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>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy
)
130 if (policy
== SCHED_FIFO
|| policy
== SCHED_RR
)
135 static inline int task_has_rt_policy(struct task_struct
*p
)
137 return rt_policy(p
->policy
);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array
{
144 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
145 struct list_head queue
[MAX_RT_PRIO
];
148 struct rt_bandwidth
{
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock
;
153 struct hrtimer rt_period_timer
;
156 static struct rt_bandwidth def_rt_bandwidth
;
158 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
160 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
162 struct rt_bandwidth
*rt_b
=
163 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
169 now
= hrtimer_cb_get_time(timer
);
170 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
175 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
178 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
182 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
184 rt_b
->rt_period
= ns_to_ktime(period
);
185 rt_b
->rt_runtime
= runtime
;
187 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
189 hrtimer_init(&rt_b
->rt_period_timer
,
190 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
191 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime
>= 0;
199 static void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
202 ktime_t soft
, hard
, now
;
205 if (hrtimer_active(period_timer
))
208 now
= hrtimer_cb_get_time(period_timer
);
209 hrtimer_forward(period_timer
, now
, period
);
211 soft
= hrtimer_get_softexpires(period_timer
);
212 hard
= hrtimer_get_expires(period_timer
);
213 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
214 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
215 HRTIMER_MODE_ABS_PINNED
, 0);
219 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
221 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
224 if (hrtimer_active(&rt_b
->rt_period_timer
))
227 raw_spin_lock(&rt_b
->rt_runtime_lock
);
228 start_bandwidth_timer(&rt_b
->rt_period_timer
, rt_b
->rt_period
);
229 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
232 #ifdef CONFIG_RT_GROUP_SCHED
233 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
235 hrtimer_cancel(&rt_b
->rt_period_timer
);
240 * sched_domains_mutex serializes calls to init_sched_domains,
241 * detach_destroy_domains and partition_sched_domains.
243 static DEFINE_MUTEX(sched_domains_mutex
);
245 #ifdef CONFIG_CGROUP_SCHED
247 #include <linux/cgroup.h>
251 static LIST_HEAD(task_groups
);
253 struct cfs_bandwidth
{
254 #ifdef CONFIG_CFS_BANDWIDTH
258 s64 hierarchal_quota
;
261 int idle
, timer_active
;
262 struct hrtimer period_timer
, slack_timer
;
263 struct list_head throttled_cfs_rq
;
266 int nr_periods
, nr_throttled
;
271 /* task group related information */
273 struct cgroup_subsys_state css
;
275 #ifdef CONFIG_FAIR_GROUP_SCHED
276 /* schedulable entities of this group on each cpu */
277 struct sched_entity
**se
;
278 /* runqueue "owned" by this group on each cpu */
279 struct cfs_rq
**cfs_rq
;
280 unsigned long shares
;
282 atomic_t load_weight
;
285 #ifdef CONFIG_RT_GROUP_SCHED
286 struct sched_rt_entity
**rt_se
;
287 struct rt_rq
**rt_rq
;
289 struct rt_bandwidth rt_bandwidth
;
293 struct list_head list
;
295 struct task_group
*parent
;
296 struct list_head siblings
;
297 struct list_head children
;
299 #ifdef CONFIG_SCHED_AUTOGROUP
300 struct autogroup
*autogroup
;
303 struct cfs_bandwidth cfs_bandwidth
;
306 /* task_group_lock serializes the addition/removal of task groups */
307 static DEFINE_SPINLOCK(task_group_lock
);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
311 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
314 * A weight of 0 or 1 can cause arithmetics problems.
315 * A weight of a cfs_rq is the sum of weights of which entities
316 * are queued on this cfs_rq, so a weight of a entity should not be
317 * too large, so as the shares value of a task group.
318 * (The default weight is 1024 - so there's no practical
319 * limitation from this.)
321 #define MIN_SHARES (1UL << 1)
322 #define MAX_SHARES (1UL << 18)
324 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
327 /* Default task group.
328 * Every task in system belong to this group at bootup.
330 struct task_group root_task_group
;
332 #endif /* CONFIG_CGROUP_SCHED */
334 /* CFS-related fields in a runqueue */
336 struct load_weight load
;
337 unsigned long nr_running
, h_nr_running
;
342 u64 min_vruntime_copy
;
345 struct rb_root tasks_timeline
;
346 struct rb_node
*rb_leftmost
;
348 struct list_head tasks
;
349 struct list_head
*balance_iterator
;
352 * 'curr' points to currently running entity on this cfs_rq.
353 * It is set to NULL otherwise (i.e when none are currently running).
355 struct sched_entity
*curr
, *next
, *last
, *skip
;
357 #ifdef CONFIG_SCHED_DEBUG
358 unsigned int nr_spread_over
;
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
365 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
366 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
367 * (like users, containers etc.)
369 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
370 * list is used during load balance.
373 struct list_head leaf_cfs_rq_list
;
374 struct task_group
*tg
; /* group that "owns" this runqueue */
378 * the part of load.weight contributed by tasks
380 unsigned long task_weight
;
383 * h_load = weight * f(tg)
385 * Where f(tg) is the recursive weight fraction assigned to
388 unsigned long h_load
;
391 * Maintaining per-cpu shares distribution for group scheduling
393 * load_stamp is the last time we updated the load average
394 * load_last is the last time we updated the load average and saw load
395 * load_unacc_exec_time is currently unaccounted execution time
399 u64 load_stamp
, load_last
, load_unacc_exec_time
;
401 unsigned long load_contribution
;
403 #ifdef CONFIG_CFS_BANDWIDTH
406 s64 runtime_remaining
;
408 u64 throttled_timestamp
;
409 int throttled
, throttle_count
;
410 struct list_head throttled_list
;
415 #ifdef CONFIG_FAIR_GROUP_SCHED
416 #ifdef CONFIG_CFS_BANDWIDTH
417 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
419 return &tg
->cfs_bandwidth
;
422 static inline u64
default_cfs_period(void);
423 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
);
424 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
);
426 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
428 struct cfs_bandwidth
*cfs_b
=
429 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
430 do_sched_cfs_slack_timer(cfs_b
);
432 return HRTIMER_NORESTART
;
435 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
437 struct cfs_bandwidth
*cfs_b
=
438 container_of(timer
, struct cfs_bandwidth
, period_timer
);
444 now
= hrtimer_cb_get_time(timer
);
445 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
450 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
453 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
456 static void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
458 raw_spin_lock_init(&cfs_b
->lock
);
460 cfs_b
->quota
= RUNTIME_INF
;
461 cfs_b
->period
= ns_to_ktime(default_cfs_period());
463 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
464 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
465 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
466 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
467 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
470 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
472 cfs_rq
->runtime_enabled
= 0;
473 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
476 /* requires cfs_b->lock, may release to reprogram timer */
477 static void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
480 * The timer may be active because we're trying to set a new bandwidth
481 * period or because we're racing with the tear-down path
482 * (timer_active==0 becomes visible before the hrtimer call-back
483 * terminates). In either case we ensure that it's re-programmed
485 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
486 raw_spin_unlock(&cfs_b
->lock
);
487 /* ensure cfs_b->lock is available while we wait */
488 hrtimer_cancel(&cfs_b
->period_timer
);
490 raw_spin_lock(&cfs_b
->lock
);
491 /* if someone else restarted the timer then we're done */
492 if (cfs_b
->timer_active
)
496 cfs_b
->timer_active
= 1;
497 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
500 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
502 hrtimer_cancel(&cfs_b
->period_timer
);
503 hrtimer_cancel(&cfs_b
->slack_timer
);
506 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
507 static void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
508 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
510 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
514 #endif /* CONFIG_CFS_BANDWIDTH */
515 #endif /* CONFIG_FAIR_GROUP_SCHED */
517 /* Real-Time classes' related field in a runqueue: */
519 struct rt_prio_array active
;
520 unsigned long rt_nr_running
;
521 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
523 int curr
; /* highest queued rt task prio */
525 int next
; /* next highest */
530 unsigned long rt_nr_migratory
;
531 unsigned long rt_nr_total
;
533 struct plist_head pushable_tasks
;
538 /* Nests inside the rq lock: */
539 raw_spinlock_t rt_runtime_lock
;
541 #ifdef CONFIG_RT_GROUP_SCHED
542 unsigned long rt_nr_boosted
;
545 struct list_head leaf_rt_rq_list
;
546 struct task_group
*tg
;
553 * We add the notion of a root-domain which will be used to define per-domain
554 * variables. Each exclusive cpuset essentially defines an island domain by
555 * fully partitioning the member cpus from any other cpuset. Whenever a new
556 * exclusive cpuset is created, we also create and attach a new root-domain
565 cpumask_var_t online
;
568 * The "RT overload" flag: it gets set if a CPU has more than
569 * one runnable RT task.
571 cpumask_var_t rto_mask
;
572 struct cpupri cpupri
;
576 * By default the system creates a single root-domain with all cpus as
577 * members (mimicking the global state we have today).
579 static struct root_domain def_root_domain
;
581 #endif /* CONFIG_SMP */
584 * This is the main, per-CPU runqueue data structure.
586 * Locking rule: those places that want to lock multiple runqueues
587 * (such as the load balancing or the thread migration code), lock
588 * acquire operations must be ordered by ascending &runqueue.
595 * nr_running and cpu_load should be in the same cacheline because
596 * remote CPUs use both these fields when doing load calculation.
598 unsigned long nr_running
;
599 #define CPU_LOAD_IDX_MAX 5
600 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
601 unsigned long last_load_update_tick
;
604 unsigned char nohz_balance_kick
;
606 int skip_clock_update
;
608 /* capture load from *all* tasks on this cpu: */
609 struct load_weight load
;
610 unsigned long nr_load_updates
;
616 #ifdef CONFIG_FAIR_GROUP_SCHED
617 /* list of leaf cfs_rq on this cpu: */
618 struct list_head leaf_cfs_rq_list
;
620 #ifdef CONFIG_RT_GROUP_SCHED
621 struct list_head leaf_rt_rq_list
;
625 * This is part of a global counter where only the total sum
626 * over all CPUs matters. A task can increase this counter on
627 * one CPU and if it got migrated afterwards it may decrease
628 * it on another CPU. Always updated under the runqueue lock:
630 unsigned long nr_uninterruptible
;
632 struct task_struct
*curr
, *idle
, *stop
;
633 unsigned long next_balance
;
634 struct mm_struct
*prev_mm
;
642 struct root_domain
*rd
;
643 struct sched_domain
*sd
;
645 unsigned long cpu_power
;
647 unsigned char idle_at_tick
;
648 /* For active balancing */
652 struct cpu_stop_work active_balance_work
;
653 /* cpu of this runqueue: */
663 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
666 #ifdef CONFIG_PARAVIRT
669 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
670 u64 prev_steal_time_rq
;
673 /* calc_load related fields */
674 unsigned long calc_load_update
;
675 long calc_load_active
;
677 #ifdef CONFIG_SCHED_HRTICK
679 int hrtick_csd_pending
;
680 struct call_single_data hrtick_csd
;
682 struct hrtimer hrtick_timer
;
685 #ifdef CONFIG_SCHEDSTATS
687 struct sched_info rq_sched_info
;
688 unsigned long long rq_cpu_time
;
689 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
691 /* sys_sched_yield() stats */
692 unsigned int yld_count
;
694 /* schedule() stats */
695 unsigned int sched_switch
;
696 unsigned int sched_count
;
697 unsigned int sched_goidle
;
699 /* try_to_wake_up() stats */
700 unsigned int ttwu_count
;
701 unsigned int ttwu_local
;
705 struct llist_head wake_list
;
709 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
712 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
714 static inline int cpu_of(struct rq
*rq
)
723 #define rcu_dereference_check_sched_domain(p) \
724 rcu_dereference_check((p), \
725 lockdep_is_held(&sched_domains_mutex))
728 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
729 * See detach_destroy_domains: synchronize_sched for details.
731 * The domain tree of any CPU may only be accessed from within
732 * preempt-disabled sections.
734 #define for_each_domain(cpu, __sd) \
735 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
737 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
738 #define this_rq() (&__get_cpu_var(runqueues))
739 #define task_rq(p) cpu_rq(task_cpu(p))
740 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
741 #define raw_rq() (&__raw_get_cpu_var(runqueues))
743 #ifdef CONFIG_CGROUP_SCHED
746 * Return the group to which this tasks belongs.
748 * We use task_subsys_state_check() and extend the RCU verification with
749 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
750 * task it moves into the cgroup. Therefore by holding either of those locks,
751 * we pin the task to the current cgroup.
753 static inline struct task_group
*task_group(struct task_struct
*p
)
755 struct task_group
*tg
;
756 struct cgroup_subsys_state
*css
;
758 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
759 lockdep_is_held(&p
->pi_lock
) ||
760 lockdep_is_held(&task_rq(p
)->lock
));
761 tg
= container_of(css
, struct task_group
, css
);
763 return autogroup_task_group(p
, tg
);
766 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
767 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
769 #ifdef CONFIG_FAIR_GROUP_SCHED
770 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
771 p
->se
.parent
= task_group(p
)->se
[cpu
];
774 #ifdef CONFIG_RT_GROUP_SCHED
775 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
776 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
780 #else /* CONFIG_CGROUP_SCHED */
782 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
783 static inline struct task_group
*task_group(struct task_struct
*p
)
788 #endif /* CONFIG_CGROUP_SCHED */
790 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
792 static void update_rq_clock(struct rq
*rq
)
796 if (rq
->skip_clock_update
> 0)
799 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
801 update_rq_clock_task(rq
, delta
);
805 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
807 #ifdef CONFIG_SCHED_DEBUG
808 # define const_debug __read_mostly
810 # define const_debug static const
814 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
815 * @cpu: the processor in question.
817 * This interface allows printk to be called with the runqueue lock
818 * held and know whether or not it is OK to wake up the klogd.
820 int runqueue_is_locked(int cpu
)
822 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
826 * Debugging: various feature bits
829 #define SCHED_FEAT(name, enabled) \
830 __SCHED_FEAT_##name ,
833 #include "sched_features.h"
838 #define SCHED_FEAT(name, enabled) \
839 (1UL << __SCHED_FEAT_##name) * enabled |
841 const_debug
unsigned int sysctl_sched_features
=
842 #include "sched_features.h"
847 #ifdef CONFIG_SCHED_DEBUG
848 #define SCHED_FEAT(name, enabled) \
851 static __read_mostly
char *sched_feat_names
[] = {
852 #include "sched_features.h"
858 static int sched_feat_show(struct seq_file
*m
, void *v
)
862 for (i
= 0; sched_feat_names
[i
]; i
++) {
863 if (!(sysctl_sched_features
& (1UL << i
)))
865 seq_printf(m
, "%s ", sched_feat_names
[i
]);
873 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
874 size_t cnt
, loff_t
*ppos
)
884 if (copy_from_user(&buf
, ubuf
, cnt
))
890 if (strncmp(cmp
, "NO_", 3) == 0) {
895 for (i
= 0; sched_feat_names
[i
]; i
++) {
896 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
898 sysctl_sched_features
&= ~(1UL << i
);
900 sysctl_sched_features
|= (1UL << i
);
905 if (!sched_feat_names
[i
])
913 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
915 return single_open(filp
, sched_feat_show
, NULL
);
918 static const struct file_operations sched_feat_fops
= {
919 .open
= sched_feat_open
,
920 .write
= sched_feat_write
,
923 .release
= single_release
,
926 static __init
int sched_init_debug(void)
928 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
933 late_initcall(sched_init_debug
);
937 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
940 * Number of tasks to iterate in a single balance run.
941 * Limited because this is done with IRQs disabled.
943 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
946 * period over which we average the RT time consumption, measured
951 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
954 * period over which we measure -rt task cpu usage in us.
957 unsigned int sysctl_sched_rt_period
= 1000000;
959 static __read_mostly
int scheduler_running
;
962 * part of the period that we allow rt tasks to run in us.
965 int sysctl_sched_rt_runtime
= 950000;
967 static inline u64
global_rt_period(void)
969 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
972 static inline u64
global_rt_runtime(void)
974 if (sysctl_sched_rt_runtime
< 0)
977 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
980 #ifndef prepare_arch_switch
981 # define prepare_arch_switch(next) do { } while (0)
983 #ifndef finish_arch_switch
984 # define finish_arch_switch(prev) do { } while (0)
987 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
989 return rq
->curr
== p
;
992 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
997 return task_current(rq
, p
);
1001 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1002 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1006 * We can optimise this out completely for !SMP, because the
1007 * SMP rebalancing from interrupt is the only thing that cares
1014 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1018 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1019 * We must ensure this doesn't happen until the switch is completely
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 calc_load_account_idle(struct rq
*this_rq
);
1888 static void update_sysctl(void);
1889 static int get_update_sysctl_factor(void);
1890 static void update_cpu_load(struct rq
*this_rq
);
1892 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1894 set_task_rq(p
, cpu
);
1897 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1898 * successfuly executed on another CPU. We must ensure that updates of
1899 * per-task data have been completed by this moment.
1902 task_thread_info(p
)->cpu
= cpu
;
1906 static const struct sched_class rt_sched_class
;
1908 #define sched_class_highest (&stop_sched_class)
1909 #define for_each_class(class) \
1910 for (class = sched_class_highest; class; class = class->next)
1912 #include "sched_stats.h"
1914 static void inc_nr_running(struct rq
*rq
)
1919 static void dec_nr_running(struct rq
*rq
)
1924 static void set_load_weight(struct task_struct
*p
)
1926 int prio
= p
->static_prio
- MAX_RT_PRIO
;
1927 struct load_weight
*load
= &p
->se
.load
;
1930 * SCHED_IDLE tasks get minimal weight:
1932 if (p
->policy
== SCHED_IDLE
) {
1933 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
1934 load
->inv_weight
= WMULT_IDLEPRIO
;
1938 load
->weight
= scale_load(prio_to_weight
[prio
]);
1939 load
->inv_weight
= prio_to_wmult
[prio
];
1942 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1944 update_rq_clock(rq
);
1945 sched_info_queued(p
);
1946 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1949 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1951 update_rq_clock(rq
);
1952 sched_info_dequeued(p
);
1953 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1957 * activate_task - move a task to the runqueue.
1959 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1961 if (task_contributes_to_load(p
))
1962 rq
->nr_uninterruptible
--;
1964 enqueue_task(rq
, p
, flags
);
1968 * deactivate_task - remove a task from the runqueue.
1970 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1972 if (task_contributes_to_load(p
))
1973 rq
->nr_uninterruptible
++;
1975 dequeue_task(rq
, p
, flags
);
1978 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1981 * There are no locks covering percpu hardirq/softirq time.
1982 * They are only modified in account_system_vtime, on corresponding CPU
1983 * with interrupts disabled. So, writes are safe.
1984 * They are read and saved off onto struct rq in update_rq_clock().
1985 * This may result in other CPU reading this CPU's irq time and can
1986 * race with irq/account_system_vtime on this CPU. We would either get old
1987 * or new value with a side effect of accounting a slice of irq time to wrong
1988 * task when irq is in progress while we read rq->clock. That is a worthy
1989 * compromise in place of having locks on each irq in account_system_time.
1991 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1992 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1994 static DEFINE_PER_CPU(u64
, irq_start_time
);
1995 static int sched_clock_irqtime
;
1997 void enable_sched_clock_irqtime(void)
1999 sched_clock_irqtime
= 1;
2002 void disable_sched_clock_irqtime(void)
2004 sched_clock_irqtime
= 0;
2007 #ifndef CONFIG_64BIT
2008 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
2010 static inline void irq_time_write_begin(void)
2012 __this_cpu_inc(irq_time_seq
.sequence
);
2016 static inline void irq_time_write_end(void)
2019 __this_cpu_inc(irq_time_seq
.sequence
);
2022 static inline u64
irq_time_read(int cpu
)
2028 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
2029 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
2030 per_cpu(cpu_hardirq_time
, cpu
);
2031 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
2035 #else /* CONFIG_64BIT */
2036 static inline void irq_time_write_begin(void)
2040 static inline void irq_time_write_end(void)
2044 static inline u64
irq_time_read(int cpu
)
2046 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
2048 #endif /* CONFIG_64BIT */
2051 * Called before incrementing preempt_count on {soft,}irq_enter
2052 * and before decrementing preempt_count on {soft,}irq_exit.
2054 void account_system_vtime(struct task_struct
*curr
)
2056 unsigned long flags
;
2060 if (!sched_clock_irqtime
)
2063 local_irq_save(flags
);
2065 cpu
= smp_processor_id();
2066 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
2067 __this_cpu_add(irq_start_time
, delta
);
2069 irq_time_write_begin();
2071 * We do not account for softirq time from ksoftirqd here.
2072 * We want to continue accounting softirq time to ksoftirqd thread
2073 * in that case, so as not to confuse scheduler with a special task
2074 * that do not consume any time, but still wants to run.
2076 if (hardirq_count())
2077 __this_cpu_add(cpu_hardirq_time
, delta
);
2078 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
2079 __this_cpu_add(cpu_softirq_time
, delta
);
2081 irq_time_write_end();
2082 local_irq_restore(flags
);
2084 EXPORT_SYMBOL_GPL(account_system_vtime
);
2086 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2088 #ifdef CONFIG_PARAVIRT
2089 static inline u64
steal_ticks(u64 steal
)
2091 if (unlikely(steal
> NSEC_PER_SEC
))
2092 return div_u64(steal
, TICK_NSEC
);
2094 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
2098 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
2101 * In theory, the compile should just see 0 here, and optimize out the call
2102 * to sched_rt_avg_update. But I don't trust it...
2104 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2105 s64 steal
= 0, irq_delta
= 0;
2107 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2108 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
2111 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2112 * this case when a previous update_rq_clock() happened inside a
2113 * {soft,}irq region.
2115 * When this happens, we stop ->clock_task and only update the
2116 * prev_irq_time stamp to account for the part that fit, so that a next
2117 * update will consume the rest. This ensures ->clock_task is
2120 * It does however cause some slight miss-attribution of {soft,}irq
2121 * time, a more accurate solution would be to update the irq_time using
2122 * the current rq->clock timestamp, except that would require using
2125 if (irq_delta
> delta
)
2128 rq
->prev_irq_time
+= irq_delta
;
2131 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2132 if (static_branch((¶virt_steal_rq_enabled
))) {
2135 steal
= paravirt_steal_clock(cpu_of(rq
));
2136 steal
-= rq
->prev_steal_time_rq
;
2138 if (unlikely(steal
> delta
))
2141 st
= steal_ticks(steal
);
2142 steal
= st
* TICK_NSEC
;
2144 rq
->prev_steal_time_rq
+= steal
;
2150 rq
->clock_task
+= delta
;
2152 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2153 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
2154 sched_rt_avg_update(rq
, irq_delta
+ steal
);
2158 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2159 static int irqtime_account_hi_update(void)
2161 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2162 unsigned long flags
;
2166 local_irq_save(flags
);
2167 latest_ns
= this_cpu_read(cpu_hardirq_time
);
2168 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->irq
))
2170 local_irq_restore(flags
);
2174 static int irqtime_account_si_update(void)
2176 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2177 unsigned long flags
;
2181 local_irq_save(flags
);
2182 latest_ns
= this_cpu_read(cpu_softirq_time
);
2183 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->softirq
))
2185 local_irq_restore(flags
);
2189 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2191 #define sched_clock_irqtime (0)
2195 #include "sched_idletask.c"
2196 #include "sched_fair.c"
2197 #include "sched_rt.c"
2198 #include "sched_autogroup.c"
2199 #include "sched_stoptask.c"
2200 #ifdef CONFIG_SCHED_DEBUG
2201 # include "sched_debug.c"
2204 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2206 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2207 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2211 * Make it appear like a SCHED_FIFO task, its something
2212 * userspace knows about and won't get confused about.
2214 * Also, it will make PI more or less work without too
2215 * much confusion -- but then, stop work should not
2216 * rely on PI working anyway.
2218 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2220 stop
->sched_class
= &stop_sched_class
;
2223 cpu_rq(cpu
)->stop
= stop
;
2227 * Reset it back to a normal scheduling class so that
2228 * it can die in pieces.
2230 old_stop
->sched_class
= &rt_sched_class
;
2235 * __normal_prio - return the priority that is based on the static prio
2237 static inline int __normal_prio(struct task_struct
*p
)
2239 return p
->static_prio
;
2243 * Calculate the expected normal priority: i.e. priority
2244 * without taking RT-inheritance into account. Might be
2245 * boosted by interactivity modifiers. Changes upon fork,
2246 * setprio syscalls, and whenever the interactivity
2247 * estimator recalculates.
2249 static inline int normal_prio(struct task_struct
*p
)
2253 if (task_has_rt_policy(p
))
2254 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2256 prio
= __normal_prio(p
);
2261 * Calculate the current priority, i.e. the priority
2262 * taken into account by the scheduler. This value might
2263 * be boosted by RT tasks, or might be boosted by
2264 * interactivity modifiers. Will be RT if the task got
2265 * RT-boosted. If not then it returns p->normal_prio.
2267 static int effective_prio(struct task_struct
*p
)
2269 p
->normal_prio
= normal_prio(p
);
2271 * If we are RT tasks or we were boosted to RT priority,
2272 * keep the priority unchanged. Otherwise, update priority
2273 * to the normal priority:
2275 if (!rt_prio(p
->prio
))
2276 return p
->normal_prio
;
2281 * task_curr - is this task currently executing on a CPU?
2282 * @p: the task in question.
2284 inline int task_curr(const struct task_struct
*p
)
2286 return cpu_curr(task_cpu(p
)) == p
;
2289 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2290 const struct sched_class
*prev_class
,
2293 if (prev_class
!= p
->sched_class
) {
2294 if (prev_class
->switched_from
)
2295 prev_class
->switched_from(rq
, p
);
2296 p
->sched_class
->switched_to(rq
, p
);
2297 } else if (oldprio
!= p
->prio
)
2298 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2301 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2303 const struct sched_class
*class;
2305 if (p
->sched_class
== rq
->curr
->sched_class
) {
2306 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2308 for_each_class(class) {
2309 if (class == rq
->curr
->sched_class
)
2311 if (class == p
->sched_class
) {
2312 resched_task(rq
->curr
);
2319 * A queue event has occurred, and we're going to schedule. In
2320 * this case, we can save a useless back to back clock update.
2322 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
2323 rq
->skip_clock_update
= 1;
2328 * Is this task likely cache-hot:
2331 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2335 if (p
->sched_class
!= &fair_sched_class
)
2338 if (unlikely(p
->policy
== SCHED_IDLE
))
2342 * Buddy candidates are cache hot:
2344 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2345 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2346 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2349 if (sysctl_sched_migration_cost
== -1)
2351 if (sysctl_sched_migration_cost
== 0)
2354 delta
= now
- p
->se
.exec_start
;
2356 return delta
< (s64
)sysctl_sched_migration_cost
;
2359 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2361 #ifdef CONFIG_SCHED_DEBUG
2363 * We should never call set_task_cpu() on a blocked task,
2364 * ttwu() will sort out the placement.
2366 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2367 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2369 #ifdef CONFIG_LOCKDEP
2371 * The caller should hold either p->pi_lock or rq->lock, when changing
2372 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2374 * sched_move_task() holds both and thus holding either pins the cgroup,
2375 * see set_task_rq().
2377 * Furthermore, all task_rq users should acquire both locks, see
2380 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2381 lockdep_is_held(&task_rq(p
)->lock
)));
2385 trace_sched_migrate_task(p
, new_cpu
);
2387 if (task_cpu(p
) != new_cpu
) {
2388 p
->se
.nr_migrations
++;
2389 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
2392 __set_task_cpu(p
, new_cpu
);
2395 struct migration_arg
{
2396 struct task_struct
*task
;
2400 static int migration_cpu_stop(void *data
);
2403 * wait_task_inactive - wait for a thread to unschedule.
2405 * If @match_state is nonzero, it's the @p->state value just checked and
2406 * not expected to change. If it changes, i.e. @p might have woken up,
2407 * then return zero. When we succeed in waiting for @p to be off its CPU,
2408 * we return a positive number (its total switch count). If a second call
2409 * a short while later returns the same number, the caller can be sure that
2410 * @p has remained unscheduled the whole time.
2412 * The caller must ensure that the task *will* unschedule sometime soon,
2413 * else this function might spin for a *long* time. This function can't
2414 * be called with interrupts off, or it may introduce deadlock with
2415 * smp_call_function() if an IPI is sent by the same process we are
2416 * waiting to become inactive.
2418 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2420 unsigned long flags
;
2427 * We do the initial early heuristics without holding
2428 * any task-queue locks at all. We'll only try to get
2429 * the runqueue lock when things look like they will
2435 * If the task is actively running on another CPU
2436 * still, just relax and busy-wait without holding
2439 * NOTE! Since we don't hold any locks, it's not
2440 * even sure that "rq" stays as the right runqueue!
2441 * But we don't care, since "task_running()" will
2442 * return false if the runqueue has changed and p
2443 * is actually now running somewhere else!
2445 while (task_running(rq
, p
)) {
2446 if (match_state
&& unlikely(p
->state
!= match_state
))
2452 * Ok, time to look more closely! We need the rq
2453 * lock now, to be *sure*. If we're wrong, we'll
2454 * just go back and repeat.
2456 rq
= task_rq_lock(p
, &flags
);
2457 trace_sched_wait_task(p
);
2458 running
= task_running(rq
, p
);
2461 if (!match_state
|| p
->state
== match_state
)
2462 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2463 task_rq_unlock(rq
, p
, &flags
);
2466 * If it changed from the expected state, bail out now.
2468 if (unlikely(!ncsw
))
2472 * Was it really running after all now that we
2473 * checked with the proper locks actually held?
2475 * Oops. Go back and try again..
2477 if (unlikely(running
)) {
2483 * It's not enough that it's not actively running,
2484 * it must be off the runqueue _entirely_, and not
2487 * So if it was still runnable (but just not actively
2488 * running right now), it's preempted, and we should
2489 * yield - it could be a while.
2491 if (unlikely(on_rq
)) {
2492 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
2494 set_current_state(TASK_UNINTERRUPTIBLE
);
2495 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2500 * Ahh, all good. It wasn't running, and it wasn't
2501 * runnable, which means that it will never become
2502 * running in the future either. We're all done!
2511 * kick_process - kick a running thread to enter/exit the kernel
2512 * @p: the to-be-kicked thread
2514 * Cause a process which is running on another CPU to enter
2515 * kernel-mode, without any delay. (to get signals handled.)
2517 * NOTE: this function doesn't have to take the runqueue lock,
2518 * because all it wants to ensure is that the remote task enters
2519 * the kernel. If the IPI races and the task has been migrated
2520 * to another CPU then no harm is done and the purpose has been
2523 void kick_process(struct task_struct
*p
)
2529 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2530 smp_send_reschedule(cpu
);
2533 EXPORT_SYMBOL_GPL(kick_process
);
2534 #endif /* CONFIG_SMP */
2538 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2540 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2543 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2545 /* Look for allowed, online CPU in same node. */
2546 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2547 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2550 /* Any allowed, online CPU? */
2551 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2552 if (dest_cpu
< nr_cpu_ids
)
2555 /* No more Mr. Nice Guy. */
2556 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2558 * Don't tell them about moving exiting tasks or
2559 * kernel threads (both mm NULL), since they never
2562 if (p
->mm
&& printk_ratelimit()) {
2563 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2564 task_pid_nr(p
), p
->comm
, cpu
);
2571 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2574 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2576 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2579 * In order not to call set_task_cpu() on a blocking task we need
2580 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2583 * Since this is common to all placement strategies, this lives here.
2585 * [ this allows ->select_task() to simply return task_cpu(p) and
2586 * not worry about this generic constraint ]
2588 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2590 cpu
= select_fallback_rq(task_cpu(p
), p
);
2595 static void update_avg(u64
*avg
, u64 sample
)
2597 s64 diff
= sample
- *avg
;
2603 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2605 #ifdef CONFIG_SCHEDSTATS
2606 struct rq
*rq
= this_rq();
2609 int this_cpu
= smp_processor_id();
2611 if (cpu
== this_cpu
) {
2612 schedstat_inc(rq
, ttwu_local
);
2613 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2615 struct sched_domain
*sd
;
2617 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2619 for_each_domain(this_cpu
, sd
) {
2620 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2621 schedstat_inc(sd
, ttwu_wake_remote
);
2628 if (wake_flags
& WF_MIGRATED
)
2629 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2631 #endif /* CONFIG_SMP */
2633 schedstat_inc(rq
, ttwu_count
);
2634 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2636 if (wake_flags
& WF_SYNC
)
2637 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2639 #endif /* CONFIG_SCHEDSTATS */
2642 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
2644 activate_task(rq
, p
, en_flags
);
2647 /* if a worker is waking up, notify workqueue */
2648 if (p
->flags
& PF_WQ_WORKER
)
2649 wq_worker_waking_up(p
, cpu_of(rq
));
2653 * Mark the task runnable and perform wakeup-preemption.
2656 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2658 trace_sched_wakeup(p
, true);
2659 check_preempt_curr(rq
, p
, wake_flags
);
2661 p
->state
= TASK_RUNNING
;
2663 if (p
->sched_class
->task_woken
)
2664 p
->sched_class
->task_woken(rq
, p
);
2666 if (rq
->idle_stamp
) {
2667 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2668 u64 max
= 2*sysctl_sched_migration_cost
;
2673 update_avg(&rq
->avg_idle
, delta
);
2680 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2683 if (p
->sched_contributes_to_load
)
2684 rq
->nr_uninterruptible
--;
2687 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
2688 ttwu_do_wakeup(rq
, p
, wake_flags
);
2692 * Called in case the task @p isn't fully descheduled from its runqueue,
2693 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2694 * since all we need to do is flip p->state to TASK_RUNNING, since
2695 * the task is still ->on_rq.
2697 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2702 rq
= __task_rq_lock(p
);
2704 ttwu_do_wakeup(rq
, p
, wake_flags
);
2707 __task_rq_unlock(rq
);
2713 static void sched_ttwu_pending(void)
2715 struct rq
*rq
= this_rq();
2716 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2717 struct task_struct
*p
;
2719 raw_spin_lock(&rq
->lock
);
2722 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
2723 llist
= llist_next(llist
);
2724 ttwu_do_activate(rq
, p
, 0);
2727 raw_spin_unlock(&rq
->lock
);
2730 void scheduler_ipi(void)
2732 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
2736 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2737 * traditionally all their work was done from the interrupt return
2738 * path. Now that we actually do some work, we need to make sure
2741 * Some archs already do call them, luckily irq_enter/exit nest
2744 * Arguably we should visit all archs and update all handlers,
2745 * however a fair share of IPIs are still resched only so this would
2746 * somewhat pessimize the simple resched case.
2749 sched_ttwu_pending();
2752 * Check if someone kicked us for doing the nohz idle load balance.
2754 if (unlikely(got_nohz_idle_kick() && !need_resched()))
2755 raise_softirq_irqoff(SCHED_SOFTIRQ
);
2759 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
2761 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
2762 smp_send_reschedule(cpu
);
2765 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2766 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
2771 rq
= __task_rq_lock(p
);
2773 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2774 ttwu_do_wakeup(rq
, p
, wake_flags
);
2777 __task_rq_unlock(rq
);
2782 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2783 #endif /* CONFIG_SMP */
2785 static void ttwu_queue(struct task_struct
*p
, int cpu
)
2787 struct rq
*rq
= cpu_rq(cpu
);
2789 #if defined(CONFIG_SMP)
2790 if (sched_feat(TTWU_QUEUE
) && cpu
!= smp_processor_id()) {
2791 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
2792 ttwu_queue_remote(p
, cpu
);
2797 raw_spin_lock(&rq
->lock
);
2798 ttwu_do_activate(rq
, p
, 0);
2799 raw_spin_unlock(&rq
->lock
);
2803 * try_to_wake_up - wake up a thread
2804 * @p: the thread to be awakened
2805 * @state: the mask of task states that can be woken
2806 * @wake_flags: wake modifier flags (WF_*)
2808 * Put it on the run-queue if it's not already there. The "current"
2809 * thread is always on the run-queue (except when the actual
2810 * re-schedule is in progress), and as such you're allowed to do
2811 * the simpler "current->state = TASK_RUNNING" to mark yourself
2812 * runnable without the overhead of this.
2814 * Returns %true if @p was woken up, %false if it was already running
2815 * or @state didn't match @p's state.
2818 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2820 unsigned long flags
;
2821 int cpu
, success
= 0;
2824 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2825 if (!(p
->state
& state
))
2828 success
= 1; /* we're going to change ->state */
2831 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2836 * If the owning (remote) cpu is still in the middle of schedule() with
2837 * this task as prev, wait until its done referencing the task.
2840 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2842 * In case the architecture enables interrupts in
2843 * context_switch(), we cannot busy wait, since that
2844 * would lead to deadlocks when an interrupt hits and
2845 * tries to wake up @prev. So bail and do a complete
2848 if (ttwu_activate_remote(p
, wake_flags
))
2855 * Pairs with the smp_wmb() in finish_lock_switch().
2859 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2860 p
->state
= TASK_WAKING
;
2862 if (p
->sched_class
->task_waking
)
2863 p
->sched_class
->task_waking(p
);
2865 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2866 if (task_cpu(p
) != cpu
) {
2867 wake_flags
|= WF_MIGRATED
;
2868 set_task_cpu(p
, cpu
);
2870 #endif /* CONFIG_SMP */
2874 ttwu_stat(p
, cpu
, wake_flags
);
2876 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2882 * try_to_wake_up_local - try to wake up a local task with rq lock held
2883 * @p: the thread to be awakened
2885 * Put @p on the run-queue if it's not already there. The caller must
2886 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2889 static void try_to_wake_up_local(struct task_struct
*p
)
2891 struct rq
*rq
= task_rq(p
);
2893 BUG_ON(rq
!= this_rq());
2894 BUG_ON(p
== current
);
2895 lockdep_assert_held(&rq
->lock
);
2897 if (!raw_spin_trylock(&p
->pi_lock
)) {
2898 raw_spin_unlock(&rq
->lock
);
2899 raw_spin_lock(&p
->pi_lock
);
2900 raw_spin_lock(&rq
->lock
);
2903 if (!(p
->state
& TASK_NORMAL
))
2907 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2909 ttwu_do_wakeup(rq
, p
, 0);
2910 ttwu_stat(p
, smp_processor_id(), 0);
2912 raw_spin_unlock(&p
->pi_lock
);
2916 * wake_up_process - Wake up a specific process
2917 * @p: The process to be woken up.
2919 * Attempt to wake up the nominated process and move it to the set of runnable
2920 * processes. Returns 1 if the process was woken up, 0 if it was already
2923 * It may be assumed that this function implies a write memory barrier before
2924 * changing the task state if and only if any tasks are woken up.
2926 int wake_up_process(struct task_struct
*p
)
2928 return try_to_wake_up(p
, TASK_ALL
, 0);
2930 EXPORT_SYMBOL(wake_up_process
);
2932 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2934 return try_to_wake_up(p
, state
, 0);
2938 * Perform scheduler related setup for a newly forked process p.
2939 * p is forked by current.
2941 * __sched_fork() is basic setup used by init_idle() too:
2943 static void __sched_fork(struct task_struct
*p
)
2948 p
->se
.exec_start
= 0;
2949 p
->se
.sum_exec_runtime
= 0;
2950 p
->se
.prev_sum_exec_runtime
= 0;
2951 p
->se
.nr_migrations
= 0;
2953 INIT_LIST_HEAD(&p
->se
.group_node
);
2955 #ifdef CONFIG_SCHEDSTATS
2956 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2959 INIT_LIST_HEAD(&p
->rt
.run_list
);
2961 #ifdef CONFIG_PREEMPT_NOTIFIERS
2962 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2967 * fork()/clone()-time setup:
2969 void sched_fork(struct task_struct
*p
)
2971 unsigned long flags
;
2972 int cpu
= get_cpu();
2976 * We mark the process as running here. This guarantees that
2977 * nobody will actually run it, and a signal or other external
2978 * event cannot wake it up and insert it on the runqueue either.
2980 p
->state
= TASK_RUNNING
;
2983 * Make sure we do not leak PI boosting priority to the child.
2985 p
->prio
= current
->normal_prio
;
2988 * Revert to default priority/policy on fork if requested.
2990 if (unlikely(p
->sched_reset_on_fork
)) {
2991 if (task_has_rt_policy(p
)) {
2992 p
->policy
= SCHED_NORMAL
;
2993 p
->static_prio
= NICE_TO_PRIO(0);
2995 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2996 p
->static_prio
= NICE_TO_PRIO(0);
2998 p
->prio
= p
->normal_prio
= __normal_prio(p
);
3002 * We don't need the reset flag anymore after the fork. It has
3003 * fulfilled its duty:
3005 p
->sched_reset_on_fork
= 0;
3008 if (!rt_prio(p
->prio
))
3009 p
->sched_class
= &fair_sched_class
;
3011 if (p
->sched_class
->task_fork
)
3012 p
->sched_class
->task_fork(p
);
3015 * The child is not yet in the pid-hash so no cgroup attach races,
3016 * and the cgroup is pinned to this child due to cgroup_fork()
3017 * is ran before sched_fork().
3019 * Silence PROVE_RCU.
3021 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3022 set_task_cpu(p
, cpu
);
3023 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3025 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3026 if (likely(sched_info_on()))
3027 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
3029 #if defined(CONFIG_SMP)
3032 #ifdef CONFIG_PREEMPT_COUNT
3033 /* Want to start with kernel preemption disabled. */
3034 task_thread_info(p
)->preempt_count
= 1;
3037 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
3044 * wake_up_new_task - wake up a newly created task for the first time.
3046 * This function will do some initial scheduler statistics housekeeping
3047 * that must be done for every newly created context, then puts the task
3048 * on the runqueue and wakes it.
3050 void wake_up_new_task(struct task_struct
*p
)
3052 unsigned long flags
;
3055 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3058 * Fork balancing, do it here and not earlier because:
3059 * - cpus_allowed can change in the fork path
3060 * - any previously selected cpu might disappear through hotplug
3062 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
3065 rq
= __task_rq_lock(p
);
3066 activate_task(rq
, p
, 0);
3068 trace_sched_wakeup_new(p
, true);
3069 check_preempt_curr(rq
, p
, WF_FORK
);
3071 if (p
->sched_class
->task_woken
)
3072 p
->sched_class
->task_woken(rq
, p
);
3074 task_rq_unlock(rq
, p
, &flags
);
3077 #ifdef CONFIG_PREEMPT_NOTIFIERS
3080 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3081 * @notifier: notifier struct to register
3083 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3085 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3087 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3090 * preempt_notifier_unregister - no longer interested in preemption notifications
3091 * @notifier: notifier struct to unregister
3093 * This is safe to call from within a preemption notifier.
3095 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3097 hlist_del(¬ifier
->link
);
3099 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3101 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3103 struct preempt_notifier
*notifier
;
3104 struct hlist_node
*node
;
3106 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
3107 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3111 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3112 struct task_struct
*next
)
3114 struct preempt_notifier
*notifier
;
3115 struct hlist_node
*node
;
3117 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
3118 notifier
->ops
->sched_out(notifier
, next
);
3121 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3123 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3128 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3129 struct task_struct
*next
)
3133 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3136 * prepare_task_switch - prepare to switch tasks
3137 * @rq: the runqueue preparing to switch
3138 * @prev: the current task that is being switched out
3139 * @next: the task we are going to switch to.
3141 * This is called with the rq lock held and interrupts off. It must
3142 * be paired with a subsequent finish_task_switch after the context
3145 * prepare_task_switch sets up locking and calls architecture specific
3149 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3150 struct task_struct
*next
)
3152 sched_info_switch(prev
, next
);
3153 perf_event_task_sched_out(prev
, next
);
3154 fire_sched_out_preempt_notifiers(prev
, next
);
3155 prepare_lock_switch(rq
, next
);
3156 prepare_arch_switch(next
);
3157 trace_sched_switch(prev
, next
);
3161 * finish_task_switch - clean up after a task-switch
3162 * @rq: runqueue associated with task-switch
3163 * @prev: the thread we just switched away from.
3165 * finish_task_switch must be called after the context switch, paired
3166 * with a prepare_task_switch call before the context switch.
3167 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3168 * and do any other architecture-specific cleanup actions.
3170 * Note that we may have delayed dropping an mm in context_switch(). If
3171 * so, we finish that here outside of the runqueue lock. (Doing it
3172 * with the lock held can cause deadlocks; see schedule() for
3175 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
3176 __releases(rq
->lock
)
3178 struct mm_struct
*mm
= rq
->prev_mm
;
3184 * A task struct has one reference for the use as "current".
3185 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3186 * schedule one last time. The schedule call will never return, and
3187 * the scheduled task must drop that reference.
3188 * The test for TASK_DEAD must occur while the runqueue locks are
3189 * still held, otherwise prev could be scheduled on another cpu, die
3190 * there before we look at prev->state, and then the reference would
3192 * Manfred Spraul <manfred@colorfullife.com>
3194 prev_state
= prev
->state
;
3195 finish_arch_switch(prev
);
3196 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3197 local_irq_disable();
3198 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3199 perf_event_task_sched_in(prev
, current
);
3200 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3202 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3203 finish_lock_switch(rq
, prev
);
3205 fire_sched_in_preempt_notifiers(current
);
3208 if (unlikely(prev_state
== TASK_DEAD
)) {
3210 * Remove function-return probe instances associated with this
3211 * task and put them back on the free list.
3213 kprobe_flush_task(prev
);
3214 put_task_struct(prev
);
3220 /* assumes rq->lock is held */
3221 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
3223 if (prev
->sched_class
->pre_schedule
)
3224 prev
->sched_class
->pre_schedule(rq
, prev
);
3227 /* rq->lock is NOT held, but preemption is disabled */
3228 static inline void post_schedule(struct rq
*rq
)
3230 if (rq
->post_schedule
) {
3231 unsigned long flags
;
3233 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3234 if (rq
->curr
->sched_class
->post_schedule
)
3235 rq
->curr
->sched_class
->post_schedule(rq
);
3236 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3238 rq
->post_schedule
= 0;
3244 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
3248 static inline void post_schedule(struct rq
*rq
)
3255 * schedule_tail - first thing a freshly forked thread must call.
3256 * @prev: the thread we just switched away from.
3258 asmlinkage
void schedule_tail(struct task_struct
*prev
)
3259 __releases(rq
->lock
)
3261 struct rq
*rq
= this_rq();
3263 finish_task_switch(rq
, prev
);
3266 * FIXME: do we need to worry about rq being invalidated by the
3271 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3272 /* In this case, finish_task_switch does not reenable preemption */
3275 if (current
->set_child_tid
)
3276 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3280 * context_switch - switch to the new MM and the new
3281 * thread's register state.
3284 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3285 struct task_struct
*next
)
3287 struct mm_struct
*mm
, *oldmm
;
3289 prepare_task_switch(rq
, prev
, next
);
3292 oldmm
= prev
->active_mm
;
3294 * For paravirt, this is coupled with an exit in switch_to to
3295 * combine the page table reload and the switch backend into
3298 arch_start_context_switch(prev
);
3301 next
->active_mm
= oldmm
;
3302 atomic_inc(&oldmm
->mm_count
);
3303 enter_lazy_tlb(oldmm
, next
);
3305 switch_mm(oldmm
, mm
, next
);
3308 prev
->active_mm
= NULL
;
3309 rq
->prev_mm
= oldmm
;
3312 * Since the runqueue lock will be released by the next
3313 * task (which is an invalid locking op but in the case
3314 * of the scheduler it's an obvious special-case), so we
3315 * do an early lockdep release here:
3317 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3318 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3321 /* Here we just switch the register state and the stack. */
3322 switch_to(prev
, next
, prev
);
3326 * this_rq must be evaluated again because prev may have moved
3327 * CPUs since it called schedule(), thus the 'rq' on its stack
3328 * frame will be invalid.
3330 finish_task_switch(this_rq(), prev
);
3334 * nr_running, nr_uninterruptible and nr_context_switches:
3336 * externally visible scheduler statistics: current number of runnable
3337 * threads, current number of uninterruptible-sleeping threads, total
3338 * number of context switches performed since bootup.
3340 unsigned long nr_running(void)
3342 unsigned long i
, sum
= 0;
3344 for_each_online_cpu(i
)
3345 sum
+= cpu_rq(i
)->nr_running
;
3350 unsigned long nr_uninterruptible(void)
3352 unsigned long i
, sum
= 0;
3354 for_each_possible_cpu(i
)
3355 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3358 * Since we read the counters lockless, it might be slightly
3359 * inaccurate. Do not allow it to go below zero though:
3361 if (unlikely((long)sum
< 0))
3367 unsigned long long nr_context_switches(void)
3370 unsigned long long sum
= 0;
3372 for_each_possible_cpu(i
)
3373 sum
+= cpu_rq(i
)->nr_switches
;
3378 unsigned long nr_iowait(void)
3380 unsigned long i
, sum
= 0;
3382 for_each_possible_cpu(i
)
3383 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3388 unsigned long nr_iowait_cpu(int cpu
)
3390 struct rq
*this = cpu_rq(cpu
);
3391 return atomic_read(&this->nr_iowait
);
3394 unsigned long this_cpu_load(void)
3396 struct rq
*this = this_rq();
3397 return this->cpu_load
[0];
3401 /* Variables and functions for calc_load */
3402 static atomic_long_t calc_load_tasks
;
3403 static unsigned long calc_load_update
;
3404 unsigned long avenrun
[3];
3405 EXPORT_SYMBOL(avenrun
);
3407 static long calc_load_fold_active(struct rq
*this_rq
)
3409 long nr_active
, delta
= 0;
3411 nr_active
= this_rq
->nr_running
;
3412 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3414 if (nr_active
!= this_rq
->calc_load_active
) {
3415 delta
= nr_active
- this_rq
->calc_load_active
;
3416 this_rq
->calc_load_active
= nr_active
;
3422 static unsigned long
3423 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3426 load
+= active
* (FIXED_1
- exp
);
3427 load
+= 1UL << (FSHIFT
- 1);
3428 return load
>> FSHIFT
;
3433 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3435 * When making the ILB scale, we should try to pull this in as well.
3437 static atomic_long_t calc_load_tasks_idle
;
3439 static void calc_load_account_idle(struct rq
*this_rq
)
3443 delta
= calc_load_fold_active(this_rq
);
3445 atomic_long_add(delta
, &calc_load_tasks_idle
);
3448 static long calc_load_fold_idle(void)
3453 * Its got a race, we don't care...
3455 if (atomic_long_read(&calc_load_tasks_idle
))
3456 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3462 * fixed_power_int - compute: x^n, in O(log n) time
3464 * @x: base of the power
3465 * @frac_bits: fractional bits of @x
3466 * @n: power to raise @x to.
3468 * By exploiting the relation between the definition of the natural power
3469 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3470 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3471 * (where: n_i \elem {0, 1}, the binary vector representing n),
3472 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3473 * of course trivially computable in O(log_2 n), the length of our binary
3476 static unsigned long
3477 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3479 unsigned long result
= 1UL << frac_bits
;
3484 result
+= 1UL << (frac_bits
- 1);
3485 result
>>= frac_bits
;
3491 x
+= 1UL << (frac_bits
- 1);
3499 * a1 = a0 * e + a * (1 - e)
3501 * a2 = a1 * e + a * (1 - e)
3502 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3503 * = a0 * e^2 + a * (1 - e) * (1 + e)
3505 * a3 = a2 * e + a * (1 - e)
3506 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3507 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3511 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3512 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3513 * = a0 * e^n + a * (1 - e^n)
3515 * [1] application of the geometric series:
3518 * S_n := \Sum x^i = -------------
3521 static unsigned long
3522 calc_load_n(unsigned long load
, unsigned long exp
,
3523 unsigned long active
, unsigned int n
)
3526 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3530 * NO_HZ can leave us missing all per-cpu ticks calling
3531 * calc_load_account_active(), but since an idle CPU folds its delta into
3532 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3533 * in the pending idle delta if our idle period crossed a load cycle boundary.
3535 * Once we've updated the global active value, we need to apply the exponential
3536 * weights adjusted to the number of cycles missed.
3538 static void calc_global_nohz(unsigned long ticks
)
3540 long delta
, active
, n
;
3542 if (time_before(jiffies
, calc_load_update
))
3546 * If we crossed a calc_load_update boundary, make sure to fold
3547 * any pending idle changes, the respective CPUs might have
3548 * missed the tick driven calc_load_account_active() update
3551 delta
= calc_load_fold_idle();
3553 atomic_long_add(delta
, &calc_load_tasks
);
3556 * If we were idle for multiple load cycles, apply them.
3558 if (ticks
>= LOAD_FREQ
) {
3559 n
= ticks
/ LOAD_FREQ
;
3561 active
= atomic_long_read(&calc_load_tasks
);
3562 active
= active
> 0 ? active
* FIXED_1
: 0;
3564 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3565 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3566 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3568 calc_load_update
+= n
* LOAD_FREQ
;
3572 * Its possible the remainder of the above division also crosses
3573 * a LOAD_FREQ period, the regular check in calc_global_load()
3574 * which comes after this will take care of that.
3576 * Consider us being 11 ticks before a cycle completion, and us
3577 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3578 * age us 4 cycles, and the test in calc_global_load() will
3579 * pick up the final one.
3583 static void calc_load_account_idle(struct rq
*this_rq
)
3587 static inline long calc_load_fold_idle(void)
3592 static void calc_global_nohz(unsigned long ticks
)
3598 * get_avenrun - get the load average array
3599 * @loads: pointer to dest load array
3600 * @offset: offset to add
3601 * @shift: shift count to shift the result left
3603 * These values are estimates at best, so no need for locking.
3605 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3607 loads
[0] = (avenrun
[0] + offset
) << shift
;
3608 loads
[1] = (avenrun
[1] + offset
) << shift
;
3609 loads
[2] = (avenrun
[2] + offset
) << shift
;
3613 * calc_load - update the avenrun load estimates 10 ticks after the
3614 * CPUs have updated calc_load_tasks.
3616 void calc_global_load(unsigned long ticks
)
3620 calc_global_nohz(ticks
);
3622 if (time_before(jiffies
, calc_load_update
+ 10))
3625 active
= atomic_long_read(&calc_load_tasks
);
3626 active
= active
> 0 ? active
* FIXED_1
: 0;
3628 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3629 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3630 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3632 calc_load_update
+= LOAD_FREQ
;
3636 * Called from update_cpu_load() to periodically update this CPU's
3639 static void calc_load_account_active(struct rq
*this_rq
)
3643 if (time_before(jiffies
, this_rq
->calc_load_update
))
3646 delta
= calc_load_fold_active(this_rq
);
3647 delta
+= calc_load_fold_idle();
3649 atomic_long_add(delta
, &calc_load_tasks
);
3651 this_rq
->calc_load_update
+= LOAD_FREQ
;
3655 * The exact cpuload at various idx values, calculated at every tick would be
3656 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3658 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3659 * on nth tick when cpu may be busy, then we have:
3660 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3661 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3663 * decay_load_missed() below does efficient calculation of
3664 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3665 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3667 * The calculation is approximated on a 128 point scale.
3668 * degrade_zero_ticks is the number of ticks after which load at any
3669 * particular idx is approximated to be zero.
3670 * degrade_factor is a precomputed table, a row for each load idx.
3671 * Each column corresponds to degradation factor for a power of two ticks,
3672 * based on 128 point scale.
3674 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3675 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3677 * With this power of 2 load factors, we can degrade the load n times
3678 * by looking at 1 bits in n and doing as many mult/shift instead of
3679 * n mult/shifts needed by the exact degradation.
3681 #define DEGRADE_SHIFT 7
3682 static const unsigned char
3683 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3684 static const unsigned char
3685 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3686 {0, 0, 0, 0, 0, 0, 0, 0},
3687 {64, 32, 8, 0, 0, 0, 0, 0},
3688 {96, 72, 40, 12, 1, 0, 0},
3689 {112, 98, 75, 43, 15, 1, 0},
3690 {120, 112, 98, 76, 45, 16, 2} };
3693 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3694 * would be when CPU is idle and so we just decay the old load without
3695 * adding any new load.
3697 static unsigned long
3698 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3702 if (!missed_updates
)
3705 if (missed_updates
>= degrade_zero_ticks
[idx
])
3709 return load
>> missed_updates
;
3711 while (missed_updates
) {
3712 if (missed_updates
% 2)
3713 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3715 missed_updates
>>= 1;
3722 * Update rq->cpu_load[] statistics. This function is usually called every
3723 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3724 * every tick. We fix it up based on jiffies.
3726 static void update_cpu_load(struct rq
*this_rq
)
3728 unsigned long this_load
= this_rq
->load
.weight
;
3729 unsigned long curr_jiffies
= jiffies
;
3730 unsigned long pending_updates
;
3733 this_rq
->nr_load_updates
++;
3735 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3736 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3739 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3740 this_rq
->last_load_update_tick
= curr_jiffies
;
3742 /* Update our load: */
3743 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3744 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3745 unsigned long old_load
, new_load
;
3747 /* scale is effectively 1 << i now, and >> i divides by scale */
3749 old_load
= this_rq
->cpu_load
[i
];
3750 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3751 new_load
= this_load
;
3753 * Round up the averaging division if load is increasing. This
3754 * prevents us from getting stuck on 9 if the load is 10, for
3757 if (new_load
> old_load
)
3758 new_load
+= scale
- 1;
3760 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3763 sched_avg_update(this_rq
);
3766 static void update_cpu_load_active(struct rq
*this_rq
)
3768 update_cpu_load(this_rq
);
3770 calc_load_account_active(this_rq
);
3776 * sched_exec - execve() is a valuable balancing opportunity, because at
3777 * this point the task has the smallest effective memory and cache footprint.
3779 void sched_exec(void)
3781 struct task_struct
*p
= current
;
3782 unsigned long flags
;
3785 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3786 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3787 if (dest_cpu
== smp_processor_id())
3790 if (likely(cpu_active(dest_cpu
))) {
3791 struct migration_arg arg
= { p
, dest_cpu
};
3793 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3794 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3798 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3803 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3805 EXPORT_PER_CPU_SYMBOL(kstat
);
3808 * Return any ns on the sched_clock that have not yet been accounted in
3809 * @p in case that task is currently running.
3811 * Called with task_rq_lock() held on @rq.
3813 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3817 if (task_current(rq
, p
)) {
3818 update_rq_clock(rq
);
3819 ns
= rq
->clock_task
- p
->se
.exec_start
;
3827 unsigned long long task_delta_exec(struct task_struct
*p
)
3829 unsigned long flags
;
3833 rq
= task_rq_lock(p
, &flags
);
3834 ns
= do_task_delta_exec(p
, rq
);
3835 task_rq_unlock(rq
, p
, &flags
);
3841 * Return accounted runtime for the task.
3842 * In case the task is currently running, return the runtime plus current's
3843 * pending runtime that have not been accounted yet.
3845 unsigned long long task_sched_runtime(struct task_struct
*p
)
3847 unsigned long flags
;
3851 rq
= task_rq_lock(p
, &flags
);
3852 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3853 task_rq_unlock(rq
, p
, &flags
);
3859 * Account user cpu time to a process.
3860 * @p: the process that the cpu time gets accounted to
3861 * @cputime: the cpu time spent in user space since the last update
3862 * @cputime_scaled: cputime scaled by cpu frequency
3864 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3865 cputime_t cputime_scaled
)
3867 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3870 /* Add user time to process. */
3871 p
->utime
= cputime_add(p
->utime
, cputime
);
3872 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3873 account_group_user_time(p
, cputime
);
3875 /* Add user time to cpustat. */
3876 tmp
= cputime_to_cputime64(cputime
);
3877 if (TASK_NICE(p
) > 0)
3878 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3880 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3882 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3883 /* Account for user time used */
3884 acct_update_integrals(p
);
3888 * Account guest cpu time to a process.
3889 * @p: the process that the cpu time gets accounted to
3890 * @cputime: the cpu time spent in virtual machine since the last update
3891 * @cputime_scaled: cputime scaled by cpu frequency
3893 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3894 cputime_t cputime_scaled
)
3897 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3899 tmp
= cputime_to_cputime64(cputime
);
3901 /* Add guest time to process. */
3902 p
->utime
= cputime_add(p
->utime
, cputime
);
3903 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3904 account_group_user_time(p
, cputime
);
3905 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3907 /* Add guest time to cpustat. */
3908 if (TASK_NICE(p
) > 0) {
3909 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3910 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3912 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3913 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3918 * Account system cpu time to a process and desired cpustat field
3919 * @p: the process that the cpu time gets accounted to
3920 * @cputime: the cpu time spent in kernel space since the last update
3921 * @cputime_scaled: cputime scaled by cpu frequency
3922 * @target_cputime64: pointer to cpustat field that has to be updated
3925 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
3926 cputime_t cputime_scaled
, cputime64_t
*target_cputime64
)
3928 cputime64_t tmp
= cputime_to_cputime64(cputime
);
3930 /* Add system time to process. */
3931 p
->stime
= cputime_add(p
->stime
, cputime
);
3932 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3933 account_group_system_time(p
, cputime
);
3935 /* Add system time to cpustat. */
3936 *target_cputime64
= cputime64_add(*target_cputime64
, tmp
);
3937 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3939 /* Account for system time used */
3940 acct_update_integrals(p
);
3944 * Account system cpu time to a process.
3945 * @p: the process that the cpu time gets accounted to
3946 * @hardirq_offset: the offset to subtract from hardirq_count()
3947 * @cputime: the cpu time spent in kernel space since the last update
3948 * @cputime_scaled: cputime scaled by cpu frequency
3950 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3951 cputime_t cputime
, cputime_t cputime_scaled
)
3953 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3954 cputime64_t
*target_cputime64
;
3956 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3957 account_guest_time(p
, cputime
, cputime_scaled
);
3961 if (hardirq_count() - hardirq_offset
)
3962 target_cputime64
= &cpustat
->irq
;
3963 else if (in_serving_softirq())
3964 target_cputime64
= &cpustat
->softirq
;
3966 target_cputime64
= &cpustat
->system
;
3968 __account_system_time(p
, cputime
, cputime_scaled
, target_cputime64
);
3972 * Account for involuntary wait time.
3973 * @cputime: the cpu time spent in involuntary wait
3975 void account_steal_time(cputime_t cputime
)
3977 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3978 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3980 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3984 * Account for idle time.
3985 * @cputime: the cpu time spent in idle wait
3987 void account_idle_time(cputime_t cputime
)
3989 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3990 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3991 struct rq
*rq
= this_rq();
3993 if (atomic_read(&rq
->nr_iowait
) > 0)
3994 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3996 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3999 static __always_inline
bool steal_account_process_tick(void)
4001 #ifdef CONFIG_PARAVIRT
4002 if (static_branch(¶virt_steal_enabled
)) {
4005 steal
= paravirt_steal_clock(smp_processor_id());
4006 steal
-= this_rq()->prev_steal_time
;
4008 st
= steal_ticks(steal
);
4009 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
4011 account_steal_time(st
);
4018 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4020 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4022 * Account a tick to a process and cpustat
4023 * @p: the process that the cpu time gets accounted to
4024 * @user_tick: is the tick from userspace
4025 * @rq: the pointer to rq
4027 * Tick demultiplexing follows the order
4028 * - pending hardirq update
4029 * - pending softirq update
4033 * - check for guest_time
4034 * - else account as system_time
4036 * Check for hardirq is done both for system and user time as there is
4037 * no timer going off while we are on hardirq and hence we may never get an
4038 * opportunity to update it solely in system time.
4039 * p->stime and friends are only updated on system time and not on irq
4040 * softirq as those do not count in task exec_runtime any more.
4042 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
4045 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
4046 cputime64_t tmp
= cputime_to_cputime64(cputime_one_jiffy
);
4047 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4049 if (steal_account_process_tick())
4052 if (irqtime_account_hi_update()) {
4053 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4054 } else if (irqtime_account_si_update()) {
4055 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4056 } else if (this_cpu_ksoftirqd() == p
) {
4058 * ksoftirqd time do not get accounted in cpu_softirq_time.
4059 * So, we have to handle it separately here.
4060 * Also, p->stime needs to be updated for ksoftirqd.
4062 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
4064 } else if (user_tick
) {
4065 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4066 } else if (p
== rq
->idle
) {
4067 account_idle_time(cputime_one_jiffy
);
4068 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
4069 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4071 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
4076 static void irqtime_account_idle_ticks(int ticks
)
4079 struct rq
*rq
= this_rq();
4081 for (i
= 0; i
< ticks
; i
++)
4082 irqtime_account_process_tick(current
, 0, rq
);
4084 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4085 static void irqtime_account_idle_ticks(int ticks
) {}
4086 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
4088 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4091 * Account a single tick of cpu time.
4092 * @p: the process that the cpu time gets accounted to
4093 * @user_tick: indicates if the tick is a user or a system tick
4095 void account_process_tick(struct task_struct
*p
, int user_tick
)
4097 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
4098 struct rq
*rq
= this_rq();
4100 if (sched_clock_irqtime
) {
4101 irqtime_account_process_tick(p
, user_tick
, rq
);
4105 if (steal_account_process_tick())
4109 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4110 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
4111 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
4114 account_idle_time(cputime_one_jiffy
);
4118 * Account multiple ticks of steal time.
4119 * @p: the process from which the cpu time has been stolen
4120 * @ticks: number of stolen ticks
4122 void account_steal_ticks(unsigned long ticks
)
4124 account_steal_time(jiffies_to_cputime(ticks
));
4128 * Account multiple ticks of idle time.
4129 * @ticks: number of stolen ticks
4131 void account_idle_ticks(unsigned long ticks
)
4134 if (sched_clock_irqtime
) {
4135 irqtime_account_idle_ticks(ticks
);
4139 account_idle_time(jiffies_to_cputime(ticks
));
4145 * Use precise platform statistics if available:
4147 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4148 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4154 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4156 struct task_cputime cputime
;
4158 thread_group_cputime(p
, &cputime
);
4160 *ut
= cputime
.utime
;
4161 *st
= cputime
.stime
;
4165 #ifndef nsecs_to_cputime
4166 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4169 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4171 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
4174 * Use CFS's precise accounting:
4176 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
4182 do_div(temp
, total
);
4183 utime
= (cputime_t
)temp
;
4188 * Compare with previous values, to keep monotonicity:
4190 p
->prev_utime
= max(p
->prev_utime
, utime
);
4191 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
4193 *ut
= p
->prev_utime
;
4194 *st
= p
->prev_stime
;
4198 * Must be called with siglock held.
4200 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4202 struct signal_struct
*sig
= p
->signal
;
4203 struct task_cputime cputime
;
4204 cputime_t rtime
, utime
, total
;
4206 thread_group_cputime(p
, &cputime
);
4208 total
= cputime_add(cputime
.utime
, cputime
.stime
);
4209 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
4214 temp
*= cputime
.utime
;
4215 do_div(temp
, total
);
4216 utime
= (cputime_t
)temp
;
4220 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
4221 sig
->prev_stime
= max(sig
->prev_stime
,
4222 cputime_sub(rtime
, sig
->prev_utime
));
4224 *ut
= sig
->prev_utime
;
4225 *st
= sig
->prev_stime
;
4230 * This function gets called by the timer code, with HZ frequency.
4231 * We call it with interrupts disabled.
4233 void scheduler_tick(void)
4235 int cpu
= smp_processor_id();
4236 struct rq
*rq
= cpu_rq(cpu
);
4237 struct task_struct
*curr
= rq
->curr
;
4241 raw_spin_lock(&rq
->lock
);
4242 update_rq_clock(rq
);
4243 update_cpu_load_active(rq
);
4244 curr
->sched_class
->task_tick(rq
, curr
, 0);
4245 raw_spin_unlock(&rq
->lock
);
4247 perf_event_task_tick();
4250 rq
->idle_at_tick
= idle_cpu(cpu
);
4251 trigger_load_balance(rq
, cpu
);
4255 notrace
unsigned long get_parent_ip(unsigned long addr
)
4257 if (in_lock_functions(addr
)) {
4258 addr
= CALLER_ADDR2
;
4259 if (in_lock_functions(addr
))
4260 addr
= CALLER_ADDR3
;
4265 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4266 defined(CONFIG_PREEMPT_TRACER))
4268 void __kprobes
add_preempt_count(int val
)
4270 #ifdef CONFIG_DEBUG_PREEMPT
4274 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4277 preempt_count() += val
;
4278 #ifdef CONFIG_DEBUG_PREEMPT
4280 * Spinlock count overflowing soon?
4282 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4285 if (preempt_count() == val
)
4286 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4288 EXPORT_SYMBOL(add_preempt_count
);
4290 void __kprobes
sub_preempt_count(int val
)
4292 #ifdef CONFIG_DEBUG_PREEMPT
4296 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4299 * Is the spinlock portion underflowing?
4301 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4302 !(preempt_count() & PREEMPT_MASK
)))
4306 if (preempt_count() == val
)
4307 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4308 preempt_count() -= val
;
4310 EXPORT_SYMBOL(sub_preempt_count
);
4315 * Print scheduling while atomic bug:
4317 static noinline
void __schedule_bug(struct task_struct
*prev
)
4319 struct pt_regs
*regs
= get_irq_regs();
4321 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4322 prev
->comm
, prev
->pid
, preempt_count());
4324 debug_show_held_locks(prev
);
4326 if (irqs_disabled())
4327 print_irqtrace_events(prev
);
4336 * Various schedule()-time debugging checks and statistics:
4338 static inline void schedule_debug(struct task_struct
*prev
)
4341 * Test if we are atomic. Since do_exit() needs to call into
4342 * schedule() atomically, we ignore that path for now.
4343 * Otherwise, whine if we are scheduling when we should not be.
4345 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4346 __schedule_bug(prev
);
4348 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4350 schedstat_inc(this_rq(), sched_count
);
4353 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4355 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
4356 update_rq_clock(rq
);
4357 prev
->sched_class
->put_prev_task(rq
, prev
);
4361 * Pick up the highest-prio task:
4363 static inline struct task_struct
*
4364 pick_next_task(struct rq
*rq
)
4366 const struct sched_class
*class;
4367 struct task_struct
*p
;
4370 * Optimization: we know that if all tasks are in
4371 * the fair class we can call that function directly:
4373 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4374 p
= fair_sched_class
.pick_next_task(rq
);
4379 for_each_class(class) {
4380 p
= class->pick_next_task(rq
);
4385 BUG(); /* the idle class will always have a runnable task */
4389 * __schedule() is the main scheduler function.
4391 static void __sched
__schedule(void)
4393 struct task_struct
*prev
, *next
;
4394 unsigned long *switch_count
;
4400 cpu
= smp_processor_id();
4402 rcu_note_context_switch(cpu
);
4405 schedule_debug(prev
);
4407 if (sched_feat(HRTICK
))
4410 raw_spin_lock_irq(&rq
->lock
);
4412 switch_count
= &prev
->nivcsw
;
4413 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4414 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
4415 prev
->state
= TASK_RUNNING
;
4417 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
4421 * If a worker went to sleep, notify and ask workqueue
4422 * whether it wants to wake up a task to maintain
4425 if (prev
->flags
& PF_WQ_WORKER
) {
4426 struct task_struct
*to_wakeup
;
4428 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
4430 try_to_wake_up_local(to_wakeup
);
4433 switch_count
= &prev
->nvcsw
;
4436 pre_schedule(rq
, prev
);
4438 if (unlikely(!rq
->nr_running
))
4439 idle_balance(cpu
, rq
);
4441 put_prev_task(rq
, prev
);
4442 next
= pick_next_task(rq
);
4443 clear_tsk_need_resched(prev
);
4444 rq
->skip_clock_update
= 0;
4446 if (likely(prev
!= next
)) {
4451 context_switch(rq
, prev
, next
); /* unlocks the rq */
4453 * The context switch have flipped the stack from under us
4454 * and restored the local variables which were saved when
4455 * this task called schedule() in the past. prev == current
4456 * is still correct, but it can be moved to another cpu/rq.
4458 cpu
= smp_processor_id();
4461 raw_spin_unlock_irq(&rq
->lock
);
4465 preempt_enable_no_resched();
4470 static inline void sched_submit_work(struct task_struct
*tsk
)
4475 * If we are going to sleep and we have plugged IO queued,
4476 * make sure to submit it to avoid deadlocks.
4478 if (blk_needs_flush_plug(tsk
))
4479 blk_schedule_flush_plug(tsk
);
4482 asmlinkage
void __sched
schedule(void)
4484 struct task_struct
*tsk
= current
;
4486 sched_submit_work(tsk
);
4489 EXPORT_SYMBOL(schedule
);
4491 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4493 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
4495 if (lock
->owner
!= owner
)
4499 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4500 * lock->owner still matches owner, if that fails, owner might
4501 * point to free()d memory, if it still matches, the rcu_read_lock()
4502 * ensures the memory stays valid.
4506 return owner
->on_cpu
;
4510 * Look out! "owner" is an entirely speculative pointer
4511 * access and not reliable.
4513 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
4515 if (!sched_feat(OWNER_SPIN
))
4519 while (owner_running(lock
, owner
)) {
4523 arch_mutex_cpu_relax();
4528 * We break out the loop above on need_resched() and when the
4529 * owner changed, which is a sign for heavy contention. Return
4530 * success only when lock->owner is NULL.
4532 return lock
->owner
== NULL
;
4536 #ifdef CONFIG_PREEMPT
4538 * this is the entry point to schedule() from in-kernel preemption
4539 * off of preempt_enable. Kernel preemptions off return from interrupt
4540 * occur there and call schedule directly.
4542 asmlinkage
void __sched notrace
preempt_schedule(void)
4544 struct thread_info
*ti
= current_thread_info();
4547 * If there is a non-zero preempt_count or interrupts are disabled,
4548 * we do not want to preempt the current task. Just return..
4550 if (likely(ti
->preempt_count
|| irqs_disabled()))
4554 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4556 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4559 * Check again in case we missed a preemption opportunity
4560 * between schedule and now.
4563 } while (need_resched());
4565 EXPORT_SYMBOL(preempt_schedule
);
4568 * this is the entry point to schedule() from kernel preemption
4569 * off of irq context.
4570 * Note, that this is called and return with irqs disabled. This will
4571 * protect us against recursive calling from irq.
4573 asmlinkage
void __sched
preempt_schedule_irq(void)
4575 struct thread_info
*ti
= current_thread_info();
4577 /* Catch callers which need to be fixed */
4578 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4581 add_preempt_count(PREEMPT_ACTIVE
);
4584 local_irq_disable();
4585 sub_preempt_count(PREEMPT_ACTIVE
);
4588 * Check again in case we missed a preemption opportunity
4589 * between schedule and now.
4592 } while (need_resched());
4595 #endif /* CONFIG_PREEMPT */
4597 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4600 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4602 EXPORT_SYMBOL(default_wake_function
);
4605 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4606 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4607 * number) then we wake all the non-exclusive tasks and one exclusive task.
4609 * There are circumstances in which we can try to wake a task which has already
4610 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4611 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4613 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4614 int nr_exclusive
, int wake_flags
, void *key
)
4616 wait_queue_t
*curr
, *next
;
4618 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4619 unsigned flags
= curr
->flags
;
4621 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4622 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4628 * __wake_up - wake up threads blocked on a waitqueue.
4630 * @mode: which threads
4631 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4632 * @key: is directly passed to the wakeup function
4634 * It may be assumed that this function implies a write memory barrier before
4635 * changing the task state if and only if any tasks are woken up.
4637 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4638 int nr_exclusive
, void *key
)
4640 unsigned long flags
;
4642 spin_lock_irqsave(&q
->lock
, flags
);
4643 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4644 spin_unlock_irqrestore(&q
->lock
, flags
);
4646 EXPORT_SYMBOL(__wake_up
);
4649 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4651 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4653 __wake_up_common(q
, mode
, 1, 0, NULL
);
4655 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4657 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4659 __wake_up_common(q
, mode
, 1, 0, key
);
4661 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
4664 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4666 * @mode: which threads
4667 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4668 * @key: opaque value to be passed to wakeup targets
4670 * The sync wakeup differs that the waker knows that it will schedule
4671 * away soon, so while the target thread will be woken up, it will not
4672 * be migrated to another CPU - ie. the two threads are 'synchronized'
4673 * with each other. This can prevent needless bouncing between CPUs.
4675 * On UP it can prevent extra preemption.
4677 * It may be assumed that this function implies a write memory barrier before
4678 * changing the task state if and only if any tasks are woken up.
4680 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4681 int nr_exclusive
, void *key
)
4683 unsigned long flags
;
4684 int wake_flags
= WF_SYNC
;
4689 if (unlikely(!nr_exclusive
))
4692 spin_lock_irqsave(&q
->lock
, flags
);
4693 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4694 spin_unlock_irqrestore(&q
->lock
, flags
);
4696 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4699 * __wake_up_sync - see __wake_up_sync_key()
4701 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4703 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4705 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4708 * complete: - signals a single thread waiting on this completion
4709 * @x: holds the state of this particular completion
4711 * This will wake up a single thread waiting on this completion. Threads will be
4712 * awakened in the same order in which they were queued.
4714 * See also complete_all(), wait_for_completion() and related routines.
4716 * It may be assumed that this function implies a write memory barrier before
4717 * changing the task state if and only if any tasks are woken up.
4719 void complete(struct completion
*x
)
4721 unsigned long flags
;
4723 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4725 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4726 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4728 EXPORT_SYMBOL(complete
);
4731 * complete_all: - signals all threads waiting on this completion
4732 * @x: holds the state of this particular completion
4734 * This will wake up all threads waiting on this particular completion event.
4736 * It may be assumed that this function implies a write memory barrier before
4737 * changing the task state if and only if any tasks are woken up.
4739 void complete_all(struct completion
*x
)
4741 unsigned long flags
;
4743 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4744 x
->done
+= UINT_MAX
/2;
4745 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4746 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4748 EXPORT_SYMBOL(complete_all
);
4750 static inline long __sched
4751 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4754 DECLARE_WAITQUEUE(wait
, current
);
4756 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4758 if (signal_pending_state(state
, current
)) {
4759 timeout
= -ERESTARTSYS
;
4762 __set_current_state(state
);
4763 spin_unlock_irq(&x
->wait
.lock
);
4764 timeout
= schedule_timeout(timeout
);
4765 spin_lock_irq(&x
->wait
.lock
);
4766 } while (!x
->done
&& timeout
);
4767 __remove_wait_queue(&x
->wait
, &wait
);
4772 return timeout
?: 1;
4776 wait_for_common(struct completion
*x
, long timeout
, int state
)
4780 spin_lock_irq(&x
->wait
.lock
);
4781 timeout
= do_wait_for_common(x
, timeout
, state
);
4782 spin_unlock_irq(&x
->wait
.lock
);
4787 * wait_for_completion: - waits for completion of a task
4788 * @x: holds the state of this particular completion
4790 * This waits to be signaled for completion of a specific task. It is NOT
4791 * interruptible and there is no timeout.
4793 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4794 * and interrupt capability. Also see complete().
4796 void __sched
wait_for_completion(struct completion
*x
)
4798 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4800 EXPORT_SYMBOL(wait_for_completion
);
4803 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4804 * @x: holds the state of this particular completion
4805 * @timeout: timeout value in jiffies
4807 * This waits for either a completion of a specific task to be signaled or for a
4808 * specified timeout to expire. The timeout is in jiffies. It is not
4811 unsigned long __sched
4812 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4814 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4816 EXPORT_SYMBOL(wait_for_completion_timeout
);
4819 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4820 * @x: holds the state of this particular completion
4822 * This waits for completion of a specific task to be signaled. It is
4825 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4827 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4828 if (t
== -ERESTARTSYS
)
4832 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4835 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4836 * @x: holds the state of this particular completion
4837 * @timeout: timeout value in jiffies
4839 * This waits for either a completion of a specific task to be signaled or for a
4840 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4843 wait_for_completion_interruptible_timeout(struct completion
*x
,
4844 unsigned long timeout
)
4846 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4848 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4851 * wait_for_completion_killable: - waits for completion of a task (killable)
4852 * @x: holds the state of this particular completion
4854 * This waits to be signaled for completion of a specific task. It can be
4855 * interrupted by a kill signal.
4857 int __sched
wait_for_completion_killable(struct completion
*x
)
4859 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4860 if (t
== -ERESTARTSYS
)
4864 EXPORT_SYMBOL(wait_for_completion_killable
);
4867 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4868 * @x: holds the state of this particular completion
4869 * @timeout: timeout value in jiffies
4871 * This waits for either a completion of a specific task to be
4872 * signaled or for a specified timeout to expire. It can be
4873 * interrupted by a kill signal. The timeout is in jiffies.
4876 wait_for_completion_killable_timeout(struct completion
*x
,
4877 unsigned long timeout
)
4879 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4881 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4884 * try_wait_for_completion - try to decrement a completion without blocking
4885 * @x: completion structure
4887 * Returns: 0 if a decrement cannot be done without blocking
4888 * 1 if a decrement succeeded.
4890 * If a completion is being used as a counting completion,
4891 * attempt to decrement the counter without blocking. This
4892 * enables us to avoid waiting if the resource the completion
4893 * is protecting is not available.
4895 bool try_wait_for_completion(struct completion
*x
)
4897 unsigned long flags
;
4900 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4905 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4908 EXPORT_SYMBOL(try_wait_for_completion
);
4911 * completion_done - Test to see if a completion has any waiters
4912 * @x: completion structure
4914 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4915 * 1 if there are no waiters.
4918 bool completion_done(struct completion
*x
)
4920 unsigned long flags
;
4923 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4926 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4929 EXPORT_SYMBOL(completion_done
);
4932 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4934 unsigned long flags
;
4937 init_waitqueue_entry(&wait
, current
);
4939 __set_current_state(state
);
4941 spin_lock_irqsave(&q
->lock
, flags
);
4942 __add_wait_queue(q
, &wait
);
4943 spin_unlock(&q
->lock
);
4944 timeout
= schedule_timeout(timeout
);
4945 spin_lock_irq(&q
->lock
);
4946 __remove_wait_queue(q
, &wait
);
4947 spin_unlock_irqrestore(&q
->lock
, flags
);
4952 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4954 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4956 EXPORT_SYMBOL(interruptible_sleep_on
);
4959 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4961 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4963 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4965 void __sched
sleep_on(wait_queue_head_t
*q
)
4967 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4969 EXPORT_SYMBOL(sleep_on
);
4971 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4973 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4975 EXPORT_SYMBOL(sleep_on_timeout
);
4977 #ifdef CONFIG_RT_MUTEXES
4980 * rt_mutex_setprio - set the current priority of a task
4982 * @prio: prio value (kernel-internal form)
4984 * This function changes the 'effective' priority of a task. It does
4985 * not touch ->normal_prio like __setscheduler().
4987 * Used by the rt_mutex code to implement priority inheritance logic.
4989 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4991 int oldprio
, on_rq
, running
;
4993 const struct sched_class
*prev_class
;
4995 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4997 rq
= __task_rq_lock(p
);
4999 trace_sched_pi_setprio(p
, prio
);
5001 prev_class
= p
->sched_class
;
5003 running
= task_current(rq
, p
);
5005 dequeue_task(rq
, p
, 0);
5007 p
->sched_class
->put_prev_task(rq
, p
);
5010 p
->sched_class
= &rt_sched_class
;
5012 p
->sched_class
= &fair_sched_class
;
5017 p
->sched_class
->set_curr_task(rq
);
5019 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
5021 check_class_changed(rq
, p
, prev_class
, oldprio
);
5022 __task_rq_unlock(rq
);
5027 void set_user_nice(struct task_struct
*p
, long nice
)
5029 int old_prio
, delta
, on_rq
;
5030 unsigned long flags
;
5033 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5036 * We have to be careful, if called from sys_setpriority(),
5037 * the task might be in the middle of scheduling on another CPU.
5039 rq
= task_rq_lock(p
, &flags
);
5041 * The RT priorities are set via sched_setscheduler(), but we still
5042 * allow the 'normal' nice value to be set - but as expected
5043 * it wont have any effect on scheduling until the task is
5044 * SCHED_FIFO/SCHED_RR:
5046 if (task_has_rt_policy(p
)) {
5047 p
->static_prio
= NICE_TO_PRIO(nice
);
5052 dequeue_task(rq
, p
, 0);
5054 p
->static_prio
= NICE_TO_PRIO(nice
);
5057 p
->prio
= effective_prio(p
);
5058 delta
= p
->prio
- old_prio
;
5061 enqueue_task(rq
, p
, 0);
5063 * If the task increased its priority or is running and
5064 * lowered its priority, then reschedule its CPU:
5066 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5067 resched_task(rq
->curr
);
5070 task_rq_unlock(rq
, p
, &flags
);
5072 EXPORT_SYMBOL(set_user_nice
);
5075 * can_nice - check if a task can reduce its nice value
5079 int can_nice(const struct task_struct
*p
, const int nice
)
5081 /* convert nice value [19,-20] to rlimit style value [1,40] */
5082 int nice_rlim
= 20 - nice
;
5084 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
5085 capable(CAP_SYS_NICE
));
5088 #ifdef __ARCH_WANT_SYS_NICE
5091 * sys_nice - change the priority of the current process.
5092 * @increment: priority increment
5094 * sys_setpriority is a more generic, but much slower function that
5095 * does similar things.
5097 SYSCALL_DEFINE1(nice
, int, increment
)
5102 * Setpriority might change our priority at the same moment.
5103 * We don't have to worry. Conceptually one call occurs first
5104 * and we have a single winner.
5106 if (increment
< -40)
5111 nice
= TASK_NICE(current
) + increment
;
5117 if (increment
< 0 && !can_nice(current
, nice
))
5120 retval
= security_task_setnice(current
, nice
);
5124 set_user_nice(current
, nice
);
5131 * task_prio - return the priority value of a given task.
5132 * @p: the task in question.
5134 * This is the priority value as seen by users in /proc.
5135 * RT tasks are offset by -200. Normal tasks are centered
5136 * around 0, value goes from -16 to +15.
5138 int task_prio(const struct task_struct
*p
)
5140 return p
->prio
- MAX_RT_PRIO
;
5144 * task_nice - return the nice value of a given task.
5145 * @p: the task in question.
5147 int task_nice(const struct task_struct
*p
)
5149 return TASK_NICE(p
);
5151 EXPORT_SYMBOL(task_nice
);
5154 * idle_cpu - is a given cpu idle currently?
5155 * @cpu: the processor in question.
5157 int idle_cpu(int cpu
)
5159 struct rq
*rq
= cpu_rq(cpu
);
5161 if (rq
->curr
!= rq
->idle
)
5168 if (!llist_empty(&rq
->wake_list
))
5176 * idle_task - return the idle task for a given cpu.
5177 * @cpu: the processor in question.
5179 struct task_struct
*idle_task(int cpu
)
5181 return cpu_rq(cpu
)->idle
;
5185 * find_process_by_pid - find a process with a matching PID value.
5186 * @pid: the pid in question.
5188 static struct task_struct
*find_process_by_pid(pid_t pid
)
5190 return pid
? find_task_by_vpid(pid
) : current
;
5193 /* Actually do priority change: must hold rq lock. */
5195 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5198 p
->rt_priority
= prio
;
5199 p
->normal_prio
= normal_prio(p
);
5200 /* we are holding p->pi_lock already */
5201 p
->prio
= rt_mutex_getprio(p
);
5202 if (rt_prio(p
->prio
))
5203 p
->sched_class
= &rt_sched_class
;
5205 p
->sched_class
= &fair_sched_class
;
5210 * check the target process has a UID that matches the current process's
5212 static bool check_same_owner(struct task_struct
*p
)
5214 const struct cred
*cred
= current_cred(), *pcred
;
5218 pcred
= __task_cred(p
);
5219 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
5220 match
= (cred
->euid
== pcred
->euid
||
5221 cred
->euid
== pcred
->uid
);
5228 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5229 const struct sched_param
*param
, bool user
)
5231 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5232 unsigned long flags
;
5233 const struct sched_class
*prev_class
;
5237 /* may grab non-irq protected spin_locks */
5238 BUG_ON(in_interrupt());
5240 /* double check policy once rq lock held */
5242 reset_on_fork
= p
->sched_reset_on_fork
;
5243 policy
= oldpolicy
= p
->policy
;
5245 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
5246 policy
&= ~SCHED_RESET_ON_FORK
;
5248 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5249 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5250 policy
!= SCHED_IDLE
)
5255 * Valid priorities for SCHED_FIFO and SCHED_RR are
5256 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5257 * SCHED_BATCH and SCHED_IDLE is 0.
5259 if (param
->sched_priority
< 0 ||
5260 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5261 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5263 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5267 * Allow unprivileged RT tasks to decrease priority:
5269 if (user
&& !capable(CAP_SYS_NICE
)) {
5270 if (rt_policy(policy
)) {
5271 unsigned long rlim_rtprio
=
5272 task_rlimit(p
, RLIMIT_RTPRIO
);
5274 /* can't set/change the rt policy */
5275 if (policy
!= p
->policy
&& !rlim_rtprio
)
5278 /* can't increase priority */
5279 if (param
->sched_priority
> p
->rt_priority
&&
5280 param
->sched_priority
> rlim_rtprio
)
5285 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5286 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5288 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
5289 if (!can_nice(p
, TASK_NICE(p
)))
5293 /* can't change other user's priorities */
5294 if (!check_same_owner(p
))
5297 /* Normal users shall not reset the sched_reset_on_fork flag */
5298 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5303 retval
= security_task_setscheduler(p
);
5309 * make sure no PI-waiters arrive (or leave) while we are
5310 * changing the priority of the task:
5312 * To be able to change p->policy safely, the appropriate
5313 * runqueue lock must be held.
5315 rq
= task_rq_lock(p
, &flags
);
5318 * Changing the policy of the stop threads its a very bad idea
5320 if (p
== rq
->stop
) {
5321 task_rq_unlock(rq
, p
, &flags
);
5326 * If not changing anything there's no need to proceed further:
5328 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
5329 param
->sched_priority
== p
->rt_priority
))) {
5331 __task_rq_unlock(rq
);
5332 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5336 #ifdef CONFIG_RT_GROUP_SCHED
5339 * Do not allow realtime tasks into groups that have no runtime
5342 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5343 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5344 !task_group_is_autogroup(task_group(p
))) {
5345 task_rq_unlock(rq
, p
, &flags
);
5351 /* recheck policy now with rq lock held */
5352 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5353 policy
= oldpolicy
= -1;
5354 task_rq_unlock(rq
, p
, &flags
);
5358 running
= task_current(rq
, p
);
5360 deactivate_task(rq
, p
, 0);
5362 p
->sched_class
->put_prev_task(rq
, p
);
5364 p
->sched_reset_on_fork
= reset_on_fork
;
5367 prev_class
= p
->sched_class
;
5368 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5371 p
->sched_class
->set_curr_task(rq
);
5373 activate_task(rq
, p
, 0);
5375 check_class_changed(rq
, p
, prev_class
, oldprio
);
5376 task_rq_unlock(rq
, p
, &flags
);
5378 rt_mutex_adjust_pi(p
);
5384 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5385 * @p: the task in question.
5386 * @policy: new policy.
5387 * @param: structure containing the new RT priority.
5389 * NOTE that the task may be already dead.
5391 int sched_setscheduler(struct task_struct
*p
, int policy
,
5392 const struct sched_param
*param
)
5394 return __sched_setscheduler(p
, policy
, param
, true);
5396 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5399 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5400 * @p: the task in question.
5401 * @policy: new policy.
5402 * @param: structure containing the new RT priority.
5404 * Just like sched_setscheduler, only don't bother checking if the
5405 * current context has permission. For example, this is needed in
5406 * stop_machine(): we create temporary high priority worker threads,
5407 * but our caller might not have that capability.
5409 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5410 const struct sched_param
*param
)
5412 return __sched_setscheduler(p
, policy
, param
, false);
5416 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5418 struct sched_param lparam
;
5419 struct task_struct
*p
;
5422 if (!param
|| pid
< 0)
5424 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5429 p
= find_process_by_pid(pid
);
5431 retval
= sched_setscheduler(p
, policy
, &lparam
);
5438 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5439 * @pid: the pid in question.
5440 * @policy: new policy.
5441 * @param: structure containing the new RT priority.
5443 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5444 struct sched_param __user
*, param
)
5446 /* negative values for policy are not valid */
5450 return do_sched_setscheduler(pid
, policy
, param
);
5454 * sys_sched_setparam - set/change the RT priority of a thread
5455 * @pid: the pid in question.
5456 * @param: structure containing the new RT priority.
5458 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5460 return do_sched_setscheduler(pid
, -1, param
);
5464 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5465 * @pid: the pid in question.
5467 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5469 struct task_struct
*p
;
5477 p
= find_process_by_pid(pid
);
5479 retval
= security_task_getscheduler(p
);
5482 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5489 * sys_sched_getparam - get the RT priority of a thread
5490 * @pid: the pid in question.
5491 * @param: structure containing the RT priority.
5493 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5495 struct sched_param lp
;
5496 struct task_struct
*p
;
5499 if (!param
|| pid
< 0)
5503 p
= find_process_by_pid(pid
);
5508 retval
= security_task_getscheduler(p
);
5512 lp
.sched_priority
= p
->rt_priority
;
5516 * This one might sleep, we cannot do it with a spinlock held ...
5518 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5527 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5529 cpumask_var_t cpus_allowed
, new_mask
;
5530 struct task_struct
*p
;
5536 p
= find_process_by_pid(pid
);
5543 /* Prevent p going away */
5547 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5551 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5553 goto out_free_cpus_allowed
;
5556 if (!check_same_owner(p
) && !task_ns_capable(p
, CAP_SYS_NICE
))
5559 retval
= security_task_setscheduler(p
);
5563 cpuset_cpus_allowed(p
, cpus_allowed
);
5564 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5566 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5569 cpuset_cpus_allowed(p
, cpus_allowed
);
5570 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5572 * We must have raced with a concurrent cpuset
5573 * update. Just reset the cpus_allowed to the
5574 * cpuset's cpus_allowed
5576 cpumask_copy(new_mask
, cpus_allowed
);
5581 free_cpumask_var(new_mask
);
5582 out_free_cpus_allowed
:
5583 free_cpumask_var(cpus_allowed
);
5590 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5591 struct cpumask
*new_mask
)
5593 if (len
< cpumask_size())
5594 cpumask_clear(new_mask
);
5595 else if (len
> cpumask_size())
5596 len
= cpumask_size();
5598 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5602 * sys_sched_setaffinity - set the cpu affinity of a process
5603 * @pid: pid of the process
5604 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5605 * @user_mask_ptr: user-space pointer to the new cpu mask
5607 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5608 unsigned long __user
*, user_mask_ptr
)
5610 cpumask_var_t new_mask
;
5613 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5616 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5618 retval
= sched_setaffinity(pid
, new_mask
);
5619 free_cpumask_var(new_mask
);
5623 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5625 struct task_struct
*p
;
5626 unsigned long flags
;
5633 p
= find_process_by_pid(pid
);
5637 retval
= security_task_getscheduler(p
);
5641 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5642 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5643 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5653 * sys_sched_getaffinity - get the cpu affinity of a process
5654 * @pid: pid of the process
5655 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5656 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5658 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5659 unsigned long __user
*, user_mask_ptr
)
5664 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5666 if (len
& (sizeof(unsigned long)-1))
5669 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5672 ret
= sched_getaffinity(pid
, mask
);
5674 size_t retlen
= min_t(size_t, len
, cpumask_size());
5676 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5681 free_cpumask_var(mask
);
5687 * sys_sched_yield - yield the current processor to other threads.
5689 * This function yields the current CPU to other tasks. If there are no
5690 * other threads running on this CPU then this function will return.
5692 SYSCALL_DEFINE0(sched_yield
)
5694 struct rq
*rq
= this_rq_lock();
5696 schedstat_inc(rq
, yld_count
);
5697 current
->sched_class
->yield_task(rq
);
5700 * Since we are going to call schedule() anyway, there's
5701 * no need to preempt or enable interrupts:
5703 __release(rq
->lock
);
5704 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5705 do_raw_spin_unlock(&rq
->lock
);
5706 preempt_enable_no_resched();
5713 static inline int should_resched(void)
5715 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5718 static void __cond_resched(void)
5720 add_preempt_count(PREEMPT_ACTIVE
);
5722 sub_preempt_count(PREEMPT_ACTIVE
);
5725 int __sched
_cond_resched(void)
5727 if (should_resched()) {
5733 EXPORT_SYMBOL(_cond_resched
);
5736 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5737 * call schedule, and on return reacquire the lock.
5739 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5740 * operations here to prevent schedule() from being called twice (once via
5741 * spin_unlock(), once by hand).
5743 int __cond_resched_lock(spinlock_t
*lock
)
5745 int resched
= should_resched();
5748 lockdep_assert_held(lock
);
5750 if (spin_needbreak(lock
) || resched
) {
5761 EXPORT_SYMBOL(__cond_resched_lock
);
5763 int __sched
__cond_resched_softirq(void)
5765 BUG_ON(!in_softirq());
5767 if (should_resched()) {
5775 EXPORT_SYMBOL(__cond_resched_softirq
);
5778 * yield - yield the current processor to other threads.
5780 * This is a shortcut for kernel-space yielding - it marks the
5781 * thread runnable and calls sys_sched_yield().
5783 void __sched
yield(void)
5785 set_current_state(TASK_RUNNING
);
5788 EXPORT_SYMBOL(yield
);
5791 * yield_to - yield the current processor to another thread in
5792 * your thread group, or accelerate that thread toward the
5793 * processor it's on.
5795 * @preempt: whether task preemption is allowed or not
5797 * It's the caller's job to ensure that the target task struct
5798 * can't go away on us before we can do any checks.
5800 * Returns true if we indeed boosted the target task.
5802 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
5804 struct task_struct
*curr
= current
;
5805 struct rq
*rq
, *p_rq
;
5806 unsigned long flags
;
5809 local_irq_save(flags
);
5814 double_rq_lock(rq
, p_rq
);
5815 while (task_rq(p
) != p_rq
) {
5816 double_rq_unlock(rq
, p_rq
);
5820 if (!curr
->sched_class
->yield_to_task
)
5823 if (curr
->sched_class
!= p
->sched_class
)
5826 if (task_running(p_rq
, p
) || p
->state
)
5829 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5831 schedstat_inc(rq
, yld_count
);
5833 * Make p's CPU reschedule; pick_next_entity takes care of
5836 if (preempt
&& rq
!= p_rq
)
5837 resched_task(p_rq
->curr
);
5841 double_rq_unlock(rq
, p_rq
);
5842 local_irq_restore(flags
);
5849 EXPORT_SYMBOL_GPL(yield_to
);
5852 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5853 * that process accounting knows that this is a task in IO wait state.
5855 void __sched
io_schedule(void)
5857 struct rq
*rq
= raw_rq();
5859 delayacct_blkio_start();
5860 atomic_inc(&rq
->nr_iowait
);
5861 blk_flush_plug(current
);
5862 current
->in_iowait
= 1;
5864 current
->in_iowait
= 0;
5865 atomic_dec(&rq
->nr_iowait
);
5866 delayacct_blkio_end();
5868 EXPORT_SYMBOL(io_schedule
);
5870 long __sched
io_schedule_timeout(long timeout
)
5872 struct rq
*rq
= raw_rq();
5875 delayacct_blkio_start();
5876 atomic_inc(&rq
->nr_iowait
);
5877 blk_flush_plug(current
);
5878 current
->in_iowait
= 1;
5879 ret
= schedule_timeout(timeout
);
5880 current
->in_iowait
= 0;
5881 atomic_dec(&rq
->nr_iowait
);
5882 delayacct_blkio_end();
5887 * sys_sched_get_priority_max - return maximum RT priority.
5888 * @policy: scheduling class.
5890 * this syscall returns the maximum rt_priority that can be used
5891 * by a given scheduling class.
5893 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5900 ret
= MAX_USER_RT_PRIO
-1;
5912 * sys_sched_get_priority_min - return minimum RT priority.
5913 * @policy: scheduling class.
5915 * this syscall returns the minimum rt_priority that can be used
5916 * by a given scheduling class.
5918 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5936 * sys_sched_rr_get_interval - return the default timeslice of a process.
5937 * @pid: pid of the process.
5938 * @interval: userspace pointer to the timeslice value.
5940 * this syscall writes the default timeslice value of a given process
5941 * into the user-space timespec buffer. A value of '0' means infinity.
5943 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5944 struct timespec __user
*, interval
)
5946 struct task_struct
*p
;
5947 unsigned int time_slice
;
5948 unsigned long flags
;
5958 p
= find_process_by_pid(pid
);
5962 retval
= security_task_getscheduler(p
);
5966 rq
= task_rq_lock(p
, &flags
);
5967 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5968 task_rq_unlock(rq
, p
, &flags
);
5971 jiffies_to_timespec(time_slice
, &t
);
5972 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5980 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5982 void sched_show_task(struct task_struct
*p
)
5984 unsigned long free
= 0;
5987 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5988 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5989 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5990 #if BITS_PER_LONG == 32
5991 if (state
== TASK_RUNNING
)
5992 printk(KERN_CONT
" running ");
5994 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5996 if (state
== TASK_RUNNING
)
5997 printk(KERN_CONT
" running task ");
5999 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6001 #ifdef CONFIG_DEBUG_STACK_USAGE
6002 free
= stack_not_used(p
);
6004 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6005 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6006 (unsigned long)task_thread_info(p
)->flags
);
6008 show_stack(p
, NULL
);
6011 void show_state_filter(unsigned long state_filter
)
6013 struct task_struct
*g
, *p
;
6015 #if BITS_PER_LONG == 32
6017 " task PC stack pid father\n");
6020 " task PC stack pid father\n");
6022 read_lock(&tasklist_lock
);
6023 do_each_thread(g
, p
) {
6025 * reset the NMI-timeout, listing all files on a slow
6026 * console might take a lot of time:
6028 touch_nmi_watchdog();
6029 if (!state_filter
|| (p
->state
& state_filter
))
6031 } while_each_thread(g
, p
);
6033 touch_all_softlockup_watchdogs();
6035 #ifdef CONFIG_SCHED_DEBUG
6036 sysrq_sched_debug_show();
6038 read_unlock(&tasklist_lock
);
6040 * Only show locks if all tasks are dumped:
6043 debug_show_all_locks();
6046 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6048 idle
->sched_class
= &idle_sched_class
;
6052 * init_idle - set up an idle thread for a given CPU
6053 * @idle: task in question
6054 * @cpu: cpu the idle task belongs to
6056 * NOTE: this function does not set the idle thread's NEED_RESCHED
6057 * flag, to make booting more robust.
6059 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6061 struct rq
*rq
= cpu_rq(cpu
);
6062 unsigned long flags
;
6064 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6067 idle
->state
= TASK_RUNNING
;
6068 idle
->se
.exec_start
= sched_clock();
6070 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
6072 * We're having a chicken and egg problem, even though we are
6073 * holding rq->lock, the cpu isn't yet set to this cpu so the
6074 * lockdep check in task_group() will fail.
6076 * Similar case to sched_fork(). / Alternatively we could
6077 * use task_rq_lock() here and obtain the other rq->lock.
6082 __set_task_cpu(idle
, cpu
);
6085 rq
->curr
= rq
->idle
= idle
;
6086 #if defined(CONFIG_SMP)
6089 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6091 /* Set the preempt count _outside_ the spinlocks! */
6092 task_thread_info(idle
)->preempt_count
= 0;
6095 * The idle tasks have their own, simple scheduling class:
6097 idle
->sched_class
= &idle_sched_class
;
6098 ftrace_graph_init_idle_task(idle
, cpu
);
6102 * In a system that switches off the HZ timer nohz_cpu_mask
6103 * indicates which cpus entered this state. This is used
6104 * in the rcu update to wait only for active cpus. For system
6105 * which do not switch off the HZ timer nohz_cpu_mask should
6106 * always be CPU_BITS_NONE.
6108 cpumask_var_t nohz_cpu_mask
;
6111 * Increase the granularity value when there are more CPUs,
6112 * because with more CPUs the 'effective latency' as visible
6113 * to users decreases. But the relationship is not linear,
6114 * so pick a second-best guess by going with the log2 of the
6117 * This idea comes from the SD scheduler of Con Kolivas:
6119 static int get_update_sysctl_factor(void)
6121 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
6122 unsigned int factor
;
6124 switch (sysctl_sched_tunable_scaling
) {
6125 case SCHED_TUNABLESCALING_NONE
:
6128 case SCHED_TUNABLESCALING_LINEAR
:
6131 case SCHED_TUNABLESCALING_LOG
:
6133 factor
= 1 + ilog2(cpus
);
6140 static void update_sysctl(void)
6142 unsigned int factor
= get_update_sysctl_factor();
6144 #define SET_SYSCTL(name) \
6145 (sysctl_##name = (factor) * normalized_sysctl_##name)
6146 SET_SYSCTL(sched_min_granularity
);
6147 SET_SYSCTL(sched_latency
);
6148 SET_SYSCTL(sched_wakeup_granularity
);
6152 static inline void sched_init_granularity(void)
6158 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
6160 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
6161 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6163 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6164 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6169 * This is how migration works:
6171 * 1) we invoke migration_cpu_stop() on the target CPU using
6173 * 2) stopper starts to run (implicitly forcing the migrated thread
6175 * 3) it checks whether the migrated task is still in the wrong runqueue.
6176 * 4) if it's in the wrong runqueue then the migration thread removes
6177 * it and puts it into the right queue.
6178 * 5) stopper completes and stop_one_cpu() returns and the migration
6183 * Change a given task's CPU affinity. Migrate the thread to a
6184 * proper CPU and schedule it away if the CPU it's executing on
6185 * is removed from the allowed bitmask.
6187 * NOTE: the caller must have a valid reference to the task, the
6188 * task must not exit() & deallocate itself prematurely. The
6189 * call is not atomic; no spinlocks may be held.
6191 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6193 unsigned long flags
;
6195 unsigned int dest_cpu
;
6198 rq
= task_rq_lock(p
, &flags
);
6200 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
6203 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
6208 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
6213 do_set_cpus_allowed(p
, new_mask
);
6215 /* Can the task run on the task's current CPU? If so, we're done */
6216 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6219 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
6221 struct migration_arg arg
= { p
, dest_cpu
};
6222 /* Need help from migration thread: drop lock and wait. */
6223 task_rq_unlock(rq
, p
, &flags
);
6224 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
6225 tlb_migrate_finish(p
->mm
);
6229 task_rq_unlock(rq
, p
, &flags
);
6233 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6236 * Move (not current) task off this cpu, onto dest cpu. We're doing
6237 * this because either it can't run here any more (set_cpus_allowed()
6238 * away from this CPU, or CPU going down), or because we're
6239 * attempting to rebalance this task on exec (sched_exec).
6241 * So we race with normal scheduler movements, but that's OK, as long
6242 * as the task is no longer on this CPU.
6244 * Returns non-zero if task was successfully migrated.
6246 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6248 struct rq
*rq_dest
, *rq_src
;
6251 if (unlikely(!cpu_active(dest_cpu
)))
6254 rq_src
= cpu_rq(src_cpu
);
6255 rq_dest
= cpu_rq(dest_cpu
);
6257 raw_spin_lock(&p
->pi_lock
);
6258 double_rq_lock(rq_src
, rq_dest
);
6259 /* Already moved. */
6260 if (task_cpu(p
) != src_cpu
)
6262 /* Affinity changed (again). */
6263 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6267 * If we're not on a rq, the next wake-up will ensure we're
6271 deactivate_task(rq_src
, p
, 0);
6272 set_task_cpu(p
, dest_cpu
);
6273 activate_task(rq_dest
, p
, 0);
6274 check_preempt_curr(rq_dest
, p
, 0);
6279 double_rq_unlock(rq_src
, rq_dest
);
6280 raw_spin_unlock(&p
->pi_lock
);
6285 * migration_cpu_stop - this will be executed by a highprio stopper thread
6286 * and performs thread migration by bumping thread off CPU then
6287 * 'pushing' onto another runqueue.
6289 static int migration_cpu_stop(void *data
)
6291 struct migration_arg
*arg
= data
;
6294 * The original target cpu might have gone down and we might
6295 * be on another cpu but it doesn't matter.
6297 local_irq_disable();
6298 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
6303 #ifdef CONFIG_HOTPLUG_CPU
6306 * Ensures that the idle task is using init_mm right before its cpu goes
6309 void idle_task_exit(void)
6311 struct mm_struct
*mm
= current
->active_mm
;
6313 BUG_ON(cpu_online(smp_processor_id()));
6316 switch_mm(mm
, &init_mm
, current
);
6321 * While a dead CPU has no uninterruptible tasks queued at this point,
6322 * it might still have a nonzero ->nr_uninterruptible counter, because
6323 * for performance reasons the counter is not stricly tracking tasks to
6324 * their home CPUs. So we just add the counter to another CPU's counter,
6325 * to keep the global sum constant after CPU-down:
6327 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6329 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
6331 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6332 rq_src
->nr_uninterruptible
= 0;
6336 * remove the tasks which were accounted by rq from calc_load_tasks.
6338 static void calc_global_load_remove(struct rq
*rq
)
6340 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
6341 rq
->calc_load_active
= 0;
6344 #ifdef CONFIG_CFS_BANDWIDTH
6345 static void unthrottle_offline_cfs_rqs(struct rq
*rq
)
6347 struct cfs_rq
*cfs_rq
;
6349 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6350 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
6352 if (!cfs_rq
->runtime_enabled
)
6356 * clock_task is not advancing so we just need to make sure
6357 * there's some valid quota amount
6359 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
6360 if (cfs_rq_throttled(cfs_rq
))
6361 unthrottle_cfs_rq(cfs_rq
);
6365 static void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
6369 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6370 * try_to_wake_up()->select_task_rq().
6372 * Called with rq->lock held even though we'er in stop_machine() and
6373 * there's no concurrency possible, we hold the required locks anyway
6374 * because of lock validation efforts.
6376 static void migrate_tasks(unsigned int dead_cpu
)
6378 struct rq
*rq
= cpu_rq(dead_cpu
);
6379 struct task_struct
*next
, *stop
= rq
->stop
;
6383 * Fudge the rq selection such that the below task selection loop
6384 * doesn't get stuck on the currently eligible stop task.
6386 * We're currently inside stop_machine() and the rq is either stuck
6387 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6388 * either way we should never end up calling schedule() until we're
6393 /* Ensure any throttled groups are reachable by pick_next_task */
6394 unthrottle_offline_cfs_rqs(rq
);
6398 * There's this thread running, bail when that's the only
6401 if (rq
->nr_running
== 1)
6404 next
= pick_next_task(rq
);
6406 next
->sched_class
->put_prev_task(rq
, next
);
6408 /* Find suitable destination for @next, with force if needed. */
6409 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
6410 raw_spin_unlock(&rq
->lock
);
6412 __migrate_task(next
, dead_cpu
, dest_cpu
);
6414 raw_spin_lock(&rq
->lock
);
6420 #endif /* CONFIG_HOTPLUG_CPU */
6422 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6424 static struct ctl_table sd_ctl_dir
[] = {
6426 .procname
= "sched_domain",
6432 static struct ctl_table sd_ctl_root
[] = {
6434 .procname
= "kernel",
6436 .child
= sd_ctl_dir
,
6441 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6443 struct ctl_table
*entry
=
6444 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6449 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6451 struct ctl_table
*entry
;
6454 * In the intermediate directories, both the child directory and
6455 * procname are dynamically allocated and could fail but the mode
6456 * will always be set. In the lowest directory the names are
6457 * static strings and all have proc handlers.
6459 for (entry
= *tablep
; entry
->mode
; entry
++) {
6461 sd_free_ctl_entry(&entry
->child
);
6462 if (entry
->proc_handler
== NULL
)
6463 kfree(entry
->procname
);
6471 set_table_entry(struct ctl_table
*entry
,
6472 const char *procname
, void *data
, int maxlen
,
6473 mode_t mode
, proc_handler
*proc_handler
)
6475 entry
->procname
= procname
;
6477 entry
->maxlen
= maxlen
;
6479 entry
->proc_handler
= proc_handler
;
6482 static struct ctl_table
*
6483 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6485 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6490 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6491 sizeof(long), 0644, proc_doulongvec_minmax
);
6492 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6493 sizeof(long), 0644, proc_doulongvec_minmax
);
6494 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6495 sizeof(int), 0644, proc_dointvec_minmax
);
6496 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6497 sizeof(int), 0644, proc_dointvec_minmax
);
6498 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6499 sizeof(int), 0644, proc_dointvec_minmax
);
6500 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6501 sizeof(int), 0644, proc_dointvec_minmax
);
6502 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6503 sizeof(int), 0644, proc_dointvec_minmax
);
6504 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6505 sizeof(int), 0644, proc_dointvec_minmax
);
6506 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6507 sizeof(int), 0644, proc_dointvec_minmax
);
6508 set_table_entry(&table
[9], "cache_nice_tries",
6509 &sd
->cache_nice_tries
,
6510 sizeof(int), 0644, proc_dointvec_minmax
);
6511 set_table_entry(&table
[10], "flags", &sd
->flags
,
6512 sizeof(int), 0644, proc_dointvec_minmax
);
6513 set_table_entry(&table
[11], "name", sd
->name
,
6514 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6515 /* &table[12] is terminator */
6520 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6522 struct ctl_table
*entry
, *table
;
6523 struct sched_domain
*sd
;
6524 int domain_num
= 0, i
;
6527 for_each_domain(cpu
, sd
)
6529 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6534 for_each_domain(cpu
, sd
) {
6535 snprintf(buf
, 32, "domain%d", i
);
6536 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6538 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6545 static struct ctl_table_header
*sd_sysctl_header
;
6546 static void register_sched_domain_sysctl(void)
6548 int i
, cpu_num
= num_possible_cpus();
6549 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6552 WARN_ON(sd_ctl_dir
[0].child
);
6553 sd_ctl_dir
[0].child
= entry
;
6558 for_each_possible_cpu(i
) {
6559 snprintf(buf
, 32, "cpu%d", i
);
6560 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6562 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6566 WARN_ON(sd_sysctl_header
);
6567 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6570 /* may be called multiple times per register */
6571 static void unregister_sched_domain_sysctl(void)
6573 if (sd_sysctl_header
)
6574 unregister_sysctl_table(sd_sysctl_header
);
6575 sd_sysctl_header
= NULL
;
6576 if (sd_ctl_dir
[0].child
)
6577 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6580 static void register_sched_domain_sysctl(void)
6583 static void unregister_sched_domain_sysctl(void)
6588 static void set_rq_online(struct rq
*rq
)
6591 const struct sched_class
*class;
6593 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6596 for_each_class(class) {
6597 if (class->rq_online
)
6598 class->rq_online(rq
);
6603 static void set_rq_offline(struct rq
*rq
)
6606 const struct sched_class
*class;
6608 for_each_class(class) {
6609 if (class->rq_offline
)
6610 class->rq_offline(rq
);
6613 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6619 * migration_call - callback that gets triggered when a CPU is added.
6620 * Here we can start up the necessary migration thread for the new CPU.
6622 static int __cpuinit
6623 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6625 int cpu
= (long)hcpu
;
6626 unsigned long flags
;
6627 struct rq
*rq
= cpu_rq(cpu
);
6629 switch (action
& ~CPU_TASKS_FROZEN
) {
6631 case CPU_UP_PREPARE
:
6632 rq
->calc_load_update
= calc_load_update
;
6636 /* Update our root-domain */
6637 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6639 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6643 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6646 #ifdef CONFIG_HOTPLUG_CPU
6648 sched_ttwu_pending();
6649 /* Update our root-domain */
6650 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6652 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6656 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6657 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6659 migrate_nr_uninterruptible(rq
);
6660 calc_global_load_remove(rq
);
6665 update_max_interval();
6671 * Register at high priority so that task migration (migrate_all_tasks)
6672 * happens before everything else. This has to be lower priority than
6673 * the notifier in the perf_event subsystem, though.
6675 static struct notifier_block __cpuinitdata migration_notifier
= {
6676 .notifier_call
= migration_call
,
6677 .priority
= CPU_PRI_MIGRATION
,
6680 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6681 unsigned long action
, void *hcpu
)
6683 switch (action
& ~CPU_TASKS_FROZEN
) {
6685 case CPU_DOWN_FAILED
:
6686 set_cpu_active((long)hcpu
, true);
6693 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6694 unsigned long action
, void *hcpu
)
6696 switch (action
& ~CPU_TASKS_FROZEN
) {
6697 case CPU_DOWN_PREPARE
:
6698 set_cpu_active((long)hcpu
, false);
6705 static int __init
migration_init(void)
6707 void *cpu
= (void *)(long)smp_processor_id();
6710 /* Initialize migration for the boot CPU */
6711 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6712 BUG_ON(err
== NOTIFY_BAD
);
6713 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6714 register_cpu_notifier(&migration_notifier
);
6716 /* Register cpu active notifiers */
6717 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6718 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6722 early_initcall(migration_init
);
6727 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
6729 #ifdef CONFIG_SCHED_DEBUG
6731 static __read_mostly
int sched_domain_debug_enabled
;
6733 static int __init
sched_domain_debug_setup(char *str
)
6735 sched_domain_debug_enabled
= 1;
6739 early_param("sched_debug", sched_domain_debug_setup
);
6741 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6742 struct cpumask
*groupmask
)
6744 struct sched_group
*group
= sd
->groups
;
6747 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6748 cpumask_clear(groupmask
);
6750 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6752 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6753 printk("does not load-balance\n");
6755 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6760 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6762 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6763 printk(KERN_ERR
"ERROR: domain->span does not contain "
6766 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6767 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6771 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6775 printk(KERN_ERR
"ERROR: group is NULL\n");
6779 if (!group
->sgp
->power
) {
6780 printk(KERN_CONT
"\n");
6781 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6786 if (!cpumask_weight(sched_group_cpus(group
))) {
6787 printk(KERN_CONT
"\n");
6788 printk(KERN_ERR
"ERROR: empty group\n");
6792 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6793 printk(KERN_CONT
"\n");
6794 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6798 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6800 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6802 printk(KERN_CONT
" %s", str
);
6803 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
6804 printk(KERN_CONT
" (cpu_power = %d)",
6808 group
= group
->next
;
6809 } while (group
!= sd
->groups
);
6810 printk(KERN_CONT
"\n");
6812 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6813 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6816 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6817 printk(KERN_ERR
"ERROR: parent span is not a superset "
6818 "of domain->span\n");
6822 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6826 if (!sched_domain_debug_enabled
)
6830 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6834 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6837 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
6845 #else /* !CONFIG_SCHED_DEBUG */
6846 # define sched_domain_debug(sd, cpu) do { } while (0)
6847 #endif /* CONFIG_SCHED_DEBUG */
6849 static int sd_degenerate(struct sched_domain
*sd
)
6851 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6854 /* Following flags need at least 2 groups */
6855 if (sd
->flags
& (SD_LOAD_BALANCE
|
6856 SD_BALANCE_NEWIDLE
|
6860 SD_SHARE_PKG_RESOURCES
)) {
6861 if (sd
->groups
!= sd
->groups
->next
)
6865 /* Following flags don't use groups */
6866 if (sd
->flags
& (SD_WAKE_AFFINE
))
6873 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6875 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6877 if (sd_degenerate(parent
))
6880 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6883 /* Flags needing groups don't count if only 1 group in parent */
6884 if (parent
->groups
== parent
->groups
->next
) {
6885 pflags
&= ~(SD_LOAD_BALANCE
|
6886 SD_BALANCE_NEWIDLE
|
6890 SD_SHARE_PKG_RESOURCES
);
6891 if (nr_node_ids
== 1)
6892 pflags
&= ~SD_SERIALIZE
;
6894 if (~cflags
& pflags
)
6900 static void free_rootdomain(struct rcu_head
*rcu
)
6902 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
6904 cpupri_cleanup(&rd
->cpupri
);
6905 free_cpumask_var(rd
->rto_mask
);
6906 free_cpumask_var(rd
->online
);
6907 free_cpumask_var(rd
->span
);
6911 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6913 struct root_domain
*old_rd
= NULL
;
6914 unsigned long flags
;
6916 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6921 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6924 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6927 * If we dont want to free the old_rt yet then
6928 * set old_rd to NULL to skip the freeing later
6931 if (!atomic_dec_and_test(&old_rd
->refcount
))
6935 atomic_inc(&rd
->refcount
);
6938 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6939 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6942 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6945 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
6948 static int init_rootdomain(struct root_domain
*rd
)
6950 memset(rd
, 0, sizeof(*rd
));
6952 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6954 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6956 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6959 if (cpupri_init(&rd
->cpupri
) != 0)
6964 free_cpumask_var(rd
->rto_mask
);
6966 free_cpumask_var(rd
->online
);
6968 free_cpumask_var(rd
->span
);
6973 static void init_defrootdomain(void)
6975 init_rootdomain(&def_root_domain
);
6977 atomic_set(&def_root_domain
.refcount
, 1);
6980 static struct root_domain
*alloc_rootdomain(void)
6982 struct root_domain
*rd
;
6984 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6988 if (init_rootdomain(rd
) != 0) {
6996 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
6998 struct sched_group
*tmp
, *first
;
7007 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
7012 } while (sg
!= first
);
7015 static void free_sched_domain(struct rcu_head
*rcu
)
7017 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
7020 * If its an overlapping domain it has private groups, iterate and
7023 if (sd
->flags
& SD_OVERLAP
) {
7024 free_sched_groups(sd
->groups
, 1);
7025 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
7026 kfree(sd
->groups
->sgp
);
7032 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
7034 call_rcu(&sd
->rcu
, free_sched_domain
);
7037 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
7039 for (; sd
; sd
= sd
->parent
)
7040 destroy_sched_domain(sd
, cpu
);
7044 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7045 * hold the hotplug lock.
7048 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7050 struct rq
*rq
= cpu_rq(cpu
);
7051 struct sched_domain
*tmp
;
7053 /* Remove the sched domains which do not contribute to scheduling. */
7054 for (tmp
= sd
; tmp
; ) {
7055 struct sched_domain
*parent
= tmp
->parent
;
7059 if (sd_parent_degenerate(tmp
, parent
)) {
7060 tmp
->parent
= parent
->parent
;
7062 parent
->parent
->child
= tmp
;
7063 destroy_sched_domain(parent
, cpu
);
7068 if (sd
&& sd_degenerate(sd
)) {
7071 destroy_sched_domain(tmp
, cpu
);
7076 sched_domain_debug(sd
, cpu
);
7078 rq_attach_root(rq
, rd
);
7080 rcu_assign_pointer(rq
->sd
, sd
);
7081 destroy_sched_domains(tmp
, cpu
);
7084 /* cpus with isolated domains */
7085 static cpumask_var_t cpu_isolated_map
;
7087 /* Setup the mask of cpus configured for isolated domains */
7088 static int __init
isolated_cpu_setup(char *str
)
7090 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
7091 cpulist_parse(str
, cpu_isolated_map
);
7095 __setup("isolcpus=", isolated_cpu_setup
);
7097 #define SD_NODES_PER_DOMAIN 16
7102 * find_next_best_node - find the next node to include in a sched_domain
7103 * @node: node whose sched_domain we're building
7104 * @used_nodes: nodes already in the sched_domain
7106 * Find the next node to include in a given scheduling domain. Simply
7107 * finds the closest node not already in the @used_nodes map.
7109 * Should use nodemask_t.
7111 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7113 int i
, n
, val
, min_val
, best_node
= -1;
7117 for (i
= 0; i
< nr_node_ids
; i
++) {
7118 /* Start at @node */
7119 n
= (node
+ i
) % nr_node_ids
;
7121 if (!nr_cpus_node(n
))
7124 /* Skip already used nodes */
7125 if (node_isset(n
, *used_nodes
))
7128 /* Simple min distance search */
7129 val
= node_distance(node
, n
);
7131 if (val
< min_val
) {
7137 if (best_node
!= -1)
7138 node_set(best_node
, *used_nodes
);
7143 * sched_domain_node_span - get a cpumask for a node's sched_domain
7144 * @node: node whose cpumask we're constructing
7145 * @span: resulting cpumask
7147 * Given a node, construct a good cpumask for its sched_domain to span. It
7148 * should be one that prevents unnecessary balancing, but also spreads tasks
7151 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7153 nodemask_t used_nodes
;
7156 cpumask_clear(span
);
7157 nodes_clear(used_nodes
);
7159 cpumask_or(span
, span
, cpumask_of_node(node
));
7160 node_set(node
, used_nodes
);
7162 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7163 int next_node
= find_next_best_node(node
, &used_nodes
);
7166 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7170 static const struct cpumask
*cpu_node_mask(int cpu
)
7172 lockdep_assert_held(&sched_domains_mutex
);
7174 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
7176 return sched_domains_tmpmask
;
7179 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
7181 return cpu_possible_mask
;
7183 #endif /* CONFIG_NUMA */
7185 static const struct cpumask
*cpu_cpu_mask(int cpu
)
7187 return cpumask_of_node(cpu_to_node(cpu
));
7190 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7193 struct sched_domain
**__percpu sd
;
7194 struct sched_group
**__percpu sg
;
7195 struct sched_group_power
**__percpu sgp
;
7199 struct sched_domain
** __percpu sd
;
7200 struct root_domain
*rd
;
7210 struct sched_domain_topology_level
;
7212 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
7213 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
7215 #define SDTL_OVERLAP 0x01
7217 struct sched_domain_topology_level
{
7218 sched_domain_init_f init
;
7219 sched_domain_mask_f mask
;
7221 struct sd_data data
;
7225 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
7227 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
7228 const struct cpumask
*span
= sched_domain_span(sd
);
7229 struct cpumask
*covered
= sched_domains_tmpmask
;
7230 struct sd_data
*sdd
= sd
->private;
7231 struct sched_domain
*child
;
7234 cpumask_clear(covered
);
7236 for_each_cpu(i
, span
) {
7237 struct cpumask
*sg_span
;
7239 if (cpumask_test_cpu(i
, covered
))
7242 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7243 GFP_KERNEL
, cpu_to_node(i
));
7248 sg_span
= sched_group_cpus(sg
);
7250 child
= *per_cpu_ptr(sdd
->sd
, i
);
7252 child
= child
->child
;
7253 cpumask_copy(sg_span
, sched_domain_span(child
));
7255 cpumask_set_cpu(i
, sg_span
);
7257 cpumask_or(covered
, covered
, sg_span
);
7259 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
7260 atomic_inc(&sg
->sgp
->ref
);
7262 if (cpumask_test_cpu(cpu
, sg_span
))
7272 sd
->groups
= groups
;
7277 free_sched_groups(first
, 0);
7282 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
7284 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
7285 struct sched_domain
*child
= sd
->child
;
7288 cpu
= cpumask_first(sched_domain_span(child
));
7291 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
7292 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
7293 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
7300 * build_sched_groups will build a circular linked list of the groups
7301 * covered by the given span, and will set each group's ->cpumask correctly,
7302 * and ->cpu_power to 0.
7304 * Assumes the sched_domain tree is fully constructed
7307 build_sched_groups(struct sched_domain
*sd
, int cpu
)
7309 struct sched_group
*first
= NULL
, *last
= NULL
;
7310 struct sd_data
*sdd
= sd
->private;
7311 const struct cpumask
*span
= sched_domain_span(sd
);
7312 struct cpumask
*covered
;
7315 get_group(cpu
, sdd
, &sd
->groups
);
7316 atomic_inc(&sd
->groups
->ref
);
7318 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
7321 lockdep_assert_held(&sched_domains_mutex
);
7322 covered
= sched_domains_tmpmask
;
7324 cpumask_clear(covered
);
7326 for_each_cpu(i
, span
) {
7327 struct sched_group
*sg
;
7328 int group
= get_group(i
, sdd
, &sg
);
7331 if (cpumask_test_cpu(i
, covered
))
7334 cpumask_clear(sched_group_cpus(sg
));
7337 for_each_cpu(j
, span
) {
7338 if (get_group(j
, sdd
, NULL
) != group
)
7341 cpumask_set_cpu(j
, covered
);
7342 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7357 * Initialize sched groups cpu_power.
7359 * cpu_power indicates the capacity of sched group, which is used while
7360 * distributing the load between different sched groups in a sched domain.
7361 * Typically cpu_power for all the groups in a sched domain will be same unless
7362 * there are asymmetries in the topology. If there are asymmetries, group
7363 * having more cpu_power will pickup more load compared to the group having
7366 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7368 struct sched_group
*sg
= sd
->groups
;
7370 WARN_ON(!sd
|| !sg
);
7373 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
7375 } while (sg
!= sd
->groups
);
7377 if (cpu
!= group_first_cpu(sg
))
7380 update_group_power(sd
, cpu
);
7384 * Initializers for schedule domains
7385 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7388 #ifdef CONFIG_SCHED_DEBUG
7389 # define SD_INIT_NAME(sd, type) sd->name = #type
7391 # define SD_INIT_NAME(sd, type) do { } while (0)
7394 #define SD_INIT_FUNC(type) \
7395 static noinline struct sched_domain * \
7396 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7398 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7399 *sd = SD_##type##_INIT; \
7400 SD_INIT_NAME(sd, type); \
7401 sd->private = &tl->data; \
7407 SD_INIT_FUNC(ALLNODES
)
7410 #ifdef CONFIG_SCHED_SMT
7411 SD_INIT_FUNC(SIBLING
)
7413 #ifdef CONFIG_SCHED_MC
7416 #ifdef CONFIG_SCHED_BOOK
7420 static int default_relax_domain_level
= -1;
7421 int sched_domain_level_max
;
7423 static int __init
setup_relax_domain_level(char *str
)
7427 val
= simple_strtoul(str
, NULL
, 0);
7428 if (val
< sched_domain_level_max
)
7429 default_relax_domain_level
= val
;
7433 __setup("relax_domain_level=", setup_relax_domain_level
);
7435 static void set_domain_attribute(struct sched_domain
*sd
,
7436 struct sched_domain_attr
*attr
)
7440 if (!attr
|| attr
->relax_domain_level
< 0) {
7441 if (default_relax_domain_level
< 0)
7444 request
= default_relax_domain_level
;
7446 request
= attr
->relax_domain_level
;
7447 if (request
< sd
->level
) {
7448 /* turn off idle balance on this domain */
7449 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7451 /* turn on idle balance on this domain */
7452 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7456 static void __sdt_free(const struct cpumask
*cpu_map
);
7457 static int __sdt_alloc(const struct cpumask
*cpu_map
);
7459 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7460 const struct cpumask
*cpu_map
)
7464 if (!atomic_read(&d
->rd
->refcount
))
7465 free_rootdomain(&d
->rd
->rcu
); /* fall through */
7467 free_percpu(d
->sd
); /* fall through */
7469 __sdt_free(cpu_map
); /* fall through */
7475 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7476 const struct cpumask
*cpu_map
)
7478 memset(d
, 0, sizeof(*d
));
7480 if (__sdt_alloc(cpu_map
))
7481 return sa_sd_storage
;
7482 d
->sd
= alloc_percpu(struct sched_domain
*);
7484 return sa_sd_storage
;
7485 d
->rd
= alloc_rootdomain();
7488 return sa_rootdomain
;
7492 * NULL the sd_data elements we've used to build the sched_domain and
7493 * sched_group structure so that the subsequent __free_domain_allocs()
7494 * will not free the data we're using.
7496 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
7498 struct sd_data
*sdd
= sd
->private;
7500 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
7501 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
7503 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
7504 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
7506 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
7507 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
7510 #ifdef CONFIG_SCHED_SMT
7511 static const struct cpumask
*cpu_smt_mask(int cpu
)
7513 return topology_thread_cpumask(cpu
);
7518 * Topology list, bottom-up.
7520 static struct sched_domain_topology_level default_topology
[] = {
7521 #ifdef CONFIG_SCHED_SMT
7522 { sd_init_SIBLING
, cpu_smt_mask
, },
7524 #ifdef CONFIG_SCHED_MC
7525 { sd_init_MC
, cpu_coregroup_mask
, },
7527 #ifdef CONFIG_SCHED_BOOK
7528 { sd_init_BOOK
, cpu_book_mask
, },
7530 { sd_init_CPU
, cpu_cpu_mask
, },
7532 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
7533 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
7538 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
7540 static int __sdt_alloc(const struct cpumask
*cpu_map
)
7542 struct sched_domain_topology_level
*tl
;
7545 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7546 struct sd_data
*sdd
= &tl
->data
;
7548 sdd
->sd
= alloc_percpu(struct sched_domain
*);
7552 sdd
->sg
= alloc_percpu(struct sched_group
*);
7556 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
7560 for_each_cpu(j
, cpu_map
) {
7561 struct sched_domain
*sd
;
7562 struct sched_group
*sg
;
7563 struct sched_group_power
*sgp
;
7565 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
7566 GFP_KERNEL
, cpu_to_node(j
));
7570 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
7572 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7573 GFP_KERNEL
, cpu_to_node(j
));
7577 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
7579 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
7580 GFP_KERNEL
, cpu_to_node(j
));
7584 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
7591 static void __sdt_free(const struct cpumask
*cpu_map
)
7593 struct sched_domain_topology_level
*tl
;
7596 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7597 struct sd_data
*sdd
= &tl
->data
;
7599 for_each_cpu(j
, cpu_map
) {
7600 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
7601 if (sd
&& (sd
->flags
& SD_OVERLAP
))
7602 free_sched_groups(sd
->groups
, 0);
7603 kfree(*per_cpu_ptr(sdd
->sd
, j
));
7604 kfree(*per_cpu_ptr(sdd
->sg
, j
));
7605 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
7607 free_percpu(sdd
->sd
);
7608 free_percpu(sdd
->sg
);
7609 free_percpu(sdd
->sgp
);
7613 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
7614 struct s_data
*d
, const struct cpumask
*cpu_map
,
7615 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
7618 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
7622 set_domain_attribute(sd
, attr
);
7623 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
7625 sd
->level
= child
->level
+ 1;
7626 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
7635 * Build sched domains for a given set of cpus and attach the sched domains
7636 * to the individual cpus
7638 static int build_sched_domains(const struct cpumask
*cpu_map
,
7639 struct sched_domain_attr
*attr
)
7641 enum s_alloc alloc_state
= sa_none
;
7642 struct sched_domain
*sd
;
7644 int i
, ret
= -ENOMEM
;
7646 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7647 if (alloc_state
!= sa_rootdomain
)
7650 /* Set up domains for cpus specified by the cpu_map. */
7651 for_each_cpu(i
, cpu_map
) {
7652 struct sched_domain_topology_level
*tl
;
7655 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7656 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
7657 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
7658 sd
->flags
|= SD_OVERLAP
;
7659 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
7666 *per_cpu_ptr(d
.sd
, i
) = sd
;
7669 /* Build the groups for the domains */
7670 for_each_cpu(i
, cpu_map
) {
7671 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7672 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
7673 if (sd
->flags
& SD_OVERLAP
) {
7674 if (build_overlap_sched_groups(sd
, i
))
7677 if (build_sched_groups(sd
, i
))
7683 /* Calculate CPU power for physical packages and nodes */
7684 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
7685 if (!cpumask_test_cpu(i
, cpu_map
))
7688 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7689 claim_allocations(i
, sd
);
7690 init_sched_groups_power(i
, sd
);
7694 /* Attach the domains */
7696 for_each_cpu(i
, cpu_map
) {
7697 sd
= *per_cpu_ptr(d
.sd
, i
);
7698 cpu_attach_domain(sd
, d
.rd
, i
);
7704 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7708 static cpumask_var_t
*doms_cur
; /* current sched domains */
7709 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7710 static struct sched_domain_attr
*dattr_cur
;
7711 /* attribues of custom domains in 'doms_cur' */
7714 * Special case: If a kmalloc of a doms_cur partition (array of
7715 * cpumask) fails, then fallback to a single sched domain,
7716 * as determined by the single cpumask fallback_doms.
7718 static cpumask_var_t fallback_doms
;
7721 * arch_update_cpu_topology lets virtualized architectures update the
7722 * cpu core maps. It is supposed to return 1 if the topology changed
7723 * or 0 if it stayed the same.
7725 int __attribute__((weak
)) arch_update_cpu_topology(void)
7730 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7733 cpumask_var_t
*doms
;
7735 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7738 for (i
= 0; i
< ndoms
; i
++) {
7739 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7740 free_sched_domains(doms
, i
);
7747 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7750 for (i
= 0; i
< ndoms
; i
++)
7751 free_cpumask_var(doms
[i
]);
7756 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7757 * For now this just excludes isolated cpus, but could be used to
7758 * exclude other special cases in the future.
7760 static int init_sched_domains(const struct cpumask
*cpu_map
)
7764 arch_update_cpu_topology();
7766 doms_cur
= alloc_sched_domains(ndoms_cur
);
7768 doms_cur
= &fallback_doms
;
7769 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7771 err
= build_sched_domains(doms_cur
[0], NULL
);
7772 register_sched_domain_sysctl();
7778 * Detach sched domains from a group of cpus specified in cpu_map
7779 * These cpus will now be attached to the NULL domain
7781 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7786 for_each_cpu(i
, cpu_map
)
7787 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7791 /* handle null as "default" */
7792 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7793 struct sched_domain_attr
*new, int idx_new
)
7795 struct sched_domain_attr tmp
;
7802 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7803 new ? (new + idx_new
) : &tmp
,
7804 sizeof(struct sched_domain_attr
));
7808 * Partition sched domains as specified by the 'ndoms_new'
7809 * cpumasks in the array doms_new[] of cpumasks. This compares
7810 * doms_new[] to the current sched domain partitioning, doms_cur[].
7811 * It destroys each deleted domain and builds each new domain.
7813 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7814 * The masks don't intersect (don't overlap.) We should setup one
7815 * sched domain for each mask. CPUs not in any of the cpumasks will
7816 * not be load balanced. If the same cpumask appears both in the
7817 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7820 * The passed in 'doms_new' should be allocated using
7821 * alloc_sched_domains. This routine takes ownership of it and will
7822 * free_sched_domains it when done with it. If the caller failed the
7823 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7824 * and partition_sched_domains() will fallback to the single partition
7825 * 'fallback_doms', it also forces the domains to be rebuilt.
7827 * If doms_new == NULL it will be replaced with cpu_online_mask.
7828 * ndoms_new == 0 is a special case for destroying existing domains,
7829 * and it will not create the default domain.
7831 * Call with hotplug lock held
7833 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7834 struct sched_domain_attr
*dattr_new
)
7839 mutex_lock(&sched_domains_mutex
);
7841 /* always unregister in case we don't destroy any domains */
7842 unregister_sched_domain_sysctl();
7844 /* Let architecture update cpu core mappings. */
7845 new_topology
= arch_update_cpu_topology();
7847 n
= doms_new
? ndoms_new
: 0;
7849 /* Destroy deleted domains */
7850 for (i
= 0; i
< ndoms_cur
; i
++) {
7851 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7852 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7853 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7856 /* no match - a current sched domain not in new doms_new[] */
7857 detach_destroy_domains(doms_cur
[i
]);
7862 if (doms_new
== NULL
) {
7864 doms_new
= &fallback_doms
;
7865 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7866 WARN_ON_ONCE(dattr_new
);
7869 /* Build new domains */
7870 for (i
= 0; i
< ndoms_new
; i
++) {
7871 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7872 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7873 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7876 /* no match - add a new doms_new */
7877 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7882 /* Remember the new sched domains */
7883 if (doms_cur
!= &fallback_doms
)
7884 free_sched_domains(doms_cur
, ndoms_cur
);
7885 kfree(dattr_cur
); /* kfree(NULL) is safe */
7886 doms_cur
= doms_new
;
7887 dattr_cur
= dattr_new
;
7888 ndoms_cur
= ndoms_new
;
7890 register_sched_domain_sysctl();
7892 mutex_unlock(&sched_domains_mutex
);
7895 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7896 static void reinit_sched_domains(void)
7900 /* Destroy domains first to force the rebuild */
7901 partition_sched_domains(0, NULL
, NULL
);
7903 rebuild_sched_domains();
7907 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7909 unsigned int level
= 0;
7911 if (sscanf(buf
, "%u", &level
) != 1)
7915 * level is always be positive so don't check for
7916 * level < POWERSAVINGS_BALANCE_NONE which is 0
7917 * What happens on 0 or 1 byte write,
7918 * need to check for count as well?
7921 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7925 sched_smt_power_savings
= level
;
7927 sched_mc_power_savings
= level
;
7929 reinit_sched_domains();
7934 #ifdef CONFIG_SCHED_MC
7935 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7936 struct sysdev_class_attribute
*attr
,
7939 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7941 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7942 struct sysdev_class_attribute
*attr
,
7943 const char *buf
, size_t count
)
7945 return sched_power_savings_store(buf
, count
, 0);
7947 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7948 sched_mc_power_savings_show
,
7949 sched_mc_power_savings_store
);
7952 #ifdef CONFIG_SCHED_SMT
7953 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7954 struct sysdev_class_attribute
*attr
,
7957 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7959 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7960 struct sysdev_class_attribute
*attr
,
7961 const char *buf
, size_t count
)
7963 return sched_power_savings_store(buf
, count
, 1);
7965 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7966 sched_smt_power_savings_show
,
7967 sched_smt_power_savings_store
);
7970 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7974 #ifdef CONFIG_SCHED_SMT
7976 err
= sysfs_create_file(&cls
->kset
.kobj
,
7977 &attr_sched_smt_power_savings
.attr
);
7979 #ifdef CONFIG_SCHED_MC
7980 if (!err
&& mc_capable())
7981 err
= sysfs_create_file(&cls
->kset
.kobj
,
7982 &attr_sched_mc_power_savings
.attr
);
7986 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7989 * Update cpusets according to cpu_active mask. If cpusets are
7990 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7991 * around partition_sched_domains().
7993 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7996 switch (action
& ~CPU_TASKS_FROZEN
) {
7998 case CPU_DOWN_FAILED
:
7999 cpuset_update_active_cpus();
8006 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
8009 switch (action
& ~CPU_TASKS_FROZEN
) {
8010 case CPU_DOWN_PREPARE
:
8011 cpuset_update_active_cpus();
8018 static int update_runtime(struct notifier_block
*nfb
,
8019 unsigned long action
, void *hcpu
)
8021 int cpu
= (int)(long)hcpu
;
8024 case CPU_DOWN_PREPARE
:
8025 case CPU_DOWN_PREPARE_FROZEN
:
8026 disable_runtime(cpu_rq(cpu
));
8029 case CPU_DOWN_FAILED
:
8030 case CPU_DOWN_FAILED_FROZEN
:
8032 case CPU_ONLINE_FROZEN
:
8033 enable_runtime(cpu_rq(cpu
));
8041 void __init
sched_init_smp(void)
8043 cpumask_var_t non_isolated_cpus
;
8045 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8046 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8049 mutex_lock(&sched_domains_mutex
);
8050 init_sched_domains(cpu_active_mask
);
8051 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8052 if (cpumask_empty(non_isolated_cpus
))
8053 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8054 mutex_unlock(&sched_domains_mutex
);
8057 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
8058 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
8060 /* RT runtime code needs to handle some hotplug events */
8061 hotcpu_notifier(update_runtime
, 0);
8065 /* Move init over to a non-isolated CPU */
8066 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8068 sched_init_granularity();
8069 free_cpumask_var(non_isolated_cpus
);
8071 init_sched_rt_class();
8074 void __init
sched_init_smp(void)
8076 sched_init_granularity();
8078 #endif /* CONFIG_SMP */
8080 const_debug
unsigned int sysctl_timer_migration
= 1;
8082 int in_sched_functions(unsigned long addr
)
8084 return in_lock_functions(addr
) ||
8085 (addr
>= (unsigned long)__sched_text_start
8086 && addr
< (unsigned long)__sched_text_end
);
8089 static void init_cfs_rq(struct cfs_rq
*cfs_rq
)
8091 cfs_rq
->tasks_timeline
= RB_ROOT
;
8092 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8093 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8094 #ifndef CONFIG_64BIT
8095 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
8099 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8101 struct rt_prio_array
*array
;
8104 array
= &rt_rq
->active
;
8105 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8106 INIT_LIST_HEAD(array
->queue
+ i
);
8107 __clear_bit(i
, array
->bitmap
);
8109 /* delimiter for bitsearch: */
8110 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8112 #if defined CONFIG_SMP
8113 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8114 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8115 rt_rq
->rt_nr_migratory
= 0;
8116 rt_rq
->overloaded
= 0;
8117 plist_head_init(&rt_rq
->pushable_tasks
);
8121 rt_rq
->rt_throttled
= 0;
8122 rt_rq
->rt_runtime
= 0;
8123 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
8126 #ifdef CONFIG_FAIR_GROUP_SCHED
8127 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8128 struct sched_entity
*se
, int cpu
,
8129 struct sched_entity
*parent
)
8131 struct rq
*rq
= cpu_rq(cpu
);
8136 /* allow initial update_cfs_load() to truncate */
8137 cfs_rq
->load_stamp
= 1;
8139 init_cfs_rq_runtime(cfs_rq
);
8141 tg
->cfs_rq
[cpu
] = cfs_rq
;
8144 /* se could be NULL for root_task_group */
8149 se
->cfs_rq
= &rq
->cfs
;
8151 se
->cfs_rq
= parent
->my_q
;
8154 update_load_set(&se
->load
, 0);
8155 se
->parent
= parent
;
8159 #ifdef CONFIG_RT_GROUP_SCHED
8160 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8161 struct sched_rt_entity
*rt_se
, int cpu
,
8162 struct sched_rt_entity
*parent
)
8164 struct rq
*rq
= cpu_rq(cpu
);
8166 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8167 rt_rq
->rt_nr_boosted
= 0;
8171 tg
->rt_rq
[cpu
] = rt_rq
;
8172 tg
->rt_se
[cpu
] = rt_se
;
8178 rt_se
->rt_rq
= &rq
->rt
;
8180 rt_se
->rt_rq
= parent
->my_q
;
8182 rt_se
->my_q
= rt_rq
;
8183 rt_se
->parent
= parent
;
8184 INIT_LIST_HEAD(&rt_se
->run_list
);
8188 void __init
sched_init(void)
8191 unsigned long alloc_size
= 0, ptr
;
8193 #ifdef CONFIG_FAIR_GROUP_SCHED
8194 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8196 #ifdef CONFIG_RT_GROUP_SCHED
8197 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8199 #ifdef CONFIG_CPUMASK_OFFSTACK
8200 alloc_size
+= num_possible_cpus() * cpumask_size();
8203 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
8205 #ifdef CONFIG_FAIR_GROUP_SCHED
8206 root_task_group
.se
= (struct sched_entity
**)ptr
;
8207 ptr
+= nr_cpu_ids
* sizeof(void **);
8209 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8210 ptr
+= nr_cpu_ids
* sizeof(void **);
8212 #endif /* CONFIG_FAIR_GROUP_SCHED */
8213 #ifdef CONFIG_RT_GROUP_SCHED
8214 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8215 ptr
+= nr_cpu_ids
* sizeof(void **);
8217 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8218 ptr
+= nr_cpu_ids
* sizeof(void **);
8220 #endif /* CONFIG_RT_GROUP_SCHED */
8221 #ifdef CONFIG_CPUMASK_OFFSTACK
8222 for_each_possible_cpu(i
) {
8223 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8224 ptr
+= cpumask_size();
8226 #endif /* CONFIG_CPUMASK_OFFSTACK */
8230 init_defrootdomain();
8233 init_rt_bandwidth(&def_rt_bandwidth
,
8234 global_rt_period(), global_rt_runtime());
8236 #ifdef CONFIG_RT_GROUP_SCHED
8237 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8238 global_rt_period(), global_rt_runtime());
8239 #endif /* CONFIG_RT_GROUP_SCHED */
8241 #ifdef CONFIG_CGROUP_SCHED
8242 list_add(&root_task_group
.list
, &task_groups
);
8243 INIT_LIST_HEAD(&root_task_group
.children
);
8244 autogroup_init(&init_task
);
8245 #endif /* CONFIG_CGROUP_SCHED */
8247 for_each_possible_cpu(i
) {
8251 raw_spin_lock_init(&rq
->lock
);
8253 rq
->calc_load_active
= 0;
8254 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8255 init_cfs_rq(&rq
->cfs
);
8256 init_rt_rq(&rq
->rt
, rq
);
8257 #ifdef CONFIG_FAIR_GROUP_SCHED
8258 root_task_group
.shares
= root_task_group_load
;
8259 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8261 * How much cpu bandwidth does root_task_group get?
8263 * In case of task-groups formed thr' the cgroup filesystem, it
8264 * gets 100% of the cpu resources in the system. This overall
8265 * system cpu resource is divided among the tasks of
8266 * root_task_group and its child task-groups in a fair manner,
8267 * based on each entity's (task or task-group's) weight
8268 * (se->load.weight).
8270 * In other words, if root_task_group has 10 tasks of weight
8271 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8272 * then A0's share of the cpu resource is:
8274 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8276 * We achieve this by letting root_task_group's tasks sit
8277 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8279 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
8280 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8281 #endif /* CONFIG_FAIR_GROUP_SCHED */
8283 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8284 #ifdef CONFIG_RT_GROUP_SCHED
8285 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8286 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8289 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8290 rq
->cpu_load
[j
] = 0;
8292 rq
->last_load_update_tick
= jiffies
;
8297 rq
->cpu_power
= SCHED_POWER_SCALE
;
8298 rq
->post_schedule
= 0;
8299 rq
->active_balance
= 0;
8300 rq
->next_balance
= jiffies
;
8305 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8306 rq_attach_root(rq
, &def_root_domain
);
8308 rq
->nohz_balance_kick
= 0;
8312 atomic_set(&rq
->nr_iowait
, 0);
8315 set_load_weight(&init_task
);
8317 #ifdef CONFIG_PREEMPT_NOTIFIERS
8318 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8322 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8325 #ifdef CONFIG_RT_MUTEXES
8326 plist_head_init(&init_task
.pi_waiters
);
8330 * The boot idle thread does lazy MMU switching as well:
8332 atomic_inc(&init_mm
.mm_count
);
8333 enter_lazy_tlb(&init_mm
, current
);
8336 * Make us the idle thread. Technically, schedule() should not be
8337 * called from this thread, however somewhere below it might be,
8338 * but because we are the idle thread, we just pick up running again
8339 * when this runqueue becomes "idle".
8341 init_idle(current
, smp_processor_id());
8343 calc_load_update
= jiffies
+ LOAD_FREQ
;
8346 * During early bootup we pretend to be a normal task:
8348 current
->sched_class
= &fair_sched_class
;
8350 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8351 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
8353 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
8355 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8356 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8357 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8358 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8359 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8361 /* May be allocated at isolcpus cmdline parse time */
8362 if (cpu_isolated_map
== NULL
)
8363 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8366 scheduler_running
= 1;
8369 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8370 static inline int preempt_count_equals(int preempt_offset
)
8372 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8374 return (nested
== preempt_offset
);
8377 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8379 static unsigned long prev_jiffy
; /* ratelimiting */
8381 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8382 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8384 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8386 prev_jiffy
= jiffies
;
8389 "BUG: sleeping function called from invalid context at %s:%d\n",
8392 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8393 in_atomic(), irqs_disabled(),
8394 current
->pid
, current
->comm
);
8396 debug_show_held_locks(current
);
8397 if (irqs_disabled())
8398 print_irqtrace_events(current
);
8401 EXPORT_SYMBOL(__might_sleep
);
8404 #ifdef CONFIG_MAGIC_SYSRQ
8405 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8407 const struct sched_class
*prev_class
= p
->sched_class
;
8408 int old_prio
= p
->prio
;
8413 deactivate_task(rq
, p
, 0);
8414 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8416 activate_task(rq
, p
, 0);
8417 resched_task(rq
->curr
);
8420 check_class_changed(rq
, p
, prev_class
, old_prio
);
8423 void normalize_rt_tasks(void)
8425 struct task_struct
*g
, *p
;
8426 unsigned long flags
;
8429 read_lock_irqsave(&tasklist_lock
, flags
);
8430 do_each_thread(g
, p
) {
8432 * Only normalize user tasks:
8437 p
->se
.exec_start
= 0;
8438 #ifdef CONFIG_SCHEDSTATS
8439 p
->se
.statistics
.wait_start
= 0;
8440 p
->se
.statistics
.sleep_start
= 0;
8441 p
->se
.statistics
.block_start
= 0;
8446 * Renice negative nice level userspace
8449 if (TASK_NICE(p
) < 0 && p
->mm
)
8450 set_user_nice(p
, 0);
8454 raw_spin_lock(&p
->pi_lock
);
8455 rq
= __task_rq_lock(p
);
8457 normalize_task(rq
, p
);
8459 __task_rq_unlock(rq
);
8460 raw_spin_unlock(&p
->pi_lock
);
8461 } while_each_thread(g
, p
);
8463 read_unlock_irqrestore(&tasklist_lock
, flags
);
8466 #endif /* CONFIG_MAGIC_SYSRQ */
8468 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8470 * These functions are only useful for the IA64 MCA handling, or kdb.
8472 * They can only be called when the whole system has been
8473 * stopped - every CPU needs to be quiescent, and no scheduling
8474 * activity can take place. Using them for anything else would
8475 * be a serious bug, and as a result, they aren't even visible
8476 * under any other configuration.
8480 * curr_task - return the current task for a given cpu.
8481 * @cpu: the processor in question.
8483 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8485 struct task_struct
*curr_task(int cpu
)
8487 return cpu_curr(cpu
);
8490 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8494 * set_curr_task - set the current task for a given cpu.
8495 * @cpu: the processor in question.
8496 * @p: the task pointer to set.
8498 * Description: This function must only be used when non-maskable interrupts
8499 * are serviced on a separate stack. It allows the architecture to switch the
8500 * notion of the current task on a cpu in a non-blocking manner. This function
8501 * must be called with all CPU's synchronized, and interrupts disabled, the
8502 * and caller must save the original value of the current task (see
8503 * curr_task() above) and restore that value before reenabling interrupts and
8504 * re-starting the system.
8506 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8508 void set_curr_task(int cpu
, struct task_struct
*p
)
8515 #ifdef CONFIG_FAIR_GROUP_SCHED
8516 static void free_fair_sched_group(struct task_group
*tg
)
8520 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8522 for_each_possible_cpu(i
) {
8524 kfree(tg
->cfs_rq
[i
]);
8534 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8536 struct cfs_rq
*cfs_rq
;
8537 struct sched_entity
*se
;
8540 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8543 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8547 tg
->shares
= NICE_0_LOAD
;
8549 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8551 for_each_possible_cpu(i
) {
8552 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8553 GFP_KERNEL
, cpu_to_node(i
));
8557 se
= kzalloc_node(sizeof(struct sched_entity
),
8558 GFP_KERNEL
, cpu_to_node(i
));
8562 init_cfs_rq(cfs_rq
);
8563 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8574 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8576 struct rq
*rq
= cpu_rq(cpu
);
8577 unsigned long flags
;
8580 * Only empty task groups can be destroyed; so we can speculatively
8581 * check on_list without danger of it being re-added.
8583 if (!tg
->cfs_rq
[cpu
]->on_list
)
8586 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8587 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8588 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8590 #else /* !CONFIG_FAIR_GROUP_SCHED */
8591 static inline void free_fair_sched_group(struct task_group
*tg
)
8596 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8601 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8604 #endif /* CONFIG_FAIR_GROUP_SCHED */
8606 #ifdef CONFIG_RT_GROUP_SCHED
8607 static void free_rt_sched_group(struct task_group
*tg
)
8612 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8614 for_each_possible_cpu(i
) {
8616 kfree(tg
->rt_rq
[i
]);
8618 kfree(tg
->rt_se
[i
]);
8626 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8628 struct rt_rq
*rt_rq
;
8629 struct sched_rt_entity
*rt_se
;
8632 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8635 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8639 init_rt_bandwidth(&tg
->rt_bandwidth
,
8640 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8642 for_each_possible_cpu(i
) {
8643 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8644 GFP_KERNEL
, cpu_to_node(i
));
8648 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8649 GFP_KERNEL
, cpu_to_node(i
));
8653 init_rt_rq(rt_rq
, cpu_rq(i
));
8654 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8655 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8665 #else /* !CONFIG_RT_GROUP_SCHED */
8666 static inline void free_rt_sched_group(struct task_group
*tg
)
8671 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8675 #endif /* CONFIG_RT_GROUP_SCHED */
8677 #ifdef CONFIG_CGROUP_SCHED
8678 static void free_sched_group(struct task_group
*tg
)
8680 free_fair_sched_group(tg
);
8681 free_rt_sched_group(tg
);
8686 /* allocate runqueue etc for a new task group */
8687 struct task_group
*sched_create_group(struct task_group
*parent
)
8689 struct task_group
*tg
;
8690 unsigned long flags
;
8692 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8694 return ERR_PTR(-ENOMEM
);
8696 if (!alloc_fair_sched_group(tg
, parent
))
8699 if (!alloc_rt_sched_group(tg
, parent
))
8702 spin_lock_irqsave(&task_group_lock
, flags
);
8703 list_add_rcu(&tg
->list
, &task_groups
);
8705 WARN_ON(!parent
); /* root should already exist */
8707 tg
->parent
= parent
;
8708 INIT_LIST_HEAD(&tg
->children
);
8709 list_add_rcu(&tg
->siblings
, &parent
->children
);
8710 spin_unlock_irqrestore(&task_group_lock
, flags
);
8715 free_sched_group(tg
);
8716 return ERR_PTR(-ENOMEM
);
8719 /* rcu callback to free various structures associated with a task group */
8720 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8722 /* now it should be safe to free those cfs_rqs */
8723 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8726 /* Destroy runqueue etc associated with a task group */
8727 void sched_destroy_group(struct task_group
*tg
)
8729 unsigned long flags
;
8732 /* end participation in shares distribution */
8733 for_each_possible_cpu(i
)
8734 unregister_fair_sched_group(tg
, i
);
8736 spin_lock_irqsave(&task_group_lock
, flags
);
8737 list_del_rcu(&tg
->list
);
8738 list_del_rcu(&tg
->siblings
);
8739 spin_unlock_irqrestore(&task_group_lock
, flags
);
8741 /* wait for possible concurrent references to cfs_rqs complete */
8742 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8745 /* change task's runqueue when it moves between groups.
8746 * The caller of this function should have put the task in its new group
8747 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8748 * reflect its new group.
8750 void sched_move_task(struct task_struct
*tsk
)
8753 unsigned long flags
;
8756 rq
= task_rq_lock(tsk
, &flags
);
8758 running
= task_current(rq
, tsk
);
8762 dequeue_task(rq
, tsk
, 0);
8763 if (unlikely(running
))
8764 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8766 #ifdef CONFIG_FAIR_GROUP_SCHED
8767 if (tsk
->sched_class
->task_move_group
)
8768 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8771 set_task_rq(tsk
, task_cpu(tsk
));
8773 if (unlikely(running
))
8774 tsk
->sched_class
->set_curr_task(rq
);
8776 enqueue_task(rq
, tsk
, 0);
8778 task_rq_unlock(rq
, tsk
, &flags
);
8780 #endif /* CONFIG_CGROUP_SCHED */
8782 #ifdef CONFIG_FAIR_GROUP_SCHED
8783 static DEFINE_MUTEX(shares_mutex
);
8785 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8788 unsigned long flags
;
8791 * We can't change the weight of the root cgroup.
8796 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8798 mutex_lock(&shares_mutex
);
8799 if (tg
->shares
== shares
)
8802 tg
->shares
= shares
;
8803 for_each_possible_cpu(i
) {
8804 struct rq
*rq
= cpu_rq(i
);
8805 struct sched_entity
*se
;
8808 /* Propagate contribution to hierarchy */
8809 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8810 for_each_sched_entity(se
)
8811 update_cfs_shares(group_cfs_rq(se
));
8812 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8816 mutex_unlock(&shares_mutex
);
8820 unsigned long sched_group_shares(struct task_group
*tg
)
8826 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8827 static unsigned long to_ratio(u64 period
, u64 runtime
)
8829 if (runtime
== RUNTIME_INF
)
8832 return div64_u64(runtime
<< 20, period
);
8836 #ifdef CONFIG_RT_GROUP_SCHED
8838 * Ensure that the real time constraints are schedulable.
8840 static DEFINE_MUTEX(rt_constraints_mutex
);
8842 /* Must be called with tasklist_lock held */
8843 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8845 struct task_struct
*g
, *p
;
8847 do_each_thread(g
, p
) {
8848 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8850 } while_each_thread(g
, p
);
8855 struct rt_schedulable_data
{
8856 struct task_group
*tg
;
8861 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
8863 struct rt_schedulable_data
*d
= data
;
8864 struct task_group
*child
;
8865 unsigned long total
, sum
= 0;
8866 u64 period
, runtime
;
8868 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8869 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8872 period
= d
->rt_period
;
8873 runtime
= d
->rt_runtime
;
8877 * Cannot have more runtime than the period.
8879 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8883 * Ensure we don't starve existing RT tasks.
8885 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8888 total
= to_ratio(period
, runtime
);
8891 * Nobody can have more than the global setting allows.
8893 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8897 * The sum of our children's runtime should not exceed our own.
8899 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8900 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8901 runtime
= child
->rt_bandwidth
.rt_runtime
;
8903 if (child
== d
->tg
) {
8904 period
= d
->rt_period
;
8905 runtime
= d
->rt_runtime
;
8908 sum
+= to_ratio(period
, runtime
);
8917 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8921 struct rt_schedulable_data data
= {
8923 .rt_period
= period
,
8924 .rt_runtime
= runtime
,
8928 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
8934 static int tg_set_rt_bandwidth(struct task_group
*tg
,
8935 u64 rt_period
, u64 rt_runtime
)
8939 mutex_lock(&rt_constraints_mutex
);
8940 read_lock(&tasklist_lock
);
8941 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8945 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8946 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8947 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8949 for_each_possible_cpu(i
) {
8950 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8952 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8953 rt_rq
->rt_runtime
= rt_runtime
;
8954 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8956 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8958 read_unlock(&tasklist_lock
);
8959 mutex_unlock(&rt_constraints_mutex
);
8964 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8966 u64 rt_runtime
, rt_period
;
8968 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8969 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8970 if (rt_runtime_us
< 0)
8971 rt_runtime
= RUNTIME_INF
;
8973 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8976 long sched_group_rt_runtime(struct task_group
*tg
)
8980 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8983 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8984 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8985 return rt_runtime_us
;
8988 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8990 u64 rt_runtime
, rt_period
;
8992 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8993 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8998 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
9001 long sched_group_rt_period(struct task_group
*tg
)
9005 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9006 do_div(rt_period_us
, NSEC_PER_USEC
);
9007 return rt_period_us
;
9010 static int sched_rt_global_constraints(void)
9012 u64 runtime
, period
;
9015 if (sysctl_sched_rt_period
<= 0)
9018 runtime
= global_rt_runtime();
9019 period
= global_rt_period();
9022 * Sanity check on the sysctl variables.
9024 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9027 mutex_lock(&rt_constraints_mutex
);
9028 read_lock(&tasklist_lock
);
9029 ret
= __rt_schedulable(NULL
, 0, 0);
9030 read_unlock(&tasklist_lock
);
9031 mutex_unlock(&rt_constraints_mutex
);
9036 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9038 /* Don't accept realtime tasks when there is no way for them to run */
9039 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9045 #else /* !CONFIG_RT_GROUP_SCHED */
9046 static int sched_rt_global_constraints(void)
9048 unsigned long flags
;
9051 if (sysctl_sched_rt_period
<= 0)
9055 * There's always some RT tasks in the root group
9056 * -- migration, kstopmachine etc..
9058 if (sysctl_sched_rt_runtime
== 0)
9061 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9062 for_each_possible_cpu(i
) {
9063 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9065 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
9066 rt_rq
->rt_runtime
= global_rt_runtime();
9067 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
9069 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9073 #endif /* CONFIG_RT_GROUP_SCHED */
9075 int sched_rt_handler(struct ctl_table
*table
, int write
,
9076 void __user
*buffer
, size_t *lenp
,
9080 int old_period
, old_runtime
;
9081 static DEFINE_MUTEX(mutex
);
9084 old_period
= sysctl_sched_rt_period
;
9085 old_runtime
= sysctl_sched_rt_runtime
;
9087 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
9089 if (!ret
&& write
) {
9090 ret
= sched_rt_global_constraints();
9092 sysctl_sched_rt_period
= old_period
;
9093 sysctl_sched_rt_runtime
= old_runtime
;
9095 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9096 def_rt_bandwidth
.rt_period
=
9097 ns_to_ktime(global_rt_period());
9100 mutex_unlock(&mutex
);
9105 #ifdef CONFIG_CGROUP_SCHED
9107 /* return corresponding task_group object of a cgroup */
9108 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9110 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9111 struct task_group
, css
);
9114 static struct cgroup_subsys_state
*
9115 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9117 struct task_group
*tg
, *parent
;
9119 if (!cgrp
->parent
) {
9120 /* This is early initialization for the top cgroup */
9121 return &root_task_group
.css
;
9124 parent
= cgroup_tg(cgrp
->parent
);
9125 tg
= sched_create_group(parent
);
9127 return ERR_PTR(-ENOMEM
);
9133 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9135 struct task_group
*tg
= cgroup_tg(cgrp
);
9137 sched_destroy_group(tg
);
9141 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
9143 #ifdef CONFIG_RT_GROUP_SCHED
9144 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9147 /* We don't support RT-tasks being in separate groups */
9148 if (tsk
->sched_class
!= &fair_sched_class
)
9155 cpu_cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
9157 sched_move_task(tsk
);
9161 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9162 struct cgroup
*old_cgrp
, struct task_struct
*task
)
9165 * cgroup_exit() is called in the copy_process() failure path.
9166 * Ignore this case since the task hasn't ran yet, this avoids
9167 * trying to poke a half freed task state from generic code.
9169 if (!(task
->flags
& PF_EXITING
))
9172 sched_move_task(task
);
9175 #ifdef CONFIG_FAIR_GROUP_SCHED
9176 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9179 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
9182 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9184 struct task_group
*tg
= cgroup_tg(cgrp
);
9186 return (u64
) scale_load_down(tg
->shares
);
9189 #ifdef CONFIG_CFS_BANDWIDTH
9190 static DEFINE_MUTEX(cfs_constraints_mutex
);
9192 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
9193 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
9195 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
9197 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
9199 int i
, ret
= 0, runtime_enabled
;
9200 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9202 if (tg
== &root_task_group
)
9206 * Ensure we have at some amount of bandwidth every period. This is
9207 * to prevent reaching a state of large arrears when throttled via
9208 * entity_tick() resulting in prolonged exit starvation.
9210 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
9214 * Likewise, bound things on the otherside by preventing insane quota
9215 * periods. This also allows us to normalize in computing quota
9218 if (period
> max_cfs_quota_period
)
9221 mutex_lock(&cfs_constraints_mutex
);
9222 ret
= __cfs_schedulable(tg
, period
, quota
);
9226 runtime_enabled
= quota
!= RUNTIME_INF
;
9227 raw_spin_lock_irq(&cfs_b
->lock
);
9228 cfs_b
->period
= ns_to_ktime(period
);
9229 cfs_b
->quota
= quota
;
9231 __refill_cfs_bandwidth_runtime(cfs_b
);
9232 /* restart the period timer (if active) to handle new period expiry */
9233 if (runtime_enabled
&& cfs_b
->timer_active
) {
9234 /* force a reprogram */
9235 cfs_b
->timer_active
= 0;
9236 __start_cfs_bandwidth(cfs_b
);
9238 raw_spin_unlock_irq(&cfs_b
->lock
);
9240 for_each_possible_cpu(i
) {
9241 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
9242 struct rq
*rq
= rq_of(cfs_rq
);
9244 raw_spin_lock_irq(&rq
->lock
);
9245 cfs_rq
->runtime_enabled
= runtime_enabled
;
9246 cfs_rq
->runtime_remaining
= 0;
9248 if (cfs_rq_throttled(cfs_rq
))
9249 unthrottle_cfs_rq(cfs_rq
);
9250 raw_spin_unlock_irq(&rq
->lock
);
9253 mutex_unlock(&cfs_constraints_mutex
);
9258 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
9262 period
= ktime_to_ns(tg_cfs_bandwidth(tg
)->period
);
9263 if (cfs_quota_us
< 0)
9264 quota
= RUNTIME_INF
;
9266 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
9268 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9271 long tg_get_cfs_quota(struct task_group
*tg
)
9275 if (tg_cfs_bandwidth(tg
)->quota
== RUNTIME_INF
)
9278 quota_us
= tg_cfs_bandwidth(tg
)->quota
;
9279 do_div(quota_us
, NSEC_PER_USEC
);
9284 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
9288 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
9289 quota
= tg_cfs_bandwidth(tg
)->quota
;
9294 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9297 long tg_get_cfs_period(struct task_group
*tg
)
9301 cfs_period_us
= ktime_to_ns(tg_cfs_bandwidth(tg
)->period
);
9302 do_div(cfs_period_us
, NSEC_PER_USEC
);
9304 return cfs_period_us
;
9307 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
9309 return tg_get_cfs_quota(cgroup_tg(cgrp
));
9312 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9315 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
9318 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9320 return tg_get_cfs_period(cgroup_tg(cgrp
));
9323 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9326 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
9329 struct cfs_schedulable_data
{
9330 struct task_group
*tg
;
9335 * normalize group quota/period to be quota/max_period
9336 * note: units are usecs
9338 static u64
normalize_cfs_quota(struct task_group
*tg
,
9339 struct cfs_schedulable_data
*d
)
9347 period
= tg_get_cfs_period(tg
);
9348 quota
= tg_get_cfs_quota(tg
);
9351 /* note: these should typically be equivalent */
9352 if (quota
== RUNTIME_INF
|| quota
== -1)
9355 return to_ratio(period
, quota
);
9358 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
9360 struct cfs_schedulable_data
*d
= data
;
9361 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9362 s64 quota
= 0, parent_quota
= -1;
9365 quota
= RUNTIME_INF
;
9367 struct cfs_bandwidth
*parent_b
= tg_cfs_bandwidth(tg
->parent
);
9369 quota
= normalize_cfs_quota(tg
, d
);
9370 parent_quota
= parent_b
->hierarchal_quota
;
9373 * ensure max(child_quota) <= parent_quota, inherit when no
9376 if (quota
== RUNTIME_INF
)
9377 quota
= parent_quota
;
9378 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
9381 cfs_b
->hierarchal_quota
= quota
;
9386 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
9389 struct cfs_schedulable_data data
= {
9395 if (quota
!= RUNTIME_INF
) {
9396 do_div(data
.period
, NSEC_PER_USEC
);
9397 do_div(data
.quota
, NSEC_PER_USEC
);
9401 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
9407 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9408 struct cgroup_map_cb
*cb
)
9410 struct task_group
*tg
= cgroup_tg(cgrp
);
9411 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9413 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
9414 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
9415 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
9419 #endif /* CONFIG_CFS_BANDWIDTH */
9420 #endif /* CONFIG_FAIR_GROUP_SCHED */
9422 #ifdef CONFIG_RT_GROUP_SCHED
9423 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9426 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9429 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9431 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9434 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9437 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9440 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9442 return sched_group_rt_period(cgroup_tg(cgrp
));
9444 #endif /* CONFIG_RT_GROUP_SCHED */
9446 static struct cftype cpu_files
[] = {
9447 #ifdef CONFIG_FAIR_GROUP_SCHED
9450 .read_u64
= cpu_shares_read_u64
,
9451 .write_u64
= cpu_shares_write_u64
,
9454 #ifdef CONFIG_CFS_BANDWIDTH
9456 .name
= "cfs_quota_us",
9457 .read_s64
= cpu_cfs_quota_read_s64
,
9458 .write_s64
= cpu_cfs_quota_write_s64
,
9461 .name
= "cfs_period_us",
9462 .read_u64
= cpu_cfs_period_read_u64
,
9463 .write_u64
= cpu_cfs_period_write_u64
,
9467 .read_map
= cpu_stats_show
,
9470 #ifdef CONFIG_RT_GROUP_SCHED
9472 .name
= "rt_runtime_us",
9473 .read_s64
= cpu_rt_runtime_read
,
9474 .write_s64
= cpu_rt_runtime_write
,
9477 .name
= "rt_period_us",
9478 .read_u64
= cpu_rt_period_read_uint
,
9479 .write_u64
= cpu_rt_period_write_uint
,
9484 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9486 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9489 struct cgroup_subsys cpu_cgroup_subsys
= {
9491 .create
= cpu_cgroup_create
,
9492 .destroy
= cpu_cgroup_destroy
,
9493 .can_attach_task
= cpu_cgroup_can_attach_task
,
9494 .attach_task
= cpu_cgroup_attach_task
,
9495 .exit
= cpu_cgroup_exit
,
9496 .populate
= cpu_cgroup_populate
,
9497 .subsys_id
= cpu_cgroup_subsys_id
,
9501 #endif /* CONFIG_CGROUP_SCHED */
9503 #ifdef CONFIG_CGROUP_CPUACCT
9506 * CPU accounting code for task groups.
9508 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9509 * (balbir@in.ibm.com).
9512 /* track cpu usage of a group of tasks and its child groups */
9514 struct cgroup_subsys_state css
;
9515 /* cpuusage holds pointer to a u64-type object on every cpu */
9516 u64 __percpu
*cpuusage
;
9517 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9518 struct cpuacct
*parent
;
9521 struct cgroup_subsys cpuacct_subsys
;
9523 /* return cpu accounting group corresponding to this container */
9524 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9526 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9527 struct cpuacct
, css
);
9530 /* return cpu accounting group to which this task belongs */
9531 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9533 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9534 struct cpuacct
, css
);
9537 /* create a new cpu accounting group */
9538 static struct cgroup_subsys_state
*cpuacct_create(
9539 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9541 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9547 ca
->cpuusage
= alloc_percpu(u64
);
9551 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9552 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9553 goto out_free_counters
;
9556 ca
->parent
= cgroup_ca(cgrp
->parent
);
9562 percpu_counter_destroy(&ca
->cpustat
[i
]);
9563 free_percpu(ca
->cpuusage
);
9567 return ERR_PTR(-ENOMEM
);
9570 /* destroy an existing cpu accounting group */
9572 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9574 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9577 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9578 percpu_counter_destroy(&ca
->cpustat
[i
]);
9579 free_percpu(ca
->cpuusage
);
9583 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9585 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9588 #ifndef CONFIG_64BIT
9590 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9592 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9594 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9602 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9604 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9606 #ifndef CONFIG_64BIT
9608 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9610 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9612 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9618 /* return total cpu usage (in nanoseconds) of a group */
9619 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9621 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9622 u64 totalcpuusage
= 0;
9625 for_each_present_cpu(i
)
9626 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9628 return totalcpuusage
;
9631 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9634 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9643 for_each_present_cpu(i
)
9644 cpuacct_cpuusage_write(ca
, i
, 0);
9650 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9653 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9657 for_each_present_cpu(i
) {
9658 percpu
= cpuacct_cpuusage_read(ca
, i
);
9659 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9661 seq_printf(m
, "\n");
9665 static const char *cpuacct_stat_desc
[] = {
9666 [CPUACCT_STAT_USER
] = "user",
9667 [CPUACCT_STAT_SYSTEM
] = "system",
9670 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9671 struct cgroup_map_cb
*cb
)
9673 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9676 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9677 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9678 val
= cputime64_to_clock_t(val
);
9679 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9684 static struct cftype files
[] = {
9687 .read_u64
= cpuusage_read
,
9688 .write_u64
= cpuusage_write
,
9691 .name
= "usage_percpu",
9692 .read_seq_string
= cpuacct_percpu_seq_read
,
9696 .read_map
= cpuacct_stats_show
,
9700 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9702 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9706 * charge this task's execution time to its accounting group.
9708 * called with rq->lock held.
9710 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9715 if (unlikely(!cpuacct_subsys
.active
))
9718 cpu
= task_cpu(tsk
);
9724 for (; ca
; ca
= ca
->parent
) {
9725 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9726 *cpuusage
+= cputime
;
9733 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9734 * in cputime_t units. As a result, cpuacct_update_stats calls
9735 * percpu_counter_add with values large enough to always overflow the
9736 * per cpu batch limit causing bad SMP scalability.
9738 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9739 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9740 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9743 #define CPUACCT_BATCH \
9744 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9746 #define CPUACCT_BATCH 0
9750 * Charge the system/user time to the task's accounting group.
9752 static void cpuacct_update_stats(struct task_struct
*tsk
,
9753 enum cpuacct_stat_index idx
, cputime_t val
)
9756 int batch
= CPUACCT_BATCH
;
9758 if (unlikely(!cpuacct_subsys
.active
))
9765 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9771 struct cgroup_subsys cpuacct_subsys
= {
9773 .create
= cpuacct_create
,
9774 .destroy
= cpuacct_destroy
,
9775 .populate
= cpuacct_populate
,
9776 .subsys_id
= cpuacct_subsys_id
,
9778 #endif /* CONFIG_CGROUP_CPUACCT */