4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
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
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
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
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
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/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
83 #include "sched_cpupri.h"
84 #include "workqueue_sched.h"
85 #include "sched_autogroup.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
95 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
104 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
109 * Helpers for converting nanosecond timing to jiffy resolution
111 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
113 #define NICE_0_LOAD SCHED_LOAD_SCALE
114 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
117 * These are the 'tuning knobs' of the scheduler:
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
122 #define DEF_TIMESLICE (100 * HZ / 1000)
125 * single value that denotes runtime == period, ie unlimited time.
127 #define RUNTIME_INF ((u64)~0ULL)
129 static inline int rt_policy(int policy)
131 if (policy == SCHED_FIFO || policy == SCHED_RR)
136 static inline int task_has_rt_policy(struct task_struct *p)
138 return rt_policy(p->policy);
142 * This is the priority-queue data structure of the RT scheduling class:
144 struct rt_prio_array {
145 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
146 struct list_head queue[MAX_RT_PRIO];
149 struct rt_bandwidth {
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock;
154 struct hrtimer rt_period_timer;
157 static struct rt_bandwidth def_rt_bandwidth;
159 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
161 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
163 struct rt_bandwidth *rt_b =
164 container_of(timer, struct rt_bandwidth, rt_period_timer);
170 now = hrtimer_cb_get_time(timer);
171 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
176 idle = do_sched_rt_period_timer(rt_b, overrun);
179 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
183 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
185 rt_b->rt_period = ns_to_ktime(period);
186 rt_b->rt_runtime = runtime;
188 raw_spin_lock_init(&rt_b->rt_runtime_lock);
190 hrtimer_init(&rt_b->rt_period_timer,
191 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
192 rt_b->rt_period_timer.function = sched_rt_period_timer;
195 static inline int rt_bandwidth_enabled(void)
197 return sysctl_sched_rt_runtime >= 0;
200 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
203 ktime_t soft, hard, now;
206 if (hrtimer_active(period_timer))
209 now = hrtimer_cb_get_time(period_timer);
210 hrtimer_forward(period_timer, now, period);
212 soft = hrtimer_get_softexpires(period_timer);
213 hard = hrtimer_get_expires(period_timer);
214 delta = ktime_to_ns(ktime_sub(hard, soft));
215 __hrtimer_start_range_ns(period_timer, soft, delta,
216 HRTIMER_MODE_ABS_PINNED, 0);
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 raw_spin_lock(&rt_b->rt_runtime_lock);
229 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
230 raw_spin_unlock(&rt_b->rt_runtime_lock);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
236 hrtimer_cancel(&rt_b->rt_period_timer);
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex);
246 #ifdef CONFIG_CGROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups);
254 struct cfs_bandwidth {
255 #ifdef CONFIG_CFS_BANDWIDTH
259 s64 hierarchal_quota;
262 int idle, timer_active;
263 struct hrtimer period_timer, slack_timer;
264 struct list_head throttled_cfs_rq;
267 int nr_periods, nr_throttled;
272 /* task group related information */
274 struct cgroup_subsys_state css;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
283 atomic_t load_weight;
286 #ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity **rt_se;
288 struct rt_rq **rt_rq;
290 struct rt_bandwidth rt_bandwidth;
294 struct list_head list;
296 struct task_group *parent;
297 struct list_head siblings;
298 struct list_head children;
300 #ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup *autogroup;
304 struct cfs_bandwidth cfs_bandwidth;
307 /* task_group_lock serializes the addition/removal of task groups */
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES (1UL << 1)
323 #define MAX_SHARES (1UL << 18)
325 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group root_task_group;
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load;
338 unsigned long nr_running, h_nr_running;
343 u64 min_vruntime_copy;
346 struct rb_root tasks_timeline;
347 struct rb_node *rb_leftmost;
349 struct list_head tasks;
350 struct list_head *balance_iterator;
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
356 struct sched_entity *curr, *next, *last, *skip;
358 #ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
374 struct list_head leaf_cfs_rq_list;
375 struct task_group *tg; /* group that "owns" this runqueue */
379 * the part of load.weight contributed by tasks
381 unsigned long task_weight;
384 * h_load = weight * f(tg)
386 * Where f(tg) is the recursive weight fraction assigned to
389 unsigned long h_load;
392 * Maintaining per-cpu shares distribution for group scheduling
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
400 u64 load_stamp, load_last, load_unacc_exec_time;
402 unsigned long load_contribution;
404 #ifdef CONFIG_CFS_BANDWIDTH
407 s64 runtime_remaining;
409 u64 throttled_timestamp;
410 int throttled, throttle_count;
411 struct list_head throttled_list;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 #ifdef CONFIG_CFS_BANDWIDTH
418 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
420 return &tg->cfs_bandwidth;
423 static inline u64 default_cfs_period(void);
424 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
427 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
429 struct cfs_bandwidth *cfs_b =
430 container_of(timer, struct cfs_bandwidth, slack_timer);
431 do_sched_cfs_slack_timer(cfs_b);
433 return HRTIMER_NORESTART;
436 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
438 struct cfs_bandwidth *cfs_b =
439 container_of(timer, struct cfs_bandwidth, period_timer);
445 now = hrtimer_cb_get_time(timer);
446 overrun = hrtimer_forward(timer, now, cfs_b->period);
451 idle = do_sched_cfs_period_timer(cfs_b, overrun);
454 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
457 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
459 raw_spin_lock_init(&cfs_b->lock);
461 cfs_b->quota = RUNTIME_INF;
462 cfs_b->period = ns_to_ktime(default_cfs_period());
464 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
465 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
466 cfs_b->period_timer.function = sched_cfs_period_timer;
467 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 cfs_b->slack_timer.function = sched_cfs_slack_timer;
471 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
473 cfs_rq->runtime_enabled = 0;
474 INIT_LIST_HEAD(&cfs_rq->throttled_list);
477 /* requires cfs_b->lock, may release to reprogram timer */
478 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
486 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
487 raw_spin_unlock(&cfs_b->lock);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b->period_timer);
491 raw_spin_lock(&cfs_b->lock);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b->timer_active)
497 cfs_b->timer_active = 1;
498 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
501 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
503 hrtimer_cancel(&cfs_b->period_timer);
504 hrtimer_cancel(&cfs_b->slack_timer);
507 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
508 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
509 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
511 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
515 #endif /* CONFIG_CFS_BANDWIDTH */
516 #endif /* CONFIG_FAIR_GROUP_SCHED */
518 /* Real-Time classes' related field in a runqueue: */
520 struct rt_prio_array active;
521 unsigned long rt_nr_running;
522 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
524 int curr; /* highest queued rt task prio */
526 int next; /* next highest */
531 unsigned long rt_nr_migratory;
532 unsigned long rt_nr_total;
534 struct plist_head pushable_tasks;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted;
546 struct list_head leaf_rt_rq_list;
547 struct task_group *tg;
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
566 cpumask_var_t online;
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
572 cpumask_var_t rto_mask;
573 struct cpupri cpupri;
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
580 static struct root_domain def_root_domain;
582 #endif /* CONFIG_SMP */
585 * This is the main, per-CPU runqueue data structure.
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
599 unsigned long nr_running;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
602 unsigned long last_load_update_tick;
605 unsigned char nohz_balance_kick;
607 int skip_clock_update;
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load;
611 unsigned long nr_load_updates;
617 #ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list;
621 #ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list;
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
631 unsigned long nr_uninterruptible;
633 struct task_struct *curr, *idle, *stop;
634 unsigned long next_balance;
635 struct mm_struct *prev_mm;
643 struct root_domain *rd;
644 struct sched_domain *sd;
646 unsigned long cpu_power;
648 unsigned char idle_balance;
649 /* For active balancing */
653 struct cpu_stop_work active_balance_work;
654 /* cpu of this runqueue: */
664 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
667 #ifdef CONFIG_PARAVIRT
670 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq;
674 /* calc_load related fields */
675 unsigned long calc_load_update;
676 long calc_load_active;
678 #ifdef CONFIG_SCHED_HRTICK
680 int hrtick_csd_pending;
681 struct call_single_data hrtick_csd;
683 struct hrtimer hrtick_timer;
686 #ifdef CONFIG_SCHEDSTATS
688 struct sched_info rq_sched_info;
689 unsigned long long rq_cpu_time;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
692 /* sys_sched_yield() stats */
693 unsigned int yld_count;
695 /* schedule() stats */
696 unsigned int sched_switch;
697 unsigned int sched_count;
698 unsigned int sched_goidle;
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count;
702 unsigned int ttwu_local;
706 struct llist_head wake_list;
710 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
713 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
715 static inline int cpu_of(struct rq *rq)
724 #define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
735 #define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
738 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739 #define this_rq() (&__get_cpu_var(runqueues))
740 #define task_rq(p) cpu_rq(task_cpu(p))
741 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742 #define raw_rq() (&__raw_get_cpu_var(runqueues))
744 #ifdef CONFIG_CGROUP_SCHED
747 * Return the group to which this tasks belongs.
749 * We cannot use task_subsys_state() and friends because the cgroup
750 * subsystem changes that value before the cgroup_subsys::attach() method
751 * is called, therefore we cannot pin it and might observe the wrong value.
753 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
754 * core changes this before calling sched_move_task().
756 * Instead we use a 'copy' which is updated from sched_move_task() while
757 * holding both task_struct::pi_lock and rq::lock.
759 static inline struct task_group *task_group(struct task_struct *p)
761 return p->sched_task_group;
764 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
767 #ifdef CONFIG_FAIR_GROUP_SCHED
768 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
769 p->se.parent = task_group(p)->se[cpu];
772 #ifdef CONFIG_RT_GROUP_SCHED
773 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
774 p->rt.parent = task_group(p)->rt_se[cpu];
778 #else /* CONFIG_CGROUP_SCHED */
780 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
781 static inline struct task_group *task_group(struct task_struct *p)
786 #endif /* CONFIG_CGROUP_SCHED */
788 static void update_rq_clock_task(struct rq *rq, s64 delta);
790 static void update_rq_clock(struct rq *rq)
794 if (rq->skip_clock_update > 0)
797 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
799 update_rq_clock_task(rq, delta);
803 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
805 #ifdef CONFIG_SCHED_DEBUG
806 # define const_debug __read_mostly
808 # define const_debug static const
812 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
813 * @cpu: the processor in question.
815 * This interface allows printk to be called with the runqueue lock
816 * held and know whether or not it is OK to wake up the klogd.
818 int runqueue_is_locked(int cpu)
820 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
824 * Debugging: various feature bits
827 #define SCHED_FEAT(name, enabled) \
828 __SCHED_FEAT_##name ,
831 #include "sched_features.h"
836 #define SCHED_FEAT(name, enabled) \
837 (1UL << __SCHED_FEAT_##name) * enabled |
839 const_debug unsigned int sysctl_sched_features =
840 #include "sched_features.h"
845 #ifdef CONFIG_SCHED_DEBUG
846 #define SCHED_FEAT(name, enabled) \
849 static __read_mostly char *sched_feat_names[] = {
850 #include "sched_features.h"
856 static int sched_feat_show(struct seq_file *m, void *v)
860 for (i = 0; sched_feat_names[i]; i++) {
861 if (!(sysctl_sched_features & (1UL << i)))
863 seq_printf(m, "%s ", sched_feat_names[i]);
871 sched_feat_write(struct file *filp, const char __user *ubuf,
872 size_t cnt, loff_t *ppos)
882 if (copy_from_user(&buf, ubuf, cnt))
888 if (strncmp(cmp, "NO_", 3) == 0) {
893 for (i = 0; sched_feat_names[i]; i++) {
894 if (strcmp(cmp, sched_feat_names[i]) == 0) {
896 sysctl_sched_features &= ~(1UL << i);
898 sysctl_sched_features |= (1UL << i);
903 if (!sched_feat_names[i])
911 static int sched_feat_open(struct inode *inode, struct file *filp)
913 return single_open(filp, sched_feat_show, NULL);
916 static const struct file_operations sched_feat_fops = {
917 .open = sched_feat_open,
918 .write = sched_feat_write,
921 .release = single_release,
924 static __init int sched_init_debug(void)
926 debugfs_create_file("sched_features", 0644, NULL, NULL,
931 late_initcall(sched_init_debug);
935 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
938 * Number of tasks to iterate in a single balance run.
939 * Limited because this is done with IRQs disabled.
941 const_debug unsigned int sysctl_sched_nr_migrate = 32;
944 * period over which we average the RT time consumption, measured
949 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
952 * period over which we measure -rt task cpu usage in us.
955 unsigned int sysctl_sched_rt_period = 1000000;
957 static __read_mostly int scheduler_running;
960 * part of the period that we allow rt tasks to run in us.
963 int sysctl_sched_rt_runtime = 950000;
965 static inline u64 global_rt_period(void)
967 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
970 static inline u64 global_rt_runtime(void)
972 if (sysctl_sched_rt_runtime < 0)
975 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
978 #ifndef prepare_arch_switch
979 # define prepare_arch_switch(next) do { } while (0)
981 #ifndef finish_arch_switch
982 # define finish_arch_switch(prev) do { } while (0)
985 static inline int task_current(struct rq *rq, struct task_struct *p)
987 return rq->curr == p;
990 static inline int task_running(struct rq *rq, struct task_struct *p)
995 return task_current(rq, p);
999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1000 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1004 * We can optimise this out completely for !SMP, because the
1005 * SMP rebalancing from interrupt is the only thing that cares
1012 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1016 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1017 * We must ensure this doesn't happen until the switch is completely
1020 * Pairs with the control dependency and rmb in try_to_wake_up().
1025 #ifdef CONFIG_DEBUG_SPINLOCK
1026 /* this is a valid case when another task releases the spinlock */
1027 rq->lock.owner = current;
1030 * If we are tracking spinlock dependencies then we have to
1031 * fix up the runqueue lock - which gets 'carried over' from
1032 * prev into current:
1034 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1036 raw_spin_unlock_irq(&rq->lock);
1039 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1040 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1044 * We can optimise this out completely for !SMP, because the
1045 * SMP rebalancing from interrupt is the only thing that cares
1050 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1051 raw_spin_unlock_irq(&rq->lock);
1053 raw_spin_unlock(&rq->lock);
1057 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1061 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1062 * We must ensure this doesn't happen until the switch is completely
1068 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1072 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1075 * __task_rq_lock - lock the rq @p resides on.
1077 static inline struct rq *__task_rq_lock(struct task_struct *p)
1078 __acquires(rq->lock)
1082 lockdep_assert_held(&p->pi_lock);
1086 raw_spin_lock(&rq->lock);
1087 if (likely(rq == task_rq(p)))
1089 raw_spin_unlock(&rq->lock);
1094 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1096 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1097 __acquires(p->pi_lock)
1098 __acquires(rq->lock)
1103 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1105 raw_spin_lock(&rq->lock);
1106 if (likely(rq == task_rq(p)))
1108 raw_spin_unlock(&rq->lock);
1109 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1113 static void __task_rq_unlock(struct rq *rq)
1114 __releases(rq->lock)
1116 raw_spin_unlock(&rq->lock);
1120 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1121 __releases(rq->lock)
1122 __releases(p->pi_lock)
1124 raw_spin_unlock(&rq->lock);
1125 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1129 * this_rq_lock - lock this runqueue and disable interrupts.
1131 static struct rq *this_rq_lock(void)
1132 __acquires(rq->lock)
1136 local_irq_disable();
1138 raw_spin_lock(&rq->lock);
1143 #ifdef CONFIG_SCHED_HRTICK
1145 * Use HR-timers to deliver accurate preemption points.
1147 * Its all a bit involved since we cannot program an hrt while holding the
1148 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1151 * When we get rescheduled we reprogram the hrtick_timer outside of the
1157 * - enabled by features
1158 * - hrtimer is actually high res
1160 static inline int hrtick_enabled(struct rq *rq)
1162 if (!sched_feat(HRTICK))
1164 if (!cpu_active(cpu_of(rq)))
1166 return hrtimer_is_hres_active(&rq->hrtick_timer);
1169 static void hrtick_clear(struct rq *rq)
1171 if (hrtimer_active(&rq->hrtick_timer))
1172 hrtimer_cancel(&rq->hrtick_timer);
1176 * High-resolution timer tick.
1177 * Runs from hardirq context with interrupts disabled.
1179 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1181 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1183 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1185 raw_spin_lock(&rq->lock);
1186 update_rq_clock(rq);
1187 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1188 raw_spin_unlock(&rq->lock);
1190 return HRTIMER_NORESTART;
1195 * called from hardirq (IPI) context
1197 static void __hrtick_start(void *arg)
1199 struct rq *rq = arg;
1201 raw_spin_lock(&rq->lock);
1202 hrtimer_restart(&rq->hrtick_timer);
1203 rq->hrtick_csd_pending = 0;
1204 raw_spin_unlock(&rq->lock);
1208 * Called to set the hrtick timer state.
1210 * called with rq->lock held and irqs disabled
1212 static void hrtick_start(struct rq *rq, u64 delay)
1214 struct hrtimer *timer = &rq->hrtick_timer;
1215 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1217 hrtimer_set_expires(timer, time);
1219 if (rq == this_rq()) {
1220 hrtimer_restart(timer);
1221 } else if (!rq->hrtick_csd_pending) {
1222 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1223 rq->hrtick_csd_pending = 1;
1228 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1230 int cpu = (int)(long)hcpu;
1233 case CPU_UP_CANCELED:
1234 case CPU_UP_CANCELED_FROZEN:
1235 case CPU_DOWN_PREPARE:
1236 case CPU_DOWN_PREPARE_FROZEN:
1238 case CPU_DEAD_FROZEN:
1239 hrtick_clear(cpu_rq(cpu));
1246 static __init void init_hrtick(void)
1248 hotcpu_notifier(hotplug_hrtick, 0);
1252 * Called to set the hrtick timer state.
1254 * called with rq->lock held and irqs disabled
1256 static void hrtick_start(struct rq *rq, u64 delay)
1258 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1259 HRTIMER_MODE_REL_PINNED, 0);
1262 static inline void init_hrtick(void)
1265 #endif /* CONFIG_SMP */
1267 static void init_rq_hrtick(struct rq *rq)
1270 rq->hrtick_csd_pending = 0;
1272 rq->hrtick_csd.flags = 0;
1273 rq->hrtick_csd.func = __hrtick_start;
1274 rq->hrtick_csd.info = rq;
1277 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1278 rq->hrtick_timer.function = hrtick;
1280 #else /* CONFIG_SCHED_HRTICK */
1281 static inline void hrtick_clear(struct rq *rq)
1285 static inline void init_rq_hrtick(struct rq *rq)
1289 static inline void init_hrtick(void)
1292 #endif /* CONFIG_SCHED_HRTICK */
1295 * resched_task - mark a task 'to be rescheduled now'.
1297 * On UP this means the setting of the need_resched flag, on SMP it
1298 * might also involve a cross-CPU call to trigger the scheduler on
1303 #ifndef tsk_is_polling
1304 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1307 static void resched_task(struct task_struct *p)
1311 assert_raw_spin_locked(&task_rq(p)->lock);
1313 if (test_tsk_need_resched(p))
1316 set_tsk_need_resched(p);
1319 if (cpu == smp_processor_id())
1322 /* NEED_RESCHED must be visible before we test polling */
1324 if (!tsk_is_polling(p))
1325 smp_send_reschedule(cpu);
1328 static void resched_cpu(int cpu)
1330 struct rq *rq = cpu_rq(cpu);
1331 unsigned long flags;
1333 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1335 resched_task(cpu_curr(cpu));
1336 raw_spin_unlock_irqrestore(&rq->lock, flags);
1341 * In the semi idle case, use the nearest busy cpu for migrating timers
1342 * from an idle cpu. This is good for power-savings.
1344 * We don't do similar optimization for completely idle system, as
1345 * selecting an idle cpu will add more delays to the timers than intended
1346 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1348 int get_nohz_timer_target(void)
1350 int cpu = smp_processor_id();
1352 struct sched_domain *sd;
1355 for_each_domain(cpu, sd) {
1356 for_each_cpu(i, sched_domain_span(sd)) {
1368 * When add_timer_on() enqueues a timer into the timer wheel of an
1369 * idle CPU then this timer might expire before the next timer event
1370 * which is scheduled to wake up that CPU. In case of a completely
1371 * idle system the next event might even be infinite time into the
1372 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1373 * leaves the inner idle loop so the newly added timer is taken into
1374 * account when the CPU goes back to idle and evaluates the timer
1375 * wheel for the next timer event.
1377 void wake_up_idle_cpu(int cpu)
1379 struct rq *rq = cpu_rq(cpu);
1381 if (cpu == smp_processor_id())
1385 * This is safe, as this function is called with the timer
1386 * wheel base lock of (cpu) held. When the CPU is on the way
1387 * to idle and has not yet set rq->curr to idle then it will
1388 * be serialized on the timer wheel base lock and take the new
1389 * timer into account automatically.
1391 if (rq->curr != rq->idle)
1395 * We can set TIF_RESCHED on the idle task of the other CPU
1396 * lockless. The worst case is that the other CPU runs the
1397 * idle task through an additional NOOP schedule()
1399 set_tsk_need_resched(rq->idle);
1401 /* NEED_RESCHED must be visible before we test polling */
1403 if (!tsk_is_polling(rq->idle))
1404 smp_send_reschedule(cpu);
1407 static inline bool got_nohz_idle_kick(void)
1409 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
1412 #else /* CONFIG_NO_HZ */
1414 static inline bool got_nohz_idle_kick(void)
1419 #endif /* CONFIG_NO_HZ */
1421 static u64 sched_avg_period(void)
1423 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1426 static void sched_avg_update(struct rq *rq)
1428 s64 period = sched_avg_period();
1430 while ((s64)(rq->clock - rq->age_stamp) > period) {
1432 * Inline assembly required to prevent the compiler
1433 * optimising this loop into a divmod call.
1434 * See __iter_div_u64_rem() for another example of this.
1436 asm("" : "+rm" (rq->age_stamp));
1437 rq->age_stamp += period;
1442 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1444 rq->rt_avg += rt_delta;
1445 sched_avg_update(rq);
1448 #else /* !CONFIG_SMP */
1449 static void resched_task(struct task_struct *p)
1451 assert_raw_spin_locked(&task_rq(p)->lock);
1452 set_tsk_need_resched(p);
1455 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1459 static void sched_avg_update(struct rq *rq)
1462 #endif /* CONFIG_SMP */
1464 #if BITS_PER_LONG == 32
1465 # define WMULT_CONST (~0UL)
1467 # define WMULT_CONST (1UL << 32)
1470 #define WMULT_SHIFT 32
1473 * Shift right and round:
1475 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1478 * delta *= weight / lw
1480 static unsigned long
1481 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1482 struct load_weight *lw)
1487 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1488 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1489 * 2^SCHED_LOAD_RESOLUTION.
1491 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1492 tmp = (u64)delta_exec * scale_load_down(weight);
1494 tmp = (u64)delta_exec;
1496 if (!lw->inv_weight) {
1497 unsigned long w = scale_load_down(lw->weight);
1499 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1501 else if (unlikely(!w))
1502 lw->inv_weight = WMULT_CONST;
1504 lw->inv_weight = WMULT_CONST / w;
1508 * Check whether we'd overflow the 64-bit multiplication:
1510 if (unlikely(tmp > WMULT_CONST))
1511 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1514 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1516 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1519 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1525 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1531 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1538 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1539 * of tasks with abnormal "nice" values across CPUs the contribution that
1540 * each task makes to its run queue's load is weighted according to its
1541 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1542 * scaled version of the new time slice allocation that they receive on time
1546 #define WEIGHT_IDLEPRIO 3
1547 #define WMULT_IDLEPRIO 1431655765
1550 * Nice levels are multiplicative, with a gentle 10% change for every
1551 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1552 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1553 * that remained on nice 0.
1555 * The "10% effect" is relative and cumulative: from _any_ nice level,
1556 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1557 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1558 * If a task goes up by ~10% and another task goes down by ~10% then
1559 * the relative distance between them is ~25%.)
1561 static const int prio_to_weight[40] = {
1562 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1563 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1564 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1565 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1566 /* 0 */ 1024, 820, 655, 526, 423,
1567 /* 5 */ 335, 272, 215, 172, 137,
1568 /* 10 */ 110, 87, 70, 56, 45,
1569 /* 15 */ 36, 29, 23, 18, 15,
1573 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1575 * In cases where the weight does not change often, we can use the
1576 * precalculated inverse to speed up arithmetics by turning divisions
1577 * into multiplications:
1579 static const u32 prio_to_wmult[40] = {
1580 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1581 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1582 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1583 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1584 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1585 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1586 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1587 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1590 /* Time spent by the tasks of the cpu accounting group executing in ... */
1591 enum cpuacct_stat_index {
1592 CPUACCT_STAT_USER, /* ... user mode */
1593 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1595 CPUACCT_STAT_NSTATS,
1598 #ifdef CONFIG_CGROUP_CPUACCT
1599 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1600 static void cpuacct_update_stats(struct task_struct *tsk,
1601 enum cpuacct_stat_index idx, cputime_t val);
1603 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1604 static inline void cpuacct_update_stats(struct task_struct *tsk,
1605 enum cpuacct_stat_index idx, cputime_t val) {}
1608 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1610 update_load_add(&rq->load, load);
1613 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1615 update_load_sub(&rq->load, load);
1618 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1619 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1620 typedef int (*tg_visitor)(struct task_group *, void *);
1623 * Iterate task_group tree rooted at *from, calling @down when first entering a
1624 * node and @up when leaving it for the final time.
1626 * Caller must hold rcu_lock or sufficient equivalent.
1628 static int walk_tg_tree_from(struct task_group *from,
1629 tg_visitor down, tg_visitor up, void *data)
1631 struct task_group *parent, *child;
1637 ret = (*down)(parent, data);
1640 list_for_each_entry_rcu(child, &parent->children, siblings) {
1647 ret = (*up)(parent, data);
1648 if (ret || parent == from)
1652 parent = parent->parent;
1660 * Iterate the full tree, calling @down when first entering a node and @up when
1661 * leaving it for the final time.
1663 * Caller must hold rcu_lock or sufficient equivalent.
1666 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1668 return walk_tg_tree_from(&root_task_group, down, up, data);
1671 static int tg_nop(struct task_group *tg, void *data)
1678 /* Used instead of source_load when we know the type == 0 */
1679 static unsigned long weighted_cpuload(const int cpu)
1681 return cpu_rq(cpu)->load.weight;
1685 * Return a low guess at the load of a migration-source cpu weighted
1686 * according to the scheduling class and "nice" value.
1688 * We want to under-estimate the load of migration sources, to
1689 * balance conservatively.
1691 static unsigned long source_load(int cpu, int type)
1693 struct rq *rq = cpu_rq(cpu);
1694 unsigned long total = weighted_cpuload(cpu);
1696 if (type == 0 || !sched_feat(LB_BIAS))
1699 return min(rq->cpu_load[type-1], total);
1703 * Return a high guess at the load of a migration-target cpu weighted
1704 * according to the scheduling class and "nice" value.
1706 static unsigned long target_load(int cpu, int type)
1708 struct rq *rq = cpu_rq(cpu);
1709 unsigned long total = weighted_cpuload(cpu);
1711 if (type == 0 || !sched_feat(LB_BIAS))
1714 return max(rq->cpu_load[type-1], total);
1717 static unsigned long power_of(int cpu)
1719 return cpu_rq(cpu)->cpu_power;
1722 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1724 static unsigned long cpu_avg_load_per_task(int cpu)
1726 struct rq *rq = cpu_rq(cpu);
1727 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1730 return rq->load.weight / nr_running;
1735 #ifdef CONFIG_PREEMPT
1737 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1740 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1741 * way at the expense of forcing extra atomic operations in all
1742 * invocations. This assures that the double_lock is acquired using the
1743 * same underlying policy as the spinlock_t on this architecture, which
1744 * reduces latency compared to the unfair variant below. However, it
1745 * also adds more overhead and therefore may reduce throughput.
1747 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1748 __releases(this_rq->lock)
1749 __acquires(busiest->lock)
1750 __acquires(this_rq->lock)
1752 raw_spin_unlock(&this_rq->lock);
1753 double_rq_lock(this_rq, busiest);
1760 * Unfair double_lock_balance: Optimizes throughput at the expense of
1761 * latency by eliminating extra atomic operations when the locks are
1762 * already in proper order on entry. This favors lower cpu-ids and will
1763 * grant the double lock to lower cpus over higher ids under contention,
1764 * regardless of entry order into the function.
1766 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1767 __releases(this_rq->lock)
1768 __acquires(busiest->lock)
1769 __acquires(this_rq->lock)
1773 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1774 if (busiest < this_rq) {
1775 raw_spin_unlock(&this_rq->lock);
1776 raw_spin_lock(&busiest->lock);
1777 raw_spin_lock_nested(&this_rq->lock,
1778 SINGLE_DEPTH_NESTING);
1781 raw_spin_lock_nested(&busiest->lock,
1782 SINGLE_DEPTH_NESTING);
1787 #endif /* CONFIG_PREEMPT */
1790 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1792 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1794 if (unlikely(!irqs_disabled())) {
1795 /* printk() doesn't work good under rq->lock */
1796 raw_spin_unlock(&this_rq->lock);
1800 return _double_lock_balance(this_rq, busiest);
1803 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1804 __releases(busiest->lock)
1806 raw_spin_unlock(&busiest->lock);
1807 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1811 * double_rq_lock - safely lock two runqueues
1813 * Note this does not disable interrupts like task_rq_lock,
1814 * you need to do so manually before calling.
1816 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1817 __acquires(rq1->lock)
1818 __acquires(rq2->lock)
1820 BUG_ON(!irqs_disabled());
1822 raw_spin_lock(&rq1->lock);
1823 __acquire(rq2->lock); /* Fake it out ;) */
1826 raw_spin_lock(&rq1->lock);
1827 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1829 raw_spin_lock(&rq2->lock);
1830 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1836 * double_rq_unlock - safely unlock two runqueues
1838 * Note this does not restore interrupts like task_rq_unlock,
1839 * you need to do so manually after calling.
1841 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1842 __releases(rq1->lock)
1843 __releases(rq2->lock)
1845 raw_spin_unlock(&rq1->lock);
1847 raw_spin_unlock(&rq2->lock);
1849 __release(rq2->lock);
1852 #else /* CONFIG_SMP */
1855 * double_rq_lock - safely lock two runqueues
1857 * Note this does not disable interrupts like task_rq_lock,
1858 * you need to do so manually before calling.
1860 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1861 __acquires(rq1->lock)
1862 __acquires(rq2->lock)
1864 BUG_ON(!irqs_disabled());
1866 raw_spin_lock(&rq1->lock);
1867 __acquire(rq2->lock); /* Fake it out ;) */
1871 * double_rq_unlock - safely unlock two runqueues
1873 * Note this does not restore interrupts like task_rq_unlock,
1874 * you need to do so manually after calling.
1876 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1877 __releases(rq1->lock)
1878 __releases(rq2->lock)
1881 raw_spin_unlock(&rq1->lock);
1882 __release(rq2->lock);
1887 static void update_sysctl(void);
1888 static int get_update_sysctl_factor(void);
1889 static void update_idle_cpu_load(struct rq *this_rq);
1891 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1893 set_task_rq(p, cpu);
1896 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1897 * successfully executed on another CPU. We must ensure that updates of
1898 * per-task data have been completed by this moment.
1901 task_thread_info(p)->cpu = cpu;
1905 static const struct sched_class rt_sched_class;
1907 #define sched_class_highest (&stop_sched_class)
1908 #define for_each_class(class) \
1909 for (class = sched_class_highest; class; class = class->next)
1911 #include "sched_stats.h"
1913 static void inc_nr_running(struct rq *rq)
1918 static void dec_nr_running(struct rq *rq)
1923 static void set_load_weight(struct task_struct *p)
1925 int prio = p->static_prio - MAX_RT_PRIO;
1926 struct load_weight *load = &p->se.load;
1929 * SCHED_IDLE tasks get minimal weight:
1931 if (p->policy == SCHED_IDLE) {
1932 load->weight = scale_load(WEIGHT_IDLEPRIO);
1933 load->inv_weight = WMULT_IDLEPRIO;
1937 load->weight = scale_load(prio_to_weight[prio]);
1938 load->inv_weight = prio_to_wmult[prio];
1941 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1943 update_rq_clock(rq);
1944 sched_info_queued(p);
1945 p->sched_class->enqueue_task(rq, p, flags);
1948 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1950 update_rq_clock(rq);
1951 sched_info_dequeued(p);
1952 p->sched_class->dequeue_task(rq, p, flags);
1956 * activate_task - move a task to the runqueue.
1958 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1963 enqueue_task(rq, p, flags);
1967 * deactivate_task - remove a task from the runqueue.
1969 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1971 if (task_contributes_to_load(p))
1972 rq->nr_uninterruptible++;
1974 dequeue_task(rq, p, flags);
1977 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1980 * There are no locks covering percpu hardirq/softirq time.
1981 * They are only modified in account_system_vtime, on corresponding CPU
1982 * with interrupts disabled. So, writes are safe.
1983 * They are read and saved off onto struct rq in update_rq_clock().
1984 * This may result in other CPU reading this CPU's irq time and can
1985 * race with irq/account_system_vtime on this CPU. We would either get old
1986 * or new value with a side effect of accounting a slice of irq time to wrong
1987 * task when irq is in progress while we read rq->clock. That is a worthy
1988 * compromise in place of having locks on each irq in account_system_time.
1990 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1991 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1993 static DEFINE_PER_CPU(u64, irq_start_time);
1994 static int sched_clock_irqtime;
1996 void enable_sched_clock_irqtime(void)
1998 sched_clock_irqtime = 1;
2001 void disable_sched_clock_irqtime(void)
2003 sched_clock_irqtime = 0;
2006 #ifndef CONFIG_64BIT
2007 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
2009 static inline void irq_time_write_begin(void)
2011 __this_cpu_inc(irq_time_seq.sequence);
2015 static inline void irq_time_write_end(void)
2018 __this_cpu_inc(irq_time_seq.sequence);
2021 static inline u64 irq_time_read(int cpu)
2027 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2028 irq_time = per_cpu(cpu_softirq_time, cpu) +
2029 per_cpu(cpu_hardirq_time, cpu);
2030 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2034 #else /* CONFIG_64BIT */
2035 static inline void irq_time_write_begin(void)
2039 static inline void irq_time_write_end(void)
2043 static inline u64 irq_time_read(int cpu)
2045 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2047 #endif /* CONFIG_64BIT */
2050 * Called before incrementing preempt_count on {soft,}irq_enter
2051 * and before decrementing preempt_count on {soft,}irq_exit.
2053 void account_system_vtime(struct task_struct *curr)
2055 unsigned long flags;
2059 if (!sched_clock_irqtime)
2062 local_irq_save(flags);
2064 cpu = smp_processor_id();
2065 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2066 __this_cpu_add(irq_start_time, delta);
2068 irq_time_write_begin();
2070 * We do not account for softirq time from ksoftirqd here.
2071 * We want to continue accounting softirq time to ksoftirqd thread
2072 * in that case, so as not to confuse scheduler with a special task
2073 * that do not consume any time, but still wants to run.
2075 if (hardirq_count())
2076 __this_cpu_add(cpu_hardirq_time, delta);
2077 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2078 __this_cpu_add(cpu_softirq_time, delta);
2080 irq_time_write_end();
2081 local_irq_restore(flags);
2083 EXPORT_SYMBOL_GPL(account_system_vtime);
2085 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2087 static inline void account_reset_rq(struct rq *rq)
2089 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2090 rq->prev_irq_time = 0;
2092 #ifdef CONFIG_PARAVIRT
2093 rq->prev_steal_time = 0;
2095 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2096 rq->prev_steal_time_rq = 0;
2100 #ifdef CONFIG_PARAVIRT
2101 static inline u64 steal_ticks(u64 steal)
2103 if (unlikely(steal > NSEC_PER_SEC))
2104 return div_u64(steal, TICK_NSEC);
2106 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2110 static void update_rq_clock_task(struct rq *rq, s64 delta)
2113 * In theory, the compile should just see 0 here, and optimize out the call
2114 * to sched_rt_avg_update. But I don't trust it...
2116 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2117 s64 steal = 0, irq_delta = 0;
2119 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2120 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2123 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2124 * this case when a previous update_rq_clock() happened inside a
2125 * {soft,}irq region.
2127 * When this happens, we stop ->clock_task and only update the
2128 * prev_irq_time stamp to account for the part that fit, so that a next
2129 * update will consume the rest. This ensures ->clock_task is
2132 * It does however cause some slight miss-attribution of {soft,}irq
2133 * time, a more accurate solution would be to update the irq_time using
2134 * the current rq->clock timestamp, except that would require using
2137 if (irq_delta > delta)
2140 rq->prev_irq_time += irq_delta;
2143 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2144 if (static_branch((¶virt_steal_rq_enabled))) {
2147 steal = paravirt_steal_clock(cpu_of(rq));
2148 steal -= rq->prev_steal_time_rq;
2150 if (unlikely(steal > delta))
2153 st = steal_ticks(steal);
2154 steal = st * TICK_NSEC;
2156 rq->prev_steal_time_rq += steal;
2162 rq->clock_task += delta;
2164 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2165 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2166 sched_rt_avg_update(rq, irq_delta + steal);
2170 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2171 static int irqtime_account_hi_update(void)
2173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2174 unsigned long flags;
2178 local_irq_save(flags);
2179 latest_ns = this_cpu_read(cpu_hardirq_time);
2180 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2182 local_irq_restore(flags);
2186 static int irqtime_account_si_update(void)
2188 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2189 unsigned long flags;
2193 local_irq_save(flags);
2194 latest_ns = this_cpu_read(cpu_softirq_time);
2195 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2197 local_irq_restore(flags);
2201 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2203 #define sched_clock_irqtime (0)
2208 static void unthrottle_offline_cfs_rqs(struct rq *rq);
2211 #include "sched_idletask.c"
2212 #include "sched_fair.c"
2213 #include "sched_rt.c"
2214 #include "sched_autogroup.c"
2215 #include "sched_stoptask.c"
2216 #ifdef CONFIG_SCHED_DEBUG
2217 # include "sched_debug.c"
2220 void sched_set_stop_task(int cpu, struct task_struct *stop)
2222 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2223 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2227 * Make it appear like a SCHED_FIFO task, its something
2228 * userspace knows about and won't get confused about.
2230 * Also, it will make PI more or less work without too
2231 * much confusion -- but then, stop work should not
2232 * rely on PI working anyway.
2234 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2236 stop->sched_class = &stop_sched_class;
2239 cpu_rq(cpu)->stop = stop;
2243 * Reset it back to a normal scheduling class so that
2244 * it can die in pieces.
2246 old_stop->sched_class = &rt_sched_class;
2251 * __normal_prio - return the priority that is based on the static prio
2253 static inline int __normal_prio(struct task_struct *p)
2255 return p->static_prio;
2259 * Calculate the expected normal priority: i.e. priority
2260 * without taking RT-inheritance into account. Might be
2261 * boosted by interactivity modifiers. Changes upon fork,
2262 * setprio syscalls, and whenever the interactivity
2263 * estimator recalculates.
2265 static inline int normal_prio(struct task_struct *p)
2269 if (task_has_rt_policy(p))
2270 prio = MAX_RT_PRIO-1 - p->rt_priority;
2272 prio = __normal_prio(p);
2277 * Calculate the current priority, i.e. the priority
2278 * taken into account by the scheduler. This value might
2279 * be boosted by RT tasks, or might be boosted by
2280 * interactivity modifiers. Will be RT if the task got
2281 * RT-boosted. If not then it returns p->normal_prio.
2283 static int effective_prio(struct task_struct *p)
2285 p->normal_prio = normal_prio(p);
2287 * If we are RT tasks or we were boosted to RT priority,
2288 * keep the priority unchanged. Otherwise, update priority
2289 * to the normal priority:
2291 if (!rt_prio(p->prio))
2292 return p->normal_prio;
2297 * task_curr - is this task currently executing on a CPU?
2298 * @p: the task in question.
2300 inline int task_curr(const struct task_struct *p)
2302 return cpu_curr(task_cpu(p)) == p;
2305 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2306 const struct sched_class *prev_class,
2309 if (prev_class != p->sched_class) {
2310 if (prev_class->switched_from)
2311 prev_class->switched_from(rq, p);
2312 p->sched_class->switched_to(rq, p);
2313 } else if (oldprio != p->prio)
2314 p->sched_class->prio_changed(rq, p, oldprio);
2317 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2319 const struct sched_class *class;
2321 if (p->sched_class == rq->curr->sched_class) {
2322 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2324 for_each_class(class) {
2325 if (class == rq->curr->sched_class)
2327 if (class == p->sched_class) {
2328 resched_task(rq->curr);
2335 * A queue event has occurred, and we're going to schedule. In
2336 * this case, we can save a useless back to back clock update.
2338 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2339 rq->skip_clock_update = 1;
2344 * Is this task likely cache-hot:
2347 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2351 if (p->sched_class != &fair_sched_class)
2354 if (unlikely(p->policy == SCHED_IDLE))
2358 * Buddy candidates are cache hot:
2360 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2361 (&p->se == cfs_rq_of(&p->se)->next ||
2362 &p->se == cfs_rq_of(&p->se)->last))
2365 if (sysctl_sched_migration_cost == -1)
2367 if (sysctl_sched_migration_cost == 0)
2370 delta = now - p->se.exec_start;
2372 return delta < (s64)sysctl_sched_migration_cost;
2375 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2377 #ifdef CONFIG_SCHED_DEBUG
2379 * We should never call set_task_cpu() on a blocked task,
2380 * ttwu() will sort out the placement.
2382 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2383 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2385 #ifdef CONFIG_LOCKDEP
2387 * The caller should hold either p->pi_lock or rq->lock, when changing
2388 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2390 * sched_move_task() holds both and thus holding either pins the cgroup,
2393 * Furthermore, all task_rq users should acquire both locks, see
2396 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2397 lockdep_is_held(&task_rq(p)->lock)));
2401 trace_sched_migrate_task(p, new_cpu);
2403 if (task_cpu(p) != new_cpu) {
2404 p->se.nr_migrations++;
2405 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2408 __set_task_cpu(p, new_cpu);
2411 struct migration_arg {
2412 struct task_struct *task;
2416 static int migration_cpu_stop(void *data);
2419 * wait_task_inactive - wait for a thread to unschedule.
2421 * If @match_state is nonzero, it's the @p->state value just checked and
2422 * not expected to change. If it changes, i.e. @p might have woken up,
2423 * then return zero. When we succeed in waiting for @p to be off its CPU,
2424 * we return a positive number (its total switch count). If a second call
2425 * a short while later returns the same number, the caller can be sure that
2426 * @p has remained unscheduled the whole time.
2428 * The caller must ensure that the task *will* unschedule sometime soon,
2429 * else this function might spin for a *long* time. This function can't
2430 * be called with interrupts off, or it may introduce deadlock with
2431 * smp_call_function() if an IPI is sent by the same process we are
2432 * waiting to become inactive.
2434 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2436 unsigned long flags;
2443 * We do the initial early heuristics without holding
2444 * any task-queue locks at all. We'll only try to get
2445 * the runqueue lock when things look like they will
2451 * If the task is actively running on another CPU
2452 * still, just relax and busy-wait without holding
2455 * NOTE! Since we don't hold any locks, it's not
2456 * even sure that "rq" stays as the right runqueue!
2457 * But we don't care, since "task_running()" will
2458 * return false if the runqueue has changed and p
2459 * is actually now running somewhere else!
2461 while (task_running(rq, p)) {
2462 if (match_state && unlikely(p->state != match_state))
2468 * Ok, time to look more closely! We need the rq
2469 * lock now, to be *sure*. If we're wrong, we'll
2470 * just go back and repeat.
2472 rq = task_rq_lock(p, &flags);
2473 trace_sched_wait_task(p);
2474 running = task_running(rq, p);
2477 if (!match_state || p->state == match_state)
2478 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2479 task_rq_unlock(rq, p, &flags);
2482 * If it changed from the expected state, bail out now.
2484 if (unlikely(!ncsw))
2488 * Was it really running after all now that we
2489 * checked with the proper locks actually held?
2491 * Oops. Go back and try again..
2493 if (unlikely(running)) {
2499 * It's not enough that it's not actively running,
2500 * it must be off the runqueue _entirely_, and not
2503 * So if it was still runnable (but just not actively
2504 * running right now), it's preempted, and we should
2505 * yield - it could be a while.
2507 if (unlikely(on_rq)) {
2508 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2510 set_current_state(TASK_UNINTERRUPTIBLE);
2511 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2516 * Ahh, all good. It wasn't running, and it wasn't
2517 * runnable, which means that it will never become
2518 * running in the future either. We're all done!
2527 * kick_process - kick a running thread to enter/exit the kernel
2528 * @p: the to-be-kicked thread
2530 * Cause a process which is running on another CPU to enter
2531 * kernel-mode, without any delay. (to get signals handled.)
2533 * NOTE: this function doesn't have to take the runqueue lock,
2534 * because all it wants to ensure is that the remote task enters
2535 * the kernel. If the IPI races and the task has been migrated
2536 * to another CPU then no harm is done and the purpose has been
2539 void kick_process(struct task_struct *p)
2545 if ((cpu != smp_processor_id()) && task_curr(p))
2546 smp_send_reschedule(cpu);
2549 EXPORT_SYMBOL_GPL(kick_process);
2550 #endif /* CONFIG_SMP */
2554 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2556 static int select_fallback_rq(int cpu, struct task_struct *p)
2559 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2561 /* Look for allowed, online CPU in same node. */
2562 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2563 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
2566 /* Any allowed, online CPU? */
2567 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
2568 if (dest_cpu < nr_cpu_ids)
2571 /* No more Mr. Nice Guy. */
2572 dest_cpu = cpuset_cpus_allowed_fallback(p);
2574 * Don't tell them about moving exiting tasks or
2575 * kernel threads (both mm NULL), since they never
2578 if (p->mm && printk_ratelimit()) {
2579 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2580 task_pid_nr(p), p->comm, cpu);
2587 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2590 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2592 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2595 * In order not to call set_task_cpu() on a blocking task we need
2596 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2599 * Since this is common to all placement strategies, this lives here.
2601 * [ this allows ->select_task() to simply return task_cpu(p) and
2602 * not worry about this generic constraint ]
2604 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
2606 cpu = select_fallback_rq(task_cpu(p), p);
2611 static void update_avg(u64 *avg, u64 sample)
2613 s64 diff = sample - *avg;
2619 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2621 #ifdef CONFIG_SCHEDSTATS
2622 struct rq *rq = this_rq();
2625 int this_cpu = smp_processor_id();
2627 if (cpu == this_cpu) {
2628 schedstat_inc(rq, ttwu_local);
2629 schedstat_inc(p, se.statistics.nr_wakeups_local);
2631 struct sched_domain *sd;
2633 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2635 for_each_domain(this_cpu, sd) {
2636 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2637 schedstat_inc(sd, ttwu_wake_remote);
2644 if (wake_flags & WF_MIGRATED)
2645 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2647 #endif /* CONFIG_SMP */
2649 schedstat_inc(rq, ttwu_count);
2650 schedstat_inc(p, se.statistics.nr_wakeups);
2652 if (wake_flags & WF_SYNC)
2653 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2655 #endif /* CONFIG_SCHEDSTATS */
2658 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2660 activate_task(rq, p, en_flags);
2663 /* if a worker is waking up, notify workqueue */
2664 if (p->flags & PF_WQ_WORKER)
2665 wq_worker_waking_up(p, cpu_of(rq));
2669 * Mark the task runnable and perform wakeup-preemption.
2672 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2674 trace_sched_wakeup(p, true);
2675 check_preempt_curr(rq, p, wake_flags);
2677 p->state = TASK_RUNNING;
2679 if (p->sched_class->task_woken)
2680 p->sched_class->task_woken(rq, p);
2682 if (rq->idle_stamp) {
2683 u64 delta = rq->clock - rq->idle_stamp;
2684 u64 max = 2*sysctl_sched_migration_cost;
2689 update_avg(&rq->avg_idle, delta);
2696 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2699 if (p->sched_contributes_to_load)
2700 rq->nr_uninterruptible--;
2703 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2704 ttwu_do_wakeup(rq, p, wake_flags);
2708 * Called in case the task @p isn't fully descheduled from its runqueue,
2709 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2710 * since all we need to do is flip p->state to TASK_RUNNING, since
2711 * the task is still ->on_rq.
2713 static int ttwu_remote(struct task_struct *p, int wake_flags)
2718 rq = __task_rq_lock(p);
2720 ttwu_do_wakeup(rq, p, wake_flags);
2723 __task_rq_unlock(rq);
2729 static void sched_ttwu_pending(void)
2731 struct rq *rq = this_rq();
2732 struct llist_node *llist = llist_del_all(&rq->wake_list);
2733 struct task_struct *p;
2735 raw_spin_lock(&rq->lock);
2738 p = llist_entry(llist, struct task_struct, wake_entry);
2739 llist = llist_next(llist);
2740 ttwu_do_activate(rq, p, 0);
2743 raw_spin_unlock(&rq->lock);
2746 void scheduler_ipi(void)
2748 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
2752 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2753 * traditionally all their work was done from the interrupt return
2754 * path. Now that we actually do some work, we need to make sure
2757 * Some archs already do call them, luckily irq_enter/exit nest
2760 * Arguably we should visit all archs and update all handlers,
2761 * however a fair share of IPIs are still resched only so this would
2762 * somewhat pessimize the simple resched case.
2765 sched_ttwu_pending();
2768 * Check if someone kicked us for doing the nohz idle load balance.
2770 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2771 this_rq()->idle_balance = 1;
2772 raise_softirq_irqoff(SCHED_SOFTIRQ);
2777 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2779 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2780 smp_send_reschedule(cpu);
2783 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2784 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2789 rq = __task_rq_lock(p);
2791 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2792 ttwu_do_wakeup(rq, p, wake_flags);
2795 __task_rq_unlock(rq);
2800 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2801 #endif /* CONFIG_SMP */
2803 static void ttwu_queue(struct task_struct *p, int cpu)
2805 struct rq *rq = cpu_rq(cpu);
2807 #if defined(CONFIG_SMP)
2808 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2809 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2810 ttwu_queue_remote(p, cpu);
2815 raw_spin_lock(&rq->lock);
2816 ttwu_do_activate(rq, p, 0);
2817 raw_spin_unlock(&rq->lock);
2821 * try_to_wake_up - wake up a thread
2822 * @p: the thread to be awakened
2823 * @state: the mask of task states that can be woken
2824 * @wake_flags: wake modifier flags (WF_*)
2826 * Put it on the run-queue if it's not already there. The "current"
2827 * thread is always on the run-queue (except when the actual
2828 * re-schedule is in progress), and as such you're allowed to do
2829 * the simpler "current->state = TASK_RUNNING" to mark yourself
2830 * runnable without the overhead of this.
2832 * Returns %true if @p was woken up, %false if it was already running
2833 * or @state didn't match @p's state.
2836 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2838 unsigned long flags;
2839 int cpu, success = 0;
2842 raw_spin_lock_irqsave(&p->pi_lock, flags);
2843 if (!(p->state & state))
2846 success = 1; /* we're going to change ->state */
2849 if (p->on_rq && ttwu_remote(p, wake_flags))
2854 * If the owning (remote) cpu is still in the middle of schedule() with
2855 * this task as prev, wait until its done referencing the task.
2858 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2860 * In case the architecture enables interrupts in
2861 * context_switch(), we cannot busy wait, since that
2862 * would lead to deadlocks when an interrupt hits and
2863 * tries to wake up @prev. So bail and do a complete
2866 if (ttwu_activate_remote(p, wake_flags))
2873 * Pairs with the smp_wmb() in finish_lock_switch().
2877 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2878 p->state = TASK_WAKING;
2880 if (p->sched_class->task_waking)
2881 p->sched_class->task_waking(p);
2883 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2884 if (task_cpu(p) != cpu) {
2885 wake_flags |= WF_MIGRATED;
2886 set_task_cpu(p, cpu);
2888 #endif /* CONFIG_SMP */
2892 ttwu_stat(p, cpu, wake_flags);
2894 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2900 * try_to_wake_up_local - try to wake up a local task with rq lock held
2901 * @p: the thread to be awakened
2903 * Put @p on the run-queue if it's not already there. The caller must
2904 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2907 static void try_to_wake_up_local(struct task_struct *p)
2909 struct rq *rq = task_rq(p);
2911 if (WARN_ON_ONCE(rq != this_rq()) ||
2912 WARN_ON_ONCE(p == current))
2915 lockdep_assert_held(&rq->lock);
2917 if (!raw_spin_trylock(&p->pi_lock)) {
2918 raw_spin_unlock(&rq->lock);
2919 raw_spin_lock(&p->pi_lock);
2920 raw_spin_lock(&rq->lock);
2923 if (!(p->state & TASK_NORMAL))
2927 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2929 ttwu_do_wakeup(rq, p, 0);
2930 ttwu_stat(p, smp_processor_id(), 0);
2932 raw_spin_unlock(&p->pi_lock);
2936 * wake_up_process - Wake up a specific process
2937 * @p: The process to be woken up.
2939 * Attempt to wake up the nominated process and move it to the set of runnable
2940 * processes. Returns 1 if the process was woken up, 0 if it was already
2943 * It may be assumed that this function implies a write memory barrier before
2944 * changing the task state if and only if any tasks are woken up.
2946 int wake_up_process(struct task_struct *p)
2948 return try_to_wake_up(p, TASK_NORMAL, 0);
2950 EXPORT_SYMBOL(wake_up_process);
2952 int wake_up_state(struct task_struct *p, unsigned int state)
2954 return try_to_wake_up(p, state, 0);
2958 * Perform scheduler related setup for a newly forked process p.
2959 * p is forked by current.
2961 * __sched_fork() is basic setup used by init_idle() too:
2963 static void __sched_fork(struct task_struct *p)
2968 p->se.exec_start = 0;
2969 p->se.sum_exec_runtime = 0;
2970 p->se.prev_sum_exec_runtime = 0;
2971 p->se.nr_migrations = 0;
2973 INIT_LIST_HEAD(&p->se.group_node);
2975 #ifdef CONFIG_SCHEDSTATS
2976 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2979 INIT_LIST_HEAD(&p->rt.run_list);
2981 #ifdef CONFIG_PREEMPT_NOTIFIERS
2982 INIT_HLIST_HEAD(&p->preempt_notifiers);
2987 * fork()/clone()-time setup:
2989 void sched_fork(struct task_struct *p)
2991 unsigned long flags;
2992 int cpu = get_cpu();
2996 * We mark the process as running here. This guarantees that
2997 * nobody will actually run it, and a signal or other external
2998 * event cannot wake it up and insert it on the runqueue either.
3000 p->state = TASK_RUNNING;
3003 * Make sure we do not leak PI boosting priority to the child.
3005 p->prio = current->normal_prio;
3008 * Revert to default priority/policy on fork if requested.
3010 if (unlikely(p->sched_reset_on_fork)) {
3011 if (task_has_rt_policy(p)) {
3012 p->policy = SCHED_NORMAL;
3013 p->static_prio = NICE_TO_PRIO(0);
3015 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3016 p->static_prio = NICE_TO_PRIO(0);
3018 p->prio = p->normal_prio = __normal_prio(p);
3022 * We don't need the reset flag anymore after the fork. It has
3023 * fulfilled its duty:
3025 p->sched_reset_on_fork = 0;
3028 if (!rt_prio(p->prio))
3029 p->sched_class = &fair_sched_class;
3031 if (p->sched_class->task_fork)
3032 p->sched_class->task_fork(p);
3035 * The child is not yet in the pid-hash so no cgroup attach races,
3036 * and the cgroup is pinned to this child due to cgroup_fork()
3037 * is ran before sched_fork().
3039 * Silence PROVE_RCU.
3041 raw_spin_lock_irqsave(&p->pi_lock, flags);
3042 set_task_cpu(p, cpu);
3043 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3045 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3046 if (likely(sched_info_on()))
3047 memset(&p->sched_info, 0, sizeof(p->sched_info));
3049 #if defined(CONFIG_SMP)
3052 #ifdef CONFIG_PREEMPT_COUNT
3053 /* Want to start with kernel preemption disabled. */
3054 task_thread_info(p)->preempt_count = 1;
3057 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3064 * wake_up_new_task - wake up a newly created task for the first time.
3066 * This function will do some initial scheduler statistics housekeeping
3067 * that must be done for every newly created context, then puts the task
3068 * on the runqueue and wakes it.
3070 void wake_up_new_task(struct task_struct *p)
3072 unsigned long flags;
3075 raw_spin_lock_irqsave(&p->pi_lock, flags);
3078 * Fork balancing, do it here and not earlier because:
3079 * - cpus_allowed can change in the fork path
3080 * - any previously selected cpu might disappear through hotplug
3082 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3085 rq = __task_rq_lock(p);
3086 activate_task(rq, p, 0);
3088 trace_sched_wakeup_new(p, true);
3089 check_preempt_curr(rq, p, WF_FORK);
3091 if (p->sched_class->task_woken)
3092 p->sched_class->task_woken(rq, p);
3094 task_rq_unlock(rq, p, &flags);
3097 #ifdef CONFIG_PREEMPT_NOTIFIERS
3100 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3101 * @notifier: notifier struct to register
3103 void preempt_notifier_register(struct preempt_notifier *notifier)
3105 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3107 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3110 * preempt_notifier_unregister - no longer interested in preemption notifications
3111 * @notifier: notifier struct to unregister
3113 * This is safe to call from within a preemption notifier.
3115 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3117 hlist_del(¬ifier->link);
3119 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3121 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3123 struct preempt_notifier *notifier;
3124 struct hlist_node *node;
3126 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3127 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3131 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3132 struct task_struct *next)
3134 struct preempt_notifier *notifier;
3135 struct hlist_node *node;
3137 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3138 notifier->ops->sched_out(notifier, next);
3141 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3143 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3148 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3149 struct task_struct *next)
3153 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3156 * prepare_task_switch - prepare to switch tasks
3157 * @rq: the runqueue preparing to switch
3158 * @prev: the current task that is being switched out
3159 * @next: the task we are going to switch to.
3161 * This is called with the rq lock held and interrupts off. It must
3162 * be paired with a subsequent finish_task_switch after the context
3165 * prepare_task_switch sets up locking and calls architecture specific
3169 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3170 struct task_struct *next)
3172 sched_info_switch(prev, next);
3173 perf_event_task_sched_out(prev, next);
3174 fire_sched_out_preempt_notifiers(prev, next);
3175 prepare_lock_switch(rq, next);
3176 prepare_arch_switch(next);
3177 trace_sched_switch(prev, next);
3181 * finish_task_switch - clean up after a task-switch
3182 * @rq: runqueue associated with task-switch
3183 * @prev: the thread we just switched away from.
3185 * finish_task_switch must be called after the context switch, paired
3186 * with a prepare_task_switch call before the context switch.
3187 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3188 * and do any other architecture-specific cleanup actions.
3190 * Note that we may have delayed dropping an mm in context_switch(). If
3191 * so, we finish that here outside of the runqueue lock. (Doing it
3192 * with the lock held can cause deadlocks; see schedule() for
3195 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3196 __releases(rq->lock)
3198 struct mm_struct *mm = rq->prev_mm;
3204 * A task struct has one reference for the use as "current".
3205 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3206 * schedule one last time. The schedule call will never return, and
3207 * the scheduled task must drop that reference.
3209 * We must observe prev->state before clearing prev->on_cpu (in
3210 * finish_lock_switch), otherwise a concurrent wakeup can get prev
3211 * running on another CPU and we could rave with its RUNNING -> DEAD
3212 * transition, resulting in a double drop.
3214 prev_state = prev->state;
3215 finish_arch_switch(prev);
3216 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3217 local_irq_disable();
3218 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3219 perf_event_task_sched_in(prev, current);
3220 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3222 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3223 finish_lock_switch(rq, prev);
3225 fire_sched_in_preempt_notifiers(current);
3228 if (unlikely(prev_state == TASK_DEAD)) {
3230 * Remove function-return probe instances associated with this
3231 * task and put them back on the free list.
3233 kprobe_flush_task(prev);
3234 put_task_struct(prev);
3240 /* assumes rq->lock is held */
3241 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3243 if (prev->sched_class->pre_schedule)
3244 prev->sched_class->pre_schedule(rq, prev);
3247 /* rq->lock is NOT held, but preemption is disabled */
3248 static inline void post_schedule(struct rq *rq)
3250 if (rq->post_schedule) {
3251 unsigned long flags;
3253 raw_spin_lock_irqsave(&rq->lock, flags);
3254 if (rq->curr->sched_class->post_schedule)
3255 rq->curr->sched_class->post_schedule(rq);
3256 raw_spin_unlock_irqrestore(&rq->lock, flags);
3258 rq->post_schedule = 0;
3264 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3268 static inline void post_schedule(struct rq *rq)
3275 * schedule_tail - first thing a freshly forked thread must call.
3276 * @prev: the thread we just switched away from.
3278 asmlinkage void schedule_tail(struct task_struct *prev)
3279 __releases(rq->lock)
3281 struct rq *rq = this_rq();
3283 finish_task_switch(rq, prev);
3286 * FIXME: do we need to worry about rq being invalidated by the
3291 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3292 /* In this case, finish_task_switch does not reenable preemption */
3295 if (current->set_child_tid)
3296 put_user(task_pid_vnr(current), current->set_child_tid);
3300 * context_switch - switch to the new MM and the new
3301 * thread's register state.
3304 context_switch(struct rq *rq, struct task_struct *prev,
3305 struct task_struct *next)
3307 struct mm_struct *mm, *oldmm;
3309 prepare_task_switch(rq, prev, next);
3312 oldmm = prev->active_mm;
3314 * For paravirt, this is coupled with an exit in switch_to to
3315 * combine the page table reload and the switch backend into
3318 arch_start_context_switch(prev);
3321 next->active_mm = oldmm;
3322 atomic_inc(&oldmm->mm_count);
3323 enter_lazy_tlb(oldmm, next);
3325 switch_mm(oldmm, mm, next);
3328 prev->active_mm = NULL;
3329 rq->prev_mm = oldmm;
3332 * Since the runqueue lock will be released by the next
3333 * task (which is an invalid locking op but in the case
3334 * of the scheduler it's an obvious special-case), so we
3335 * do an early lockdep release here:
3337 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3338 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3341 /* Here we just switch the register state and the stack. */
3342 switch_to(prev, next, prev);
3346 * this_rq must be evaluated again because prev may have moved
3347 * CPUs since it called schedule(), thus the 'rq' on its stack
3348 * frame will be invalid.
3350 finish_task_switch(this_rq(), prev);
3354 * nr_running, nr_uninterruptible and nr_context_switches:
3356 * externally visible scheduler statistics: current number of runnable
3357 * threads, current number of uninterruptible-sleeping threads, total
3358 * number of context switches performed since bootup.
3360 unsigned long nr_running(void)
3362 unsigned long i, sum = 0;
3364 for_each_online_cpu(i)
3365 sum += cpu_rq(i)->nr_running;
3370 unsigned long nr_uninterruptible(void)
3372 unsigned long i, sum = 0;
3374 for_each_possible_cpu(i)
3375 sum += cpu_rq(i)->nr_uninterruptible;
3378 * Since we read the counters lockless, it might be slightly
3379 * inaccurate. Do not allow it to go below zero though:
3381 if (unlikely((long)sum < 0))
3387 unsigned long long nr_context_switches(void)
3390 unsigned long long sum = 0;
3392 for_each_possible_cpu(i)
3393 sum += cpu_rq(i)->nr_switches;
3398 unsigned long nr_iowait(void)
3400 unsigned long i, sum = 0;
3402 for_each_possible_cpu(i)
3403 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3408 unsigned long nr_iowait_cpu(int cpu)
3410 struct rq *this = cpu_rq(cpu);
3411 return atomic_read(&this->nr_iowait);
3414 unsigned long this_cpu_load(void)
3416 struct rq *this = this_rq();
3417 return this->cpu_load[0];
3422 * Global load-average calculations
3424 * We take a distributed and async approach to calculating the global load-avg
3425 * in order to minimize overhead.
3427 * The global load average is an exponentially decaying average of nr_running +
3428 * nr_uninterruptible.
3430 * Once every LOAD_FREQ:
3433 * for_each_possible_cpu(cpu)
3434 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
3436 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
3438 * Due to a number of reasons the above turns in the mess below:
3440 * - for_each_possible_cpu() is prohibitively expensive on machines with
3441 * serious number of cpus, therefore we need to take a distributed approach
3442 * to calculating nr_active.
3444 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
3445 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
3447 * So assuming nr_active := 0 when we start out -- true per definition, we
3448 * can simply take per-cpu deltas and fold those into a global accumulate
3449 * to obtain the same result. See calc_load_fold_active().
3451 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
3452 * across the machine, we assume 10 ticks is sufficient time for every
3453 * cpu to have completed this task.
3455 * This places an upper-bound on the IRQ-off latency of the machine. Then
3456 * again, being late doesn't loose the delta, just wrecks the sample.
3458 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
3459 * this would add another cross-cpu cacheline miss and atomic operation
3460 * to the wakeup path. Instead we increment on whatever cpu the task ran
3461 * when it went into uninterruptible state and decrement on whatever cpu
3462 * did the wakeup. This means that only the sum of nr_uninterruptible over
3463 * all cpus yields the correct result.
3465 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
3468 /* Variables and functions for calc_load */
3469 static atomic_long_t calc_load_tasks;
3470 static unsigned long calc_load_update;
3471 unsigned long avenrun[3];
3472 EXPORT_SYMBOL(avenrun); /* should be removed */
3475 * get_avenrun - get the load average array
3476 * @loads: pointer to dest load array
3477 * @offset: offset to add
3478 * @shift: shift count to shift the result left
3480 * These values are estimates at best, so no need for locking.
3482 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3484 loads[0] = (avenrun[0] + offset) << shift;
3485 loads[1] = (avenrun[1] + offset) << shift;
3486 loads[2] = (avenrun[2] + offset) << shift;
3489 static long calc_load_fold_active(struct rq *this_rq)
3491 long nr_active, delta = 0;
3493 nr_active = this_rq->nr_running;
3494 nr_active += (long) this_rq->nr_uninterruptible;
3496 if (nr_active != this_rq->calc_load_active) {
3497 delta = nr_active - this_rq->calc_load_active;
3498 this_rq->calc_load_active = nr_active;
3505 * a1 = a0 * e + a * (1 - e)
3507 static unsigned long
3508 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3511 load += active * (FIXED_1 - exp);
3512 load += 1UL << (FSHIFT - 1);
3513 return load >> FSHIFT;
3518 * Handle NO_HZ for the global load-average.
3520 * Since the above described distributed algorithm to compute the global
3521 * load-average relies on per-cpu sampling from the tick, it is affected by
3524 * The basic idea is to fold the nr_active delta into a global idle-delta upon
3525 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
3526 * when we read the global state.
3528 * Obviously reality has to ruin such a delightfully simple scheme:
3530 * - When we go NO_HZ idle during the window, we can negate our sample
3531 * contribution, causing under-accounting.
3533 * We avoid this by keeping two idle-delta counters and flipping them
3534 * when the window starts, thus separating old and new NO_HZ load.
3536 * The only trick is the slight shift in index flip for read vs write.
3540 * |-|-----------|-|-----------|-|-----------|-|
3541 * r:0 0 1 1 0 0 1 1 0
3542 * w:0 1 1 0 0 1 1 0 0
3544 * This ensures we'll fold the old idle contribution in this window while
3545 * accumlating the new one.
3547 * - When we wake up from NO_HZ idle during the window, we push up our
3548 * contribution, since we effectively move our sample point to a known
3551 * This is solved by pushing the window forward, and thus skipping the
3552 * sample, for this cpu (effectively using the idle-delta for this cpu which
3553 * was in effect at the time the window opened). This also solves the issue
3554 * of having to deal with a cpu having been in NOHZ idle for multiple
3555 * LOAD_FREQ intervals.
3557 * When making the ILB scale, we should try to pull this in as well.
3559 static atomic_long_t calc_load_idle[2];
3560 static int calc_load_idx;
3562 static inline int calc_load_write_idx(void)
3564 int idx = calc_load_idx;
3567 * See calc_global_nohz(), if we observe the new index, we also
3568 * need to observe the new update time.
3573 * If the folding window started, make sure we start writing in the
3576 if (!time_before(jiffies, calc_load_update))
3582 static inline int calc_load_read_idx(void)
3584 return calc_load_idx & 1;
3587 void calc_load_enter_idle(void)
3589 struct rq *this_rq = this_rq();
3593 * We're going into NOHZ mode, if there's any pending delta, fold it
3594 * into the pending idle delta.
3596 delta = calc_load_fold_active(this_rq);
3598 int idx = calc_load_write_idx();
3599 atomic_long_add(delta, &calc_load_idle[idx]);
3603 void calc_load_exit_idle(void)
3605 struct rq *this_rq = this_rq();
3608 * If we're still before the sample window, we're done.
3610 if (time_before(jiffies, this_rq->calc_load_update))
3614 * We woke inside or after the sample window, this means we're already
3615 * accounted through the nohz accounting, so skip the entire deal and
3616 * sync up for the next window.
3618 this_rq->calc_load_update = calc_load_update;
3619 if (time_before(jiffies, this_rq->calc_load_update + 10))
3620 this_rq->calc_load_update += LOAD_FREQ;
3623 static long calc_load_fold_idle(void)
3625 int idx = calc_load_read_idx();
3628 if (atomic_long_read(&calc_load_idle[idx]))
3629 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
3635 * fixed_power_int - compute: x^n, in O(log n) time
3637 * @x: base of the power
3638 * @frac_bits: fractional bits of @x
3639 * @n: power to raise @x to.
3641 * By exploiting the relation between the definition of the natural power
3642 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3643 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3644 * (where: n_i \elem {0, 1}, the binary vector representing n),
3645 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3646 * of course trivially computable in O(log_2 n), the length of our binary
3649 static unsigned long
3650 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3652 unsigned long result = 1UL << frac_bits;
3657 result += 1UL << (frac_bits - 1);
3658 result >>= frac_bits;
3664 x += 1UL << (frac_bits - 1);
3672 * a1 = a0 * e + a * (1 - e)
3674 * a2 = a1 * e + a * (1 - e)
3675 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3676 * = a0 * e^2 + a * (1 - e) * (1 + e)
3678 * a3 = a2 * e + a * (1 - e)
3679 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3680 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3684 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3685 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3686 * = a0 * e^n + a * (1 - e^n)
3688 * [1] application of the geometric series:
3691 * S_n := \Sum x^i = -------------
3694 static unsigned long
3695 calc_load_n(unsigned long load, unsigned long exp,
3696 unsigned long active, unsigned int n)
3699 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3703 * NO_HZ can leave us missing all per-cpu ticks calling
3704 * calc_load_account_active(), but since an idle CPU folds its delta into
3705 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3706 * in the pending idle delta if our idle period crossed a load cycle boundary.
3708 * Once we've updated the global active value, we need to apply the exponential
3709 * weights adjusted to the number of cycles missed.
3711 static void calc_global_nohz(void)
3713 long delta, active, n;
3715 if (!time_before(jiffies, calc_load_update + 10)) {
3717 * Catch-up, fold however many we are behind still
3719 delta = jiffies - calc_load_update - 10;
3720 n = 1 + (delta / LOAD_FREQ);
3722 active = atomic_long_read(&calc_load_tasks);
3723 active = active > 0 ? active * FIXED_1 : 0;
3725 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3726 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3727 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3729 calc_load_update += n * LOAD_FREQ;
3733 * Flip the idle index...
3735 * Make sure we first write the new time then flip the index, so that
3736 * calc_load_write_idx() will see the new time when it reads the new
3737 * index, this avoids a double flip messing things up.
3742 #else /* !CONFIG_NO_HZ */
3744 static inline long calc_load_fold_idle(void) { return 0; }
3745 static inline void calc_global_nohz(void) { }
3747 #endif /* CONFIG_NO_HZ */
3750 * calc_load - update the avenrun load estimates 10 ticks after the
3751 * CPUs have updated calc_load_tasks.
3753 void calc_global_load(unsigned long ticks)
3757 if (time_before(jiffies, calc_load_update + 10))
3761 * Fold the 'old' idle-delta to include all NO_HZ cpus.
3763 delta = calc_load_fold_idle();
3765 atomic_long_add(delta, &calc_load_tasks);
3767 active = atomic_long_read(&calc_load_tasks);
3768 active = active > 0 ? active * FIXED_1 : 0;
3770 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3771 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3772 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3774 calc_load_update += LOAD_FREQ;
3777 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
3783 * Called from update_cpu_load() to periodically update this CPU's
3786 static void calc_load_account_active(struct rq *this_rq)
3790 if (time_before(jiffies, this_rq->calc_load_update))
3793 delta = calc_load_fold_active(this_rq);
3795 atomic_long_add(delta, &calc_load_tasks);
3797 this_rq->calc_load_update += LOAD_FREQ;
3801 * End of global load-average stuff
3805 * The exact cpuload at various idx values, calculated at every tick would be
3806 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3808 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3809 * on nth tick when cpu may be busy, then we have:
3810 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3811 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3813 * decay_load_missed() below does efficient calculation of
3814 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3815 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3817 * The calculation is approximated on a 128 point scale.
3818 * degrade_zero_ticks is the number of ticks after which load at any
3819 * particular idx is approximated to be zero.
3820 * degrade_factor is a precomputed table, a row for each load idx.
3821 * Each column corresponds to degradation factor for a power of two ticks,
3822 * based on 128 point scale.
3824 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3825 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3827 * With this power of 2 load factors, we can degrade the load n times
3828 * by looking at 1 bits in n and doing as many mult/shift instead of
3829 * n mult/shifts needed by the exact degradation.
3831 #define DEGRADE_SHIFT 7
3832 static const unsigned char
3833 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3834 static const unsigned char
3835 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3836 {0, 0, 0, 0, 0, 0, 0, 0},
3837 {64, 32, 8, 0, 0, 0, 0, 0},
3838 {96, 72, 40, 12, 1, 0, 0},
3839 {112, 98, 75, 43, 15, 1, 0},
3840 {120, 112, 98, 76, 45, 16, 2} };
3843 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3844 * would be when CPU is idle and so we just decay the old load without
3845 * adding any new load.
3847 static unsigned long
3848 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3852 if (!missed_updates)
3855 if (missed_updates >= degrade_zero_ticks[idx])
3859 return load >> missed_updates;
3861 while (missed_updates) {
3862 if (missed_updates % 2)
3863 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3865 missed_updates >>= 1;
3872 * Update rq->cpu_load[] statistics. This function is usually called every
3873 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3874 * every tick. We fix it up based on jiffies.
3876 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
3877 unsigned long pending_updates)
3881 this_rq->nr_load_updates++;
3883 /* Update our load: */
3884 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3885 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3886 unsigned long old_load, new_load;
3888 /* scale is effectively 1 << i now, and >> i divides by scale */
3890 old_load = this_rq->cpu_load[i];
3891 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3892 new_load = this_load;
3894 * Round up the averaging division if load is increasing. This
3895 * prevents us from getting stuck on 9 if the load is 10, for
3898 if (new_load > old_load)
3899 new_load += scale - 1;
3901 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3904 sched_avg_update(this_rq);
3909 * There is no sane way to deal with nohz on smp when using jiffies because the
3910 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
3911 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
3913 * Therefore we cannot use the delta approach from the regular tick since that
3914 * would seriously skew the load calculation. However we'll make do for those
3915 * updates happening while idle (nohz_idle_balance) or coming out of idle
3916 * (tick_nohz_idle_exit).
3918 * This means we might still be one tick off for nohz periods.
3922 * Called from nohz_idle_balance() to update the load ratings before doing the
3925 static void update_idle_cpu_load(struct rq *this_rq)
3927 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3928 unsigned long load = this_rq->load.weight;
3929 unsigned long pending_updates;
3932 * bail if there's load or we're actually up-to-date.
3934 if (load || curr_jiffies == this_rq->last_load_update_tick)
3937 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3938 this_rq->last_load_update_tick = curr_jiffies;
3940 __update_cpu_load(this_rq, load, pending_updates);
3944 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
3946 void update_cpu_load_nohz(void)
3948 struct rq *this_rq = this_rq();
3949 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
3950 unsigned long pending_updates;
3952 if (curr_jiffies == this_rq->last_load_update_tick)
3955 raw_spin_lock(&this_rq->lock);
3956 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3957 if (pending_updates) {
3958 this_rq->last_load_update_tick = curr_jiffies;
3960 * We were idle, this means load 0, the current load might be
3961 * !0 due to remote wakeups and the sort.
3963 __update_cpu_load(this_rq, 0, pending_updates);
3965 raw_spin_unlock(&this_rq->lock);
3967 #endif /* CONFIG_NO_HZ */
3970 * Called from scheduler_tick()
3972 static void update_cpu_load_active(struct rq *this_rq)
3975 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
3977 this_rq->last_load_update_tick = jiffies;
3978 __update_cpu_load(this_rq, this_rq->load.weight, 1);
3980 calc_load_account_active(this_rq);
3986 * sched_exec - execve() is a valuable balancing opportunity, because at
3987 * this point the task has the smallest effective memory and cache footprint.
3989 void sched_exec(void)
3991 struct task_struct *p = current;
3992 unsigned long flags;
3995 raw_spin_lock_irqsave(&p->pi_lock, flags);
3996 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3997 if (dest_cpu == smp_processor_id())
4000 if (likely(cpu_active(dest_cpu))) {
4001 struct migration_arg arg = { p, dest_cpu };
4003 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4004 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
4008 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4013 DEFINE_PER_CPU(struct kernel_stat, kstat);
4015 EXPORT_PER_CPU_SYMBOL(kstat);
4018 * Return any ns on the sched_clock that have not yet been accounted in
4019 * @p in case that task is currently running.
4021 * Called with task_rq_lock() held on @rq.
4023 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4027 if (task_current(rq, p)) {
4028 update_rq_clock(rq);
4029 ns = rq->clock_task - p->se.exec_start;
4037 unsigned long long task_delta_exec(struct task_struct *p)
4039 unsigned long flags;
4043 rq = task_rq_lock(p, &flags);
4044 ns = do_task_delta_exec(p, rq);
4045 task_rq_unlock(rq, p, &flags);
4051 * Return accounted runtime for the task.
4052 * In case the task is currently running, return the runtime plus current's
4053 * pending runtime that have not been accounted yet.
4055 unsigned long long task_sched_runtime(struct task_struct *p)
4057 unsigned long flags;
4061 rq = task_rq_lock(p, &flags);
4062 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4063 task_rq_unlock(rq, p, &flags);
4069 * Account user cpu time to a process.
4070 * @p: the process that the cpu time gets accounted to
4071 * @cputime: the cpu time spent in user space since the last update
4072 * @cputime_scaled: cputime scaled by cpu frequency
4074 void account_user_time(struct task_struct *p, cputime_t cputime,
4075 cputime_t cputime_scaled)
4077 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4080 /* Add user time to process. */
4081 p->utime = cputime_add(p->utime, cputime);
4082 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4083 account_group_user_time(p, cputime);
4085 /* Add user time to cpustat. */
4086 tmp = cputime_to_cputime64(cputime);
4087 if (TASK_NICE(p) > 0)
4088 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4090 cpustat->user = cputime64_add(cpustat->user, tmp);
4092 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4093 /* Account for user time used */
4094 acct_update_integrals(p);
4098 * Account guest cpu time to a process.
4099 * @p: the process that the cpu time gets accounted to
4100 * @cputime: the cpu time spent in virtual machine since the last update
4101 * @cputime_scaled: cputime scaled by cpu frequency
4103 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4104 cputime_t cputime_scaled)
4107 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4109 tmp = cputime_to_cputime64(cputime);
4111 /* Add guest time to process. */
4112 p->utime = cputime_add(p->utime, cputime);
4113 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4114 account_group_user_time(p, cputime);
4115 p->gtime = cputime_add(p->gtime, cputime);
4117 /* Add guest time to cpustat. */
4118 if (TASK_NICE(p) > 0) {
4119 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4120 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
4122 cpustat->user = cputime64_add(cpustat->user, tmp);
4123 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4128 * Account system cpu time to a process and desired cpustat field
4129 * @p: the process that the cpu time gets accounted to
4130 * @cputime: the cpu time spent in kernel space since the last update
4131 * @cputime_scaled: cputime scaled by cpu frequency
4132 * @target_cputime64: pointer to cpustat field that has to be updated
4135 void __account_system_time(struct task_struct *p, cputime_t cputime,
4136 cputime_t cputime_scaled, cputime64_t *target_cputime64)
4138 cputime64_t tmp = cputime_to_cputime64(cputime);
4140 /* Add system time to process. */
4141 p->stime = cputime_add(p->stime, cputime);
4142 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4143 account_group_system_time(p, cputime);
4145 /* Add system time to cpustat. */
4146 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
4147 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4149 /* Account for system time used */
4150 acct_update_integrals(p);
4154 * Account system cpu time to a process.
4155 * @p: the process that the cpu time gets accounted to
4156 * @hardirq_offset: the offset to subtract from hardirq_count()
4157 * @cputime: the cpu time spent in kernel space since the last update
4158 * @cputime_scaled: cputime scaled by cpu frequency
4160 void account_system_time(struct task_struct *p, int hardirq_offset,
4161 cputime_t cputime, cputime_t cputime_scaled)
4163 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4164 cputime64_t *target_cputime64;
4166 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4167 account_guest_time(p, cputime, cputime_scaled);
4171 if (hardirq_count() - hardirq_offset)
4172 target_cputime64 = &cpustat->irq;
4173 else if (in_serving_softirq())
4174 target_cputime64 = &cpustat->softirq;
4176 target_cputime64 = &cpustat->system;
4178 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
4182 * Account for involuntary wait time.
4183 * @cputime: the cpu time spent in involuntary wait
4185 void account_steal_time(cputime_t cputime)
4187 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4188 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4190 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4194 * Account for idle time.
4195 * @cputime: the cpu time spent in idle wait
4197 void account_idle_time(cputime_t cputime)
4199 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4200 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4201 struct rq *rq = this_rq();
4203 if (atomic_read(&rq->nr_iowait) > 0)
4204 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4206 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4209 static __always_inline bool steal_account_process_tick(void)
4211 #ifdef CONFIG_PARAVIRT
4212 if (static_branch(¶virt_steal_enabled)) {
4215 steal = paravirt_steal_clock(smp_processor_id());
4216 steal -= this_rq()->prev_steal_time;
4218 st = steal_ticks(steal);
4219 this_rq()->prev_steal_time += st * TICK_NSEC;
4221 account_steal_time(st);
4228 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4230 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4232 * Account a tick to a process and cpustat
4233 * @p: the process that the cpu time gets accounted to
4234 * @user_tick: is the tick from userspace
4235 * @rq: the pointer to rq
4237 * Tick demultiplexing follows the order
4238 * - pending hardirq update
4239 * - pending softirq update
4243 * - check for guest_time
4244 * - else account as system_time
4246 * Check for hardirq is done both for system and user time as there is
4247 * no timer going off while we are on hardirq and hence we may never get an
4248 * opportunity to update it solely in system time.
4249 * p->stime and friends are only updated on system time and not on irq
4250 * softirq as those do not count in task exec_runtime any more.
4252 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4255 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4256 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4257 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4259 if (steal_account_process_tick())
4262 if (irqtime_account_hi_update()) {
4263 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4264 } else if (irqtime_account_si_update()) {
4265 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4266 } else if (this_cpu_ksoftirqd() == p) {
4268 * ksoftirqd time do not get accounted in cpu_softirq_time.
4269 * So, we have to handle it separately here.
4270 * Also, p->stime needs to be updated for ksoftirqd.
4272 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4274 } else if (user_tick) {
4275 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4276 } else if (p == rq->idle) {
4277 account_idle_time(cputime_one_jiffy);
4278 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4279 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4281 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4286 static void irqtime_account_idle_ticks(int ticks)
4289 struct rq *rq = this_rq();
4291 for (i = 0; i < ticks; i++)
4292 irqtime_account_process_tick(current, 0, rq);
4294 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4295 static void irqtime_account_idle_ticks(int ticks) {}
4296 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4298 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4301 * Account a single tick of cpu time.
4302 * @p: the process that the cpu time gets accounted to
4303 * @user_tick: indicates if the tick is a user or a system tick
4305 void account_process_tick(struct task_struct *p, int user_tick)
4307 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4308 struct rq *rq = this_rq();
4310 if (sched_clock_irqtime) {
4311 irqtime_account_process_tick(p, user_tick, rq);
4315 if (steal_account_process_tick())
4319 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4320 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4321 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4324 account_idle_time(cputime_one_jiffy);
4328 * Account multiple ticks of steal time.
4329 * @p: the process from which the cpu time has been stolen
4330 * @ticks: number of stolen ticks
4332 void account_steal_ticks(unsigned long ticks)
4334 account_steal_time(jiffies_to_cputime(ticks));
4338 * Account multiple ticks of idle time.
4339 * @ticks: number of stolen ticks
4341 void account_idle_ticks(unsigned long ticks)
4344 if (sched_clock_irqtime) {
4345 irqtime_account_idle_ticks(ticks);
4349 account_idle_time(jiffies_to_cputime(ticks));
4355 * Use precise platform statistics if available:
4357 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4358 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4364 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4366 struct task_cputime cputime;
4368 thread_group_cputime(p, &cputime);
4370 *ut = cputime.utime;
4371 *st = cputime.stime;
4375 #ifndef nsecs_to_cputime
4376 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4379 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
4381 u64 temp = (__force u64) rtime;
4383 temp *= (__force u64) utime;
4385 if (sizeof(cputime_t) == 4)
4386 temp = div_u64(temp, (__force u32) total);
4388 temp = div64_u64(temp, (__force u64) total);
4390 return (__force cputime_t) temp;
4393 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4395 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4398 * Use CFS's precise accounting:
4400 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4403 utime = scale_utime(utime, rtime, total);
4408 * Compare with previous values, to keep monotonicity:
4410 p->prev_utime = max(p->prev_utime, utime);
4411 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4413 *ut = p->prev_utime;
4414 *st = p->prev_stime;
4418 * Must be called with siglock held.
4420 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4422 struct signal_struct *sig = p->signal;
4423 struct task_cputime cputime;
4424 cputime_t rtime, utime, total;
4426 thread_group_cputime(p, &cputime);
4428 total = cputime_add(cputime.utime, cputime.stime);
4429 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4432 utime = scale_utime(cputime.utime, rtime, total);
4436 sig->prev_utime = max(sig->prev_utime, utime);
4437 sig->prev_stime = max(sig->prev_stime,
4438 cputime_sub(rtime, sig->prev_utime));
4440 *ut = sig->prev_utime;
4441 *st = sig->prev_stime;
4446 * This function gets called by the timer code, with HZ frequency.
4447 * We call it with interrupts disabled.
4449 void scheduler_tick(void)
4451 int cpu = smp_processor_id();
4452 struct rq *rq = cpu_rq(cpu);
4453 struct task_struct *curr = rq->curr;
4457 raw_spin_lock(&rq->lock);
4458 update_rq_clock(rq);
4459 update_cpu_load_active(rq);
4460 curr->sched_class->task_tick(rq, curr, 0);
4461 raw_spin_unlock(&rq->lock);
4463 perf_event_task_tick();
4466 rq->idle_balance = idle_cpu(cpu);
4467 trigger_load_balance(rq, cpu);
4471 notrace unsigned long get_parent_ip(unsigned long addr)
4473 if (in_lock_functions(addr)) {
4474 addr = CALLER_ADDR2;
4475 if (in_lock_functions(addr))
4476 addr = CALLER_ADDR3;
4481 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4482 defined(CONFIG_PREEMPT_TRACER))
4484 void __kprobes add_preempt_count(int val)
4486 #ifdef CONFIG_DEBUG_PREEMPT
4490 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4493 preempt_count() += val;
4494 #ifdef CONFIG_DEBUG_PREEMPT
4496 * Spinlock count overflowing soon?
4498 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4501 if (preempt_count() == val)
4502 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4504 EXPORT_SYMBOL(add_preempt_count);
4506 void __kprobes sub_preempt_count(int val)
4508 #ifdef CONFIG_DEBUG_PREEMPT
4512 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4515 * Is the spinlock portion underflowing?
4517 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4518 !(preempt_count() & PREEMPT_MASK)))
4522 if (preempt_count() == val)
4523 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4524 preempt_count() -= val;
4526 EXPORT_SYMBOL(sub_preempt_count);
4531 * Print scheduling while atomic bug:
4533 static noinline void __schedule_bug(struct task_struct *prev)
4535 struct pt_regs *regs = get_irq_regs();
4537 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4538 prev->comm, prev->pid, preempt_count());
4540 debug_show_held_locks(prev);
4542 if (irqs_disabled())
4543 print_irqtrace_events(prev);
4552 * Various schedule()-time debugging checks and statistics:
4554 static inline void schedule_debug(struct task_struct *prev)
4557 * Test if we are atomic. Since do_exit() needs to call into
4558 * schedule() atomically, we ignore that path for now.
4559 * Otherwise, whine if we are scheduling when we should not be.
4561 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4562 __schedule_bug(prev);
4565 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4567 schedstat_inc(this_rq(), sched_count);
4570 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4572 if (prev->on_rq || rq->skip_clock_update < 0)
4573 update_rq_clock(rq);
4574 prev->sched_class->put_prev_task(rq, prev);
4578 * Pick up the highest-prio task:
4580 static inline struct task_struct *
4581 pick_next_task(struct rq *rq)
4583 const struct sched_class *class;
4584 struct task_struct *p;
4587 * Optimization: we know that if all tasks are in
4588 * the fair class we can call that function directly:
4590 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4591 p = fair_sched_class.pick_next_task(rq);
4596 for_each_class(class) {
4597 p = class->pick_next_task(rq);
4602 BUG(); /* the idle class will always have a runnable task */
4606 * __schedule() is the main scheduler function.
4608 static void __sched __schedule(void)
4610 struct task_struct *prev, *next;
4611 unsigned long *switch_count;
4617 cpu = smp_processor_id();
4619 rcu_note_context_switch(cpu);
4622 schedule_debug(prev);
4624 if (sched_feat(HRTICK))
4627 raw_spin_lock_irq(&rq->lock);
4629 switch_count = &prev->nivcsw;
4630 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4631 if (unlikely(signal_pending_state(prev->state, prev))) {
4632 prev->state = TASK_RUNNING;
4634 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4638 * If a worker went to sleep, notify and ask workqueue
4639 * whether it wants to wake up a task to maintain
4642 if (prev->flags & PF_WQ_WORKER) {
4643 struct task_struct *to_wakeup;
4645 to_wakeup = wq_worker_sleeping(prev, cpu);
4647 try_to_wake_up_local(to_wakeup);
4650 switch_count = &prev->nvcsw;
4653 pre_schedule(rq, prev);
4655 if (unlikely(!rq->nr_running))
4656 idle_balance(cpu, rq);
4658 put_prev_task(rq, prev);
4659 next = pick_next_task(rq);
4660 clear_tsk_need_resched(prev);
4661 rq->skip_clock_update = 0;
4663 if (likely(prev != next)) {
4668 context_switch(rq, prev, next); /* unlocks the rq */
4670 * The context switch have flipped the stack from under us
4671 * and restored the local variables which were saved when
4672 * this task called schedule() in the past. prev == current
4673 * is still correct, but it can be moved to another cpu/rq.
4675 cpu = smp_processor_id();
4678 raw_spin_unlock_irq(&rq->lock);
4682 preempt_enable_no_resched();
4687 static inline void sched_submit_work(struct task_struct *tsk)
4692 * If we are going to sleep and we have plugged IO queued,
4693 * make sure to submit it to avoid deadlocks.
4695 if (blk_needs_flush_plug(tsk))
4696 blk_schedule_flush_plug(tsk);
4699 asmlinkage void __sched schedule(void)
4701 struct task_struct *tsk = current;
4703 sched_submit_work(tsk);
4706 EXPORT_SYMBOL(schedule);
4708 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4710 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4712 if (lock->owner != owner)
4716 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4717 * lock->owner still matches owner, if that fails, owner might
4718 * point to free()d memory, if it still matches, the rcu_read_lock()
4719 * ensures the memory stays valid.
4723 return owner->on_cpu;
4727 * Look out! "owner" is an entirely speculative pointer
4728 * access and not reliable.
4730 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4732 if (!sched_feat(OWNER_SPIN))
4736 while (owner_running(lock, owner)) {
4740 arch_mutex_cpu_relax();
4745 * We break out the loop above on need_resched() and when the
4746 * owner changed, which is a sign for heavy contention. Return
4747 * success only when lock->owner is NULL.
4749 return lock->owner == NULL;
4753 #ifdef CONFIG_PREEMPT
4755 * this is the entry point to schedule() from in-kernel preemption
4756 * off of preempt_enable. Kernel preemptions off return from interrupt
4757 * occur there and call schedule directly.
4759 asmlinkage void __sched notrace preempt_schedule(void)
4761 struct thread_info *ti = current_thread_info();
4764 * If there is a non-zero preempt_count or interrupts are disabled,
4765 * we do not want to preempt the current task. Just return..
4767 if (likely(ti->preempt_count || irqs_disabled()))
4771 add_preempt_count_notrace(PREEMPT_ACTIVE);
4773 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4776 * Check again in case we missed a preemption opportunity
4777 * between schedule and now.
4780 } while (need_resched());
4782 EXPORT_SYMBOL(preempt_schedule);
4785 * this is the entry point to schedule() from kernel preemption
4786 * off of irq context.
4787 * Note, that this is called and return with irqs disabled. This will
4788 * protect us against recursive calling from irq.
4790 asmlinkage void __sched preempt_schedule_irq(void)
4792 struct thread_info *ti = current_thread_info();
4794 /* Catch callers which need to be fixed */
4795 BUG_ON(ti->preempt_count || !irqs_disabled());
4798 add_preempt_count(PREEMPT_ACTIVE);
4801 local_irq_disable();
4802 sub_preempt_count(PREEMPT_ACTIVE);
4805 * Check again in case we missed a preemption opportunity
4806 * between schedule and now.
4809 } while (need_resched());
4812 #endif /* CONFIG_PREEMPT */
4814 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4817 return try_to_wake_up(curr->private, mode, wake_flags);
4819 EXPORT_SYMBOL(default_wake_function);
4822 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4823 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4824 * number) then we wake all the non-exclusive tasks and one exclusive task.
4826 * There are circumstances in which we can try to wake a task which has already
4827 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4828 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4830 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4831 int nr_exclusive, int wake_flags, void *key)
4833 wait_queue_t *curr, *next;
4835 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4836 unsigned flags = curr->flags;
4838 if (curr->func(curr, mode, wake_flags, key) &&
4839 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4845 * __wake_up - wake up threads blocked on a waitqueue.
4847 * @mode: which threads
4848 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4849 * @key: is directly passed to the wakeup function
4851 * It may be assumed that this function implies a write memory barrier before
4852 * changing the task state if and only if any tasks are woken up.
4854 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4855 int nr_exclusive, void *key)
4857 unsigned long flags;
4859 spin_lock_irqsave(&q->lock, flags);
4860 __wake_up_common(q, mode, nr_exclusive, 0, key);
4861 spin_unlock_irqrestore(&q->lock, flags);
4863 EXPORT_SYMBOL(__wake_up);
4866 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4868 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4870 __wake_up_common(q, mode, 1, 0, NULL);
4872 EXPORT_SYMBOL_GPL(__wake_up_locked);
4874 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4876 __wake_up_common(q, mode, 1, 0, key);
4878 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4881 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4883 * @mode: which threads
4884 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4885 * @key: opaque value to be passed to wakeup targets
4887 * The sync wakeup differs that the waker knows that it will schedule
4888 * away soon, so while the target thread will be woken up, it will not
4889 * be migrated to another CPU - ie. the two threads are 'synchronized'
4890 * with each other. This can prevent needless bouncing between CPUs.
4892 * On UP it can prevent extra preemption.
4894 * It may be assumed that this function implies a write memory barrier before
4895 * changing the task state if and only if any tasks are woken up.
4897 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4898 int nr_exclusive, void *key)
4900 unsigned long flags;
4901 int wake_flags = WF_SYNC;
4906 if (unlikely(!nr_exclusive))
4909 spin_lock_irqsave(&q->lock, flags);
4910 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4911 spin_unlock_irqrestore(&q->lock, flags);
4913 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4916 * __wake_up_sync - see __wake_up_sync_key()
4918 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4920 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4922 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4925 * complete: - signals a single thread waiting on this completion
4926 * @x: holds the state of this particular completion
4928 * This will wake up a single thread waiting on this completion. Threads will be
4929 * awakened in the same order in which they were queued.
4931 * See also complete_all(), wait_for_completion() and related routines.
4933 * It may be assumed that this function implies a write memory barrier before
4934 * changing the task state if and only if any tasks are woken up.
4936 void complete(struct completion *x)
4938 unsigned long flags;
4940 spin_lock_irqsave(&x->wait.lock, flags);
4942 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4943 spin_unlock_irqrestore(&x->wait.lock, flags);
4945 EXPORT_SYMBOL(complete);
4948 * complete_all: - signals all threads waiting on this completion
4949 * @x: holds the state of this particular completion
4951 * This will wake up all threads waiting on this particular completion event.
4953 * It may be assumed that this function implies a write memory barrier before
4954 * changing the task state if and only if any tasks are woken up.
4956 void complete_all(struct completion *x)
4958 unsigned long flags;
4960 spin_lock_irqsave(&x->wait.lock, flags);
4961 x->done += UINT_MAX/2;
4962 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4963 spin_unlock_irqrestore(&x->wait.lock, flags);
4965 EXPORT_SYMBOL(complete_all);
4967 static inline long __sched
4968 do_wait_for_common(struct completion *x, long timeout, int state)
4971 DECLARE_WAITQUEUE(wait, current);
4973 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4975 if (signal_pending_state(state, current)) {
4976 timeout = -ERESTARTSYS;
4979 __set_current_state(state);
4980 spin_unlock_irq(&x->wait.lock);
4981 timeout = schedule_timeout(timeout);
4982 spin_lock_irq(&x->wait.lock);
4983 } while (!x->done && timeout);
4984 __remove_wait_queue(&x->wait, &wait);
4989 return timeout ?: 1;
4993 wait_for_common(struct completion *x, long timeout, int state)
4997 spin_lock_irq(&x->wait.lock);
4998 timeout = do_wait_for_common(x, timeout, state);
4999 spin_unlock_irq(&x->wait.lock);
5004 * wait_for_completion: - waits for completion of a task
5005 * @x: holds the state of this particular completion
5007 * This waits to be signaled for completion of a specific task. It is NOT
5008 * interruptible and there is no timeout.
5010 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5011 * and interrupt capability. Also see complete().
5013 void __sched wait_for_completion(struct completion *x)
5015 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5017 EXPORT_SYMBOL(wait_for_completion);
5020 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5021 * @x: holds the state of this particular completion
5022 * @timeout: timeout value in jiffies
5024 * This waits for either a completion of a specific task to be signaled or for a
5025 * specified timeout to expire. The timeout is in jiffies. It is not
5028 * The return value is 0 if timed out, and positive (at least 1, or number of
5029 * jiffies left till timeout) if completed.
5031 unsigned long __sched
5032 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5034 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5036 EXPORT_SYMBOL(wait_for_completion_timeout);
5039 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5040 * @x: holds the state of this particular completion
5042 * This waits for completion of a specific task to be signaled. It is
5045 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5047 int __sched wait_for_completion_interruptible(struct completion *x)
5049 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5050 if (t == -ERESTARTSYS)
5054 EXPORT_SYMBOL(wait_for_completion_interruptible);
5057 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5058 * @x: holds the state of this particular completion
5059 * @timeout: timeout value in jiffies
5061 * This waits for either a completion of a specific task to be signaled or for a
5062 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5064 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5065 * positive (at least 1, or number of jiffies left till timeout) if completed.
5068 wait_for_completion_interruptible_timeout(struct completion *x,
5069 unsigned long timeout)
5071 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5073 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5076 * wait_for_completion_killable: - waits for completion of a task (killable)
5077 * @x: holds the state of this particular completion
5079 * This waits to be signaled for completion of a specific task. It can be
5080 * interrupted by a kill signal.
5082 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
5084 int __sched wait_for_completion_killable(struct completion *x)
5086 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5087 if (t == -ERESTARTSYS)
5091 EXPORT_SYMBOL(wait_for_completion_killable);
5094 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
5095 * @x: holds the state of this particular completion
5096 * @timeout: timeout value in jiffies
5098 * This waits for either a completion of a specific task to be
5099 * signaled or for a specified timeout to expire. It can be
5100 * interrupted by a kill signal. The timeout is in jiffies.
5102 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
5103 * positive (at least 1, or number of jiffies left till timeout) if completed.
5106 wait_for_completion_killable_timeout(struct completion *x,
5107 unsigned long timeout)
5109 return wait_for_common(x, timeout, TASK_KILLABLE);
5111 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
5114 * try_wait_for_completion - try to decrement a completion without blocking
5115 * @x: completion structure
5117 * Returns: 0 if a decrement cannot be done without blocking
5118 * 1 if a decrement succeeded.
5120 * If a completion is being used as a counting completion,
5121 * attempt to decrement the counter without blocking. This
5122 * enables us to avoid waiting if the resource the completion
5123 * is protecting is not available.
5125 bool try_wait_for_completion(struct completion *x)
5127 unsigned long flags;
5130 spin_lock_irqsave(&x->wait.lock, flags);
5135 spin_unlock_irqrestore(&x->wait.lock, flags);
5138 EXPORT_SYMBOL(try_wait_for_completion);
5141 * completion_done - Test to see if a completion has any waiters
5142 * @x: completion structure
5144 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5145 * 1 if there are no waiters.
5148 bool completion_done(struct completion *x)
5150 unsigned long flags;
5153 spin_lock_irqsave(&x->wait.lock, flags);
5156 spin_unlock_irqrestore(&x->wait.lock, flags);
5159 EXPORT_SYMBOL(completion_done);
5162 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5164 unsigned long flags;
5167 init_waitqueue_entry(&wait, current);
5169 __set_current_state(state);
5171 spin_lock_irqsave(&q->lock, flags);
5172 __add_wait_queue(q, &wait);
5173 spin_unlock(&q->lock);
5174 timeout = schedule_timeout(timeout);
5175 spin_lock_irq(&q->lock);
5176 __remove_wait_queue(q, &wait);
5177 spin_unlock_irqrestore(&q->lock, flags);
5182 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5184 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5186 EXPORT_SYMBOL(interruptible_sleep_on);
5189 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5191 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5193 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5195 void __sched sleep_on(wait_queue_head_t *q)
5197 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5199 EXPORT_SYMBOL(sleep_on);
5201 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5203 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5205 EXPORT_SYMBOL(sleep_on_timeout);
5207 #ifdef CONFIG_RT_MUTEXES
5210 * rt_mutex_setprio - set the current priority of a task
5212 * @prio: prio value (kernel-internal form)
5214 * This function changes the 'effective' priority of a task. It does
5215 * not touch ->normal_prio like __setscheduler().
5217 * Used by the rt_mutex code to implement priority inheritance logic.
5219 void rt_mutex_setprio(struct task_struct *p, int prio)
5221 int oldprio, on_rq, running;
5223 const struct sched_class *prev_class;
5225 BUG_ON(prio < 0 || prio > MAX_PRIO);
5227 rq = __task_rq_lock(p);
5229 trace_sched_pi_setprio(p, prio);
5231 prev_class = p->sched_class;
5233 running = task_current(rq, p);
5235 dequeue_task(rq, p, 0);
5237 p->sched_class->put_prev_task(rq, p);
5240 p->sched_class = &rt_sched_class;
5242 if (rt_prio(oldprio))
5244 p->sched_class = &fair_sched_class;
5250 p->sched_class->set_curr_task(rq);
5252 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5254 check_class_changed(rq, p, prev_class, oldprio);
5255 __task_rq_unlock(rq);
5260 void set_user_nice(struct task_struct *p, long nice)
5262 int old_prio, delta, on_rq;
5263 unsigned long flags;
5266 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5269 * We have to be careful, if called from sys_setpriority(),
5270 * the task might be in the middle of scheduling on another CPU.
5272 rq = task_rq_lock(p, &flags);
5274 * The RT priorities are set via sched_setscheduler(), but we still
5275 * allow the 'normal' nice value to be set - but as expected
5276 * it wont have any effect on scheduling until the task is
5277 * SCHED_FIFO/SCHED_RR:
5279 if (task_has_rt_policy(p)) {
5280 p->static_prio = NICE_TO_PRIO(nice);
5285 dequeue_task(rq, p, 0);
5287 p->static_prio = NICE_TO_PRIO(nice);
5290 p->prio = effective_prio(p);
5291 delta = p->prio - old_prio;
5294 enqueue_task(rq, p, 0);
5296 * If the task increased its priority or is running and
5297 * lowered its priority, then reschedule its CPU:
5299 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5300 resched_task(rq->curr);
5303 task_rq_unlock(rq, p, &flags);
5305 EXPORT_SYMBOL(set_user_nice);
5308 * can_nice - check if a task can reduce its nice value
5312 int can_nice(const struct task_struct *p, const int nice)
5314 /* convert nice value [19,-20] to rlimit style value [1,40] */
5315 int nice_rlim = 20 - nice;
5317 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5318 capable(CAP_SYS_NICE));
5321 #ifdef __ARCH_WANT_SYS_NICE
5324 * sys_nice - change the priority of the current process.
5325 * @increment: priority increment
5327 * sys_setpriority is a more generic, but much slower function that
5328 * does similar things.
5330 SYSCALL_DEFINE1(nice, int, increment)
5335 * Setpriority might change our priority at the same moment.
5336 * We don't have to worry. Conceptually one call occurs first
5337 * and we have a single winner.
5339 if (increment < -40)
5344 nice = TASK_NICE(current) + increment;
5350 if (increment < 0 && !can_nice(current, nice))
5353 retval = security_task_setnice(current, nice);
5357 set_user_nice(current, nice);
5364 * task_prio - return the priority value of a given task.
5365 * @p: the task in question.
5367 * This is the priority value as seen by users in /proc.
5368 * RT tasks are offset by -200. Normal tasks are centered
5369 * around 0, value goes from -16 to +15.
5371 int task_prio(const struct task_struct *p)
5373 return p->prio - MAX_RT_PRIO;
5377 * task_nice - return the nice value of a given task.
5378 * @p: the task in question.
5380 int task_nice(const struct task_struct *p)
5382 return TASK_NICE(p);
5384 EXPORT_SYMBOL(task_nice);
5387 * idle_cpu - is a given cpu idle currently?
5388 * @cpu: the processor in question.
5390 int idle_cpu(int cpu)
5392 struct rq *rq = cpu_rq(cpu);
5394 if (rq->curr != rq->idle)
5401 if (!llist_empty(&rq->wake_list))
5409 * idle_task - return the idle task for a given cpu.
5410 * @cpu: the processor in question.
5412 struct task_struct *idle_task(int cpu)
5414 return cpu_rq(cpu)->idle;
5418 * find_process_by_pid - find a process with a matching PID value.
5419 * @pid: the pid in question.
5421 static struct task_struct *find_process_by_pid(pid_t pid)
5423 return pid ? find_task_by_vpid(pid) : current;
5426 /* Actually do priority change: must hold rq lock. */
5428 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5431 p->rt_priority = prio;
5432 p->normal_prio = normal_prio(p);
5433 /* we are holding p->pi_lock already */
5434 p->prio = rt_mutex_getprio(p);
5435 if (rt_prio(p->prio))
5436 p->sched_class = &rt_sched_class;
5438 p->sched_class = &fair_sched_class;
5443 * check the target process has a UID that matches the current process's
5445 static bool check_same_owner(struct task_struct *p)
5447 const struct cred *cred = current_cred(), *pcred;
5451 pcred = __task_cred(p);
5452 if (cred->user->user_ns == pcred->user->user_ns)
5453 match = (cred->euid == pcred->euid ||
5454 cred->euid == pcred->uid);
5461 static int __sched_setscheduler(struct task_struct *p, int policy,
5462 const struct sched_param *param, bool user)
5464 int retval, oldprio, oldpolicy = -1, on_rq, running;
5465 unsigned long flags;
5466 const struct sched_class *prev_class;
5470 /* may grab non-irq protected spin_locks */
5471 BUG_ON(in_interrupt());
5473 /* double check policy once rq lock held */
5475 reset_on_fork = p->sched_reset_on_fork;
5476 policy = oldpolicy = p->policy;
5478 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5479 policy &= ~SCHED_RESET_ON_FORK;
5481 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5482 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5483 policy != SCHED_IDLE)
5488 * Valid priorities for SCHED_FIFO and SCHED_RR are
5489 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5490 * SCHED_BATCH and SCHED_IDLE is 0.
5492 if (param->sched_priority < 0 ||
5493 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5494 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5496 if (rt_policy(policy) != (param->sched_priority != 0))
5500 * Allow unprivileged RT tasks to decrease priority:
5502 if (user && !capable(CAP_SYS_NICE)) {
5503 if (rt_policy(policy)) {
5504 unsigned long rlim_rtprio =
5505 task_rlimit(p, RLIMIT_RTPRIO);
5507 /* can't set/change the rt policy */
5508 if (policy != p->policy && !rlim_rtprio)
5511 /* can't increase priority */
5512 if (param->sched_priority > p->rt_priority &&
5513 param->sched_priority > rlim_rtprio)
5518 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5519 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5521 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5522 if (!can_nice(p, TASK_NICE(p)))
5526 /* can't change other user's priorities */
5527 if (!check_same_owner(p))
5530 /* Normal users shall not reset the sched_reset_on_fork flag */
5531 if (p->sched_reset_on_fork && !reset_on_fork)
5536 retval = security_task_setscheduler(p);
5542 * make sure no PI-waiters arrive (or leave) while we are
5543 * changing the priority of the task:
5545 * To be able to change p->policy safely, the appropriate
5546 * runqueue lock must be held.
5548 rq = task_rq_lock(p, &flags);
5551 * Changing the policy of the stop threads its a very bad idea
5553 if (p == rq->stop) {
5554 task_rq_unlock(rq, p, &flags);
5559 * If not changing anything there's no need to proceed further:
5561 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5562 param->sched_priority == p->rt_priority))) {
5564 __task_rq_unlock(rq);
5565 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5569 #ifdef CONFIG_RT_GROUP_SCHED
5572 * Do not allow realtime tasks into groups that have no runtime
5575 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5576 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5577 !task_group_is_autogroup(task_group(p))) {
5578 task_rq_unlock(rq, p, &flags);
5584 /* recheck policy now with rq lock held */
5585 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5586 policy = oldpolicy = -1;
5587 task_rq_unlock(rq, p, &flags);
5591 running = task_current(rq, p);
5593 deactivate_task(rq, p, 0);
5595 p->sched_class->put_prev_task(rq, p);
5597 p->sched_reset_on_fork = reset_on_fork;
5600 prev_class = p->sched_class;
5601 __setscheduler(rq, p, policy, param->sched_priority);
5604 p->sched_class->set_curr_task(rq);
5606 activate_task(rq, p, 0);
5608 check_class_changed(rq, p, prev_class, oldprio);
5609 task_rq_unlock(rq, p, &flags);
5611 rt_mutex_adjust_pi(p);
5617 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5618 * @p: the task in question.
5619 * @policy: new policy.
5620 * @param: structure containing the new RT priority.
5622 * NOTE that the task may be already dead.
5624 int sched_setscheduler(struct task_struct *p, int policy,
5625 const struct sched_param *param)
5627 return __sched_setscheduler(p, policy, param, true);
5629 EXPORT_SYMBOL_GPL(sched_setscheduler);
5632 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5633 * @p: the task in question.
5634 * @policy: new policy.
5635 * @param: structure containing the new RT priority.
5637 * Just like sched_setscheduler, only don't bother checking if the
5638 * current context has permission. For example, this is needed in
5639 * stop_machine(): we create temporary high priority worker threads,
5640 * but our caller might not have that capability.
5642 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5643 const struct sched_param *param)
5645 return __sched_setscheduler(p, policy, param, false);
5649 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5651 struct sched_param lparam;
5652 struct task_struct *p;
5655 if (!param || pid < 0)
5657 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5662 p = find_process_by_pid(pid);
5664 retval = sched_setscheduler(p, policy, &lparam);
5671 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5672 * @pid: the pid in question.
5673 * @policy: new policy.
5674 * @param: structure containing the new RT priority.
5676 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5677 struct sched_param __user *, param)
5679 /* negative values for policy are not valid */
5683 return do_sched_setscheduler(pid, policy, param);
5687 * sys_sched_setparam - set/change the RT priority of a thread
5688 * @pid: the pid in question.
5689 * @param: structure containing the new RT priority.
5691 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5693 return do_sched_setscheduler(pid, -1, param);
5697 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5698 * @pid: the pid in question.
5700 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5702 struct task_struct *p;
5710 p = find_process_by_pid(pid);
5712 retval = security_task_getscheduler(p);
5715 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5722 * sys_sched_getparam - get the RT priority of a thread
5723 * @pid: the pid in question.
5724 * @param: structure containing the RT priority.
5726 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5728 struct sched_param lp;
5729 struct task_struct *p;
5732 if (!param || pid < 0)
5736 p = find_process_by_pid(pid);
5741 retval = security_task_getscheduler(p);
5745 lp.sched_priority = p->rt_priority;
5749 * This one might sleep, we cannot do it with a spinlock held ...
5751 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5760 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5762 cpumask_var_t cpus_allowed, new_mask;
5763 struct task_struct *p;
5769 p = find_process_by_pid(pid);
5776 /* Prevent p going away */
5780 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5784 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5786 goto out_free_cpus_allowed;
5789 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5792 retval = security_task_setscheduler(p);
5796 cpuset_cpus_allowed(p, cpus_allowed);
5797 cpumask_and(new_mask, in_mask, cpus_allowed);
5799 retval = set_cpus_allowed_ptr(p, new_mask);
5802 cpuset_cpus_allowed(p, cpus_allowed);
5803 if (!cpumask_subset(new_mask, cpus_allowed)) {
5805 * We must have raced with a concurrent cpuset
5806 * update. Just reset the cpus_allowed to the
5807 * cpuset's cpus_allowed
5809 cpumask_copy(new_mask, cpus_allowed);
5814 free_cpumask_var(new_mask);
5815 out_free_cpus_allowed:
5816 free_cpumask_var(cpus_allowed);
5823 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5824 struct cpumask *new_mask)
5826 if (len < cpumask_size())
5827 cpumask_clear(new_mask);
5828 else if (len > cpumask_size())
5829 len = cpumask_size();
5831 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5835 * sys_sched_setaffinity - set the cpu affinity of a process
5836 * @pid: pid of the process
5837 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5838 * @user_mask_ptr: user-space pointer to the new cpu mask
5840 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5841 unsigned long __user *, user_mask_ptr)
5843 cpumask_var_t new_mask;
5846 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5849 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5851 retval = sched_setaffinity(pid, new_mask);
5852 free_cpumask_var(new_mask);
5856 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5858 struct task_struct *p;
5859 unsigned long flags;
5866 p = find_process_by_pid(pid);
5870 retval = security_task_getscheduler(p);
5874 raw_spin_lock_irqsave(&p->pi_lock, flags);
5875 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5876 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5886 * sys_sched_getaffinity - get the cpu affinity of a process
5887 * @pid: pid of the process
5888 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5889 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5891 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5892 unsigned long __user *, user_mask_ptr)
5897 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5899 if (len & (sizeof(unsigned long)-1))
5902 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5905 ret = sched_getaffinity(pid, mask);
5907 size_t retlen = min_t(size_t, len, cpumask_size());
5909 if (copy_to_user(user_mask_ptr, mask, retlen))
5914 free_cpumask_var(mask);
5920 * sys_sched_yield - yield the current processor to other threads.
5922 * This function yields the current CPU to other tasks. If there are no
5923 * other threads running on this CPU then this function will return.
5925 SYSCALL_DEFINE0(sched_yield)
5927 struct rq *rq = this_rq_lock();
5929 schedstat_inc(rq, yld_count);
5930 current->sched_class->yield_task(rq);
5933 * Since we are going to call schedule() anyway, there's
5934 * no need to preempt or enable interrupts:
5936 __release(rq->lock);
5937 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5938 do_raw_spin_unlock(&rq->lock);
5939 preempt_enable_no_resched();
5946 static inline int should_resched(void)
5948 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5951 static void __cond_resched(void)
5953 add_preempt_count(PREEMPT_ACTIVE);
5955 sub_preempt_count(PREEMPT_ACTIVE);
5958 int __sched _cond_resched(void)
5960 if (should_resched()) {
5966 EXPORT_SYMBOL(_cond_resched);
5969 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5970 * call schedule, and on return reacquire the lock.
5972 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5973 * operations here to prevent schedule() from being called twice (once via
5974 * spin_unlock(), once by hand).
5976 int __cond_resched_lock(spinlock_t *lock)
5978 int resched = should_resched();
5981 lockdep_assert_held(lock);
5983 if (spin_needbreak(lock) || resched) {
5994 EXPORT_SYMBOL(__cond_resched_lock);
5996 int __sched __cond_resched_softirq(void)
5998 BUG_ON(!in_softirq());
6000 if (should_resched()) {
6008 EXPORT_SYMBOL(__cond_resched_softirq);
6011 * yield - yield the current processor to other threads.
6013 * This is a shortcut for kernel-space yielding - it marks the
6014 * thread runnable and calls sys_sched_yield().
6016 void __sched yield(void)
6018 set_current_state(TASK_RUNNING);
6021 EXPORT_SYMBOL(yield);
6024 * yield_to - yield the current processor to another thread in
6025 * your thread group, or accelerate that thread toward the
6026 * processor it's on.
6028 * @preempt: whether task preemption is allowed or not
6030 * It's the caller's job to ensure that the target task struct
6031 * can't go away on us before we can do any checks.
6033 * Returns true if we indeed boosted the target task.
6035 bool __sched yield_to(struct task_struct *p, bool preempt)
6037 struct task_struct *curr = current;
6038 struct rq *rq, *p_rq;
6039 unsigned long flags;
6042 local_irq_save(flags);
6047 double_rq_lock(rq, p_rq);
6048 while (task_rq(p) != p_rq) {
6049 double_rq_unlock(rq, p_rq);
6053 if (!curr->sched_class->yield_to_task)
6056 if (curr->sched_class != p->sched_class)
6059 if (task_running(p_rq, p) || p->state)
6062 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
6064 schedstat_inc(rq, yld_count);
6066 * Make p's CPU reschedule; pick_next_entity takes care of
6069 if (preempt && rq != p_rq)
6070 resched_task(p_rq->curr);
6074 double_rq_unlock(rq, p_rq);
6075 local_irq_restore(flags);
6082 EXPORT_SYMBOL_GPL(yield_to);
6085 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6086 * that process accounting knows that this is a task in IO wait state.
6088 void __sched io_schedule(void)
6090 struct rq *rq = raw_rq();
6092 delayacct_blkio_start();
6093 atomic_inc(&rq->nr_iowait);
6094 blk_flush_plug(current);
6095 current->in_iowait = 1;
6097 current->in_iowait = 0;
6098 atomic_dec(&rq->nr_iowait);
6099 delayacct_blkio_end();
6101 EXPORT_SYMBOL(io_schedule);
6103 long __sched io_schedule_timeout(long timeout)
6105 struct rq *rq = raw_rq();
6108 delayacct_blkio_start();
6109 atomic_inc(&rq->nr_iowait);
6110 blk_flush_plug(current);
6111 current->in_iowait = 1;
6112 ret = schedule_timeout(timeout);
6113 current->in_iowait = 0;
6114 atomic_dec(&rq->nr_iowait);
6115 delayacct_blkio_end();
6120 * sys_sched_get_priority_max - return maximum RT priority.
6121 * @policy: scheduling class.
6123 * this syscall returns the maximum rt_priority that can be used
6124 * by a given scheduling class.
6126 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6133 ret = MAX_USER_RT_PRIO-1;
6145 * sys_sched_get_priority_min - return minimum RT priority.
6146 * @policy: scheduling class.
6148 * this syscall returns the minimum rt_priority that can be used
6149 * by a given scheduling class.
6151 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6169 * sys_sched_rr_get_interval - return the default timeslice of a process.
6170 * @pid: pid of the process.
6171 * @interval: userspace pointer to the timeslice value.
6173 * this syscall writes the default timeslice value of a given process
6174 * into the user-space timespec buffer. A value of '0' means infinity.
6176 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6177 struct timespec __user *, interval)
6179 struct task_struct *p;
6180 unsigned int time_slice;
6181 unsigned long flags;
6191 p = find_process_by_pid(pid);
6195 retval = security_task_getscheduler(p);
6199 rq = task_rq_lock(p, &flags);
6200 time_slice = p->sched_class->get_rr_interval(rq, p);
6201 task_rq_unlock(rq, p, &flags);
6204 jiffies_to_timespec(time_slice, &t);
6205 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6213 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6215 void sched_show_task(struct task_struct *p)
6217 unsigned long free = 0;
6220 state = p->state ? __ffs(p->state) + 1 : 0;
6221 printk(KERN_INFO "%-15.15s %c", p->comm,
6222 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6223 #if BITS_PER_LONG == 32
6224 if (state == TASK_RUNNING)
6225 printk(KERN_CONT " running ");
6227 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6229 if (state == TASK_RUNNING)
6230 printk(KERN_CONT " running task ");
6232 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6234 #ifdef CONFIG_DEBUG_STACK_USAGE
6235 free = stack_not_used(p);
6237 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6238 task_pid_nr(p), task_pid_nr(p->real_parent),
6239 (unsigned long)task_thread_info(p)->flags);
6241 show_stack(p, NULL);
6244 void show_state_filter(unsigned long state_filter)
6246 struct task_struct *g, *p;
6248 #if BITS_PER_LONG == 32
6250 " task PC stack pid father\n");
6253 " task PC stack pid father\n");
6256 do_each_thread(g, p) {
6258 * reset the NMI-timeout, listing all files on a slow
6259 * console might take a lot of time:
6261 touch_nmi_watchdog();
6262 if (!state_filter || (p->state & state_filter))
6264 } while_each_thread(g, p);
6266 touch_all_softlockup_watchdogs();
6268 #ifdef CONFIG_SCHED_DEBUG
6269 sysrq_sched_debug_show();
6273 * Only show locks if all tasks are dumped:
6276 debug_show_all_locks();
6279 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6281 idle->sched_class = &idle_sched_class;
6285 * init_idle - set up an idle thread for a given CPU
6286 * @idle: task in question
6287 * @cpu: cpu the idle task belongs to
6289 * NOTE: this function does not set the idle thread's NEED_RESCHED
6290 * flag, to make booting more robust.
6292 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6294 struct rq *rq = cpu_rq(cpu);
6295 unsigned long flags;
6297 raw_spin_lock_irqsave(&rq->lock, flags);
6300 idle->state = TASK_RUNNING;
6301 idle->se.exec_start = sched_clock();
6303 do_set_cpus_allowed(idle, cpumask_of(cpu));
6305 * We're having a chicken and egg problem, even though we are
6306 * holding rq->lock, the cpu isn't yet set to this cpu so the
6307 * lockdep check in task_group() will fail.
6309 * Similar case to sched_fork(). / Alternatively we could
6310 * use task_rq_lock() here and obtain the other rq->lock.
6315 __set_task_cpu(idle, cpu);
6318 rq->curr = rq->idle = idle;
6319 #if defined(CONFIG_SMP)
6322 raw_spin_unlock_irqrestore(&rq->lock, flags);
6324 /* Set the preempt count _outside_ the spinlocks! */
6325 task_thread_info(idle)->preempt_count = 0;
6328 * The idle tasks have their own, simple scheduling class:
6330 idle->sched_class = &idle_sched_class;
6331 ftrace_graph_init_idle_task(idle, cpu);
6332 #if defined(CONFIG_SMP)
6333 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
6338 * Increase the granularity value when there are more CPUs,
6339 * because with more CPUs the 'effective latency' as visible
6340 * to users decreases. But the relationship is not linear,
6341 * so pick a second-best guess by going with the log2 of the
6344 * This idea comes from the SD scheduler of Con Kolivas:
6346 static int get_update_sysctl_factor(void)
6348 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6349 unsigned int factor;
6351 switch (sysctl_sched_tunable_scaling) {
6352 case SCHED_TUNABLESCALING_NONE:
6355 case SCHED_TUNABLESCALING_LINEAR:
6358 case SCHED_TUNABLESCALING_LOG:
6360 factor = 1 + ilog2(cpus);
6367 static void update_sysctl(void)
6369 unsigned int factor = get_update_sysctl_factor();
6371 #define SET_SYSCTL(name) \
6372 (sysctl_##name = (factor) * normalized_sysctl_##name)
6373 SET_SYSCTL(sched_min_granularity);
6374 SET_SYSCTL(sched_latency);
6375 SET_SYSCTL(sched_wakeup_granularity);
6379 static inline void sched_init_granularity(void)
6385 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6387 if (p->sched_class && p->sched_class->set_cpus_allowed)
6388 p->sched_class->set_cpus_allowed(p, new_mask);
6390 cpumask_copy(&p->cpus_allowed, new_mask);
6391 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6395 * This is how migration works:
6397 * 1) we invoke migration_cpu_stop() on the target CPU using
6399 * 2) stopper starts to run (implicitly forcing the migrated thread
6401 * 3) it checks whether the migrated task is still in the wrong runqueue.
6402 * 4) if it's in the wrong runqueue then the migration thread removes
6403 * it and puts it into the right queue.
6404 * 5) stopper completes and stop_one_cpu() returns and the migration
6409 * Change a given task's CPU affinity. Migrate the thread to a
6410 * proper CPU and schedule it away if the CPU it's executing on
6411 * is removed from the allowed bitmask.
6413 * NOTE: the caller must have a valid reference to the task, the
6414 * task must not exit() & deallocate itself prematurely. The
6415 * call is not atomic; no spinlocks may be held.
6417 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6419 unsigned long flags;
6421 unsigned int dest_cpu;
6424 rq = task_rq_lock(p, &flags);
6426 if (cpumask_equal(&p->cpus_allowed, new_mask))
6429 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6434 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6439 do_set_cpus_allowed(p, new_mask);
6441 /* Can the task run on the task's current CPU? If so, we're done */
6442 if (cpumask_test_cpu(task_cpu(p), new_mask))
6445 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6447 struct migration_arg arg = { p, dest_cpu };
6448 /* Need help from migration thread: drop lock and wait. */
6449 task_rq_unlock(rq, p, &flags);
6450 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6451 tlb_migrate_finish(p->mm);
6455 task_rq_unlock(rq, p, &flags);
6459 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6462 * Move (not current) task off this cpu, onto dest cpu. We're doing
6463 * this because either it can't run here any more (set_cpus_allowed()
6464 * away from this CPU, or CPU going down), or because we're
6465 * attempting to rebalance this task on exec (sched_exec).
6467 * So we race with normal scheduler movements, but that's OK, as long
6468 * as the task is no longer on this CPU.
6470 * Returns non-zero if task was successfully migrated.
6472 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6474 struct rq *rq_dest, *rq_src;
6477 if (unlikely(!cpu_active(dest_cpu)))
6480 rq_src = cpu_rq(src_cpu);
6481 rq_dest = cpu_rq(dest_cpu);
6483 raw_spin_lock(&p->pi_lock);
6484 double_rq_lock(rq_src, rq_dest);
6485 /* Already moved. */
6486 if (task_cpu(p) != src_cpu)
6488 /* Affinity changed (again). */
6489 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
6493 * If we're not on a rq, the next wake-up will ensure we're
6497 deactivate_task(rq_src, p, 0);
6498 set_task_cpu(p, dest_cpu);
6499 activate_task(rq_dest, p, 0);
6500 check_preempt_curr(rq_dest, p, 0);
6505 double_rq_unlock(rq_src, rq_dest);
6506 raw_spin_unlock(&p->pi_lock);
6511 * migration_cpu_stop - this will be executed by a highprio stopper thread
6512 * and performs thread migration by bumping thread off CPU then
6513 * 'pushing' onto another runqueue.
6515 static int migration_cpu_stop(void *data)
6517 struct migration_arg *arg = data;
6520 * The original target cpu might have gone down and we might
6521 * be on another cpu but it doesn't matter.
6523 local_irq_disable();
6524 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6529 #ifdef CONFIG_HOTPLUG_CPU
6532 * Ensures that the idle task is using init_mm right before its cpu goes
6535 void idle_task_exit(void)
6537 struct mm_struct *mm = current->active_mm;
6539 BUG_ON(cpu_online(smp_processor_id()));
6542 switch_mm(mm, &init_mm, current);
6547 * While a dead CPU has no uninterruptible tasks queued at this point,
6548 * it might still have a nonzero ->nr_uninterruptible counter, because
6549 * for performance reasons the counter is not stricly tracking tasks to
6550 * their home CPUs. So we just add the counter to another CPU's counter,
6551 * to keep the global sum constant after CPU-down:
6553 static void migrate_nr_uninterruptible(struct rq *rq_src)
6555 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6557 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6558 rq_src->nr_uninterruptible = 0;
6562 * remove the tasks which were accounted by rq from calc_load_tasks.
6564 static void calc_global_load_remove(struct rq *rq)
6566 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6567 rq->calc_load_active = 0;
6570 #ifdef CONFIG_CFS_BANDWIDTH
6571 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6573 struct cfs_rq *cfs_rq;
6575 for_each_leaf_cfs_rq(rq, cfs_rq) {
6576 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6578 if (!cfs_rq->runtime_enabled)
6582 * clock_task is not advancing so we just need to make sure
6583 * there's some valid quota amount
6585 cfs_rq->runtime_remaining = cfs_b->quota;
6586 if (cfs_rq_throttled(cfs_rq))
6587 unthrottle_cfs_rq(cfs_rq);
6593 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6594 * try_to_wake_up()->select_task_rq().
6596 * Called with rq->lock held even though we'er in stop_machine() and
6597 * there's no concurrency possible, we hold the required locks anyway
6598 * because of lock validation efforts.
6600 static void migrate_tasks(unsigned int dead_cpu)
6602 struct rq *rq = cpu_rq(dead_cpu);
6603 struct task_struct *next, *stop = rq->stop;
6607 * Fudge the rq selection such that the below task selection loop
6608 * doesn't get stuck on the currently eligible stop task.
6610 * We're currently inside stop_machine() and the rq is either stuck
6611 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6612 * either way we should never end up calling schedule() until we're
6619 * There's this thread running, bail when that's the only
6622 if (rq->nr_running == 1)
6625 next = pick_next_task(rq);
6627 next->sched_class->put_prev_task(rq, next);
6629 /* Find suitable destination for @next, with force if needed. */
6630 dest_cpu = select_fallback_rq(dead_cpu, next);
6631 raw_spin_unlock(&rq->lock);
6633 __migrate_task(next, dead_cpu, dest_cpu);
6635 raw_spin_lock(&rq->lock);
6641 #endif /* CONFIG_HOTPLUG_CPU */
6643 #if !defined(CONFIG_HOTPLUG_CPU) || !defined(CONFIG_CFS_BANDWIDTH)
6644 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6647 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6649 static struct ctl_table sd_ctl_dir[] = {
6651 .procname = "sched_domain",
6657 static struct ctl_table sd_ctl_root[] = {
6659 .procname = "kernel",
6661 .child = sd_ctl_dir,
6666 static struct ctl_table *sd_alloc_ctl_entry(int n)
6668 struct ctl_table *entry =
6669 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6674 static void sd_free_ctl_entry(struct ctl_table **tablep)
6676 struct ctl_table *entry;
6679 * In the intermediate directories, both the child directory and
6680 * procname are dynamically allocated and could fail but the mode
6681 * will always be set. In the lowest directory the names are
6682 * static strings and all have proc handlers.
6684 for (entry = *tablep; entry->mode; entry++) {
6686 sd_free_ctl_entry(&entry->child);
6687 if (entry->proc_handler == NULL)
6688 kfree(entry->procname);
6695 static int min_load_idx = 0;
6696 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
6699 set_table_entry(struct ctl_table *entry,
6700 const char *procname, void *data, int maxlen,
6701 mode_t mode, proc_handler *proc_handler,
6704 entry->procname = procname;
6706 entry->maxlen = maxlen;
6708 entry->proc_handler = proc_handler;
6711 entry->extra1 = &min_load_idx;
6712 entry->extra2 = &max_load_idx;
6716 static struct ctl_table *
6717 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6719 struct ctl_table *table = sd_alloc_ctl_entry(13);
6724 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6725 sizeof(long), 0644, proc_doulongvec_minmax, false);
6726 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6727 sizeof(long), 0644, proc_doulongvec_minmax, false);
6728 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6729 sizeof(int), 0644, proc_dointvec_minmax, true);
6730 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6731 sizeof(int), 0644, proc_dointvec_minmax, true);
6732 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6733 sizeof(int), 0644, proc_dointvec_minmax, true);
6734 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6735 sizeof(int), 0644, proc_dointvec_minmax, true);
6736 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6737 sizeof(int), 0644, proc_dointvec_minmax, true);
6738 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6739 sizeof(int), 0644, proc_dointvec_minmax, false);
6740 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6741 sizeof(int), 0644, proc_dointvec_minmax, false);
6742 set_table_entry(&table[9], "cache_nice_tries",
6743 &sd->cache_nice_tries,
6744 sizeof(int), 0644, proc_dointvec_minmax, false);
6745 set_table_entry(&table[10], "flags", &sd->flags,
6746 sizeof(int), 0644, proc_dointvec_minmax, false);
6747 set_table_entry(&table[11], "name", sd->name,
6748 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
6749 /* &table[12] is terminator */
6754 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6756 struct ctl_table *entry, *table;
6757 struct sched_domain *sd;
6758 int domain_num = 0, i;
6761 for_each_domain(cpu, sd)
6763 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6768 for_each_domain(cpu, sd) {
6769 snprintf(buf, 32, "domain%d", i);
6770 entry->procname = kstrdup(buf, GFP_KERNEL);
6772 entry->child = sd_alloc_ctl_domain_table(sd);
6779 static struct ctl_table_header *sd_sysctl_header;
6780 static void register_sched_domain_sysctl(void)
6782 int i, cpu_num = num_possible_cpus();
6783 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6786 WARN_ON(sd_ctl_dir[0].child);
6787 sd_ctl_dir[0].child = entry;
6792 for_each_possible_cpu(i) {
6793 snprintf(buf, 32, "cpu%d", i);
6794 entry->procname = kstrdup(buf, GFP_KERNEL);
6796 entry->child = sd_alloc_ctl_cpu_table(i);
6800 WARN_ON(sd_sysctl_header);
6801 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6804 /* may be called multiple times per register */
6805 static void unregister_sched_domain_sysctl(void)
6807 if (sd_sysctl_header)
6808 unregister_sysctl_table(sd_sysctl_header);
6809 sd_sysctl_header = NULL;
6810 if (sd_ctl_dir[0].child)
6811 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6814 static void register_sched_domain_sysctl(void)
6817 static void unregister_sched_domain_sysctl(void)
6822 static void set_rq_online(struct rq *rq)
6825 const struct sched_class *class;
6827 cpumask_set_cpu(rq->cpu, rq->rd->online);
6830 for_each_class(class) {
6831 if (class->rq_online)
6832 class->rq_online(rq);
6837 static void set_rq_offline(struct rq *rq)
6840 const struct sched_class *class;
6842 for_each_class(class) {
6843 if (class->rq_offline)
6844 class->rq_offline(rq);
6847 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6853 * migration_call - callback that gets triggered when a CPU is added.
6854 * Here we can start up the necessary migration thread for the new CPU.
6856 static int __cpuinit
6857 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6859 int cpu = (long)hcpu;
6860 unsigned long flags;
6861 struct rq *rq = cpu_rq(cpu);
6863 switch (action & ~CPU_TASKS_FROZEN) {
6865 case CPU_UP_PREPARE:
6866 rq->calc_load_update = calc_load_update;
6867 account_reset_rq(rq);
6871 /* Update our root-domain */
6872 raw_spin_lock_irqsave(&rq->lock, flags);
6874 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6878 raw_spin_unlock_irqrestore(&rq->lock, flags);
6881 #ifdef CONFIG_HOTPLUG_CPU
6883 sched_ttwu_pending();
6884 /* Update our root-domain */
6885 raw_spin_lock_irqsave(&rq->lock, flags);
6887 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6891 BUG_ON(rq->nr_running != 1); /* the migration thread */
6892 raw_spin_unlock_irqrestore(&rq->lock, flags);
6894 migrate_nr_uninterruptible(rq);
6895 calc_global_load_remove(rq);
6900 update_max_interval();
6906 * Register at high priority so that task migration (migrate_all_tasks)
6907 * happens before everything else. This has to be lower priority than
6908 * the notifier in the perf_event subsystem, though.
6910 static struct notifier_block __cpuinitdata migration_notifier = {
6911 .notifier_call = migration_call,
6912 .priority = CPU_PRI_MIGRATION,
6915 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6916 unsigned long action, void *hcpu)
6918 switch (action & ~CPU_TASKS_FROZEN) {
6920 case CPU_DOWN_FAILED:
6921 set_cpu_active((long)hcpu, true);
6928 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6929 unsigned long action, void *hcpu)
6931 switch (action & ~CPU_TASKS_FROZEN) {
6932 case CPU_DOWN_PREPARE:
6933 set_cpu_active((long)hcpu, false);
6940 static int __init migration_init(void)
6942 void *cpu = (void *)(long)smp_processor_id();
6945 /* Initialize migration for the boot CPU */
6946 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6947 BUG_ON(err == NOTIFY_BAD);
6948 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6949 register_cpu_notifier(&migration_notifier);
6951 /* Register cpu active notifiers */
6952 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6953 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6957 early_initcall(migration_init);
6962 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6964 #ifdef CONFIG_SCHED_DEBUG
6966 static __read_mostly int sched_domain_debug_enabled;
6968 static int __init sched_domain_debug_setup(char *str)
6970 sched_domain_debug_enabled = 1;
6974 early_param("sched_debug", sched_domain_debug_setup);
6976 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6977 struct cpumask *groupmask)
6979 struct sched_group *group = sd->groups;
6982 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6983 cpumask_clear(groupmask);
6985 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6987 if (!(sd->flags & SD_LOAD_BALANCE)) {
6988 printk("does not load-balance\n");
6990 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6995 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6997 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6998 printk(KERN_ERR "ERROR: domain->span does not contain "
7001 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7002 printk(KERN_ERR "ERROR: domain->groups does not contain"
7006 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7010 printk(KERN_ERR "ERROR: group is NULL\n");
7014 if (!group->sgp->power) {
7015 printk(KERN_CONT "\n");
7016 printk(KERN_ERR "ERROR: domain->cpu_power not "
7021 if (!cpumask_weight(sched_group_cpus(group))) {
7022 printk(KERN_CONT "\n");
7023 printk(KERN_ERR "ERROR: empty group\n");
7027 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7028 printk(KERN_CONT "\n");
7029 printk(KERN_ERR "ERROR: repeated CPUs\n");
7033 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7035 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7037 printk(KERN_CONT " %s", str);
7038 if (group->sgp->power != SCHED_POWER_SCALE) {
7039 printk(KERN_CONT " (cpu_power = %d)",
7043 group = group->next;
7044 } while (group != sd->groups);
7045 printk(KERN_CONT "\n");
7047 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7048 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7051 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7052 printk(KERN_ERR "ERROR: parent span is not a superset "
7053 "of domain->span\n");
7057 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7061 if (!sched_domain_debug_enabled)
7065 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7069 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7072 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
7080 #else /* !CONFIG_SCHED_DEBUG */
7081 # define sched_domain_debug(sd, cpu) do { } while (0)
7082 #endif /* CONFIG_SCHED_DEBUG */
7084 static int sd_degenerate(struct sched_domain *sd)
7086 if (cpumask_weight(sched_domain_span(sd)) == 1)
7089 /* Following flags need at least 2 groups */
7090 if (sd->flags & (SD_LOAD_BALANCE |
7091 SD_BALANCE_NEWIDLE |
7095 SD_SHARE_PKG_RESOURCES)) {
7096 if (sd->groups != sd->groups->next)
7100 /* Following flags don't use groups */
7101 if (sd->flags & (SD_WAKE_AFFINE))
7108 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7110 unsigned long cflags = sd->flags, pflags = parent->flags;
7112 if (sd_degenerate(parent))
7115 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7118 /* Flags needing groups don't count if only 1 group in parent */
7119 if (parent->groups == parent->groups->next) {
7120 pflags &= ~(SD_LOAD_BALANCE |
7121 SD_BALANCE_NEWIDLE |
7125 SD_SHARE_PKG_RESOURCES);
7126 if (nr_node_ids == 1)
7127 pflags &= ~SD_SERIALIZE;
7129 if (~cflags & pflags)
7135 static void free_rootdomain(struct rcu_head *rcu)
7137 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
7139 cpupri_cleanup(&rd->cpupri);
7140 free_cpumask_var(rd->rto_mask);
7141 free_cpumask_var(rd->online);
7142 free_cpumask_var(rd->span);
7146 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7148 struct root_domain *old_rd = NULL;
7149 unsigned long flags;
7151 raw_spin_lock_irqsave(&rq->lock, flags);
7156 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7159 cpumask_clear_cpu(rq->cpu, old_rd->span);
7162 * If we dont want to free the old_rt yet then
7163 * set old_rd to NULL to skip the freeing later
7166 if (!atomic_dec_and_test(&old_rd->refcount))
7170 atomic_inc(&rd->refcount);
7173 cpumask_set_cpu(rq->cpu, rd->span);
7174 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7177 raw_spin_unlock_irqrestore(&rq->lock, flags);
7180 call_rcu_sched(&old_rd->rcu, free_rootdomain);
7183 static int init_rootdomain(struct root_domain *rd)
7185 memset(rd, 0, sizeof(*rd));
7187 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
7189 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
7191 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7194 if (cpupri_init(&rd->cpupri) != 0)
7199 free_cpumask_var(rd->rto_mask);
7201 free_cpumask_var(rd->online);
7203 free_cpumask_var(rd->span);
7208 static void init_defrootdomain(void)
7210 init_rootdomain(&def_root_domain);
7212 atomic_set(&def_root_domain.refcount, 1);
7215 static struct root_domain *alloc_rootdomain(void)
7217 struct root_domain *rd;
7219 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7223 if (init_rootdomain(rd) != 0) {
7231 static void free_sched_groups(struct sched_group *sg, int free_sgp)
7233 struct sched_group *tmp, *first;
7242 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7247 } while (sg != first);
7250 static void free_sched_domain(struct rcu_head *rcu)
7252 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7255 * If its an overlapping domain it has private groups, iterate and
7258 if (sd->flags & SD_OVERLAP) {
7259 free_sched_groups(sd->groups, 1);
7260 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7261 kfree(sd->groups->sgp);
7267 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7269 call_rcu(&sd->rcu, free_sched_domain);
7272 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7274 for (; sd; sd = sd->parent)
7275 destroy_sched_domain(sd, cpu);
7279 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7280 * hold the hotplug lock.
7283 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7285 struct rq *rq = cpu_rq(cpu);
7286 struct sched_domain *tmp;
7288 /* Remove the sched domains which do not contribute to scheduling. */
7289 for (tmp = sd; tmp; ) {
7290 struct sched_domain *parent = tmp->parent;
7294 if (sd_parent_degenerate(tmp, parent)) {
7295 tmp->parent = parent->parent;
7297 parent->parent->child = tmp;
7298 destroy_sched_domain(parent, cpu);
7303 if (sd && sd_degenerate(sd)) {
7306 destroy_sched_domain(tmp, cpu);
7311 sched_domain_debug(sd, cpu);
7313 rq_attach_root(rq, rd);
7315 rcu_assign_pointer(rq->sd, sd);
7316 destroy_sched_domains(tmp, cpu);
7319 /* cpus with isolated domains */
7320 static cpumask_var_t cpu_isolated_map;
7322 /* Setup the mask of cpus configured for isolated domains */
7323 static int __init isolated_cpu_setup(char *str)
7325 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7326 cpulist_parse(str, cpu_isolated_map);
7330 __setup("isolcpus=", isolated_cpu_setup);
7335 * find_next_best_node - find the next node to include in a sched_domain
7336 * @node: node whose sched_domain we're building
7337 * @used_nodes: nodes already in the sched_domain
7339 * Find the next node to include in a given scheduling domain. Simply
7340 * finds the closest node not already in the @used_nodes map.
7342 * Should use nodemask_t.
7344 static int find_next_best_node(int node, nodemask_t *used_nodes)
7346 int i, n, val, min_val, best_node = -1;
7350 for (i = 0; i < nr_node_ids; i++) {
7351 /* Start at @node */
7352 n = (node + i) % nr_node_ids;
7354 if (!nr_cpus_node(n))
7357 /* Skip already used nodes */
7358 if (node_isset(n, *used_nodes))
7361 /* Simple min distance search */
7362 val = node_distance(node, n);
7364 if (val < min_val) {
7370 if (best_node != -1)
7371 node_set(best_node, *used_nodes);
7376 * sched_domain_node_span - get a cpumask for a node's sched_domain
7377 * @node: node whose cpumask we're constructing
7378 * @span: resulting cpumask
7380 * Given a node, construct a good cpumask for its sched_domain to span. It
7381 * should be one that prevents unnecessary balancing, but also spreads tasks
7384 static void sched_domain_node_span(int node, struct cpumask *span)
7386 nodemask_t used_nodes;
7389 cpumask_clear(span);
7390 nodes_clear(used_nodes);
7392 cpumask_or(span, span, cpumask_of_node(node));
7393 node_set(node, used_nodes);
7395 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7396 int next_node = find_next_best_node(node, &used_nodes);
7399 cpumask_or(span, span, cpumask_of_node(next_node));
7403 static const struct cpumask *cpu_node_mask(int cpu)
7405 lockdep_assert_held(&sched_domains_mutex);
7407 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7409 return sched_domains_tmpmask;
7412 static const struct cpumask *cpu_allnodes_mask(int cpu)
7414 return cpu_possible_mask;
7416 #endif /* CONFIG_NUMA */
7418 static const struct cpumask *cpu_cpu_mask(int cpu)
7420 return cpumask_of_node(cpu_to_node(cpu));
7423 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7426 struct sched_domain **__percpu sd;
7427 struct sched_group **__percpu sg;
7428 struct sched_group_power **__percpu sgp;
7432 struct sched_domain ** __percpu sd;
7433 struct root_domain *rd;
7443 struct sched_domain_topology_level;
7445 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7446 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7448 #define SDTL_OVERLAP 0x01
7450 struct sched_domain_topology_level {
7451 sched_domain_init_f init;
7452 sched_domain_mask_f mask;
7454 struct sd_data data;
7458 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7460 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7461 const struct cpumask *span = sched_domain_span(sd);
7462 struct cpumask *covered = sched_domains_tmpmask;
7463 struct sd_data *sdd = sd->private;
7464 struct sched_domain *child;
7467 cpumask_clear(covered);
7469 for_each_cpu(i, span) {
7470 struct cpumask *sg_span;
7472 if (cpumask_test_cpu(i, covered))
7475 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7476 GFP_KERNEL, cpu_to_node(i));
7481 sg_span = sched_group_cpus(sg);
7483 child = *per_cpu_ptr(sdd->sd, i);
7485 child = child->child;
7486 cpumask_copy(sg_span, sched_domain_span(child));
7488 cpumask_set_cpu(i, sg_span);
7490 cpumask_or(covered, covered, sg_span);
7492 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7493 atomic_inc(&sg->sgp->ref);
7495 if (cpumask_test_cpu(cpu, sg_span))
7505 sd->groups = groups;
7510 free_sched_groups(first, 0);
7515 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7517 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7518 struct sched_domain *child = sd->child;
7521 cpu = cpumask_first(sched_domain_span(child));
7524 *sg = *per_cpu_ptr(sdd->sg, cpu);
7525 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7526 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7533 * build_sched_groups will build a circular linked list of the groups
7534 * covered by the given span, and will set each group's ->cpumask correctly,
7535 * and ->cpu_power to 0.
7537 * Assumes the sched_domain tree is fully constructed
7540 build_sched_groups(struct sched_domain *sd, int cpu)
7542 struct sched_group *first = NULL, *last = NULL;
7543 struct sd_data *sdd = sd->private;
7544 const struct cpumask *span = sched_domain_span(sd);
7545 struct cpumask *covered;
7548 get_group(cpu, sdd, &sd->groups);
7549 atomic_inc(&sd->groups->ref);
7551 if (cpu != cpumask_first(sched_domain_span(sd)))
7554 lockdep_assert_held(&sched_domains_mutex);
7555 covered = sched_domains_tmpmask;
7557 cpumask_clear(covered);
7559 for_each_cpu(i, span) {
7560 struct sched_group *sg;
7561 int group = get_group(i, sdd, &sg);
7564 if (cpumask_test_cpu(i, covered))
7567 cpumask_clear(sched_group_cpus(sg));
7570 for_each_cpu(j, span) {
7571 if (get_group(j, sdd, NULL) != group)
7574 cpumask_set_cpu(j, covered);
7575 cpumask_set_cpu(j, sched_group_cpus(sg));
7590 * Initialize sched groups cpu_power.
7592 * cpu_power indicates the capacity of sched group, which is used while
7593 * distributing the load between different sched groups in a sched domain.
7594 * Typically cpu_power for all the groups in a sched domain will be same unless
7595 * there are asymmetries in the topology. If there are asymmetries, group
7596 * having more cpu_power will pickup more load compared to the group having
7599 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7601 struct sched_group *sg = sd->groups;
7603 WARN_ON(!sd || !sg);
7606 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7608 } while (sg != sd->groups);
7610 if (cpu != group_first_cpu(sg))
7613 update_group_power(sd, cpu);
7617 * Initializers for schedule domains
7618 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7621 #ifdef CONFIG_SCHED_DEBUG
7622 # define SD_INIT_NAME(sd, type) sd->name = #type
7624 # define SD_INIT_NAME(sd, type) do { } while (0)
7627 #define SD_INIT_FUNC(type) \
7628 static noinline struct sched_domain * \
7629 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7631 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7632 *sd = SD_##type##_INIT; \
7633 SD_INIT_NAME(sd, type); \
7634 sd->private = &tl->data; \
7640 SD_INIT_FUNC(ALLNODES)
7643 #ifdef CONFIG_SCHED_SMT
7644 SD_INIT_FUNC(SIBLING)
7646 #ifdef CONFIG_SCHED_MC
7649 #ifdef CONFIG_SCHED_BOOK
7653 static int default_relax_domain_level = -1;
7654 int sched_domain_level_max;
7656 static int __init setup_relax_domain_level(char *str)
7658 if (kstrtoint(str, 0, &default_relax_domain_level))
7659 pr_warn("Unable to set relax_domain_level\n");
7663 __setup("relax_domain_level=", setup_relax_domain_level);
7665 static void set_domain_attribute(struct sched_domain *sd,
7666 struct sched_domain_attr *attr)
7670 if (!attr || attr->relax_domain_level < 0) {
7671 if (default_relax_domain_level < 0)
7674 request = default_relax_domain_level;
7676 request = attr->relax_domain_level;
7677 if (request < sd->level) {
7678 /* turn off idle balance on this domain */
7679 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7681 /* turn on idle balance on this domain */
7682 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7686 static void __sdt_free(const struct cpumask *cpu_map);
7687 static int __sdt_alloc(const struct cpumask *cpu_map);
7689 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7690 const struct cpumask *cpu_map)
7694 if (!atomic_read(&d->rd->refcount))
7695 free_rootdomain(&d->rd->rcu); /* fall through */
7697 free_percpu(d->sd); /* fall through */
7699 __sdt_free(cpu_map); /* fall through */
7705 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7706 const struct cpumask *cpu_map)
7708 memset(d, 0, sizeof(*d));
7710 if (__sdt_alloc(cpu_map))
7711 return sa_sd_storage;
7712 d->sd = alloc_percpu(struct sched_domain *);
7714 return sa_sd_storage;
7715 d->rd = alloc_rootdomain();
7718 return sa_rootdomain;
7722 * NULL the sd_data elements we've used to build the sched_domain and
7723 * sched_group structure so that the subsequent __free_domain_allocs()
7724 * will not free the data we're using.
7726 static void claim_allocations(int cpu, struct sched_domain *sd)
7728 struct sd_data *sdd = sd->private;
7730 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7731 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7733 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7734 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7736 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7737 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7740 #ifdef CONFIG_SCHED_SMT
7741 static const struct cpumask *cpu_smt_mask(int cpu)
7743 return topology_thread_cpumask(cpu);
7748 * Topology list, bottom-up.
7750 static struct sched_domain_topology_level default_topology[] = {
7751 #ifdef CONFIG_SCHED_SMT
7752 { sd_init_SIBLING, cpu_smt_mask, },
7754 #ifdef CONFIG_SCHED_MC
7755 { sd_init_MC, cpu_coregroup_mask, },
7757 #ifdef CONFIG_SCHED_BOOK
7758 { sd_init_BOOK, cpu_book_mask, },
7760 { sd_init_CPU, cpu_cpu_mask, },
7762 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7763 { sd_init_ALLNODES, cpu_allnodes_mask, },
7768 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7770 static int __sdt_alloc(const struct cpumask *cpu_map)
7772 struct sched_domain_topology_level *tl;
7775 for (tl = sched_domain_topology; tl->init; tl++) {
7776 struct sd_data *sdd = &tl->data;
7778 sdd->sd = alloc_percpu(struct sched_domain *);
7782 sdd->sg = alloc_percpu(struct sched_group *);
7786 sdd->sgp = alloc_percpu(struct sched_group_power *);
7790 for_each_cpu(j, cpu_map) {
7791 struct sched_domain *sd;
7792 struct sched_group *sg;
7793 struct sched_group_power *sgp;
7795 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7796 GFP_KERNEL, cpu_to_node(j));
7800 *per_cpu_ptr(sdd->sd, j) = sd;
7802 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7803 GFP_KERNEL, cpu_to_node(j));
7807 *per_cpu_ptr(sdd->sg, j) = sg;
7809 sgp = kzalloc_node(sizeof(struct sched_group_power),
7810 GFP_KERNEL, cpu_to_node(j));
7814 *per_cpu_ptr(sdd->sgp, j) = sgp;
7821 static void __sdt_free(const struct cpumask *cpu_map)
7823 struct sched_domain_topology_level *tl;
7826 for (tl = sched_domain_topology; tl->init; tl++) {
7827 struct sd_data *sdd = &tl->data;
7829 for_each_cpu(j, cpu_map) {
7830 struct sched_domain *sd;
7833 sd = *per_cpu_ptr(sdd->sd, j);
7834 if (sd && (sd->flags & SD_OVERLAP))
7835 free_sched_groups(sd->groups, 0);
7836 kfree(*per_cpu_ptr(sdd->sd, j));
7840 kfree(*per_cpu_ptr(sdd->sg, j));
7842 kfree(*per_cpu_ptr(sdd->sgp, j));
7844 free_percpu(sdd->sd);
7846 free_percpu(sdd->sg);
7848 free_percpu(sdd->sgp);
7853 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7854 struct s_data *d, const struct cpumask *cpu_map,
7855 struct sched_domain_attr *attr, struct sched_domain *child,
7858 struct sched_domain *sd = tl->init(tl, cpu);
7862 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7864 sd->level = child->level + 1;
7865 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7869 set_domain_attribute(sd, attr);
7875 * Build sched domains for a given set of cpus and attach the sched domains
7876 * to the individual cpus
7878 static int build_sched_domains(const struct cpumask *cpu_map,
7879 struct sched_domain_attr *attr)
7881 enum s_alloc alloc_state = sa_none;
7882 struct sched_domain *sd;
7884 int i, ret = -ENOMEM;
7886 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7887 if (alloc_state != sa_rootdomain)
7890 /* Set up domains for cpus specified by the cpu_map. */
7891 for_each_cpu(i, cpu_map) {
7892 struct sched_domain_topology_level *tl;
7895 for (tl = sched_domain_topology; tl->init; tl++) {
7896 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7897 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7898 sd->flags |= SD_OVERLAP;
7899 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7906 *per_cpu_ptr(d.sd, i) = sd;
7909 /* Build the groups for the domains */
7910 for_each_cpu(i, cpu_map) {
7911 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7912 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7913 if (sd->flags & SD_OVERLAP) {
7914 if (build_overlap_sched_groups(sd, i))
7917 if (build_sched_groups(sd, i))
7923 /* Calculate CPU power for physical packages and nodes */
7924 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7925 if (!cpumask_test_cpu(i, cpu_map))
7928 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7929 claim_allocations(i, sd);
7930 init_sched_groups_power(i, sd);
7934 /* Attach the domains */
7936 for_each_cpu(i, cpu_map) {
7937 sd = *per_cpu_ptr(d.sd, i);
7938 cpu_attach_domain(sd, d.rd, i);
7944 __free_domain_allocs(&d, alloc_state, cpu_map);
7948 static cpumask_var_t *doms_cur; /* current sched domains */
7949 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7950 static struct sched_domain_attr *dattr_cur;
7951 /* attribues of custom domains in 'doms_cur' */
7954 * Special case: If a kmalloc of a doms_cur partition (array of
7955 * cpumask) fails, then fallback to a single sched domain,
7956 * as determined by the single cpumask fallback_doms.
7958 static cpumask_var_t fallback_doms;
7961 * arch_update_cpu_topology lets virtualized architectures update the
7962 * cpu core maps. It is supposed to return 1 if the topology changed
7963 * or 0 if it stayed the same.
7965 int __attribute__((weak)) arch_update_cpu_topology(void)
7970 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7973 cpumask_var_t *doms;
7975 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7978 for (i = 0; i < ndoms; i++) {
7979 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7980 free_sched_domains(doms, i);
7987 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7990 for (i = 0; i < ndoms; i++)
7991 free_cpumask_var(doms[i]);
7996 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7997 * For now this just excludes isolated cpus, but could be used to
7998 * exclude other special cases in the future.
8000 static int init_sched_domains(const struct cpumask *cpu_map)
8004 arch_update_cpu_topology();
8006 doms_cur = alloc_sched_domains(ndoms_cur);
8008 doms_cur = &fallback_doms;
8009 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8011 err = build_sched_domains(doms_cur[0], NULL);
8012 register_sched_domain_sysctl();
8018 * Detach sched domains from a group of cpus specified in cpu_map
8019 * These cpus will now be attached to the NULL domain
8021 static void detach_destroy_domains(const struct cpumask *cpu_map)
8026 for_each_cpu(i, cpu_map)
8027 cpu_attach_domain(NULL, &def_root_domain, i);
8031 /* handle null as "default" */
8032 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8033 struct sched_domain_attr *new, int idx_new)
8035 struct sched_domain_attr tmp;
8042 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8043 new ? (new + idx_new) : &tmp,
8044 sizeof(struct sched_domain_attr));
8048 * Partition sched domains as specified by the 'ndoms_new'
8049 * cpumasks in the array doms_new[] of cpumasks. This compares
8050 * doms_new[] to the current sched domain partitioning, doms_cur[].
8051 * It destroys each deleted domain and builds each new domain.
8053 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
8054 * The masks don't intersect (don't overlap.) We should setup one
8055 * sched domain for each mask. CPUs not in any of the cpumasks will
8056 * not be load balanced. If the same cpumask appears both in the
8057 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8060 * The passed in 'doms_new' should be allocated using
8061 * alloc_sched_domains. This routine takes ownership of it and will
8062 * free_sched_domains it when done with it. If the caller failed the
8063 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
8064 * and partition_sched_domains() will fallback to the single partition
8065 * 'fallback_doms', it also forces the domains to be rebuilt.
8067 * If doms_new == NULL it will be replaced with cpu_online_mask.
8068 * ndoms_new == 0 is a special case for destroying existing domains,
8069 * and it will not create the default domain.
8071 * Call with hotplug lock held
8073 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
8074 struct sched_domain_attr *dattr_new)
8079 mutex_lock(&sched_domains_mutex);
8081 /* always unregister in case we don't destroy any domains */
8082 unregister_sched_domain_sysctl();
8084 /* Let architecture update cpu core mappings. */
8085 new_topology = arch_update_cpu_topology();
8087 n = doms_new ? ndoms_new : 0;
8089 /* Destroy deleted domains */
8090 for (i = 0; i < ndoms_cur; i++) {
8091 for (j = 0; j < n && !new_topology; j++) {
8092 if (cpumask_equal(doms_cur[i], doms_new[j])
8093 && dattrs_equal(dattr_cur, i, dattr_new, j))
8096 /* no match - a current sched domain not in new doms_new[] */
8097 detach_destroy_domains(doms_cur[i]);
8102 if (doms_new == NULL) {
8104 doms_new = &fallback_doms;
8105 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
8106 WARN_ON_ONCE(dattr_new);
8109 /* Build new domains */
8110 for (i = 0; i < ndoms_new; i++) {
8111 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8112 if (cpumask_equal(doms_new[i], doms_cur[j])
8113 && dattrs_equal(dattr_new, i, dattr_cur, j))
8116 /* no match - add a new doms_new */
8117 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
8122 /* Remember the new sched domains */
8123 if (doms_cur != &fallback_doms)
8124 free_sched_domains(doms_cur, ndoms_cur);
8125 kfree(dattr_cur); /* kfree(NULL) is safe */
8126 doms_cur = doms_new;
8127 dattr_cur = dattr_new;
8128 ndoms_cur = ndoms_new;
8130 register_sched_domain_sysctl();
8132 mutex_unlock(&sched_domains_mutex);
8135 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8136 static void reinit_sched_domains(void)
8140 /* Destroy domains first to force the rebuild */
8141 partition_sched_domains(0, NULL, NULL);
8143 rebuild_sched_domains();
8147 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8149 unsigned int level = 0;
8151 if (sscanf(buf, "%u", &level) != 1)
8155 * level is always be positive so don't check for
8156 * level < POWERSAVINGS_BALANCE_NONE which is 0
8157 * What happens on 0 or 1 byte write,
8158 * need to check for count as well?
8161 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8165 sched_smt_power_savings = level;
8167 sched_mc_power_savings = level;
8169 reinit_sched_domains();
8174 #ifdef CONFIG_SCHED_MC
8175 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8176 struct sysdev_class_attribute *attr,
8179 return sprintf(page, "%u\n", sched_mc_power_savings);
8181 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8182 struct sysdev_class_attribute *attr,
8183 const char *buf, size_t count)
8185 return sched_power_savings_store(buf, count, 0);
8187 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8188 sched_mc_power_savings_show,
8189 sched_mc_power_savings_store);
8192 #ifdef CONFIG_SCHED_SMT
8193 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8194 struct sysdev_class_attribute *attr,
8197 return sprintf(page, "%u\n", sched_smt_power_savings);
8199 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8200 struct sysdev_class_attribute *attr,
8201 const char *buf, size_t count)
8203 return sched_power_savings_store(buf, count, 1);
8205 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8206 sched_smt_power_savings_show,
8207 sched_smt_power_savings_store);
8210 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8214 #ifdef CONFIG_SCHED_SMT
8216 err = sysfs_create_file(&cls->kset.kobj,
8217 &attr_sched_smt_power_savings.attr);
8219 #ifdef CONFIG_SCHED_MC
8220 if (!err && mc_capable())
8221 err = sysfs_create_file(&cls->kset.kobj,
8222 &attr_sched_mc_power_savings.attr);
8226 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8228 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
8231 * Update cpusets according to cpu_active mask. If cpusets are
8232 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
8233 * around partition_sched_domains().
8235 * If we come here as part of a suspend/resume, don't touch cpusets because we
8236 * want to restore it back to its original state upon resume anyway.
8238 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
8242 case CPU_ONLINE_FROZEN:
8243 case CPU_DOWN_FAILED_FROZEN:
8246 * num_cpus_frozen tracks how many CPUs are involved in suspend
8247 * resume sequence. As long as this is not the last online
8248 * operation in the resume sequence, just build a single sched
8249 * domain, ignoring cpusets.
8252 if (likely(num_cpus_frozen)) {
8253 partition_sched_domains(1, NULL, NULL);
8258 * This is the last CPU online operation. So fall through and
8259 * restore the original sched domains by considering the
8260 * cpuset configurations.
8264 case CPU_DOWN_FAILED:
8265 cpuset_update_active_cpus();
8273 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8277 case CPU_DOWN_PREPARE:
8278 cpuset_update_active_cpus();
8280 case CPU_DOWN_PREPARE_FROZEN:
8282 partition_sched_domains(1, NULL, NULL);
8290 static int update_runtime(struct notifier_block *nfb,
8291 unsigned long action, void *hcpu)
8293 int cpu = (int)(long)hcpu;
8296 case CPU_DOWN_PREPARE:
8297 case CPU_DOWN_PREPARE_FROZEN:
8298 disable_runtime(cpu_rq(cpu));
8301 case CPU_DOWN_FAILED:
8302 case CPU_DOWN_FAILED_FROZEN:
8304 case CPU_ONLINE_FROZEN:
8305 enable_runtime(cpu_rq(cpu));
8313 void __init sched_init_smp(void)
8315 cpumask_var_t non_isolated_cpus;
8317 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8318 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8321 mutex_lock(&sched_domains_mutex);
8322 init_sched_domains(cpu_active_mask);
8323 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8324 if (cpumask_empty(non_isolated_cpus))
8325 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8326 mutex_unlock(&sched_domains_mutex);
8329 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8330 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8332 /* RT runtime code needs to handle some hotplug events */
8333 hotcpu_notifier(update_runtime, 0);
8337 /* Move init over to a non-isolated CPU */
8338 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8340 sched_init_granularity();
8341 free_cpumask_var(non_isolated_cpus);
8343 init_sched_rt_class();
8346 void __init sched_init_smp(void)
8348 sched_init_granularity();
8350 #endif /* CONFIG_SMP */
8352 const_debug unsigned int sysctl_timer_migration = 1;
8354 int in_sched_functions(unsigned long addr)
8356 return in_lock_functions(addr) ||
8357 (addr >= (unsigned long)__sched_text_start
8358 && addr < (unsigned long)__sched_text_end);
8361 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8363 cfs_rq->tasks_timeline = RB_ROOT;
8364 INIT_LIST_HEAD(&cfs_rq->tasks);
8365 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8366 #ifndef CONFIG_64BIT
8367 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8371 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8373 struct rt_prio_array *array;
8376 array = &rt_rq->active;
8377 for (i = 0; i < MAX_RT_PRIO; i++) {
8378 INIT_LIST_HEAD(array->queue + i);
8379 __clear_bit(i, array->bitmap);
8381 /* delimiter for bitsearch: */
8382 __set_bit(MAX_RT_PRIO, array->bitmap);
8384 #if defined CONFIG_SMP
8385 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8386 rt_rq->highest_prio.next = MAX_RT_PRIO;
8387 rt_rq->rt_nr_migratory = 0;
8388 rt_rq->overloaded = 0;
8389 plist_head_init(&rt_rq->pushable_tasks);
8393 rt_rq->rt_throttled = 0;
8394 rt_rq->rt_runtime = 0;
8395 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8398 #ifdef CONFIG_FAIR_GROUP_SCHED
8399 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8400 struct sched_entity *se, int cpu,
8401 struct sched_entity *parent)
8403 struct rq *rq = cpu_rq(cpu);
8408 /* allow initial update_cfs_load() to truncate */
8409 cfs_rq->load_stamp = 1;
8411 init_cfs_rq_runtime(cfs_rq);
8413 tg->cfs_rq[cpu] = cfs_rq;
8416 /* se could be NULL for root_task_group */
8421 se->cfs_rq = &rq->cfs;
8423 se->cfs_rq = parent->my_q;
8426 update_load_set(&se->load, 0);
8427 se->parent = parent;
8431 #ifdef CONFIG_RT_GROUP_SCHED
8432 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8433 struct sched_rt_entity *rt_se, int cpu,
8434 struct sched_rt_entity *parent)
8436 struct rq *rq = cpu_rq(cpu);
8438 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8439 rt_rq->rt_nr_boosted = 0;
8443 tg->rt_rq[cpu] = rt_rq;
8444 tg->rt_se[cpu] = rt_se;
8450 rt_se->rt_rq = &rq->rt;
8452 rt_se->rt_rq = parent->my_q;
8454 rt_se->my_q = rt_rq;
8455 rt_se->parent = parent;
8456 INIT_LIST_HEAD(&rt_se->run_list);
8460 void __init sched_init(void)
8463 unsigned long alloc_size = 0, ptr;
8465 #ifdef CONFIG_FAIR_GROUP_SCHED
8466 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8468 #ifdef CONFIG_RT_GROUP_SCHED
8469 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8471 #ifdef CONFIG_CPUMASK_OFFSTACK
8472 alloc_size += num_possible_cpus() * cpumask_size();
8475 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8477 #ifdef CONFIG_FAIR_GROUP_SCHED
8478 root_task_group.se = (struct sched_entity **)ptr;
8479 ptr += nr_cpu_ids * sizeof(void **);
8481 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8482 ptr += nr_cpu_ids * sizeof(void **);
8484 #endif /* CONFIG_FAIR_GROUP_SCHED */
8485 #ifdef CONFIG_RT_GROUP_SCHED
8486 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8487 ptr += nr_cpu_ids * sizeof(void **);
8489 root_task_group.rt_rq = (struct rt_rq **)ptr;
8490 ptr += nr_cpu_ids * sizeof(void **);
8492 #endif /* CONFIG_RT_GROUP_SCHED */
8493 #ifdef CONFIG_CPUMASK_OFFSTACK
8494 for_each_possible_cpu(i) {
8495 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8496 ptr += cpumask_size();
8498 #endif /* CONFIG_CPUMASK_OFFSTACK */
8502 init_defrootdomain();
8505 init_rt_bandwidth(&def_rt_bandwidth,
8506 global_rt_period(), global_rt_runtime());
8508 #ifdef CONFIG_RT_GROUP_SCHED
8509 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8510 global_rt_period(), global_rt_runtime());
8511 #endif /* CONFIG_RT_GROUP_SCHED */
8513 #ifdef CONFIG_CGROUP_SCHED
8514 list_add(&root_task_group.list, &task_groups);
8515 INIT_LIST_HEAD(&root_task_group.children);
8516 autogroup_init(&init_task);
8517 #endif /* CONFIG_CGROUP_SCHED */
8519 for_each_possible_cpu(i) {
8523 raw_spin_lock_init(&rq->lock);
8525 rq->calc_load_active = 0;
8526 rq->calc_load_update = jiffies + LOAD_FREQ;
8527 init_cfs_rq(&rq->cfs);
8528 init_rt_rq(&rq->rt, rq);
8529 #ifdef CONFIG_FAIR_GROUP_SCHED
8530 root_task_group.shares = root_task_group_load;
8531 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8533 * How much cpu bandwidth does root_task_group get?
8535 * In case of task-groups formed thr' the cgroup filesystem, it
8536 * gets 100% of the cpu resources in the system. This overall
8537 * system cpu resource is divided among the tasks of
8538 * root_task_group and its child task-groups in a fair manner,
8539 * based on each entity's (task or task-group's) weight
8540 * (se->load.weight).
8542 * In other words, if root_task_group has 10 tasks of weight
8543 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8544 * then A0's share of the cpu resource is:
8546 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8548 * We achieve this by letting root_task_group's tasks sit
8549 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8551 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8552 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8553 #endif /* CONFIG_FAIR_GROUP_SCHED */
8555 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8556 #ifdef CONFIG_RT_GROUP_SCHED
8557 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8558 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8561 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8562 rq->cpu_load[j] = 0;
8564 rq->last_load_update_tick = jiffies;
8569 rq->cpu_power = SCHED_POWER_SCALE;
8570 rq->post_schedule = 0;
8571 rq->active_balance = 0;
8572 rq->next_balance = jiffies;
8577 rq->avg_idle = 2*sysctl_sched_migration_cost;
8578 rq_attach_root(rq, &def_root_domain);
8580 rq->nohz_balance_kick = 0;
8584 atomic_set(&rq->nr_iowait, 0);
8587 set_load_weight(&init_task);
8589 #ifdef CONFIG_PREEMPT_NOTIFIERS
8590 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8594 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8597 #ifdef CONFIG_RT_MUTEXES
8598 plist_head_init(&init_task.pi_waiters);
8602 * The boot idle thread does lazy MMU switching as well:
8604 atomic_inc(&init_mm.mm_count);
8605 enter_lazy_tlb(&init_mm, current);
8608 * Make us the idle thread. Technically, schedule() should not be
8609 * called from this thread, however somewhere below it might be,
8610 * but because we are the idle thread, we just pick up running again
8611 * when this runqueue becomes "idle".
8613 init_idle(current, smp_processor_id());
8615 calc_load_update = jiffies + LOAD_FREQ;
8618 * During early bootup we pretend to be a normal task:
8620 current->sched_class = &fair_sched_class;
8623 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8625 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8626 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8627 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8628 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8629 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8631 /* May be allocated at isolcpus cmdline parse time */
8632 if (cpu_isolated_map == NULL)
8633 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8636 scheduler_running = 1;
8639 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8640 static inline int preempt_count_equals(int preempt_offset)
8642 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8644 return (nested == preempt_offset);
8647 void __might_sleep(const char *file, int line, int preempt_offset)
8649 static unsigned long prev_jiffy; /* ratelimiting */
8651 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8652 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8653 system_state != SYSTEM_RUNNING || oops_in_progress)
8655 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8657 prev_jiffy = jiffies;
8660 "BUG: sleeping function called from invalid context at %s:%d\n",
8663 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8664 in_atomic(), irqs_disabled(),
8665 current->pid, current->comm);
8667 debug_show_held_locks(current);
8668 if (irqs_disabled())
8669 print_irqtrace_events(current);
8672 EXPORT_SYMBOL(__might_sleep);
8675 #ifdef CONFIG_MAGIC_SYSRQ
8676 static void normalize_task(struct rq *rq, struct task_struct *p)
8678 const struct sched_class *prev_class = p->sched_class;
8679 int old_prio = p->prio;
8684 deactivate_task(rq, p, 0);
8685 __setscheduler(rq, p, SCHED_NORMAL, 0);
8687 activate_task(rq, p, 0);
8688 resched_task(rq->curr);
8691 check_class_changed(rq, p, prev_class, old_prio);
8694 void normalize_rt_tasks(void)
8696 struct task_struct *g, *p;
8697 unsigned long flags;
8700 read_lock_irqsave(&tasklist_lock, flags);
8701 do_each_thread(g, p) {
8703 * Only normalize user tasks:
8708 p->se.exec_start = 0;
8709 #ifdef CONFIG_SCHEDSTATS
8710 p->se.statistics.wait_start = 0;
8711 p->se.statistics.sleep_start = 0;
8712 p->se.statistics.block_start = 0;
8717 * Renice negative nice level userspace
8720 if (TASK_NICE(p) < 0 && p->mm)
8721 set_user_nice(p, 0);
8725 raw_spin_lock(&p->pi_lock);
8726 rq = __task_rq_lock(p);
8728 normalize_task(rq, p);
8730 __task_rq_unlock(rq);
8731 raw_spin_unlock(&p->pi_lock);
8732 } while_each_thread(g, p);
8734 read_unlock_irqrestore(&tasklist_lock, flags);
8737 #endif /* CONFIG_MAGIC_SYSRQ */
8739 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8741 * These functions are only useful for the IA64 MCA handling, or kdb.
8743 * They can only be called when the whole system has been
8744 * stopped - every CPU needs to be quiescent, and no scheduling
8745 * activity can take place. Using them for anything else would
8746 * be a serious bug, and as a result, they aren't even visible
8747 * under any other configuration.
8751 * curr_task - return the current task for a given cpu.
8752 * @cpu: the processor in question.
8754 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8756 struct task_struct *curr_task(int cpu)
8758 return cpu_curr(cpu);
8761 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8765 * set_curr_task - set the current task for a given cpu.
8766 * @cpu: the processor in question.
8767 * @p: the task pointer to set.
8769 * Description: This function must only be used when non-maskable interrupts
8770 * are serviced on a separate stack. It allows the architecture to switch the
8771 * notion of the current task on a cpu in a non-blocking manner. This function
8772 * must be called with all CPU's synchronized, and interrupts disabled, the
8773 * and caller must save the original value of the current task (see
8774 * curr_task() above) and restore that value before reenabling interrupts and
8775 * re-starting the system.
8777 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8779 void set_curr_task(int cpu, struct task_struct *p)
8786 #ifdef CONFIG_FAIR_GROUP_SCHED
8787 static void free_fair_sched_group(struct task_group *tg)
8791 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8793 for_each_possible_cpu(i) {
8795 kfree(tg->cfs_rq[i]);
8805 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8807 struct cfs_rq *cfs_rq;
8808 struct sched_entity *se;
8811 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8814 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8818 tg->shares = NICE_0_LOAD;
8820 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8822 for_each_possible_cpu(i) {
8823 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8824 GFP_KERNEL, cpu_to_node(i));
8828 se = kzalloc_node(sizeof(struct sched_entity),
8829 GFP_KERNEL, cpu_to_node(i));
8833 init_cfs_rq(cfs_rq);
8834 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8845 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8847 struct rq *rq = cpu_rq(cpu);
8848 unsigned long flags;
8851 * Only empty task groups can be destroyed; so we can speculatively
8852 * check on_list without danger of it being re-added.
8854 if (!tg->cfs_rq[cpu]->on_list)
8857 raw_spin_lock_irqsave(&rq->lock, flags);
8858 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8859 raw_spin_unlock_irqrestore(&rq->lock, flags);
8861 #else /* !CONFIG_FAIR_GROUP_SCHED */
8862 static inline void free_fair_sched_group(struct task_group *tg)
8867 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8872 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8875 #endif /* CONFIG_FAIR_GROUP_SCHED */
8877 #ifdef CONFIG_RT_GROUP_SCHED
8878 static void free_rt_sched_group(struct task_group *tg)
8883 destroy_rt_bandwidth(&tg->rt_bandwidth);
8885 for_each_possible_cpu(i) {
8887 kfree(tg->rt_rq[i]);
8889 kfree(tg->rt_se[i]);
8897 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8899 struct rt_rq *rt_rq;
8900 struct sched_rt_entity *rt_se;
8903 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8906 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8910 init_rt_bandwidth(&tg->rt_bandwidth,
8911 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8913 for_each_possible_cpu(i) {
8914 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8915 GFP_KERNEL, cpu_to_node(i));
8919 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8920 GFP_KERNEL, cpu_to_node(i));
8924 init_rt_rq(rt_rq, cpu_rq(i));
8925 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8926 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8936 #else /* !CONFIG_RT_GROUP_SCHED */
8937 static inline void free_rt_sched_group(struct task_group *tg)
8942 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8946 #endif /* CONFIG_RT_GROUP_SCHED */
8948 #ifdef CONFIG_CGROUP_SCHED
8949 static void free_sched_group(struct task_group *tg)
8951 free_fair_sched_group(tg);
8952 free_rt_sched_group(tg);
8957 /* allocate runqueue etc for a new task group */
8958 struct task_group *sched_create_group(struct task_group *parent)
8960 struct task_group *tg;
8961 unsigned long flags;
8963 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8965 return ERR_PTR(-ENOMEM);
8967 if (!alloc_fair_sched_group(tg, parent))
8970 if (!alloc_rt_sched_group(tg, parent))
8973 spin_lock_irqsave(&task_group_lock, flags);
8974 list_add_rcu(&tg->list, &task_groups);
8976 WARN_ON(!parent); /* root should already exist */
8978 tg->parent = parent;
8979 INIT_LIST_HEAD(&tg->children);
8980 list_add_rcu(&tg->siblings, &parent->children);
8981 spin_unlock_irqrestore(&task_group_lock, flags);
8986 free_sched_group(tg);
8987 return ERR_PTR(-ENOMEM);
8990 /* rcu callback to free various structures associated with a task group */
8991 static void free_sched_group_rcu(struct rcu_head *rhp)
8993 /* now it should be safe to free those cfs_rqs */
8994 free_sched_group(container_of(rhp, struct task_group, rcu));
8997 /* Destroy runqueue etc associated with a task group */
8998 void sched_destroy_group(struct task_group *tg)
9000 unsigned long flags;
9003 /* end participation in shares distribution */
9004 for_each_possible_cpu(i)
9005 unregister_fair_sched_group(tg, i);
9007 spin_lock_irqsave(&task_group_lock, flags);
9008 list_del_rcu(&tg->list);
9009 list_del_rcu(&tg->siblings);
9010 spin_unlock_irqrestore(&task_group_lock, flags);
9012 /* wait for possible concurrent references to cfs_rqs complete */
9013 call_rcu(&tg->rcu, free_sched_group_rcu);
9016 /* change task's runqueue when it moves between groups.
9017 * The caller of this function should have put the task in its new group
9018 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9019 * reflect its new group.
9021 void sched_move_task(struct task_struct *tsk)
9023 struct task_group *tg;
9025 unsigned long flags;
9028 rq = task_rq_lock(tsk, &flags);
9030 running = task_current(rq, tsk);
9034 dequeue_task(rq, tsk, 0);
9035 if (unlikely(running))
9036 tsk->sched_class->put_prev_task(rq, tsk);
9038 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
9039 lockdep_is_held(&tsk->sighand->siglock)),
9040 struct task_group, css);
9041 tg = autogroup_task_group(tsk, tg);
9042 tsk->sched_task_group = tg;
9044 #ifdef CONFIG_FAIR_GROUP_SCHED
9045 if (tsk->sched_class->task_move_group)
9046 tsk->sched_class->task_move_group(tsk, on_rq);
9049 set_task_rq(tsk, task_cpu(tsk));
9051 if (unlikely(running))
9052 tsk->sched_class->set_curr_task(rq);
9054 enqueue_task(rq, tsk, 0);
9056 task_rq_unlock(rq, tsk, &flags);
9058 #endif /* CONFIG_CGROUP_SCHED */
9060 #ifdef CONFIG_FAIR_GROUP_SCHED
9061 static DEFINE_MUTEX(shares_mutex);
9063 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9066 unsigned long flags;
9069 * We can't change the weight of the root cgroup.
9074 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9076 mutex_lock(&shares_mutex);
9077 if (tg->shares == shares)
9080 tg->shares = shares;
9081 for_each_possible_cpu(i) {
9082 struct rq *rq = cpu_rq(i);
9083 struct sched_entity *se;
9086 /* Propagate contribution to hierarchy */
9087 raw_spin_lock_irqsave(&rq->lock, flags);
9088 for_each_sched_entity(se)
9089 update_cfs_shares(group_cfs_rq(se));
9090 raw_spin_unlock_irqrestore(&rq->lock, flags);
9094 mutex_unlock(&shares_mutex);
9098 unsigned long sched_group_shares(struct task_group *tg)
9104 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
9105 static unsigned long to_ratio(u64 period, u64 runtime)
9107 if (runtime == RUNTIME_INF)
9110 return div64_u64(runtime << 20, period);
9114 #ifdef CONFIG_RT_GROUP_SCHED
9116 * Ensure that the real time constraints are schedulable.
9118 static DEFINE_MUTEX(rt_constraints_mutex);
9120 /* Must be called with tasklist_lock held */
9121 static inline int tg_has_rt_tasks(struct task_group *tg)
9123 struct task_struct *g, *p;
9126 * Autogroups do not have RT tasks; see autogroup_create().
9128 if (task_group_is_autogroup(tg))
9131 do_each_thread(g, p) {
9132 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9134 } while_each_thread(g, p);
9139 struct rt_schedulable_data {
9140 struct task_group *tg;
9145 static int tg_rt_schedulable(struct task_group *tg, void *data)
9147 struct rt_schedulable_data *d = data;
9148 struct task_group *child;
9149 unsigned long total, sum = 0;
9150 u64 period, runtime;
9152 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9153 runtime = tg->rt_bandwidth.rt_runtime;
9156 period = d->rt_period;
9157 runtime = d->rt_runtime;
9161 * Cannot have more runtime than the period.
9163 if (runtime > period && runtime != RUNTIME_INF)
9167 * Ensure we don't starve existing RT tasks.
9169 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9172 total = to_ratio(period, runtime);
9175 * Nobody can have more than the global setting allows.
9177 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9181 * The sum of our children's runtime should not exceed our own.
9183 list_for_each_entry_rcu(child, &tg->children, siblings) {
9184 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9185 runtime = child->rt_bandwidth.rt_runtime;
9187 if (child == d->tg) {
9188 period = d->rt_period;
9189 runtime = d->rt_runtime;
9192 sum += to_ratio(period, runtime);
9201 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9205 struct rt_schedulable_data data = {
9207 .rt_period = period,
9208 .rt_runtime = runtime,
9212 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
9218 static int tg_set_rt_bandwidth(struct task_group *tg,
9219 u64 rt_period, u64 rt_runtime)
9223 mutex_lock(&rt_constraints_mutex);
9224 read_lock(&tasklist_lock);
9225 err = __rt_schedulable(tg, rt_period, rt_runtime);
9229 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9230 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9231 tg->rt_bandwidth.rt_runtime = rt_runtime;
9233 for_each_possible_cpu(i) {
9234 struct rt_rq *rt_rq = tg->rt_rq[i];
9236 raw_spin_lock(&rt_rq->rt_runtime_lock);
9237 rt_rq->rt_runtime = rt_runtime;
9238 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9240 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9242 read_unlock(&tasklist_lock);
9243 mutex_unlock(&rt_constraints_mutex);
9248 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9250 u64 rt_runtime, rt_period;
9252 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9253 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9254 if (rt_runtime_us < 0)
9255 rt_runtime = RUNTIME_INF;
9257 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9260 long sched_group_rt_runtime(struct task_group *tg)
9264 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9267 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9268 do_div(rt_runtime_us, NSEC_PER_USEC);
9269 return rt_runtime_us;
9272 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9274 u64 rt_runtime, rt_period;
9276 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9277 rt_runtime = tg->rt_bandwidth.rt_runtime;
9282 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
9285 long sched_group_rt_period(struct task_group *tg)
9289 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9290 do_div(rt_period_us, NSEC_PER_USEC);
9291 return rt_period_us;
9294 static int sched_rt_global_constraints(void)
9296 u64 runtime, period;
9299 if (sysctl_sched_rt_period <= 0)
9302 runtime = global_rt_runtime();
9303 period = global_rt_period();
9306 * Sanity check on the sysctl variables.
9308 if (runtime > period && runtime != RUNTIME_INF)
9311 mutex_lock(&rt_constraints_mutex);
9312 read_lock(&tasklist_lock);
9313 ret = __rt_schedulable(NULL, 0, 0);
9314 read_unlock(&tasklist_lock);
9315 mutex_unlock(&rt_constraints_mutex);
9320 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9322 /* Don't accept realtime tasks when there is no way for them to run */
9323 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9329 #else /* !CONFIG_RT_GROUP_SCHED */
9330 static int sched_rt_global_constraints(void)
9332 unsigned long flags;
9335 if (sysctl_sched_rt_period <= 0)
9339 * There's always some RT tasks in the root group
9340 * -- migration, kstopmachine etc..
9342 if (sysctl_sched_rt_runtime == 0)
9345 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9346 for_each_possible_cpu(i) {
9347 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9349 raw_spin_lock(&rt_rq->rt_runtime_lock);
9350 rt_rq->rt_runtime = global_rt_runtime();
9351 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9353 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9357 #endif /* CONFIG_RT_GROUP_SCHED */
9359 int sched_rt_handler(struct ctl_table *table, int write,
9360 void __user *buffer, size_t *lenp,
9364 int old_period, old_runtime;
9365 static DEFINE_MUTEX(mutex);
9368 old_period = sysctl_sched_rt_period;
9369 old_runtime = sysctl_sched_rt_runtime;
9371 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9373 if (!ret && write) {
9374 ret = sched_rt_global_constraints();
9376 sysctl_sched_rt_period = old_period;
9377 sysctl_sched_rt_runtime = old_runtime;
9379 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9380 def_rt_bandwidth.rt_period =
9381 ns_to_ktime(global_rt_period());
9384 mutex_unlock(&mutex);
9389 #ifdef CONFIG_CGROUP_SCHED
9391 /* return corresponding task_group object of a cgroup */
9392 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9394 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9395 struct task_group, css);
9398 static struct cgroup_subsys_state *
9399 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9401 struct task_group *tg, *parent;
9403 if (!cgrp->parent) {
9404 /* This is early initialization for the top cgroup */
9405 return &root_task_group.css;
9408 parent = cgroup_tg(cgrp->parent);
9409 tg = sched_create_group(parent);
9411 return ERR_PTR(-ENOMEM);
9417 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9419 struct task_group *tg = cgroup_tg(cgrp);
9421 sched_destroy_group(tg);
9425 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9427 #ifdef CONFIG_RT_GROUP_SCHED
9428 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9431 /* We don't support RT-tasks being in separate groups */
9432 if (tsk->sched_class != &fair_sched_class)
9439 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9441 sched_move_task(tsk);
9445 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9446 struct cgroup *old_cgrp, struct task_struct *task)
9449 * cgroup_exit() is called in the copy_process() failure path.
9450 * Ignore this case since the task hasn't ran yet, this avoids
9451 * trying to poke a half freed task state from generic code.
9453 if (!(task->flags & PF_EXITING))
9456 sched_move_task(task);
9459 #ifdef CONFIG_FAIR_GROUP_SCHED
9460 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9463 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9466 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9468 struct task_group *tg = cgroup_tg(cgrp);
9470 return (u64) scale_load_down(tg->shares);
9473 #ifdef CONFIG_CFS_BANDWIDTH
9474 static DEFINE_MUTEX(cfs_constraints_mutex);
9476 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9477 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9479 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9481 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9483 int i, ret = 0, runtime_enabled;
9484 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9486 if (tg == &root_task_group)
9490 * Ensure we have at some amount of bandwidth every period. This is
9491 * to prevent reaching a state of large arrears when throttled via
9492 * entity_tick() resulting in prolonged exit starvation.
9494 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9498 * Likewise, bound things on the otherside by preventing insane quota
9499 * periods. This also allows us to normalize in computing quota
9502 if (period > max_cfs_quota_period)
9505 mutex_lock(&cfs_constraints_mutex);
9506 ret = __cfs_schedulable(tg, period, quota);
9510 runtime_enabled = quota != RUNTIME_INF;
9511 raw_spin_lock_irq(&cfs_b->lock);
9512 cfs_b->period = ns_to_ktime(period);
9513 cfs_b->quota = quota;
9515 __refill_cfs_bandwidth_runtime(cfs_b);
9516 /* restart the period timer (if active) to handle new period expiry */
9517 if (runtime_enabled && cfs_b->timer_active) {
9518 /* force a reprogram */
9519 cfs_b->timer_active = 0;
9520 __start_cfs_bandwidth(cfs_b);
9522 raw_spin_unlock_irq(&cfs_b->lock);
9524 for_each_possible_cpu(i) {
9525 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9526 struct rq *rq = rq_of(cfs_rq);
9528 raw_spin_lock_irq(&rq->lock);
9529 cfs_rq->runtime_enabled = runtime_enabled;
9530 cfs_rq->runtime_remaining = 0;
9532 if (cfs_rq_throttled(cfs_rq))
9533 unthrottle_cfs_rq(cfs_rq);
9534 raw_spin_unlock_irq(&rq->lock);
9537 mutex_unlock(&cfs_constraints_mutex);
9542 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9546 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9547 if (cfs_quota_us < 0)
9548 quota = RUNTIME_INF;
9550 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9552 return tg_set_cfs_bandwidth(tg, period, quota);
9555 long tg_get_cfs_quota(struct task_group *tg)
9559 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9562 quota_us = tg_cfs_bandwidth(tg)->quota;
9563 do_div(quota_us, NSEC_PER_USEC);
9568 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9572 period = (u64)cfs_period_us * NSEC_PER_USEC;
9573 quota = tg_cfs_bandwidth(tg)->quota;
9578 return tg_set_cfs_bandwidth(tg, period, quota);
9581 long tg_get_cfs_period(struct task_group *tg)
9585 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9586 do_div(cfs_period_us, NSEC_PER_USEC);
9588 return cfs_period_us;
9591 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9593 return tg_get_cfs_quota(cgroup_tg(cgrp));
9596 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9599 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9602 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9604 return tg_get_cfs_period(cgroup_tg(cgrp));
9607 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9610 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9613 struct cfs_schedulable_data {
9614 struct task_group *tg;
9619 * normalize group quota/period to be quota/max_period
9620 * note: units are usecs
9622 static u64 normalize_cfs_quota(struct task_group *tg,
9623 struct cfs_schedulable_data *d)
9631 period = tg_get_cfs_period(tg);
9632 quota = tg_get_cfs_quota(tg);
9635 /* note: these should typically be equivalent */
9636 if (quota == RUNTIME_INF || quota == -1)
9639 return to_ratio(period, quota);
9642 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9644 struct cfs_schedulable_data *d = data;
9645 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9646 s64 quota = 0, parent_quota = -1;
9649 quota = RUNTIME_INF;
9651 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9653 quota = normalize_cfs_quota(tg, d);
9654 parent_quota = parent_b->hierarchal_quota;
9657 * ensure max(child_quota) <= parent_quota, inherit when no
9660 if (quota == RUNTIME_INF)
9661 quota = parent_quota;
9662 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9665 cfs_b->hierarchal_quota = quota;
9670 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9673 struct cfs_schedulable_data data = {
9679 if (quota != RUNTIME_INF) {
9680 do_div(data.period, NSEC_PER_USEC);
9681 do_div(data.quota, NSEC_PER_USEC);
9685 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9691 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9692 struct cgroup_map_cb *cb)
9694 struct task_group *tg = cgroup_tg(cgrp);
9695 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9697 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9698 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9699 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9703 #endif /* CONFIG_CFS_BANDWIDTH */
9704 #endif /* CONFIG_FAIR_GROUP_SCHED */
9706 #ifdef CONFIG_RT_GROUP_SCHED
9707 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9710 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9713 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9715 return sched_group_rt_runtime(cgroup_tg(cgrp));
9718 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9721 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9724 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9726 return sched_group_rt_period(cgroup_tg(cgrp));
9728 #endif /* CONFIG_RT_GROUP_SCHED */
9730 static struct cftype cpu_files[] = {
9731 #ifdef CONFIG_FAIR_GROUP_SCHED
9734 .read_u64 = cpu_shares_read_u64,
9735 .write_u64 = cpu_shares_write_u64,
9738 #ifdef CONFIG_CFS_BANDWIDTH
9740 .name = "cfs_quota_us",
9741 .read_s64 = cpu_cfs_quota_read_s64,
9742 .write_s64 = cpu_cfs_quota_write_s64,
9745 .name = "cfs_period_us",
9746 .read_u64 = cpu_cfs_period_read_u64,
9747 .write_u64 = cpu_cfs_period_write_u64,
9751 .read_map = cpu_stats_show,
9754 #ifdef CONFIG_RT_GROUP_SCHED
9756 .name = "rt_runtime_us",
9757 .read_s64 = cpu_rt_runtime_read,
9758 .write_s64 = cpu_rt_runtime_write,
9761 .name = "rt_period_us",
9762 .read_u64 = cpu_rt_period_read_uint,
9763 .write_u64 = cpu_rt_period_write_uint,
9768 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9770 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9773 struct cgroup_subsys cpu_cgroup_subsys = {
9775 .create = cpu_cgroup_create,
9776 .destroy = cpu_cgroup_destroy,
9777 .can_attach_task = cpu_cgroup_can_attach_task,
9778 .attach_task = cpu_cgroup_attach_task,
9779 .exit = cpu_cgroup_exit,
9780 .populate = cpu_cgroup_populate,
9781 .subsys_id = cpu_cgroup_subsys_id,
9785 #endif /* CONFIG_CGROUP_SCHED */
9787 #ifdef CONFIG_CGROUP_CPUACCT
9790 * CPU accounting code for task groups.
9792 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9793 * (balbir@in.ibm.com).
9796 /* track cpu usage of a group of tasks and its child groups */
9798 struct cgroup_subsys_state css;
9799 /* cpuusage holds pointer to a u64-type object on every cpu */
9800 u64 __percpu *cpuusage;
9801 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9802 struct cpuacct *parent;
9805 struct cgroup_subsys cpuacct_subsys;
9807 /* return cpu accounting group corresponding to this container */
9808 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9810 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9811 struct cpuacct, css);
9814 /* return cpu accounting group to which this task belongs */
9815 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9817 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9818 struct cpuacct, css);
9821 /* create a new cpu accounting group */
9822 static struct cgroup_subsys_state *cpuacct_create(
9823 struct cgroup_subsys *ss, struct cgroup *cgrp)
9825 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9831 ca->cpuusage = alloc_percpu(u64);
9835 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9836 if (percpu_counter_init(&ca->cpustat[i], 0))
9837 goto out_free_counters;
9840 ca->parent = cgroup_ca(cgrp->parent);
9846 percpu_counter_destroy(&ca->cpustat[i]);
9847 free_percpu(ca->cpuusage);
9851 return ERR_PTR(-ENOMEM);
9854 /* destroy an existing cpu accounting group */
9856 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9858 struct cpuacct *ca = cgroup_ca(cgrp);
9861 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9862 percpu_counter_destroy(&ca->cpustat[i]);
9863 free_percpu(ca->cpuusage);
9867 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9869 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9872 #ifndef CONFIG_64BIT
9874 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9876 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9878 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9886 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9888 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9890 #ifndef CONFIG_64BIT
9892 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9894 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9896 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9902 /* return total cpu usage (in nanoseconds) of a group */
9903 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9905 struct cpuacct *ca = cgroup_ca(cgrp);
9906 u64 totalcpuusage = 0;
9909 for_each_present_cpu(i)
9910 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9912 return totalcpuusage;
9915 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9918 struct cpuacct *ca = cgroup_ca(cgrp);
9927 for_each_present_cpu(i)
9928 cpuacct_cpuusage_write(ca, i, 0);
9934 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9937 struct cpuacct *ca = cgroup_ca(cgroup);
9941 for_each_present_cpu(i) {
9942 percpu = cpuacct_cpuusage_read(ca, i);
9943 seq_printf(m, "%llu ", (unsigned long long) percpu);
9945 seq_printf(m, "\n");
9949 static const char *cpuacct_stat_desc[] = {
9950 [CPUACCT_STAT_USER] = "user",
9951 [CPUACCT_STAT_SYSTEM] = "system",
9954 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9955 struct cgroup_map_cb *cb)
9957 struct cpuacct *ca = cgroup_ca(cgrp);
9960 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9961 s64 val = percpu_counter_read(&ca->cpustat[i]);
9962 val = cputime64_to_clock_t(val);
9963 cb->fill(cb, cpuacct_stat_desc[i], val);
9968 static struct cftype files[] = {
9971 .read_u64 = cpuusage_read,
9972 .write_u64 = cpuusage_write,
9975 .name = "usage_percpu",
9976 .read_seq_string = cpuacct_percpu_seq_read,
9980 .read_map = cpuacct_stats_show,
9984 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9986 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9990 * charge this task's execution time to its accounting group.
9992 * called with rq->lock held.
9994 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9999 if (unlikely(!cpuacct_subsys.active))
10002 cpu = task_cpu(tsk);
10008 for (; ca; ca = ca->parent) {
10009 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10010 *cpuusage += cputime;
10017 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
10018 * in cputime_t units. As a result, cpuacct_update_stats calls
10019 * percpu_counter_add with values large enough to always overflow the
10020 * per cpu batch limit causing bad SMP scalability.
10022 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
10023 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10024 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10027 #define CPUACCT_BATCH \
10028 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10030 #define CPUACCT_BATCH 0
10034 * Charge the system/user time to the task's accounting group.
10036 static void cpuacct_update_stats(struct task_struct *tsk,
10037 enum cpuacct_stat_index idx, cputime_t val)
10039 struct cpuacct *ca;
10040 int batch = CPUACCT_BATCH;
10042 if (unlikely(!cpuacct_subsys.active))
10049 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10055 struct cgroup_subsys cpuacct_subsys = {
10057 .create = cpuacct_create,
10058 .destroy = cpuacct_destroy,
10059 .populate = cpuacct_populate,
10060 .subsys_id = cpuacct_subsys_id,
10062 #endif /* CONFIG_CGROUP_CPUACCT */