]> bbs.cooldavid.org Git - net-next-2.6.git/blame - kernel/sched.c
sched: no need for 'affine wakeup' balancing
[net-next-2.6.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>
58#include <linux/seq_file.h>
e692ab53 59#include <linux/sysctl.h>
1da177e4
LT
60#include <linux/syscalls.h>
61#include <linux/times.h>
8f0ab514 62#include <linux/tsacct_kern.h>
c6fd91f0 63#include <linux/kprobes.h>
0ff92245 64#include <linux/delayacct.h>
5517d86b 65#include <linux/reciprocal_div.h>
dff06c15 66#include <linux/unistd.h>
f5ff8422 67#include <linux/pagemap.h>
1da177e4 68
5517d86b 69#include <asm/tlb.h>
838225b4 70#include <asm/irq_regs.h>
1da177e4 71
b035b6de
AD
72/*
73 * Scheduler clock - returns current time in nanosec units.
74 * This is default implementation.
75 * Architectures and sub-architectures can override this.
76 */
77unsigned long long __attribute__((weak)) sched_clock(void)
78{
d6322faf 79 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
b035b6de
AD
80}
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
5517d86b
ED
116#ifdef CONFIG_SMP
117/*
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
120 */
121static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
122{
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
124}
125
126/*
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
129 */
130static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
131{
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
134}
135#endif
136
e05606d3
IM
137static inline int rt_policy(int policy)
138{
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
140 return 1;
141 return 0;
142}
143
144static inline int task_has_rt_policy(struct task_struct *p)
145{
146 return rt_policy(p->policy);
147}
148
1da177e4 149/*
6aa645ea 150 * This is the priority-queue data structure of the RT scheduling class:
1da177e4 151 */
6aa645ea
IM
152struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
155};
156
29f59db3
SV
157#ifdef CONFIG_FAIR_GROUP_SCHED
158
68318b8e
SV
159#include <linux/cgroup.h>
160
29f59db3
SV
161struct cfs_rq;
162
163/* task group related information */
4cf86d77 164struct task_group {
68318b8e
SV
165#ifdef CONFIG_FAIR_CGROUP_SCHED
166 struct cgroup_subsys_state css;
167#endif
29f59db3
SV
168 /* schedulable entities of this group on each cpu */
169 struct sched_entity **se;
170 /* runqueue "owned" by this group on each cpu */
171 struct cfs_rq **cfs_rq;
6b2d7700
SV
172
173 /*
174 * shares assigned to a task group governs how much of cpu bandwidth
175 * is allocated to the group. The more shares a group has, the more is
176 * the cpu bandwidth allocated to it.
177 *
178 * For ex, lets say that there are three task groups, A, B and C which
179 * have been assigned shares 1000, 2000 and 3000 respectively. Then,
180 * cpu bandwidth allocated by the scheduler to task groups A, B and C
181 * should be:
182 *
183 * Bw(A) = 1000/(1000+2000+3000) * 100 = 16.66%
184 * Bw(B) = 2000/(1000+2000+3000) * 100 = 33.33%
185 * Bw(C) = 3000/(1000+2000+3000) * 100 = 50%
186 *
187 * The weight assigned to a task group's schedulable entities on every
188 * cpu (task_group.se[a_cpu]->load.weight) is derived from the task
189 * group's shares. For ex: lets say that task group A has been
190 * assigned shares of 1000 and there are two CPUs in a system. Then,
191 *
192 * tg_A->se[0]->load.weight = tg_A->se[1]->load.weight = 1000;
193 *
194 * Note: It's not necessary that each of a task's group schedulable
195 * entity have the same weight on all CPUs. If the group
196 * has 2 of its tasks on CPU0 and 1 task on CPU1, then a
197 * better distribution of weight could be:
198 *
199 * tg_A->se[0]->load.weight = 2/3 * 2000 = 1333
200 * tg_A->se[1]->load.weight = 1/2 * 2000 = 667
201 *
202 * rebalance_shares() is responsible for distributing the shares of a
203 * task groups like this among the group's schedulable entities across
204 * cpus.
205 *
206 */
29f59db3 207 unsigned long shares;
6b2d7700 208
ae8393e5 209 struct rcu_head rcu;
29f59db3
SV
210};
211
212/* Default task group's sched entity on each cpu */
213static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
214/* Default task group's cfs_rq on each cpu */
215static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
216
9b5b7751
SV
217static struct sched_entity *init_sched_entity_p[NR_CPUS];
218static struct cfs_rq *init_cfs_rq_p[NR_CPUS];
29f59db3 219
ec2c507f
SV
220/* task_group_mutex serializes add/remove of task groups and also changes to
221 * a task group's cpu shares.
222 */
223static DEFINE_MUTEX(task_group_mutex);
224
a1835615
SV
225/* doms_cur_mutex serializes access to doms_cur[] array */
226static DEFINE_MUTEX(doms_cur_mutex);
227
6b2d7700
SV
228#ifdef CONFIG_SMP
229/* kernel thread that runs rebalance_shares() periodically */
230static struct task_struct *lb_monitor_task;
231static int load_balance_monitor(void *unused);
232#endif
233
234static void set_se_shares(struct sched_entity *se, unsigned long shares);
235
29f59db3 236/* Default task group.
3a252015 237 * Every task in system belong to this group at bootup.
29f59db3 238 */
4cf86d77 239struct task_group init_task_group = {
0eab9146 240 .se = init_sched_entity_p,
3a252015
IM
241 .cfs_rq = init_cfs_rq_p,
242};
9b5b7751 243
24e377a8 244#ifdef CONFIG_FAIR_USER_SCHED
0eab9146 245# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
24e377a8 246#else
93f992cc 247# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
24e377a8
SV
248#endif
249
0eab9146 250#define MIN_GROUP_SHARES 2
6b2d7700 251
93f992cc 252static int init_task_group_load = INIT_TASK_GROUP_LOAD;
29f59db3
SV
253
254/* return group to which a task belongs */
4cf86d77 255static inline struct task_group *task_group(struct task_struct *p)
29f59db3 256{
4cf86d77 257 struct task_group *tg;
9b5b7751 258
24e377a8
SV
259#ifdef CONFIG_FAIR_USER_SCHED
260 tg = p->user->tg;
68318b8e
SV
261#elif defined(CONFIG_FAIR_CGROUP_SCHED)
262 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
263 struct task_group, css);
24e377a8 264#else
41a2d6cf 265 tg = &init_task_group;
24e377a8 266#endif
9b5b7751 267 return tg;
29f59db3
SV
268}
269
270/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
ce96b5ac 271static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu)
29f59db3 272{
ce96b5ac
DA
273 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
274 p->se.parent = task_group(p)->se[cpu];
29f59db3
SV
275}
276
ec2c507f
SV
277static inline void lock_task_group_list(void)
278{
279 mutex_lock(&task_group_mutex);
280}
281
282static inline void unlock_task_group_list(void)
283{
284 mutex_unlock(&task_group_mutex);
285}
286
a1835615
SV
287static inline void lock_doms_cur(void)
288{
289 mutex_lock(&doms_cur_mutex);
290}
291
292static inline void unlock_doms_cur(void)
293{
294 mutex_unlock(&doms_cur_mutex);
295}
296
29f59db3
SV
297#else
298
ce96b5ac 299static inline void set_task_cfs_rq(struct task_struct *p, unsigned int cpu) { }
ec2c507f
SV
300static inline void lock_task_group_list(void) { }
301static inline void unlock_task_group_list(void) { }
a1835615
SV
302static inline void lock_doms_cur(void) { }
303static inline void unlock_doms_cur(void) { }
29f59db3
SV
304
305#endif /* CONFIG_FAIR_GROUP_SCHED */
306
6aa645ea
IM
307/* CFS-related fields in a runqueue */
308struct cfs_rq {
309 struct load_weight load;
310 unsigned long nr_running;
311
6aa645ea 312 u64 exec_clock;
e9acbff6 313 u64 min_vruntime;
6aa645ea
IM
314
315 struct rb_root tasks_timeline;
316 struct rb_node *rb_leftmost;
317 struct rb_node *rb_load_balance_curr;
6aa645ea
IM
318 /* 'curr' points to currently running entity on this cfs_rq.
319 * It is set to NULL otherwise (i.e when none are currently running).
320 */
321 struct sched_entity *curr;
ddc97297
PZ
322
323 unsigned long nr_spread_over;
324
62160e3f 325#ifdef CONFIG_FAIR_GROUP_SCHED
6aa645ea
IM
326 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
327
41a2d6cf
IM
328 /*
329 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
6aa645ea
IM
330 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
331 * (like users, containers etc.)
332 *
333 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
334 * list is used during load balance.
335 */
41a2d6cf
IM
336 struct list_head leaf_cfs_rq_list;
337 struct task_group *tg; /* group that "owns" this runqueue */
6aa645ea
IM
338#endif
339};
1da177e4 340
6aa645ea
IM
341/* Real-Time classes' related field in a runqueue: */
342struct rt_rq {
343 struct rt_prio_array active;
344 int rt_load_balance_idx;
345 struct list_head *rt_load_balance_head, *rt_load_balance_curr;
63489e45 346 unsigned long rt_nr_running;
73fe6aae 347 unsigned long rt_nr_migratory;
764a9d6f
SR
348 /* highest queued rt task prio */
349 int highest_prio;
a22d7fc1 350 int overloaded;
6aa645ea
IM
351};
352
57d885fe
GH
353#ifdef CONFIG_SMP
354
355/*
356 * We add the notion of a root-domain which will be used to define per-domain
0eab9146
IM
357 * variables. Each exclusive cpuset essentially defines an island domain by
358 * fully partitioning the member cpus from any other cpuset. Whenever a new
57d885fe
GH
359 * exclusive cpuset is created, we also create and attach a new root-domain
360 * object.
361 *
362 * By default the system creates a single root-domain with all cpus as
363 * members (mimicking the global state we have today).
364 */
365struct root_domain {
366 atomic_t refcount;
367 cpumask_t span;
368 cpumask_t online;
637f5085 369
0eab9146 370 /*
637f5085
GH
371 * The "RT overload" flag: it gets set if a CPU has more than
372 * one runnable RT task.
373 */
374 cpumask_t rto_mask;
0eab9146 375 atomic_t rto_count;
57d885fe
GH
376};
377
378static struct root_domain def_root_domain;
379
380#endif
381
1da177e4
LT
382/*
383 * This is the main, per-CPU runqueue data structure.
384 *
385 * Locking rule: those places that want to lock multiple runqueues
386 * (such as the load balancing or the thread migration code), lock
387 * acquire operations must be ordered by ascending &runqueue.
388 */
70b97a7f 389struct rq {
d8016491
IM
390 /* runqueue lock: */
391 spinlock_t lock;
1da177e4
LT
392
393 /*
394 * nr_running and cpu_load should be in the same cacheline because
395 * remote CPUs use both these fields when doing load calculation.
396 */
397 unsigned long nr_running;
6aa645ea
IM
398 #define CPU_LOAD_IDX_MAX 5
399 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
bdecea3a 400 unsigned char idle_at_tick;
46cb4b7c
SS
401#ifdef CONFIG_NO_HZ
402 unsigned char in_nohz_recently;
403#endif
d8016491
IM
404 /* capture load from *all* tasks on this cpu: */
405 struct load_weight load;
6aa645ea
IM
406 unsigned long nr_load_updates;
407 u64 nr_switches;
408
409 struct cfs_rq cfs;
410#ifdef CONFIG_FAIR_GROUP_SCHED
d8016491
IM
411 /* list of leaf cfs_rq on this cpu: */
412 struct list_head leaf_cfs_rq_list;
1da177e4 413#endif
41a2d6cf 414 struct rt_rq rt;
1da177e4
LT
415
416 /*
417 * This is part of a global counter where only the total sum
418 * over all CPUs matters. A task can increase this counter on
419 * one CPU and if it got migrated afterwards it may decrease
420 * it on another CPU. Always updated under the runqueue lock:
421 */
422 unsigned long nr_uninterruptible;
423
36c8b586 424 struct task_struct *curr, *idle;
c9819f45 425 unsigned long next_balance;
1da177e4 426 struct mm_struct *prev_mm;
6aa645ea 427
6aa645ea
IM
428 u64 clock, prev_clock_raw;
429 s64 clock_max_delta;
430
431 unsigned int clock_warps, clock_overflows;
2aa44d05
IM
432 u64 idle_clock;
433 unsigned int clock_deep_idle_events;
529c7726 434 u64 tick_timestamp;
6aa645ea 435
1da177e4
LT
436 atomic_t nr_iowait;
437
438#ifdef CONFIG_SMP
0eab9146 439 struct root_domain *rd;
1da177e4
LT
440 struct sched_domain *sd;
441
442 /* For active balancing */
443 int active_balance;
444 int push_cpu;
d8016491
IM
445 /* cpu of this runqueue: */
446 int cpu;
1da177e4 447
36c8b586 448 struct task_struct *migration_thread;
1da177e4
LT
449 struct list_head migration_queue;
450#endif
451
452#ifdef CONFIG_SCHEDSTATS
453 /* latency stats */
454 struct sched_info rq_sched_info;
455
456 /* sys_sched_yield() stats */
480b9434
KC
457 unsigned int yld_exp_empty;
458 unsigned int yld_act_empty;
459 unsigned int yld_both_empty;
460 unsigned int yld_count;
1da177e4
LT
461
462 /* schedule() stats */
480b9434
KC
463 unsigned int sched_switch;
464 unsigned int sched_count;
465 unsigned int sched_goidle;
1da177e4
LT
466
467 /* try_to_wake_up() stats */
480b9434
KC
468 unsigned int ttwu_count;
469 unsigned int ttwu_local;
b8efb561
IM
470
471 /* BKL stats */
480b9434 472 unsigned int bkl_count;
1da177e4 473#endif
fcb99371 474 struct lock_class_key rq_lock_key;
1da177e4
LT
475};
476
f34e3b61 477static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
1da177e4 478
dd41f596
IM
479static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
480{
481 rq->curr->sched_class->check_preempt_curr(rq, p);
482}
483
0a2966b4
CL
484static inline int cpu_of(struct rq *rq)
485{
486#ifdef CONFIG_SMP
487 return rq->cpu;
488#else
489 return 0;
490#endif
491}
492
20d315d4 493/*
b04a0f4c
IM
494 * Update the per-runqueue clock, as finegrained as the platform can give
495 * us, but without assuming monotonicity, etc.:
20d315d4 496 */
b04a0f4c 497static void __update_rq_clock(struct rq *rq)
20d315d4
IM
498{
499 u64 prev_raw = rq->prev_clock_raw;
500 u64 now = sched_clock();
501 s64 delta = now - prev_raw;
502 u64 clock = rq->clock;
503
b04a0f4c
IM
504#ifdef CONFIG_SCHED_DEBUG
505 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
506#endif
20d315d4
IM
507 /*
508 * Protect against sched_clock() occasionally going backwards:
509 */
510 if (unlikely(delta < 0)) {
511 clock++;
512 rq->clock_warps++;
513 } else {
514 /*
515 * Catch too large forward jumps too:
516 */
529c7726
IM
517 if (unlikely(clock + delta > rq->tick_timestamp + TICK_NSEC)) {
518 if (clock < rq->tick_timestamp + TICK_NSEC)
519 clock = rq->tick_timestamp + TICK_NSEC;
520 else
521 clock++;
20d315d4
IM
522 rq->clock_overflows++;
523 } else {
524 if (unlikely(delta > rq->clock_max_delta))
525 rq->clock_max_delta = delta;
526 clock += delta;
527 }
528 }
529
530 rq->prev_clock_raw = now;
531 rq->clock = clock;
b04a0f4c 532}
20d315d4 533
b04a0f4c
IM
534static void update_rq_clock(struct rq *rq)
535{
536 if (likely(smp_processor_id() == cpu_of(rq)))
537 __update_rq_clock(rq);
20d315d4
IM
538}
539
674311d5
NP
540/*
541 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
1a20ff27 542 * See detach_destroy_domains: synchronize_sched for details.
674311d5
NP
543 *
544 * The domain tree of any CPU may only be accessed from within
545 * preempt-disabled sections.
546 */
48f24c4d
IM
547#define for_each_domain(cpu, __sd) \
548 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
1da177e4
LT
549
550#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
551#define this_rq() (&__get_cpu_var(runqueues))
552#define task_rq(p) cpu_rq(task_cpu(p))
553#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
554
bf5c91ba
IM
555/*
556 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
557 */
558#ifdef CONFIG_SCHED_DEBUG
559# define const_debug __read_mostly
560#else
561# define const_debug static const
562#endif
563
564/*
565 * Debugging: various feature bits
566 */
567enum {
bbdba7c0 568 SCHED_FEAT_NEW_FAIR_SLEEPERS = 1,
9612633a
IM
569 SCHED_FEAT_WAKEUP_PREEMPT = 2,
570 SCHED_FEAT_START_DEBIT = 4,
41a2d6cf
IM
571 SCHED_FEAT_TREE_AVG = 8,
572 SCHED_FEAT_APPROX_AVG = 16,
bf5c91ba
IM
573};
574
575const_debug unsigned int sysctl_sched_features =
8401f775 576 SCHED_FEAT_NEW_FAIR_SLEEPERS * 1 |
9612633a 577 SCHED_FEAT_WAKEUP_PREEMPT * 1 |
8401f775
IM
578 SCHED_FEAT_START_DEBIT * 1 |
579 SCHED_FEAT_TREE_AVG * 0 |
9612633a 580 SCHED_FEAT_APPROX_AVG * 0;
bf5c91ba
IM
581
582#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
583
b82d9fdd
PZ
584/*
585 * Number of tasks to iterate in a single balance run.
586 * Limited because this is done with IRQs disabled.
587 */
588const_debug unsigned int sysctl_sched_nr_migrate = 32;
589
e436d800
IM
590/*
591 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
592 * clock constructed from sched_clock():
593 */
594unsigned long long cpu_clock(int cpu)
595{
e436d800
IM
596 unsigned long long now;
597 unsigned long flags;
b04a0f4c 598 struct rq *rq;
e436d800 599
2cd4d0ea 600 local_irq_save(flags);
b04a0f4c 601 rq = cpu_rq(cpu);
8ced5f69
IM
602 /*
603 * Only call sched_clock() if the scheduler has already been
604 * initialized (some code might call cpu_clock() very early):
605 */
606 if (rq->idle)
607 update_rq_clock(rq);
b04a0f4c 608 now = rq->clock;
2cd4d0ea 609 local_irq_restore(flags);
e436d800
IM
610
611 return now;
612}
a58f6f25 613EXPORT_SYMBOL_GPL(cpu_clock);
e436d800 614
1da177e4 615#ifndef prepare_arch_switch
4866cde0
NP
616# define prepare_arch_switch(next) do { } while (0)
617#endif
618#ifndef finish_arch_switch
619# define finish_arch_switch(prev) do { } while (0)
620#endif
621
051a1d1a
DA
622static inline int task_current(struct rq *rq, struct task_struct *p)
623{
624 return rq->curr == p;
625}
626
4866cde0 627#ifndef __ARCH_WANT_UNLOCKED_CTXSW
70b97a7f 628static inline int task_running(struct rq *rq, struct task_struct *p)
4866cde0 629{
051a1d1a 630 return task_current(rq, p);
4866cde0
NP
631}
632
70b97a7f 633static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
4866cde0
NP
634{
635}
636
70b97a7f 637static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
4866cde0 638{
da04c035
IM
639#ifdef CONFIG_DEBUG_SPINLOCK
640 /* this is a valid case when another task releases the spinlock */
641 rq->lock.owner = current;
642#endif
8a25d5de
IM
643 /*
644 * If we are tracking spinlock dependencies then we have to
645 * fix up the runqueue lock - which gets 'carried over' from
646 * prev into current:
647 */
648 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
649
4866cde0
NP
650 spin_unlock_irq(&rq->lock);
651}
652
653#else /* __ARCH_WANT_UNLOCKED_CTXSW */
70b97a7f 654static inline int task_running(struct rq *rq, struct task_struct *p)
4866cde0
NP
655{
656#ifdef CONFIG_SMP
657 return p->oncpu;
658#else
051a1d1a 659 return task_current(rq, p);
4866cde0
NP
660#endif
661}
662
70b97a7f 663static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
4866cde0
NP
664{
665#ifdef CONFIG_SMP
666 /*
667 * We can optimise this out completely for !SMP, because the
668 * SMP rebalancing from interrupt is the only thing that cares
669 * here.
670 */
671 next->oncpu = 1;
672#endif
673#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
674 spin_unlock_irq(&rq->lock);
675#else
676 spin_unlock(&rq->lock);
677#endif
678}
679
70b97a7f 680static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
4866cde0
NP
681{
682#ifdef CONFIG_SMP
683 /*
684 * After ->oncpu is cleared, the task can be moved to a different CPU.
685 * We must ensure this doesn't happen until the switch is completely
686 * finished.
687 */
688 smp_wmb();
689 prev->oncpu = 0;
690#endif
691#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
692 local_irq_enable();
1da177e4 693#endif
4866cde0
NP
694}
695#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1da177e4 696
b29739f9
IM
697/*
698 * __task_rq_lock - lock the runqueue a given task resides on.
699 * Must be called interrupts disabled.
700 */
70b97a7f 701static inline struct rq *__task_rq_lock(struct task_struct *p)
b29739f9
IM
702 __acquires(rq->lock)
703{
3a5c359a
AK
704 for (;;) {
705 struct rq *rq = task_rq(p);
706 spin_lock(&rq->lock);
707 if (likely(rq == task_rq(p)))
708 return rq;
b29739f9 709 spin_unlock(&rq->lock);
b29739f9 710 }
b29739f9
IM
711}
712
1da177e4
LT
713/*
714 * task_rq_lock - lock the runqueue a given task resides on and disable
41a2d6cf 715 * interrupts. Note the ordering: we can safely lookup the task_rq without
1da177e4
LT
716 * explicitly disabling preemption.
717 */
70b97a7f 718static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1da177e4
LT
719 __acquires(rq->lock)
720{
70b97a7f 721 struct rq *rq;
1da177e4 722
3a5c359a
AK
723 for (;;) {
724 local_irq_save(*flags);
725 rq = task_rq(p);
726 spin_lock(&rq->lock);
727 if (likely(rq == task_rq(p)))
728 return rq;
1da177e4 729 spin_unlock_irqrestore(&rq->lock, *flags);
1da177e4 730 }
1da177e4
LT
731}
732
a9957449 733static void __task_rq_unlock(struct rq *rq)
b29739f9
IM
734 __releases(rq->lock)
735{
736 spin_unlock(&rq->lock);
737}
738
70b97a7f 739static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1da177e4
LT
740 __releases(rq->lock)
741{
742 spin_unlock_irqrestore(&rq->lock, *flags);
743}
744
1da177e4 745/*
cc2a73b5 746 * this_rq_lock - lock this runqueue and disable interrupts.
1da177e4 747 */
a9957449 748static struct rq *this_rq_lock(void)
1da177e4
LT
749 __acquires(rq->lock)
750{
70b97a7f 751 struct rq *rq;
1da177e4
LT
752
753 local_irq_disable();
754 rq = this_rq();
755 spin_lock(&rq->lock);
756
757 return rq;
758}
759
1b9f19c2 760/*
2aa44d05 761 * We are going deep-idle (irqs are disabled):
1b9f19c2 762 */
2aa44d05 763void sched_clock_idle_sleep_event(void)
1b9f19c2 764{
2aa44d05
IM
765 struct rq *rq = cpu_rq(smp_processor_id());
766
767 spin_lock(&rq->lock);
768 __update_rq_clock(rq);
769 spin_unlock(&rq->lock);
770 rq->clock_deep_idle_events++;
771}
772EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
773
774/*
775 * We just idled delta nanoseconds (called with irqs disabled):
776 */
777void sched_clock_idle_wakeup_event(u64 delta_ns)
778{
779 struct rq *rq = cpu_rq(smp_processor_id());
780 u64 now = sched_clock();
1b9f19c2 781
2bacec8c 782 touch_softlockup_watchdog();
2aa44d05
IM
783 rq->idle_clock += delta_ns;
784 /*
785 * Override the previous timestamp and ignore all
786 * sched_clock() deltas that occured while we idled,
787 * and use the PM-provided delta_ns to advance the
788 * rq clock:
789 */
790 spin_lock(&rq->lock);
791 rq->prev_clock_raw = now;
792 rq->clock += delta_ns;
793 spin_unlock(&rq->lock);
1b9f19c2 794}
2aa44d05 795EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
1b9f19c2 796
c24d20db
IM
797/*
798 * resched_task - mark a task 'to be rescheduled now'.
799 *
800 * On UP this means the setting of the need_resched flag, on SMP it
801 * might also involve a cross-CPU call to trigger the scheduler on
802 * the target CPU.
803 */
804#ifdef CONFIG_SMP
805
806#ifndef tsk_is_polling
807#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
808#endif
809
810static void resched_task(struct task_struct *p)
811{
812 int cpu;
813
814 assert_spin_locked(&task_rq(p)->lock);
815
816 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
817 return;
818
819 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
820
821 cpu = task_cpu(p);
822 if (cpu == smp_processor_id())
823 return;
824
825 /* NEED_RESCHED must be visible before we test polling */
826 smp_mb();
827 if (!tsk_is_polling(p))
828 smp_send_reschedule(cpu);
829}
830
831static void resched_cpu(int cpu)
832{
833 struct rq *rq = cpu_rq(cpu);
834 unsigned long flags;
835
836 if (!spin_trylock_irqsave(&rq->lock, flags))
837 return;
838 resched_task(cpu_curr(cpu));
839 spin_unlock_irqrestore(&rq->lock, flags);
840}
841#else
842static inline void resched_task(struct task_struct *p)
843{
844 assert_spin_locked(&task_rq(p)->lock);
845 set_tsk_need_resched(p);
846}
847#endif
848
45bf76df
IM
849#if BITS_PER_LONG == 32
850# define WMULT_CONST (~0UL)
851#else
852# define WMULT_CONST (1UL << 32)
853#endif
854
855#define WMULT_SHIFT 32
856
194081eb
IM
857/*
858 * Shift right and round:
859 */
cf2ab469 860#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
194081eb 861
cb1c4fc9 862static unsigned long
45bf76df
IM
863calc_delta_mine(unsigned long delta_exec, unsigned long weight,
864 struct load_weight *lw)
865{
866 u64 tmp;
867
868 if (unlikely(!lw->inv_weight))
194081eb 869 lw->inv_weight = (WMULT_CONST - lw->weight/2) / lw->weight + 1;
45bf76df
IM
870
871 tmp = (u64)delta_exec * weight;
872 /*
873 * Check whether we'd overflow the 64-bit multiplication:
874 */
194081eb 875 if (unlikely(tmp > WMULT_CONST))
cf2ab469 876 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
194081eb
IM
877 WMULT_SHIFT/2);
878 else
cf2ab469 879 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
45bf76df 880
ecf691da 881 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
45bf76df
IM
882}
883
884static inline unsigned long
885calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
886{
887 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
888}
889
1091985b 890static inline void update_load_add(struct load_weight *lw, unsigned long inc)
45bf76df
IM
891{
892 lw->weight += inc;
45bf76df
IM
893}
894
1091985b 895static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
45bf76df
IM
896{
897 lw->weight -= dec;
45bf76df
IM
898}
899
2dd73a4f
PW
900/*
901 * To aid in avoiding the subversion of "niceness" due to uneven distribution
902 * of tasks with abnormal "nice" values across CPUs the contribution that
903 * each task makes to its run queue's load is weighted according to its
41a2d6cf 904 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
2dd73a4f
PW
905 * scaled version of the new time slice allocation that they receive on time
906 * slice expiry etc.
907 */
908
dd41f596
IM
909#define WEIGHT_IDLEPRIO 2
910#define WMULT_IDLEPRIO (1 << 31)
911
912/*
913 * Nice levels are multiplicative, with a gentle 10% change for every
914 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
915 * nice 1, it will get ~10% less CPU time than another CPU-bound task
916 * that remained on nice 0.
917 *
918 * The "10% effect" is relative and cumulative: from _any_ nice level,
919 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
f9153ee6
IM
920 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
921 * If a task goes up by ~10% and another task goes down by ~10% then
922 * the relative distance between them is ~25%.)
dd41f596
IM
923 */
924static const int prio_to_weight[40] = {
254753dc
IM
925 /* -20 */ 88761, 71755, 56483, 46273, 36291,
926 /* -15 */ 29154, 23254, 18705, 14949, 11916,
927 /* -10 */ 9548, 7620, 6100, 4904, 3906,
928 /* -5 */ 3121, 2501, 1991, 1586, 1277,
929 /* 0 */ 1024, 820, 655, 526, 423,
930 /* 5 */ 335, 272, 215, 172, 137,
931 /* 10 */ 110, 87, 70, 56, 45,
932 /* 15 */ 36, 29, 23, 18, 15,
dd41f596
IM
933};
934
5714d2de
IM
935/*
936 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
937 *
938 * In cases where the weight does not change often, we can use the
939 * precalculated inverse to speed up arithmetics by turning divisions
940 * into multiplications:
941 */
dd41f596 942static const u32 prio_to_wmult[40] = {
254753dc
IM
943 /* -20 */ 48388, 59856, 76040, 92818, 118348,
944 /* -15 */ 147320, 184698, 229616, 287308, 360437,
945 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
946 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
947 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
948 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
949 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
950 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
dd41f596 951};
2dd73a4f 952
dd41f596
IM
953static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
954
955/*
956 * runqueue iterator, to support SMP load-balancing between different
957 * scheduling classes, without having to expose their internal data
958 * structures to the load-balancing proper:
959 */
960struct rq_iterator {
961 void *arg;
962 struct task_struct *(*start)(void *);
963 struct task_struct *(*next)(void *);
964};
965
e1d1484f
PW
966#ifdef CONFIG_SMP
967static unsigned long
968balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
969 unsigned long max_load_move, struct sched_domain *sd,
970 enum cpu_idle_type idle, int *all_pinned,
971 int *this_best_prio, struct rq_iterator *iterator);
972
973static int
974iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
975 struct sched_domain *sd, enum cpu_idle_type idle,
976 struct rq_iterator *iterator);
e1d1484f 977#endif
dd41f596 978
d842de87
SV
979#ifdef CONFIG_CGROUP_CPUACCT
980static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
981#else
982static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
983#endif
984
58e2d4ca
SV
985static inline void inc_cpu_load(struct rq *rq, unsigned long load)
986{
987 update_load_add(&rq->load, load);
988}
989
990static inline void dec_cpu_load(struct rq *rq, unsigned long load)
991{
992 update_load_sub(&rq->load, load);
993}
994
e7693a36
GH
995#ifdef CONFIG_SMP
996static unsigned long source_load(int cpu, int type);
997static unsigned long target_load(int cpu, int type);
998static unsigned long cpu_avg_load_per_task(int cpu);
999static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1000#endif /* CONFIG_SMP */
1001
dd41f596 1002#include "sched_stats.h"
dd41f596 1003#include "sched_idletask.c"
5522d5d5
IM
1004#include "sched_fair.c"
1005#include "sched_rt.c"
dd41f596
IM
1006#ifdef CONFIG_SCHED_DEBUG
1007# include "sched_debug.c"
1008#endif
1009
1010#define sched_class_highest (&rt_sched_class)
1011
e5fa2237 1012static void inc_nr_running(struct task_struct *p, struct rq *rq)
9c217245
IM
1013{
1014 rq->nr_running++;
9c217245
IM
1015}
1016
db53181e 1017static void dec_nr_running(struct task_struct *p, struct rq *rq)
9c217245
IM
1018{
1019 rq->nr_running--;
9c217245
IM
1020}
1021
45bf76df
IM
1022static void set_load_weight(struct task_struct *p)
1023{
1024 if (task_has_rt_policy(p)) {
dd41f596
IM
1025 p->se.load.weight = prio_to_weight[0] * 2;
1026 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1027 return;
1028 }
45bf76df 1029
dd41f596
IM
1030 /*
1031 * SCHED_IDLE tasks get minimal weight:
1032 */
1033 if (p->policy == SCHED_IDLE) {
1034 p->se.load.weight = WEIGHT_IDLEPRIO;
1035 p->se.load.inv_weight = WMULT_IDLEPRIO;
1036 return;
1037 }
71f8bd46 1038
dd41f596
IM
1039 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1040 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
71f8bd46
IM
1041}
1042
8159f87e 1043static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
71f8bd46 1044{
dd41f596 1045 sched_info_queued(p);
fd390f6a 1046 p->sched_class->enqueue_task(rq, p, wakeup);
dd41f596 1047 p->se.on_rq = 1;
71f8bd46
IM
1048}
1049
69be72c1 1050static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
71f8bd46 1051{
f02231e5 1052 p->sched_class->dequeue_task(rq, p, sleep);
dd41f596 1053 p->se.on_rq = 0;
71f8bd46
IM
1054}
1055
14531189 1056/*
dd41f596 1057 * __normal_prio - return the priority that is based on the static prio
14531189 1058 */
14531189
IM
1059static inline int __normal_prio(struct task_struct *p)
1060{
dd41f596 1061 return p->static_prio;
14531189
IM
1062}
1063
b29739f9
IM
1064/*
1065 * Calculate the expected normal priority: i.e. priority
1066 * without taking RT-inheritance into account. Might be
1067 * boosted by interactivity modifiers. Changes upon fork,
1068 * setprio syscalls, and whenever the interactivity
1069 * estimator recalculates.
1070 */
36c8b586 1071static inline int normal_prio(struct task_struct *p)
b29739f9
IM
1072{
1073 int prio;
1074
e05606d3 1075 if (task_has_rt_policy(p))
b29739f9
IM
1076 prio = MAX_RT_PRIO-1 - p->rt_priority;
1077 else
1078 prio = __normal_prio(p);
1079 return prio;
1080}
1081
1082/*
1083 * Calculate the current priority, i.e. the priority
1084 * taken into account by the scheduler. This value might
1085 * be boosted by RT tasks, or might be boosted by
1086 * interactivity modifiers. Will be RT if the task got
1087 * RT-boosted. If not then it returns p->normal_prio.
1088 */
36c8b586 1089static int effective_prio(struct task_struct *p)
b29739f9
IM
1090{
1091 p->normal_prio = normal_prio(p);
1092 /*
1093 * If we are RT tasks or we were boosted to RT priority,
1094 * keep the priority unchanged. Otherwise, update priority
1095 * to the normal priority:
1096 */
1097 if (!rt_prio(p->prio))
1098 return p->normal_prio;
1099 return p->prio;
1100}
1101
1da177e4 1102/*
dd41f596 1103 * activate_task - move a task to the runqueue.
1da177e4 1104 */
dd41f596 1105static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1da177e4 1106{
dd41f596
IM
1107 if (p->state == TASK_UNINTERRUPTIBLE)
1108 rq->nr_uninterruptible--;
1da177e4 1109
8159f87e 1110 enqueue_task(rq, p, wakeup);
e5fa2237 1111 inc_nr_running(p, rq);
1da177e4
LT
1112}
1113
1da177e4
LT
1114/*
1115 * deactivate_task - remove a task from the runqueue.
1116 */
2e1cb74a 1117static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1da177e4 1118{
dd41f596
IM
1119 if (p->state == TASK_UNINTERRUPTIBLE)
1120 rq->nr_uninterruptible++;
1121
69be72c1 1122 dequeue_task(rq, p, sleep);
db53181e 1123 dec_nr_running(p, rq);
1da177e4
LT
1124}
1125
1da177e4
LT
1126/**
1127 * task_curr - is this task currently executing on a CPU?
1128 * @p: the task in question.
1129 */
36c8b586 1130inline int task_curr(const struct task_struct *p)
1da177e4
LT
1131{
1132 return cpu_curr(task_cpu(p)) == p;
1133}
1134
2dd73a4f
PW
1135/* Used instead of source_load when we know the type == 0 */
1136unsigned long weighted_cpuload(const int cpu)
1137{
495eca49 1138 return cpu_rq(cpu)->load.weight;
dd41f596
IM
1139}
1140
1141static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1142{
ce96b5ac 1143 set_task_cfs_rq(p, cpu);
dd41f596 1144#ifdef CONFIG_SMP
ce96b5ac
DA
1145 /*
1146 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1147 * successfuly executed on another CPU. We must ensure that updates of
1148 * per-task data have been completed by this moment.
1149 */
1150 smp_wmb();
dd41f596 1151 task_thread_info(p)->cpu = cpu;
dd41f596 1152#endif
2dd73a4f
PW
1153}
1154
1da177e4 1155#ifdef CONFIG_SMP
c65cc870 1156
cc367732
IM
1157/*
1158 * Is this task likely cache-hot:
1159 */
e7693a36 1160static int
cc367732
IM
1161task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1162{
1163 s64 delta;
1164
1165 if (p->sched_class != &fair_sched_class)
1166 return 0;
1167
6bc1665b
IM
1168 if (sysctl_sched_migration_cost == -1)
1169 return 1;
1170 if (sysctl_sched_migration_cost == 0)
1171 return 0;
1172
cc367732
IM
1173 delta = now - p->se.exec_start;
1174
1175 return delta < (s64)sysctl_sched_migration_cost;
1176}
1177
1178
dd41f596 1179void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
c65cc870 1180{
dd41f596
IM
1181 int old_cpu = task_cpu(p);
1182 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2830cf8c
SV
1183 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1184 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
bbdba7c0 1185 u64 clock_offset;
dd41f596
IM
1186
1187 clock_offset = old_rq->clock - new_rq->clock;
6cfb0d5d
IM
1188
1189#ifdef CONFIG_SCHEDSTATS
1190 if (p->se.wait_start)
1191 p->se.wait_start -= clock_offset;
dd41f596
IM
1192 if (p->se.sleep_start)
1193 p->se.sleep_start -= clock_offset;
1194 if (p->se.block_start)
1195 p->se.block_start -= clock_offset;
cc367732
IM
1196 if (old_cpu != new_cpu) {
1197 schedstat_inc(p, se.nr_migrations);
1198 if (task_hot(p, old_rq->clock, NULL))
1199 schedstat_inc(p, se.nr_forced2_migrations);
1200 }
6cfb0d5d 1201#endif
2830cf8c
SV
1202 p->se.vruntime -= old_cfsrq->min_vruntime -
1203 new_cfsrq->min_vruntime;
dd41f596
IM
1204
1205 __set_task_cpu(p, new_cpu);
c65cc870
IM
1206}
1207
70b97a7f 1208struct migration_req {
1da177e4 1209 struct list_head list;
1da177e4 1210
36c8b586 1211 struct task_struct *task;
1da177e4
LT
1212 int dest_cpu;
1213
1da177e4 1214 struct completion done;
70b97a7f 1215};
1da177e4
LT
1216
1217/*
1218 * The task's runqueue lock must be held.
1219 * Returns true if you have to wait for migration thread.
1220 */
36c8b586 1221static int
70b97a7f 1222migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1da177e4 1223{
70b97a7f 1224 struct rq *rq = task_rq(p);
1da177e4
LT
1225
1226 /*
1227 * If the task is not on a runqueue (and not running), then
1228 * it is sufficient to simply update the task's cpu field.
1229 */
dd41f596 1230 if (!p->se.on_rq && !task_running(rq, p)) {
1da177e4
LT
1231 set_task_cpu(p, dest_cpu);
1232 return 0;
1233 }
1234
1235 init_completion(&req->done);
1da177e4
LT
1236 req->task = p;
1237 req->dest_cpu = dest_cpu;
1238 list_add(&req->list, &rq->migration_queue);
48f24c4d 1239
1da177e4
LT
1240 return 1;
1241}
1242
1243/*
1244 * wait_task_inactive - wait for a thread to unschedule.
1245 *
1246 * The caller must ensure that the task *will* unschedule sometime soon,
1247 * else this function might spin for a *long* time. This function can't
1248 * be called with interrupts off, or it may introduce deadlock with
1249 * smp_call_function() if an IPI is sent by the same process we are
1250 * waiting to become inactive.
1251 */
36c8b586 1252void wait_task_inactive(struct task_struct *p)
1da177e4
LT
1253{
1254 unsigned long flags;
dd41f596 1255 int running, on_rq;
70b97a7f 1256 struct rq *rq;
1da177e4 1257
3a5c359a
AK
1258 for (;;) {
1259 /*
1260 * We do the initial early heuristics without holding
1261 * any task-queue locks at all. We'll only try to get
1262 * the runqueue lock when things look like they will
1263 * work out!
1264 */
1265 rq = task_rq(p);
fa490cfd 1266
3a5c359a
AK
1267 /*
1268 * If the task is actively running on another CPU
1269 * still, just relax and busy-wait without holding
1270 * any locks.
1271 *
1272 * NOTE! Since we don't hold any locks, it's not
1273 * even sure that "rq" stays as the right runqueue!
1274 * But we don't care, since "task_running()" will
1275 * return false if the runqueue has changed and p
1276 * is actually now running somewhere else!
1277 */
1278 while (task_running(rq, p))
1279 cpu_relax();
fa490cfd 1280
3a5c359a
AK
1281 /*
1282 * Ok, time to look more closely! We need the rq
1283 * lock now, to be *sure*. If we're wrong, we'll
1284 * just go back and repeat.
1285 */
1286 rq = task_rq_lock(p, &flags);
1287 running = task_running(rq, p);
1288 on_rq = p->se.on_rq;
1289 task_rq_unlock(rq, &flags);
fa490cfd 1290
3a5c359a
AK
1291 /*
1292 * Was it really running after all now that we
1293 * checked with the proper locks actually held?
1294 *
1295 * Oops. Go back and try again..
1296 */
1297 if (unlikely(running)) {
1298 cpu_relax();
1299 continue;
1300 }
fa490cfd 1301
3a5c359a
AK
1302 /*
1303 * It's not enough that it's not actively running,
1304 * it must be off the runqueue _entirely_, and not
1305 * preempted!
1306 *
1307 * So if it wa still runnable (but just not actively
1308 * running right now), it's preempted, and we should
1309 * yield - it could be a while.
1310 */
1311 if (unlikely(on_rq)) {
1312 schedule_timeout_uninterruptible(1);
1313 continue;
1314 }
fa490cfd 1315
3a5c359a
AK
1316 /*
1317 * Ahh, all good. It wasn't running, and it wasn't
1318 * runnable, which means that it will never become
1319 * running in the future either. We're all done!
1320 */
1321 break;
1322 }
1da177e4
LT
1323}
1324
1325/***
1326 * kick_process - kick a running thread to enter/exit the kernel
1327 * @p: the to-be-kicked thread
1328 *
1329 * Cause a process which is running on another CPU to enter
1330 * kernel-mode, without any delay. (to get signals handled.)
1331 *
1332 * NOTE: this function doesnt have to take the runqueue lock,
1333 * because all it wants to ensure is that the remote task enters
1334 * the kernel. If the IPI races and the task has been migrated
1335 * to another CPU then no harm is done and the purpose has been
1336 * achieved as well.
1337 */
36c8b586 1338void kick_process(struct task_struct *p)
1da177e4
LT
1339{
1340 int cpu;
1341
1342 preempt_disable();
1343 cpu = task_cpu(p);
1344 if ((cpu != smp_processor_id()) && task_curr(p))
1345 smp_send_reschedule(cpu);
1346 preempt_enable();
1347}
1348
1349/*
2dd73a4f
PW
1350 * Return a low guess at the load of a migration-source cpu weighted
1351 * according to the scheduling class and "nice" value.
1da177e4
LT
1352 *
1353 * We want to under-estimate the load of migration sources, to
1354 * balance conservatively.
1355 */
a9957449 1356static unsigned long source_load(int cpu, int type)
1da177e4 1357{
70b97a7f 1358 struct rq *rq = cpu_rq(cpu);
dd41f596 1359 unsigned long total = weighted_cpuload(cpu);
2dd73a4f 1360
3b0bd9bc 1361 if (type == 0)
dd41f596 1362 return total;
b910472d 1363
dd41f596 1364 return min(rq->cpu_load[type-1], total);
1da177e4
LT
1365}
1366
1367/*
2dd73a4f
PW
1368 * Return a high guess at the load of a migration-target cpu weighted
1369 * according to the scheduling class and "nice" value.
1da177e4 1370 */
a9957449 1371static unsigned long target_load(int cpu, int type)
1da177e4 1372{
70b97a7f 1373 struct rq *rq = cpu_rq(cpu);
dd41f596 1374 unsigned long total = weighted_cpuload(cpu);
2dd73a4f 1375
7897986b 1376 if (type == 0)
dd41f596 1377 return total;
3b0bd9bc 1378
dd41f596 1379 return max(rq->cpu_load[type-1], total);
2dd73a4f
PW
1380}
1381
1382/*
1383 * Return the average load per task on the cpu's run queue
1384 */
e7693a36 1385static unsigned long cpu_avg_load_per_task(int cpu)
2dd73a4f 1386{
70b97a7f 1387 struct rq *rq = cpu_rq(cpu);
dd41f596 1388 unsigned long total = weighted_cpuload(cpu);
2dd73a4f
PW
1389 unsigned long n = rq->nr_running;
1390
dd41f596 1391 return n ? total / n : SCHED_LOAD_SCALE;
1da177e4
LT
1392}
1393
147cbb4b
NP
1394/*
1395 * find_idlest_group finds and returns the least busy CPU group within the
1396 * domain.
1397 */
1398static struct sched_group *
1399find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1400{
1401 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1402 unsigned long min_load = ULONG_MAX, this_load = 0;
1403 int load_idx = sd->forkexec_idx;
1404 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1405
1406 do {
1407 unsigned long load, avg_load;
1408 int local_group;
1409 int i;
1410
da5a5522
BD
1411 /* Skip over this group if it has no CPUs allowed */
1412 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
3a5c359a 1413 continue;
da5a5522 1414
147cbb4b 1415 local_group = cpu_isset(this_cpu, group->cpumask);
147cbb4b
NP
1416
1417 /* Tally up the load of all CPUs in the group */
1418 avg_load = 0;
1419
1420 for_each_cpu_mask(i, group->cpumask) {
1421 /* Bias balancing toward cpus of our domain */
1422 if (local_group)
1423 load = source_load(i, load_idx);
1424 else
1425 load = target_load(i, load_idx);
1426
1427 avg_load += load;
1428 }
1429
1430 /* Adjust by relative CPU power of the group */
5517d86b
ED
1431 avg_load = sg_div_cpu_power(group,
1432 avg_load * SCHED_LOAD_SCALE);
147cbb4b
NP
1433
1434 if (local_group) {
1435 this_load = avg_load;
1436 this = group;
1437 } else if (avg_load < min_load) {
1438 min_load = avg_load;
1439 idlest = group;
1440 }
3a5c359a 1441 } while (group = group->next, group != sd->groups);
147cbb4b
NP
1442
1443 if (!idlest || 100*this_load < imbalance*min_load)
1444 return NULL;
1445 return idlest;
1446}
1447
1448/*
0feaece9 1449 * find_idlest_cpu - find the idlest cpu among the cpus in group.
147cbb4b 1450 */
95cdf3b7
IM
1451static int
1452find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
147cbb4b 1453{
da5a5522 1454 cpumask_t tmp;
147cbb4b
NP
1455 unsigned long load, min_load = ULONG_MAX;
1456 int idlest = -1;
1457 int i;
1458
da5a5522
BD
1459 /* Traverse only the allowed CPUs */
1460 cpus_and(tmp, group->cpumask, p->cpus_allowed);
1461
1462 for_each_cpu_mask(i, tmp) {
2dd73a4f 1463 load = weighted_cpuload(i);
147cbb4b
NP
1464
1465 if (load < min_load || (load == min_load && i == this_cpu)) {
1466 min_load = load;
1467 idlest = i;
1468 }
1469 }
1470
1471 return idlest;
1472}
1473
476d139c
NP
1474/*
1475 * sched_balance_self: balance the current task (running on cpu) in domains
1476 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1477 * SD_BALANCE_EXEC.
1478 *
1479 * Balance, ie. select the least loaded group.
1480 *
1481 * Returns the target CPU number, or the same CPU if no balancing is needed.
1482 *
1483 * preempt must be disabled.
1484 */
1485static int sched_balance_self(int cpu, int flag)
1486{
1487 struct task_struct *t = current;
1488 struct sched_domain *tmp, *sd = NULL;
147cbb4b 1489
c96d145e 1490 for_each_domain(cpu, tmp) {
9761eea8
IM
1491 /*
1492 * If power savings logic is enabled for a domain, stop there.
1493 */
5c45bf27
SS
1494 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1495 break;
476d139c
NP
1496 if (tmp->flags & flag)
1497 sd = tmp;
c96d145e 1498 }
476d139c
NP
1499
1500 while (sd) {
1501 cpumask_t span;
1502 struct sched_group *group;
1a848870
SS
1503 int new_cpu, weight;
1504
1505 if (!(sd->flags & flag)) {
1506 sd = sd->child;
1507 continue;
1508 }
476d139c
NP
1509
1510 span = sd->span;
1511 group = find_idlest_group(sd, t, cpu);
1a848870
SS
1512 if (!group) {
1513 sd = sd->child;
1514 continue;
1515 }
476d139c 1516
da5a5522 1517 new_cpu = find_idlest_cpu(group, t, cpu);
1a848870
SS
1518 if (new_cpu == -1 || new_cpu == cpu) {
1519 /* Now try balancing at a lower domain level of cpu */
1520 sd = sd->child;
1521 continue;
1522 }
476d139c 1523
1a848870 1524 /* Now try balancing at a lower domain level of new_cpu */
476d139c 1525 cpu = new_cpu;
476d139c
NP
1526 sd = NULL;
1527 weight = cpus_weight(span);
1528 for_each_domain(cpu, tmp) {
1529 if (weight <= cpus_weight(tmp->span))
1530 break;
1531 if (tmp->flags & flag)
1532 sd = tmp;
1533 }
1534 /* while loop will break here if sd == NULL */
1535 }
1536
1537 return cpu;
1538}
1539
1540#endif /* CONFIG_SMP */
1da177e4 1541
1da177e4
LT
1542/***
1543 * try_to_wake_up - wake up a thread
1544 * @p: the to-be-woken-up thread
1545 * @state: the mask of task states that can be woken
1546 * @sync: do a synchronous wakeup?
1547 *
1548 * Put it on the run-queue if it's not already there. The "current"
1549 * thread is always on the run-queue (except when the actual
1550 * re-schedule is in progress), and as such you're allowed to do
1551 * the simpler "current->state = TASK_RUNNING" to mark yourself
1552 * runnable without the overhead of this.
1553 *
1554 * returns failure only if the task is already active.
1555 */
36c8b586 1556static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
1da177e4 1557{
cc367732 1558 int cpu, orig_cpu, this_cpu, success = 0;
1da177e4
LT
1559 unsigned long flags;
1560 long old_state;
70b97a7f 1561 struct rq *rq;
1da177e4 1562#ifdef CONFIG_SMP
1da177e4
LT
1563 int new_cpu;
1564#endif
1565
1566 rq = task_rq_lock(p, &flags);
1567 old_state = p->state;
1568 if (!(old_state & state))
1569 goto out;
1570
dd41f596 1571 if (p->se.on_rq)
1da177e4
LT
1572 goto out_running;
1573
1574 cpu = task_cpu(p);
cc367732 1575 orig_cpu = cpu;
1da177e4
LT
1576 this_cpu = smp_processor_id();
1577
1578#ifdef CONFIG_SMP
1579 if (unlikely(task_running(rq, p)))
1580 goto out_activate;
1581
e7693a36 1582 new_cpu = p->sched_class->select_task_rq(p, sync);
1da177e4
LT
1583 if (new_cpu != cpu) {
1584 set_task_cpu(p, new_cpu);
1585 task_rq_unlock(rq, &flags);
1586 /* might preempt at this point */
1587 rq = task_rq_lock(p, &flags);
1588 old_state = p->state;
1589 if (!(old_state & state))
1590 goto out;
dd41f596 1591 if (p->se.on_rq)
1da177e4
LT
1592 goto out_running;
1593
1594 this_cpu = smp_processor_id();
1595 cpu = task_cpu(p);
1596 }
1597
e7693a36
GH
1598#ifdef CONFIG_SCHEDSTATS
1599 schedstat_inc(rq, ttwu_count);
1600 if (cpu == this_cpu)
1601 schedstat_inc(rq, ttwu_local);
1602 else {
1603 struct sched_domain *sd;
1604 for_each_domain(this_cpu, sd) {
1605 if (cpu_isset(cpu, sd->span)) {
1606 schedstat_inc(sd, ttwu_wake_remote);
1607 break;
1608 }
1609 }
1610 }
1611
1612#endif
1613
1614
1da177e4
LT
1615out_activate:
1616#endif /* CONFIG_SMP */
cc367732
IM
1617 schedstat_inc(p, se.nr_wakeups);
1618 if (sync)
1619 schedstat_inc(p, se.nr_wakeups_sync);
1620 if (orig_cpu != cpu)
1621 schedstat_inc(p, se.nr_wakeups_migrate);
1622 if (cpu == this_cpu)
1623 schedstat_inc(p, se.nr_wakeups_local);
1624 else
1625 schedstat_inc(p, se.nr_wakeups_remote);
2daa3577 1626 update_rq_clock(rq);
dd41f596 1627 activate_task(rq, p, 1);
9c63d9c0 1628 check_preempt_curr(rq, p);
1da177e4
LT
1629 success = 1;
1630
1631out_running:
1632 p->state = TASK_RUNNING;
4642dafd 1633 wakeup_balance_rt(rq, p);
1da177e4
LT
1634out:
1635 task_rq_unlock(rq, &flags);
1636
1637 return success;
1638}
1639
36c8b586 1640int fastcall wake_up_process(struct task_struct *p)
1da177e4
LT
1641{
1642 return try_to_wake_up(p, TASK_STOPPED | TASK_TRACED |
1643 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1644}
1da177e4
LT
1645EXPORT_SYMBOL(wake_up_process);
1646
36c8b586 1647int fastcall wake_up_state(struct task_struct *p, unsigned int state)
1da177e4
LT
1648{
1649 return try_to_wake_up(p, state, 0);
1650}
1651
1da177e4
LT
1652/*
1653 * Perform scheduler related setup for a newly forked process p.
1654 * p is forked by current.
dd41f596
IM
1655 *
1656 * __sched_fork() is basic setup used by init_idle() too:
1657 */
1658static void __sched_fork(struct task_struct *p)
1659{
dd41f596
IM
1660 p->se.exec_start = 0;
1661 p->se.sum_exec_runtime = 0;
f6cf891c 1662 p->se.prev_sum_exec_runtime = 0;
6cfb0d5d
IM
1663
1664#ifdef CONFIG_SCHEDSTATS
1665 p->se.wait_start = 0;
dd41f596
IM
1666 p->se.sum_sleep_runtime = 0;
1667 p->se.sleep_start = 0;
dd41f596
IM
1668 p->se.block_start = 0;
1669 p->se.sleep_max = 0;
1670 p->se.block_max = 0;
1671 p->se.exec_max = 0;
eba1ed4b 1672 p->se.slice_max = 0;
dd41f596 1673 p->se.wait_max = 0;
6cfb0d5d 1674#endif
476d139c 1675
dd41f596
IM
1676 INIT_LIST_HEAD(&p->run_list);
1677 p->se.on_rq = 0;
476d139c 1678
e107be36
AK
1679#ifdef CONFIG_PREEMPT_NOTIFIERS
1680 INIT_HLIST_HEAD(&p->preempt_notifiers);
1681#endif
1682
1da177e4
LT
1683 /*
1684 * We mark the process as running here, but have not actually
1685 * inserted it onto the runqueue yet. This guarantees that
1686 * nobody will actually run it, and a signal or other external
1687 * event cannot wake it up and insert it on the runqueue either.
1688 */
1689 p->state = TASK_RUNNING;
dd41f596
IM
1690}
1691
1692/*
1693 * fork()/clone()-time setup:
1694 */
1695void sched_fork(struct task_struct *p, int clone_flags)
1696{
1697 int cpu = get_cpu();
1698
1699 __sched_fork(p);
1700
1701#ifdef CONFIG_SMP
1702 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
1703#endif
02e4bac2 1704 set_task_cpu(p, cpu);
b29739f9
IM
1705
1706 /*
1707 * Make sure we do not leak PI boosting priority to the child:
1708 */
1709 p->prio = current->normal_prio;
2ddbf952
HS
1710 if (!rt_prio(p->prio))
1711 p->sched_class = &fair_sched_class;
b29739f9 1712
52f17b6c 1713#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
dd41f596 1714 if (likely(sched_info_on()))
52f17b6c 1715 memset(&p->sched_info, 0, sizeof(p->sched_info));
1da177e4 1716#endif
d6077cb8 1717#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4866cde0
NP
1718 p->oncpu = 0;
1719#endif
1da177e4 1720#ifdef CONFIG_PREEMPT
4866cde0 1721 /* Want to start with kernel preemption disabled. */
a1261f54 1722 task_thread_info(p)->preempt_count = 1;
1da177e4 1723#endif
476d139c 1724 put_cpu();
1da177e4
LT
1725}
1726
1727/*
1728 * wake_up_new_task - wake up a newly created task for the first time.
1729 *
1730 * This function will do some initial scheduler statistics housekeeping
1731 * that must be done for every newly created context, then puts the task
1732 * on the runqueue and wakes it.
1733 */
36c8b586 1734void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
1da177e4
LT
1735{
1736 unsigned long flags;
dd41f596 1737 struct rq *rq;
1da177e4
LT
1738
1739 rq = task_rq_lock(p, &flags);
147cbb4b 1740 BUG_ON(p->state != TASK_RUNNING);
a8e504d2 1741 update_rq_clock(rq);
1da177e4
LT
1742
1743 p->prio = effective_prio(p);
1744
b9dca1e0 1745 if (!p->sched_class->task_new || !current->se.on_rq) {
dd41f596 1746 activate_task(rq, p, 0);
1da177e4 1747 } else {
1da177e4 1748 /*
dd41f596
IM
1749 * Let the scheduling class do new task startup
1750 * management (if any):
1da177e4 1751 */
ee0827d8 1752 p->sched_class->task_new(rq, p);
e5fa2237 1753 inc_nr_running(p, rq);
1da177e4 1754 }
dd41f596 1755 check_preempt_curr(rq, p);
0d1311a5 1756 wakeup_balance_rt(rq, p);
dd41f596 1757 task_rq_unlock(rq, &flags);
1da177e4
LT
1758}
1759
e107be36
AK
1760#ifdef CONFIG_PREEMPT_NOTIFIERS
1761
1762/**
421cee29
RD
1763 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
1764 * @notifier: notifier struct to register
e107be36
AK
1765 */
1766void preempt_notifier_register(struct preempt_notifier *notifier)
1767{
1768 hlist_add_head(&notifier->link, &current->preempt_notifiers);
1769}
1770EXPORT_SYMBOL_GPL(preempt_notifier_register);
1771
1772/**
1773 * preempt_notifier_unregister - no longer interested in preemption notifications
421cee29 1774 * @notifier: notifier struct to unregister
e107be36
AK
1775 *
1776 * This is safe to call from within a preemption notifier.
1777 */
1778void preempt_notifier_unregister(struct preempt_notifier *notifier)
1779{
1780 hlist_del(&notifier->link);
1781}
1782EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1783
1784static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1785{
1786 struct preempt_notifier *notifier;
1787 struct hlist_node *node;
1788
1789 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1790 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1791}
1792
1793static void
1794fire_sched_out_preempt_notifiers(struct task_struct *curr,
1795 struct task_struct *next)
1796{
1797 struct preempt_notifier *notifier;
1798 struct hlist_node *node;
1799
1800 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1801 notifier->ops->sched_out(notifier, next);
1802}
1803
1804#else
1805
1806static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1807{
1808}
1809
1810static void
1811fire_sched_out_preempt_notifiers(struct task_struct *curr,
1812 struct task_struct *next)
1813{
1814}
1815
1816#endif
1817
4866cde0
NP
1818/**
1819 * prepare_task_switch - prepare to switch tasks
1820 * @rq: the runqueue preparing to switch
421cee29 1821 * @prev: the current task that is being switched out
4866cde0
NP
1822 * @next: the task we are going to switch to.
1823 *
1824 * This is called with the rq lock held and interrupts off. It must
1825 * be paired with a subsequent finish_task_switch after the context
1826 * switch.
1827 *
1828 * prepare_task_switch sets up locking and calls architecture specific
1829 * hooks.
1830 */
e107be36
AK
1831static inline void
1832prepare_task_switch(struct rq *rq, struct task_struct *prev,
1833 struct task_struct *next)
4866cde0 1834{
e107be36 1835 fire_sched_out_preempt_notifiers(prev, next);
4866cde0
NP
1836 prepare_lock_switch(rq, next);
1837 prepare_arch_switch(next);
1838}
1839
1da177e4
LT
1840/**
1841 * finish_task_switch - clean up after a task-switch
344babaa 1842 * @rq: runqueue associated with task-switch
1da177e4
LT
1843 * @prev: the thread we just switched away from.
1844 *
4866cde0
NP
1845 * finish_task_switch must be called after the context switch, paired
1846 * with a prepare_task_switch call before the context switch.
1847 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1848 * and do any other architecture-specific cleanup actions.
1da177e4
LT
1849 *
1850 * Note that we may have delayed dropping an mm in context_switch(). If
41a2d6cf 1851 * so, we finish that here outside of the runqueue lock. (Doing it
1da177e4
LT
1852 * with the lock held can cause deadlocks; see schedule() for
1853 * details.)
1854 */
a9957449 1855static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1da177e4
LT
1856 __releases(rq->lock)
1857{
1da177e4 1858 struct mm_struct *mm = rq->prev_mm;
55a101f8 1859 long prev_state;
1da177e4
LT
1860
1861 rq->prev_mm = NULL;
1862
1863 /*
1864 * A task struct has one reference for the use as "current".
c394cc9f 1865 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
55a101f8
ON
1866 * schedule one last time. The schedule call will never return, and
1867 * the scheduled task must drop that reference.
c394cc9f 1868 * The test for TASK_DEAD must occur while the runqueue locks are
1da177e4
LT
1869 * still held, otherwise prev could be scheduled on another cpu, die
1870 * there before we look at prev->state, and then the reference would
1871 * be dropped twice.
1872 * Manfred Spraul <manfred@colorfullife.com>
1873 */
55a101f8 1874 prev_state = prev->state;
4866cde0
NP
1875 finish_arch_switch(prev);
1876 finish_lock_switch(rq, prev);
e8fa1362
SR
1877 schedule_tail_balance_rt(rq);
1878
e107be36 1879 fire_sched_in_preempt_notifiers(current);
1da177e4
LT
1880 if (mm)
1881 mmdrop(mm);
c394cc9f 1882 if (unlikely(prev_state == TASK_DEAD)) {
c6fd91f0 1883 /*
1884 * Remove function-return probe instances associated with this
1885 * task and put them back on the free list.
9761eea8 1886 */
c6fd91f0 1887 kprobe_flush_task(prev);
1da177e4 1888 put_task_struct(prev);
c6fd91f0 1889 }
1da177e4
LT
1890}
1891
1892/**
1893 * schedule_tail - first thing a freshly forked thread must call.
1894 * @prev: the thread we just switched away from.
1895 */
36c8b586 1896asmlinkage void schedule_tail(struct task_struct *prev)
1da177e4
LT
1897 __releases(rq->lock)
1898{
70b97a7f
IM
1899 struct rq *rq = this_rq();
1900
4866cde0
NP
1901 finish_task_switch(rq, prev);
1902#ifdef __ARCH_WANT_UNLOCKED_CTXSW
1903 /* In this case, finish_task_switch does not reenable preemption */
1904 preempt_enable();
1905#endif
1da177e4 1906 if (current->set_child_tid)
b488893a 1907 put_user(task_pid_vnr(current), current->set_child_tid);
1da177e4
LT
1908}
1909
1910/*
1911 * context_switch - switch to the new MM and the new
1912 * thread's register state.
1913 */
dd41f596 1914static inline void
70b97a7f 1915context_switch(struct rq *rq, struct task_struct *prev,
36c8b586 1916 struct task_struct *next)
1da177e4 1917{
dd41f596 1918 struct mm_struct *mm, *oldmm;
1da177e4 1919
e107be36 1920 prepare_task_switch(rq, prev, next);
dd41f596
IM
1921 mm = next->mm;
1922 oldmm = prev->active_mm;
9226d125
ZA
1923 /*
1924 * For paravirt, this is coupled with an exit in switch_to to
1925 * combine the page table reload and the switch backend into
1926 * one hypercall.
1927 */
1928 arch_enter_lazy_cpu_mode();
1929
dd41f596 1930 if (unlikely(!mm)) {
1da177e4
LT
1931 next->active_mm = oldmm;
1932 atomic_inc(&oldmm->mm_count);
1933 enter_lazy_tlb(oldmm, next);
1934 } else
1935 switch_mm(oldmm, mm, next);
1936
dd41f596 1937 if (unlikely(!prev->mm)) {
1da177e4 1938 prev->active_mm = NULL;
1da177e4
LT
1939 rq->prev_mm = oldmm;
1940 }
3a5f5e48
IM
1941 /*
1942 * Since the runqueue lock will be released by the next
1943 * task (which is an invalid locking op but in the case
1944 * of the scheduler it's an obvious special-case), so we
1945 * do an early lockdep release here:
1946 */
1947#ifndef __ARCH_WANT_UNLOCKED_CTXSW
8a25d5de 1948 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3a5f5e48 1949#endif
1da177e4
LT
1950
1951 /* Here we just switch the register state and the stack. */
1952 switch_to(prev, next, prev);
1953
dd41f596
IM
1954 barrier();
1955 /*
1956 * this_rq must be evaluated again because prev may have moved
1957 * CPUs since it called schedule(), thus the 'rq' on its stack
1958 * frame will be invalid.
1959 */
1960 finish_task_switch(this_rq(), prev);
1da177e4
LT
1961}
1962
1963/*
1964 * nr_running, nr_uninterruptible and nr_context_switches:
1965 *
1966 * externally visible scheduler statistics: current number of runnable
1967 * threads, current number of uninterruptible-sleeping threads, total
1968 * number of context switches performed since bootup.
1969 */
1970unsigned long nr_running(void)
1971{
1972 unsigned long i, sum = 0;
1973
1974 for_each_online_cpu(i)
1975 sum += cpu_rq(i)->nr_running;
1976
1977 return sum;
1978}
1979
1980unsigned long nr_uninterruptible(void)
1981{
1982 unsigned long i, sum = 0;
1983
0a945022 1984 for_each_possible_cpu(i)
1da177e4
LT
1985 sum += cpu_rq(i)->nr_uninterruptible;
1986
1987 /*
1988 * Since we read the counters lockless, it might be slightly
1989 * inaccurate. Do not allow it to go below zero though:
1990 */
1991 if (unlikely((long)sum < 0))
1992 sum = 0;
1993
1994 return sum;
1995}
1996
1997unsigned long long nr_context_switches(void)
1998{
cc94abfc
SR
1999 int i;
2000 unsigned long long sum = 0;
1da177e4 2001
0a945022 2002 for_each_possible_cpu(i)
1da177e4
LT
2003 sum += cpu_rq(i)->nr_switches;
2004
2005 return sum;
2006}
2007
2008unsigned long nr_iowait(void)
2009{
2010 unsigned long i, sum = 0;
2011
0a945022 2012 for_each_possible_cpu(i)
1da177e4
LT
2013 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2014
2015 return sum;
2016}
2017
db1b1fef
JS
2018unsigned long nr_active(void)
2019{
2020 unsigned long i, running = 0, uninterruptible = 0;
2021
2022 for_each_online_cpu(i) {
2023 running += cpu_rq(i)->nr_running;
2024 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2025 }
2026
2027 if (unlikely((long)uninterruptible < 0))
2028 uninterruptible = 0;
2029
2030 return running + uninterruptible;
2031}
2032
48f24c4d 2033/*
dd41f596
IM
2034 * Update rq->cpu_load[] statistics. This function is usually called every
2035 * scheduler tick (TICK_NSEC).
48f24c4d 2036 */
dd41f596 2037static void update_cpu_load(struct rq *this_rq)
48f24c4d 2038{
495eca49 2039 unsigned long this_load = this_rq->load.weight;
dd41f596
IM
2040 int i, scale;
2041
2042 this_rq->nr_load_updates++;
dd41f596
IM
2043
2044 /* Update our load: */
2045 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2046 unsigned long old_load, new_load;
2047
2048 /* scale is effectively 1 << i now, and >> i divides by scale */
2049
2050 old_load = this_rq->cpu_load[i];
2051 new_load = this_load;
a25707f3
IM
2052 /*
2053 * Round up the averaging division if load is increasing. This
2054 * prevents us from getting stuck on 9 if the load is 10, for
2055 * example.
2056 */
2057 if (new_load > old_load)
2058 new_load += scale-1;
dd41f596
IM
2059 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2060 }
48f24c4d
IM
2061}
2062
dd41f596
IM
2063#ifdef CONFIG_SMP
2064
1da177e4
LT
2065/*
2066 * double_rq_lock - safely lock two runqueues
2067 *
2068 * Note this does not disable interrupts like task_rq_lock,
2069 * you need to do so manually before calling.
2070 */
70b97a7f 2071static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1da177e4
LT
2072 __acquires(rq1->lock)
2073 __acquires(rq2->lock)
2074{
054b9108 2075 BUG_ON(!irqs_disabled());
1da177e4
LT
2076 if (rq1 == rq2) {
2077 spin_lock(&rq1->lock);
2078 __acquire(rq2->lock); /* Fake it out ;) */
2079 } else {
c96d145e 2080 if (rq1 < rq2) {
1da177e4
LT
2081 spin_lock(&rq1->lock);
2082 spin_lock(&rq2->lock);
2083 } else {
2084 spin_lock(&rq2->lock);
2085 spin_lock(&rq1->lock);
2086 }
2087 }
6e82a3be
IM
2088 update_rq_clock(rq1);
2089 update_rq_clock(rq2);
1da177e4
LT
2090}
2091
2092/*
2093 * double_rq_unlock - safely unlock two runqueues
2094 *
2095 * Note this does not restore interrupts like task_rq_unlock,
2096 * you need to do so manually after calling.
2097 */
70b97a7f 2098static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1da177e4
LT
2099 __releases(rq1->lock)
2100 __releases(rq2->lock)
2101{
2102 spin_unlock(&rq1->lock);
2103 if (rq1 != rq2)
2104 spin_unlock(&rq2->lock);
2105 else
2106 __release(rq2->lock);
2107}
2108
2109/*
2110 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2111 */
e8fa1362 2112static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1da177e4
LT
2113 __releases(this_rq->lock)
2114 __acquires(busiest->lock)
2115 __acquires(this_rq->lock)
2116{
e8fa1362
SR
2117 int ret = 0;
2118
054b9108
KK
2119 if (unlikely(!irqs_disabled())) {
2120 /* printk() doesn't work good under rq->lock */
2121 spin_unlock(&this_rq->lock);
2122 BUG_ON(1);
2123 }
1da177e4 2124 if (unlikely(!spin_trylock(&busiest->lock))) {
c96d145e 2125 if (busiest < this_rq) {
1da177e4
LT
2126 spin_unlock(&this_rq->lock);
2127 spin_lock(&busiest->lock);
2128 spin_lock(&this_rq->lock);
e8fa1362 2129 ret = 1;
1da177e4
LT
2130 } else
2131 spin_lock(&busiest->lock);
2132 }
e8fa1362 2133 return ret;
1da177e4
LT
2134}
2135
1da177e4
LT
2136/*
2137 * If dest_cpu is allowed for this process, migrate the task to it.
2138 * This is accomplished by forcing the cpu_allowed mask to only
41a2d6cf 2139 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1da177e4
LT
2140 * the cpu_allowed mask is restored.
2141 */
36c8b586 2142static void sched_migrate_task(struct task_struct *p, int dest_cpu)
1da177e4 2143{
70b97a7f 2144 struct migration_req req;
1da177e4 2145 unsigned long flags;
70b97a7f 2146 struct rq *rq;
1da177e4
LT
2147
2148 rq = task_rq_lock(p, &flags);
2149 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2150 || unlikely(cpu_is_offline(dest_cpu)))
2151 goto out;
2152
2153 /* force the process onto the specified CPU */
2154 if (migrate_task(p, dest_cpu, &req)) {
2155 /* Need to wait for migration thread (might exit: take ref). */
2156 struct task_struct *mt = rq->migration_thread;
36c8b586 2157
1da177e4
LT
2158 get_task_struct(mt);
2159 task_rq_unlock(rq, &flags);
2160 wake_up_process(mt);
2161 put_task_struct(mt);
2162 wait_for_completion(&req.done);
36c8b586 2163
1da177e4
LT
2164 return;
2165 }
2166out:
2167 task_rq_unlock(rq, &flags);
2168}
2169
2170/*
476d139c
NP
2171 * sched_exec - execve() is a valuable balancing opportunity, because at
2172 * this point the task has the smallest effective memory and cache footprint.
1da177e4
LT
2173 */
2174void sched_exec(void)
2175{
1da177e4 2176 int new_cpu, this_cpu = get_cpu();
476d139c 2177 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
1da177e4 2178 put_cpu();
476d139c
NP
2179 if (new_cpu != this_cpu)
2180 sched_migrate_task(current, new_cpu);
1da177e4
LT
2181}
2182
2183/*
2184 * pull_task - move a task from a remote runqueue to the local runqueue.
2185 * Both runqueues must be locked.
2186 */
dd41f596
IM
2187static void pull_task(struct rq *src_rq, struct task_struct *p,
2188 struct rq *this_rq, int this_cpu)
1da177e4 2189{
2e1cb74a 2190 deactivate_task(src_rq, p, 0);
1da177e4 2191 set_task_cpu(p, this_cpu);
dd41f596 2192 activate_task(this_rq, p, 0);
1da177e4
LT
2193 /*
2194 * Note that idle threads have a prio of MAX_PRIO, for this test
2195 * to be always true for them.
2196 */
dd41f596 2197 check_preempt_curr(this_rq, p);
1da177e4
LT
2198}
2199
2200/*
2201 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2202 */
858119e1 2203static
70b97a7f 2204int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
d15bcfdb 2205 struct sched_domain *sd, enum cpu_idle_type idle,
95cdf3b7 2206 int *all_pinned)
1da177e4
LT
2207{
2208 /*
2209 * We do not migrate tasks that are:
2210 * 1) running (obviously), or
2211 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2212 * 3) are cache-hot on their current CPU.
2213 */
cc367732
IM
2214 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2215 schedstat_inc(p, se.nr_failed_migrations_affine);
1da177e4 2216 return 0;
cc367732 2217 }
81026794
NP
2218 *all_pinned = 0;
2219
cc367732
IM
2220 if (task_running(rq, p)) {
2221 schedstat_inc(p, se.nr_failed_migrations_running);
81026794 2222 return 0;
cc367732 2223 }
1da177e4 2224
da84d961
IM
2225 /*
2226 * Aggressive migration if:
2227 * 1) task is cache cold, or
2228 * 2) too many balance attempts have failed.
2229 */
2230
6bc1665b
IM
2231 if (!task_hot(p, rq->clock, sd) ||
2232 sd->nr_balance_failed > sd->cache_nice_tries) {
da84d961 2233#ifdef CONFIG_SCHEDSTATS
cc367732 2234 if (task_hot(p, rq->clock, sd)) {
da84d961 2235 schedstat_inc(sd, lb_hot_gained[idle]);
cc367732
IM
2236 schedstat_inc(p, se.nr_forced_migrations);
2237 }
da84d961
IM
2238#endif
2239 return 1;
2240 }
2241
cc367732
IM
2242 if (task_hot(p, rq->clock, sd)) {
2243 schedstat_inc(p, se.nr_failed_migrations_hot);
da84d961 2244 return 0;
cc367732 2245 }
1da177e4
LT
2246 return 1;
2247}
2248
e1d1484f
PW
2249static unsigned long
2250balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2251 unsigned long max_load_move, struct sched_domain *sd,
2252 enum cpu_idle_type idle, int *all_pinned,
2253 int *this_best_prio, struct rq_iterator *iterator)
1da177e4 2254{
b82d9fdd 2255 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
dd41f596
IM
2256 struct task_struct *p;
2257 long rem_load_move = max_load_move;
1da177e4 2258
e1d1484f 2259 if (max_load_move == 0)
1da177e4
LT
2260 goto out;
2261
81026794
NP
2262 pinned = 1;
2263
1da177e4 2264 /*
dd41f596 2265 * Start the load-balancing iterator:
1da177e4 2266 */
dd41f596
IM
2267 p = iterator->start(iterator->arg);
2268next:
b82d9fdd 2269 if (!p || loops++ > sysctl_sched_nr_migrate)
1da177e4 2270 goto out;
50ddd969 2271 /*
b82d9fdd 2272 * To help distribute high priority tasks across CPUs we don't
50ddd969
PW
2273 * skip a task if it will be the highest priority task (i.e. smallest
2274 * prio value) on its new queue regardless of its load weight
2275 */
dd41f596
IM
2276 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2277 SCHED_LOAD_SCALE_FUZZ;
a4ac01c3 2278 if ((skip_for_load && p->prio >= *this_best_prio) ||
dd41f596 2279 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
dd41f596
IM
2280 p = iterator->next(iterator->arg);
2281 goto next;
1da177e4
LT
2282 }
2283
dd41f596 2284 pull_task(busiest, p, this_rq, this_cpu);
1da177e4 2285 pulled++;
dd41f596 2286 rem_load_move -= p->se.load.weight;
1da177e4 2287
2dd73a4f 2288 /*
b82d9fdd 2289 * We only want to steal up to the prescribed amount of weighted load.
2dd73a4f 2290 */
e1d1484f 2291 if (rem_load_move > 0) {
a4ac01c3
PW
2292 if (p->prio < *this_best_prio)
2293 *this_best_prio = p->prio;
dd41f596
IM
2294 p = iterator->next(iterator->arg);
2295 goto next;
1da177e4
LT
2296 }
2297out:
2298 /*
e1d1484f 2299 * Right now, this is one of only two places pull_task() is called,
1da177e4
LT
2300 * so we can safely collect pull_task() stats here rather than
2301 * inside pull_task().
2302 */
2303 schedstat_add(sd, lb_gained[idle], pulled);
81026794
NP
2304
2305 if (all_pinned)
2306 *all_pinned = pinned;
e1d1484f
PW
2307
2308 return max_load_move - rem_load_move;
1da177e4
LT
2309}
2310
dd41f596 2311/*
43010659
PW
2312 * move_tasks tries to move up to max_load_move weighted load from busiest to
2313 * this_rq, as part of a balancing operation within domain "sd".
2314 * Returns 1 if successful and 0 otherwise.
dd41f596
IM
2315 *
2316 * Called with both runqueues locked.
2317 */
2318static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
43010659 2319 unsigned long max_load_move,
dd41f596
IM
2320 struct sched_domain *sd, enum cpu_idle_type idle,
2321 int *all_pinned)
2322{
5522d5d5 2323 const struct sched_class *class = sched_class_highest;
43010659 2324 unsigned long total_load_moved = 0;
a4ac01c3 2325 int this_best_prio = this_rq->curr->prio;
dd41f596
IM
2326
2327 do {
43010659
PW
2328 total_load_moved +=
2329 class->load_balance(this_rq, this_cpu, busiest,
e1d1484f 2330 max_load_move - total_load_moved,
a4ac01c3 2331 sd, idle, all_pinned, &this_best_prio);
dd41f596 2332 class = class->next;
43010659 2333 } while (class && max_load_move > total_load_moved);
dd41f596 2334
43010659
PW
2335 return total_load_moved > 0;
2336}
2337
e1d1484f
PW
2338static int
2339iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2340 struct sched_domain *sd, enum cpu_idle_type idle,
2341 struct rq_iterator *iterator)
2342{
2343 struct task_struct *p = iterator->start(iterator->arg);
2344 int pinned = 0;
2345
2346 while (p) {
2347 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2348 pull_task(busiest, p, this_rq, this_cpu);
2349 /*
2350 * Right now, this is only the second place pull_task()
2351 * is called, so we can safely collect pull_task()
2352 * stats here rather than inside pull_task().
2353 */
2354 schedstat_inc(sd, lb_gained[idle]);
2355
2356 return 1;
2357 }
2358 p = iterator->next(iterator->arg);
2359 }
2360
2361 return 0;
2362}
2363
43010659
PW
2364/*
2365 * move_one_task tries to move exactly one task from busiest to this_rq, as
2366 * part of active balancing operations within "domain".
2367 * Returns 1 if successful and 0 otherwise.
2368 *
2369 * Called with both runqueues locked.
2370 */
2371static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2372 struct sched_domain *sd, enum cpu_idle_type idle)
2373{
5522d5d5 2374 const struct sched_class *class;
43010659
PW
2375
2376 for (class = sched_class_highest; class; class = class->next)
e1d1484f 2377 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
43010659
PW
2378 return 1;
2379
2380 return 0;
dd41f596
IM
2381}
2382
1da177e4
LT
2383/*
2384 * find_busiest_group finds and returns the busiest CPU group within the
48f24c4d
IM
2385 * domain. It calculates and returns the amount of weighted load which
2386 * should be moved to restore balance via the imbalance parameter.
1da177e4
LT
2387 */
2388static struct sched_group *
2389find_busiest_group(struct sched_domain *sd, int this_cpu,
dd41f596
IM
2390 unsigned long *imbalance, enum cpu_idle_type idle,
2391 int *sd_idle, cpumask_t *cpus, int *balance)
1da177e4
LT
2392{
2393 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2394 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
0c117f1b 2395 unsigned long max_pull;
2dd73a4f
PW
2396 unsigned long busiest_load_per_task, busiest_nr_running;
2397 unsigned long this_load_per_task, this_nr_running;
908a7c1b 2398 int load_idx, group_imb = 0;
5c45bf27
SS
2399#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2400 int power_savings_balance = 1;
2401 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2402 unsigned long min_nr_running = ULONG_MAX;
2403 struct sched_group *group_min = NULL, *group_leader = NULL;
2404#endif
1da177e4
LT
2405
2406 max_load = this_load = total_load = total_pwr = 0;
2dd73a4f
PW
2407 busiest_load_per_task = busiest_nr_running = 0;
2408 this_load_per_task = this_nr_running = 0;
d15bcfdb 2409 if (idle == CPU_NOT_IDLE)
7897986b 2410 load_idx = sd->busy_idx;
d15bcfdb 2411 else if (idle == CPU_NEWLY_IDLE)
7897986b
NP
2412 load_idx = sd->newidle_idx;
2413 else
2414 load_idx = sd->idle_idx;
1da177e4
LT
2415
2416 do {
908a7c1b 2417 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
1da177e4
LT
2418 int local_group;
2419 int i;
908a7c1b 2420 int __group_imb = 0;
783609c6 2421 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2dd73a4f 2422 unsigned long sum_nr_running, sum_weighted_load;
1da177e4
LT
2423
2424 local_group = cpu_isset(this_cpu, group->cpumask);
2425
783609c6
SS
2426 if (local_group)
2427 balance_cpu = first_cpu(group->cpumask);
2428
1da177e4 2429 /* Tally up the load of all CPUs in the group */
2dd73a4f 2430 sum_weighted_load = sum_nr_running = avg_load = 0;
908a7c1b
KC
2431 max_cpu_load = 0;
2432 min_cpu_load = ~0UL;
1da177e4
LT
2433
2434 for_each_cpu_mask(i, group->cpumask) {
0a2966b4
CL
2435 struct rq *rq;
2436
2437 if (!cpu_isset(i, *cpus))
2438 continue;
2439
2440 rq = cpu_rq(i);
2dd73a4f 2441
9439aab8 2442 if (*sd_idle && rq->nr_running)
5969fe06
NP
2443 *sd_idle = 0;
2444
1da177e4 2445 /* Bias balancing toward cpus of our domain */
783609c6
SS
2446 if (local_group) {
2447 if (idle_cpu(i) && !first_idle_cpu) {
2448 first_idle_cpu = 1;
2449 balance_cpu = i;
2450 }
2451
a2000572 2452 load = target_load(i, load_idx);
908a7c1b 2453 } else {
a2000572 2454 load = source_load(i, load_idx);
908a7c1b
KC
2455 if (load > max_cpu_load)
2456 max_cpu_load = load;
2457 if (min_cpu_load > load)
2458 min_cpu_load = load;
2459 }
1da177e4
LT
2460
2461 avg_load += load;
2dd73a4f 2462 sum_nr_running += rq->nr_running;
dd41f596 2463 sum_weighted_load += weighted_cpuload(i);
1da177e4
LT
2464 }
2465
783609c6
SS
2466 /*
2467 * First idle cpu or the first cpu(busiest) in this sched group
2468 * is eligible for doing load balancing at this and above
9439aab8
SS
2469 * domains. In the newly idle case, we will allow all the cpu's
2470 * to do the newly idle load balance.
783609c6 2471 */
9439aab8
SS
2472 if (idle != CPU_NEWLY_IDLE && local_group &&
2473 balance_cpu != this_cpu && balance) {
783609c6
SS
2474 *balance = 0;
2475 goto ret;
2476 }
2477
1da177e4 2478 total_load += avg_load;
5517d86b 2479 total_pwr += group->__cpu_power;
1da177e4
LT
2480
2481 /* Adjust by relative CPU power of the group */
5517d86b
ED
2482 avg_load = sg_div_cpu_power(group,
2483 avg_load * SCHED_LOAD_SCALE);
1da177e4 2484
908a7c1b
KC
2485 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
2486 __group_imb = 1;
2487
5517d86b 2488 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
5c45bf27 2489
1da177e4
LT
2490 if (local_group) {
2491 this_load = avg_load;
2492 this = group;
2dd73a4f
PW
2493 this_nr_running = sum_nr_running;
2494 this_load_per_task = sum_weighted_load;
2495 } else if (avg_load > max_load &&
908a7c1b 2496 (sum_nr_running > group_capacity || __group_imb)) {
1da177e4
LT
2497 max_load = avg_load;
2498 busiest = group;
2dd73a4f
PW
2499 busiest_nr_running = sum_nr_running;
2500 busiest_load_per_task = sum_weighted_load;
908a7c1b 2501 group_imb = __group_imb;
1da177e4 2502 }
5c45bf27
SS
2503
2504#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2505 /*
2506 * Busy processors will not participate in power savings
2507 * balance.
2508 */
dd41f596
IM
2509 if (idle == CPU_NOT_IDLE ||
2510 !(sd->flags & SD_POWERSAVINGS_BALANCE))
2511 goto group_next;
5c45bf27
SS
2512
2513 /*
2514 * If the local group is idle or completely loaded
2515 * no need to do power savings balance at this domain
2516 */
2517 if (local_group && (this_nr_running >= group_capacity ||
2518 !this_nr_running))
2519 power_savings_balance = 0;
2520
dd41f596 2521 /*
5c45bf27
SS
2522 * If a group is already running at full capacity or idle,
2523 * don't include that group in power savings calculations
dd41f596
IM
2524 */
2525 if (!power_savings_balance || sum_nr_running >= group_capacity
5c45bf27 2526 || !sum_nr_running)
dd41f596 2527 goto group_next;
5c45bf27 2528
dd41f596 2529 /*
5c45bf27 2530 * Calculate the group which has the least non-idle load.
dd41f596
IM
2531 * This is the group from where we need to pick up the load
2532 * for saving power
2533 */
2534 if ((sum_nr_running < min_nr_running) ||
2535 (sum_nr_running == min_nr_running &&
5c45bf27
SS
2536 first_cpu(group->cpumask) <
2537 first_cpu(group_min->cpumask))) {
dd41f596
IM
2538 group_min = group;
2539 min_nr_running = sum_nr_running;
5c45bf27
SS
2540 min_load_per_task = sum_weighted_load /
2541 sum_nr_running;
dd41f596 2542 }
5c45bf27 2543
dd41f596 2544 /*
5c45bf27 2545 * Calculate the group which is almost near its
dd41f596
IM
2546 * capacity but still has some space to pick up some load
2547 * from other group and save more power
2548 */
2549 if (sum_nr_running <= group_capacity - 1) {
2550 if (sum_nr_running > leader_nr_running ||
2551 (sum_nr_running == leader_nr_running &&
2552 first_cpu(group->cpumask) >
2553 first_cpu(group_leader->cpumask))) {
2554 group_leader = group;
2555 leader_nr_running = sum_nr_running;
2556 }
48f24c4d 2557 }
5c45bf27
SS
2558group_next:
2559#endif
1da177e4
LT
2560 group = group->next;
2561 } while (group != sd->groups);
2562
2dd73a4f 2563 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
1da177e4
LT
2564 goto out_balanced;
2565
2566 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
2567
2568 if (this_load >= avg_load ||
2569 100*max_load <= sd->imbalance_pct*this_load)
2570 goto out_balanced;
2571
2dd73a4f 2572 busiest_load_per_task /= busiest_nr_running;
908a7c1b
KC
2573 if (group_imb)
2574 busiest_load_per_task = min(busiest_load_per_task, avg_load);
2575
1da177e4
LT
2576 /*
2577 * We're trying to get all the cpus to the average_load, so we don't
2578 * want to push ourselves above the average load, nor do we wish to
2579 * reduce the max loaded cpu below the average load, as either of these
2580 * actions would just result in more rebalancing later, and ping-pong
2581 * tasks around. Thus we look for the minimum possible imbalance.
2582 * Negative imbalances (*we* are more loaded than anyone else) will
2583 * be counted as no imbalance for these purposes -- we can't fix that
41a2d6cf 2584 * by pulling tasks to us. Be careful of negative numbers as they'll
1da177e4
LT
2585 * appear as very large values with unsigned longs.
2586 */
2dd73a4f
PW
2587 if (max_load <= busiest_load_per_task)
2588 goto out_balanced;
2589
2590 /*
2591 * In the presence of smp nice balancing, certain scenarios can have
2592 * max load less than avg load(as we skip the groups at or below
2593 * its cpu_power, while calculating max_load..)
2594 */
2595 if (max_load < avg_load) {
2596 *imbalance = 0;
2597 goto small_imbalance;
2598 }
0c117f1b
SS
2599
2600 /* Don't want to pull so many tasks that a group would go idle */
2dd73a4f 2601 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
0c117f1b 2602
1da177e4 2603 /* How much load to actually move to equalise the imbalance */
5517d86b
ED
2604 *imbalance = min(max_pull * busiest->__cpu_power,
2605 (avg_load - this_load) * this->__cpu_power)
1da177e4
LT
2606 / SCHED_LOAD_SCALE;
2607
2dd73a4f
PW
2608 /*
2609 * if *imbalance is less than the average load per runnable task
2610 * there is no gaurantee that any tasks will be moved so we'll have
2611 * a think about bumping its value to force at least one task to be
2612 * moved
2613 */
7fd0d2dd 2614 if (*imbalance < busiest_load_per_task) {
48f24c4d 2615 unsigned long tmp, pwr_now, pwr_move;
2dd73a4f
PW
2616 unsigned int imbn;
2617
2618small_imbalance:
2619 pwr_move = pwr_now = 0;
2620 imbn = 2;
2621 if (this_nr_running) {
2622 this_load_per_task /= this_nr_running;
2623 if (busiest_load_per_task > this_load_per_task)
2624 imbn = 1;
2625 } else
2626 this_load_per_task = SCHED_LOAD_SCALE;
1da177e4 2627
dd41f596
IM
2628 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
2629 busiest_load_per_task * imbn) {
2dd73a4f 2630 *imbalance = busiest_load_per_task;
1da177e4
LT
2631 return busiest;
2632 }
2633
2634 /*
2635 * OK, we don't have enough imbalance to justify moving tasks,
2636 * however we may be able to increase total CPU power used by
2637 * moving them.
2638 */
2639
5517d86b
ED
2640 pwr_now += busiest->__cpu_power *
2641 min(busiest_load_per_task, max_load);
2642 pwr_now += this->__cpu_power *
2643 min(this_load_per_task, this_load);
1da177e4
LT
2644 pwr_now /= SCHED_LOAD_SCALE;
2645
2646 /* Amount of load we'd subtract */
5517d86b
ED
2647 tmp = sg_div_cpu_power(busiest,
2648 busiest_load_per_task * SCHED_LOAD_SCALE);
1da177e4 2649 if (max_load > tmp)
5517d86b 2650 pwr_move += busiest->__cpu_power *
2dd73a4f 2651 min(busiest_load_per_task, max_load - tmp);
1da177e4
LT
2652
2653 /* Amount of load we'd add */
5517d86b 2654 if (max_load * busiest->__cpu_power <
33859f7f 2655 busiest_load_per_task * SCHED_LOAD_SCALE)
5517d86b
ED
2656 tmp = sg_div_cpu_power(this,
2657 max_load * busiest->__cpu_power);
1da177e4 2658 else
5517d86b
ED
2659 tmp = sg_div_cpu_power(this,
2660 busiest_load_per_task * SCHED_LOAD_SCALE);
2661 pwr_move += this->__cpu_power *
2662 min(this_load_per_task, this_load + tmp);
1da177e4
LT
2663 pwr_move /= SCHED_LOAD_SCALE;
2664
2665 /* Move if we gain throughput */
7fd0d2dd
SS
2666 if (pwr_move > pwr_now)
2667 *imbalance = busiest_load_per_task;
1da177e4
LT
2668 }
2669
1da177e4
LT
2670 return busiest;
2671
2672out_balanced:
5c45bf27 2673#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
d15bcfdb 2674 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
5c45bf27 2675 goto ret;
1da177e4 2676
5c45bf27
SS
2677 if (this == group_leader && group_leader != group_min) {
2678 *imbalance = min_load_per_task;
2679 return group_min;
2680 }
5c45bf27 2681#endif
783609c6 2682ret:
1da177e4
LT
2683 *imbalance = 0;
2684 return NULL;
2685}
2686
2687/*
2688 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2689 */
70b97a7f 2690static struct rq *
d15bcfdb 2691find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
0a2966b4 2692 unsigned long imbalance, cpumask_t *cpus)
1da177e4 2693{
70b97a7f 2694 struct rq *busiest = NULL, *rq;
2dd73a4f 2695 unsigned long max_load = 0;
1da177e4
LT
2696 int i;
2697
2698 for_each_cpu_mask(i, group->cpumask) {
dd41f596 2699 unsigned long wl;
0a2966b4
CL
2700
2701 if (!cpu_isset(i, *cpus))
2702 continue;
2703
48f24c4d 2704 rq = cpu_rq(i);
dd41f596 2705 wl = weighted_cpuload(i);
2dd73a4f 2706
dd41f596 2707 if (rq->nr_running == 1 && wl > imbalance)
2dd73a4f 2708 continue;
1da177e4 2709
dd41f596
IM
2710 if (wl > max_load) {
2711 max_load = wl;
48f24c4d 2712 busiest = rq;
1da177e4
LT
2713 }
2714 }
2715
2716 return busiest;
2717}
2718
77391d71
NP
2719/*
2720 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2721 * so long as it is large enough.
2722 */
2723#define MAX_PINNED_INTERVAL 512
2724
1da177e4
LT
2725/*
2726 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2727 * tasks if there is an imbalance.
1da177e4 2728 */
70b97a7f 2729static int load_balance(int this_cpu, struct rq *this_rq,
d15bcfdb 2730 struct sched_domain *sd, enum cpu_idle_type idle,
783609c6 2731 int *balance)
1da177e4 2732{
43010659 2733 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
1da177e4 2734 struct sched_group *group;
1da177e4 2735 unsigned long imbalance;
70b97a7f 2736 struct rq *busiest;
0a2966b4 2737 cpumask_t cpus = CPU_MASK_ALL;
fe2eea3f 2738 unsigned long flags;
5969fe06 2739
89c4710e
SS
2740 /*
2741 * When power savings policy is enabled for the parent domain, idle
2742 * sibling can pick up load irrespective of busy siblings. In this case,
dd41f596 2743 * let the state of idle sibling percolate up as CPU_IDLE, instead of
d15bcfdb 2744 * portraying it as CPU_NOT_IDLE.
89c4710e 2745 */
d15bcfdb 2746 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 2747 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 2748 sd_idle = 1;
1da177e4 2749
2d72376b 2750 schedstat_inc(sd, lb_count[idle]);
1da177e4 2751
0a2966b4
CL
2752redo:
2753 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
783609c6
SS
2754 &cpus, balance);
2755
06066714 2756 if (*balance == 0)
783609c6 2757 goto out_balanced;
783609c6 2758
1da177e4
LT
2759 if (!group) {
2760 schedstat_inc(sd, lb_nobusyg[idle]);
2761 goto out_balanced;
2762 }
2763
0a2966b4 2764 busiest = find_busiest_queue(group, idle, imbalance, &cpus);
1da177e4
LT
2765 if (!busiest) {
2766 schedstat_inc(sd, lb_nobusyq[idle]);
2767 goto out_balanced;
2768 }
2769
db935dbd 2770 BUG_ON(busiest == this_rq);
1da177e4
LT
2771
2772 schedstat_add(sd, lb_imbalance[idle], imbalance);
2773
43010659 2774 ld_moved = 0;
1da177e4
LT
2775 if (busiest->nr_running > 1) {
2776 /*
2777 * Attempt to move tasks. If find_busiest_group has found
2778 * an imbalance but busiest->nr_running <= 1, the group is
43010659 2779 * still unbalanced. ld_moved simply stays zero, so it is
1da177e4
LT
2780 * correctly treated as an imbalance.
2781 */
fe2eea3f 2782 local_irq_save(flags);
e17224bf 2783 double_rq_lock(this_rq, busiest);
43010659 2784 ld_moved = move_tasks(this_rq, this_cpu, busiest,
48f24c4d 2785 imbalance, sd, idle, &all_pinned);
e17224bf 2786 double_rq_unlock(this_rq, busiest);
fe2eea3f 2787 local_irq_restore(flags);
81026794 2788
46cb4b7c
SS
2789 /*
2790 * some other cpu did the load balance for us.
2791 */
43010659 2792 if (ld_moved && this_cpu != smp_processor_id())
46cb4b7c
SS
2793 resched_cpu(this_cpu);
2794
81026794 2795 /* All tasks on this runqueue were pinned by CPU affinity */
0a2966b4
CL
2796 if (unlikely(all_pinned)) {
2797 cpu_clear(cpu_of(busiest), cpus);
2798 if (!cpus_empty(cpus))
2799 goto redo;
81026794 2800 goto out_balanced;
0a2966b4 2801 }
1da177e4 2802 }
81026794 2803
43010659 2804 if (!ld_moved) {
1da177e4
LT
2805 schedstat_inc(sd, lb_failed[idle]);
2806 sd->nr_balance_failed++;
2807
2808 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
1da177e4 2809
fe2eea3f 2810 spin_lock_irqsave(&busiest->lock, flags);
fa3b6ddc
SS
2811
2812 /* don't kick the migration_thread, if the curr
2813 * task on busiest cpu can't be moved to this_cpu
2814 */
2815 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
fe2eea3f 2816 spin_unlock_irqrestore(&busiest->lock, flags);
fa3b6ddc
SS
2817 all_pinned = 1;
2818 goto out_one_pinned;
2819 }
2820
1da177e4
LT
2821 if (!busiest->active_balance) {
2822 busiest->active_balance = 1;
2823 busiest->push_cpu = this_cpu;
81026794 2824 active_balance = 1;
1da177e4 2825 }
fe2eea3f 2826 spin_unlock_irqrestore(&busiest->lock, flags);
81026794 2827 if (active_balance)
1da177e4
LT
2828 wake_up_process(busiest->migration_thread);
2829
2830 /*
2831 * We've kicked active balancing, reset the failure
2832 * counter.
2833 */
39507451 2834 sd->nr_balance_failed = sd->cache_nice_tries+1;
1da177e4 2835 }
81026794 2836 } else
1da177e4
LT
2837 sd->nr_balance_failed = 0;
2838
81026794 2839 if (likely(!active_balance)) {
1da177e4
LT
2840 /* We were unbalanced, so reset the balancing interval */
2841 sd->balance_interval = sd->min_interval;
81026794
NP
2842 } else {
2843 /*
2844 * If we've begun active balancing, start to back off. This
2845 * case may not be covered by the all_pinned logic if there
2846 * is only 1 task on the busy runqueue (because we don't call
2847 * move_tasks).
2848 */
2849 if (sd->balance_interval < sd->max_interval)
2850 sd->balance_interval *= 2;
1da177e4
LT
2851 }
2852
43010659 2853 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 2854 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 2855 return -1;
43010659 2856 return ld_moved;
1da177e4
LT
2857
2858out_balanced:
1da177e4
LT
2859 schedstat_inc(sd, lb_balanced[idle]);
2860
16cfb1c0 2861 sd->nr_balance_failed = 0;
fa3b6ddc
SS
2862
2863out_one_pinned:
1da177e4 2864 /* tune up the balancing interval */
77391d71
NP
2865 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2866 (sd->balance_interval < sd->max_interval))
1da177e4
LT
2867 sd->balance_interval *= 2;
2868
48f24c4d 2869 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 2870 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 2871 return -1;
1da177e4
LT
2872 return 0;
2873}
2874
2875/*
2876 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2877 * tasks if there is an imbalance.
2878 *
d15bcfdb 2879 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
1da177e4
LT
2880 * this_rq is locked.
2881 */
48f24c4d 2882static int
70b97a7f 2883load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
1da177e4
LT
2884{
2885 struct sched_group *group;
70b97a7f 2886 struct rq *busiest = NULL;
1da177e4 2887 unsigned long imbalance;
43010659 2888 int ld_moved = 0;
5969fe06 2889 int sd_idle = 0;
969bb4e4 2890 int all_pinned = 0;
0a2966b4 2891 cpumask_t cpus = CPU_MASK_ALL;
5969fe06 2892
89c4710e
SS
2893 /*
2894 * When power savings policy is enabled for the parent domain, idle
2895 * sibling can pick up load irrespective of busy siblings. In this case,
2896 * let the state of idle sibling percolate up as IDLE, instead of
d15bcfdb 2897 * portraying it as CPU_NOT_IDLE.
89c4710e
SS
2898 */
2899 if (sd->flags & SD_SHARE_CPUPOWER &&
2900 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 2901 sd_idle = 1;
1da177e4 2902
2d72376b 2903 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
0a2966b4 2904redo:
d15bcfdb 2905 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
783609c6 2906 &sd_idle, &cpus, NULL);
1da177e4 2907 if (!group) {
d15bcfdb 2908 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
16cfb1c0 2909 goto out_balanced;
1da177e4
LT
2910 }
2911
d15bcfdb 2912 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
0a2966b4 2913 &cpus);
db935dbd 2914 if (!busiest) {
d15bcfdb 2915 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
16cfb1c0 2916 goto out_balanced;
1da177e4
LT
2917 }
2918
db935dbd
NP
2919 BUG_ON(busiest == this_rq);
2920
d15bcfdb 2921 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
d6d5cfaf 2922
43010659 2923 ld_moved = 0;
d6d5cfaf
NP
2924 if (busiest->nr_running > 1) {
2925 /* Attempt to move tasks */
2926 double_lock_balance(this_rq, busiest);
6e82a3be
IM
2927 /* this_rq->clock is already updated */
2928 update_rq_clock(busiest);
43010659 2929 ld_moved = move_tasks(this_rq, this_cpu, busiest,
969bb4e4
SS
2930 imbalance, sd, CPU_NEWLY_IDLE,
2931 &all_pinned);
d6d5cfaf 2932 spin_unlock(&busiest->lock);
0a2966b4 2933
969bb4e4 2934 if (unlikely(all_pinned)) {
0a2966b4
CL
2935 cpu_clear(cpu_of(busiest), cpus);
2936 if (!cpus_empty(cpus))
2937 goto redo;
2938 }
d6d5cfaf
NP
2939 }
2940
43010659 2941 if (!ld_moved) {
d15bcfdb 2942 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
89c4710e
SS
2943 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2944 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06
NP
2945 return -1;
2946 } else
16cfb1c0 2947 sd->nr_balance_failed = 0;
1da177e4 2948
43010659 2949 return ld_moved;
16cfb1c0
NP
2950
2951out_balanced:
d15bcfdb 2952 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
48f24c4d 2953 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
89c4710e 2954 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
5969fe06 2955 return -1;
16cfb1c0 2956 sd->nr_balance_failed = 0;
48f24c4d 2957
16cfb1c0 2958 return 0;
1da177e4
LT
2959}
2960
2961/*
2962 * idle_balance is called by schedule() if this_cpu is about to become
2963 * idle. Attempts to pull tasks from other CPUs.
2964 */
70b97a7f 2965static void idle_balance(int this_cpu, struct rq *this_rq)
1da177e4
LT
2966{
2967 struct sched_domain *sd;
dd41f596
IM
2968 int pulled_task = -1;
2969 unsigned long next_balance = jiffies + HZ;
1da177e4
LT
2970
2971 for_each_domain(this_cpu, sd) {
92c4ca5c
CL
2972 unsigned long interval;
2973
2974 if (!(sd->flags & SD_LOAD_BALANCE))
2975 continue;
2976
2977 if (sd->flags & SD_BALANCE_NEWIDLE)
48f24c4d 2978 /* If we've pulled tasks over stop searching: */
1bd77f2d 2979 pulled_task = load_balance_newidle(this_cpu,
92c4ca5c
CL
2980 this_rq, sd);
2981
2982 interval = msecs_to_jiffies(sd->balance_interval);
2983 if (time_after(next_balance, sd->last_balance + interval))
2984 next_balance = sd->last_balance + interval;
2985 if (pulled_task)
2986 break;
1da177e4 2987 }
dd41f596 2988 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
1bd77f2d
CL
2989 /*
2990 * We are going idle. next_balance may be set based on
2991 * a busy processor. So reset next_balance.
2992 */
2993 this_rq->next_balance = next_balance;
dd41f596 2994 }
1da177e4
LT
2995}
2996
2997/*
2998 * active_load_balance is run by migration threads. It pushes running tasks
2999 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3000 * running on each physical CPU where possible, and avoids physical /
3001 * logical imbalances.
3002 *
3003 * Called with busiest_rq locked.
3004 */
70b97a7f 3005static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
1da177e4 3006{
39507451 3007 int target_cpu = busiest_rq->push_cpu;
70b97a7f
IM
3008 struct sched_domain *sd;
3009 struct rq *target_rq;
39507451 3010
48f24c4d 3011 /* Is there any task to move? */
39507451 3012 if (busiest_rq->nr_running <= 1)
39507451
NP
3013 return;
3014
3015 target_rq = cpu_rq(target_cpu);
1da177e4
LT
3016
3017 /*
39507451 3018 * This condition is "impossible", if it occurs
41a2d6cf 3019 * we need to fix it. Originally reported by
39507451 3020 * Bjorn Helgaas on a 128-cpu setup.
1da177e4 3021 */
39507451 3022 BUG_ON(busiest_rq == target_rq);
1da177e4 3023
39507451
NP
3024 /* move a task from busiest_rq to target_rq */
3025 double_lock_balance(busiest_rq, target_rq);
6e82a3be
IM
3026 update_rq_clock(busiest_rq);
3027 update_rq_clock(target_rq);
39507451
NP
3028
3029 /* Search for an sd spanning us and the target CPU. */
c96d145e 3030 for_each_domain(target_cpu, sd) {
39507451 3031 if ((sd->flags & SD_LOAD_BALANCE) &&
48f24c4d 3032 cpu_isset(busiest_cpu, sd->span))
39507451 3033 break;
c96d145e 3034 }
39507451 3035
48f24c4d 3036 if (likely(sd)) {
2d72376b 3037 schedstat_inc(sd, alb_count);
39507451 3038
43010659
PW
3039 if (move_one_task(target_rq, target_cpu, busiest_rq,
3040 sd, CPU_IDLE))
48f24c4d
IM
3041 schedstat_inc(sd, alb_pushed);
3042 else
3043 schedstat_inc(sd, alb_failed);
3044 }
39507451 3045 spin_unlock(&target_rq->lock);
1da177e4
LT
3046}
3047
46cb4b7c
SS
3048#ifdef CONFIG_NO_HZ
3049static struct {
3050 atomic_t load_balancer;
41a2d6cf 3051 cpumask_t cpu_mask;
46cb4b7c
SS
3052} nohz ____cacheline_aligned = {
3053 .load_balancer = ATOMIC_INIT(-1),
3054 .cpu_mask = CPU_MASK_NONE,
3055};
3056
7835b98b 3057/*
46cb4b7c
SS
3058 * This routine will try to nominate the ilb (idle load balancing)
3059 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3060 * load balancing on behalf of all those cpus. If all the cpus in the system
3061 * go into this tickless mode, then there will be no ilb owner (as there is
3062 * no need for one) and all the cpus will sleep till the next wakeup event
3063 * arrives...
3064 *
3065 * For the ilb owner, tick is not stopped. And this tick will be used
3066 * for idle load balancing. ilb owner will still be part of
3067 * nohz.cpu_mask..
7835b98b 3068 *
46cb4b7c
SS
3069 * While stopping the tick, this cpu will become the ilb owner if there
3070 * is no other owner. And will be the owner till that cpu becomes busy
3071 * or if all cpus in the system stop their ticks at which point
3072 * there is no need for ilb owner.
3073 *
3074 * When the ilb owner becomes busy, it nominates another owner, during the
3075 * next busy scheduler_tick()
3076 */
3077int select_nohz_load_balancer(int stop_tick)
3078{
3079 int cpu = smp_processor_id();
3080
3081 if (stop_tick) {
3082 cpu_set(cpu, nohz.cpu_mask);
3083 cpu_rq(cpu)->in_nohz_recently = 1;
3084
3085 /*
3086 * If we are going offline and still the leader, give up!
3087 */
3088 if (cpu_is_offline(cpu) &&
3089 atomic_read(&nohz.load_balancer) == cpu) {
3090 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3091 BUG();
3092 return 0;
3093 }
3094
3095 /* time for ilb owner also to sleep */
3096 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3097 if (atomic_read(&nohz.load_balancer) == cpu)
3098 atomic_set(&nohz.load_balancer, -1);
3099 return 0;
3100 }
3101
3102 if (atomic_read(&nohz.load_balancer) == -1) {
3103 /* make me the ilb owner */
3104 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3105 return 1;
3106 } else if (atomic_read(&nohz.load_balancer) == cpu)
3107 return 1;
3108 } else {
3109 if (!cpu_isset(cpu, nohz.cpu_mask))
3110 return 0;
3111
3112 cpu_clear(cpu, nohz.cpu_mask);
3113
3114 if (atomic_read(&nohz.load_balancer) == cpu)
3115 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3116 BUG();
3117 }
3118 return 0;
3119}
3120#endif
3121
3122static DEFINE_SPINLOCK(balancing);
3123
3124/*
7835b98b
CL
3125 * It checks each scheduling domain to see if it is due to be balanced,
3126 * and initiates a balancing operation if so.
3127 *
3128 * Balancing parameters are set up in arch_init_sched_domains.
3129 */
a9957449 3130static void rebalance_domains(int cpu, enum cpu_idle_type idle)
7835b98b 3131{
46cb4b7c
SS
3132 int balance = 1;
3133 struct rq *rq = cpu_rq(cpu);
7835b98b
CL
3134 unsigned long interval;
3135 struct sched_domain *sd;
46cb4b7c 3136 /* Earliest time when we have to do rebalance again */
c9819f45 3137 unsigned long next_balance = jiffies + 60*HZ;
f549da84 3138 int update_next_balance = 0;
1da177e4 3139
46cb4b7c 3140 for_each_domain(cpu, sd) {
1da177e4
LT
3141 if (!(sd->flags & SD_LOAD_BALANCE))
3142 continue;
3143
3144 interval = sd->balance_interval;
d15bcfdb 3145 if (idle != CPU_IDLE)
1da177e4
LT
3146 interval *= sd->busy_factor;
3147
3148 /* scale ms to jiffies */
3149 interval = msecs_to_jiffies(interval);
3150 if (unlikely(!interval))
3151 interval = 1;
dd41f596
IM
3152 if (interval > HZ*NR_CPUS/10)
3153 interval = HZ*NR_CPUS/10;
3154
1da177e4 3155
08c183f3
CL
3156 if (sd->flags & SD_SERIALIZE) {
3157 if (!spin_trylock(&balancing))
3158 goto out;
3159 }
3160
c9819f45 3161 if (time_after_eq(jiffies, sd->last_balance + interval)) {
46cb4b7c 3162 if (load_balance(cpu, rq, sd, idle, &balance)) {
fa3b6ddc
SS
3163 /*
3164 * We've pulled tasks over so either we're no
5969fe06
NP
3165 * longer idle, or one of our SMT siblings is
3166 * not idle.
3167 */
d15bcfdb 3168 idle = CPU_NOT_IDLE;
1da177e4 3169 }
1bd77f2d 3170 sd->last_balance = jiffies;
1da177e4 3171 }
08c183f3
CL
3172 if (sd->flags & SD_SERIALIZE)
3173 spin_unlock(&balancing);
3174out:
f549da84 3175 if (time_after(next_balance, sd->last_balance + interval)) {
c9819f45 3176 next_balance = sd->last_balance + interval;
f549da84
SS
3177 update_next_balance = 1;
3178 }
783609c6
SS
3179
3180 /*
3181 * Stop the load balance at this level. There is another
3182 * CPU in our sched group which is doing load balancing more
3183 * actively.
3184 */
3185 if (!balance)
3186 break;
1da177e4 3187 }
f549da84
SS
3188
3189 /*
3190 * next_balance will be updated only when there is a need.
3191 * When the cpu is attached to null domain for ex, it will not be
3192 * updated.
3193 */
3194 if (likely(update_next_balance))
3195 rq->next_balance = next_balance;
46cb4b7c
SS
3196}
3197
3198/*
3199 * run_rebalance_domains is triggered when needed from the scheduler tick.
3200 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3201 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3202 */
3203static void run_rebalance_domains(struct softirq_action *h)
3204{
dd41f596
IM
3205 int this_cpu = smp_processor_id();
3206 struct rq *this_rq = cpu_rq(this_cpu);
3207 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3208 CPU_IDLE : CPU_NOT_IDLE;
46cb4b7c 3209
dd41f596 3210 rebalance_domains(this_cpu, idle);
46cb4b7c
SS
3211
3212#ifdef CONFIG_NO_HZ
3213 /*
3214 * If this cpu is the owner for idle load balancing, then do the
3215 * balancing on behalf of the other idle cpus whose ticks are
3216 * stopped.
3217 */
dd41f596
IM
3218 if (this_rq->idle_at_tick &&
3219 atomic_read(&nohz.load_balancer) == this_cpu) {
46cb4b7c
SS
3220 cpumask_t cpus = nohz.cpu_mask;
3221 struct rq *rq;
3222 int balance_cpu;
3223
dd41f596 3224 cpu_clear(this_cpu, cpus);
46cb4b7c
SS
3225 for_each_cpu_mask(balance_cpu, cpus) {
3226 /*
3227 * If this cpu gets work to do, stop the load balancing
3228 * work being done for other cpus. Next load
3229 * balancing owner will pick it up.
3230 */
3231 if (need_resched())
3232 break;
3233
de0cf899 3234 rebalance_domains(balance_cpu, CPU_IDLE);
46cb4b7c
SS
3235
3236 rq = cpu_rq(balance_cpu);
dd41f596
IM
3237 if (time_after(this_rq->next_balance, rq->next_balance))
3238 this_rq->next_balance = rq->next_balance;
46cb4b7c
SS
3239 }
3240 }
3241#endif
3242}
3243
3244/*
3245 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3246 *
3247 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3248 * idle load balancing owner or decide to stop the periodic load balancing,
3249 * if the whole system is idle.
3250 */
dd41f596 3251static inline void trigger_load_balance(struct rq *rq, int cpu)
46cb4b7c 3252{
46cb4b7c
SS
3253#ifdef CONFIG_NO_HZ
3254 /*
3255 * If we were in the nohz mode recently and busy at the current
3256 * scheduler tick, then check if we need to nominate new idle
3257 * load balancer.
3258 */
3259 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3260 rq->in_nohz_recently = 0;
3261
3262 if (atomic_read(&nohz.load_balancer) == cpu) {
3263 cpu_clear(cpu, nohz.cpu_mask);
3264 atomic_set(&nohz.load_balancer, -1);
3265 }
3266
3267 if (atomic_read(&nohz.load_balancer) == -1) {
3268 /*
3269 * simple selection for now: Nominate the
3270 * first cpu in the nohz list to be the next
3271 * ilb owner.
3272 *
3273 * TBD: Traverse the sched domains and nominate
3274 * the nearest cpu in the nohz.cpu_mask.
3275 */
3276 int ilb = first_cpu(nohz.cpu_mask);
3277
3278 if (ilb != NR_CPUS)
3279 resched_cpu(ilb);
3280 }
3281 }
3282
3283 /*
3284 * If this cpu is idle and doing idle load balancing for all the
3285 * cpus with ticks stopped, is it time for that to stop?
3286 */
3287 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3288 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3289 resched_cpu(cpu);
3290 return;
3291 }
3292
3293 /*
3294 * If this cpu is idle and the idle load balancing is done by
3295 * someone else, then no need raise the SCHED_SOFTIRQ
3296 */
3297 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3298 cpu_isset(cpu, nohz.cpu_mask))
3299 return;
3300#endif
3301 if (time_after_eq(jiffies, rq->next_balance))
3302 raise_softirq(SCHED_SOFTIRQ);
1da177e4 3303}
dd41f596
IM
3304
3305#else /* CONFIG_SMP */
3306
1da177e4
LT
3307/*
3308 * on UP we do not need to balance between CPUs:
3309 */
70b97a7f 3310static inline void idle_balance(int cpu, struct rq *rq)
1da177e4
LT
3311{
3312}
dd41f596 3313
1da177e4
LT
3314#endif
3315
1da177e4
LT
3316DEFINE_PER_CPU(struct kernel_stat, kstat);
3317
3318EXPORT_PER_CPU_SYMBOL(kstat);
3319
3320/*
41b86e9c
IM
3321 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3322 * that have not yet been banked in case the task is currently running.
1da177e4 3323 */
41b86e9c 3324unsigned long long task_sched_runtime(struct task_struct *p)
1da177e4 3325{
1da177e4 3326 unsigned long flags;
41b86e9c
IM
3327 u64 ns, delta_exec;
3328 struct rq *rq;
48f24c4d 3329
41b86e9c
IM
3330 rq = task_rq_lock(p, &flags);
3331 ns = p->se.sum_exec_runtime;
051a1d1a 3332 if (task_current(rq, p)) {
a8e504d2
IM
3333 update_rq_clock(rq);
3334 delta_exec = rq->clock - p->se.exec_start;
41b86e9c
IM
3335 if ((s64)delta_exec > 0)
3336 ns += delta_exec;
3337 }
3338 task_rq_unlock(rq, &flags);
48f24c4d 3339
1da177e4
LT
3340 return ns;
3341}
3342
1da177e4
LT
3343/*
3344 * Account user cpu time to a process.
3345 * @p: the process that the cpu time gets accounted to
1da177e4
LT
3346 * @cputime: the cpu time spent in user space since the last update
3347 */
3348void account_user_time(struct task_struct *p, cputime_t cputime)
3349{
3350 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3351 cputime64_t tmp;
3352
3353 p->utime = cputime_add(p->utime, cputime);
3354
3355 /* Add user time to cpustat. */
3356 tmp = cputime_to_cputime64(cputime);
3357 if (TASK_NICE(p) > 0)
3358 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3359 else
3360 cpustat->user = cputime64_add(cpustat->user, tmp);
3361}
3362
94886b84
LV
3363/*
3364 * Account guest cpu time to a process.
3365 * @p: the process that the cpu time gets accounted to
3366 * @cputime: the cpu time spent in virtual machine since the last update
3367 */
f7402e03 3368static void account_guest_time(struct task_struct *p, cputime_t cputime)
94886b84
LV
3369{
3370 cputime64_t tmp;
3371 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3372
3373 tmp = cputime_to_cputime64(cputime);
3374
3375 p->utime = cputime_add(p->utime, cputime);
3376 p->gtime = cputime_add(p->gtime, cputime);
3377
3378 cpustat->user = cputime64_add(cpustat->user, tmp);
3379 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3380}
3381
c66f08be
MN
3382/*
3383 * Account scaled user cpu time to a process.
3384 * @p: the process that the cpu time gets accounted to
3385 * @cputime: the cpu time spent in user space since the last update
3386 */
3387void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3388{
3389 p->utimescaled = cputime_add(p->utimescaled, cputime);
3390}
3391
1da177e4
LT
3392/*
3393 * Account system cpu time to a process.
3394 * @p: the process that the cpu time gets accounted to
3395 * @hardirq_offset: the offset to subtract from hardirq_count()
3396 * @cputime: the cpu time spent in kernel space since the last update
3397 */
3398void account_system_time(struct task_struct *p, int hardirq_offset,
3399 cputime_t cputime)
3400{
3401 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
70b97a7f 3402 struct rq *rq = this_rq();
1da177e4
LT
3403 cputime64_t tmp;
3404
9778385d
CB
3405 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
3406 return account_guest_time(p, cputime);
94886b84 3407
1da177e4
LT
3408 p->stime = cputime_add(p->stime, cputime);
3409
3410 /* Add system time to cpustat. */
3411 tmp = cputime_to_cputime64(cputime);
3412 if (hardirq_count() - hardirq_offset)
3413 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3414 else if (softirq_count())
3415 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
cfb52856 3416 else if (p != rq->idle)
1da177e4 3417 cpustat->system = cputime64_add(cpustat->system, tmp);
cfb52856 3418 else if (atomic_read(&rq->nr_iowait) > 0)
1da177e4
LT
3419 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3420 else
3421 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3422 /* Account for system time used */
3423 acct_update_integrals(p);
1da177e4
LT
3424}
3425
c66f08be
MN
3426/*
3427 * Account scaled system cpu time to a process.
3428 * @p: the process that the cpu time gets accounted to
3429 * @hardirq_offset: the offset to subtract from hardirq_count()
3430 * @cputime: the cpu time spent in kernel space since the last update
3431 */
3432void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3433{
3434 p->stimescaled = cputime_add(p->stimescaled, cputime);
3435}
3436
1da177e4
LT
3437/*
3438 * Account for involuntary wait time.
3439 * @p: the process from which the cpu time has been stolen
3440 * @steal: the cpu time spent in involuntary wait
3441 */
3442void account_steal_time(struct task_struct *p, cputime_t steal)
3443{
3444 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3445 cputime64_t tmp = cputime_to_cputime64(steal);
70b97a7f 3446 struct rq *rq = this_rq();
1da177e4
LT
3447
3448 if (p == rq->idle) {
3449 p->stime = cputime_add(p->stime, steal);
3450 if (atomic_read(&rq->nr_iowait) > 0)
3451 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3452 else
3453 cpustat->idle = cputime64_add(cpustat->idle, tmp);
cfb52856 3454 } else
1da177e4
LT
3455 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3456}
3457
7835b98b
CL
3458/*
3459 * This function gets called by the timer code, with HZ frequency.
3460 * We call it with interrupts disabled.
3461 *
3462 * It also gets called by the fork code, when changing the parent's
3463 * timeslices.
3464 */
3465void scheduler_tick(void)
3466{
7835b98b
CL
3467 int cpu = smp_processor_id();
3468 struct rq *rq = cpu_rq(cpu);
dd41f596 3469 struct task_struct *curr = rq->curr;
529c7726 3470 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
dd41f596
IM
3471
3472 spin_lock(&rq->lock);
546fe3c9 3473 __update_rq_clock(rq);
529c7726
IM
3474 /*
3475 * Let rq->clock advance by at least TICK_NSEC:
3476 */
3477 if (unlikely(rq->clock < next_tick))
3478 rq->clock = next_tick;
3479 rq->tick_timestamp = rq->clock;
f1a438d8 3480 update_cpu_load(rq);
dd41f596
IM
3481 if (curr != rq->idle) /* FIXME: needed? */
3482 curr->sched_class->task_tick(rq, curr);
dd41f596 3483 spin_unlock(&rq->lock);
7835b98b 3484
e418e1c2 3485#ifdef CONFIG_SMP
dd41f596
IM
3486 rq->idle_at_tick = idle_cpu(cpu);
3487 trigger_load_balance(rq, cpu);
e418e1c2 3488#endif
1da177e4
LT
3489}
3490
1da177e4
LT
3491#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3492
3493void fastcall add_preempt_count(int val)
3494{
3495 /*
3496 * Underflow?
3497 */
9a11b49a
IM
3498 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3499 return;
1da177e4
LT
3500 preempt_count() += val;
3501 /*
3502 * Spinlock count overflowing soon?
3503 */
33859f7f
MOS
3504 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3505 PREEMPT_MASK - 10);
1da177e4
LT
3506}
3507EXPORT_SYMBOL(add_preempt_count);
3508
3509void fastcall sub_preempt_count(int val)
3510{
3511 /*
3512 * Underflow?
3513 */
9a11b49a
IM
3514 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3515 return;
1da177e4
LT
3516 /*
3517 * Is the spinlock portion underflowing?
3518 */
9a11b49a
IM
3519 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3520 !(preempt_count() & PREEMPT_MASK)))
3521 return;
3522
1da177e4
LT
3523 preempt_count() -= val;
3524}
3525EXPORT_SYMBOL(sub_preempt_count);
3526
3527#endif
3528
3529/*
dd41f596 3530 * Print scheduling while atomic bug:
1da177e4 3531 */
dd41f596 3532static noinline void __schedule_bug(struct task_struct *prev)
1da177e4 3533{
838225b4
SS
3534 struct pt_regs *regs = get_irq_regs();
3535
3536 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3537 prev->comm, prev->pid, preempt_count());
3538
dd41f596
IM
3539 debug_show_held_locks(prev);
3540 if (irqs_disabled())
3541 print_irqtrace_events(prev);
838225b4
SS
3542
3543 if (regs)
3544 show_regs(regs);
3545 else
3546 dump_stack();
dd41f596 3547}
1da177e4 3548
dd41f596
IM
3549/*
3550 * Various schedule()-time debugging checks and statistics:
3551 */
3552static inline void schedule_debug(struct task_struct *prev)
3553{
1da177e4 3554 /*
41a2d6cf 3555 * Test if we are atomic. Since do_exit() needs to call into
1da177e4
LT
3556 * schedule() atomically, we ignore that path for now.
3557 * Otherwise, whine if we are scheduling when we should not be.
3558 */
dd41f596
IM
3559 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
3560 __schedule_bug(prev);
3561
1da177e4
LT
3562 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3563
2d72376b 3564 schedstat_inc(this_rq(), sched_count);
b8efb561
IM
3565#ifdef CONFIG_SCHEDSTATS
3566 if (unlikely(prev->lock_depth >= 0)) {
2d72376b
IM
3567 schedstat_inc(this_rq(), bkl_count);
3568 schedstat_inc(prev, sched_info.bkl_count);
b8efb561
IM
3569 }
3570#endif
dd41f596
IM
3571}
3572
3573/*
3574 * Pick up the highest-prio task:
3575 */
3576static inline struct task_struct *
ff95f3df 3577pick_next_task(struct rq *rq, struct task_struct *prev)
dd41f596 3578{
5522d5d5 3579 const struct sched_class *class;
dd41f596 3580 struct task_struct *p;
1da177e4
LT
3581
3582 /*
dd41f596
IM
3583 * Optimization: we know that if all tasks are in
3584 * the fair class we can call that function directly:
1da177e4 3585 */
dd41f596 3586 if (likely(rq->nr_running == rq->cfs.nr_running)) {
fb8d4724 3587 p = fair_sched_class.pick_next_task(rq);
dd41f596
IM
3588 if (likely(p))
3589 return p;
1da177e4
LT
3590 }
3591
dd41f596
IM
3592 class = sched_class_highest;
3593 for ( ; ; ) {
fb8d4724 3594 p = class->pick_next_task(rq);
dd41f596
IM
3595 if (p)
3596 return p;
3597 /*
3598 * Will never be NULL as the idle class always
3599 * returns a non-NULL p:
3600 */
3601 class = class->next;
3602 }
3603}
1da177e4 3604
dd41f596
IM
3605/*
3606 * schedule() is the main scheduler function.
3607 */
3608asmlinkage void __sched schedule(void)
3609{
3610 struct task_struct *prev, *next;
3611 long *switch_count;
3612 struct rq *rq;
dd41f596
IM
3613 int cpu;
3614
3615need_resched:
3616 preempt_disable();
3617 cpu = smp_processor_id();
3618 rq = cpu_rq(cpu);
3619 rcu_qsctr_inc(cpu);
3620 prev = rq->curr;
3621 switch_count = &prev->nivcsw;
3622
3623 release_kernel_lock(prev);
3624need_resched_nonpreemptible:
3625
3626 schedule_debug(prev);
1da177e4 3627
1e819950
IM
3628 /*
3629 * Do the rq-clock update outside the rq lock:
3630 */
3631 local_irq_disable();
c1b3da3e 3632 __update_rq_clock(rq);
1e819950
IM
3633 spin_lock(&rq->lock);
3634 clear_tsk_need_resched(prev);
1da177e4 3635
1da177e4 3636 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
1da177e4 3637 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
dd41f596 3638 unlikely(signal_pending(prev)))) {
1da177e4 3639 prev->state = TASK_RUNNING;
dd41f596 3640 } else {
2e1cb74a 3641 deactivate_task(rq, prev, 1);
1da177e4 3642 }
dd41f596 3643 switch_count = &prev->nvcsw;
1da177e4
LT
3644 }
3645
f65eda4f
SR
3646 schedule_balance_rt(rq, prev);
3647
dd41f596 3648 if (unlikely(!rq->nr_running))
1da177e4 3649 idle_balance(cpu, rq);
1da177e4 3650
31ee529c 3651 prev->sched_class->put_prev_task(rq, prev);
ff95f3df 3652 next = pick_next_task(rq, prev);
1da177e4
LT
3653
3654 sched_info_switch(prev, next);
dd41f596 3655
1da177e4 3656 if (likely(prev != next)) {
1da177e4
LT
3657 rq->nr_switches++;
3658 rq->curr = next;
3659 ++*switch_count;
3660
dd41f596 3661 context_switch(rq, prev, next); /* unlocks the rq */
1da177e4
LT
3662 } else
3663 spin_unlock_irq(&rq->lock);
3664
dd41f596
IM
3665 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3666 cpu = smp_processor_id();
3667 rq = cpu_rq(cpu);
1da177e4 3668 goto need_resched_nonpreemptible;
dd41f596 3669 }
1da177e4
LT
3670 preempt_enable_no_resched();
3671 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
3672 goto need_resched;
3673}
1da177e4
LT
3674EXPORT_SYMBOL(schedule);
3675
3676#ifdef CONFIG_PREEMPT
3677/*
2ed6e34f 3678 * this is the entry point to schedule() from in-kernel preemption
41a2d6cf 3679 * off of preempt_enable. Kernel preemptions off return from interrupt
1da177e4
LT
3680 * occur there and call schedule directly.
3681 */
3682asmlinkage void __sched preempt_schedule(void)
3683{
3684 struct thread_info *ti = current_thread_info();
3685#ifdef CONFIG_PREEMPT_BKL
3686 struct task_struct *task = current;
3687 int saved_lock_depth;
3688#endif
3689 /*
3690 * If there is a non-zero preempt_count or interrupts are disabled,
41a2d6cf 3691 * we do not want to preempt the current task. Just return..
1da177e4 3692 */
beed33a8 3693 if (likely(ti->preempt_count || irqs_disabled()))
1da177e4
LT
3694 return;
3695
3a5c359a
AK
3696 do {
3697 add_preempt_count(PREEMPT_ACTIVE);
3698
3699 /*
3700 * We keep the big kernel semaphore locked, but we
3701 * clear ->lock_depth so that schedule() doesnt
3702 * auto-release the semaphore:
3703 */
1da177e4 3704#ifdef CONFIG_PREEMPT_BKL
3a5c359a
AK
3705 saved_lock_depth = task->lock_depth;
3706 task->lock_depth = -1;
1da177e4 3707#endif
3a5c359a 3708 schedule();
1da177e4 3709#ifdef CONFIG_PREEMPT_BKL
3a5c359a 3710 task->lock_depth = saved_lock_depth;
1da177e4 3711#endif
3a5c359a 3712 sub_preempt_count(PREEMPT_ACTIVE);
1da177e4 3713
3a5c359a
AK
3714 /*
3715 * Check again in case we missed a preemption opportunity
3716 * between schedule and now.
3717 */
3718 barrier();
3719 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
1da177e4 3720}
1da177e4
LT
3721EXPORT_SYMBOL(preempt_schedule);
3722
3723/*
2ed6e34f 3724 * this is the entry point to schedule() from kernel preemption
1da177e4
LT
3725 * off of irq context.
3726 * Note, that this is called and return with irqs disabled. This will
3727 * protect us against recursive calling from irq.
3728 */
3729asmlinkage void __sched preempt_schedule_irq(void)
3730{
3731 struct thread_info *ti = current_thread_info();
3732#ifdef CONFIG_PREEMPT_BKL
3733 struct task_struct *task = current;
3734 int saved_lock_depth;
3735#endif
2ed6e34f 3736 /* Catch callers which need to be fixed */
1da177e4
LT
3737 BUG_ON(ti->preempt_count || !irqs_disabled());
3738
3a5c359a
AK
3739 do {
3740 add_preempt_count(PREEMPT_ACTIVE);
3741
3742 /*
3743 * We keep the big kernel semaphore locked, but we
3744 * clear ->lock_depth so that schedule() doesnt
3745 * auto-release the semaphore:
3746 */
1da177e4 3747#ifdef CONFIG_PREEMPT_BKL
3a5c359a
AK
3748 saved_lock_depth = task->lock_depth;
3749 task->lock_depth = -1;
1da177e4 3750#endif
3a5c359a
AK
3751 local_irq_enable();
3752 schedule();
3753 local_irq_disable();
1da177e4 3754#ifdef CONFIG_PREEMPT_BKL
3a5c359a 3755 task->lock_depth = saved_lock_depth;
1da177e4 3756#endif
3a5c359a 3757 sub_preempt_count(PREEMPT_ACTIVE);
1da177e4 3758
3a5c359a
AK
3759 /*
3760 * Check again in case we missed a preemption opportunity
3761 * between schedule and now.
3762 */
3763 barrier();
3764 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
1da177e4
LT
3765}
3766
3767#endif /* CONFIG_PREEMPT */
3768
95cdf3b7
IM
3769int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
3770 void *key)
1da177e4 3771{
48f24c4d 3772 return try_to_wake_up(curr->private, mode, sync);
1da177e4 3773}
1da177e4
LT
3774EXPORT_SYMBOL(default_wake_function);
3775
3776/*
41a2d6cf
IM
3777 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3778 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
1da177e4
LT
3779 * number) then we wake all the non-exclusive tasks and one exclusive task.
3780 *
3781 * There are circumstances in which we can try to wake a task which has already
41a2d6cf 3782 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
1da177e4
LT
3783 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3784 */
3785static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3786 int nr_exclusive, int sync, void *key)
3787{
2e45874c 3788 wait_queue_t *curr, *next;
1da177e4 3789
2e45874c 3790 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
48f24c4d
IM
3791 unsigned flags = curr->flags;
3792
1da177e4 3793 if (curr->func(curr, mode, sync, key) &&
48f24c4d 3794 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
1da177e4
LT
3795 break;
3796 }
3797}
3798
3799/**
3800 * __wake_up - wake up threads blocked on a waitqueue.
3801 * @q: the waitqueue
3802 * @mode: which threads
3803 * @nr_exclusive: how many wake-one or wake-many threads to wake up
67be2dd1 3804 * @key: is directly passed to the wakeup function
1da177e4
LT
3805 */
3806void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
95cdf3b7 3807 int nr_exclusive, void *key)
1da177e4
LT
3808{
3809 unsigned long flags;
3810
3811 spin_lock_irqsave(&q->lock, flags);
3812 __wake_up_common(q, mode, nr_exclusive, 0, key);
3813 spin_unlock_irqrestore(&q->lock, flags);
3814}
1da177e4
LT
3815EXPORT_SYMBOL(__wake_up);
3816
3817/*
3818 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3819 */
3820void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3821{
3822 __wake_up_common(q, mode, 1, 0, NULL);
3823}
3824
3825/**
67be2dd1 3826 * __wake_up_sync - wake up threads blocked on a waitqueue.
1da177e4
LT
3827 * @q: the waitqueue
3828 * @mode: which threads
3829 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3830 *
3831 * The sync wakeup differs that the waker knows that it will schedule
3832 * away soon, so while the target thread will be woken up, it will not
3833 * be migrated to another CPU - ie. the two threads are 'synchronized'
3834 * with each other. This can prevent needless bouncing between CPUs.
3835 *
3836 * On UP it can prevent extra preemption.
3837 */
95cdf3b7
IM
3838void fastcall
3839__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
1da177e4
LT
3840{
3841 unsigned long flags;
3842 int sync = 1;
3843
3844 if (unlikely(!q))
3845 return;
3846
3847 if (unlikely(!nr_exclusive))
3848 sync = 0;
3849
3850 spin_lock_irqsave(&q->lock, flags);
3851 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
3852 spin_unlock_irqrestore(&q->lock, flags);
3853}
3854EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3855
b15136e9 3856void complete(struct completion *x)
1da177e4
LT
3857{
3858 unsigned long flags;
3859
3860 spin_lock_irqsave(&x->wait.lock, flags);
3861 x->done++;
3862 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3863 1, 0, NULL);
3864 spin_unlock_irqrestore(&x->wait.lock, flags);
3865}
3866EXPORT_SYMBOL(complete);
3867
b15136e9 3868void complete_all(struct completion *x)
1da177e4
LT
3869{
3870 unsigned long flags;
3871
3872 spin_lock_irqsave(&x->wait.lock, flags);
3873 x->done += UINT_MAX/2;
3874 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
3875 0, 0, NULL);
3876 spin_unlock_irqrestore(&x->wait.lock, flags);
3877}
3878EXPORT_SYMBOL(complete_all);
3879
8cbbe86d
AK
3880static inline long __sched
3881do_wait_for_common(struct completion *x, long timeout, int state)
1da177e4 3882{
1da177e4
LT
3883 if (!x->done) {
3884 DECLARE_WAITQUEUE(wait, current);
3885
3886 wait.flags |= WQ_FLAG_EXCLUSIVE;
3887 __add_wait_queue_tail(&x->wait, &wait);
3888 do {
8cbbe86d
AK
3889 if (state == TASK_INTERRUPTIBLE &&
3890 signal_pending(current)) {
3891 __remove_wait_queue(&x->wait, &wait);
3892 return -ERESTARTSYS;
3893 }
3894 __set_current_state(state);
1da177e4
LT
3895 spin_unlock_irq(&x->wait.lock);
3896 timeout = schedule_timeout(timeout);
3897 spin_lock_irq(&x->wait.lock);
3898 if (!timeout) {
3899 __remove_wait_queue(&x->wait, &wait);
8cbbe86d 3900 return timeout;
1da177e4
LT
3901 }
3902 } while (!x->done);
3903 __remove_wait_queue(&x->wait, &wait);
3904 }
3905 x->done--;
1da177e4
LT
3906 return timeout;
3907}
1da177e4 3908
8cbbe86d
AK
3909static long __sched
3910wait_for_common(struct completion *x, long timeout, int state)
1da177e4 3911{
1da177e4
LT
3912 might_sleep();
3913
3914 spin_lock_irq(&x->wait.lock);
8cbbe86d 3915 timeout = do_wait_for_common(x, timeout, state);
1da177e4 3916 spin_unlock_irq(&x->wait.lock);
8cbbe86d
AK
3917 return timeout;
3918}
1da177e4 3919
b15136e9 3920void __sched wait_for_completion(struct completion *x)
8cbbe86d
AK
3921{
3922 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
1da177e4 3923}
8cbbe86d 3924EXPORT_SYMBOL(wait_for_completion);
1da177e4 3925
b15136e9 3926unsigned long __sched
8cbbe86d 3927wait_for_completion_timeout(struct completion *x, unsigned long timeout)
1da177e4 3928{
8cbbe86d 3929 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
1da177e4 3930}
8cbbe86d 3931EXPORT_SYMBOL(wait_for_completion_timeout);
1da177e4 3932
8cbbe86d 3933int __sched wait_for_completion_interruptible(struct completion *x)
0fec171c 3934{
51e97990
AK
3935 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3936 if (t == -ERESTARTSYS)
3937 return t;
3938 return 0;
0fec171c 3939}
8cbbe86d 3940EXPORT_SYMBOL(wait_for_completion_interruptible);
1da177e4 3941
b15136e9 3942unsigned long __sched
8cbbe86d
AK
3943wait_for_completion_interruptible_timeout(struct completion *x,
3944 unsigned long timeout)
0fec171c 3945{
8cbbe86d 3946 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
0fec171c 3947}
8cbbe86d 3948EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
1da177e4 3949
8cbbe86d
AK
3950static long __sched
3951sleep_on_common(wait_queue_head_t *q, int state, long timeout)
1da177e4 3952{
0fec171c
IM
3953 unsigned long flags;
3954 wait_queue_t wait;
3955
3956 init_waitqueue_entry(&wait, current);
1da177e4 3957
8cbbe86d 3958 __set_current_state(state);
1da177e4 3959
8cbbe86d
AK
3960 spin_lock_irqsave(&q->lock, flags);
3961 __add_wait_queue(q, &wait);
3962 spin_unlock(&q->lock);
3963 timeout = schedule_timeout(timeout);
3964 spin_lock_irq(&q->lock);
3965 __remove_wait_queue(q, &wait);
3966 spin_unlock_irqrestore(&q->lock, flags);
3967
3968 return timeout;
3969}
3970
3971void __sched interruptible_sleep_on(wait_queue_head_t *q)
3972{
3973 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
1da177e4 3974}
1da177e4
LT
3975EXPORT_SYMBOL(interruptible_sleep_on);
3976
0fec171c 3977long __sched
95cdf3b7 3978interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
1da177e4 3979{
8cbbe86d 3980 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
1da177e4 3981}
1da177e4
LT
3982EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3983
0fec171c 3984void __sched sleep_on(wait_queue_head_t *q)
1da177e4 3985{
8cbbe86d 3986 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
1da177e4 3987}
1da177e4
LT
3988EXPORT_SYMBOL(sleep_on);
3989
0fec171c 3990long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
1da177e4 3991{
8cbbe86d 3992 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
1da177e4 3993}
1da177e4
LT
3994EXPORT_SYMBOL(sleep_on_timeout);
3995
b29739f9
IM
3996#ifdef CONFIG_RT_MUTEXES
3997
3998/*
3999 * rt_mutex_setprio - set the current priority of a task
4000 * @p: task
4001 * @prio: prio value (kernel-internal form)
4002 *
4003 * This function changes the 'effective' priority of a task. It does
4004 * not touch ->normal_prio like __setscheduler().
4005 *
4006 * Used by the rt_mutex code to implement priority inheritance logic.
4007 */
36c8b586 4008void rt_mutex_setprio(struct task_struct *p, int prio)
b29739f9
IM
4009{
4010 unsigned long flags;
83b699ed 4011 int oldprio, on_rq, running;
70b97a7f 4012 struct rq *rq;
b29739f9
IM
4013
4014 BUG_ON(prio < 0 || prio > MAX_PRIO);
4015
4016 rq = task_rq_lock(p, &flags);
a8e504d2 4017 update_rq_clock(rq);
b29739f9 4018
d5f9f942 4019 oldprio = p->prio;
dd41f596 4020 on_rq = p->se.on_rq;
051a1d1a 4021 running = task_current(rq, p);
83b699ed 4022 if (on_rq) {
69be72c1 4023 dequeue_task(rq, p, 0);
83b699ed
SV
4024 if (running)
4025 p->sched_class->put_prev_task(rq, p);
4026 }
dd41f596
IM
4027
4028 if (rt_prio(prio))
4029 p->sched_class = &rt_sched_class;
4030 else
4031 p->sched_class = &fair_sched_class;
4032
b29739f9
IM
4033 p->prio = prio;
4034
dd41f596 4035 if (on_rq) {
83b699ed
SV
4036 if (running)
4037 p->sched_class->set_curr_task(rq);
8159f87e 4038 enqueue_task(rq, p, 0);
b29739f9
IM
4039 /*
4040 * Reschedule if we are currently running on this runqueue and
d5f9f942
AM
4041 * our priority decreased, or if we are not currently running on
4042 * this runqueue and our priority is higher than the current's
b29739f9 4043 */
83b699ed 4044 if (running) {
d5f9f942
AM
4045 if (p->prio > oldprio)
4046 resched_task(rq->curr);
dd41f596
IM
4047 } else {
4048 check_preempt_curr(rq, p);
4049 }
b29739f9
IM
4050 }
4051 task_rq_unlock(rq, &flags);
4052}
4053
4054#endif
4055
36c8b586 4056void set_user_nice(struct task_struct *p, long nice)
1da177e4 4057{
dd41f596 4058 int old_prio, delta, on_rq;
1da177e4 4059 unsigned long flags;
70b97a7f 4060 struct rq *rq;
1da177e4
LT
4061
4062 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4063 return;
4064 /*
4065 * We have to be careful, if called from sys_setpriority(),
4066 * the task might be in the middle of scheduling on another CPU.
4067 */
4068 rq = task_rq_lock(p, &flags);
a8e504d2 4069 update_rq_clock(rq);
1da177e4
LT
4070 /*
4071 * The RT priorities are set via sched_setscheduler(), but we still
4072 * allow the 'normal' nice value to be set - but as expected
4073 * it wont have any effect on scheduling until the task is
dd41f596 4074 * SCHED_FIFO/SCHED_RR:
1da177e4 4075 */
e05606d3 4076 if (task_has_rt_policy(p)) {
1da177e4
LT
4077 p->static_prio = NICE_TO_PRIO(nice);
4078 goto out_unlock;
4079 }
dd41f596 4080 on_rq = p->se.on_rq;
58e2d4ca 4081 if (on_rq)
69be72c1 4082 dequeue_task(rq, p, 0);
1da177e4 4083
1da177e4 4084 p->static_prio = NICE_TO_PRIO(nice);
2dd73a4f 4085 set_load_weight(p);
b29739f9
IM
4086 old_prio = p->prio;
4087 p->prio = effective_prio(p);
4088 delta = p->prio - old_prio;
1da177e4 4089
dd41f596 4090 if (on_rq) {
8159f87e 4091 enqueue_task(rq, p, 0);
1da177e4 4092 /*
d5f9f942
AM
4093 * If the task increased its priority or is running and
4094 * lowered its priority, then reschedule its CPU:
1da177e4 4095 */
d5f9f942 4096 if (delta < 0 || (delta > 0 && task_running(rq, p)))
1da177e4
LT
4097 resched_task(rq->curr);
4098 }
4099out_unlock:
4100 task_rq_unlock(rq, &flags);
4101}
1da177e4
LT
4102EXPORT_SYMBOL(set_user_nice);
4103
e43379f1
MM
4104/*
4105 * can_nice - check if a task can reduce its nice value
4106 * @p: task
4107 * @nice: nice value
4108 */
36c8b586 4109int can_nice(const struct task_struct *p, const int nice)
e43379f1 4110{
024f4747
MM
4111 /* convert nice value [19,-20] to rlimit style value [1,40] */
4112 int nice_rlim = 20 - nice;
48f24c4d 4113
e43379f1
MM
4114 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4115 capable(CAP_SYS_NICE));
4116}
4117
1da177e4
LT
4118#ifdef __ARCH_WANT_SYS_NICE
4119
4120/*
4121 * sys_nice - change the priority of the current process.
4122 * @increment: priority increment
4123 *
4124 * sys_setpriority is a more generic, but much slower function that
4125 * does similar things.
4126 */
4127asmlinkage long sys_nice(int increment)
4128{
48f24c4d 4129 long nice, retval;
1da177e4
LT
4130
4131 /*
4132 * Setpriority might change our priority at the same moment.
4133 * We don't have to worry. Conceptually one call occurs first
4134 * and we have a single winner.
4135 */
e43379f1
MM
4136 if (increment < -40)
4137 increment = -40;
1da177e4
LT
4138 if (increment > 40)
4139 increment = 40;
4140
4141 nice = PRIO_TO_NICE(current->static_prio) + increment;
4142 if (nice < -20)
4143 nice = -20;
4144 if (nice > 19)
4145 nice = 19;
4146
e43379f1
MM
4147 if (increment < 0 && !can_nice(current, nice))
4148 return -EPERM;
4149
1da177e4
LT
4150 retval = security_task_setnice(current, nice);
4151 if (retval)
4152 return retval;
4153
4154 set_user_nice(current, nice);
4155 return 0;
4156}
4157
4158#endif
4159
4160/**
4161 * task_prio - return the priority value of a given task.
4162 * @p: the task in question.
4163 *
4164 * This is the priority value as seen by users in /proc.
4165 * RT tasks are offset by -200. Normal tasks are centered
4166 * around 0, value goes from -16 to +15.
4167 */
36c8b586 4168int task_prio(const struct task_struct *p)
1da177e4
LT
4169{
4170 return p->prio - MAX_RT_PRIO;
4171}
4172
4173/**
4174 * task_nice - return the nice value of a given task.
4175 * @p: the task in question.
4176 */
36c8b586 4177int task_nice(const struct task_struct *p)
1da177e4
LT
4178{
4179 return TASK_NICE(p);
4180}
1da177e4 4181EXPORT_SYMBOL_GPL(task_nice);
1da177e4
LT
4182
4183/**
4184 * idle_cpu - is a given cpu idle currently?
4185 * @cpu: the processor in question.
4186 */
4187int idle_cpu(int cpu)
4188{
4189 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4190}
4191
1da177e4
LT
4192/**
4193 * idle_task - return the idle task for a given cpu.
4194 * @cpu: the processor in question.
4195 */
36c8b586 4196struct task_struct *idle_task(int cpu)
1da177e4
LT
4197{
4198 return cpu_rq(cpu)->idle;
4199}
4200
4201/**
4202 * find_process_by_pid - find a process with a matching PID value.
4203 * @pid: the pid in question.
4204 */
a9957449 4205static struct task_struct *find_process_by_pid(pid_t pid)
1da177e4 4206{
228ebcbe 4207 return pid ? find_task_by_vpid(pid) : current;
1da177e4
LT
4208}
4209
4210/* Actually do priority change: must hold rq lock. */
dd41f596
IM
4211static void
4212__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
1da177e4 4213{
dd41f596 4214 BUG_ON(p->se.on_rq);
48f24c4d 4215
1da177e4 4216 p->policy = policy;
dd41f596
IM
4217 switch (p->policy) {
4218 case SCHED_NORMAL:
4219 case SCHED_BATCH:
4220 case SCHED_IDLE:
4221 p->sched_class = &fair_sched_class;
4222 break;
4223 case SCHED_FIFO:
4224 case SCHED_RR:
4225 p->sched_class = &rt_sched_class;
4226 break;
4227 }
4228
1da177e4 4229 p->rt_priority = prio;
b29739f9
IM
4230 p->normal_prio = normal_prio(p);
4231 /* we are holding p->pi_lock already */
4232 p->prio = rt_mutex_getprio(p);
2dd73a4f 4233 set_load_weight(p);
1da177e4
LT
4234}
4235
4236/**
72fd4a35 4237 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
1da177e4
LT
4238 * @p: the task in question.
4239 * @policy: new policy.
4240 * @param: structure containing the new RT priority.
5fe1d75f 4241 *
72fd4a35 4242 * NOTE that the task may be already dead.
1da177e4 4243 */
95cdf3b7
IM
4244int sched_setscheduler(struct task_struct *p, int policy,
4245 struct sched_param *param)
1da177e4 4246{
83b699ed 4247 int retval, oldprio, oldpolicy = -1, on_rq, running;
1da177e4 4248 unsigned long flags;
70b97a7f 4249 struct rq *rq;
1da177e4 4250
66e5393a
SR
4251 /* may grab non-irq protected spin_locks */
4252 BUG_ON(in_interrupt());
1da177e4
LT
4253recheck:
4254 /* double check policy once rq lock held */
4255 if (policy < 0)
4256 policy = oldpolicy = p->policy;
4257 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
dd41f596
IM
4258 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4259 policy != SCHED_IDLE)
b0a9499c 4260 return -EINVAL;
1da177e4
LT
4261 /*
4262 * Valid priorities for SCHED_FIFO and SCHED_RR are
dd41f596
IM
4263 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4264 * SCHED_BATCH and SCHED_IDLE is 0.
1da177e4
LT
4265 */
4266 if (param->sched_priority < 0 ||
95cdf3b7 4267 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
d46523ea 4268 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
1da177e4 4269 return -EINVAL;
e05606d3 4270 if (rt_policy(policy) != (param->sched_priority != 0))
1da177e4
LT
4271 return -EINVAL;
4272
37e4ab3f
OC
4273 /*
4274 * Allow unprivileged RT tasks to decrease priority:
4275 */
4276 if (!capable(CAP_SYS_NICE)) {
e05606d3 4277 if (rt_policy(policy)) {
8dc3e909 4278 unsigned long rlim_rtprio;
8dc3e909
ON
4279
4280 if (!lock_task_sighand(p, &flags))
4281 return -ESRCH;
4282 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4283 unlock_task_sighand(p, &flags);
4284
4285 /* can't set/change the rt policy */
4286 if (policy != p->policy && !rlim_rtprio)
4287 return -EPERM;
4288
4289 /* can't increase priority */
4290 if (param->sched_priority > p->rt_priority &&
4291 param->sched_priority > rlim_rtprio)
4292 return -EPERM;
4293 }
dd41f596
IM
4294 /*
4295 * Like positive nice levels, dont allow tasks to
4296 * move out of SCHED_IDLE either:
4297 */
4298 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4299 return -EPERM;
5fe1d75f 4300
37e4ab3f
OC
4301 /* can't change other user's priorities */
4302 if ((current->euid != p->euid) &&
4303 (current->euid != p->uid))
4304 return -EPERM;
4305 }
1da177e4
LT
4306
4307 retval = security_task_setscheduler(p, policy, param);
4308 if (retval)
4309 return retval;
b29739f9
IM
4310 /*
4311 * make sure no PI-waiters arrive (or leave) while we are
4312 * changing the priority of the task:
4313 */
4314 spin_lock_irqsave(&p->pi_lock, flags);
1da177e4
LT
4315 /*
4316 * To be able to change p->policy safely, the apropriate
4317 * runqueue lock must be held.
4318 */
b29739f9 4319 rq = __task_rq_lock(p);
1da177e4
LT
4320 /* recheck policy now with rq lock held */
4321 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4322 policy = oldpolicy = -1;
b29739f9
IM
4323 __task_rq_unlock(rq);
4324 spin_unlock_irqrestore(&p->pi_lock, flags);
1da177e4
LT
4325 goto recheck;
4326 }
2daa3577 4327 update_rq_clock(rq);
dd41f596 4328 on_rq = p->se.on_rq;
051a1d1a 4329 running = task_current(rq, p);
83b699ed 4330 if (on_rq) {
2e1cb74a 4331 deactivate_task(rq, p, 0);
83b699ed
SV
4332 if (running)
4333 p->sched_class->put_prev_task(rq, p);
4334 }
f6b53205 4335
1da177e4 4336 oldprio = p->prio;
dd41f596 4337 __setscheduler(rq, p, policy, param->sched_priority);
f6b53205 4338
dd41f596 4339 if (on_rq) {
83b699ed
SV
4340 if (running)
4341 p->sched_class->set_curr_task(rq);
dd41f596 4342 activate_task(rq, p, 0);
1da177e4
LT
4343 /*
4344 * Reschedule if we are currently running on this runqueue and
d5f9f942
AM
4345 * our priority decreased, or if we are not currently running on
4346 * this runqueue and our priority is higher than the current's
1da177e4 4347 */
83b699ed 4348 if (running) {
d5f9f942
AM
4349 if (p->prio > oldprio)
4350 resched_task(rq->curr);
dd41f596
IM
4351 } else {
4352 check_preempt_curr(rq, p);
4353 }
1da177e4 4354 }
b29739f9
IM
4355 __task_rq_unlock(rq);
4356 spin_unlock_irqrestore(&p->pi_lock, flags);
4357
95e02ca9
TG
4358 rt_mutex_adjust_pi(p);
4359
1da177e4
LT
4360 return 0;
4361}
4362EXPORT_SYMBOL_GPL(sched_setscheduler);
4363
95cdf3b7
IM
4364static int
4365do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
1da177e4 4366{
1da177e4
LT
4367 struct sched_param lparam;
4368 struct task_struct *p;
36c8b586 4369 int retval;
1da177e4
LT
4370
4371 if (!param || pid < 0)
4372 return -EINVAL;
4373 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4374 return -EFAULT;
5fe1d75f
ON
4375
4376 rcu_read_lock();
4377 retval = -ESRCH;
1da177e4 4378 p = find_process_by_pid(pid);
5fe1d75f
ON
4379 if (p != NULL)
4380 retval = sched_setscheduler(p, policy, &lparam);
4381 rcu_read_unlock();
36c8b586 4382
1da177e4
LT
4383 return retval;
4384}
4385
4386/**
4387 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4388 * @pid: the pid in question.
4389 * @policy: new policy.
4390 * @param: structure containing the new RT priority.
4391 */
41a2d6cf
IM
4392asmlinkage long
4393sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
1da177e4 4394{
c21761f1
JB
4395 /* negative values for policy are not valid */
4396 if (policy < 0)
4397 return -EINVAL;
4398
1da177e4
LT
4399 return do_sched_setscheduler(pid, policy, param);
4400}
4401
4402/**
4403 * sys_sched_setparam - set/change the RT priority of a thread
4404 * @pid: the pid in question.
4405 * @param: structure containing the new RT priority.
4406 */
4407asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4408{
4409 return do_sched_setscheduler(pid, -1, param);
4410}
4411
4412/**
4413 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4414 * @pid: the pid in question.
4415 */
4416asmlinkage long sys_sched_getscheduler(pid_t pid)
4417{
36c8b586 4418 struct task_struct *p;
3a5c359a 4419 int retval;
1da177e4
LT
4420
4421 if (pid < 0)
3a5c359a 4422 return -EINVAL;
1da177e4
LT
4423
4424 retval = -ESRCH;
4425 read_lock(&tasklist_lock);
4426 p = find_process_by_pid(pid);
4427 if (p) {
4428 retval = security_task_getscheduler(p);
4429 if (!retval)
4430 retval = p->policy;
4431 }
4432 read_unlock(&tasklist_lock);
1da177e4
LT
4433 return retval;
4434}
4435
4436/**
4437 * sys_sched_getscheduler - get the RT priority of a thread
4438 * @pid: the pid in question.
4439 * @param: structure containing the RT priority.
4440 */
4441asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4442{
4443 struct sched_param lp;
36c8b586 4444 struct task_struct *p;
3a5c359a 4445 int retval;
1da177e4
LT
4446
4447 if (!param || pid < 0)
3a5c359a 4448 return -EINVAL;
1da177e4
LT
4449
4450 read_lock(&tasklist_lock);
4451 p = find_process_by_pid(pid);
4452 retval = -ESRCH;
4453 if (!p)
4454 goto out_unlock;
4455
4456 retval = security_task_getscheduler(p);
4457 if (retval)
4458 goto out_unlock;
4459
4460 lp.sched_priority = p->rt_priority;
4461 read_unlock(&tasklist_lock);
4462
4463 /*
4464 * This one might sleep, we cannot do it with a spinlock held ...
4465 */
4466 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4467
1da177e4
LT
4468 return retval;
4469
4470out_unlock:
4471 read_unlock(&tasklist_lock);
4472 return retval;
4473}
4474
4475long sched_setaffinity(pid_t pid, cpumask_t new_mask)
4476{
1da177e4 4477 cpumask_t cpus_allowed;
36c8b586
IM
4478 struct task_struct *p;
4479 int retval;
1da177e4 4480
95402b38 4481 get_online_cpus();
1da177e4
LT
4482 read_lock(&tasklist_lock);
4483
4484 p = find_process_by_pid(pid);
4485 if (!p) {
4486 read_unlock(&tasklist_lock);
95402b38 4487 put_online_cpus();
1da177e4
LT
4488 return -ESRCH;
4489 }
4490
4491 /*
4492 * It is not safe to call set_cpus_allowed with the
41a2d6cf 4493 * tasklist_lock held. We will bump the task_struct's
1da177e4
LT
4494 * usage count and then drop tasklist_lock.
4495 */
4496 get_task_struct(p);
4497 read_unlock(&tasklist_lock);
4498
4499 retval = -EPERM;
4500 if ((current->euid != p->euid) && (current->euid != p->uid) &&
4501 !capable(CAP_SYS_NICE))
4502 goto out_unlock;
4503
e7834f8f
DQ
4504 retval = security_task_setscheduler(p, 0, NULL);
4505 if (retval)
4506 goto out_unlock;
4507
1da177e4
LT
4508 cpus_allowed = cpuset_cpus_allowed(p);
4509 cpus_and(new_mask, new_mask, cpus_allowed);
8707d8b8 4510 again:
1da177e4
LT
4511 retval = set_cpus_allowed(p, new_mask);
4512
8707d8b8
PM
4513 if (!retval) {
4514 cpus_allowed = cpuset_cpus_allowed(p);
4515 if (!cpus_subset(new_mask, cpus_allowed)) {
4516 /*
4517 * We must have raced with a concurrent cpuset
4518 * update. Just reset the cpus_allowed to the
4519 * cpuset's cpus_allowed
4520 */
4521 new_mask = cpus_allowed;
4522 goto again;
4523 }
4524 }
1da177e4
LT
4525out_unlock:
4526 put_task_struct(p);
95402b38 4527 put_online_cpus();
1da177e4
LT
4528 return retval;
4529}
4530
4531static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4532 cpumask_t *new_mask)
4533{
4534 if (len < sizeof(cpumask_t)) {
4535 memset(new_mask, 0, sizeof(cpumask_t));
4536 } else if (len > sizeof(cpumask_t)) {
4537 len = sizeof(cpumask_t);
4538 }
4539 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4540}
4541
4542/**
4543 * sys_sched_setaffinity - set the cpu affinity of a process
4544 * @pid: pid of the process
4545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4546 * @user_mask_ptr: user-space pointer to the new cpu mask
4547 */
4548asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
4549 unsigned long __user *user_mask_ptr)
4550{
4551 cpumask_t new_mask;
4552 int retval;
4553
4554 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
4555 if (retval)
4556 return retval;
4557
4558 return sched_setaffinity(pid, new_mask);
4559}
4560
4561/*
4562 * Represents all cpu's present in the system
4563 * In systems capable of hotplug, this map could dynamically grow
4564 * as new cpu's are detected in the system via any platform specific
4565 * method, such as ACPI for e.g.
4566 */
4567
4cef0c61 4568cpumask_t cpu_present_map __read_mostly;
1da177e4
LT
4569EXPORT_SYMBOL(cpu_present_map);
4570
4571#ifndef CONFIG_SMP
4cef0c61 4572cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
e16b38f7
GB
4573EXPORT_SYMBOL(cpu_online_map);
4574
4cef0c61 4575cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
e16b38f7 4576EXPORT_SYMBOL(cpu_possible_map);
1da177e4
LT
4577#endif
4578
4579long sched_getaffinity(pid_t pid, cpumask_t *mask)
4580{
36c8b586 4581 struct task_struct *p;
1da177e4 4582 int retval;
1da177e4 4583
95402b38 4584 get_online_cpus();
1da177e4
LT
4585 read_lock(&tasklist_lock);
4586
4587 retval = -ESRCH;
4588 p = find_process_by_pid(pid);
4589 if (!p)
4590 goto out_unlock;
4591
e7834f8f
DQ
4592 retval = security_task_getscheduler(p);
4593 if (retval)
4594 goto out_unlock;
4595
2f7016d9 4596 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
1da177e4
LT
4597
4598out_unlock:
4599 read_unlock(&tasklist_lock);
95402b38 4600 put_online_cpus();
1da177e4 4601
9531b62f 4602 return retval;
1da177e4
LT
4603}
4604
4605/**
4606 * sys_sched_getaffinity - get the cpu affinity of a process
4607 * @pid: pid of the process
4608 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4609 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4610 */
4611asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
4612 unsigned long __user *user_mask_ptr)
4613{
4614 int ret;
4615 cpumask_t mask;
4616
4617 if (len < sizeof(cpumask_t))
4618 return -EINVAL;
4619
4620 ret = sched_getaffinity(pid, &mask);
4621 if (ret < 0)
4622 return ret;
4623
4624 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
4625 return -EFAULT;
4626
4627 return sizeof(cpumask_t);
4628}
4629
4630/**
4631 * sys_sched_yield - yield the current processor to other threads.
4632 *
dd41f596
IM
4633 * This function yields the current CPU to other tasks. If there are no
4634 * other threads running on this CPU then this function will return.
1da177e4
LT
4635 */
4636asmlinkage long sys_sched_yield(void)
4637{
70b97a7f 4638 struct rq *rq = this_rq_lock();
1da177e4 4639
2d72376b 4640 schedstat_inc(rq, yld_count);
4530d7ab 4641 current->sched_class->yield_task(rq);
1da177e4
LT
4642
4643 /*
4644 * Since we are going to call schedule() anyway, there's
4645 * no need to preempt or enable interrupts:
4646 */
4647 __release(rq->lock);
8a25d5de 4648 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1da177e4
LT
4649 _raw_spin_unlock(&rq->lock);
4650 preempt_enable_no_resched();
4651
4652 schedule();
4653
4654 return 0;
4655}
4656
e7b38404 4657static void __cond_resched(void)
1da177e4 4658{
8e0a43d8
IM
4659#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4660 __might_sleep(__FILE__, __LINE__);
4661#endif
5bbcfd90
IM
4662 /*
4663 * The BKS might be reacquired before we have dropped
4664 * PREEMPT_ACTIVE, which could trigger a second
4665 * cond_resched() call.
4666 */
1da177e4
LT
4667 do {
4668 add_preempt_count(PREEMPT_ACTIVE);
4669 schedule();
4670 sub_preempt_count(PREEMPT_ACTIVE);
4671 } while (need_resched());
4672}
4673
4674int __sched cond_resched(void)
4675{
9414232f
IM
4676 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
4677 system_state == SYSTEM_RUNNING) {
1da177e4
LT
4678 __cond_resched();
4679 return 1;
4680 }
4681 return 0;
4682}
1da177e4
LT
4683EXPORT_SYMBOL(cond_resched);
4684
4685/*
4686 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4687 * call schedule, and on return reacquire the lock.
4688 *
41a2d6cf 4689 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
1da177e4
LT
4690 * operations here to prevent schedule() from being called twice (once via
4691 * spin_unlock(), once by hand).
4692 */
95cdf3b7 4693int cond_resched_lock(spinlock_t *lock)
1da177e4 4694{
6df3cecb
JK
4695 int ret = 0;
4696
1da177e4
LT
4697 if (need_lockbreak(lock)) {
4698 spin_unlock(lock);
4699 cpu_relax();
6df3cecb 4700 ret = 1;
1da177e4
LT
4701 spin_lock(lock);
4702 }
9414232f 4703 if (need_resched() && system_state == SYSTEM_RUNNING) {
8a25d5de 4704 spin_release(&lock->dep_map, 1, _THIS_IP_);
1da177e4
LT
4705 _raw_spin_unlock(lock);
4706 preempt_enable_no_resched();
4707 __cond_resched();
6df3cecb 4708 ret = 1;
1da177e4 4709 spin_lock(lock);
1da177e4 4710 }
6df3cecb 4711 return ret;
1da177e4 4712}
1da177e4
LT
4713EXPORT_SYMBOL(cond_resched_lock);
4714
4715int __sched cond_resched_softirq(void)
4716{
4717 BUG_ON(!in_softirq());
4718
9414232f 4719 if (need_resched() && system_state == SYSTEM_RUNNING) {
98d82567 4720 local_bh_enable();
1da177e4
LT
4721 __cond_resched();
4722 local_bh_disable();
4723 return 1;
4724 }
4725 return 0;
4726}
1da177e4
LT
4727EXPORT_SYMBOL(cond_resched_softirq);
4728
1da177e4
LT
4729/**
4730 * yield - yield the current processor to other threads.
4731 *
72fd4a35 4732 * This is a shortcut for kernel-space yielding - it marks the
1da177e4
LT
4733 * thread runnable and calls sys_sched_yield().
4734 */
4735void __sched yield(void)
4736{
4737 set_current_state(TASK_RUNNING);
4738 sys_sched_yield();
4739}
1da177e4
LT
4740EXPORT_SYMBOL(yield);
4741
4742/*
41a2d6cf 4743 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
1da177e4
LT
4744 * that process accounting knows that this is a task in IO wait state.
4745 *
4746 * But don't do that if it is a deliberate, throttling IO wait (this task
4747 * has set its backing_dev_info: the queue against which it should throttle)
4748 */
4749void __sched io_schedule(void)
4750{
70b97a7f 4751 struct rq *rq = &__raw_get_cpu_var(runqueues);
1da177e4 4752
0ff92245 4753 delayacct_blkio_start();
1da177e4
LT
4754 atomic_inc(&rq->nr_iowait);
4755 schedule();
4756 atomic_dec(&rq->nr_iowait);
0ff92245 4757 delayacct_blkio_end();
1da177e4 4758}
1da177e4
LT
4759EXPORT_SYMBOL(io_schedule);
4760
4761long __sched io_schedule_timeout(long timeout)
4762{
70b97a7f 4763 struct rq *rq = &__raw_get_cpu_var(runqueues);
1da177e4
LT
4764 long ret;
4765
0ff92245 4766 delayacct_blkio_start();
1da177e4
LT
4767 atomic_inc(&rq->nr_iowait);
4768 ret = schedule_timeout(timeout);
4769 atomic_dec(&rq->nr_iowait);
0ff92245 4770 delayacct_blkio_end();
1da177e4
LT
4771 return ret;
4772}
4773
4774/**
4775 * sys_sched_get_priority_max - return maximum RT priority.
4776 * @policy: scheduling class.
4777 *
4778 * this syscall returns the maximum rt_priority that can be used
4779 * by a given scheduling class.
4780 */
4781asmlinkage long sys_sched_get_priority_max(int policy)
4782{
4783 int ret = -EINVAL;
4784
4785 switch (policy) {
4786 case SCHED_FIFO:
4787 case SCHED_RR:
4788 ret = MAX_USER_RT_PRIO-1;
4789 break;
4790 case SCHED_NORMAL:
b0a9499c 4791 case SCHED_BATCH:
dd41f596 4792 case SCHED_IDLE:
1da177e4
LT
4793 ret = 0;
4794 break;
4795 }
4796 return ret;
4797}
4798
4799/**
4800 * sys_sched_get_priority_min - return minimum RT priority.
4801 * @policy: scheduling class.
4802 *
4803 * this syscall returns the minimum rt_priority that can be used
4804 * by a given scheduling class.
4805 */
4806asmlinkage long sys_sched_get_priority_min(int policy)
4807{
4808 int ret = -EINVAL;
4809
4810 switch (policy) {
4811 case SCHED_FIFO:
4812 case SCHED_RR:
4813 ret = 1;
4814 break;
4815 case SCHED_NORMAL:
b0a9499c 4816 case SCHED_BATCH:
dd41f596 4817 case SCHED_IDLE:
1da177e4
LT
4818 ret = 0;
4819 }
4820 return ret;
4821}
4822
4823/**
4824 * sys_sched_rr_get_interval - return the default timeslice of a process.
4825 * @pid: pid of the process.
4826 * @interval: userspace pointer to the timeslice value.
4827 *
4828 * this syscall writes the default timeslice value of a given process
4829 * into the user-space timespec buffer. A value of '0' means infinity.
4830 */
4831asmlinkage
4832long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
4833{
36c8b586 4834 struct task_struct *p;
a4ec24b4 4835 unsigned int time_slice;
3a5c359a 4836 int retval;
1da177e4 4837 struct timespec t;
1da177e4
LT
4838
4839 if (pid < 0)
3a5c359a 4840 return -EINVAL;
1da177e4
LT
4841
4842 retval = -ESRCH;
4843 read_lock(&tasklist_lock);
4844 p = find_process_by_pid(pid);
4845 if (!p)
4846 goto out_unlock;
4847
4848 retval = security_task_getscheduler(p);
4849 if (retval)
4850 goto out_unlock;
4851
77034937
IM
4852 /*
4853 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
4854 * tasks that are on an otherwise idle runqueue:
4855 */
4856 time_slice = 0;
4857 if (p->policy == SCHED_RR) {
a4ec24b4 4858 time_slice = DEF_TIMESLICE;
77034937 4859 } else {
a4ec24b4
DA
4860 struct sched_entity *se = &p->se;
4861 unsigned long flags;
4862 struct rq *rq;
4863
4864 rq = task_rq_lock(p, &flags);
77034937
IM
4865 if (rq->cfs.load.weight)
4866 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
a4ec24b4
DA
4867 task_rq_unlock(rq, &flags);
4868 }
1da177e4 4869 read_unlock(&tasklist_lock);
a4ec24b4 4870 jiffies_to_timespec(time_slice, &t);
1da177e4 4871 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
1da177e4 4872 return retval;
3a5c359a 4873
1da177e4
LT
4874out_unlock:
4875 read_unlock(&tasklist_lock);
4876 return retval;
4877}
4878
2ed6e34f 4879static const char stat_nam[] = "RSDTtZX";
36c8b586 4880
82a1fcb9 4881void sched_show_task(struct task_struct *p)
1da177e4 4882{
1da177e4 4883 unsigned long free = 0;
36c8b586 4884 unsigned state;
1da177e4 4885
1da177e4 4886 state = p->state ? __ffs(p->state) + 1 : 0;
cc4ea795 4887 printk(KERN_INFO "%-13.13s %c", p->comm,
2ed6e34f 4888 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4bd77321 4889#if BITS_PER_LONG == 32
1da177e4 4890 if (state == TASK_RUNNING)
cc4ea795 4891 printk(KERN_CONT " running ");
1da177e4 4892 else
cc4ea795 4893 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
1da177e4
LT
4894#else
4895 if (state == TASK_RUNNING)
cc4ea795 4896 printk(KERN_CONT " running task ");
1da177e4 4897 else
cc4ea795 4898 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
1da177e4
LT
4899#endif
4900#ifdef CONFIG_DEBUG_STACK_USAGE
4901 {
10ebffde 4902 unsigned long *n = end_of_stack(p);
1da177e4
LT
4903 while (!*n)
4904 n++;
10ebffde 4905 free = (unsigned long)n - (unsigned long)end_of_stack(p);
1da177e4
LT
4906 }
4907#endif
ba25f9dc 4908 printk(KERN_CONT "%5lu %5d %6d\n", free,
fcfd50af 4909 task_pid_nr(p), task_pid_nr(p->real_parent));
1da177e4
LT
4910
4911 if (state != TASK_RUNNING)
4912 show_stack(p, NULL);
4913}
4914
e59e2ae2 4915void show_state_filter(unsigned long state_filter)
1da177e4 4916{
36c8b586 4917 struct task_struct *g, *p;
1da177e4 4918
4bd77321
IM
4919#if BITS_PER_LONG == 32
4920 printk(KERN_INFO
4921 " task PC stack pid father\n");
1da177e4 4922#else
4bd77321
IM
4923 printk(KERN_INFO
4924 " task PC stack pid father\n");
1da177e4
LT
4925#endif
4926 read_lock(&tasklist_lock);
4927 do_each_thread(g, p) {
4928 /*
4929 * reset the NMI-timeout, listing all files on a slow
4930 * console might take alot of time:
4931 */
4932 touch_nmi_watchdog();
39bc89fd 4933 if (!state_filter || (p->state & state_filter))
82a1fcb9 4934 sched_show_task(p);
1da177e4
LT
4935 } while_each_thread(g, p);
4936
04c9167f
JF
4937 touch_all_softlockup_watchdogs();
4938
dd41f596
IM
4939#ifdef CONFIG_SCHED_DEBUG
4940 sysrq_sched_debug_show();
4941#endif
1da177e4 4942 read_unlock(&tasklist_lock);
e59e2ae2
IM
4943 /*
4944 * Only show locks if all tasks are dumped:
4945 */
4946 if (state_filter == -1)
4947 debug_show_all_locks();
1da177e4
LT
4948}
4949
1df21055
IM
4950void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4951{
dd41f596 4952 idle->sched_class = &idle_sched_class;
1df21055
IM
4953}
4954
f340c0d1
IM
4955/**
4956 * init_idle - set up an idle thread for a given CPU
4957 * @idle: task in question
4958 * @cpu: cpu the idle task belongs to
4959 *
4960 * NOTE: this function does not set the idle thread's NEED_RESCHED
4961 * flag, to make booting more robust.
4962 */
5c1e1767 4963void __cpuinit init_idle(struct task_struct *idle, int cpu)
1da177e4 4964{
70b97a7f 4965 struct rq *rq = cpu_rq(cpu);
1da177e4
LT
4966 unsigned long flags;
4967
dd41f596
IM
4968 __sched_fork(idle);
4969 idle->se.exec_start = sched_clock();
4970
b29739f9 4971 idle->prio = idle->normal_prio = MAX_PRIO;
1da177e4 4972 idle->cpus_allowed = cpumask_of_cpu(cpu);
dd41f596 4973 __set_task_cpu(idle, cpu);
1da177e4
LT
4974
4975 spin_lock_irqsave(&rq->lock, flags);
4976 rq->curr = rq->idle = idle;
4866cde0
NP
4977#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4978 idle->oncpu = 1;
4979#endif
1da177e4
LT
4980 spin_unlock_irqrestore(&rq->lock, flags);
4981
4982 /* Set the preempt count _outside_ the spinlocks! */
4983#if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
a1261f54 4984 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
1da177e4 4985#else
a1261f54 4986 task_thread_info(idle)->preempt_count = 0;
1da177e4 4987#endif
dd41f596
IM
4988 /*
4989 * The idle tasks have their own, simple scheduling class:
4990 */
4991 idle->sched_class = &idle_sched_class;
1da177e4
LT
4992}
4993
4994/*
4995 * In a system that switches off the HZ timer nohz_cpu_mask
4996 * indicates which cpus entered this state. This is used
4997 * in the rcu update to wait only for active cpus. For system
4998 * which do not switch off the HZ timer nohz_cpu_mask should
4999 * always be CPU_MASK_NONE.
5000 */
5001cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5002
19978ca6
IM
5003/*
5004 * Increase the granularity value when there are more CPUs,
5005 * because with more CPUs the 'effective latency' as visible
5006 * to users decreases. But the relationship is not linear,
5007 * so pick a second-best guess by going with the log2 of the
5008 * number of CPUs.
5009 *
5010 * This idea comes from the SD scheduler of Con Kolivas:
5011 */
5012static inline void sched_init_granularity(void)
5013{
5014 unsigned int factor = 1 + ilog2(num_online_cpus());
5015 const unsigned long limit = 200000000;
5016
5017 sysctl_sched_min_granularity *= factor;
5018 if (sysctl_sched_min_granularity > limit)
5019 sysctl_sched_min_granularity = limit;
5020
5021 sysctl_sched_latency *= factor;
5022 if (sysctl_sched_latency > limit)
5023 sysctl_sched_latency = limit;
5024
5025 sysctl_sched_wakeup_granularity *= factor;
5026 sysctl_sched_batch_wakeup_granularity *= factor;
5027}
5028
1da177e4
LT
5029#ifdef CONFIG_SMP
5030/*
5031 * This is how migration works:
5032 *
70b97a7f 5033 * 1) we queue a struct migration_req structure in the source CPU's
1da177e4
LT
5034 * runqueue and wake up that CPU's migration thread.
5035 * 2) we down() the locked semaphore => thread blocks.
5036 * 3) migration thread wakes up (implicitly it forces the migrated
5037 * thread off the CPU)
5038 * 4) it gets the migration request and checks whether the migrated
5039 * task is still in the wrong runqueue.
5040 * 5) if it's in the wrong runqueue then the migration thread removes
5041 * it and puts it into the right queue.
5042 * 6) migration thread up()s the semaphore.
5043 * 7) we wake up and the migration is done.
5044 */
5045
5046/*
5047 * Change a given task's CPU affinity. Migrate the thread to a
5048 * proper CPU and schedule it away if the CPU it's executing on
5049 * is removed from the allowed bitmask.
5050 *
5051 * NOTE: the caller must have a valid reference to the task, the
41a2d6cf 5052 * task must not exit() & deallocate itself prematurely. The
1da177e4
LT
5053 * call is not atomic; no spinlocks may be held.
5054 */
36c8b586 5055int set_cpus_allowed(struct task_struct *p, cpumask_t new_mask)
1da177e4 5056{
70b97a7f 5057 struct migration_req req;
1da177e4 5058 unsigned long flags;
70b97a7f 5059 struct rq *rq;
48f24c4d 5060 int ret = 0;
1da177e4
LT
5061
5062 rq = task_rq_lock(p, &flags);
5063 if (!cpus_intersects(new_mask, cpu_online_map)) {
5064 ret = -EINVAL;
5065 goto out;
5066 }
5067
73fe6aae
GH
5068 if (p->sched_class->set_cpus_allowed)
5069 p->sched_class->set_cpus_allowed(p, &new_mask);
5070 else {
0eab9146 5071 p->cpus_allowed = new_mask;
73fe6aae
GH
5072 p->nr_cpus_allowed = cpus_weight(new_mask);
5073 }
5074
1da177e4
LT
5075 /* Can the task run on the task's current CPU? If so, we're done */
5076 if (cpu_isset(task_cpu(p), new_mask))
5077 goto out;
5078
5079 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
5080 /* Need help from migration thread: drop lock and wait. */
5081 task_rq_unlock(rq, &flags);
5082 wake_up_process(rq->migration_thread);
5083 wait_for_completion(&req.done);
5084 tlb_migrate_finish(p->mm);
5085 return 0;
5086 }
5087out:
5088 task_rq_unlock(rq, &flags);
48f24c4d 5089
1da177e4
LT
5090 return ret;
5091}
1da177e4
LT
5092EXPORT_SYMBOL_GPL(set_cpus_allowed);
5093
5094/*
41a2d6cf 5095 * Move (not current) task off this cpu, onto dest cpu. We're doing
1da177e4
LT
5096 * this because either it can't run here any more (set_cpus_allowed()
5097 * away from this CPU, or CPU going down), or because we're
5098 * attempting to rebalance this task on exec (sched_exec).
5099 *
5100 * So we race with normal scheduler movements, but that's OK, as long
5101 * as the task is no longer on this CPU.
efc30814
KK
5102 *
5103 * Returns non-zero if task was successfully migrated.
1da177e4 5104 */
efc30814 5105static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
1da177e4 5106{
70b97a7f 5107 struct rq *rq_dest, *rq_src;
dd41f596 5108 int ret = 0, on_rq;
1da177e4
LT
5109
5110 if (unlikely(cpu_is_offline(dest_cpu)))
efc30814 5111 return ret;
1da177e4
LT
5112
5113 rq_src = cpu_rq(src_cpu);
5114 rq_dest = cpu_rq(dest_cpu);
5115
5116 double_rq_lock(rq_src, rq_dest);
5117 /* Already moved. */
5118 if (task_cpu(p) != src_cpu)
5119 goto out;
5120 /* Affinity changed (again). */
5121 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5122 goto out;
5123
dd41f596 5124 on_rq = p->se.on_rq;
6e82a3be 5125 if (on_rq)
2e1cb74a 5126 deactivate_task(rq_src, p, 0);
6e82a3be 5127
1da177e4 5128 set_task_cpu(p, dest_cpu);
dd41f596
IM
5129 if (on_rq) {
5130 activate_task(rq_dest, p, 0);
5131 check_preempt_curr(rq_dest, p);
1da177e4 5132 }
efc30814 5133 ret = 1;
1da177e4
LT
5134out:
5135 double_rq_unlock(rq_src, rq_dest);
efc30814 5136 return ret;
1da177e4
LT
5137}
5138
5139/*
5140 * migration_thread - this is a highprio system thread that performs
5141 * thread migration by bumping thread off CPU then 'pushing' onto
5142 * another runqueue.
5143 */
95cdf3b7 5144static int migration_thread(void *data)
1da177e4 5145{
1da177e4 5146 int cpu = (long)data;
70b97a7f 5147 struct rq *rq;
1da177e4
LT
5148
5149 rq = cpu_rq(cpu);
5150 BUG_ON(rq->migration_thread != current);
5151
5152 set_current_state(TASK_INTERRUPTIBLE);
5153 while (!kthread_should_stop()) {
70b97a7f 5154 struct migration_req *req;
1da177e4 5155 struct list_head *head;
1da177e4 5156
1da177e4
LT
5157 spin_lock_irq(&rq->lock);
5158
5159 if (cpu_is_offline(cpu)) {
5160 spin_unlock_irq(&rq->lock);
5161 goto wait_to_die;
5162 }
5163
5164 if (rq->active_balance) {
5165 active_load_balance(rq, cpu);
5166 rq->active_balance = 0;
5167 }
5168
5169 head = &rq->migration_queue;
5170
5171 if (list_empty(head)) {
5172 spin_unlock_irq(&rq->lock);
5173 schedule();
5174 set_current_state(TASK_INTERRUPTIBLE);
5175 continue;
5176 }
70b97a7f 5177 req = list_entry(head->next, struct migration_req, list);
1da177e4
LT
5178 list_del_init(head->next);
5179
674311d5
NP
5180 spin_unlock(&rq->lock);
5181 __migrate_task(req->task, cpu, req->dest_cpu);
5182 local_irq_enable();
1da177e4
LT
5183
5184 complete(&req->done);
5185 }
5186 __set_current_state(TASK_RUNNING);
5187 return 0;
5188
5189wait_to_die:
5190 /* Wait for kthread_stop */
5191 set_current_state(TASK_INTERRUPTIBLE);
5192 while (!kthread_should_stop()) {
5193 schedule();
5194 set_current_state(TASK_INTERRUPTIBLE);
5195 }
5196 __set_current_state(TASK_RUNNING);
5197 return 0;
5198}
5199
5200#ifdef CONFIG_HOTPLUG_CPU
f7b4cddc
ON
5201
5202static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5203{
5204 int ret;
5205
5206 local_irq_disable();
5207 ret = __migrate_task(p, src_cpu, dest_cpu);
5208 local_irq_enable();
5209 return ret;
5210}
5211
054b9108 5212/*
3a4fa0a2 5213 * Figure out where task on dead CPU should go, use force if necessary.
054b9108
KK
5214 * NOTE: interrupts should be disabled by the caller
5215 */
48f24c4d 5216static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
1da177e4 5217{
efc30814 5218 unsigned long flags;
1da177e4 5219 cpumask_t mask;
70b97a7f
IM
5220 struct rq *rq;
5221 int dest_cpu;
1da177e4 5222
3a5c359a
AK
5223 do {
5224 /* On same node? */
5225 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5226 cpus_and(mask, mask, p->cpus_allowed);
5227 dest_cpu = any_online_cpu(mask);
5228
5229 /* On any allowed CPU? */
5230 if (dest_cpu == NR_CPUS)
5231 dest_cpu = any_online_cpu(p->cpus_allowed);
5232
5233 /* No more Mr. Nice Guy. */
5234 if (dest_cpu == NR_CPUS) {
470fd646
CW
5235 cpumask_t cpus_allowed = cpuset_cpus_allowed_locked(p);
5236 /*
5237 * Try to stay on the same cpuset, where the
5238 * current cpuset may be a subset of all cpus.
5239 * The cpuset_cpus_allowed_locked() variant of
41a2d6cf 5240 * cpuset_cpus_allowed() will not block. It must be
470fd646
CW
5241 * called within calls to cpuset_lock/cpuset_unlock.
5242 */
3a5c359a 5243 rq = task_rq_lock(p, &flags);
470fd646 5244 p->cpus_allowed = cpus_allowed;
3a5c359a
AK
5245 dest_cpu = any_online_cpu(p->cpus_allowed);
5246 task_rq_unlock(rq, &flags);
1da177e4 5247
3a5c359a
AK
5248 /*
5249 * Don't tell them about moving exiting tasks or
5250 * kernel threads (both mm NULL), since they never
5251 * leave kernel.
5252 */
41a2d6cf 5253 if (p->mm && printk_ratelimit()) {
3a5c359a
AK
5254 printk(KERN_INFO "process %d (%s) no "
5255 "longer affine to cpu%d\n",
41a2d6cf
IM
5256 task_pid_nr(p), p->comm, dead_cpu);
5257 }
3a5c359a 5258 }
f7b4cddc 5259 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
1da177e4
LT
5260}
5261
5262/*
5263 * While a dead CPU has no uninterruptible tasks queued at this point,
5264 * it might still have a nonzero ->nr_uninterruptible counter, because
5265 * for performance reasons the counter is not stricly tracking tasks to
5266 * their home CPUs. So we just add the counter to another CPU's counter,
5267 * to keep the global sum constant after CPU-down:
5268 */
70b97a7f 5269static void migrate_nr_uninterruptible(struct rq *rq_src)
1da177e4 5270{
70b97a7f 5271 struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
1da177e4
LT
5272 unsigned long flags;
5273
5274 local_irq_save(flags);
5275 double_rq_lock(rq_src, rq_dest);
5276 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5277 rq_src->nr_uninterruptible = 0;
5278 double_rq_unlock(rq_src, rq_dest);
5279 local_irq_restore(flags);
5280}
5281
5282/* Run through task list and migrate tasks from the dead cpu. */
5283static void migrate_live_tasks(int src_cpu)
5284{
48f24c4d 5285 struct task_struct *p, *t;
1da177e4 5286
f7b4cddc 5287 read_lock(&tasklist_lock);
1da177e4 5288
48f24c4d
IM
5289 do_each_thread(t, p) {
5290 if (p == current)
1da177e4
LT
5291 continue;
5292
48f24c4d
IM
5293 if (task_cpu(p) == src_cpu)
5294 move_task_off_dead_cpu(src_cpu, p);
5295 } while_each_thread(t, p);
1da177e4 5296
f7b4cddc 5297 read_unlock(&tasklist_lock);
1da177e4
LT
5298}
5299
dd41f596
IM
5300/*
5301 * Schedules idle task to be the next runnable task on current CPU.
94bc9a7b
DA
5302 * It does so by boosting its priority to highest possible.
5303 * Used by CPU offline code.
1da177e4
LT
5304 */
5305void sched_idle_next(void)
5306{
48f24c4d 5307 int this_cpu = smp_processor_id();
70b97a7f 5308 struct rq *rq = cpu_rq(this_cpu);
1da177e4
LT
5309 struct task_struct *p = rq->idle;
5310 unsigned long flags;
5311
5312 /* cpu has to be offline */
48f24c4d 5313 BUG_ON(cpu_online(this_cpu));
1da177e4 5314
48f24c4d
IM
5315 /*
5316 * Strictly not necessary since rest of the CPUs are stopped by now
5317 * and interrupts disabled on the current cpu.
1da177e4
LT
5318 */
5319 spin_lock_irqsave(&rq->lock, flags);
5320
dd41f596 5321 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
48f24c4d 5322
94bc9a7b
DA
5323 update_rq_clock(rq);
5324 activate_task(rq, p, 0);
1da177e4
LT
5325
5326 spin_unlock_irqrestore(&rq->lock, flags);
5327}
5328
48f24c4d
IM
5329/*
5330 * Ensures that the idle task is using init_mm right before its cpu goes
1da177e4
LT
5331 * offline.
5332 */
5333void idle_task_exit(void)
5334{
5335 struct mm_struct *mm = current->active_mm;
5336
5337 BUG_ON(cpu_online(smp_processor_id()));
5338
5339 if (mm != &init_mm)
5340 switch_mm(mm, &init_mm, current);
5341 mmdrop(mm);
5342}
5343
054b9108 5344/* called under rq->lock with disabled interrupts */
36c8b586 5345static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
1da177e4 5346{
70b97a7f 5347 struct rq *rq = cpu_rq(dead_cpu);
1da177e4
LT
5348
5349 /* Must be exiting, otherwise would be on tasklist. */
270f722d 5350 BUG_ON(!p->exit_state);
1da177e4
LT
5351
5352 /* Cannot have done final schedule yet: would have vanished. */
c394cc9f 5353 BUG_ON(p->state == TASK_DEAD);
1da177e4 5354
48f24c4d 5355 get_task_struct(p);
1da177e4
LT
5356
5357 /*
5358 * Drop lock around migration; if someone else moves it,
41a2d6cf 5359 * that's OK. No task can be added to this CPU, so iteration is
1da177e4
LT
5360 * fine.
5361 */
f7b4cddc 5362 spin_unlock_irq(&rq->lock);
48f24c4d 5363 move_task_off_dead_cpu(dead_cpu, p);
f7b4cddc 5364 spin_lock_irq(&rq->lock);
1da177e4 5365
48f24c4d 5366 put_task_struct(p);
1da177e4
LT
5367}
5368
5369/* release_task() removes task from tasklist, so we won't find dead tasks. */
5370static void migrate_dead_tasks(unsigned int dead_cpu)
5371{
70b97a7f 5372 struct rq *rq = cpu_rq(dead_cpu);
dd41f596 5373 struct task_struct *next;
48f24c4d 5374
dd41f596
IM
5375 for ( ; ; ) {
5376 if (!rq->nr_running)
5377 break;
a8e504d2 5378 update_rq_clock(rq);
ff95f3df 5379 next = pick_next_task(rq, rq->curr);
dd41f596
IM
5380 if (!next)
5381 break;
5382 migrate_dead(dead_cpu, next);
e692ab53 5383
1da177e4
LT
5384 }
5385}
5386#endif /* CONFIG_HOTPLUG_CPU */
5387
e692ab53
NP
5388#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5389
5390static struct ctl_table sd_ctl_dir[] = {
e0361851
AD
5391 {
5392 .procname = "sched_domain",
c57baf1e 5393 .mode = 0555,
e0361851 5394 },
38605cae 5395 {0, },
e692ab53
NP
5396};
5397
5398static struct ctl_table sd_ctl_root[] = {
e0361851 5399 {
c57baf1e 5400 .ctl_name = CTL_KERN,
e0361851 5401 .procname = "kernel",
c57baf1e 5402 .mode = 0555,
e0361851
AD
5403 .child = sd_ctl_dir,
5404 },
38605cae 5405 {0, },
e692ab53
NP
5406};
5407
5408static struct ctl_table *sd_alloc_ctl_entry(int n)
5409{
5410 struct ctl_table *entry =
5cf9f062 5411 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
e692ab53 5412
e692ab53
NP
5413 return entry;
5414}
5415
6382bc90
MM
5416static void sd_free_ctl_entry(struct ctl_table **tablep)
5417{
cd790076 5418 struct ctl_table *entry;
6382bc90 5419
cd790076
MM
5420 /*
5421 * In the intermediate directories, both the child directory and
5422 * procname are dynamically allocated and could fail but the mode
41a2d6cf 5423 * will always be set. In the lowest directory the names are
cd790076
MM
5424 * static strings and all have proc handlers.
5425 */
5426 for (entry = *tablep; entry->mode; entry++) {
6382bc90
MM
5427 if (entry->child)
5428 sd_free_ctl_entry(&entry->child);
cd790076
MM
5429 if (entry->proc_handler == NULL)
5430 kfree(entry->procname);
5431 }
6382bc90
MM
5432
5433 kfree(*tablep);
5434 *tablep = NULL;
5435}
5436
e692ab53 5437static void
e0361851 5438set_table_entry(struct ctl_table *entry,
e692ab53
NP
5439 const char *procname, void *data, int maxlen,
5440 mode_t mode, proc_handler *proc_handler)
5441{
e692ab53
NP
5442 entry->procname = procname;
5443 entry->data = data;
5444 entry->maxlen = maxlen;
5445 entry->mode = mode;
5446 entry->proc_handler = proc_handler;
5447}
5448
5449static struct ctl_table *
5450sd_alloc_ctl_domain_table(struct sched_domain *sd)
5451{
ace8b3d6 5452 struct ctl_table *table = sd_alloc_ctl_entry(12);
e692ab53 5453
ad1cdc1d
MM
5454 if (table == NULL)
5455 return NULL;
5456
e0361851 5457 set_table_entry(&table[0], "min_interval", &sd->min_interval,
e692ab53 5458 sizeof(long), 0644, proc_doulongvec_minmax);
e0361851 5459 set_table_entry(&table[1], "max_interval", &sd->max_interval,
e692ab53 5460 sizeof(long), 0644, proc_doulongvec_minmax);
e0361851 5461 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
e692ab53 5462 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5463 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
e692ab53 5464 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5465 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
e692ab53 5466 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5467 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
e692ab53 5468 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5469 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
e692ab53 5470 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5471 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
e692ab53 5472 sizeof(int), 0644, proc_dointvec_minmax);
e0361851 5473 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
e692ab53 5474 sizeof(int), 0644, proc_dointvec_minmax);
ace8b3d6 5475 set_table_entry(&table[9], "cache_nice_tries",
e692ab53
NP
5476 &sd->cache_nice_tries,
5477 sizeof(int), 0644, proc_dointvec_minmax);
ace8b3d6 5478 set_table_entry(&table[10], "flags", &sd->flags,
e692ab53 5479 sizeof(int), 0644, proc_dointvec_minmax);
6323469f 5480 /* &table[11] is terminator */
e692ab53
NP
5481
5482 return table;
5483}
5484
9a4e7159 5485static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
e692ab53
NP
5486{
5487 struct ctl_table *entry, *table;
5488 struct sched_domain *sd;
5489 int domain_num = 0, i;
5490 char buf[32];
5491
5492 for_each_domain(cpu, sd)
5493 domain_num++;
5494 entry = table = sd_alloc_ctl_entry(domain_num + 1);
ad1cdc1d
MM
5495 if (table == NULL)
5496 return NULL;
e692ab53
NP
5497
5498 i = 0;
5499 for_each_domain(cpu, sd) {
5500 snprintf(buf, 32, "domain%d", i);
e692ab53 5501 entry->procname = kstrdup(buf, GFP_KERNEL);
c57baf1e 5502 entry->mode = 0555;
e692ab53
NP
5503 entry->child = sd_alloc_ctl_domain_table(sd);
5504 entry++;
5505 i++;
5506 }
5507 return table;
5508}
5509
5510static struct ctl_table_header *sd_sysctl_header;
6382bc90 5511static void register_sched_domain_sysctl(void)
e692ab53
NP
5512{
5513 int i, cpu_num = num_online_cpus();
5514 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5515 char buf[32];
5516
7378547f
MM
5517 WARN_ON(sd_ctl_dir[0].child);
5518 sd_ctl_dir[0].child = entry;
5519
ad1cdc1d
MM
5520 if (entry == NULL)
5521 return;
5522
97b6ea7b 5523 for_each_online_cpu(i) {
e692ab53 5524 snprintf(buf, 32, "cpu%d", i);
e692ab53 5525 entry->procname = kstrdup(buf, GFP_KERNEL);
c57baf1e 5526 entry->mode = 0555;
e692ab53 5527 entry->child = sd_alloc_ctl_cpu_table(i);
97b6ea7b 5528 entry++;
e692ab53 5529 }
7378547f
MM
5530
5531 WARN_ON(sd_sysctl_header);
e692ab53
NP
5532 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5533}
6382bc90 5534
7378547f 5535/* may be called multiple times per register */
6382bc90
MM
5536static void unregister_sched_domain_sysctl(void)
5537{
7378547f
MM
5538 if (sd_sysctl_header)
5539 unregister_sysctl_table(sd_sysctl_header);
6382bc90 5540 sd_sysctl_header = NULL;
7378547f
MM
5541 if (sd_ctl_dir[0].child)
5542 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6382bc90 5543}
e692ab53 5544#else
6382bc90
MM
5545static void register_sched_domain_sysctl(void)
5546{
5547}
5548static void unregister_sched_domain_sysctl(void)
e692ab53
NP
5549{
5550}
5551#endif
5552
1da177e4
LT
5553/*
5554 * migration_call - callback that gets triggered when a CPU is added.
5555 * Here we can start up the necessary migration thread for the new CPU.
5556 */
48f24c4d
IM
5557static int __cpuinit
5558migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
1da177e4 5559{
1da177e4 5560 struct task_struct *p;
48f24c4d 5561 int cpu = (long)hcpu;
1da177e4 5562 unsigned long flags;
70b97a7f 5563 struct rq *rq;
1da177e4
LT
5564
5565 switch (action) {
5be9361c 5566
1da177e4 5567 case CPU_UP_PREPARE:
8bb78442 5568 case CPU_UP_PREPARE_FROZEN:
dd41f596 5569 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
1da177e4
LT
5570 if (IS_ERR(p))
5571 return NOTIFY_BAD;
1da177e4
LT
5572 kthread_bind(p, cpu);
5573 /* Must be high prio: stop_machine expects to yield to it. */
5574 rq = task_rq_lock(p, &flags);
dd41f596 5575 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
1da177e4
LT
5576 task_rq_unlock(rq, &flags);
5577 cpu_rq(cpu)->migration_thread = p;
5578 break;
48f24c4d 5579
1da177e4 5580 case CPU_ONLINE:
8bb78442 5581 case CPU_ONLINE_FROZEN:
3a4fa0a2 5582 /* Strictly unnecessary, as first user will wake it. */
1da177e4 5583 wake_up_process(cpu_rq(cpu)->migration_thread);
57d885fe
GH
5584
5585 /* Update our root-domain */
5586 rq = cpu_rq(cpu);
5587 spin_lock_irqsave(&rq->lock, flags);
5588 if (rq->rd) {
5589 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5590 cpu_set(cpu, rq->rd->online);
5591 }
5592 spin_unlock_irqrestore(&rq->lock, flags);
1da177e4 5593 break;
48f24c4d 5594
1da177e4
LT
5595#ifdef CONFIG_HOTPLUG_CPU
5596 case CPU_UP_CANCELED:
8bb78442 5597 case CPU_UP_CANCELED_FROZEN:
fc75cdfa
HC
5598 if (!cpu_rq(cpu)->migration_thread)
5599 break;
41a2d6cf 5600 /* Unbind it from offline cpu so it can run. Fall thru. */
a4c4af7c
HC
5601 kthread_bind(cpu_rq(cpu)->migration_thread,
5602 any_online_cpu(cpu_online_map));
1da177e4
LT
5603 kthread_stop(cpu_rq(cpu)->migration_thread);
5604 cpu_rq(cpu)->migration_thread = NULL;
5605 break;
48f24c4d 5606
1da177e4 5607 case CPU_DEAD:
8bb78442 5608 case CPU_DEAD_FROZEN:
470fd646 5609 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
1da177e4
LT
5610 migrate_live_tasks(cpu);
5611 rq = cpu_rq(cpu);
5612 kthread_stop(rq->migration_thread);
5613 rq->migration_thread = NULL;
5614 /* Idle task back to normal (off runqueue, low prio) */
d2da272a 5615 spin_lock_irq(&rq->lock);
a8e504d2 5616 update_rq_clock(rq);
2e1cb74a 5617 deactivate_task(rq, rq->idle, 0);
1da177e4 5618 rq->idle->static_prio = MAX_PRIO;
dd41f596
IM
5619 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5620 rq->idle->sched_class = &idle_sched_class;
1da177e4 5621 migrate_dead_tasks(cpu);
d2da272a 5622 spin_unlock_irq(&rq->lock);
470fd646 5623 cpuset_unlock();
1da177e4
LT
5624 migrate_nr_uninterruptible(rq);
5625 BUG_ON(rq->nr_running != 0);
5626
41a2d6cf
IM
5627 /*
5628 * No need to migrate the tasks: it was best-effort if
5629 * they didn't take sched_hotcpu_mutex. Just wake up
5630 * the requestors.
5631 */
1da177e4
LT
5632 spin_lock_irq(&rq->lock);
5633 while (!list_empty(&rq->migration_queue)) {
70b97a7f
IM
5634 struct migration_req *req;
5635
1da177e4 5636 req = list_entry(rq->migration_queue.next,
70b97a7f 5637 struct migration_req, list);
1da177e4
LT
5638 list_del_init(&req->list);
5639 complete(&req->done);
5640 }
5641 spin_unlock_irq(&rq->lock);
5642 break;
57d885fe
GH
5643
5644 case CPU_DOWN_PREPARE:
5645 /* Update our root-domain */
5646 rq = cpu_rq(cpu);
5647 spin_lock_irqsave(&rq->lock, flags);
5648 if (rq->rd) {
5649 BUG_ON(!cpu_isset(cpu, rq->rd->span));
5650 cpu_clear(cpu, rq->rd->online);
5651 }
5652 spin_unlock_irqrestore(&rq->lock, flags);
5653 break;
1da177e4
LT
5654#endif
5655 }
5656 return NOTIFY_OK;
5657}
5658
5659/* Register at highest priority so that task migration (migrate_all_tasks)
5660 * happens before everything else.
5661 */
26c2143b 5662static struct notifier_block __cpuinitdata migration_notifier = {
1da177e4
LT
5663 .notifier_call = migration_call,
5664 .priority = 10
5665};
5666
e6fe6649 5667void __init migration_init(void)
1da177e4
LT
5668{
5669 void *cpu = (void *)(long)smp_processor_id();
07dccf33 5670 int err;
48f24c4d
IM
5671
5672 /* Start one for the boot CPU: */
07dccf33
AM
5673 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5674 BUG_ON(err == NOTIFY_BAD);
1da177e4
LT
5675 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5676 register_cpu_notifier(&migration_notifier);
1da177e4
LT
5677}
5678#endif
5679
5680#ifdef CONFIG_SMP
476f3534
CL
5681
5682/* Number of possible processor ids */
5683int nr_cpu_ids __read_mostly = NR_CPUS;
5684EXPORT_SYMBOL(nr_cpu_ids);
5685
3e9830dc 5686#ifdef CONFIG_SCHED_DEBUG
4dcf6aff
IM
5687
5688static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
1da177e4 5689{
4dcf6aff
IM
5690 struct sched_group *group = sd->groups;
5691 cpumask_t groupmask;
5692 char str[NR_CPUS];
1da177e4 5693
4dcf6aff
IM
5694 cpumask_scnprintf(str, NR_CPUS, sd->span);
5695 cpus_clear(groupmask);
5696
5697 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5698
5699 if (!(sd->flags & SD_LOAD_BALANCE)) {
5700 printk("does not load-balance\n");
5701 if (sd->parent)
5702 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5703 " has parent");
5704 return -1;
41c7ce9a
NP
5705 }
5706
4dcf6aff
IM
5707 printk(KERN_CONT "span %s\n", str);
5708
5709 if (!cpu_isset(cpu, sd->span)) {
5710 printk(KERN_ERR "ERROR: domain->span does not contain "
5711 "CPU%d\n", cpu);
5712 }
5713 if (!cpu_isset(cpu, group->cpumask)) {
5714 printk(KERN_ERR "ERROR: domain->groups does not contain"
5715 " CPU%d\n", cpu);
5716 }
1da177e4 5717
4dcf6aff 5718 printk(KERN_DEBUG "%*s groups:", level + 1, "");
1da177e4 5719 do {
4dcf6aff
IM
5720 if (!group) {
5721 printk("\n");
5722 printk(KERN_ERR "ERROR: group is NULL\n");
1da177e4
LT
5723 break;
5724 }
5725
4dcf6aff
IM
5726 if (!group->__cpu_power) {
5727 printk(KERN_CONT "\n");
5728 printk(KERN_ERR "ERROR: domain->cpu_power not "
5729 "set\n");
5730 break;
5731 }
1da177e4 5732
4dcf6aff
IM
5733 if (!cpus_weight(group->cpumask)) {
5734 printk(KERN_CONT "\n");
5735 printk(KERN_ERR "ERROR: empty group\n");
5736 break;
5737 }
1da177e4 5738
4dcf6aff
IM
5739 if (cpus_intersects(groupmask, group->cpumask)) {
5740 printk(KERN_CONT "\n");
5741 printk(KERN_ERR "ERROR: repeated CPUs\n");
5742 break;
5743 }
1da177e4 5744
4dcf6aff 5745 cpus_or(groupmask, groupmask, group->cpumask);
1da177e4 5746
4dcf6aff
IM
5747 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
5748 printk(KERN_CONT " %s", str);
1da177e4 5749
4dcf6aff
IM
5750 group = group->next;
5751 } while (group != sd->groups);
5752 printk(KERN_CONT "\n");
1da177e4 5753
4dcf6aff
IM
5754 if (!cpus_equal(sd->span, groupmask))
5755 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
1da177e4 5756
4dcf6aff
IM
5757 if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
5758 printk(KERN_ERR "ERROR: parent span is not a superset "
5759 "of domain->span\n");
5760 return 0;
5761}
1da177e4 5762
4dcf6aff
IM
5763static void sched_domain_debug(struct sched_domain *sd, int cpu)
5764{
5765 int level = 0;
1da177e4 5766
4dcf6aff
IM
5767 if (!sd) {
5768 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5769 return;
5770 }
1da177e4 5771
4dcf6aff
IM
5772 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5773
5774 for (;;) {
5775 if (sched_domain_debug_one(sd, cpu, level))
5776 break;
1da177e4
LT
5777 level++;
5778 sd = sd->parent;
33859f7f 5779 if (!sd)
4dcf6aff
IM
5780 break;
5781 }
1da177e4
LT
5782}
5783#else
48f24c4d 5784# define sched_domain_debug(sd, cpu) do { } while (0)
1da177e4
LT
5785#endif
5786
1a20ff27 5787static int sd_degenerate(struct sched_domain *sd)
245af2c7
SS
5788{
5789 if (cpus_weight(sd->span) == 1)
5790 return 1;
5791
5792 /* Following flags need at least 2 groups */
5793 if (sd->flags & (SD_LOAD_BALANCE |
5794 SD_BALANCE_NEWIDLE |
5795 SD_BALANCE_FORK |
89c4710e
SS
5796 SD_BALANCE_EXEC |
5797 SD_SHARE_CPUPOWER |
5798 SD_SHARE_PKG_RESOURCES)) {
245af2c7
SS
5799 if (sd->groups != sd->groups->next)
5800 return 0;
5801 }
5802
5803 /* Following flags don't use groups */
5804 if (sd->flags & (SD_WAKE_IDLE |
5805 SD_WAKE_AFFINE |
5806 SD_WAKE_BALANCE))
5807 return 0;
5808
5809 return 1;
5810}
5811
48f24c4d
IM
5812static int
5813sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
245af2c7
SS
5814{
5815 unsigned long cflags = sd->flags, pflags = parent->flags;
5816
5817 if (sd_degenerate(parent))
5818 return 1;
5819
5820 if (!cpus_equal(sd->span, parent->span))
5821 return 0;
5822
5823 /* Does parent contain flags not in child? */
5824 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5825 if (cflags & SD_WAKE_AFFINE)
5826 pflags &= ~SD_WAKE_BALANCE;
5827 /* Flags needing groups don't count if only 1 group in parent */
5828 if (parent->groups == parent->groups->next) {
5829 pflags &= ~(SD_LOAD_BALANCE |
5830 SD_BALANCE_NEWIDLE |
5831 SD_BALANCE_FORK |
89c4710e
SS
5832 SD_BALANCE_EXEC |
5833 SD_SHARE_CPUPOWER |
5834 SD_SHARE_PKG_RESOURCES);
245af2c7
SS
5835 }
5836 if (~cflags & pflags)
5837 return 0;
5838
5839 return 1;
5840}
5841
57d885fe
GH
5842static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5843{
5844 unsigned long flags;
5845 const struct sched_class *class;
5846
5847 spin_lock_irqsave(&rq->lock, flags);
5848
5849 if (rq->rd) {
5850 struct root_domain *old_rd = rq->rd;
5851
0eab9146 5852 for (class = sched_class_highest; class; class = class->next) {
57d885fe
GH
5853 if (class->leave_domain)
5854 class->leave_domain(rq);
0eab9146 5855 }
57d885fe
GH
5856
5857 if (atomic_dec_and_test(&old_rd->refcount))
5858 kfree(old_rd);
5859 }
5860
5861 atomic_inc(&rd->refcount);
5862 rq->rd = rd;
5863
0eab9146 5864 for (class = sched_class_highest; class; class = class->next) {
57d885fe
GH
5865 if (class->join_domain)
5866 class->join_domain(rq);
0eab9146 5867 }
57d885fe
GH
5868
5869 spin_unlock_irqrestore(&rq->lock, flags);
5870}
5871
5872static void init_rootdomain(struct root_domain *rd, const cpumask_t *map)
5873{
5874 memset(rd, 0, sizeof(*rd));
5875
5876 rd->span = *map;
5877 cpus_and(rd->online, rd->span, cpu_online_map);
5878}
5879
5880static void init_defrootdomain(void)
5881{
5882 cpumask_t cpus = CPU_MASK_ALL;
5883
5884 init_rootdomain(&def_root_domain, &cpus);
5885 atomic_set(&def_root_domain.refcount, 1);
5886}
5887
5888static struct root_domain *alloc_rootdomain(const cpumask_t *map)
5889{
5890 struct root_domain *rd;
5891
5892 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5893 if (!rd)
5894 return NULL;
5895
5896 init_rootdomain(rd, map);
5897
5898 return rd;
5899}
5900
1da177e4 5901/*
0eab9146 5902 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
1da177e4
LT
5903 * hold the hotplug lock.
5904 */
0eab9146
IM
5905static void
5906cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
1da177e4 5907{
70b97a7f 5908 struct rq *rq = cpu_rq(cpu);
245af2c7
SS
5909 struct sched_domain *tmp;
5910
5911 /* Remove the sched domains which do not contribute to scheduling. */
5912 for (tmp = sd; tmp; tmp = tmp->parent) {
5913 struct sched_domain *parent = tmp->parent;
5914 if (!parent)
5915 break;
1a848870 5916 if (sd_parent_degenerate(tmp, parent)) {
245af2c7 5917 tmp->parent = parent->parent;
1a848870
SS
5918 if (parent->parent)
5919 parent->parent->child = tmp;
5920 }
245af2c7
SS
5921 }
5922
1a848870 5923 if (sd && sd_degenerate(sd)) {
245af2c7 5924 sd = sd->parent;
1a848870
SS
5925 if (sd)
5926 sd->child = NULL;
5927 }
1da177e4
LT
5928
5929 sched_domain_debug(sd, cpu);
5930
57d885fe 5931 rq_attach_root(rq, rd);
674311d5 5932 rcu_assign_pointer(rq->sd, sd);
1da177e4
LT
5933}
5934
5935/* cpus with isolated domains */
67af63a6 5936static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
1da177e4
LT
5937
5938/* Setup the mask of cpus configured for isolated domains */
5939static int __init isolated_cpu_setup(char *str)
5940{
5941 int ints[NR_CPUS], i;
5942
5943 str = get_options(str, ARRAY_SIZE(ints), ints);
5944 cpus_clear(cpu_isolated_map);
5945 for (i = 1; i <= ints[0]; i++)
5946 if (ints[i] < NR_CPUS)
5947 cpu_set(ints[i], cpu_isolated_map);
5948 return 1;
5949}
5950
8927f494 5951__setup("isolcpus=", isolated_cpu_setup);
1da177e4
LT
5952
5953/*
6711cab4
SS
5954 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5955 * to a function which identifies what group(along with sched group) a CPU
5956 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5957 * (due to the fact that we keep track of groups covered with a cpumask_t).
1da177e4
LT
5958 *
5959 * init_sched_build_groups will build a circular linked list of the groups
5960 * covered by the given span, and will set each group's ->cpumask correctly,
5961 * and ->cpu_power to 0.
5962 */
a616058b 5963static void
6711cab4
SS
5964init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
5965 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
5966 struct sched_group **sg))
1da177e4
LT
5967{
5968 struct sched_group *first = NULL, *last = NULL;
5969 cpumask_t covered = CPU_MASK_NONE;
5970 int i;
5971
5972 for_each_cpu_mask(i, span) {
6711cab4
SS
5973 struct sched_group *sg;
5974 int group = group_fn(i, cpu_map, &sg);
1da177e4
LT
5975 int j;
5976
5977 if (cpu_isset(i, covered))
5978 continue;
5979
5980 sg->cpumask = CPU_MASK_NONE;
5517d86b 5981 sg->__cpu_power = 0;
1da177e4
LT
5982
5983 for_each_cpu_mask(j, span) {
6711cab4 5984 if (group_fn(j, cpu_map, NULL) != group)
1da177e4
LT
5985 continue;
5986
5987 cpu_set(j, covered);
5988 cpu_set(j, sg->cpumask);
5989 }
5990 if (!first)
5991 first = sg;
5992 if (last)
5993 last->next = sg;
5994 last = sg;
5995 }
5996 last->next = first;
5997}
5998
9c1cfda2 5999#define SD_NODES_PER_DOMAIN 16
1da177e4 6000
9c1cfda2 6001#ifdef CONFIG_NUMA
198e2f18 6002
9c1cfda2
JH
6003/**
6004 * find_next_best_node - find the next node to include in a sched_domain
6005 * @node: node whose sched_domain we're building
6006 * @used_nodes: nodes already in the sched_domain
6007 *
41a2d6cf 6008 * Find the next node to include in a given scheduling domain. Simply
9c1cfda2
JH
6009 * finds the closest node not already in the @used_nodes map.
6010 *
6011 * Should use nodemask_t.
6012 */
6013static int find_next_best_node(int node, unsigned long *used_nodes)
6014{
6015 int i, n, val, min_val, best_node = 0;
6016
6017 min_val = INT_MAX;
6018
6019 for (i = 0; i < MAX_NUMNODES; i++) {
6020 /* Start at @node */
6021 n = (node + i) % MAX_NUMNODES;
6022
6023 if (!nr_cpus_node(n))
6024 continue;
6025
6026 /* Skip already used nodes */
6027 if (test_bit(n, used_nodes))
6028 continue;
6029
6030 /* Simple min distance search */
6031 val = node_distance(node, n);
6032
6033 if (val < min_val) {
6034 min_val = val;
6035 best_node = n;
6036 }
6037 }
6038
6039 set_bit(best_node, used_nodes);
6040 return best_node;
6041}
6042
6043/**
6044 * sched_domain_node_span - get a cpumask for a node's sched_domain
6045 * @node: node whose cpumask we're constructing
6046 * @size: number of nodes to include in this span
6047 *
41a2d6cf 6048 * Given a node, construct a good cpumask for its sched_domain to span. It
9c1cfda2
JH
6049 * should be one that prevents unnecessary balancing, but also spreads tasks
6050 * out optimally.
6051 */
6052static cpumask_t sched_domain_node_span(int node)
6053{
9c1cfda2 6054 DECLARE_BITMAP(used_nodes, MAX_NUMNODES);
48f24c4d
IM
6055 cpumask_t span, nodemask;
6056 int i;
9c1cfda2
JH
6057
6058 cpus_clear(span);
6059 bitmap_zero(used_nodes, MAX_NUMNODES);
6060
6061 nodemask = node_to_cpumask(node);
6062 cpus_or(span, span, nodemask);
6063 set_bit(node, used_nodes);
6064
6065 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6066 int next_node = find_next_best_node(node, used_nodes);
48f24c4d 6067
9c1cfda2
JH
6068 nodemask = node_to_cpumask(next_node);
6069 cpus_or(span, span, nodemask);
6070 }
6071
6072 return span;
6073}
6074#endif
6075
5c45bf27 6076int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
48f24c4d 6077
9c1cfda2 6078/*
48f24c4d 6079 * SMT sched-domains:
9c1cfda2 6080 */
1da177e4
LT
6081#ifdef CONFIG_SCHED_SMT
6082static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6711cab4 6083static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
48f24c4d 6084
41a2d6cf
IM
6085static int
6086cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
1da177e4 6087{
6711cab4
SS
6088 if (sg)
6089 *sg = &per_cpu(sched_group_cpus, cpu);
1da177e4
LT
6090 return cpu;
6091}
6092#endif
6093
48f24c4d
IM
6094/*
6095 * multi-core sched-domains:
6096 */
1e9f28fa
SS
6097#ifdef CONFIG_SCHED_MC
6098static DEFINE_PER_CPU(struct sched_domain, core_domains);
6711cab4 6099static DEFINE_PER_CPU(struct sched_group, sched_group_core);
1e9f28fa
SS
6100#endif
6101
6102#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
41a2d6cf
IM
6103static int
6104cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
1e9f28fa 6105{
6711cab4 6106 int group;
d5a7430d 6107 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
a616058b 6108 cpus_and(mask, mask, *cpu_map);
6711cab4
SS
6109 group = first_cpu(mask);
6110 if (sg)
6111 *sg = &per_cpu(sched_group_core, group);
6112 return group;
1e9f28fa
SS
6113}
6114#elif defined(CONFIG_SCHED_MC)
41a2d6cf
IM
6115static int
6116cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
1e9f28fa 6117{
6711cab4
SS
6118 if (sg)
6119 *sg = &per_cpu(sched_group_core, cpu);
1e9f28fa
SS
6120 return cpu;
6121}
6122#endif
6123
1da177e4 6124static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6711cab4 6125static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
48f24c4d 6126
41a2d6cf
IM
6127static int
6128cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
1da177e4 6129{
6711cab4 6130 int group;
48f24c4d 6131#ifdef CONFIG_SCHED_MC
1e9f28fa 6132 cpumask_t mask = cpu_coregroup_map(cpu);
a616058b 6133 cpus_and(mask, mask, *cpu_map);
6711cab4 6134 group = first_cpu(mask);
1e9f28fa 6135#elif defined(CONFIG_SCHED_SMT)
d5a7430d 6136 cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
a616058b 6137 cpus_and(mask, mask, *cpu_map);
6711cab4 6138 group = first_cpu(mask);
1da177e4 6139#else
6711cab4 6140 group = cpu;
1da177e4 6141#endif
6711cab4
SS
6142 if (sg)
6143 *sg = &per_cpu(sched_group_phys, group);
6144 return group;
1da177e4
LT
6145}
6146
6147#ifdef CONFIG_NUMA
1da177e4 6148/*
9c1cfda2
JH
6149 * The init_sched_build_groups can't handle what we want to do with node
6150 * groups, so roll our own. Now each node has its own list of groups which
6151 * gets dynamically allocated.
1da177e4 6152 */
9c1cfda2 6153static DEFINE_PER_CPU(struct sched_domain, node_domains);
d1b55138 6154static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
1da177e4 6155
9c1cfda2 6156static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6711cab4 6157static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
9c1cfda2 6158
6711cab4
SS
6159static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6160 struct sched_group **sg)
9c1cfda2 6161{
6711cab4
SS
6162 cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
6163 int group;
6164
6165 cpus_and(nodemask, nodemask, *cpu_map);
6166 group = first_cpu(nodemask);
6167
6168 if (sg)
6169 *sg = &per_cpu(sched_group_allnodes, group);
6170 return group;
1da177e4 6171}
6711cab4 6172
08069033
SS
6173static void init_numa_sched_groups_power(struct sched_group *group_head)
6174{
6175 struct sched_group *sg = group_head;
6176 int j;
6177
6178 if (!sg)
6179 return;
3a5c359a
AK
6180 do {
6181 for_each_cpu_mask(j, sg->cpumask) {
6182 struct sched_domain *sd;
08069033 6183
3a5c359a
AK
6184 sd = &per_cpu(phys_domains, j);
6185 if (j != first_cpu(sd->groups->cpumask)) {
6186 /*
6187 * Only add "power" once for each
6188 * physical package.
6189 */
6190 continue;
6191 }
08069033 6192
3a5c359a
AK
6193 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6194 }
6195 sg = sg->next;
6196 } while (sg != group_head);
08069033 6197}
1da177e4
LT
6198#endif
6199
a616058b 6200#ifdef CONFIG_NUMA
51888ca2
SV
6201/* Free memory allocated for various sched_group structures */
6202static void free_sched_groups(const cpumask_t *cpu_map)
6203{
a616058b 6204 int cpu, i;
51888ca2
SV
6205
6206 for_each_cpu_mask(cpu, *cpu_map) {
51888ca2
SV
6207 struct sched_group **sched_group_nodes
6208 = sched_group_nodes_bycpu[cpu];
6209
51888ca2
SV
6210 if (!sched_group_nodes)
6211 continue;
6212
6213 for (i = 0; i < MAX_NUMNODES; i++) {
6214 cpumask_t nodemask = node_to_cpumask(i);
6215 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6216
6217 cpus_and(nodemask, nodemask, *cpu_map);
6218 if (cpus_empty(nodemask))
6219 continue;
6220
6221 if (sg == NULL)
6222 continue;
6223 sg = sg->next;
6224next_sg:
6225 oldsg = sg;
6226 sg = sg->next;
6227 kfree(oldsg);
6228 if (oldsg != sched_group_nodes[i])
6229 goto next_sg;
6230 }
6231 kfree(sched_group_nodes);
6232 sched_group_nodes_bycpu[cpu] = NULL;
6233 }
51888ca2 6234}
a616058b
SS
6235#else
6236static void free_sched_groups(const cpumask_t *cpu_map)
6237{
6238}
6239#endif
51888ca2 6240
89c4710e
SS
6241/*
6242 * Initialize sched groups cpu_power.
6243 *
6244 * cpu_power indicates the capacity of sched group, which is used while
6245 * distributing the load between different sched groups in a sched domain.
6246 * Typically cpu_power for all the groups in a sched domain will be same unless
6247 * there are asymmetries in the topology. If there are asymmetries, group
6248 * having more cpu_power will pickup more load compared to the group having
6249 * less cpu_power.
6250 *
6251 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6252 * the maximum number of tasks a group can handle in the presence of other idle
6253 * or lightly loaded groups in the same sched domain.
6254 */
6255static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6256{
6257 struct sched_domain *child;
6258 struct sched_group *group;
6259
6260 WARN_ON(!sd || !sd->groups);
6261
6262 if (cpu != first_cpu(sd->groups->cpumask))
6263 return;
6264
6265 child = sd->child;
6266
5517d86b
ED
6267 sd->groups->__cpu_power = 0;
6268
89c4710e
SS
6269 /*
6270 * For perf policy, if the groups in child domain share resources
6271 * (for example cores sharing some portions of the cache hierarchy
6272 * or SMT), then set this domain groups cpu_power such that each group
6273 * can handle only one task, when there are other idle groups in the
6274 * same sched domain.
6275 */
6276 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6277 (child->flags &
6278 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
5517d86b 6279 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
89c4710e
SS
6280 return;
6281 }
6282
89c4710e
SS
6283 /*
6284 * add cpu_power of each child group to this groups cpu_power
6285 */
6286 group = child->groups;
6287 do {
5517d86b 6288 sg_inc_cpu_power(sd->groups, group->__cpu_power);
89c4710e
SS
6289 group = group->next;
6290 } while (group != child->groups);
6291}
6292
1da177e4 6293/*
1a20ff27
DG
6294 * Build sched domains for a given set of cpus and attach the sched domains
6295 * to the individual cpus
1da177e4 6296 */
51888ca2 6297static int build_sched_domains(const cpumask_t *cpu_map)
1da177e4
LT
6298{
6299 int i;
57d885fe 6300 struct root_domain *rd;
d1b55138
JH
6301#ifdef CONFIG_NUMA
6302 struct sched_group **sched_group_nodes = NULL;
6711cab4 6303 int sd_allnodes = 0;
d1b55138
JH
6304
6305 /*
6306 * Allocate the per-node list of sched groups
6307 */
5cf9f062 6308 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
41a2d6cf 6309 GFP_KERNEL);
d1b55138
JH
6310 if (!sched_group_nodes) {
6311 printk(KERN_WARNING "Can not alloc sched group node list\n");
51888ca2 6312 return -ENOMEM;
d1b55138
JH
6313 }
6314 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6315#endif
1da177e4 6316
57d885fe
GH
6317 rd = alloc_rootdomain(cpu_map);
6318 if (!rd) {
6319 printk(KERN_WARNING "Cannot alloc root domain\n");
6320 return -ENOMEM;
6321 }
6322
1da177e4 6323 /*
1a20ff27 6324 * Set up domains for cpus specified by the cpu_map.
1da177e4 6325 */
1a20ff27 6326 for_each_cpu_mask(i, *cpu_map) {
1da177e4
LT
6327 struct sched_domain *sd = NULL, *p;
6328 cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
6329
1a20ff27 6330 cpus_and(nodemask, nodemask, *cpu_map);
1da177e4
LT
6331
6332#ifdef CONFIG_NUMA
dd41f596
IM
6333 if (cpus_weight(*cpu_map) >
6334 SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
9c1cfda2
JH
6335 sd = &per_cpu(allnodes_domains, i);
6336 *sd = SD_ALLNODES_INIT;
6337 sd->span = *cpu_map;
6711cab4 6338 cpu_to_allnodes_group(i, cpu_map, &sd->groups);
9c1cfda2 6339 p = sd;
6711cab4 6340 sd_allnodes = 1;
9c1cfda2
JH
6341 } else
6342 p = NULL;
6343
1da177e4 6344 sd = &per_cpu(node_domains, i);
1da177e4 6345 *sd = SD_NODE_INIT;
9c1cfda2
JH
6346 sd->span = sched_domain_node_span(cpu_to_node(i));
6347 sd->parent = p;
1a848870
SS
6348 if (p)
6349 p->child = sd;
9c1cfda2 6350 cpus_and(sd->span, sd->span, *cpu_map);
1da177e4
LT
6351#endif
6352
6353 p = sd;
6354 sd = &per_cpu(phys_domains, i);
1da177e4
LT
6355 *sd = SD_CPU_INIT;
6356 sd->span = nodemask;
6357 sd->parent = p;
1a848870
SS
6358 if (p)
6359 p->child = sd;
6711cab4 6360 cpu_to_phys_group(i, cpu_map, &sd->groups);
1da177e4 6361
1e9f28fa
SS
6362#ifdef CONFIG_SCHED_MC
6363 p = sd;
6364 sd = &per_cpu(core_domains, i);
1e9f28fa
SS
6365 *sd = SD_MC_INIT;
6366 sd->span = cpu_coregroup_map(i);
6367 cpus_and(sd->span, sd->span, *cpu_map);
6368 sd->parent = p;
1a848870 6369 p->child = sd;
6711cab4 6370 cpu_to_core_group(i, cpu_map, &sd->groups);
1e9f28fa
SS
6371#endif
6372
1da177e4
LT
6373#ifdef CONFIG_SCHED_SMT
6374 p = sd;
6375 sd = &per_cpu(cpu_domains, i);
1da177e4 6376 *sd = SD_SIBLING_INIT;
d5a7430d 6377 sd->span = per_cpu(cpu_sibling_map, i);
1a20ff27 6378 cpus_and(sd->span, sd->span, *cpu_map);
1da177e4 6379 sd->parent = p;
1a848870 6380 p->child = sd;
6711cab4 6381 cpu_to_cpu_group(i, cpu_map, &sd->groups);
1da177e4
LT
6382#endif
6383 }
6384
6385#ifdef CONFIG_SCHED_SMT
6386 /* Set up CPU (sibling) groups */
9c1cfda2 6387 for_each_cpu_mask(i, *cpu_map) {
d5a7430d 6388 cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
1a20ff27 6389 cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
1da177e4
LT
6390 if (i != first_cpu(this_sibling_map))
6391 continue;
6392
dd41f596
IM
6393 init_sched_build_groups(this_sibling_map, cpu_map,
6394 &cpu_to_cpu_group);
1da177e4
LT
6395 }
6396#endif
6397
1e9f28fa
SS
6398#ifdef CONFIG_SCHED_MC
6399 /* Set up multi-core groups */
6400 for_each_cpu_mask(i, *cpu_map) {
6401 cpumask_t this_core_map = cpu_coregroup_map(i);
6402 cpus_and(this_core_map, this_core_map, *cpu_map);
6403 if (i != first_cpu(this_core_map))
6404 continue;
dd41f596
IM
6405 init_sched_build_groups(this_core_map, cpu_map,
6406 &cpu_to_core_group);
1e9f28fa
SS
6407 }
6408#endif
6409
1da177e4
LT
6410 /* Set up physical groups */
6411 for (i = 0; i < MAX_NUMNODES; i++) {
6412 cpumask_t nodemask = node_to_cpumask(i);
6413
1a20ff27 6414 cpus_and(nodemask, nodemask, *cpu_map);
1da177e4
LT
6415 if (cpus_empty(nodemask))
6416 continue;
6417
6711cab4 6418 init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
1da177e4
LT
6419 }
6420
6421#ifdef CONFIG_NUMA
6422 /* Set up node groups */
6711cab4 6423 if (sd_allnodes)
dd41f596
IM
6424 init_sched_build_groups(*cpu_map, cpu_map,
6425 &cpu_to_allnodes_group);
9c1cfda2
JH
6426
6427 for (i = 0; i < MAX_NUMNODES; i++) {
6428 /* Set up node groups */
6429 struct sched_group *sg, *prev;
6430 cpumask_t nodemask = node_to_cpumask(i);
6431 cpumask_t domainspan;
6432 cpumask_t covered = CPU_MASK_NONE;
6433 int j;
6434
6435 cpus_and(nodemask, nodemask, *cpu_map);
d1b55138
JH
6436 if (cpus_empty(nodemask)) {
6437 sched_group_nodes[i] = NULL;
9c1cfda2 6438 continue;
d1b55138 6439 }
9c1cfda2
JH
6440
6441 domainspan = sched_domain_node_span(i);
6442 cpus_and(domainspan, domainspan, *cpu_map);
6443
15f0b676 6444 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
51888ca2
SV
6445 if (!sg) {
6446 printk(KERN_WARNING "Can not alloc domain group for "
6447 "node %d\n", i);
6448 goto error;
6449 }
9c1cfda2
JH
6450 sched_group_nodes[i] = sg;
6451 for_each_cpu_mask(j, nodemask) {
6452 struct sched_domain *sd;
9761eea8 6453
9c1cfda2
JH
6454 sd = &per_cpu(node_domains, j);
6455 sd->groups = sg;
9c1cfda2 6456 }
5517d86b 6457 sg->__cpu_power = 0;
9c1cfda2 6458 sg->cpumask = nodemask;
51888ca2 6459 sg->next = sg;
9c1cfda2
JH
6460 cpus_or(covered, covered, nodemask);
6461 prev = sg;
6462
6463 for (j = 0; j < MAX_NUMNODES; j++) {
6464 cpumask_t tmp, notcovered;
6465 int n = (i + j) % MAX_NUMNODES;
6466
6467 cpus_complement(notcovered, covered);
6468 cpus_and(tmp, notcovered, *cpu_map);
6469 cpus_and(tmp, tmp, domainspan);
6470 if (cpus_empty(tmp))
6471 break;
6472
6473 nodemask = node_to_cpumask(n);
6474 cpus_and(tmp, tmp, nodemask);
6475 if (cpus_empty(tmp))
6476 continue;
6477
15f0b676
SV
6478 sg = kmalloc_node(sizeof(struct sched_group),
6479 GFP_KERNEL, i);
9c1cfda2
JH
6480 if (!sg) {
6481 printk(KERN_WARNING
6482 "Can not alloc domain group for node %d\n", j);
51888ca2 6483 goto error;
9c1cfda2 6484 }
5517d86b 6485 sg->__cpu_power = 0;
9c1cfda2 6486 sg->cpumask = tmp;
51888ca2 6487 sg->next = prev->next;
9c1cfda2
JH
6488 cpus_or(covered, covered, tmp);
6489 prev->next = sg;
6490 prev = sg;
6491 }
9c1cfda2 6492 }
1da177e4
LT
6493#endif
6494
6495 /* Calculate CPU power for physical packages and nodes */
5c45bf27 6496#ifdef CONFIG_SCHED_SMT
1a20ff27 6497 for_each_cpu_mask(i, *cpu_map) {
dd41f596
IM
6498 struct sched_domain *sd = &per_cpu(cpu_domains, i);
6499
89c4710e 6500 init_sched_groups_power(i, sd);
5c45bf27 6501 }
1da177e4 6502#endif
1e9f28fa 6503#ifdef CONFIG_SCHED_MC
5c45bf27 6504 for_each_cpu_mask(i, *cpu_map) {
dd41f596
IM
6505 struct sched_domain *sd = &per_cpu(core_domains, i);
6506
89c4710e 6507 init_sched_groups_power(i, sd);
5c45bf27
SS
6508 }
6509#endif
1e9f28fa 6510
5c45bf27 6511 for_each_cpu_mask(i, *cpu_map) {
dd41f596
IM
6512 struct sched_domain *sd = &per_cpu(phys_domains, i);
6513
89c4710e 6514 init_sched_groups_power(i, sd);
1da177e4
LT
6515 }
6516
9c1cfda2 6517#ifdef CONFIG_NUMA
08069033
SS
6518 for (i = 0; i < MAX_NUMNODES; i++)
6519 init_numa_sched_groups_power(sched_group_nodes[i]);
9c1cfda2 6520
6711cab4
SS
6521 if (sd_allnodes) {
6522 struct sched_group *sg;
f712c0c7 6523
6711cab4 6524 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
f712c0c7
SS
6525 init_numa_sched_groups_power(sg);
6526 }
9c1cfda2
JH
6527#endif
6528
1da177e4 6529 /* Attach the domains */
1a20ff27 6530 for_each_cpu_mask(i, *cpu_map) {
1da177e4
LT
6531 struct sched_domain *sd;
6532#ifdef CONFIG_SCHED_SMT
6533 sd = &per_cpu(cpu_domains, i);
1e9f28fa
SS
6534#elif defined(CONFIG_SCHED_MC)
6535 sd = &per_cpu(core_domains, i);
1da177e4
LT
6536#else
6537 sd = &per_cpu(phys_domains, i);
6538#endif
57d885fe 6539 cpu_attach_domain(sd, rd, i);
1da177e4 6540 }
51888ca2
SV
6541
6542 return 0;
6543
a616058b 6544#ifdef CONFIG_NUMA
51888ca2
SV
6545error:
6546 free_sched_groups(cpu_map);
6547 return -ENOMEM;
a616058b 6548#endif
1da177e4 6549}
029190c5
PJ
6550
6551static cpumask_t *doms_cur; /* current sched domains */
6552static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6553
6554/*
6555 * Special case: If a kmalloc of a doms_cur partition (array of
6556 * cpumask_t) fails, then fallback to a single sched domain,
6557 * as determined by the single cpumask_t fallback_doms.
6558 */
6559static cpumask_t fallback_doms;
6560
1a20ff27 6561/*
41a2d6cf 6562 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
029190c5
PJ
6563 * For now this just excludes isolated cpus, but could be used to
6564 * exclude other special cases in the future.
1a20ff27 6565 */
51888ca2 6566static int arch_init_sched_domains(const cpumask_t *cpu_map)
1a20ff27 6567{
7378547f
MM
6568 int err;
6569
029190c5
PJ
6570 ndoms_cur = 1;
6571 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6572 if (!doms_cur)
6573 doms_cur = &fallback_doms;
6574 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7378547f 6575 err = build_sched_domains(doms_cur);
6382bc90 6576 register_sched_domain_sysctl();
7378547f
MM
6577
6578 return err;
1a20ff27
DG
6579}
6580
6581static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
1da177e4 6582{
51888ca2 6583 free_sched_groups(cpu_map);
9c1cfda2 6584}
1da177e4 6585
1a20ff27
DG
6586/*
6587 * Detach sched domains from a group of cpus specified in cpu_map
6588 * These cpus will now be attached to the NULL domain
6589 */
858119e1 6590static void detach_destroy_domains(const cpumask_t *cpu_map)
1a20ff27
DG
6591{
6592 int i;
6593
6382bc90
MM
6594 unregister_sched_domain_sysctl();
6595
1a20ff27 6596 for_each_cpu_mask(i, *cpu_map)
57d885fe 6597 cpu_attach_domain(NULL, &def_root_domain, i);
1a20ff27
DG
6598 synchronize_sched();
6599 arch_destroy_sched_domains(cpu_map);
6600}
6601
029190c5
PJ
6602/*
6603 * Partition sched domains as specified by the 'ndoms_new'
41a2d6cf 6604 * cpumasks in the array doms_new[] of cpumasks. This compares
029190c5
PJ
6605 * doms_new[] to the current sched domain partitioning, doms_cur[].
6606 * It destroys each deleted domain and builds each new domain.
6607 *
6608 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
41a2d6cf
IM
6609 * The masks don't intersect (don't overlap.) We should setup one
6610 * sched domain for each mask. CPUs not in any of the cpumasks will
6611 * not be load balanced. If the same cpumask appears both in the
029190c5
PJ
6612 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6613 * it as it is.
6614 *
41a2d6cf
IM
6615 * The passed in 'doms_new' should be kmalloc'd. This routine takes
6616 * ownership of it and will kfree it when done with it. If the caller
029190c5
PJ
6617 * failed the kmalloc call, then it can pass in doms_new == NULL,
6618 * and partition_sched_domains() will fallback to the single partition
6619 * 'fallback_doms'.
6620 *
6621 * Call with hotplug lock held
6622 */
6623void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
6624{
6625 int i, j;
6626
a1835615
SV
6627 lock_doms_cur();
6628
7378547f
MM
6629 /* always unregister in case we don't destroy any domains */
6630 unregister_sched_domain_sysctl();
6631
029190c5
PJ
6632 if (doms_new == NULL) {
6633 ndoms_new = 1;
6634 doms_new = &fallback_doms;
6635 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
6636 }
6637
6638 /* Destroy deleted domains */
6639 for (i = 0; i < ndoms_cur; i++) {
6640 for (j = 0; j < ndoms_new; j++) {
6641 if (cpus_equal(doms_cur[i], doms_new[j]))
6642 goto match1;
6643 }
6644 /* no match - a current sched domain not in new doms_new[] */
6645 detach_destroy_domains(doms_cur + i);
6646match1:
6647 ;
6648 }
6649
6650 /* Build new domains */
6651 for (i = 0; i < ndoms_new; i++) {
6652 for (j = 0; j < ndoms_cur; j++) {
6653 if (cpus_equal(doms_new[i], doms_cur[j]))
6654 goto match2;
6655 }
6656 /* no match - add a new doms_new */
6657 build_sched_domains(doms_new + i);
6658match2:
6659 ;
6660 }
6661
6662 /* Remember the new sched domains */
6663 if (doms_cur != &fallback_doms)
6664 kfree(doms_cur);
6665 doms_cur = doms_new;
6666 ndoms_cur = ndoms_new;
7378547f
MM
6667
6668 register_sched_domain_sysctl();
a1835615
SV
6669
6670 unlock_doms_cur();
029190c5
PJ
6671}
6672
5c45bf27 6673#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6707de00 6674static int arch_reinit_sched_domains(void)
5c45bf27
SS
6675{
6676 int err;
6677
95402b38 6678 get_online_cpus();
5c45bf27
SS
6679 detach_destroy_domains(&cpu_online_map);
6680 err = arch_init_sched_domains(&cpu_online_map);
95402b38 6681 put_online_cpus();
5c45bf27
SS
6682
6683 return err;
6684}
6685
6686static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6687{
6688 int ret;
6689
6690 if (buf[0] != '0' && buf[0] != '1')
6691 return -EINVAL;
6692
6693 if (smt)
6694 sched_smt_power_savings = (buf[0] == '1');
6695 else
6696 sched_mc_power_savings = (buf[0] == '1');
6697
6698 ret = arch_reinit_sched_domains();
6699
6700 return ret ? ret : count;
6701}
6702
5c45bf27
SS
6703#ifdef CONFIG_SCHED_MC
6704static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
6705{
6706 return sprintf(page, "%u\n", sched_mc_power_savings);
6707}
48f24c4d
IM
6708static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
6709 const char *buf, size_t count)
5c45bf27
SS
6710{
6711 return sched_power_savings_store(buf, count, 0);
6712}
6707de00
AB
6713static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
6714 sched_mc_power_savings_store);
5c45bf27
SS
6715#endif
6716
6717#ifdef CONFIG_SCHED_SMT
6718static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
6719{
6720 return sprintf(page, "%u\n", sched_smt_power_savings);
6721}
48f24c4d
IM
6722static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
6723 const char *buf, size_t count)
5c45bf27
SS
6724{
6725 return sched_power_savings_store(buf, count, 1);
6726}
6707de00
AB
6727static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
6728 sched_smt_power_savings_store);
6729#endif
6730
6731int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6732{
6733 int err = 0;
6734
6735#ifdef CONFIG_SCHED_SMT
6736 if (smt_capable())
6737 err = sysfs_create_file(&cls->kset.kobj,
6738 &attr_sched_smt_power_savings.attr);
6739#endif
6740#ifdef CONFIG_SCHED_MC
6741 if (!err && mc_capable())
6742 err = sysfs_create_file(&cls->kset.kobj,
6743 &attr_sched_mc_power_savings.attr);
6744#endif
6745 return err;
6746}
5c45bf27
SS
6747#endif
6748
1da177e4 6749/*
41a2d6cf 6750 * Force a reinitialization of the sched domains hierarchy. The domains
1da177e4 6751 * and groups cannot be updated in place without racing with the balancing
41c7ce9a 6752 * code, so we temporarily attach all running cpus to the NULL domain
1da177e4
LT
6753 * which will prevent rebalancing while the sched domains are recalculated.
6754 */
6755static int update_sched_domains(struct notifier_block *nfb,
6756 unsigned long action, void *hcpu)
6757{
1da177e4
LT
6758 switch (action) {
6759 case CPU_UP_PREPARE:
8bb78442 6760 case CPU_UP_PREPARE_FROZEN:
1da177e4 6761 case CPU_DOWN_PREPARE:
8bb78442 6762 case CPU_DOWN_PREPARE_FROZEN:
1a20ff27 6763 detach_destroy_domains(&cpu_online_map);
1da177e4
LT
6764 return NOTIFY_OK;
6765
6766 case CPU_UP_CANCELED:
8bb78442 6767 case CPU_UP_CANCELED_FROZEN:
1da177e4 6768 case CPU_DOWN_FAILED:
8bb78442 6769 case CPU_DOWN_FAILED_FROZEN:
1da177e4 6770 case CPU_ONLINE:
8bb78442 6771 case CPU_ONLINE_FROZEN:
1da177e4 6772 case CPU_DEAD:
8bb78442 6773 case CPU_DEAD_FROZEN:
1da177e4
LT
6774 /*
6775 * Fall through and re-initialise the domains.
6776 */
6777 break;
6778 default:
6779 return NOTIFY_DONE;
6780 }
6781
6782 /* The hotplug lock is already held by cpu_up/cpu_down */
1a20ff27 6783 arch_init_sched_domains(&cpu_online_map);
1da177e4
LT
6784
6785 return NOTIFY_OK;
6786}
1da177e4
LT
6787
6788void __init sched_init_smp(void)
6789{
5c1e1767
NP
6790 cpumask_t non_isolated_cpus;
6791
95402b38 6792 get_online_cpus();
1a20ff27 6793 arch_init_sched_domains(&cpu_online_map);
e5e5673f 6794 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
5c1e1767
NP
6795 if (cpus_empty(non_isolated_cpus))
6796 cpu_set(smp_processor_id(), non_isolated_cpus);
95402b38 6797 put_online_cpus();
1da177e4
LT
6798 /* XXX: Theoretical race here - CPU may be hotplugged now */
6799 hotcpu_notifier(update_sched_domains, 0);
5c1e1767
NP
6800
6801 /* Move init over to a non-isolated CPU */
6802 if (set_cpus_allowed(current, non_isolated_cpus) < 0)
6803 BUG();
19978ca6 6804 sched_init_granularity();
6b2d7700
SV
6805
6806#ifdef CONFIG_FAIR_GROUP_SCHED
6807 if (nr_cpu_ids == 1)
6808 return;
6809
6810 lb_monitor_task = kthread_create(load_balance_monitor, NULL,
6811 "group_balance");
6812 if (!IS_ERR(lb_monitor_task)) {
6813 lb_monitor_task->flags |= PF_NOFREEZE;
6814 wake_up_process(lb_monitor_task);
6815 } else {
6816 printk(KERN_ERR "Could not create load balance monitor thread"
6817 "(error = %ld) \n", PTR_ERR(lb_monitor_task));
6818 }
6819#endif
1da177e4
LT
6820}
6821#else
6822void __init sched_init_smp(void)
6823{
19978ca6 6824 sched_init_granularity();
1da177e4
LT
6825}
6826#endif /* CONFIG_SMP */
6827
6828int in_sched_functions(unsigned long addr)
6829{
1da177e4
LT
6830 return in_lock_functions(addr) ||
6831 (addr >= (unsigned long)__sched_text_start
6832 && addr < (unsigned long)__sched_text_end);
6833}
6834
a9957449 6835static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
dd41f596
IM
6836{
6837 cfs_rq->tasks_timeline = RB_ROOT;
dd41f596
IM
6838#ifdef CONFIG_FAIR_GROUP_SCHED
6839 cfs_rq->rq = rq;
6840#endif
67e9fb2a 6841 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
dd41f596
IM
6842}
6843
1da177e4
LT
6844void __init sched_init(void)
6845{
476f3534 6846 int highest_cpu = 0;
dd41f596
IM
6847 int i, j;
6848
57d885fe
GH
6849#ifdef CONFIG_SMP
6850 init_defrootdomain();
6851#endif
6852
0a945022 6853 for_each_possible_cpu(i) {
dd41f596 6854 struct rt_prio_array *array;
70b97a7f 6855 struct rq *rq;
1da177e4
LT
6856
6857 rq = cpu_rq(i);
6858 spin_lock_init(&rq->lock);
fcb99371 6859 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7897986b 6860 rq->nr_running = 0;
dd41f596
IM
6861 rq->clock = 1;
6862 init_cfs_rq(&rq->cfs, rq);
6863#ifdef CONFIG_FAIR_GROUP_SCHED
6864 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
3a252015
IM
6865 {
6866 struct cfs_rq *cfs_rq = &per_cpu(init_cfs_rq, i);
6867 struct sched_entity *se =
6868 &per_cpu(init_sched_entity, i);
6869
6870 init_cfs_rq_p[i] = cfs_rq;
6871 init_cfs_rq(cfs_rq, rq);
4cf86d77 6872 cfs_rq->tg = &init_task_group;
3a252015 6873 list_add(&cfs_rq->leaf_cfs_rq_list,
29f59db3
SV
6874 &rq->leaf_cfs_rq_list);
6875
3a252015
IM
6876 init_sched_entity_p[i] = se;
6877 se->cfs_rq = &rq->cfs;
6878 se->my_q = cfs_rq;
4cf86d77 6879 se->load.weight = init_task_group_load;
9b5b7751 6880 se->load.inv_weight =
4cf86d77 6881 div64_64(1ULL<<32, init_task_group_load);
3a252015
IM
6882 se->parent = NULL;
6883 }
4cf86d77 6884 init_task_group.shares = init_task_group_load;
dd41f596 6885#endif
1da177e4 6886
dd41f596
IM
6887 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6888 rq->cpu_load[j] = 0;
1da177e4 6889#ifdef CONFIG_SMP
41c7ce9a 6890 rq->sd = NULL;
57d885fe
GH
6891 rq->rd = NULL;
6892 rq_attach_root(rq, &def_root_domain);
1da177e4 6893 rq->active_balance = 0;
dd41f596 6894 rq->next_balance = jiffies;
1da177e4 6895 rq->push_cpu = 0;
0a2966b4 6896 rq->cpu = i;
1da177e4
LT
6897 rq->migration_thread = NULL;
6898 INIT_LIST_HEAD(&rq->migration_queue);
764a9d6f 6899 rq->rt.highest_prio = MAX_RT_PRIO;
a22d7fc1 6900 rq->rt.overloaded = 0;
1da177e4
LT
6901#endif
6902 atomic_set(&rq->nr_iowait, 0);
6903
dd41f596
IM
6904 array = &rq->rt.active;
6905 for (j = 0; j < MAX_RT_PRIO; j++) {
6906 INIT_LIST_HEAD(array->queue + j);
6907 __clear_bit(j, array->bitmap);
1da177e4 6908 }
476f3534 6909 highest_cpu = i;
dd41f596
IM
6910 /* delimiter for bitsearch: */
6911 __set_bit(MAX_RT_PRIO, array->bitmap);
1da177e4
LT
6912 }
6913
2dd73a4f 6914 set_load_weight(&init_task);
b50f60ce 6915
e107be36
AK
6916#ifdef CONFIG_PREEMPT_NOTIFIERS
6917 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6918#endif
6919
c9819f45 6920#ifdef CONFIG_SMP
476f3534 6921 nr_cpu_ids = highest_cpu + 1;
c9819f45
CL
6922 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
6923#endif
6924
b50f60ce
HC
6925#ifdef CONFIG_RT_MUTEXES
6926 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
6927#endif
6928
1da177e4
LT
6929 /*
6930 * The boot idle thread does lazy MMU switching as well:
6931 */
6932 atomic_inc(&init_mm.mm_count);
6933 enter_lazy_tlb(&init_mm, current);
6934
6935 /*
6936 * Make us the idle thread. Technically, schedule() should not be
6937 * called from this thread, however somewhere below it might be,
6938 * but because we are the idle thread, we just pick up running again
6939 * when this runqueue becomes "idle".
6940 */
6941 init_idle(current, smp_processor_id());
dd41f596
IM
6942 /*
6943 * During early bootup we pretend to be a normal task:
6944 */
6945 current->sched_class = &fair_sched_class;
1da177e4
LT
6946}
6947
6948#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6949void __might_sleep(char *file, int line)
6950{
48f24c4d 6951#ifdef in_atomic
1da177e4
LT
6952 static unsigned long prev_jiffy; /* ratelimiting */
6953
6954 if ((in_atomic() || irqs_disabled()) &&
6955 system_state == SYSTEM_RUNNING && !oops_in_progress) {
6956 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6957 return;
6958 prev_jiffy = jiffies;
91368d73 6959 printk(KERN_ERR "BUG: sleeping function called from invalid"
1da177e4
LT
6960 " context at %s:%d\n", file, line);
6961 printk("in_atomic():%d, irqs_disabled():%d\n",
6962 in_atomic(), irqs_disabled());
a4c410f0 6963 debug_show_held_locks(current);
3117df04
IM
6964 if (irqs_disabled())
6965 print_irqtrace_events(current);
1da177e4
LT
6966 dump_stack();
6967 }
6968#endif
6969}
6970EXPORT_SYMBOL(__might_sleep);
6971#endif
6972
6973#ifdef CONFIG_MAGIC_SYSRQ
3a5e4dc1
AK
6974static void normalize_task(struct rq *rq, struct task_struct *p)
6975{
6976 int on_rq;
6977 update_rq_clock(rq);
6978 on_rq = p->se.on_rq;
6979 if (on_rq)
6980 deactivate_task(rq, p, 0);
6981 __setscheduler(rq, p, SCHED_NORMAL, 0);
6982 if (on_rq) {
6983 activate_task(rq, p, 0);
6984 resched_task(rq->curr);
6985 }
6986}
6987
1da177e4
LT
6988void normalize_rt_tasks(void)
6989{
a0f98a1c 6990 struct task_struct *g, *p;
1da177e4 6991 unsigned long flags;
70b97a7f 6992 struct rq *rq;
1da177e4
LT
6993
6994 read_lock_irq(&tasklist_lock);
a0f98a1c 6995 do_each_thread(g, p) {
178be793
IM
6996 /*
6997 * Only normalize user tasks:
6998 */
6999 if (!p->mm)
7000 continue;
7001
6cfb0d5d 7002 p->se.exec_start = 0;
6cfb0d5d 7003#ifdef CONFIG_SCHEDSTATS
dd41f596 7004 p->se.wait_start = 0;
dd41f596 7005 p->se.sleep_start = 0;
dd41f596 7006 p->se.block_start = 0;
6cfb0d5d 7007#endif
dd41f596
IM
7008 task_rq(p)->clock = 0;
7009
7010 if (!rt_task(p)) {
7011 /*
7012 * Renice negative nice level userspace
7013 * tasks back to 0:
7014 */
7015 if (TASK_NICE(p) < 0 && p->mm)
7016 set_user_nice(p, 0);
1da177e4 7017 continue;
dd41f596 7018 }
1da177e4 7019
b29739f9
IM
7020 spin_lock_irqsave(&p->pi_lock, flags);
7021 rq = __task_rq_lock(p);
1da177e4 7022
178be793 7023 normalize_task(rq, p);
3a5e4dc1 7024
b29739f9
IM
7025 __task_rq_unlock(rq);
7026 spin_unlock_irqrestore(&p->pi_lock, flags);
a0f98a1c
IM
7027 } while_each_thread(g, p);
7028
1da177e4
LT
7029 read_unlock_irq(&tasklist_lock);
7030}
7031
7032#endif /* CONFIG_MAGIC_SYSRQ */
1df5c10a
LT
7033
7034#ifdef CONFIG_IA64
7035/*
7036 * These functions are only useful for the IA64 MCA handling.
7037 *
7038 * They can only be called when the whole system has been
7039 * stopped - every CPU needs to be quiescent, and no scheduling
7040 * activity can take place. Using them for anything else would
7041 * be a serious bug, and as a result, they aren't even visible
7042 * under any other configuration.
7043 */
7044
7045/**
7046 * curr_task - return the current task for a given cpu.
7047 * @cpu: the processor in question.
7048 *
7049 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7050 */
36c8b586 7051struct task_struct *curr_task(int cpu)
1df5c10a
LT
7052{
7053 return cpu_curr(cpu);
7054}
7055
7056/**
7057 * set_curr_task - set the current task for a given cpu.
7058 * @cpu: the processor in question.
7059 * @p: the task pointer to set.
7060 *
7061 * Description: This function must only be used when non-maskable interrupts
41a2d6cf
IM
7062 * are serviced on a separate stack. It allows the architecture to switch the
7063 * notion of the current task on a cpu in a non-blocking manner. This function
1df5c10a
LT
7064 * must be called with all CPU's synchronized, and interrupts disabled, the
7065 * and caller must save the original value of the current task (see
7066 * curr_task() above) and restore that value before reenabling interrupts and
7067 * re-starting the system.
7068 *
7069 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7070 */
36c8b586 7071void set_curr_task(int cpu, struct task_struct *p)
1df5c10a
LT
7072{
7073 cpu_curr(cpu) = p;
7074}
7075
7076#endif
29f59db3
SV
7077
7078#ifdef CONFIG_FAIR_GROUP_SCHED
7079
6b2d7700
SV
7080#ifdef CONFIG_SMP
7081/*
7082 * distribute shares of all task groups among their schedulable entities,
7083 * to reflect load distrbution across cpus.
7084 */
7085static int rebalance_shares(struct sched_domain *sd, int this_cpu)
7086{
7087 struct cfs_rq *cfs_rq;
7088 struct rq *rq = cpu_rq(this_cpu);
7089 cpumask_t sdspan = sd->span;
7090 int balanced = 1;
7091
7092 /* Walk thr' all the task groups that we have */
7093 for_each_leaf_cfs_rq(rq, cfs_rq) {
7094 int i;
7095 unsigned long total_load = 0, total_shares;
7096 struct task_group *tg = cfs_rq->tg;
7097
7098 /* Gather total task load of this group across cpus */
7099 for_each_cpu_mask(i, sdspan)
7100 total_load += tg->cfs_rq[i]->load.weight;
7101
0eab9146 7102 /* Nothing to do if this group has no load */
6b2d7700
SV
7103 if (!total_load)
7104 continue;
7105
7106 /*
7107 * tg->shares represents the number of cpu shares the task group
7108 * is eligible to hold on a single cpu. On N cpus, it is
7109 * eligible to hold (N * tg->shares) number of cpu shares.
7110 */
7111 total_shares = tg->shares * cpus_weight(sdspan);
7112
7113 /*
7114 * redistribute total_shares across cpus as per the task load
7115 * distribution.
7116 */
7117 for_each_cpu_mask(i, sdspan) {
7118 unsigned long local_load, local_shares;
7119
7120 local_load = tg->cfs_rq[i]->load.weight;
7121 local_shares = (local_load * total_shares) / total_load;
7122 if (!local_shares)
7123 local_shares = MIN_GROUP_SHARES;
7124 if (local_shares == tg->se[i]->load.weight)
7125 continue;
7126
7127 spin_lock_irq(&cpu_rq(i)->lock);
7128 set_se_shares(tg->se[i], local_shares);
7129 spin_unlock_irq(&cpu_rq(i)->lock);
7130 balanced = 0;
7131 }
7132 }
7133
7134 return balanced;
7135}
7136
7137/*
7138 * How frequently should we rebalance_shares() across cpus?
7139 *
7140 * The more frequently we rebalance shares, the more accurate is the fairness
7141 * of cpu bandwidth distribution between task groups. However higher frequency
7142 * also implies increased scheduling overhead.
7143 *
7144 * sysctl_sched_min_bal_int_shares represents the minimum interval between
7145 * consecutive calls to rebalance_shares() in the same sched domain.
7146 *
7147 * sysctl_sched_max_bal_int_shares represents the maximum interval between
7148 * consecutive calls to rebalance_shares() in the same sched domain.
7149 *
7150 * These settings allows for the appropriate tradeoff between accuracy of
7151 * fairness and the associated overhead.
7152 *
7153 */
7154
7155/* default: 8ms, units: milliseconds */
7156const_debug unsigned int sysctl_sched_min_bal_int_shares = 8;
7157
7158/* default: 128ms, units: milliseconds */
7159const_debug unsigned int sysctl_sched_max_bal_int_shares = 128;
7160
7161/* kernel thread that runs rebalance_shares() periodically */
7162static int load_balance_monitor(void *unused)
7163{
7164 unsigned int timeout = sysctl_sched_min_bal_int_shares;
7165 struct sched_param schedparm;
7166 int ret;
7167
7168 /*
7169 * We don't want this thread's execution to be limited by the shares
7170 * assigned to default group (init_task_group). Hence make it run
7171 * as a SCHED_RR RT task at the lowest priority.
7172 */
7173 schedparm.sched_priority = 1;
7174 ret = sched_setscheduler(current, SCHED_RR, &schedparm);
7175 if (ret)
7176 printk(KERN_ERR "Couldn't set SCHED_RR policy for load balance"
7177 " monitor thread (error = %d) \n", ret);
7178
7179 while (!kthread_should_stop()) {
7180 int i, cpu, balanced = 1;
7181
7182 /* Prevent cpus going down or coming up */
86ef5c9a 7183 get_online_cpus();
6b2d7700
SV
7184 /* lockout changes to doms_cur[] array */
7185 lock_doms_cur();
7186 /*
7187 * Enter a rcu read-side critical section to safely walk rq->sd
7188 * chain on various cpus and to walk task group list
7189 * (rq->leaf_cfs_rq_list) in rebalance_shares().
7190 */
7191 rcu_read_lock();
7192
7193 for (i = 0; i < ndoms_cur; i++) {
7194 cpumask_t cpumap = doms_cur[i];
7195 struct sched_domain *sd = NULL, *sd_prev = NULL;
7196
7197 cpu = first_cpu(cpumap);
7198
7199 /* Find the highest domain at which to balance shares */
7200 for_each_domain(cpu, sd) {
7201 if (!(sd->flags & SD_LOAD_BALANCE))
7202 continue;
7203 sd_prev = sd;
7204 }
7205
7206 sd = sd_prev;
7207 /* sd == NULL? No load balance reqd in this domain */
7208 if (!sd)
7209 continue;
7210
7211 balanced &= rebalance_shares(sd, cpu);
7212 }
7213
7214 rcu_read_unlock();
7215
7216 unlock_doms_cur();
86ef5c9a 7217 put_online_cpus();
6b2d7700
SV
7218
7219 if (!balanced)
7220 timeout = sysctl_sched_min_bal_int_shares;
7221 else if (timeout < sysctl_sched_max_bal_int_shares)
7222 timeout *= 2;
7223
7224 msleep_interruptible(timeout);
7225 }
7226
7227 return 0;
7228}
7229#endif /* CONFIG_SMP */
7230
29f59db3 7231/* allocate runqueue etc for a new task group */
4cf86d77 7232struct task_group *sched_create_group(void)
29f59db3 7233{
4cf86d77 7234 struct task_group *tg;
29f59db3
SV
7235 struct cfs_rq *cfs_rq;
7236 struct sched_entity *se;
9b5b7751 7237 struct rq *rq;
29f59db3
SV
7238 int i;
7239
29f59db3
SV
7240 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7241 if (!tg)
7242 return ERR_PTR(-ENOMEM);
7243
9b5b7751 7244 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * NR_CPUS, GFP_KERNEL);
29f59db3
SV
7245 if (!tg->cfs_rq)
7246 goto err;
9b5b7751 7247 tg->se = kzalloc(sizeof(se) * NR_CPUS, GFP_KERNEL);
29f59db3
SV
7248 if (!tg->se)
7249 goto err;
7250
7251 for_each_possible_cpu(i) {
9b5b7751 7252 rq = cpu_rq(i);
29f59db3
SV
7253
7254 cfs_rq = kmalloc_node(sizeof(struct cfs_rq), GFP_KERNEL,
7255 cpu_to_node(i));
7256 if (!cfs_rq)
7257 goto err;
7258
7259 se = kmalloc_node(sizeof(struct sched_entity), GFP_KERNEL,
7260 cpu_to_node(i));
7261 if (!se)
7262 goto err;
7263
7264 memset(cfs_rq, 0, sizeof(struct cfs_rq));
7265 memset(se, 0, sizeof(struct sched_entity));
7266
7267 tg->cfs_rq[i] = cfs_rq;
7268 init_cfs_rq(cfs_rq, rq);
7269 cfs_rq->tg = tg;
29f59db3
SV
7270
7271 tg->se[i] = se;
7272 se->cfs_rq = &rq->cfs;
7273 se->my_q = cfs_rq;
7274 se->load.weight = NICE_0_LOAD;
7275 se->load.inv_weight = div64_64(1ULL<<32, NICE_0_LOAD);
7276 se->parent = NULL;
7277 }
7278
ec2c507f
SV
7279 tg->shares = NICE_0_LOAD;
7280
7281 lock_task_group_list();
9b5b7751
SV
7282 for_each_possible_cpu(i) {
7283 rq = cpu_rq(i);
7284 cfs_rq = tg->cfs_rq[i];
7285 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7286 }
ec2c507f 7287 unlock_task_group_list();
29f59db3 7288
9b5b7751 7289 return tg;
29f59db3
SV
7290
7291err:
7292 for_each_possible_cpu(i) {
a65914b3 7293 if (tg->cfs_rq)
29f59db3 7294 kfree(tg->cfs_rq[i]);
a65914b3 7295 if (tg->se)
29f59db3
SV
7296 kfree(tg->se[i]);
7297 }
a65914b3
IM
7298 kfree(tg->cfs_rq);
7299 kfree(tg->se);
7300 kfree(tg);
29f59db3
SV
7301
7302 return ERR_PTR(-ENOMEM);
7303}
7304
9b5b7751
SV
7305/* rcu callback to free various structures associated with a task group */
7306static void free_sched_group(struct rcu_head *rhp)
29f59db3 7307{
ae8393e5
SV
7308 struct task_group *tg = container_of(rhp, struct task_group, rcu);
7309 struct cfs_rq *cfs_rq;
29f59db3
SV
7310 struct sched_entity *se;
7311 int i;
7312
29f59db3
SV
7313 /* now it should be safe to free those cfs_rqs */
7314 for_each_possible_cpu(i) {
7315 cfs_rq = tg->cfs_rq[i];
7316 kfree(cfs_rq);
7317
7318 se = tg->se[i];
7319 kfree(se);
7320 }
7321
7322 kfree(tg->cfs_rq);
7323 kfree(tg->se);
7324 kfree(tg);
7325}
7326
9b5b7751 7327/* Destroy runqueue etc associated with a task group */
4cf86d77 7328void sched_destroy_group(struct task_group *tg)
29f59db3 7329{
7bae49d4 7330 struct cfs_rq *cfs_rq = NULL;
9b5b7751 7331 int i;
29f59db3 7332
ec2c507f 7333 lock_task_group_list();
9b5b7751
SV
7334 for_each_possible_cpu(i) {
7335 cfs_rq = tg->cfs_rq[i];
7336 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7337 }
ec2c507f 7338 unlock_task_group_list();
9b5b7751 7339
7bae49d4 7340 BUG_ON(!cfs_rq);
9b5b7751
SV
7341
7342 /* wait for possible concurrent references to cfs_rqs complete */
ae8393e5 7343 call_rcu(&tg->rcu, free_sched_group);
29f59db3
SV
7344}
7345
9b5b7751 7346/* change task's runqueue when it moves between groups.
3a252015
IM
7347 * The caller of this function should have put the task in its new group
7348 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7349 * reflect its new group.
9b5b7751
SV
7350 */
7351void sched_move_task(struct task_struct *tsk)
29f59db3
SV
7352{
7353 int on_rq, running;
7354 unsigned long flags;
7355 struct rq *rq;
7356
7357 rq = task_rq_lock(tsk, &flags);
7358
dae51f56 7359 if (tsk->sched_class != &fair_sched_class) {
ce96b5ac 7360 set_task_cfs_rq(tsk, task_cpu(tsk));
29f59db3 7361 goto done;
dae51f56 7362 }
29f59db3
SV
7363
7364 update_rq_clock(rq);
7365
051a1d1a 7366 running = task_current(rq, tsk);
29f59db3
SV
7367 on_rq = tsk->se.on_rq;
7368
83b699ed 7369 if (on_rq) {
29f59db3 7370 dequeue_task(rq, tsk, 0);
83b699ed
SV
7371 if (unlikely(running))
7372 tsk->sched_class->put_prev_task(rq, tsk);
7373 }
29f59db3 7374
ce96b5ac 7375 set_task_cfs_rq(tsk, task_cpu(tsk));
29f59db3 7376
83b699ed
SV
7377 if (on_rq) {
7378 if (unlikely(running))
7379 tsk->sched_class->set_curr_task(rq);
7074badb 7380 enqueue_task(rq, tsk, 0);
83b699ed 7381 }
29f59db3
SV
7382
7383done:
7384 task_rq_unlock(rq, &flags);
7385}
7386
6b2d7700 7387/* rq->lock to be locked by caller */
29f59db3
SV
7388static void set_se_shares(struct sched_entity *se, unsigned long shares)
7389{
7390 struct cfs_rq *cfs_rq = se->cfs_rq;
7391 struct rq *rq = cfs_rq->rq;
7392 int on_rq;
7393
6b2d7700
SV
7394 if (!shares)
7395 shares = MIN_GROUP_SHARES;
29f59db3
SV
7396
7397 on_rq = se->on_rq;
6b2d7700 7398 if (on_rq) {
29f59db3 7399 dequeue_entity(cfs_rq, se, 0);
6b2d7700
SV
7400 dec_cpu_load(rq, se->load.weight);
7401 }
29f59db3
SV
7402
7403 se->load.weight = shares;
7404 se->load.inv_weight = div64_64((1ULL<<32), shares);
7405
6b2d7700 7406 if (on_rq) {
29f59db3 7407 enqueue_entity(cfs_rq, se, 0);
6b2d7700
SV
7408 inc_cpu_load(rq, se->load.weight);
7409 }
29f59db3
SV
7410}
7411
4cf86d77 7412int sched_group_set_shares(struct task_group *tg, unsigned long shares)
29f59db3
SV
7413{
7414 int i;
6b2d7700
SV
7415 struct cfs_rq *cfs_rq;
7416 struct rq *rq;
c61935fd 7417
ec2c507f 7418 lock_task_group_list();
9b5b7751 7419 if (tg->shares == shares)
5cb350ba 7420 goto done;
29f59db3 7421
6b2d7700
SV
7422 if (shares < MIN_GROUP_SHARES)
7423 shares = MIN_GROUP_SHARES;
7424
7425 /*
7426 * Prevent any load balance activity (rebalance_shares,
7427 * load_balance_fair) from referring to this group first,
7428 * by taking it off the rq->leaf_cfs_rq_list on each cpu.
7429 */
7430 for_each_possible_cpu(i) {
7431 cfs_rq = tg->cfs_rq[i];
7432 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
7433 }
7434
7435 /* wait for any ongoing reference to this group to finish */
7436 synchronize_sched();
7437
7438 /*
7439 * Now we are free to modify the group's share on each cpu
7440 * w/o tripping rebalance_share or load_balance_fair.
7441 */
9b5b7751 7442 tg->shares = shares;
6b2d7700
SV
7443 for_each_possible_cpu(i) {
7444 spin_lock_irq(&cpu_rq(i)->lock);
9b5b7751 7445 set_se_shares(tg->se[i], shares);
6b2d7700
SV
7446 spin_unlock_irq(&cpu_rq(i)->lock);
7447 }
29f59db3 7448
6b2d7700
SV
7449 /*
7450 * Enable load balance activity on this group, by inserting it back on
7451 * each cpu's rq->leaf_cfs_rq_list.
7452 */
7453 for_each_possible_cpu(i) {
7454 rq = cpu_rq(i);
7455 cfs_rq = tg->cfs_rq[i];
7456 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7457 }
5cb350ba 7458done:
ec2c507f 7459 unlock_task_group_list();
9b5b7751 7460 return 0;
29f59db3
SV
7461}
7462
5cb350ba
DG
7463unsigned long sched_group_shares(struct task_group *tg)
7464{
7465 return tg->shares;
7466}
7467
3a252015 7468#endif /* CONFIG_FAIR_GROUP_SCHED */
68318b8e
SV
7469
7470#ifdef CONFIG_FAIR_CGROUP_SCHED
7471
7472/* return corresponding task_group object of a cgroup */
2b01dfe3 7473static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
68318b8e 7474{
2b01dfe3
PM
7475 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7476 struct task_group, css);
68318b8e
SV
7477}
7478
7479static struct cgroup_subsys_state *
2b01dfe3 7480cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
68318b8e
SV
7481{
7482 struct task_group *tg;
7483
2b01dfe3 7484 if (!cgrp->parent) {
68318b8e 7485 /* This is early initialization for the top cgroup */
2b01dfe3 7486 init_task_group.css.cgroup = cgrp;
68318b8e
SV
7487 return &init_task_group.css;
7488 }
7489
7490 /* we support only 1-level deep hierarchical scheduler atm */
2b01dfe3 7491 if (cgrp->parent->parent)
68318b8e
SV
7492 return ERR_PTR(-EINVAL);
7493
7494 tg = sched_create_group();
7495 if (IS_ERR(tg))
7496 return ERR_PTR(-ENOMEM);
7497
7498 /* Bind the cgroup to task_group object we just created */
2b01dfe3 7499 tg->css.cgroup = cgrp;
68318b8e
SV
7500
7501 return &tg->css;
7502}
7503
41a2d6cf
IM
7504static void
7505cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
68318b8e 7506{
2b01dfe3 7507 struct task_group *tg = cgroup_tg(cgrp);
68318b8e
SV
7508
7509 sched_destroy_group(tg);
7510}
7511
41a2d6cf
IM
7512static int
7513cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7514 struct task_struct *tsk)
68318b8e
SV
7515{
7516 /* We don't support RT-tasks being in separate groups */
7517 if (tsk->sched_class != &fair_sched_class)
7518 return -EINVAL;
7519
7520 return 0;
7521}
7522
7523static void
2b01dfe3 7524cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
68318b8e
SV
7525 struct cgroup *old_cont, struct task_struct *tsk)
7526{
7527 sched_move_task(tsk);
7528}
7529
2b01dfe3
PM
7530static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7531 u64 shareval)
68318b8e 7532{
2b01dfe3 7533 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
68318b8e
SV
7534}
7535
2b01dfe3 7536static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
68318b8e 7537{
2b01dfe3 7538 struct task_group *tg = cgroup_tg(cgrp);
68318b8e
SV
7539
7540 return (u64) tg->shares;
7541}
7542
fe5c7cc2
PM
7543static struct cftype cpu_files[] = {
7544 {
7545 .name = "shares",
7546 .read_uint = cpu_shares_read_uint,
7547 .write_uint = cpu_shares_write_uint,
7548 },
68318b8e
SV
7549};
7550
7551static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7552{
fe5c7cc2 7553 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
68318b8e
SV
7554}
7555
7556struct cgroup_subsys cpu_cgroup_subsys = {
38605cae
IM
7557 .name = "cpu",
7558 .create = cpu_cgroup_create,
7559 .destroy = cpu_cgroup_destroy,
7560 .can_attach = cpu_cgroup_can_attach,
7561 .attach = cpu_cgroup_attach,
7562 .populate = cpu_cgroup_populate,
7563 .subsys_id = cpu_cgroup_subsys_id,
68318b8e
SV
7564 .early_init = 1,
7565};
7566
7567#endif /* CONFIG_FAIR_CGROUP_SCHED */
d842de87
SV
7568
7569#ifdef CONFIG_CGROUP_CPUACCT
7570
7571/*
7572 * CPU accounting code for task groups.
7573 *
7574 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7575 * (balbir@in.ibm.com).
7576 */
7577
7578/* track cpu usage of a group of tasks */
7579struct cpuacct {
7580 struct cgroup_subsys_state css;
7581 /* cpuusage holds pointer to a u64-type object on every cpu */
7582 u64 *cpuusage;
7583};
7584
7585struct cgroup_subsys cpuacct_subsys;
7586
7587/* return cpu accounting group corresponding to this container */
7588static inline struct cpuacct *cgroup_ca(struct cgroup *cont)
7589{
7590 return container_of(cgroup_subsys_state(cont, cpuacct_subsys_id),
7591 struct cpuacct, css);
7592}
7593
7594/* return cpu accounting group to which this task belongs */
7595static inline struct cpuacct *task_ca(struct task_struct *tsk)
7596{
7597 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7598 struct cpuacct, css);
7599}
7600
7601/* create a new cpu accounting group */
7602static struct cgroup_subsys_state *cpuacct_create(
7603 struct cgroup_subsys *ss, struct cgroup *cont)
7604{
7605 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7606
7607 if (!ca)
7608 return ERR_PTR(-ENOMEM);
7609
7610 ca->cpuusage = alloc_percpu(u64);
7611 if (!ca->cpuusage) {
7612 kfree(ca);
7613 return ERR_PTR(-ENOMEM);
7614 }
7615
7616 return &ca->css;
7617}
7618
7619/* destroy an existing cpu accounting group */
41a2d6cf
IM
7620static void
7621cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
d842de87
SV
7622{
7623 struct cpuacct *ca = cgroup_ca(cont);
7624
7625 free_percpu(ca->cpuusage);
7626 kfree(ca);
7627}
7628
7629/* return total cpu usage (in nanoseconds) of a group */
7630static u64 cpuusage_read(struct cgroup *cont, struct cftype *cft)
7631{
7632 struct cpuacct *ca = cgroup_ca(cont);
7633 u64 totalcpuusage = 0;
7634 int i;
7635
7636 for_each_possible_cpu(i) {
7637 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
7638
7639 /*
7640 * Take rq->lock to make 64-bit addition safe on 32-bit
7641 * platforms.
7642 */
7643 spin_lock_irq(&cpu_rq(i)->lock);
7644 totalcpuusage += *cpuusage;
7645 spin_unlock_irq(&cpu_rq(i)->lock);
7646 }
7647
7648 return totalcpuusage;
7649}
7650
7651static struct cftype files[] = {
7652 {
7653 .name = "usage",
7654 .read_uint = cpuusage_read,
7655 },
7656};
7657
7658static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7659{
7660 return cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
7661}
7662
7663/*
7664 * charge this task's execution time to its accounting group.
7665 *
7666 * called with rq->lock held.
7667 */
7668static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
7669{
7670 struct cpuacct *ca;
7671
7672 if (!cpuacct_subsys.active)
7673 return;
7674
7675 ca = task_ca(tsk);
7676 if (ca) {
7677 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
7678
7679 *cpuusage += cputime;
7680 }
7681}
7682
7683struct cgroup_subsys cpuacct_subsys = {
7684 .name = "cpuacct",
7685 .create = cpuacct_create,
7686 .destroy = cpuacct_destroy,
7687 .populate = cpuacct_populate,
7688 .subsys_id = cpuacct_subsys_id,
7689};
7690#endif /* CONFIG_CGROUP_CPUACCT */