345 .populate = cpuacct_populate,
9346 .subsys_id = cpuacct_subsys_id,
9347};
9348#endif /* CONFIG_CGROUP_CPUACCT */