4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/init_task.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
83 #include "sched_cpupri.h"
84 #include "workqueue_sched.h"
85 #include "sched_autogroup.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
91 * Convert user-nice values [ -20 ... 0 ... 19 ]
92 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
95 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
96 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
97 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
100 * 'User priority' is the nice value converted to something we
101 * can work with better when scaling various scheduler parameters,
102 * it's a [ 0 ... 39 ] range.
104 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
105 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
106 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
109 * Helpers for converting nanosecond timing to jiffy resolution
111 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
113 #define NICE_0_LOAD SCHED_LOAD_SCALE
114 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
117 * These are the 'tuning knobs' of the scheduler:
119 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
120 * Timeslices get refilled after they expire.
122 #define DEF_TIMESLICE (100 * HZ / 1000)
125 * single value that denotes runtime == period, ie unlimited time.
127 #define RUNTIME_INF ((u64)~0ULL)
129 static inline int rt_policy(int policy
)
131 if (policy
== SCHED_FIFO
|| policy
== SCHED_RR
)
136 static inline int task_has_rt_policy(struct task_struct
*p
)
138 return rt_policy(p
->policy
);
142 * This is the priority-queue data structure of the RT scheduling class:
144 struct rt_prio_array
{
145 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
146 struct list_head queue
[MAX_RT_PRIO
];
149 struct rt_bandwidth
{
150 /* nests inside the rq lock: */
151 raw_spinlock_t rt_runtime_lock
;
154 struct hrtimer rt_period_timer
;
157 static struct rt_bandwidth def_rt_bandwidth
;
159 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
161 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
163 struct rt_bandwidth
*rt_b
=
164 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
170 now
= hrtimer_cb_get_time(timer
);
171 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
176 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
179 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
183 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
185 rt_b
->rt_period
= ns_to_ktime(period
);
186 rt_b
->rt_runtime
= runtime
;
188 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
190 hrtimer_init(&rt_b
->rt_period_timer
,
191 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
192 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
195 static inline int rt_bandwidth_enabled(void)
197 return sysctl_sched_rt_runtime
>= 0;
200 static void start_bandwidth_timer(struct hrtimer
*period_timer
, ktime_t period
)
203 ktime_t soft
, hard
, now
;
206 if (hrtimer_active(period_timer
))
209 now
= hrtimer_cb_get_time(period_timer
);
210 hrtimer_forward(period_timer
, now
, period
);
212 soft
= hrtimer_get_softexpires(period_timer
);
213 hard
= hrtimer_get_expires(period_timer
);
214 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
215 __hrtimer_start_range_ns(period_timer
, soft
, delta
,
216 HRTIMER_MODE_ABS_PINNED
, 0);
220 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
222 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
225 if (hrtimer_active(&rt_b
->rt_period_timer
))
228 raw_spin_lock(&rt_b
->rt_runtime_lock
);
229 start_bandwidth_timer(&rt_b
->rt_period_timer
, rt_b
->rt_period
);
230 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
233 #ifdef CONFIG_RT_GROUP_SCHED
234 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
236 hrtimer_cancel(&rt_b
->rt_period_timer
);
241 * sched_domains_mutex serializes calls to init_sched_domains,
242 * detach_destroy_domains and partition_sched_domains.
244 static DEFINE_MUTEX(sched_domains_mutex
);
246 #ifdef CONFIG_CGROUP_SCHED
248 #include <linux/cgroup.h>
252 static LIST_HEAD(task_groups
);
254 struct cfs_bandwidth
{
255 #ifdef CONFIG_CFS_BANDWIDTH
259 s64 hierarchal_quota
;
262 int idle
, timer_active
;
263 struct hrtimer period_timer
, slack_timer
;
264 struct list_head throttled_cfs_rq
;
267 int nr_periods
, nr_throttled
;
272 /* task group related information */
274 struct cgroup_subsys_state css
;
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity
**se
;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq
**cfs_rq
;
281 unsigned long shares
;
283 atomic_t load_weight
;
286 #ifdef CONFIG_RT_GROUP_SCHED
287 struct sched_rt_entity
**rt_se
;
288 struct rt_rq
**rt_rq
;
290 struct rt_bandwidth rt_bandwidth
;
294 struct list_head list
;
296 struct task_group
*parent
;
297 struct list_head siblings
;
298 struct list_head children
;
300 #ifdef CONFIG_SCHED_AUTOGROUP
301 struct autogroup
*autogroup
;
304 struct cfs_bandwidth cfs_bandwidth
;
307 /* task_group_lock serializes the addition/removal of task groups */
308 static DEFINE_SPINLOCK(task_group_lock
);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0 or 1 can cause arithmetics problems.
316 * A weight of a cfs_rq is the sum of weights of which entities
317 * are queued on this cfs_rq, so a weight of a entity should not be
318 * too large, so as the shares value of a task group.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
322 #define MIN_SHARES (1UL << 1)
323 #define MAX_SHARES (1UL << 18)
325 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group root_task_group
;
333 #endif /* CONFIG_CGROUP_SCHED */
335 /* CFS-related fields in a runqueue */
337 struct load_weight load
;
338 unsigned long nr_running
, h_nr_running
;
343 u64 min_vruntime_copy
;
346 struct rb_root tasks_timeline
;
347 struct rb_node
*rb_leftmost
;
349 struct list_head tasks
;
350 struct list_head
*balance_iterator
;
353 * 'curr' points to currently running entity on this cfs_rq.
354 * It is set to NULL otherwise (i.e when none are currently running).
356 struct sched_entity
*curr
, *next
, *last
, *skip
;
358 #ifdef CONFIG_SCHED_DEBUG
359 unsigned int nr_spread_over
;
362 #ifdef CONFIG_FAIR_GROUP_SCHED
363 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
366 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
367 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
368 * (like users, containers etc.)
370 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
371 * list is used during load balance.
374 struct list_head leaf_cfs_rq_list
;
375 struct task_group
*tg
; /* group that "owns" this runqueue */
379 * the part of load.weight contributed by tasks
381 unsigned long task_weight
;
384 * h_load = weight * f(tg)
386 * Where f(tg) is the recursive weight fraction assigned to
389 unsigned long h_load
;
392 * Maintaining per-cpu shares distribution for group scheduling
394 * load_stamp is the last time we updated the load average
395 * load_last is the last time we updated the load average and saw load
396 * load_unacc_exec_time is currently unaccounted execution time
400 u64 load_stamp
, load_last
, load_unacc_exec_time
;
402 unsigned long load_contribution
;
404 #ifdef CONFIG_CFS_BANDWIDTH
407 s64 runtime_remaining
;
409 u64 throttled_timestamp
;
410 int throttled
, throttle_count
;
411 struct list_head throttled_list
;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 #ifdef CONFIG_CFS_BANDWIDTH
418 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
420 return &tg
->cfs_bandwidth
;
423 static inline u64
default_cfs_period(void);
424 static int do_sched_cfs_period_timer(struct cfs_bandwidth
*cfs_b
, int overrun
);
425 static void do_sched_cfs_slack_timer(struct cfs_bandwidth
*cfs_b
);
427 static enum hrtimer_restart
sched_cfs_slack_timer(struct hrtimer
*timer
)
429 struct cfs_bandwidth
*cfs_b
=
430 container_of(timer
, struct cfs_bandwidth
, slack_timer
);
431 do_sched_cfs_slack_timer(cfs_b
);
433 return HRTIMER_NORESTART
;
436 static enum hrtimer_restart
sched_cfs_period_timer(struct hrtimer
*timer
)
438 struct cfs_bandwidth
*cfs_b
=
439 container_of(timer
, struct cfs_bandwidth
, period_timer
);
445 now
= hrtimer_cb_get_time(timer
);
446 overrun
= hrtimer_forward(timer
, now
, cfs_b
->period
);
451 idle
= do_sched_cfs_period_timer(cfs_b
, overrun
);
454 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
457 static void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
459 raw_spin_lock_init(&cfs_b
->lock
);
461 cfs_b
->quota
= RUNTIME_INF
;
462 cfs_b
->period
= ns_to_ktime(default_cfs_period());
464 INIT_LIST_HEAD(&cfs_b
->throttled_cfs_rq
);
465 hrtimer_init(&cfs_b
->period_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
466 cfs_b
->period_timer
.function
= sched_cfs_period_timer
;
467 hrtimer_init(&cfs_b
->slack_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
468 cfs_b
->slack_timer
.function
= sched_cfs_slack_timer
;
471 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
)
473 cfs_rq
->runtime_enabled
= 0;
474 INIT_LIST_HEAD(&cfs_rq
->throttled_list
);
477 /* requires cfs_b->lock, may release to reprogram timer */
478 static void __start_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
481 * The timer may be active because we're trying to set a new bandwidth
482 * period or because we're racing with the tear-down path
483 * (timer_active==0 becomes visible before the hrtimer call-back
484 * terminates). In either case we ensure that it's re-programmed
486 while (unlikely(hrtimer_active(&cfs_b
->period_timer
))) {
487 raw_spin_unlock(&cfs_b
->lock
);
488 /* ensure cfs_b->lock is available while we wait */
489 hrtimer_cancel(&cfs_b
->period_timer
);
491 raw_spin_lock(&cfs_b
->lock
);
492 /* if someone else restarted the timer then we're done */
493 if (cfs_b
->timer_active
)
497 cfs_b
->timer_active
= 1;
498 start_bandwidth_timer(&cfs_b
->period_timer
, cfs_b
->period
);
501 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
)
503 hrtimer_cancel(&cfs_b
->period_timer
);
504 hrtimer_cancel(&cfs_b
->slack_timer
);
507 static void init_cfs_rq_runtime(struct cfs_rq
*cfs_rq
) {}
508 static void init_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
509 static void destroy_cfs_bandwidth(struct cfs_bandwidth
*cfs_b
) {}
511 static inline struct cfs_bandwidth
*tg_cfs_bandwidth(struct task_group
*tg
)
515 #endif /* CONFIG_CFS_BANDWIDTH */
516 #endif /* CONFIG_FAIR_GROUP_SCHED */
518 /* Real-Time classes' related field in a runqueue: */
520 struct rt_prio_array active
;
521 unsigned long rt_nr_running
;
522 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
524 int curr
; /* highest queued rt task prio */
526 int next
; /* next highest */
531 unsigned long rt_nr_migratory
;
532 unsigned long rt_nr_total
;
534 struct plist_head pushable_tasks
;
539 /* Nests inside the rq lock: */
540 raw_spinlock_t rt_runtime_lock
;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 unsigned long rt_nr_boosted
;
546 struct list_head leaf_rt_rq_list
;
547 struct task_group
*tg
;
554 * We add the notion of a root-domain which will be used to define per-domain
555 * variables. Each exclusive cpuset essentially defines an island domain by
556 * fully partitioning the member cpus from any other cpuset. Whenever a new
557 * exclusive cpuset is created, we also create and attach a new root-domain
566 cpumask_var_t online
;
569 * The "RT overload" flag: it gets set if a CPU has more than
570 * one runnable RT task.
572 cpumask_var_t rto_mask
;
573 struct cpupri cpupri
;
577 * By default the system creates a single root-domain with all cpus as
578 * members (mimicking the global state we have today).
580 static struct root_domain def_root_domain
;
582 #endif /* CONFIG_SMP */
585 * This is the main, per-CPU runqueue data structure.
587 * Locking rule: those places that want to lock multiple runqueues
588 * (such as the load balancing or the thread migration code), lock
589 * acquire operations must be ordered by ascending &runqueue.
596 * nr_running and cpu_load should be in the same cacheline because
597 * remote CPUs use both these fields when doing load calculation.
599 unsigned long nr_running
;
600 #define CPU_LOAD_IDX_MAX 5
601 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
602 unsigned long last_load_update_tick
;
605 unsigned char nohz_balance_kick
;
607 int skip_clock_update
;
609 /* capture load from *all* tasks on this cpu: */
610 struct load_weight load
;
611 unsigned long nr_load_updates
;
617 #ifdef CONFIG_FAIR_GROUP_SCHED
618 /* list of leaf cfs_rq on this cpu: */
619 struct list_head leaf_cfs_rq_list
;
621 #ifdef CONFIG_RT_GROUP_SCHED
622 struct list_head leaf_rt_rq_list
;
626 * This is part of a global counter where only the total sum
627 * over all CPUs matters. A task can increase this counter on
628 * one CPU and if it got migrated afterwards it may decrease
629 * it on another CPU. Always updated under the runqueue lock:
631 unsigned long nr_uninterruptible
;
633 struct task_struct
*curr
, *idle
, *stop
;
634 unsigned long next_balance
;
635 struct mm_struct
*prev_mm
;
643 struct root_domain
*rd
;
644 struct sched_domain
*sd
;
646 unsigned long cpu_power
;
648 unsigned char idle_balance
;
649 /* For active balancing */
653 struct cpu_stop_work active_balance_work
;
654 /* cpu of this runqueue: */
664 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
667 #ifdef CONFIG_PARAVIRT
670 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
671 u64 prev_steal_time_rq
;
674 /* calc_load related fields */
675 unsigned long calc_load_update
;
676 long calc_load_active
;
678 #ifdef CONFIG_SCHED_HRTICK
680 int hrtick_csd_pending
;
681 struct call_single_data hrtick_csd
;
683 struct hrtimer hrtick_timer
;
686 #ifdef CONFIG_SCHEDSTATS
688 struct sched_info rq_sched_info
;
689 unsigned long long rq_cpu_time
;
690 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
692 /* sys_sched_yield() stats */
693 unsigned int yld_count
;
695 /* schedule() stats */
696 unsigned int sched_switch
;
697 unsigned int sched_count
;
698 unsigned int sched_goidle
;
700 /* try_to_wake_up() stats */
701 unsigned int ttwu_count
;
702 unsigned int ttwu_local
;
706 struct llist_head wake_list
;
710 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
713 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
715 static inline int cpu_of(struct rq
*rq
)
724 #define rcu_dereference_check_sched_domain(p) \
725 rcu_dereference_check((p), \
726 lockdep_is_held(&sched_domains_mutex))
729 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
730 * See detach_destroy_domains: synchronize_sched for details.
732 * The domain tree of any CPU may only be accessed from within
733 * preempt-disabled sections.
735 #define for_each_domain(cpu, __sd) \
736 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
738 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
739 #define this_rq() (&__get_cpu_var(runqueues))
740 #define task_rq(p) cpu_rq(task_cpu(p))
741 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
742 #define raw_rq() (&__raw_get_cpu_var(runqueues))
744 #ifdef CONFIG_CGROUP_SCHED
747 * Return the group to which this tasks belongs.
749 * We use task_subsys_state_check() and extend the RCU verification with
750 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
751 * task it moves into the cgroup. Therefore by holding either of those locks,
752 * we pin the task to the current cgroup.
754 static inline struct task_group
*task_group(struct task_struct
*p
)
756 struct task_group
*tg
;
757 struct cgroup_subsys_state
*css
;
759 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
760 lockdep_is_held(&p
->pi_lock
) ||
761 lockdep_is_held(&task_rq(p
)->lock
));
762 tg
= container_of(css
, struct task_group
, css
);
764 return autogroup_task_group(p
, tg
);
767 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
768 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
770 #ifdef CONFIG_FAIR_GROUP_SCHED
771 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
772 p
->se
.parent
= task_group(p
)->se
[cpu
];
775 #ifdef CONFIG_RT_GROUP_SCHED
776 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
777 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
781 #else /* CONFIG_CGROUP_SCHED */
783 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
784 static inline struct task_group
*task_group(struct task_struct
*p
)
789 #endif /* CONFIG_CGROUP_SCHED */
791 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
793 static void update_rq_clock(struct rq
*rq
)
797 if (rq
->skip_clock_update
> 0)
800 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
802 update_rq_clock_task(rq
, delta
);
806 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
808 #ifdef CONFIG_SCHED_DEBUG
809 # define const_debug __read_mostly
811 # define const_debug static const
815 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
816 * @cpu: the processor in question.
818 * This interface allows printk to be called with the runqueue lock
819 * held and know whether or not it is OK to wake up the klogd.
821 int runqueue_is_locked(int cpu
)
823 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
827 * Debugging: various feature bits
830 #define SCHED_FEAT(name, enabled) \
831 __SCHED_FEAT_##name ,
834 #include "sched_features.h"
839 #define SCHED_FEAT(name, enabled) \
840 (1UL << __SCHED_FEAT_##name) * enabled |
842 const_debug
unsigned int sysctl_sched_features
=
843 #include "sched_features.h"
848 #ifdef CONFIG_SCHED_DEBUG
849 #define SCHED_FEAT(name, enabled) \
852 static __read_mostly
char *sched_feat_names
[] = {
853 #include "sched_features.h"
859 static int sched_feat_show(struct seq_file
*m
, void *v
)
863 for (i
= 0; sched_feat_names
[i
]; i
++) {
864 if (!(sysctl_sched_features
& (1UL << i
)))
866 seq_printf(m
, "%s ", sched_feat_names
[i
]);
874 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
875 size_t cnt
, loff_t
*ppos
)
885 if (copy_from_user(&buf
, ubuf
, cnt
))
891 if (strncmp(cmp
, "NO_", 3) == 0) {
896 for (i
= 0; sched_feat_names
[i
]; i
++) {
897 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
899 sysctl_sched_features
&= ~(1UL << i
);
901 sysctl_sched_features
|= (1UL << i
);
906 if (!sched_feat_names
[i
])
914 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
916 return single_open(filp
, sched_feat_show
, NULL
);
919 static const struct file_operations sched_feat_fops
= {
920 .open
= sched_feat_open
,
921 .write
= sched_feat_write
,
924 .release
= single_release
,
927 static __init
int sched_init_debug(void)
929 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
934 late_initcall(sched_init_debug
);
938 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
941 * Number of tasks to iterate in a single balance run.
942 * Limited because this is done with IRQs disabled.
944 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
947 * period over which we average the RT time consumption, measured
952 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
955 * period over which we measure -rt task cpu usage in us.
958 unsigned int sysctl_sched_rt_period
= 1000000;
960 static __read_mostly
int scheduler_running
;
963 * part of the period that we allow rt tasks to run in us.
966 int sysctl_sched_rt_runtime
= 950000;
968 static inline u64
global_rt_period(void)
970 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
973 static inline u64
global_rt_runtime(void)
975 if (sysctl_sched_rt_runtime
< 0)
978 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
981 #ifndef prepare_arch_switch
982 # define prepare_arch_switch(next) do { } while (0)
984 #ifndef finish_arch_switch
985 # define finish_arch_switch(prev) do { } while (0)
988 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
990 return rq
->curr
== p
;
993 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
998 return task_current(rq
, p
);
1002 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1003 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1007 * We can optimise this out completely for !SMP, because the
1008 * SMP rebalancing from interrupt is the only thing that cares
1015 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1019 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1020 * We must ensure this doesn't happen until the switch is completely
1026 #ifdef CONFIG_DEBUG_SPINLOCK
1027 /* this is a valid case when another task releases the spinlock */
1028 rq
->lock
.owner
= current
;
1031 * If we are tracking spinlock dependencies then we have to
1032 * fix up the runqueue lock - which gets 'carried over' from
1033 * prev into current:
1035 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
1037 raw_spin_unlock_irq(&rq
->lock
);
1040 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1041 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1045 * We can optimise this out completely for !SMP, because the
1046 * SMP rebalancing from interrupt is the only thing that cares
1051 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1052 raw_spin_unlock_irq(&rq
->lock
);
1054 raw_spin_unlock(&rq
->lock
);
1058 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1062 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1063 * We must ensure this doesn't happen until the switch is completely
1069 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1073 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1076 * __task_rq_lock - lock the rq @p resides on.
1078 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
1079 __acquires(rq
->lock
)
1083 lockdep_assert_held(&p
->pi_lock
);
1087 raw_spin_lock(&rq
->lock
);
1088 if (likely(rq
== task_rq(p
)))
1090 raw_spin_unlock(&rq
->lock
);
1095 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1097 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1098 __acquires(p
->pi_lock
)
1099 __acquires(rq
->lock
)
1104 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
1106 raw_spin_lock(&rq
->lock
);
1107 if (likely(rq
== task_rq(p
)))
1109 raw_spin_unlock(&rq
->lock
);
1110 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
1114 static void __task_rq_unlock(struct rq
*rq
)
1115 __releases(rq
->lock
)
1117 raw_spin_unlock(&rq
->lock
);
1121 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
1122 __releases(rq
->lock
)
1123 __releases(p
->pi_lock
)
1125 raw_spin_unlock(&rq
->lock
);
1126 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
1130 * this_rq_lock - lock this runqueue and disable interrupts.
1132 static struct rq
*this_rq_lock(void)
1133 __acquires(rq
->lock
)
1137 local_irq_disable();
1139 raw_spin_lock(&rq
->lock
);
1144 #ifdef CONFIG_SCHED_HRTICK
1146 * Use HR-timers to deliver accurate preemption points.
1148 * Its all a bit involved since we cannot program an hrt while holding the
1149 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1152 * When we get rescheduled we reprogram the hrtick_timer outside of the
1158 * - enabled by features
1159 * - hrtimer is actually high res
1161 static inline int hrtick_enabled(struct rq
*rq
)
1163 if (!sched_feat(HRTICK
))
1165 if (!cpu_active(cpu_of(rq
)))
1167 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1170 static void hrtick_clear(struct rq
*rq
)
1172 if (hrtimer_active(&rq
->hrtick_timer
))
1173 hrtimer_cancel(&rq
->hrtick_timer
);
1177 * High-resolution timer tick.
1178 * Runs from hardirq context with interrupts disabled.
1180 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1182 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1184 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1186 raw_spin_lock(&rq
->lock
);
1187 update_rq_clock(rq
);
1188 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1189 raw_spin_unlock(&rq
->lock
);
1191 return HRTIMER_NORESTART
;
1196 * called from hardirq (IPI) context
1198 static void __hrtick_start(void *arg
)
1200 struct rq
*rq
= arg
;
1202 raw_spin_lock(&rq
->lock
);
1203 hrtimer_restart(&rq
->hrtick_timer
);
1204 rq
->hrtick_csd_pending
= 0;
1205 raw_spin_unlock(&rq
->lock
);
1209 * Called to set the hrtick timer state.
1211 * called with rq->lock held and irqs disabled
1213 static void hrtick_start(struct rq
*rq
, u64 delay
)
1215 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1216 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1218 hrtimer_set_expires(timer
, time
);
1220 if (rq
== this_rq()) {
1221 hrtimer_restart(timer
);
1222 } else if (!rq
->hrtick_csd_pending
) {
1223 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1224 rq
->hrtick_csd_pending
= 1;
1229 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1231 int cpu
= (int)(long)hcpu
;
1234 case CPU_UP_CANCELED
:
1235 case CPU_UP_CANCELED_FROZEN
:
1236 case CPU_DOWN_PREPARE
:
1237 case CPU_DOWN_PREPARE_FROZEN
:
1239 case CPU_DEAD_FROZEN
:
1240 hrtick_clear(cpu_rq(cpu
));
1247 static __init
void init_hrtick(void)
1249 hotcpu_notifier(hotplug_hrtick
, 0);
1253 * Called to set the hrtick timer state.
1255 * called with rq->lock held and irqs disabled
1257 static void hrtick_start(struct rq
*rq
, u64 delay
)
1259 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1260 HRTIMER_MODE_REL_PINNED
, 0);
1263 static inline void init_hrtick(void)
1266 #endif /* CONFIG_SMP */
1268 static void init_rq_hrtick(struct rq
*rq
)
1271 rq
->hrtick_csd_pending
= 0;
1273 rq
->hrtick_csd
.flags
= 0;
1274 rq
->hrtick_csd
.func
= __hrtick_start
;
1275 rq
->hrtick_csd
.info
= rq
;
1278 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1279 rq
->hrtick_timer
.function
= hrtick
;
1281 #else /* CONFIG_SCHED_HRTICK */
1282 static inline void hrtick_clear(struct rq
*rq
)
1286 static inline void init_rq_hrtick(struct rq
*rq
)
1290 static inline void init_hrtick(void)
1293 #endif /* CONFIG_SCHED_HRTICK */
1296 * resched_task - mark a task 'to be rescheduled now'.
1298 * On UP this means the setting of the need_resched flag, on SMP it
1299 * might also involve a cross-CPU call to trigger the scheduler on
1304 #ifndef tsk_is_polling
1305 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1308 static void resched_task(struct task_struct
*p
)
1312 assert_raw_spin_locked(&task_rq(p
)->lock
);
1314 if (test_tsk_need_resched(p
))
1317 set_tsk_need_resched(p
);
1320 if (cpu
== smp_processor_id())
1323 /* NEED_RESCHED must be visible before we test polling */
1325 if (!tsk_is_polling(p
))
1326 smp_send_reschedule(cpu
);
1329 static void resched_cpu(int cpu
)
1331 struct rq
*rq
= cpu_rq(cpu
);
1332 unsigned long flags
;
1334 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1336 resched_task(cpu_curr(cpu
));
1337 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1342 * In the semi idle case, use the nearest busy cpu for migrating timers
1343 * from an idle cpu. This is good for power-savings.
1345 * We don't do similar optimization for completely idle system, as
1346 * selecting an idle cpu will add more delays to the timers than intended
1347 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1349 int get_nohz_timer_target(void)
1351 int cpu
= smp_processor_id();
1353 struct sched_domain
*sd
;
1356 for_each_domain(cpu
, sd
) {
1357 for_each_cpu(i
, sched_domain_span(sd
)) {
1369 * When add_timer_on() enqueues a timer into the timer wheel of an
1370 * idle CPU then this timer might expire before the next timer event
1371 * which is scheduled to wake up that CPU. In case of a completely
1372 * idle system the next event might even be infinite time into the
1373 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1374 * leaves the inner idle loop so the newly added timer is taken into
1375 * account when the CPU goes back to idle and evaluates the timer
1376 * wheel for the next timer event.
1378 void wake_up_idle_cpu(int cpu
)
1380 struct rq
*rq
= cpu_rq(cpu
);
1382 if (cpu
== smp_processor_id())
1386 * This is safe, as this function is called with the timer
1387 * wheel base lock of (cpu) held. When the CPU is on the way
1388 * to idle and has not yet set rq->curr to idle then it will
1389 * be serialized on the timer wheel base lock and take the new
1390 * timer into account automatically.
1392 if (rq
->curr
!= rq
->idle
)
1396 * We can set TIF_RESCHED on the idle task of the other CPU
1397 * lockless. The worst case is that the other CPU runs the
1398 * idle task through an additional NOOP schedule()
1400 set_tsk_need_resched(rq
->idle
);
1402 /* NEED_RESCHED must be visible before we test polling */
1404 if (!tsk_is_polling(rq
->idle
))
1405 smp_send_reschedule(cpu
);
1408 static inline bool got_nohz_idle_kick(void)
1410 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick
;
1413 #else /* CONFIG_NO_HZ */
1415 static inline bool got_nohz_idle_kick(void)
1420 #endif /* CONFIG_NO_HZ */
1422 static u64
sched_avg_period(void)
1424 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1427 static void sched_avg_update(struct rq
*rq
)
1429 s64 period
= sched_avg_period();
1431 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1433 * Inline assembly required to prevent the compiler
1434 * optimising this loop into a divmod call.
1435 * See __iter_div_u64_rem() for another example of this.
1437 asm("" : "+rm" (rq
->age_stamp
));
1438 rq
->age_stamp
+= period
;
1443 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1445 rq
->rt_avg
+= rt_delta
;
1446 sched_avg_update(rq
);
1449 #else /* !CONFIG_SMP */
1450 static void resched_task(struct task_struct
*p
)
1452 assert_raw_spin_locked(&task_rq(p
)->lock
);
1453 set_tsk_need_resched(p
);
1456 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1460 static void sched_avg_update(struct rq
*rq
)
1463 #endif /* CONFIG_SMP */
1465 #if BITS_PER_LONG == 32
1466 # define WMULT_CONST (~0UL)
1468 # define WMULT_CONST (1UL << 32)
1471 #define WMULT_SHIFT 32
1474 * Shift right and round:
1476 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1479 * delta *= weight / lw
1481 static unsigned long
1482 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1483 struct load_weight
*lw
)
1488 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1489 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1490 * 2^SCHED_LOAD_RESOLUTION.
1492 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
1493 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
1495 tmp
= (u64
)delta_exec
;
1497 if (!lw
->inv_weight
) {
1498 unsigned long w
= scale_load_down(lw
->weight
);
1500 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
1502 else if (unlikely(!w
))
1503 lw
->inv_weight
= WMULT_CONST
;
1505 lw
->inv_weight
= WMULT_CONST
/ w
;
1509 * Check whether we'd overflow the 64-bit multiplication:
1511 if (unlikely(tmp
> WMULT_CONST
))
1512 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1515 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1517 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1520 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1526 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1532 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
1539 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1540 * of tasks with abnormal "nice" values across CPUs the contribution that
1541 * each task makes to its run queue's load is weighted according to its
1542 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1543 * scaled version of the new time slice allocation that they receive on time
1547 #define WEIGHT_IDLEPRIO 3
1548 #define WMULT_IDLEPRIO 1431655765
1551 * Nice levels are multiplicative, with a gentle 10% change for every
1552 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1553 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1554 * that remained on nice 0.
1556 * The "10% effect" is relative and cumulative: from _any_ nice level,
1557 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1558 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1559 * If a task goes up by ~10% and another task goes down by ~10% then
1560 * the relative distance between them is ~25%.)
1562 static const int prio_to_weight
[40] = {
1563 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1564 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1565 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1566 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1567 /* 0 */ 1024, 820, 655, 526, 423,
1568 /* 5 */ 335, 272, 215, 172, 137,
1569 /* 10 */ 110, 87, 70, 56, 45,
1570 /* 15 */ 36, 29, 23, 18, 15,
1574 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1576 * In cases where the weight does not change often, we can use the
1577 * precalculated inverse to speed up arithmetics by turning divisions
1578 * into multiplications:
1580 static const u32 prio_to_wmult
[40] = {
1581 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1582 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1583 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1584 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1585 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1586 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1587 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1588 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1591 /* Time spent by the tasks of the cpu accounting group executing in ... */
1592 enum cpuacct_stat_index
{
1593 CPUACCT_STAT_USER
, /* ... user mode */
1594 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1596 CPUACCT_STAT_NSTATS
,
1599 #ifdef CONFIG_CGROUP_CPUACCT
1600 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1601 static void cpuacct_update_stats(struct task_struct
*tsk
,
1602 enum cpuacct_stat_index idx
, cputime_t val
);
1604 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1605 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1606 enum cpuacct_stat_index idx
, cputime_t val
) {}
1609 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1611 update_load_add(&rq
->load
, load
);
1614 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1616 update_load_sub(&rq
->load
, load
);
1619 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1620 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1621 typedef int (*tg_visitor
)(struct task_group
*, void *);
1624 * Iterate task_group tree rooted at *from, calling @down when first entering a
1625 * node and @up when leaving it for the final time.
1627 * Caller must hold rcu_lock or sufficient equivalent.
1629 static int walk_tg_tree_from(struct task_group
*from
,
1630 tg_visitor down
, tg_visitor up
, void *data
)
1632 struct task_group
*parent
, *child
;
1638 ret
= (*down
)(parent
, data
);
1641 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1648 ret
= (*up
)(parent
, data
);
1649 if (ret
|| parent
== from
)
1653 parent
= parent
->parent
;
1661 * Iterate the full tree, calling @down when first entering a node and @up when
1662 * leaving it for the final time.
1664 * Caller must hold rcu_lock or sufficient equivalent.
1667 static inline int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1669 return walk_tg_tree_from(&root_task_group
, down
, up
, data
);
1672 static int tg_nop(struct task_group
*tg
, void *data
)
1679 /* Used instead of source_load when we know the type == 0 */
1680 static unsigned long weighted_cpuload(const int cpu
)
1682 return cpu_rq(cpu
)->load
.weight
;
1686 * Return a low guess at the load of a migration-source cpu weighted
1687 * according to the scheduling class and "nice" value.
1689 * We want to under-estimate the load of migration sources, to
1690 * balance conservatively.
1692 static unsigned long source_load(int cpu
, int type
)
1694 struct rq
*rq
= cpu_rq(cpu
);
1695 unsigned long total
= weighted_cpuload(cpu
);
1697 if (type
== 0 || !sched_feat(LB_BIAS
))
1700 return min(rq
->cpu_load
[type
-1], total
);
1704 * Return a high guess at the load of a migration-target cpu weighted
1705 * according to the scheduling class and "nice" value.
1707 static unsigned long target_load(int cpu
, int type
)
1709 struct rq
*rq
= cpu_rq(cpu
);
1710 unsigned long total
= weighted_cpuload(cpu
);
1712 if (type
== 0 || !sched_feat(LB_BIAS
))
1715 return max(rq
->cpu_load
[type
-1], total
);
1718 static unsigned long power_of(int cpu
)
1720 return cpu_rq(cpu
)->cpu_power
;
1723 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1725 static unsigned long cpu_avg_load_per_task(int cpu
)
1727 struct rq
*rq
= cpu_rq(cpu
);
1728 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1731 return rq
->load
.weight
/ nr_running
;
1736 #ifdef CONFIG_PREEMPT
1738 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1748 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1749 __releases(this_rq
->lock
)
1750 __acquires(busiest
->lock
)
1751 __acquires(this_rq
->lock
)
1753 raw_spin_unlock(&this_rq
->lock
);
1754 double_rq_lock(this_rq
, busiest
);
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1767 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1768 __releases(this_rq
->lock
)
1769 __acquires(busiest
->lock
)
1770 __acquires(this_rq
->lock
)
1774 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1775 if (busiest
< this_rq
) {
1776 raw_spin_unlock(&this_rq
->lock
);
1777 raw_spin_lock(&busiest
->lock
);
1778 raw_spin_lock_nested(&this_rq
->lock
,
1779 SINGLE_DEPTH_NESTING
);
1782 raw_spin_lock_nested(&busiest
->lock
,
1783 SINGLE_DEPTH_NESTING
);
1788 #endif /* CONFIG_PREEMPT */
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1793 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq
->lock
);
1801 return _double_lock_balance(this_rq
, busiest
);
1804 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1805 __releases(busiest
->lock
)
1807 raw_spin_unlock(&busiest
->lock
);
1808 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1812 * double_rq_lock - safely lock two runqueues
1814 * Note this does not disable interrupts like task_rq_lock,
1815 * you need to do so manually before calling.
1817 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1818 __acquires(rq1
->lock
)
1819 __acquires(rq2
->lock
)
1821 BUG_ON(!irqs_disabled());
1823 raw_spin_lock(&rq1
->lock
);
1824 __acquire(rq2
->lock
); /* Fake it out ;) */
1827 raw_spin_lock(&rq1
->lock
);
1828 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1830 raw_spin_lock(&rq2
->lock
);
1831 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1837 * double_rq_unlock - safely unlock two runqueues
1839 * Note this does not restore interrupts like task_rq_unlock,
1840 * you need to do so manually after calling.
1842 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1843 __releases(rq1
->lock
)
1844 __releases(rq2
->lock
)
1846 raw_spin_unlock(&rq1
->lock
);
1848 raw_spin_unlock(&rq2
->lock
);
1850 __release(rq2
->lock
);
1853 #else /* CONFIG_SMP */
1856 * double_rq_lock - safely lock two runqueues
1858 * Note this does not disable interrupts like task_rq_lock,
1859 * you need to do so manually before calling.
1861 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1862 __acquires(rq1
->lock
)
1863 __acquires(rq2
->lock
)
1865 BUG_ON(!irqs_disabled());
1867 raw_spin_lock(&rq1
->lock
);
1868 __acquire(rq2
->lock
); /* Fake it out ;) */
1872 * double_rq_unlock - safely unlock two runqueues
1874 * Note this does not restore interrupts like task_rq_unlock,
1875 * you need to do so manually after calling.
1877 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1878 __releases(rq1
->lock
)
1879 __releases(rq2
->lock
)
1882 raw_spin_unlock(&rq1
->lock
);
1883 __release(rq2
->lock
);
1888 static void calc_load_account_idle(struct rq
*this_rq
);
1889 static void update_sysctl(void);
1890 static int get_update_sysctl_factor(void);
1891 static void update_cpu_load(struct rq
*this_rq
);
1893 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1895 set_task_rq(p
, cpu
);
1898 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1899 * successfully executed on another CPU. We must ensure that updates of
1900 * per-task data have been completed by this moment.
1903 task_thread_info(p
)->cpu
= cpu
;
1907 static const struct sched_class rt_sched_class
;
1909 #define sched_class_highest (&stop_sched_class)
1910 #define for_each_class(class) \
1911 for (class = sched_class_highest; class; class = class->next)
1913 #include "sched_stats.h"
1915 static void inc_nr_running(struct rq
*rq
)
1920 static void dec_nr_running(struct rq
*rq
)
1925 static void set_load_weight(struct task_struct
*p
)
1927 int prio
= p
->static_prio
- MAX_RT_PRIO
;
1928 struct load_weight
*load
= &p
->se
.load
;
1931 * SCHED_IDLE tasks get minimal weight:
1933 if (p
->policy
== SCHED_IDLE
) {
1934 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
1935 load
->inv_weight
= WMULT_IDLEPRIO
;
1939 load
->weight
= scale_load(prio_to_weight
[prio
]);
1940 load
->inv_weight
= prio_to_wmult
[prio
];
1943 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1945 update_rq_clock(rq
);
1946 sched_info_queued(p
);
1947 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1950 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1952 update_rq_clock(rq
);
1953 sched_info_dequeued(p
);
1954 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1958 * activate_task - move a task to the runqueue.
1960 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1962 if (task_contributes_to_load(p
))
1963 rq
->nr_uninterruptible
--;
1965 enqueue_task(rq
, p
, flags
);
1969 * deactivate_task - remove a task from the runqueue.
1971 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1973 if (task_contributes_to_load(p
))
1974 rq
->nr_uninterruptible
++;
1976 dequeue_task(rq
, p
, flags
);
1979 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1982 * There are no locks covering percpu hardirq/softirq time.
1983 * They are only modified in account_system_vtime, on corresponding CPU
1984 * with interrupts disabled. So, writes are safe.
1985 * They are read and saved off onto struct rq in update_rq_clock().
1986 * This may result in other CPU reading this CPU's irq time and can
1987 * race with irq/account_system_vtime on this CPU. We would either get old
1988 * or new value with a side effect of accounting a slice of irq time to wrong
1989 * task when irq is in progress while we read rq->clock. That is a worthy
1990 * compromise in place of having locks on each irq in account_system_time.
1992 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1993 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1995 static DEFINE_PER_CPU(u64
, irq_start_time
);
1996 static int sched_clock_irqtime
;
1998 void enable_sched_clock_irqtime(void)
2000 sched_clock_irqtime
= 1;
2003 void disable_sched_clock_irqtime(void)
2005 sched_clock_irqtime
= 0;
2008 #ifndef CONFIG_64BIT
2009 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
2011 static inline void irq_time_write_begin(void)
2013 __this_cpu_inc(irq_time_seq
.sequence
);
2017 static inline void irq_time_write_end(void)
2020 __this_cpu_inc(irq_time_seq
.sequence
);
2023 static inline u64
irq_time_read(int cpu
)
2029 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
2030 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
2031 per_cpu(cpu_hardirq_time
, cpu
);
2032 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
2036 #else /* CONFIG_64BIT */
2037 static inline void irq_time_write_begin(void)
2041 static inline void irq_time_write_end(void)
2045 static inline u64
irq_time_read(int cpu
)
2047 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
2049 #endif /* CONFIG_64BIT */
2052 * Called before incrementing preempt_count on {soft,}irq_enter
2053 * and before decrementing preempt_count on {soft,}irq_exit.
2055 void account_system_vtime(struct task_struct
*curr
)
2057 unsigned long flags
;
2061 if (!sched_clock_irqtime
)
2064 local_irq_save(flags
);
2066 cpu
= smp_processor_id();
2067 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
2068 __this_cpu_add(irq_start_time
, delta
);
2070 irq_time_write_begin();
2072 * We do not account for softirq time from ksoftirqd here.
2073 * We want to continue accounting softirq time to ksoftirqd thread
2074 * in that case, so as not to confuse scheduler with a special task
2075 * that do not consume any time, but still wants to run.
2077 if (hardirq_count())
2078 __this_cpu_add(cpu_hardirq_time
, delta
);
2079 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
2080 __this_cpu_add(cpu_softirq_time
, delta
);
2082 irq_time_write_end();
2083 local_irq_restore(flags
);
2085 EXPORT_SYMBOL_GPL(account_system_vtime
);
2087 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2089 #ifdef CONFIG_PARAVIRT
2090 static inline u64
steal_ticks(u64 steal
)
2092 if (unlikely(steal
> NSEC_PER_SEC
))
2093 return div_u64(steal
, TICK_NSEC
);
2095 return __iter_div_u64_rem(steal
, TICK_NSEC
, &steal
);
2099 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
2102 * In theory, the compile should just see 0 here, and optimize out the call
2103 * to sched_rt_avg_update. But I don't trust it...
2105 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2106 s64 steal
= 0, irq_delta
= 0;
2108 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2109 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
2112 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2113 * this case when a previous update_rq_clock() happened inside a
2114 * {soft,}irq region.
2116 * When this happens, we stop ->clock_task and only update the
2117 * prev_irq_time stamp to account for the part that fit, so that a next
2118 * update will consume the rest. This ensures ->clock_task is
2121 * It does however cause some slight miss-attribution of {soft,}irq
2122 * time, a more accurate solution would be to update the irq_time using
2123 * the current rq->clock timestamp, except that would require using
2126 if (irq_delta
> delta
)
2129 rq
->prev_irq_time
+= irq_delta
;
2132 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2133 if (static_branch((¶virt_steal_rq_enabled
))) {
2136 steal
= paravirt_steal_clock(cpu_of(rq
));
2137 steal
-= rq
->prev_steal_time_rq
;
2139 if (unlikely(steal
> delta
))
2142 st
= steal_ticks(steal
);
2143 steal
= st
* TICK_NSEC
;
2145 rq
->prev_steal_time_rq
+= steal
;
2151 rq
->clock_task
+= delta
;
2153 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2154 if ((irq_delta
+ steal
) && sched_feat(NONTASK_POWER
))
2155 sched_rt_avg_update(rq
, irq_delta
+ steal
);
2159 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2160 static int irqtime_account_hi_update(void)
2162 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2163 unsigned long flags
;
2167 local_irq_save(flags
);
2168 latest_ns
= this_cpu_read(cpu_hardirq_time
);
2169 if (nsecs_to_cputime64(latest_ns
) > cpustat
->irq
)
2171 local_irq_restore(flags
);
2175 static int irqtime_account_si_update(void)
2177 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2178 unsigned long flags
;
2182 local_irq_save(flags
);
2183 latest_ns
= this_cpu_read(cpu_softirq_time
);
2184 if (nsecs_to_cputime64(latest_ns
) > cpustat
->softirq
)
2186 local_irq_restore(flags
);
2190 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2192 #define sched_clock_irqtime (0)
2196 #include "sched_idletask.c"
2197 #include "sched_fair.c"
2198 #include "sched_rt.c"
2199 #include "sched_autogroup.c"
2200 #include "sched_stoptask.c"
2201 #ifdef CONFIG_SCHED_DEBUG
2202 # include "sched_debug.c"
2205 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2207 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2208 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2212 * Make it appear like a SCHED_FIFO task, its something
2213 * userspace knows about and won't get confused about.
2215 * Also, it will make PI more or less work without too
2216 * much confusion -- but then, stop work should not
2217 * rely on PI working anyway.
2219 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2221 stop
->sched_class
= &stop_sched_class
;
2224 cpu_rq(cpu
)->stop
= stop
;
2228 * Reset it back to a normal scheduling class so that
2229 * it can die in pieces.
2231 old_stop
->sched_class
= &rt_sched_class
;
2236 * __normal_prio - return the priority that is based on the static prio
2238 static inline int __normal_prio(struct task_struct
*p
)
2240 return p
->static_prio
;
2244 * Calculate the expected normal priority: i.e. priority
2245 * without taking RT-inheritance into account. Might be
2246 * boosted by interactivity modifiers. Changes upon fork,
2247 * setprio syscalls, and whenever the interactivity
2248 * estimator recalculates.
2250 static inline int normal_prio(struct task_struct
*p
)
2254 if (task_has_rt_policy(p
))
2255 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2257 prio
= __normal_prio(p
);
2262 * Calculate the current priority, i.e. the priority
2263 * taken into account by the scheduler. This value might
2264 * be boosted by RT tasks, or might be boosted by
2265 * interactivity modifiers. Will be RT if the task got
2266 * RT-boosted. If not then it returns p->normal_prio.
2268 static int effective_prio(struct task_struct
*p
)
2270 p
->normal_prio
= normal_prio(p
);
2272 * If we are RT tasks or we were boosted to RT priority,
2273 * keep the priority unchanged. Otherwise, update priority
2274 * to the normal priority:
2276 if (!rt_prio(p
->prio
))
2277 return p
->normal_prio
;
2282 * task_curr - is this task currently executing on a CPU?
2283 * @p: the task in question.
2285 inline int task_curr(const struct task_struct
*p
)
2287 return cpu_curr(task_cpu(p
)) == p
;
2290 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2291 const struct sched_class
*prev_class
,
2294 if (prev_class
!= p
->sched_class
) {
2295 if (prev_class
->switched_from
)
2296 prev_class
->switched_from(rq
, p
);
2297 p
->sched_class
->switched_to(rq
, p
);
2298 } else if (oldprio
!= p
->prio
)
2299 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2302 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2304 const struct sched_class
*class;
2306 if (p
->sched_class
== rq
->curr
->sched_class
) {
2307 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2309 for_each_class(class) {
2310 if (class == rq
->curr
->sched_class
)
2312 if (class == p
->sched_class
) {
2313 resched_task(rq
->curr
);
2320 * A queue event has occurred, and we're going to schedule. In
2321 * this case, we can save a useless back to back clock update.
2323 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
2324 rq
->skip_clock_update
= 1;
2329 * Is this task likely cache-hot:
2332 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2336 if (p
->sched_class
!= &fair_sched_class
)
2339 if (unlikely(p
->policy
== SCHED_IDLE
))
2343 * Buddy candidates are cache hot:
2345 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2346 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2347 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2350 if (sysctl_sched_migration_cost
== -1)
2352 if (sysctl_sched_migration_cost
== 0)
2355 delta
= now
- p
->se
.exec_start
;
2357 return delta
< (s64
)sysctl_sched_migration_cost
;
2360 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2362 #ifdef CONFIG_SCHED_DEBUG
2364 * We should never call set_task_cpu() on a blocked task,
2365 * ttwu() will sort out the placement.
2367 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2368 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2370 #ifdef CONFIG_LOCKDEP
2372 * The caller should hold either p->pi_lock or rq->lock, when changing
2373 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2375 * sched_move_task() holds both and thus holding either pins the cgroup,
2376 * see set_task_rq().
2378 * Furthermore, all task_rq users should acquire both locks, see
2381 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2382 lockdep_is_held(&task_rq(p
)->lock
)));
2386 trace_sched_migrate_task(p
, new_cpu
);
2388 if (task_cpu(p
) != new_cpu
) {
2389 p
->se
.nr_migrations
++;
2390 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, NULL
, 0);
2393 __set_task_cpu(p
, new_cpu
);
2396 struct migration_arg
{
2397 struct task_struct
*task
;
2401 static int migration_cpu_stop(void *data
);
2404 * wait_task_inactive - wait for a thread to unschedule.
2406 * If @match_state is nonzero, it's the @p->state value just checked and
2407 * not expected to change. If it changes, i.e. @p might have woken up,
2408 * then return zero. When we succeed in waiting for @p to be off its CPU,
2409 * we return a positive number (its total switch count). If a second call
2410 * a short while later returns the same number, the caller can be sure that
2411 * @p has remained unscheduled the whole time.
2413 * The caller must ensure that the task *will* unschedule sometime soon,
2414 * else this function might spin for a *long* time. This function can't
2415 * be called with interrupts off, or it may introduce deadlock with
2416 * smp_call_function() if an IPI is sent by the same process we are
2417 * waiting to become inactive.
2419 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2421 unsigned long flags
;
2428 * We do the initial early heuristics without holding
2429 * any task-queue locks at all. We'll only try to get
2430 * the runqueue lock when things look like they will
2436 * If the task is actively running on another CPU
2437 * still, just relax and busy-wait without holding
2440 * NOTE! Since we don't hold any locks, it's not
2441 * even sure that "rq" stays as the right runqueue!
2442 * But we don't care, since "task_running()" will
2443 * return false if the runqueue has changed and p
2444 * is actually now running somewhere else!
2446 while (task_running(rq
, p
)) {
2447 if (match_state
&& unlikely(p
->state
!= match_state
))
2453 * Ok, time to look more closely! We need the rq
2454 * lock now, to be *sure*. If we're wrong, we'll
2455 * just go back and repeat.
2457 rq
= task_rq_lock(p
, &flags
);
2458 trace_sched_wait_task(p
);
2459 running
= task_running(rq
, p
);
2462 if (!match_state
|| p
->state
== match_state
)
2463 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2464 task_rq_unlock(rq
, p
, &flags
);
2467 * If it changed from the expected state, bail out now.
2469 if (unlikely(!ncsw
))
2473 * Was it really running after all now that we
2474 * checked with the proper locks actually held?
2476 * Oops. Go back and try again..
2478 if (unlikely(running
)) {
2484 * It's not enough that it's not actively running,
2485 * it must be off the runqueue _entirely_, and not
2488 * So if it was still runnable (but just not actively
2489 * running right now), it's preempted, and we should
2490 * yield - it could be a while.
2492 if (unlikely(on_rq
)) {
2493 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
2495 set_current_state(TASK_UNINTERRUPTIBLE
);
2496 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2501 * Ahh, all good. It wasn't running, and it wasn't
2502 * runnable, which means that it will never become
2503 * running in the future either. We're all done!
2512 * kick_process - kick a running thread to enter/exit the kernel
2513 * @p: the to-be-kicked thread
2515 * Cause a process which is running on another CPU to enter
2516 * kernel-mode, without any delay. (to get signals handled.)
2518 * NOTE: this function doesn't have to take the runqueue lock,
2519 * because all it wants to ensure is that the remote task enters
2520 * the kernel. If the IPI races and the task has been migrated
2521 * to another CPU then no harm is done and the purpose has been
2524 void kick_process(struct task_struct
*p
)
2530 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2531 smp_send_reschedule(cpu
);
2534 EXPORT_SYMBOL_GPL(kick_process
);
2535 #endif /* CONFIG_SMP */
2539 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2541 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2544 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2546 /* Look for allowed, online CPU in same node. */
2547 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2548 if (cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
2551 /* Any allowed, online CPU? */
2552 dest_cpu
= cpumask_any_and(tsk_cpus_allowed(p
), cpu_active_mask
);
2553 if (dest_cpu
< nr_cpu_ids
)
2556 /* No more Mr. Nice Guy. */
2557 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2559 * Don't tell them about moving exiting tasks or
2560 * kernel threads (both mm NULL), since they never
2563 if (p
->mm
&& printk_ratelimit()) {
2564 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2565 task_pid_nr(p
), p
->comm
, cpu
);
2572 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2575 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2577 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2580 * In order not to call set_task_cpu() on a blocking task we need
2581 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2584 * Since this is common to all placement strategies, this lives here.
2586 * [ this allows ->select_task() to simply return task_cpu(p) and
2587 * not worry about this generic constraint ]
2589 if (unlikely(!cpumask_test_cpu(cpu
, tsk_cpus_allowed(p
)) ||
2591 cpu
= select_fallback_rq(task_cpu(p
), p
);
2596 static void update_avg(u64
*avg
, u64 sample
)
2598 s64 diff
= sample
- *avg
;
2604 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2606 #ifdef CONFIG_SCHEDSTATS
2607 struct rq
*rq
= this_rq();
2610 int this_cpu
= smp_processor_id();
2612 if (cpu
== this_cpu
) {
2613 schedstat_inc(rq
, ttwu_local
);
2614 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2616 struct sched_domain
*sd
;
2618 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2620 for_each_domain(this_cpu
, sd
) {
2621 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2622 schedstat_inc(sd
, ttwu_wake_remote
);
2629 if (wake_flags
& WF_MIGRATED
)
2630 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2632 #endif /* CONFIG_SMP */
2634 schedstat_inc(rq
, ttwu_count
);
2635 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2637 if (wake_flags
& WF_SYNC
)
2638 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2640 #endif /* CONFIG_SCHEDSTATS */
2643 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
2645 activate_task(rq
, p
, en_flags
);
2648 /* if a worker is waking up, notify workqueue */
2649 if (p
->flags
& PF_WQ_WORKER
)
2650 wq_worker_waking_up(p
, cpu_of(rq
));
2654 * Mark the task runnable and perform wakeup-preemption.
2657 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2659 trace_sched_wakeup(p
, true);
2660 check_preempt_curr(rq
, p
, wake_flags
);
2662 p
->state
= TASK_RUNNING
;
2664 if (p
->sched_class
->task_woken
)
2665 p
->sched_class
->task_woken(rq
, p
);
2667 if (rq
->idle_stamp
) {
2668 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2669 u64 max
= 2*sysctl_sched_migration_cost
;
2674 update_avg(&rq
->avg_idle
, delta
);
2681 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2684 if (p
->sched_contributes_to_load
)
2685 rq
->nr_uninterruptible
--;
2688 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
2689 ttwu_do_wakeup(rq
, p
, wake_flags
);
2693 * Called in case the task @p isn't fully descheduled from its runqueue,
2694 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2695 * since all we need to do is flip p->state to TASK_RUNNING, since
2696 * the task is still ->on_rq.
2698 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2703 rq
= __task_rq_lock(p
);
2705 ttwu_do_wakeup(rq
, p
, wake_flags
);
2708 __task_rq_unlock(rq
);
2714 static void sched_ttwu_pending(void)
2716 struct rq
*rq
= this_rq();
2717 struct llist_node
*llist
= llist_del_all(&rq
->wake_list
);
2718 struct task_struct
*p
;
2720 raw_spin_lock(&rq
->lock
);
2723 p
= llist_entry(llist
, struct task_struct
, wake_entry
);
2724 llist
= llist_next(llist
);
2725 ttwu_do_activate(rq
, p
, 0);
2728 raw_spin_unlock(&rq
->lock
);
2731 void scheduler_ipi(void)
2733 if (llist_empty(&this_rq()->wake_list
) && !got_nohz_idle_kick())
2737 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2738 * traditionally all their work was done from the interrupt return
2739 * path. Now that we actually do some work, we need to make sure
2742 * Some archs already do call them, luckily irq_enter/exit nest
2745 * Arguably we should visit all archs and update all handlers,
2746 * however a fair share of IPIs are still resched only so this would
2747 * somewhat pessimize the simple resched case.
2750 sched_ttwu_pending();
2753 * Check if someone kicked us for doing the nohz idle load balance.
2755 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
2756 this_rq()->idle_balance
= 1;
2757 raise_softirq_irqoff(SCHED_SOFTIRQ
);
2762 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
2764 if (llist_add(&p
->wake_entry
, &cpu_rq(cpu
)->wake_list
))
2765 smp_send_reschedule(cpu
);
2768 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
2774 rq
= __task_rq_lock(p
);
2776 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2777 ttwu_do_wakeup(rq
, p
, wake_flags
);
2780 __task_rq_unlock(rq
);
2785 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2786 #endif /* CONFIG_SMP */
2788 static void ttwu_queue(struct task_struct
*p
, int cpu
)
2790 struct rq
*rq
= cpu_rq(cpu
);
2792 #if defined(CONFIG_SMP)
2793 if (sched_feat(TTWU_QUEUE
) && cpu
!= smp_processor_id()) {
2794 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
2795 ttwu_queue_remote(p
, cpu
);
2800 raw_spin_lock(&rq
->lock
);
2801 ttwu_do_activate(rq
, p
, 0);
2802 raw_spin_unlock(&rq
->lock
);
2806 * try_to_wake_up - wake up a thread
2807 * @p: the thread to be awakened
2808 * @state: the mask of task states that can be woken
2809 * @wake_flags: wake modifier flags (WF_*)
2811 * Put it on the run-queue if it's not already there. The "current"
2812 * thread is always on the run-queue (except when the actual
2813 * re-schedule is in progress), and as such you're allowed to do
2814 * the simpler "current->state = TASK_RUNNING" to mark yourself
2815 * runnable without the overhead of this.
2817 * Returns %true if @p was woken up, %false if it was already running
2818 * or @state didn't match @p's state.
2821 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2823 unsigned long flags
;
2824 int cpu
, success
= 0;
2827 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2828 if (!(p
->state
& state
))
2831 success
= 1; /* we're going to change ->state */
2834 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2839 * If the owning (remote) cpu is still in the middle of schedule() with
2840 * this task as prev, wait until its done referencing the task.
2843 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2845 * In case the architecture enables interrupts in
2846 * context_switch(), we cannot busy wait, since that
2847 * would lead to deadlocks when an interrupt hits and
2848 * tries to wake up @prev. So bail and do a complete
2851 if (ttwu_activate_remote(p
, wake_flags
))
2858 * Pairs with the smp_wmb() in finish_lock_switch().
2862 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2863 p
->state
= TASK_WAKING
;
2865 if (p
->sched_class
->task_waking
)
2866 p
->sched_class
->task_waking(p
);
2868 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2869 if (task_cpu(p
) != cpu
) {
2870 wake_flags
|= WF_MIGRATED
;
2871 set_task_cpu(p
, cpu
);
2873 #endif /* CONFIG_SMP */
2877 ttwu_stat(p
, cpu
, wake_flags
);
2879 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2885 * try_to_wake_up_local - try to wake up a local task with rq lock held
2886 * @p: the thread to be awakened
2888 * Put @p on the run-queue if it's not already there. The caller must
2889 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2892 static void try_to_wake_up_local(struct task_struct
*p
)
2894 struct rq
*rq
= task_rq(p
);
2896 BUG_ON(rq
!= this_rq());
2897 BUG_ON(p
== current
);
2898 lockdep_assert_held(&rq
->lock
);
2900 if (!raw_spin_trylock(&p
->pi_lock
)) {
2901 raw_spin_unlock(&rq
->lock
);
2902 raw_spin_lock(&p
->pi_lock
);
2903 raw_spin_lock(&rq
->lock
);
2906 if (!(p
->state
& TASK_NORMAL
))
2910 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2912 ttwu_do_wakeup(rq
, p
, 0);
2913 ttwu_stat(p
, smp_processor_id(), 0);
2915 raw_spin_unlock(&p
->pi_lock
);
2919 * wake_up_process - Wake up a specific process
2920 * @p: The process to be woken up.
2922 * Attempt to wake up the nominated process and move it to the set of runnable
2923 * processes. Returns 1 if the process was woken up, 0 if it was already
2926 * It may be assumed that this function implies a write memory barrier before
2927 * changing the task state if and only if any tasks are woken up.
2929 int wake_up_process(struct task_struct
*p
)
2931 return try_to_wake_up(p
, TASK_ALL
, 0);
2933 EXPORT_SYMBOL(wake_up_process
);
2935 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2937 return try_to_wake_up(p
, state
, 0);
2941 * Perform scheduler related setup for a newly forked process p.
2942 * p is forked by current.
2944 * __sched_fork() is basic setup used by init_idle() too:
2946 static void __sched_fork(struct task_struct
*p
)
2951 p
->se
.exec_start
= 0;
2952 p
->se
.sum_exec_runtime
= 0;
2953 p
->se
.prev_sum_exec_runtime
= 0;
2954 p
->se
.nr_migrations
= 0;
2956 INIT_LIST_HEAD(&p
->se
.group_node
);
2958 #ifdef CONFIG_SCHEDSTATS
2959 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2962 INIT_LIST_HEAD(&p
->rt
.run_list
);
2964 #ifdef CONFIG_PREEMPT_NOTIFIERS
2965 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2970 * fork()/clone()-time setup:
2972 void sched_fork(struct task_struct
*p
)
2974 unsigned long flags
;
2975 int cpu
= get_cpu();
2979 * We mark the process as running here. This guarantees that
2980 * nobody will actually run it, and a signal or other external
2981 * event cannot wake it up and insert it on the runqueue either.
2983 p
->state
= TASK_RUNNING
;
2986 * Make sure we do not leak PI boosting priority to the child.
2988 p
->prio
= current
->normal_prio
;
2991 * Revert to default priority/policy on fork if requested.
2993 if (unlikely(p
->sched_reset_on_fork
)) {
2994 if (task_has_rt_policy(p
)) {
2995 p
->policy
= SCHED_NORMAL
;
2996 p
->static_prio
= NICE_TO_PRIO(0);
2998 } else if (PRIO_TO_NICE(p
->static_prio
) < 0)
2999 p
->static_prio
= NICE_TO_PRIO(0);
3001 p
->prio
= p
->normal_prio
= __normal_prio(p
);
3005 * We don't need the reset flag anymore after the fork. It has
3006 * fulfilled its duty:
3008 p
->sched_reset_on_fork
= 0;
3011 if (!rt_prio(p
->prio
))
3012 p
->sched_class
= &fair_sched_class
;
3014 if (p
->sched_class
->task_fork
)
3015 p
->sched_class
->task_fork(p
);
3018 * The child is not yet in the pid-hash so no cgroup attach races,
3019 * and the cgroup is pinned to this child due to cgroup_fork()
3020 * is ran before sched_fork().
3022 * Silence PROVE_RCU.
3024 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3025 set_task_cpu(p
, cpu
);
3026 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3028 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3029 if (likely(sched_info_on()))
3030 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
3032 #if defined(CONFIG_SMP)
3035 #ifdef CONFIG_PREEMPT_COUNT
3036 /* Want to start with kernel preemption disabled. */
3037 task_thread_info(p
)->preempt_count
= 1;
3040 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
3047 * wake_up_new_task - wake up a newly created task for the first time.
3049 * This function will do some initial scheduler statistics housekeeping
3050 * that must be done for every newly created context, then puts the task
3051 * on the runqueue and wakes it.
3053 void wake_up_new_task(struct task_struct
*p
)
3055 unsigned long flags
;
3058 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3061 * Fork balancing, do it here and not earlier because:
3062 * - cpus_allowed can change in the fork path
3063 * - any previously selected cpu might disappear through hotplug
3065 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
3068 rq
= __task_rq_lock(p
);
3069 activate_task(rq
, p
, 0);
3071 trace_sched_wakeup_new(p
, true);
3072 check_preempt_curr(rq
, p
, WF_FORK
);
3074 if (p
->sched_class
->task_woken
)
3075 p
->sched_class
->task_woken(rq
, p
);
3077 task_rq_unlock(rq
, p
, &flags
);
3080 #ifdef CONFIG_PREEMPT_NOTIFIERS
3083 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3084 * @notifier: notifier struct to register
3086 void preempt_notifier_register(struct preempt_notifier
*notifier
)
3088 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
3090 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
3093 * preempt_notifier_unregister - no longer interested in preemption notifications
3094 * @notifier: notifier struct to unregister
3096 * This is safe to call from within a preemption notifier.
3098 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
3100 hlist_del(¬ifier
->link
);
3102 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
3104 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3106 struct preempt_notifier
*notifier
;
3107 struct hlist_node
*node
;
3109 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
3110 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
3114 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3115 struct task_struct
*next
)
3117 struct preempt_notifier
*notifier
;
3118 struct hlist_node
*node
;
3120 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
3121 notifier
->ops
->sched_out(notifier
, next
);
3124 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3126 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
3131 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
3132 struct task_struct
*next
)
3136 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3139 * prepare_task_switch - prepare to switch tasks
3140 * @rq: the runqueue preparing to switch
3141 * @prev: the current task that is being switched out
3142 * @next: the task we are going to switch to.
3144 * This is called with the rq lock held and interrupts off. It must
3145 * be paired with a subsequent finish_task_switch after the context
3148 * prepare_task_switch sets up locking and calls architecture specific
3152 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3153 struct task_struct
*next
)
3155 sched_info_switch(prev
, next
);
3156 perf_event_task_sched_out(prev
, next
);
3157 fire_sched_out_preempt_notifiers(prev
, next
);
3158 prepare_lock_switch(rq
, next
);
3159 prepare_arch_switch(next
);
3160 trace_sched_switch(prev
, next
);
3164 * finish_task_switch - clean up after a task-switch
3165 * @rq: runqueue associated with task-switch
3166 * @prev: the thread we just switched away from.
3168 * finish_task_switch must be called after the context switch, paired
3169 * with a prepare_task_switch call before the context switch.
3170 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3171 * and do any other architecture-specific cleanup actions.
3173 * Note that we may have delayed dropping an mm in context_switch(). If
3174 * so, we finish that here outside of the runqueue lock. (Doing it
3175 * with the lock held can cause deadlocks; see schedule() for
3178 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
3179 __releases(rq
->lock
)
3181 struct mm_struct
*mm
= rq
->prev_mm
;
3187 * A task struct has one reference for the use as "current".
3188 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3189 * schedule one last time. The schedule call will never return, and
3190 * the scheduled task must drop that reference.
3191 * The test for TASK_DEAD must occur while the runqueue locks are
3192 * still held, otherwise prev could be scheduled on another cpu, die
3193 * there before we look at prev->state, and then the reference would
3195 * Manfred Spraul <manfred@colorfullife.com>
3197 prev_state
= prev
->state
;
3198 finish_arch_switch(prev
);
3199 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3200 local_irq_disable();
3201 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3202 perf_event_task_sched_in(prev
, current
);
3203 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3205 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3206 finish_lock_switch(rq
, prev
);
3208 fire_sched_in_preempt_notifiers(current
);
3211 if (unlikely(prev_state
== TASK_DEAD
)) {
3213 * Remove function-return probe instances associated with this
3214 * task and put them back on the free list.
3216 kprobe_flush_task(prev
);
3217 put_task_struct(prev
);
3223 /* assumes rq->lock is held */
3224 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
3226 if (prev
->sched_class
->pre_schedule
)
3227 prev
->sched_class
->pre_schedule(rq
, prev
);
3230 /* rq->lock is NOT held, but preemption is disabled */
3231 static inline void post_schedule(struct rq
*rq
)
3233 if (rq
->post_schedule
) {
3234 unsigned long flags
;
3236 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3237 if (rq
->curr
->sched_class
->post_schedule
)
3238 rq
->curr
->sched_class
->post_schedule(rq
);
3239 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3241 rq
->post_schedule
= 0;
3247 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
3251 static inline void post_schedule(struct rq
*rq
)
3258 * schedule_tail - first thing a freshly forked thread must call.
3259 * @prev: the thread we just switched away from.
3261 asmlinkage
void schedule_tail(struct task_struct
*prev
)
3262 __releases(rq
->lock
)
3264 struct rq
*rq
= this_rq();
3266 finish_task_switch(rq
, prev
);
3269 * FIXME: do we need to worry about rq being invalidated by the
3274 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3275 /* In this case, finish_task_switch does not reenable preemption */
3278 if (current
->set_child_tid
)
3279 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3283 * context_switch - switch to the new MM and the new
3284 * thread's register state.
3287 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3288 struct task_struct
*next
)
3290 struct mm_struct
*mm
, *oldmm
;
3292 prepare_task_switch(rq
, prev
, next
);
3295 oldmm
= prev
->active_mm
;
3297 * For paravirt, this is coupled with an exit in switch_to to
3298 * combine the page table reload and the switch backend into
3301 arch_start_context_switch(prev
);
3304 next
->active_mm
= oldmm
;
3305 atomic_inc(&oldmm
->mm_count
);
3306 enter_lazy_tlb(oldmm
, next
);
3308 switch_mm(oldmm
, mm
, next
);
3311 prev
->active_mm
= NULL
;
3312 rq
->prev_mm
= oldmm
;
3315 * Since the runqueue lock will be released by the next
3316 * task (which is an invalid locking op but in the case
3317 * of the scheduler it's an obvious special-case), so we
3318 * do an early lockdep release here:
3320 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3321 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3324 /* Here we just switch the register state and the stack. */
3325 switch_to(prev
, next
, prev
);
3329 * this_rq must be evaluated again because prev may have moved
3330 * CPUs since it called schedule(), thus the 'rq' on its stack
3331 * frame will be invalid.
3333 finish_task_switch(this_rq(), prev
);
3337 * nr_running, nr_uninterruptible and nr_context_switches:
3339 * externally visible scheduler statistics: current number of runnable
3340 * threads, current number of uninterruptible-sleeping threads, total
3341 * number of context switches performed since bootup.
3343 unsigned long nr_running(void)
3345 unsigned long i
, sum
= 0;
3347 for_each_online_cpu(i
)
3348 sum
+= cpu_rq(i
)->nr_running
;
3353 unsigned long nr_uninterruptible(void)
3355 unsigned long i
, sum
= 0;
3357 for_each_possible_cpu(i
)
3358 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3361 * Since we read the counters lockless, it might be slightly
3362 * inaccurate. Do not allow it to go below zero though:
3364 if (unlikely((long)sum
< 0))
3370 unsigned long long nr_context_switches(void)
3373 unsigned long long sum
= 0;
3375 for_each_possible_cpu(i
)
3376 sum
+= cpu_rq(i
)->nr_switches
;
3381 unsigned long nr_iowait(void)
3383 unsigned long i
, sum
= 0;
3385 for_each_possible_cpu(i
)
3386 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3391 unsigned long nr_iowait_cpu(int cpu
)
3393 struct rq
*this = cpu_rq(cpu
);
3394 return atomic_read(&this->nr_iowait
);
3397 unsigned long this_cpu_load(void)
3399 struct rq
*this = this_rq();
3400 return this->cpu_load
[0];
3404 /* Variables and functions for calc_load */
3405 static atomic_long_t calc_load_tasks
;
3406 static unsigned long calc_load_update
;
3407 unsigned long avenrun
[3];
3408 EXPORT_SYMBOL(avenrun
);
3410 static long calc_load_fold_active(struct rq
*this_rq
)
3412 long nr_active
, delta
= 0;
3414 nr_active
= this_rq
->nr_running
;
3415 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3417 if (nr_active
!= this_rq
->calc_load_active
) {
3418 delta
= nr_active
- this_rq
->calc_load_active
;
3419 this_rq
->calc_load_active
= nr_active
;
3425 static unsigned long
3426 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3429 load
+= active
* (FIXED_1
- exp
);
3430 load
+= 1UL << (FSHIFT
- 1);
3431 return load
>> FSHIFT
;
3436 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3438 * When making the ILB scale, we should try to pull this in as well.
3440 static atomic_long_t calc_load_tasks_idle
;
3442 static void calc_load_account_idle(struct rq
*this_rq
)
3446 delta
= calc_load_fold_active(this_rq
);
3448 atomic_long_add(delta
, &calc_load_tasks_idle
);
3451 static long calc_load_fold_idle(void)
3456 * Its got a race, we don't care...
3458 if (atomic_long_read(&calc_load_tasks_idle
))
3459 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3465 * fixed_power_int - compute: x^n, in O(log n) time
3467 * @x: base of the power
3468 * @frac_bits: fractional bits of @x
3469 * @n: power to raise @x to.
3471 * By exploiting the relation between the definition of the natural power
3472 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3473 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3474 * (where: n_i \elem {0, 1}, the binary vector representing n),
3475 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3476 * of course trivially computable in O(log_2 n), the length of our binary
3479 static unsigned long
3480 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3482 unsigned long result
= 1UL << frac_bits
;
3487 result
+= 1UL << (frac_bits
- 1);
3488 result
>>= frac_bits
;
3494 x
+= 1UL << (frac_bits
- 1);
3502 * a1 = a0 * e + a * (1 - e)
3504 * a2 = a1 * e + a * (1 - e)
3505 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3506 * = a0 * e^2 + a * (1 - e) * (1 + e)
3508 * a3 = a2 * e + a * (1 - e)
3509 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3510 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3514 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3515 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3516 * = a0 * e^n + a * (1 - e^n)
3518 * [1] application of the geometric series:
3521 * S_n := \Sum x^i = -------------
3524 static unsigned long
3525 calc_load_n(unsigned long load
, unsigned long exp
,
3526 unsigned long active
, unsigned int n
)
3529 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3533 * NO_HZ can leave us missing all per-cpu ticks calling
3534 * calc_load_account_active(), but since an idle CPU folds its delta into
3535 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3536 * in the pending idle delta if our idle period crossed a load cycle boundary.
3538 * Once we've updated the global active value, we need to apply the exponential
3539 * weights adjusted to the number of cycles missed.
3541 static void calc_global_nohz(unsigned long ticks
)
3543 long delta
, active
, n
;
3545 if (time_before(jiffies
, calc_load_update
))
3549 * If we crossed a calc_load_update boundary, make sure to fold
3550 * any pending idle changes, the respective CPUs might have
3551 * missed the tick driven calc_load_account_active() update
3554 delta
= calc_load_fold_idle();
3556 atomic_long_add(delta
, &calc_load_tasks
);
3559 * If we were idle for multiple load cycles, apply them.
3561 if (ticks
>= LOAD_FREQ
) {
3562 n
= ticks
/ LOAD_FREQ
;
3564 active
= atomic_long_read(&calc_load_tasks
);
3565 active
= active
> 0 ? active
* FIXED_1
: 0;
3567 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3568 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3569 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3571 calc_load_update
+= n
* LOAD_FREQ
;
3575 * Its possible the remainder of the above division also crosses
3576 * a LOAD_FREQ period, the regular check in calc_global_load()
3577 * which comes after this will take care of that.
3579 * Consider us being 11 ticks before a cycle completion, and us
3580 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3581 * age us 4 cycles, and the test in calc_global_load() will
3582 * pick up the final one.
3586 static void calc_load_account_idle(struct rq
*this_rq
)
3590 static inline long calc_load_fold_idle(void)
3595 static void calc_global_nohz(unsigned long ticks
)
3601 * get_avenrun - get the load average array
3602 * @loads: pointer to dest load array
3603 * @offset: offset to add
3604 * @shift: shift count to shift the result left
3606 * These values are estimates at best, so no need for locking.
3608 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3610 loads
[0] = (avenrun
[0] + offset
) << shift
;
3611 loads
[1] = (avenrun
[1] + offset
) << shift
;
3612 loads
[2] = (avenrun
[2] + offset
) << shift
;
3616 * calc_load - update the avenrun load estimates 10 ticks after the
3617 * CPUs have updated calc_load_tasks.
3619 void calc_global_load(unsigned long ticks
)
3623 calc_global_nohz(ticks
);
3625 if (time_before(jiffies
, calc_load_update
+ 10))
3628 active
= atomic_long_read(&calc_load_tasks
);
3629 active
= active
> 0 ? active
* FIXED_1
: 0;
3631 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3632 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3633 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3635 calc_load_update
+= LOAD_FREQ
;
3639 * Called from update_cpu_load() to periodically update this CPU's
3642 static void calc_load_account_active(struct rq
*this_rq
)
3646 if (time_before(jiffies
, this_rq
->calc_load_update
))
3649 delta
= calc_load_fold_active(this_rq
);
3650 delta
+= calc_load_fold_idle();
3652 atomic_long_add(delta
, &calc_load_tasks
);
3654 this_rq
->calc_load_update
+= LOAD_FREQ
;
3658 * The exact cpuload at various idx values, calculated at every tick would be
3659 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3661 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3662 * on nth tick when cpu may be busy, then we have:
3663 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3664 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3666 * decay_load_missed() below does efficient calculation of
3667 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3668 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3670 * The calculation is approximated on a 128 point scale.
3671 * degrade_zero_ticks is the number of ticks after which load at any
3672 * particular idx is approximated to be zero.
3673 * degrade_factor is a precomputed table, a row for each load idx.
3674 * Each column corresponds to degradation factor for a power of two ticks,
3675 * based on 128 point scale.
3677 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3678 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3680 * With this power of 2 load factors, we can degrade the load n times
3681 * by looking at 1 bits in n and doing as many mult/shift instead of
3682 * n mult/shifts needed by the exact degradation.
3684 #define DEGRADE_SHIFT 7
3685 static const unsigned char
3686 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3687 static const unsigned char
3688 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3689 {0, 0, 0, 0, 0, 0, 0, 0},
3690 {64, 32, 8, 0, 0, 0, 0, 0},
3691 {96, 72, 40, 12, 1, 0, 0},
3692 {112, 98, 75, 43, 15, 1, 0},
3693 {120, 112, 98, 76, 45, 16, 2} };
3696 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3697 * would be when CPU is idle and so we just decay the old load without
3698 * adding any new load.
3700 static unsigned long
3701 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3705 if (!missed_updates
)
3708 if (missed_updates
>= degrade_zero_ticks
[idx
])
3712 return load
>> missed_updates
;
3714 while (missed_updates
) {
3715 if (missed_updates
% 2)
3716 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3718 missed_updates
>>= 1;
3725 * Update rq->cpu_load[] statistics. This function is usually called every
3726 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3727 * every tick. We fix it up based on jiffies.
3729 static void update_cpu_load(struct rq
*this_rq
)
3731 unsigned long this_load
= this_rq
->load
.weight
;
3732 unsigned long curr_jiffies
= jiffies
;
3733 unsigned long pending_updates
;
3736 this_rq
->nr_load_updates
++;
3738 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3739 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3742 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3743 this_rq
->last_load_update_tick
= curr_jiffies
;
3745 /* Update our load: */
3746 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3747 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3748 unsigned long old_load
, new_load
;
3750 /* scale is effectively 1 << i now, and >> i divides by scale */
3752 old_load
= this_rq
->cpu_load
[i
];
3753 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3754 new_load
= this_load
;
3756 * Round up the averaging division if load is increasing. This
3757 * prevents us from getting stuck on 9 if the load is 10, for
3760 if (new_load
> old_load
)
3761 new_load
+= scale
- 1;
3763 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3766 sched_avg_update(this_rq
);
3769 static void update_cpu_load_active(struct rq
*this_rq
)
3771 update_cpu_load(this_rq
);
3773 calc_load_account_active(this_rq
);
3779 * sched_exec - execve() is a valuable balancing opportunity, because at
3780 * this point the task has the smallest effective memory and cache footprint.
3782 void sched_exec(void)
3784 struct task_struct
*p
= current
;
3785 unsigned long flags
;
3788 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3789 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3790 if (dest_cpu
== smp_processor_id())
3793 if (likely(cpu_active(dest_cpu
))) {
3794 struct migration_arg arg
= { p
, dest_cpu
};
3796 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3797 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3801 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3806 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3808 EXPORT_PER_CPU_SYMBOL(kstat
);
3811 * Return any ns on the sched_clock that have not yet been accounted in
3812 * @p in case that task is currently running.
3814 * Called with task_rq_lock() held on @rq.
3816 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3820 if (task_current(rq
, p
)) {
3821 update_rq_clock(rq
);
3822 ns
= rq
->clock_task
- p
->se
.exec_start
;
3830 unsigned long long task_delta_exec(struct task_struct
*p
)
3832 unsigned long flags
;
3836 rq
= task_rq_lock(p
, &flags
);
3837 ns
= do_task_delta_exec(p
, rq
);
3838 task_rq_unlock(rq
, p
, &flags
);
3844 * Return accounted runtime for the task.
3845 * In case the task is currently running, return the runtime plus current's
3846 * pending runtime that have not been accounted yet.
3848 unsigned long long task_sched_runtime(struct task_struct
*p
)
3850 unsigned long flags
;
3854 rq
= task_rq_lock(p
, &flags
);
3855 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3856 task_rq_unlock(rq
, p
, &flags
);
3862 * Account user cpu time to a process.
3863 * @p: the process that the cpu time gets accounted to
3864 * @cputime: the cpu time spent in user space since the last update
3865 * @cputime_scaled: cputime scaled by cpu frequency
3867 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3868 cputime_t cputime_scaled
)
3870 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3872 /* Add user time to process. */
3873 p
->utime
+= cputime
;
3874 p
->utimescaled
+= cputime_scaled
;
3875 account_group_user_time(p
, cputime
);
3877 /* Add user time to cpustat. */
3878 if (TASK_NICE(p
) > 0)
3879 cpustat
->nice
+= (__force cputime64_t
) cputime
;
3881 cpustat
->user
+= (__force cputime64_t
) cputime
;
3883 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3884 /* Account for user time used */
3885 acct_update_integrals(p
);
3889 * Account guest cpu time to a process.
3890 * @p: the process that the cpu time gets accounted to
3891 * @cputime: the cpu time spent in virtual machine since the last update
3892 * @cputime_scaled: cputime scaled by cpu frequency
3894 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3895 cputime_t cputime_scaled
)
3897 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3899 /* Add guest time to process. */
3900 p
->utime
+= cputime
;
3901 p
->utimescaled
+= cputime_scaled
;
3902 account_group_user_time(p
, cputime
);
3903 p
->gtime
+= cputime
;
3905 /* Add guest time to cpustat. */
3906 if (TASK_NICE(p
) > 0) {
3907 cpustat
->nice
+= (__force cputime64_t
) cputime
;
3908 cpustat
->guest_nice
+= (__force cputime64_t
) cputime
;
3910 cpustat
->user
+= (__force cputime64_t
) cputime
;
3911 cpustat
->guest
+= (__force cputime64_t
) cputime
;
3916 * Account system cpu time to a process and desired cpustat field
3917 * @p: the process that the cpu time gets accounted to
3918 * @cputime: the cpu time spent in kernel space since the last update
3919 * @cputime_scaled: cputime scaled by cpu frequency
3920 * @target_cputime64: pointer to cpustat field that has to be updated
3923 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
3924 cputime_t cputime_scaled
, cputime64_t
*target_cputime64
)
3926 /* Add system time to process. */
3927 p
->stime
+= cputime
;
3928 p
->stimescaled
+= cputime_scaled
;
3929 account_group_system_time(p
, cputime
);
3931 /* Add system time to cpustat. */
3932 *target_cputime64
+= (__force cputime64_t
) cputime
;
3933 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3935 /* Account for system time used */
3936 acct_update_integrals(p
);
3940 * Account system cpu time to a process.
3941 * @p: the process that the cpu time gets accounted to
3942 * @hardirq_offset: the offset to subtract from hardirq_count()
3943 * @cputime: the cpu time spent in kernel space since the last update
3944 * @cputime_scaled: cputime scaled by cpu frequency
3946 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3947 cputime_t cputime
, cputime_t cputime_scaled
)
3949 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3950 cputime64_t
*target_cputime64
;
3952 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3953 account_guest_time(p
, cputime
, cputime_scaled
);
3957 if (hardirq_count() - hardirq_offset
)
3958 target_cputime64
= &cpustat
->irq
;
3959 else if (in_serving_softirq())
3960 target_cputime64
= &cpustat
->softirq
;
3962 target_cputime64
= &cpustat
->system
;
3964 __account_system_time(p
, cputime
, cputime_scaled
, target_cputime64
);
3968 * Account for involuntary wait time.
3969 * @cputime: the cpu time spent in involuntary wait
3971 void account_steal_time(cputime_t cputime
)
3973 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3975 cpustat
->steal
+= (__force cputime64_t
) cputime
;
3979 * Account for idle time.
3980 * @cputime: the cpu time spent in idle wait
3982 void account_idle_time(cputime_t cputime
)
3984 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3985 struct rq
*rq
= this_rq();
3987 if (atomic_read(&rq
->nr_iowait
) > 0)
3988 cpustat
->iowait
+= (__force cputime64_t
) cputime
;
3990 cpustat
->idle
+= (__force cputime64_t
) cputime
;
3993 static __always_inline
bool steal_account_process_tick(void)
3995 #ifdef CONFIG_PARAVIRT
3996 if (static_branch(¶virt_steal_enabled
)) {
3999 steal
= paravirt_steal_clock(smp_processor_id());
4000 steal
-= this_rq()->prev_steal_time
;
4002 st
= steal_ticks(steal
);
4003 this_rq()->prev_steal_time
+= st
* TICK_NSEC
;
4005 account_steal_time(st
);
4012 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4014 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4016 * Account a tick to a process and cpustat
4017 * @p: the process that the cpu time gets accounted to
4018 * @user_tick: is the tick from userspace
4019 * @rq: the pointer to rq
4021 * Tick demultiplexing follows the order
4022 * - pending hardirq update
4023 * - pending softirq update
4027 * - check for guest_time
4028 * - else account as system_time
4030 * Check for hardirq is done both for system and user time as there is
4031 * no timer going off while we are on hardirq and hence we may never get an
4032 * opportunity to update it solely in system time.
4033 * p->stime and friends are only updated on system time and not on irq
4034 * softirq as those do not count in task exec_runtime any more.
4036 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
4039 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
4040 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4042 if (steal_account_process_tick())
4045 if (irqtime_account_hi_update()) {
4046 cpustat
->irq
+= (__force cputime64_t
) cputime_one_jiffy
;
4047 } else if (irqtime_account_si_update()) {
4048 cpustat
->softirq
+= (__force cputime64_t
) cputime_one_jiffy
;
4049 } else if (this_cpu_ksoftirqd() == p
) {
4051 * ksoftirqd time do not get accounted in cpu_softirq_time.
4052 * So, we have to handle it separately here.
4053 * Also, p->stime needs to be updated for ksoftirqd.
4055 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
4057 } else if (user_tick
) {
4058 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4059 } else if (p
== rq
->idle
) {
4060 account_idle_time(cputime_one_jiffy
);
4061 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
4062 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4064 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
4069 static void irqtime_account_idle_ticks(int ticks
)
4072 struct rq
*rq
= this_rq();
4074 for (i
= 0; i
< ticks
; i
++)
4075 irqtime_account_process_tick(current
, 0, rq
);
4077 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4078 static void irqtime_account_idle_ticks(int ticks
) {}
4079 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
4081 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4084 * Account a single tick of cpu time.
4085 * @p: the process that the cpu time gets accounted to
4086 * @user_tick: indicates if the tick is a user or a system tick
4088 void account_process_tick(struct task_struct
*p
, int user_tick
)
4090 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
4091 struct rq
*rq
= this_rq();
4093 if (sched_clock_irqtime
) {
4094 irqtime_account_process_tick(p
, user_tick
, rq
);
4098 if (steal_account_process_tick())
4102 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
4103 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
4104 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
4107 account_idle_time(cputime_one_jiffy
);
4111 * Account multiple ticks of steal time.
4112 * @p: the process from which the cpu time has been stolen
4113 * @ticks: number of stolen ticks
4115 void account_steal_ticks(unsigned long ticks
)
4117 account_steal_time(jiffies_to_cputime(ticks
));
4121 * Account multiple ticks of idle time.
4122 * @ticks: number of stolen ticks
4124 void account_idle_ticks(unsigned long ticks
)
4127 if (sched_clock_irqtime
) {
4128 irqtime_account_idle_ticks(ticks
);
4132 account_idle_time(jiffies_to_cputime(ticks
));
4138 * Use precise platform statistics if available:
4140 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4141 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4147 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4149 struct task_cputime cputime
;
4151 thread_group_cputime(p
, &cputime
);
4153 *ut
= cputime
.utime
;
4154 *st
= cputime
.stime
;
4158 #ifndef nsecs_to_cputime
4159 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4162 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4164 cputime_t rtime
, utime
= p
->utime
, total
= utime
+ p
->stime
;
4167 * Use CFS's precise accounting:
4169 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
4172 u64 temp
= (__force u64
) rtime
;
4174 temp
*= (__force u64
) utime
;
4175 do_div(temp
, (__force u32
) total
);
4176 utime
= (__force cputime_t
) temp
;
4181 * Compare with previous values, to keep monotonicity:
4183 p
->prev_utime
= max(p
->prev_utime
, utime
);
4184 p
->prev_stime
= max(p
->prev_stime
, rtime
- p
->prev_utime
);
4186 *ut
= p
->prev_utime
;
4187 *st
= p
->prev_stime
;
4191 * Must be called with siglock held.
4193 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4195 struct signal_struct
*sig
= p
->signal
;
4196 struct task_cputime cputime
;
4197 cputime_t rtime
, utime
, total
;
4199 thread_group_cputime(p
, &cputime
);
4201 total
= cputime
.utime
+ cputime
.stime
;
4202 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
4205 u64 temp
= (__force u64
) rtime
;
4207 temp
*= (__force u64
) cputime
.utime
;
4208 do_div(temp
, (__force u32
) total
);
4209 utime
= (__force cputime_t
) temp
;
4213 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
4214 sig
->prev_stime
= max(sig
->prev_stime
, rtime
- sig
->prev_utime
);
4216 *ut
= sig
->prev_utime
;
4217 *st
= sig
->prev_stime
;
4222 * This function gets called by the timer code, with HZ frequency.
4223 * We call it with interrupts disabled.
4225 void scheduler_tick(void)
4227 int cpu
= smp_processor_id();
4228 struct rq
*rq
= cpu_rq(cpu
);
4229 struct task_struct
*curr
= rq
->curr
;
4233 raw_spin_lock(&rq
->lock
);
4234 update_rq_clock(rq
);
4235 update_cpu_load_active(rq
);
4236 curr
->sched_class
->task_tick(rq
, curr
, 0);
4237 raw_spin_unlock(&rq
->lock
);
4239 perf_event_task_tick();
4242 rq
->idle_balance
= idle_cpu(cpu
);
4243 trigger_load_balance(rq
, cpu
);
4247 notrace
unsigned long get_parent_ip(unsigned long addr
)
4249 if (in_lock_functions(addr
)) {
4250 addr
= CALLER_ADDR2
;
4251 if (in_lock_functions(addr
))
4252 addr
= CALLER_ADDR3
;
4257 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4258 defined(CONFIG_PREEMPT_TRACER))
4260 void __kprobes
add_preempt_count(int val
)
4262 #ifdef CONFIG_DEBUG_PREEMPT
4266 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4269 preempt_count() += val
;
4270 #ifdef CONFIG_DEBUG_PREEMPT
4272 * Spinlock count overflowing soon?
4274 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4277 if (preempt_count() == val
)
4278 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4280 EXPORT_SYMBOL(add_preempt_count
);
4282 void __kprobes
sub_preempt_count(int val
)
4284 #ifdef CONFIG_DEBUG_PREEMPT
4288 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4291 * Is the spinlock portion underflowing?
4293 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4294 !(preempt_count() & PREEMPT_MASK
)))
4298 if (preempt_count() == val
)
4299 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4300 preempt_count() -= val
;
4302 EXPORT_SYMBOL(sub_preempt_count
);
4307 * Print scheduling while atomic bug:
4309 static noinline
void __schedule_bug(struct task_struct
*prev
)
4311 struct pt_regs
*regs
= get_irq_regs();
4313 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4314 prev
->comm
, prev
->pid
, preempt_count());
4316 debug_show_held_locks(prev
);
4318 if (irqs_disabled())
4319 print_irqtrace_events(prev
);
4328 * Various schedule()-time debugging checks and statistics:
4330 static inline void schedule_debug(struct task_struct
*prev
)
4333 * Test if we are atomic. Since do_exit() needs to call into
4334 * schedule() atomically, we ignore that path for now.
4335 * Otherwise, whine if we are scheduling when we should not be.
4337 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4338 __schedule_bug(prev
);
4341 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4343 schedstat_inc(this_rq(), sched_count
);
4346 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4348 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
4349 update_rq_clock(rq
);
4350 prev
->sched_class
->put_prev_task(rq
, prev
);
4354 * Pick up the highest-prio task:
4356 static inline struct task_struct
*
4357 pick_next_task(struct rq
*rq
)
4359 const struct sched_class
*class;
4360 struct task_struct
*p
;
4363 * Optimization: we know that if all tasks are in
4364 * the fair class we can call that function directly:
4366 if (likely(rq
->nr_running
== rq
->cfs
.h_nr_running
)) {
4367 p
= fair_sched_class
.pick_next_task(rq
);
4372 for_each_class(class) {
4373 p
= class->pick_next_task(rq
);
4378 BUG(); /* the idle class will always have a runnable task */
4382 * __schedule() is the main scheduler function.
4384 static void __sched
__schedule(void)
4386 struct task_struct
*prev
, *next
;
4387 unsigned long *switch_count
;
4393 cpu
= smp_processor_id();
4395 rcu_note_context_switch(cpu
);
4398 schedule_debug(prev
);
4400 if (sched_feat(HRTICK
))
4403 raw_spin_lock_irq(&rq
->lock
);
4405 switch_count
= &prev
->nivcsw
;
4406 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4407 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
4408 prev
->state
= TASK_RUNNING
;
4410 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
4414 * If a worker went to sleep, notify and ask workqueue
4415 * whether it wants to wake up a task to maintain
4418 if (prev
->flags
& PF_WQ_WORKER
) {
4419 struct task_struct
*to_wakeup
;
4421 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
4423 try_to_wake_up_local(to_wakeup
);
4426 switch_count
= &prev
->nvcsw
;
4429 pre_schedule(rq
, prev
);
4431 if (unlikely(!rq
->nr_running
))
4432 idle_balance(cpu
, rq
);
4434 put_prev_task(rq
, prev
);
4435 next
= pick_next_task(rq
);
4436 clear_tsk_need_resched(prev
);
4437 rq
->skip_clock_update
= 0;
4439 if (likely(prev
!= next
)) {
4444 context_switch(rq
, prev
, next
); /* unlocks the rq */
4446 * The context switch have flipped the stack from under us
4447 * and restored the local variables which were saved when
4448 * this task called schedule() in the past. prev == current
4449 * is still correct, but it can be moved to another cpu/rq.
4451 cpu
= smp_processor_id();
4454 raw_spin_unlock_irq(&rq
->lock
);
4458 preempt_enable_no_resched();
4463 static inline void sched_submit_work(struct task_struct
*tsk
)
4468 * If we are going to sleep and we have plugged IO queued,
4469 * make sure to submit it to avoid deadlocks.
4471 if (blk_needs_flush_plug(tsk
))
4472 blk_schedule_flush_plug(tsk
);
4475 asmlinkage
void __sched
schedule(void)
4477 struct task_struct
*tsk
= current
;
4479 sched_submit_work(tsk
);
4482 EXPORT_SYMBOL(schedule
);
4484 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4486 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
4488 if (lock
->owner
!= owner
)
4492 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4493 * lock->owner still matches owner, if that fails, owner might
4494 * point to free()d memory, if it still matches, the rcu_read_lock()
4495 * ensures the memory stays valid.
4499 return owner
->on_cpu
;
4503 * Look out! "owner" is an entirely speculative pointer
4504 * access and not reliable.
4506 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
4508 if (!sched_feat(OWNER_SPIN
))
4512 while (owner_running(lock
, owner
)) {
4516 arch_mutex_cpu_relax();
4521 * We break out the loop above on need_resched() and when the
4522 * owner changed, which is a sign for heavy contention. Return
4523 * success only when lock->owner is NULL.
4525 return lock
->owner
== NULL
;
4529 #ifdef CONFIG_PREEMPT
4531 * this is the entry point to schedule() from in-kernel preemption
4532 * off of preempt_enable. Kernel preemptions off return from interrupt
4533 * occur there and call schedule directly.
4535 asmlinkage
void __sched notrace
preempt_schedule(void)
4537 struct thread_info
*ti
= current_thread_info();
4540 * If there is a non-zero preempt_count or interrupts are disabled,
4541 * we do not want to preempt the current task. Just return..
4543 if (likely(ti
->preempt_count
|| irqs_disabled()))
4547 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4549 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4552 * Check again in case we missed a preemption opportunity
4553 * between schedule and now.
4556 } while (need_resched());
4558 EXPORT_SYMBOL(preempt_schedule
);
4561 * this is the entry point to schedule() from kernel preemption
4562 * off of irq context.
4563 * Note, that this is called and return with irqs disabled. This will
4564 * protect us against recursive calling from irq.
4566 asmlinkage
void __sched
preempt_schedule_irq(void)
4568 struct thread_info
*ti
= current_thread_info();
4570 /* Catch callers which need to be fixed */
4571 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4574 add_preempt_count(PREEMPT_ACTIVE
);
4577 local_irq_disable();
4578 sub_preempt_count(PREEMPT_ACTIVE
);
4581 * Check again in case we missed a preemption opportunity
4582 * between schedule and now.
4585 } while (need_resched());
4588 #endif /* CONFIG_PREEMPT */
4590 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4593 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4595 EXPORT_SYMBOL(default_wake_function
);
4598 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4599 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4600 * number) then we wake all the non-exclusive tasks and one exclusive task.
4602 * There are circumstances in which we can try to wake a task which has already
4603 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4604 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4606 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4607 int nr_exclusive
, int wake_flags
, void *key
)
4609 wait_queue_t
*curr
, *next
;
4611 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4612 unsigned flags
= curr
->flags
;
4614 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4615 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4621 * __wake_up - wake up threads blocked on a waitqueue.
4623 * @mode: which threads
4624 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4625 * @key: is directly passed to the wakeup function
4627 * It may be assumed that this function implies a write memory barrier before
4628 * changing the task state if and only if any tasks are woken up.
4630 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4631 int nr_exclusive
, void *key
)
4633 unsigned long flags
;
4635 spin_lock_irqsave(&q
->lock
, flags
);
4636 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4637 spin_unlock_irqrestore(&q
->lock
, flags
);
4639 EXPORT_SYMBOL(__wake_up
);
4642 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4644 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4646 __wake_up_common(q
, mode
, 1, 0, NULL
);
4648 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4650 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4652 __wake_up_common(q
, mode
, 1, 0, key
);
4654 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
4657 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4659 * @mode: which threads
4660 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4661 * @key: opaque value to be passed to wakeup targets
4663 * The sync wakeup differs that the waker knows that it will schedule
4664 * away soon, so while the target thread will be woken up, it will not
4665 * be migrated to another CPU - ie. the two threads are 'synchronized'
4666 * with each other. This can prevent needless bouncing between CPUs.
4668 * On UP it can prevent extra preemption.
4670 * It may be assumed that this function implies a write memory barrier before
4671 * changing the task state if and only if any tasks are woken up.
4673 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4674 int nr_exclusive
, void *key
)
4676 unsigned long flags
;
4677 int wake_flags
= WF_SYNC
;
4682 if (unlikely(!nr_exclusive
))
4685 spin_lock_irqsave(&q
->lock
, flags
);
4686 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4687 spin_unlock_irqrestore(&q
->lock
, flags
);
4689 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4692 * __wake_up_sync - see __wake_up_sync_key()
4694 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4696 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4698 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4701 * complete: - signals a single thread waiting on this completion
4702 * @x: holds the state of this particular completion
4704 * This will wake up a single thread waiting on this completion. Threads will be
4705 * awakened in the same order in which they were queued.
4707 * See also complete_all(), wait_for_completion() and related routines.
4709 * It may be assumed that this function implies a write memory barrier before
4710 * changing the task state if and only if any tasks are woken up.
4712 void complete(struct completion
*x
)
4714 unsigned long flags
;
4716 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4718 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4719 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4721 EXPORT_SYMBOL(complete
);
4724 * complete_all: - signals all threads waiting on this completion
4725 * @x: holds the state of this particular completion
4727 * This will wake up all threads waiting on this particular completion event.
4729 * It may be assumed that this function implies a write memory barrier before
4730 * changing the task state if and only if any tasks are woken up.
4732 void complete_all(struct completion
*x
)
4734 unsigned long flags
;
4736 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4737 x
->done
+= UINT_MAX
/2;
4738 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4739 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4741 EXPORT_SYMBOL(complete_all
);
4743 static inline long __sched
4744 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4747 DECLARE_WAITQUEUE(wait
, current
);
4749 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4751 if (signal_pending_state(state
, current
)) {
4752 timeout
= -ERESTARTSYS
;
4755 __set_current_state(state
);
4756 spin_unlock_irq(&x
->wait
.lock
);
4757 timeout
= schedule_timeout(timeout
);
4758 spin_lock_irq(&x
->wait
.lock
);
4759 } while (!x
->done
&& timeout
);
4760 __remove_wait_queue(&x
->wait
, &wait
);
4765 return timeout
?: 1;
4769 wait_for_common(struct completion
*x
, long timeout
, int state
)
4773 spin_lock_irq(&x
->wait
.lock
);
4774 timeout
= do_wait_for_common(x
, timeout
, state
);
4775 spin_unlock_irq(&x
->wait
.lock
);
4780 * wait_for_completion: - waits for completion of a task
4781 * @x: holds the state of this particular completion
4783 * This waits to be signaled for completion of a specific task. It is NOT
4784 * interruptible and there is no timeout.
4786 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4787 * and interrupt capability. Also see complete().
4789 void __sched
wait_for_completion(struct completion
*x
)
4791 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4793 EXPORT_SYMBOL(wait_for_completion
);
4796 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4797 * @x: holds the state of this particular completion
4798 * @timeout: timeout value in jiffies
4800 * This waits for either a completion of a specific task to be signaled or for a
4801 * specified timeout to expire. The timeout is in jiffies. It is not
4804 * The return value is 0 if timed out, and positive (at least 1, or number of
4805 * jiffies left till timeout) if completed.
4807 unsigned long __sched
4808 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4810 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4812 EXPORT_SYMBOL(wait_for_completion_timeout
);
4815 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4816 * @x: holds the state of this particular completion
4818 * This waits for completion of a specific task to be signaled. It is
4821 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
4823 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4825 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4826 if (t
== -ERESTARTSYS
)
4830 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4833 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4834 * @x: holds the state of this particular completion
4835 * @timeout: timeout value in jiffies
4837 * This waits for either a completion of a specific task to be signaled or for a
4838 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4840 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
4841 * positive (at least 1, or number of jiffies left till timeout) if completed.
4844 wait_for_completion_interruptible_timeout(struct completion
*x
,
4845 unsigned long timeout
)
4847 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4849 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4852 * wait_for_completion_killable: - waits for completion of a task (killable)
4853 * @x: holds the state of this particular completion
4855 * This waits to be signaled for completion of a specific task. It can be
4856 * interrupted by a kill signal.
4858 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
4860 int __sched
wait_for_completion_killable(struct completion
*x
)
4862 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4863 if (t
== -ERESTARTSYS
)
4867 EXPORT_SYMBOL(wait_for_completion_killable
);
4870 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4871 * @x: holds the state of this particular completion
4872 * @timeout: timeout value in jiffies
4874 * This waits for either a completion of a specific task to be
4875 * signaled or for a specified timeout to expire. It can be
4876 * interrupted by a kill signal. The timeout is in jiffies.
4878 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
4879 * positive (at least 1, or number of jiffies left till timeout) if completed.
4882 wait_for_completion_killable_timeout(struct completion
*x
,
4883 unsigned long timeout
)
4885 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4887 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4890 * try_wait_for_completion - try to decrement a completion without blocking
4891 * @x: completion structure
4893 * Returns: 0 if a decrement cannot be done without blocking
4894 * 1 if a decrement succeeded.
4896 * If a completion is being used as a counting completion,
4897 * attempt to decrement the counter without blocking. This
4898 * enables us to avoid waiting if the resource the completion
4899 * is protecting is not available.
4901 bool try_wait_for_completion(struct completion
*x
)
4903 unsigned long flags
;
4906 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4911 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4914 EXPORT_SYMBOL(try_wait_for_completion
);
4917 * completion_done - Test to see if a completion has any waiters
4918 * @x: completion structure
4920 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4921 * 1 if there are no waiters.
4924 bool completion_done(struct completion
*x
)
4926 unsigned long flags
;
4929 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4932 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4935 EXPORT_SYMBOL(completion_done
);
4938 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4940 unsigned long flags
;
4943 init_waitqueue_entry(&wait
, current
);
4945 __set_current_state(state
);
4947 spin_lock_irqsave(&q
->lock
, flags
);
4948 __add_wait_queue(q
, &wait
);
4949 spin_unlock(&q
->lock
);
4950 timeout
= schedule_timeout(timeout
);
4951 spin_lock_irq(&q
->lock
);
4952 __remove_wait_queue(q
, &wait
);
4953 spin_unlock_irqrestore(&q
->lock
, flags
);
4958 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4960 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4962 EXPORT_SYMBOL(interruptible_sleep_on
);
4965 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4967 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4969 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4971 void __sched
sleep_on(wait_queue_head_t
*q
)
4973 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4975 EXPORT_SYMBOL(sleep_on
);
4977 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4979 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4981 EXPORT_SYMBOL(sleep_on_timeout
);
4983 #ifdef CONFIG_RT_MUTEXES
4986 * rt_mutex_setprio - set the current priority of a task
4988 * @prio: prio value (kernel-internal form)
4990 * This function changes the 'effective' priority of a task. It does
4991 * not touch ->normal_prio like __setscheduler().
4993 * Used by the rt_mutex code to implement priority inheritance logic.
4995 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4997 int oldprio
, on_rq
, running
;
4999 const struct sched_class
*prev_class
;
5001 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5003 rq
= __task_rq_lock(p
);
5005 trace_sched_pi_setprio(p
, prio
);
5007 prev_class
= p
->sched_class
;
5009 running
= task_current(rq
, p
);
5011 dequeue_task(rq
, p
, 0);
5013 p
->sched_class
->put_prev_task(rq
, p
);
5016 p
->sched_class
= &rt_sched_class
;
5018 p
->sched_class
= &fair_sched_class
;
5023 p
->sched_class
->set_curr_task(rq
);
5025 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
5027 check_class_changed(rq
, p
, prev_class
, oldprio
);
5028 __task_rq_unlock(rq
);
5033 void set_user_nice(struct task_struct
*p
, long nice
)
5035 int old_prio
, delta
, on_rq
;
5036 unsigned long flags
;
5039 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5042 * We have to be careful, if called from sys_setpriority(),
5043 * the task might be in the middle of scheduling on another CPU.
5045 rq
= task_rq_lock(p
, &flags
);
5047 * The RT priorities are set via sched_setscheduler(), but we still
5048 * allow the 'normal' nice value to be set - but as expected
5049 * it wont have any effect on scheduling until the task is
5050 * SCHED_FIFO/SCHED_RR:
5052 if (task_has_rt_policy(p
)) {
5053 p
->static_prio
= NICE_TO_PRIO(nice
);
5058 dequeue_task(rq
, p
, 0);
5060 p
->static_prio
= NICE_TO_PRIO(nice
);
5063 p
->prio
= effective_prio(p
);
5064 delta
= p
->prio
- old_prio
;
5067 enqueue_task(rq
, p
, 0);
5069 * If the task increased its priority or is running and
5070 * lowered its priority, then reschedule its CPU:
5072 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5073 resched_task(rq
->curr
);
5076 task_rq_unlock(rq
, p
, &flags
);
5078 EXPORT_SYMBOL(set_user_nice
);
5081 * can_nice - check if a task can reduce its nice value
5085 int can_nice(const struct task_struct
*p
, const int nice
)
5087 /* convert nice value [19,-20] to rlimit style value [1,40] */
5088 int nice_rlim
= 20 - nice
;
5090 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
5091 capable(CAP_SYS_NICE
));
5094 #ifdef __ARCH_WANT_SYS_NICE
5097 * sys_nice - change the priority of the current process.
5098 * @increment: priority increment
5100 * sys_setpriority is a more generic, but much slower function that
5101 * does similar things.
5103 SYSCALL_DEFINE1(nice
, int, increment
)
5108 * Setpriority might change our priority at the same moment.
5109 * We don't have to worry. Conceptually one call occurs first
5110 * and we have a single winner.
5112 if (increment
< -40)
5117 nice
= TASK_NICE(current
) + increment
;
5123 if (increment
< 0 && !can_nice(current
, nice
))
5126 retval
= security_task_setnice(current
, nice
);
5130 set_user_nice(current
, nice
);
5137 * task_prio - return the priority value of a given task.
5138 * @p: the task in question.
5140 * This is the priority value as seen by users in /proc.
5141 * RT tasks are offset by -200. Normal tasks are centered
5142 * around 0, value goes from -16 to +15.
5144 int task_prio(const struct task_struct
*p
)
5146 return p
->prio
- MAX_RT_PRIO
;
5150 * task_nice - return the nice value of a given task.
5151 * @p: the task in question.
5153 int task_nice(const struct task_struct
*p
)
5155 return TASK_NICE(p
);
5157 EXPORT_SYMBOL(task_nice
);
5160 * idle_cpu - is a given cpu idle currently?
5161 * @cpu: the processor in question.
5163 int idle_cpu(int cpu
)
5165 struct rq
*rq
= cpu_rq(cpu
);
5167 if (rq
->curr
!= rq
->idle
)
5174 if (!llist_empty(&rq
->wake_list
))
5182 * idle_task - return the idle task for a given cpu.
5183 * @cpu: the processor in question.
5185 struct task_struct
*idle_task(int cpu
)
5187 return cpu_rq(cpu
)->idle
;
5191 * find_process_by_pid - find a process with a matching PID value.
5192 * @pid: the pid in question.
5194 static struct task_struct
*find_process_by_pid(pid_t pid
)
5196 return pid
? find_task_by_vpid(pid
) : current
;
5199 /* Actually do priority change: must hold rq lock. */
5201 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5204 p
->rt_priority
= prio
;
5205 p
->normal_prio
= normal_prio(p
);
5206 /* we are holding p->pi_lock already */
5207 p
->prio
= rt_mutex_getprio(p
);
5208 if (rt_prio(p
->prio
))
5209 p
->sched_class
= &rt_sched_class
;
5211 p
->sched_class
= &fair_sched_class
;
5216 * check the target process has a UID that matches the current process's
5218 static bool check_same_owner(struct task_struct
*p
)
5220 const struct cred
*cred
= current_cred(), *pcred
;
5224 pcred
= __task_cred(p
);
5225 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
5226 match
= (cred
->euid
== pcred
->euid
||
5227 cred
->euid
== pcred
->uid
);
5234 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5235 const struct sched_param
*param
, bool user
)
5237 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5238 unsigned long flags
;
5239 const struct sched_class
*prev_class
;
5243 /* may grab non-irq protected spin_locks */
5244 BUG_ON(in_interrupt());
5246 /* double check policy once rq lock held */
5248 reset_on_fork
= p
->sched_reset_on_fork
;
5249 policy
= oldpolicy
= p
->policy
;
5251 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
5252 policy
&= ~SCHED_RESET_ON_FORK
;
5254 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5255 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5256 policy
!= SCHED_IDLE
)
5261 * Valid priorities for SCHED_FIFO and SCHED_RR are
5262 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5263 * SCHED_BATCH and SCHED_IDLE is 0.
5265 if (param
->sched_priority
< 0 ||
5266 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5267 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5269 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5273 * Allow unprivileged RT tasks to decrease priority:
5275 if (user
&& !capable(CAP_SYS_NICE
)) {
5276 if (rt_policy(policy
)) {
5277 unsigned long rlim_rtprio
=
5278 task_rlimit(p
, RLIMIT_RTPRIO
);
5280 /* can't set/change the rt policy */
5281 if (policy
!= p
->policy
&& !rlim_rtprio
)
5284 /* can't increase priority */
5285 if (param
->sched_priority
> p
->rt_priority
&&
5286 param
->sched_priority
> rlim_rtprio
)
5291 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5292 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5294 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
5295 if (!can_nice(p
, TASK_NICE(p
)))
5299 /* can't change other user's priorities */
5300 if (!check_same_owner(p
))
5303 /* Normal users shall not reset the sched_reset_on_fork flag */
5304 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5309 retval
= security_task_setscheduler(p
);
5315 * make sure no PI-waiters arrive (or leave) while we are
5316 * changing the priority of the task:
5318 * To be able to change p->policy safely, the appropriate
5319 * runqueue lock must be held.
5321 rq
= task_rq_lock(p
, &flags
);
5324 * Changing the policy of the stop threads its a very bad idea
5326 if (p
== rq
->stop
) {
5327 task_rq_unlock(rq
, p
, &flags
);
5332 * If not changing anything there's no need to proceed further:
5334 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
5335 param
->sched_priority
== p
->rt_priority
))) {
5337 __task_rq_unlock(rq
);
5338 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5342 #ifdef CONFIG_RT_GROUP_SCHED
5345 * Do not allow realtime tasks into groups that have no runtime
5348 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5349 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5350 !task_group_is_autogroup(task_group(p
))) {
5351 task_rq_unlock(rq
, p
, &flags
);
5357 /* recheck policy now with rq lock held */
5358 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5359 policy
= oldpolicy
= -1;
5360 task_rq_unlock(rq
, p
, &flags
);
5364 running
= task_current(rq
, p
);
5366 deactivate_task(rq
, p
, 0);
5368 p
->sched_class
->put_prev_task(rq
, p
);
5370 p
->sched_reset_on_fork
= reset_on_fork
;
5373 prev_class
= p
->sched_class
;
5374 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5377 p
->sched_class
->set_curr_task(rq
);
5379 activate_task(rq
, p
, 0);
5381 check_class_changed(rq
, p
, prev_class
, oldprio
);
5382 task_rq_unlock(rq
, p
, &flags
);
5384 rt_mutex_adjust_pi(p
);
5390 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5391 * @p: the task in question.
5392 * @policy: new policy.
5393 * @param: structure containing the new RT priority.
5395 * NOTE that the task may be already dead.
5397 int sched_setscheduler(struct task_struct
*p
, int policy
,
5398 const struct sched_param
*param
)
5400 return __sched_setscheduler(p
, policy
, param
, true);
5402 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5405 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5406 * @p: the task in question.
5407 * @policy: new policy.
5408 * @param: structure containing the new RT priority.
5410 * Just like sched_setscheduler, only don't bother checking if the
5411 * current context has permission. For example, this is needed in
5412 * stop_machine(): we create temporary high priority worker threads,
5413 * but our caller might not have that capability.
5415 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5416 const struct sched_param
*param
)
5418 return __sched_setscheduler(p
, policy
, param
, false);
5422 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5424 struct sched_param lparam
;
5425 struct task_struct
*p
;
5428 if (!param
|| pid
< 0)
5430 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5435 p
= find_process_by_pid(pid
);
5437 retval
= sched_setscheduler(p
, policy
, &lparam
);
5444 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5445 * @pid: the pid in question.
5446 * @policy: new policy.
5447 * @param: structure containing the new RT priority.
5449 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5450 struct sched_param __user
*, param
)
5452 /* negative values for policy are not valid */
5456 return do_sched_setscheduler(pid
, policy
, param
);
5460 * sys_sched_setparam - set/change the RT priority of a thread
5461 * @pid: the pid in question.
5462 * @param: structure containing the new RT priority.
5464 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5466 return do_sched_setscheduler(pid
, -1, param
);
5470 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5471 * @pid: the pid in question.
5473 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5475 struct task_struct
*p
;
5483 p
= find_process_by_pid(pid
);
5485 retval
= security_task_getscheduler(p
);
5488 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5495 * sys_sched_getparam - get the RT priority of a thread
5496 * @pid: the pid in question.
5497 * @param: structure containing the RT priority.
5499 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5501 struct sched_param lp
;
5502 struct task_struct
*p
;
5505 if (!param
|| pid
< 0)
5509 p
= find_process_by_pid(pid
);
5514 retval
= security_task_getscheduler(p
);
5518 lp
.sched_priority
= p
->rt_priority
;
5522 * This one might sleep, we cannot do it with a spinlock held ...
5524 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5533 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5535 cpumask_var_t cpus_allowed
, new_mask
;
5536 struct task_struct
*p
;
5542 p
= find_process_by_pid(pid
);
5549 /* Prevent p going away */
5553 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5557 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5559 goto out_free_cpus_allowed
;
5562 if (!check_same_owner(p
) && !task_ns_capable(p
, CAP_SYS_NICE
))
5565 retval
= security_task_setscheduler(p
);
5569 cpuset_cpus_allowed(p
, cpus_allowed
);
5570 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5572 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5575 cpuset_cpus_allowed(p
, cpus_allowed
);
5576 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5578 * We must have raced with a concurrent cpuset
5579 * update. Just reset the cpus_allowed to the
5580 * cpuset's cpus_allowed
5582 cpumask_copy(new_mask
, cpus_allowed
);
5587 free_cpumask_var(new_mask
);
5588 out_free_cpus_allowed
:
5589 free_cpumask_var(cpus_allowed
);
5596 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5597 struct cpumask
*new_mask
)
5599 if (len
< cpumask_size())
5600 cpumask_clear(new_mask
);
5601 else if (len
> cpumask_size())
5602 len
= cpumask_size();
5604 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5608 * sys_sched_setaffinity - set the cpu affinity of a process
5609 * @pid: pid of the process
5610 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5611 * @user_mask_ptr: user-space pointer to the new cpu mask
5613 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5614 unsigned long __user
*, user_mask_ptr
)
5616 cpumask_var_t new_mask
;
5619 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5622 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5624 retval
= sched_setaffinity(pid
, new_mask
);
5625 free_cpumask_var(new_mask
);
5629 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5631 struct task_struct
*p
;
5632 unsigned long flags
;
5639 p
= find_process_by_pid(pid
);
5643 retval
= security_task_getscheduler(p
);
5647 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5648 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5649 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5659 * sys_sched_getaffinity - get the cpu affinity of a process
5660 * @pid: pid of the process
5661 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5662 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5664 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5665 unsigned long __user
*, user_mask_ptr
)
5670 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5672 if (len
& (sizeof(unsigned long)-1))
5675 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5678 ret
= sched_getaffinity(pid
, mask
);
5680 size_t retlen
= min_t(size_t, len
, cpumask_size());
5682 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5687 free_cpumask_var(mask
);
5693 * sys_sched_yield - yield the current processor to other threads.
5695 * This function yields the current CPU to other tasks. If there are no
5696 * other threads running on this CPU then this function will return.
5698 SYSCALL_DEFINE0(sched_yield
)
5700 struct rq
*rq
= this_rq_lock();
5702 schedstat_inc(rq
, yld_count
);
5703 current
->sched_class
->yield_task(rq
);
5706 * Since we are going to call schedule() anyway, there's
5707 * no need to preempt or enable interrupts:
5709 __release(rq
->lock
);
5710 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5711 do_raw_spin_unlock(&rq
->lock
);
5712 preempt_enable_no_resched();
5719 static inline int should_resched(void)
5721 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5724 static void __cond_resched(void)
5726 add_preempt_count(PREEMPT_ACTIVE
);
5728 sub_preempt_count(PREEMPT_ACTIVE
);
5731 int __sched
_cond_resched(void)
5733 if (should_resched()) {
5739 EXPORT_SYMBOL(_cond_resched
);
5742 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5743 * call schedule, and on return reacquire the lock.
5745 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5746 * operations here to prevent schedule() from being called twice (once via
5747 * spin_unlock(), once by hand).
5749 int __cond_resched_lock(spinlock_t
*lock
)
5751 int resched
= should_resched();
5754 lockdep_assert_held(lock
);
5756 if (spin_needbreak(lock
) || resched
) {
5767 EXPORT_SYMBOL(__cond_resched_lock
);
5769 int __sched
__cond_resched_softirq(void)
5771 BUG_ON(!in_softirq());
5773 if (should_resched()) {
5781 EXPORT_SYMBOL(__cond_resched_softirq
);
5784 * yield - yield the current processor to other threads.
5786 * This is a shortcut for kernel-space yielding - it marks the
5787 * thread runnable and calls sys_sched_yield().
5789 void __sched
yield(void)
5791 set_current_state(TASK_RUNNING
);
5794 EXPORT_SYMBOL(yield
);
5797 * yield_to - yield the current processor to another thread in
5798 * your thread group, or accelerate that thread toward the
5799 * processor it's on.
5801 * @preempt: whether task preemption is allowed or not
5803 * It's the caller's job to ensure that the target task struct
5804 * can't go away on us before we can do any checks.
5806 * Returns true if we indeed boosted the target task.
5808 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
5810 struct task_struct
*curr
= current
;
5811 struct rq
*rq
, *p_rq
;
5812 unsigned long flags
;
5815 local_irq_save(flags
);
5820 double_rq_lock(rq
, p_rq
);
5821 while (task_rq(p
) != p_rq
) {
5822 double_rq_unlock(rq
, p_rq
);
5826 if (!curr
->sched_class
->yield_to_task
)
5829 if (curr
->sched_class
!= p
->sched_class
)
5832 if (task_running(p_rq
, p
) || p
->state
)
5835 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5837 schedstat_inc(rq
, yld_count
);
5839 * Make p's CPU reschedule; pick_next_entity takes care of
5842 if (preempt
&& rq
!= p_rq
)
5843 resched_task(p_rq
->curr
);
5847 double_rq_unlock(rq
, p_rq
);
5848 local_irq_restore(flags
);
5855 EXPORT_SYMBOL_GPL(yield_to
);
5858 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5859 * that process accounting knows that this is a task in IO wait state.
5861 void __sched
io_schedule(void)
5863 struct rq
*rq
= raw_rq();
5865 delayacct_blkio_start();
5866 atomic_inc(&rq
->nr_iowait
);
5867 blk_flush_plug(current
);
5868 current
->in_iowait
= 1;
5870 current
->in_iowait
= 0;
5871 atomic_dec(&rq
->nr_iowait
);
5872 delayacct_blkio_end();
5874 EXPORT_SYMBOL(io_schedule
);
5876 long __sched
io_schedule_timeout(long timeout
)
5878 struct rq
*rq
= raw_rq();
5881 delayacct_blkio_start();
5882 atomic_inc(&rq
->nr_iowait
);
5883 blk_flush_plug(current
);
5884 current
->in_iowait
= 1;
5885 ret
= schedule_timeout(timeout
);
5886 current
->in_iowait
= 0;
5887 atomic_dec(&rq
->nr_iowait
);
5888 delayacct_blkio_end();
5893 * sys_sched_get_priority_max - return maximum RT priority.
5894 * @policy: scheduling class.
5896 * this syscall returns the maximum rt_priority that can be used
5897 * by a given scheduling class.
5899 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5906 ret
= MAX_USER_RT_PRIO
-1;
5918 * sys_sched_get_priority_min - return minimum RT priority.
5919 * @policy: scheduling class.
5921 * this syscall returns the minimum rt_priority that can be used
5922 * by a given scheduling class.
5924 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5942 * sys_sched_rr_get_interval - return the default timeslice of a process.
5943 * @pid: pid of the process.
5944 * @interval: userspace pointer to the timeslice value.
5946 * this syscall writes the default timeslice value of a given process
5947 * into the user-space timespec buffer. A value of '0' means infinity.
5949 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5950 struct timespec __user
*, interval
)
5952 struct task_struct
*p
;
5953 unsigned int time_slice
;
5954 unsigned long flags
;
5964 p
= find_process_by_pid(pid
);
5968 retval
= security_task_getscheduler(p
);
5972 rq
= task_rq_lock(p
, &flags
);
5973 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5974 task_rq_unlock(rq
, p
, &flags
);
5977 jiffies_to_timespec(time_slice
, &t
);
5978 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5986 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5988 void sched_show_task(struct task_struct
*p
)
5990 unsigned long free
= 0;
5993 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5994 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5995 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5996 #if BITS_PER_LONG == 32
5997 if (state
== TASK_RUNNING
)
5998 printk(KERN_CONT
" running ");
6000 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6002 if (state
== TASK_RUNNING
)
6003 printk(KERN_CONT
" running task ");
6005 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6007 #ifdef CONFIG_DEBUG_STACK_USAGE
6008 free
= stack_not_used(p
);
6010 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6011 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6012 (unsigned long)task_thread_info(p
)->flags
);
6014 show_stack(p
, NULL
);
6017 void show_state_filter(unsigned long state_filter
)
6019 struct task_struct
*g
, *p
;
6021 #if BITS_PER_LONG == 32
6023 " task PC stack pid father\n");
6026 " task PC stack pid father\n");
6029 do_each_thread(g
, p
) {
6031 * reset the NMI-timeout, listing all files on a slow
6032 * console might take a lot of time:
6034 touch_nmi_watchdog();
6035 if (!state_filter
|| (p
->state
& state_filter
))
6037 } while_each_thread(g
, p
);
6039 touch_all_softlockup_watchdogs();
6041 #ifdef CONFIG_SCHED_DEBUG
6042 sysrq_sched_debug_show();
6046 * Only show locks if all tasks are dumped:
6049 debug_show_all_locks();
6052 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6054 idle
->sched_class
= &idle_sched_class
;
6058 * init_idle - set up an idle thread for a given CPU
6059 * @idle: task in question
6060 * @cpu: cpu the idle task belongs to
6062 * NOTE: this function does not set the idle thread's NEED_RESCHED
6063 * flag, to make booting more robust.
6065 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6067 struct rq
*rq
= cpu_rq(cpu
);
6068 unsigned long flags
;
6070 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6073 idle
->state
= TASK_RUNNING
;
6074 idle
->se
.exec_start
= sched_clock();
6076 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
6078 * We're having a chicken and egg problem, even though we are
6079 * holding rq->lock, the cpu isn't yet set to this cpu so the
6080 * lockdep check in task_group() will fail.
6082 * Similar case to sched_fork(). / Alternatively we could
6083 * use task_rq_lock() here and obtain the other rq->lock.
6088 __set_task_cpu(idle
, cpu
);
6091 rq
->curr
= rq
->idle
= idle
;
6092 #if defined(CONFIG_SMP)
6095 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6097 /* Set the preempt count _outside_ the spinlocks! */
6098 task_thread_info(idle
)->preempt_count
= 0;
6101 * The idle tasks have their own, simple scheduling class:
6103 idle
->sched_class
= &idle_sched_class
;
6104 ftrace_graph_init_idle_task(idle
, cpu
);
6105 #if defined(CONFIG_SMP)
6106 sprintf(idle
->comm
, "%s/%d", INIT_TASK_COMM
, cpu
);
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
);
6168 * This is how migration works:
6170 * 1) we invoke migration_cpu_stop() on the target CPU using
6172 * 2) stopper starts to run (implicitly forcing the migrated thread
6174 * 3) it checks whether the migrated task is still in the wrong runqueue.
6175 * 4) if it's in the wrong runqueue then the migration thread removes
6176 * it and puts it into the right queue.
6177 * 5) stopper completes and stop_one_cpu() returns and the migration
6182 * Change a given task's CPU affinity. Migrate the thread to a
6183 * proper CPU and schedule it away if the CPU it's executing on
6184 * is removed from the allowed bitmask.
6186 * NOTE: the caller must have a valid reference to the task, the
6187 * task must not exit() & deallocate itself prematurely. The
6188 * call is not atomic; no spinlocks may be held.
6190 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6192 unsigned long flags
;
6194 unsigned int dest_cpu
;
6197 rq
= task_rq_lock(p
, &flags
);
6199 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
6202 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
6207 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
6212 do_set_cpus_allowed(p
, new_mask
);
6214 /* Can the task run on the task's current CPU? If so, we're done */
6215 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6218 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
6220 struct migration_arg arg
= { p
, dest_cpu
};
6221 /* Need help from migration thread: drop lock and wait. */
6222 task_rq_unlock(rq
, p
, &flags
);
6223 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
6224 tlb_migrate_finish(p
->mm
);
6228 task_rq_unlock(rq
, p
, &flags
);
6232 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6235 * Move (not current) task off this cpu, onto dest cpu. We're doing
6236 * this because either it can't run here any more (set_cpus_allowed()
6237 * away from this CPU, or CPU going down), or because we're
6238 * attempting to rebalance this task on exec (sched_exec).
6240 * So we race with normal scheduler movements, but that's OK, as long
6241 * as the task is no longer on this CPU.
6243 * Returns non-zero if task was successfully migrated.
6245 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6247 struct rq
*rq_dest
, *rq_src
;
6250 if (unlikely(!cpu_active(dest_cpu
)))
6253 rq_src
= cpu_rq(src_cpu
);
6254 rq_dest
= cpu_rq(dest_cpu
);
6256 raw_spin_lock(&p
->pi_lock
);
6257 double_rq_lock(rq_src
, rq_dest
);
6258 /* Already moved. */
6259 if (task_cpu(p
) != src_cpu
)
6261 /* Affinity changed (again). */
6262 if (!cpumask_test_cpu(dest_cpu
, tsk_cpus_allowed(p
)))
6266 * If we're not on a rq, the next wake-up will ensure we're
6270 deactivate_task(rq_src
, p
, 0);
6271 set_task_cpu(p
, dest_cpu
);
6272 activate_task(rq_dest
, p
, 0);
6273 check_preempt_curr(rq_dest
, p
, 0);
6278 double_rq_unlock(rq_src
, rq_dest
);
6279 raw_spin_unlock(&p
->pi_lock
);
6284 * migration_cpu_stop - this will be executed by a highprio stopper thread
6285 * and performs thread migration by bumping thread off CPU then
6286 * 'pushing' onto another runqueue.
6288 static int migration_cpu_stop(void *data
)
6290 struct migration_arg
*arg
= data
;
6293 * The original target cpu might have gone down and we might
6294 * be on another cpu but it doesn't matter.
6296 local_irq_disable();
6297 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
6302 #ifdef CONFIG_HOTPLUG_CPU
6305 * Ensures that the idle task is using init_mm right before its cpu goes
6308 void idle_task_exit(void)
6310 struct mm_struct
*mm
= current
->active_mm
;
6312 BUG_ON(cpu_online(smp_processor_id()));
6315 switch_mm(mm
, &init_mm
, current
);
6320 * While a dead CPU has no uninterruptible tasks queued at this point,
6321 * it might still have a nonzero ->nr_uninterruptible counter, because
6322 * for performance reasons the counter is not stricly tracking tasks to
6323 * their home CPUs. So we just add the counter to another CPU's counter,
6324 * to keep the global sum constant after CPU-down:
6326 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6328 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
6330 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6331 rq_src
->nr_uninterruptible
= 0;
6335 * remove the tasks which were accounted by rq from calc_load_tasks.
6337 static void calc_global_load_remove(struct rq
*rq
)
6339 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
6340 rq
->calc_load_active
= 0;
6343 #ifdef CONFIG_CFS_BANDWIDTH
6344 static void unthrottle_offline_cfs_rqs(struct rq
*rq
)
6346 struct cfs_rq
*cfs_rq
;
6348 for_each_leaf_cfs_rq(rq
, cfs_rq
) {
6349 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(cfs_rq
->tg
);
6351 if (!cfs_rq
->runtime_enabled
)
6355 * clock_task is not advancing so we just need to make sure
6356 * there's some valid quota amount
6358 cfs_rq
->runtime_remaining
= cfs_b
->quota
;
6359 if (cfs_rq_throttled(cfs_rq
))
6360 unthrottle_cfs_rq(cfs_rq
);
6364 static void unthrottle_offline_cfs_rqs(struct rq
*rq
) {}
6368 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6369 * try_to_wake_up()->select_task_rq().
6371 * Called with rq->lock held even though we'er in stop_machine() and
6372 * there's no concurrency possible, we hold the required locks anyway
6373 * because of lock validation efforts.
6375 static void migrate_tasks(unsigned int dead_cpu
)
6377 struct rq
*rq
= cpu_rq(dead_cpu
);
6378 struct task_struct
*next
, *stop
= rq
->stop
;
6382 * Fudge the rq selection such that the below task selection loop
6383 * doesn't get stuck on the currently eligible stop task.
6385 * We're currently inside stop_machine() and the rq is either stuck
6386 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6387 * either way we should never end up calling schedule() until we're
6392 /* Ensure any throttled groups are reachable by pick_next_task */
6393 unthrottle_offline_cfs_rqs(rq
);
6397 * There's this thread running, bail when that's the only
6400 if (rq
->nr_running
== 1)
6403 next
= pick_next_task(rq
);
6405 next
->sched_class
->put_prev_task(rq
, next
);
6407 /* Find suitable destination for @next, with force if needed. */
6408 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
6409 raw_spin_unlock(&rq
->lock
);
6411 __migrate_task(next
, dead_cpu
, dest_cpu
);
6413 raw_spin_lock(&rq
->lock
);
6419 #endif /* CONFIG_HOTPLUG_CPU */
6421 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6423 static struct ctl_table sd_ctl_dir
[] = {
6425 .procname
= "sched_domain",
6431 static struct ctl_table sd_ctl_root
[] = {
6433 .procname
= "kernel",
6435 .child
= sd_ctl_dir
,
6440 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6442 struct ctl_table
*entry
=
6443 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6448 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6450 struct ctl_table
*entry
;
6453 * In the intermediate directories, both the child directory and
6454 * procname are dynamically allocated and could fail but the mode
6455 * will always be set. In the lowest directory the names are
6456 * static strings and all have proc handlers.
6458 for (entry
= *tablep
; entry
->mode
; entry
++) {
6460 sd_free_ctl_entry(&entry
->child
);
6461 if (entry
->proc_handler
== NULL
)
6462 kfree(entry
->procname
);
6470 set_table_entry(struct ctl_table
*entry
,
6471 const char *procname
, void *data
, int maxlen
,
6472 mode_t mode
, proc_handler
*proc_handler
)
6474 entry
->procname
= procname
;
6476 entry
->maxlen
= maxlen
;
6478 entry
->proc_handler
= proc_handler
;
6481 static struct ctl_table
*
6482 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6484 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6489 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6490 sizeof(long), 0644, proc_doulongvec_minmax
);
6491 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6492 sizeof(long), 0644, proc_doulongvec_minmax
);
6493 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6494 sizeof(int), 0644, proc_dointvec_minmax
);
6495 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6496 sizeof(int), 0644, proc_dointvec_minmax
);
6497 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6498 sizeof(int), 0644, proc_dointvec_minmax
);
6499 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6500 sizeof(int), 0644, proc_dointvec_minmax
);
6501 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6502 sizeof(int), 0644, proc_dointvec_minmax
);
6503 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6504 sizeof(int), 0644, proc_dointvec_minmax
);
6505 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6506 sizeof(int), 0644, proc_dointvec_minmax
);
6507 set_table_entry(&table
[9], "cache_nice_tries",
6508 &sd
->cache_nice_tries
,
6509 sizeof(int), 0644, proc_dointvec_minmax
);
6510 set_table_entry(&table
[10], "flags", &sd
->flags
,
6511 sizeof(int), 0644, proc_dointvec_minmax
);
6512 set_table_entry(&table
[11], "name", sd
->name
,
6513 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6514 /* &table[12] is terminator */
6519 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6521 struct ctl_table
*entry
, *table
;
6522 struct sched_domain
*sd
;
6523 int domain_num
= 0, i
;
6526 for_each_domain(cpu
, sd
)
6528 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6533 for_each_domain(cpu
, sd
) {
6534 snprintf(buf
, 32, "domain%d", i
);
6535 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6537 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6544 static struct ctl_table_header
*sd_sysctl_header
;
6545 static void register_sched_domain_sysctl(void)
6547 int i
, cpu_num
= num_possible_cpus();
6548 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6551 WARN_ON(sd_ctl_dir
[0].child
);
6552 sd_ctl_dir
[0].child
= entry
;
6557 for_each_possible_cpu(i
) {
6558 snprintf(buf
, 32, "cpu%d", i
);
6559 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6561 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6565 WARN_ON(sd_sysctl_header
);
6566 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6569 /* may be called multiple times per register */
6570 static void unregister_sched_domain_sysctl(void)
6572 if (sd_sysctl_header
)
6573 unregister_sysctl_table(sd_sysctl_header
);
6574 sd_sysctl_header
= NULL
;
6575 if (sd_ctl_dir
[0].child
)
6576 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6579 static void register_sched_domain_sysctl(void)
6582 static void unregister_sched_domain_sysctl(void)
6587 static void set_rq_online(struct rq
*rq
)
6590 const struct sched_class
*class;
6592 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6595 for_each_class(class) {
6596 if (class->rq_online
)
6597 class->rq_online(rq
);
6602 static void set_rq_offline(struct rq
*rq
)
6605 const struct sched_class
*class;
6607 for_each_class(class) {
6608 if (class->rq_offline
)
6609 class->rq_offline(rq
);
6612 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6618 * migration_call - callback that gets triggered when a CPU is added.
6619 * Here we can start up the necessary migration thread for the new CPU.
6621 static int __cpuinit
6622 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6624 int cpu
= (long)hcpu
;
6625 unsigned long flags
;
6626 struct rq
*rq
= cpu_rq(cpu
);
6628 switch (action
& ~CPU_TASKS_FROZEN
) {
6630 case CPU_UP_PREPARE
:
6631 rq
->calc_load_update
= calc_load_update
;
6635 /* Update our root-domain */
6636 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6638 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6642 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6645 #ifdef CONFIG_HOTPLUG_CPU
6647 sched_ttwu_pending();
6648 /* Update our root-domain */
6649 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6651 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6655 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6656 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6658 migrate_nr_uninterruptible(rq
);
6659 calc_global_load_remove(rq
);
6664 update_max_interval();
6670 * Register at high priority so that task migration (migrate_all_tasks)
6671 * happens before everything else. This has to be lower priority than
6672 * the notifier in the perf_event subsystem, though.
6674 static struct notifier_block __cpuinitdata migration_notifier
= {
6675 .notifier_call
= migration_call
,
6676 .priority
= CPU_PRI_MIGRATION
,
6679 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6680 unsigned long action
, void *hcpu
)
6682 switch (action
& ~CPU_TASKS_FROZEN
) {
6684 case CPU_DOWN_FAILED
:
6685 set_cpu_active((long)hcpu
, true);
6692 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6693 unsigned long action
, void *hcpu
)
6695 switch (action
& ~CPU_TASKS_FROZEN
) {
6696 case CPU_DOWN_PREPARE
:
6697 set_cpu_active((long)hcpu
, false);
6704 static int __init
migration_init(void)
6706 void *cpu
= (void *)(long)smp_processor_id();
6709 /* Initialize migration for the boot CPU */
6710 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6711 BUG_ON(err
== NOTIFY_BAD
);
6712 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6713 register_cpu_notifier(&migration_notifier
);
6715 /* Register cpu active notifiers */
6716 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6717 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6721 early_initcall(migration_init
);
6726 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
6728 #ifdef CONFIG_SCHED_DEBUG
6730 static __read_mostly
int sched_domain_debug_enabled
;
6732 static int __init
sched_domain_debug_setup(char *str
)
6734 sched_domain_debug_enabled
= 1;
6738 early_param("sched_debug", sched_domain_debug_setup
);
6740 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6741 struct cpumask
*groupmask
)
6743 struct sched_group
*group
= sd
->groups
;
6746 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6747 cpumask_clear(groupmask
);
6749 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6751 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6752 printk("does not load-balance\n");
6754 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6759 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6761 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6762 printk(KERN_ERR
"ERROR: domain->span does not contain "
6765 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6766 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6770 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6774 printk(KERN_ERR
"ERROR: group is NULL\n");
6778 if (!group
->sgp
->power
) {
6779 printk(KERN_CONT
"\n");
6780 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6785 if (!cpumask_weight(sched_group_cpus(group
))) {
6786 printk(KERN_CONT
"\n");
6787 printk(KERN_ERR
"ERROR: empty group\n");
6791 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6792 printk(KERN_CONT
"\n");
6793 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6797 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6799 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6801 printk(KERN_CONT
" %s", str
);
6802 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
6803 printk(KERN_CONT
" (cpu_power = %d)",
6807 group
= group
->next
;
6808 } while (group
!= sd
->groups
);
6809 printk(KERN_CONT
"\n");
6811 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6812 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6815 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6816 printk(KERN_ERR
"ERROR: parent span is not a superset "
6817 "of domain->span\n");
6821 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6825 if (!sched_domain_debug_enabled
)
6829 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6833 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6836 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
6844 #else /* !CONFIG_SCHED_DEBUG */
6845 # define sched_domain_debug(sd, cpu) do { } while (0)
6846 #endif /* CONFIG_SCHED_DEBUG */
6848 static int sd_degenerate(struct sched_domain
*sd
)
6850 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6853 /* Following flags need at least 2 groups */
6854 if (sd
->flags
& (SD_LOAD_BALANCE
|
6855 SD_BALANCE_NEWIDLE
|
6859 SD_SHARE_PKG_RESOURCES
)) {
6860 if (sd
->groups
!= sd
->groups
->next
)
6864 /* Following flags don't use groups */
6865 if (sd
->flags
& (SD_WAKE_AFFINE
))
6872 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6874 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6876 if (sd_degenerate(parent
))
6879 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6882 /* Flags needing groups don't count if only 1 group in parent */
6883 if (parent
->groups
== parent
->groups
->next
) {
6884 pflags
&= ~(SD_LOAD_BALANCE
|
6885 SD_BALANCE_NEWIDLE
|
6889 SD_SHARE_PKG_RESOURCES
);
6890 if (nr_node_ids
== 1)
6891 pflags
&= ~SD_SERIALIZE
;
6893 if (~cflags
& pflags
)
6899 static void free_rootdomain(struct rcu_head
*rcu
)
6901 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
6903 cpupri_cleanup(&rd
->cpupri
);
6904 free_cpumask_var(rd
->rto_mask
);
6905 free_cpumask_var(rd
->online
);
6906 free_cpumask_var(rd
->span
);
6910 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6912 struct root_domain
*old_rd
= NULL
;
6913 unsigned long flags
;
6915 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6920 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6923 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6926 * If we dont want to free the old_rt yet then
6927 * set old_rd to NULL to skip the freeing later
6930 if (!atomic_dec_and_test(&old_rd
->refcount
))
6934 atomic_inc(&rd
->refcount
);
6937 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6938 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6941 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6944 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
6947 static int init_rootdomain(struct root_domain
*rd
)
6949 memset(rd
, 0, sizeof(*rd
));
6951 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6953 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6955 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6958 if (cpupri_init(&rd
->cpupri
) != 0)
6963 free_cpumask_var(rd
->rto_mask
);
6965 free_cpumask_var(rd
->online
);
6967 free_cpumask_var(rd
->span
);
6972 static void init_defrootdomain(void)
6974 init_rootdomain(&def_root_domain
);
6976 atomic_set(&def_root_domain
.refcount
, 1);
6979 static struct root_domain
*alloc_rootdomain(void)
6981 struct root_domain
*rd
;
6983 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6987 if (init_rootdomain(rd
) != 0) {
6995 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
6997 struct sched_group
*tmp
, *first
;
7006 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
7011 } while (sg
!= first
);
7014 static void free_sched_domain(struct rcu_head
*rcu
)
7016 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
7019 * If its an overlapping domain it has private groups, iterate and
7022 if (sd
->flags
& SD_OVERLAP
) {
7023 free_sched_groups(sd
->groups
, 1);
7024 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
7025 kfree(sd
->groups
->sgp
);
7031 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
7033 call_rcu(&sd
->rcu
, free_sched_domain
);
7036 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
7038 for (; sd
; sd
= sd
->parent
)
7039 destroy_sched_domain(sd
, cpu
);
7043 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7044 * hold the hotplug lock.
7047 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7049 struct rq
*rq
= cpu_rq(cpu
);
7050 struct sched_domain
*tmp
;
7052 /* Remove the sched domains which do not contribute to scheduling. */
7053 for (tmp
= sd
; tmp
; ) {
7054 struct sched_domain
*parent
= tmp
->parent
;
7058 if (sd_parent_degenerate(tmp
, parent
)) {
7059 tmp
->parent
= parent
->parent
;
7061 parent
->parent
->child
= tmp
;
7062 destroy_sched_domain(parent
, cpu
);
7067 if (sd
&& sd_degenerate(sd
)) {
7070 destroy_sched_domain(tmp
, cpu
);
7075 sched_domain_debug(sd
, cpu
);
7077 rq_attach_root(rq
, rd
);
7079 rcu_assign_pointer(rq
->sd
, sd
);
7080 destroy_sched_domains(tmp
, cpu
);
7083 /* cpus with isolated domains */
7084 static cpumask_var_t cpu_isolated_map
;
7086 /* Setup the mask of cpus configured for isolated domains */
7087 static int __init
isolated_cpu_setup(char *str
)
7089 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
7090 cpulist_parse(str
, cpu_isolated_map
);
7094 __setup("isolcpus=", isolated_cpu_setup
);
7099 * find_next_best_node - find the next node to include in a sched_domain
7100 * @node: node whose sched_domain we're building
7101 * @used_nodes: nodes already in the sched_domain
7103 * Find the next node to include in a given scheduling domain. Simply
7104 * finds the closest node not already in the @used_nodes map.
7106 * Should use nodemask_t.
7108 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7110 int i
, n
, val
, min_val
, best_node
= -1;
7114 for (i
= 0; i
< nr_node_ids
; i
++) {
7115 /* Start at @node */
7116 n
= (node
+ i
) % nr_node_ids
;
7118 if (!nr_cpus_node(n
))
7121 /* Skip already used nodes */
7122 if (node_isset(n
, *used_nodes
))
7125 /* Simple min distance search */
7126 val
= node_distance(node
, n
);
7128 if (val
< min_val
) {
7134 if (best_node
!= -1)
7135 node_set(best_node
, *used_nodes
);
7140 * sched_domain_node_span - get a cpumask for a node's sched_domain
7141 * @node: node whose cpumask we're constructing
7142 * @span: resulting cpumask
7144 * Given a node, construct a good cpumask for its sched_domain to span. It
7145 * should be one that prevents unnecessary balancing, but also spreads tasks
7148 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7150 nodemask_t used_nodes
;
7153 cpumask_clear(span
);
7154 nodes_clear(used_nodes
);
7156 cpumask_or(span
, span
, cpumask_of_node(node
));
7157 node_set(node
, used_nodes
);
7159 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7160 int next_node
= find_next_best_node(node
, &used_nodes
);
7163 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7167 static const struct cpumask
*cpu_node_mask(int cpu
)
7169 lockdep_assert_held(&sched_domains_mutex
);
7171 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
7173 return sched_domains_tmpmask
;
7176 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
7178 return cpu_possible_mask
;
7180 #endif /* CONFIG_NUMA */
7182 static const struct cpumask
*cpu_cpu_mask(int cpu
)
7184 return cpumask_of_node(cpu_to_node(cpu
));
7187 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7190 struct sched_domain
**__percpu sd
;
7191 struct sched_group
**__percpu sg
;
7192 struct sched_group_power
**__percpu sgp
;
7196 struct sched_domain
** __percpu sd
;
7197 struct root_domain
*rd
;
7207 struct sched_domain_topology_level
;
7209 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
7210 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
7212 #define SDTL_OVERLAP 0x01
7214 struct sched_domain_topology_level
{
7215 sched_domain_init_f init
;
7216 sched_domain_mask_f mask
;
7218 struct sd_data data
;
7222 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
7224 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
7225 const struct cpumask
*span
= sched_domain_span(sd
);
7226 struct cpumask
*covered
= sched_domains_tmpmask
;
7227 struct sd_data
*sdd
= sd
->private;
7228 struct sched_domain
*child
;
7231 cpumask_clear(covered
);
7233 for_each_cpu(i
, span
) {
7234 struct cpumask
*sg_span
;
7236 if (cpumask_test_cpu(i
, covered
))
7239 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7240 GFP_KERNEL
, cpu_to_node(i
));
7245 sg_span
= sched_group_cpus(sg
);
7247 child
= *per_cpu_ptr(sdd
->sd
, i
);
7249 child
= child
->child
;
7250 cpumask_copy(sg_span
, sched_domain_span(child
));
7252 cpumask_set_cpu(i
, sg_span
);
7254 cpumask_or(covered
, covered
, sg_span
);
7256 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
7257 atomic_inc(&sg
->sgp
->ref
);
7259 if (cpumask_test_cpu(cpu
, sg_span
))
7269 sd
->groups
= groups
;
7274 free_sched_groups(first
, 0);
7279 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
7281 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
7282 struct sched_domain
*child
= sd
->child
;
7285 cpu
= cpumask_first(sched_domain_span(child
));
7288 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
7289 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
7290 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
7297 * build_sched_groups will build a circular linked list of the groups
7298 * covered by the given span, and will set each group's ->cpumask correctly,
7299 * and ->cpu_power to 0.
7301 * Assumes the sched_domain tree is fully constructed
7304 build_sched_groups(struct sched_domain
*sd
, int cpu
)
7306 struct sched_group
*first
= NULL
, *last
= NULL
;
7307 struct sd_data
*sdd
= sd
->private;
7308 const struct cpumask
*span
= sched_domain_span(sd
);
7309 struct cpumask
*covered
;
7312 get_group(cpu
, sdd
, &sd
->groups
);
7313 atomic_inc(&sd
->groups
->ref
);
7315 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
7318 lockdep_assert_held(&sched_domains_mutex
);
7319 covered
= sched_domains_tmpmask
;
7321 cpumask_clear(covered
);
7323 for_each_cpu(i
, span
) {
7324 struct sched_group
*sg
;
7325 int group
= get_group(i
, sdd
, &sg
);
7328 if (cpumask_test_cpu(i
, covered
))
7331 cpumask_clear(sched_group_cpus(sg
));
7334 for_each_cpu(j
, span
) {
7335 if (get_group(j
, sdd
, NULL
) != group
)
7338 cpumask_set_cpu(j
, covered
);
7339 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7354 * Initialize sched groups cpu_power.
7356 * cpu_power indicates the capacity of sched group, which is used while
7357 * distributing the load between different sched groups in a sched domain.
7358 * Typically cpu_power for all the groups in a sched domain will be same unless
7359 * there are asymmetries in the topology. If there are asymmetries, group
7360 * having more cpu_power will pickup more load compared to the group having
7363 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7365 struct sched_group
*sg
= sd
->groups
;
7367 WARN_ON(!sd
|| !sg
);
7370 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
7372 } while (sg
!= sd
->groups
);
7374 if (cpu
!= group_first_cpu(sg
))
7377 update_group_power(sd
, cpu
);
7381 * Initializers for schedule domains
7382 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7385 #ifdef CONFIG_SCHED_DEBUG
7386 # define SD_INIT_NAME(sd, type) sd->name = #type
7388 # define SD_INIT_NAME(sd, type) do { } while (0)
7391 #define SD_INIT_FUNC(type) \
7392 static noinline struct sched_domain * \
7393 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7395 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7396 *sd = SD_##type##_INIT; \
7397 SD_INIT_NAME(sd, type); \
7398 sd->private = &tl->data; \
7404 SD_INIT_FUNC(ALLNODES
)
7407 #ifdef CONFIG_SCHED_SMT
7408 SD_INIT_FUNC(SIBLING
)
7410 #ifdef CONFIG_SCHED_MC
7413 #ifdef CONFIG_SCHED_BOOK
7417 static int default_relax_domain_level
= -1;
7418 int sched_domain_level_max
;
7420 static int __init
setup_relax_domain_level(char *str
)
7424 val
= simple_strtoul(str
, NULL
, 0);
7425 if (val
< sched_domain_level_max
)
7426 default_relax_domain_level
= val
;
7430 __setup("relax_domain_level=", setup_relax_domain_level
);
7432 static void set_domain_attribute(struct sched_domain
*sd
,
7433 struct sched_domain_attr
*attr
)
7437 if (!attr
|| attr
->relax_domain_level
< 0) {
7438 if (default_relax_domain_level
< 0)
7441 request
= default_relax_domain_level
;
7443 request
= attr
->relax_domain_level
;
7444 if (request
< sd
->level
) {
7445 /* turn off idle balance on this domain */
7446 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7448 /* turn on idle balance on this domain */
7449 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7453 static void __sdt_free(const struct cpumask
*cpu_map
);
7454 static int __sdt_alloc(const struct cpumask
*cpu_map
);
7456 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7457 const struct cpumask
*cpu_map
)
7461 if (!atomic_read(&d
->rd
->refcount
))
7462 free_rootdomain(&d
->rd
->rcu
); /* fall through */
7464 free_percpu(d
->sd
); /* fall through */
7466 __sdt_free(cpu_map
); /* fall through */
7472 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7473 const struct cpumask
*cpu_map
)
7475 memset(d
, 0, sizeof(*d
));
7477 if (__sdt_alloc(cpu_map
))
7478 return sa_sd_storage
;
7479 d
->sd
= alloc_percpu(struct sched_domain
*);
7481 return sa_sd_storage
;
7482 d
->rd
= alloc_rootdomain();
7485 return sa_rootdomain
;
7489 * NULL the sd_data elements we've used to build the sched_domain and
7490 * sched_group structure so that the subsequent __free_domain_allocs()
7491 * will not free the data we're using.
7493 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
7495 struct sd_data
*sdd
= sd
->private;
7497 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
7498 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
7500 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
7501 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
7503 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
7504 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
7507 #ifdef CONFIG_SCHED_SMT
7508 static const struct cpumask
*cpu_smt_mask(int cpu
)
7510 return topology_thread_cpumask(cpu
);
7515 * Topology list, bottom-up.
7517 static struct sched_domain_topology_level default_topology
[] = {
7518 #ifdef CONFIG_SCHED_SMT
7519 { sd_init_SIBLING
, cpu_smt_mask
, },
7521 #ifdef CONFIG_SCHED_MC
7522 { sd_init_MC
, cpu_coregroup_mask
, },
7524 #ifdef CONFIG_SCHED_BOOK
7525 { sd_init_BOOK
, cpu_book_mask
, },
7527 { sd_init_CPU
, cpu_cpu_mask
, },
7529 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
7530 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
7535 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
7537 static int __sdt_alloc(const struct cpumask
*cpu_map
)
7539 struct sched_domain_topology_level
*tl
;
7542 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7543 struct sd_data
*sdd
= &tl
->data
;
7545 sdd
->sd
= alloc_percpu(struct sched_domain
*);
7549 sdd
->sg
= alloc_percpu(struct sched_group
*);
7553 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
7557 for_each_cpu(j
, cpu_map
) {
7558 struct sched_domain
*sd
;
7559 struct sched_group
*sg
;
7560 struct sched_group_power
*sgp
;
7562 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
7563 GFP_KERNEL
, cpu_to_node(j
));
7567 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
7569 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7570 GFP_KERNEL
, cpu_to_node(j
));
7574 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
7576 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
7577 GFP_KERNEL
, cpu_to_node(j
));
7581 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
7588 static void __sdt_free(const struct cpumask
*cpu_map
)
7590 struct sched_domain_topology_level
*tl
;
7593 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7594 struct sd_data
*sdd
= &tl
->data
;
7596 for_each_cpu(j
, cpu_map
) {
7597 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
7598 if (sd
&& (sd
->flags
& SD_OVERLAP
))
7599 free_sched_groups(sd
->groups
, 0);
7600 kfree(*per_cpu_ptr(sdd
->sd
, j
));
7601 kfree(*per_cpu_ptr(sdd
->sg
, j
));
7602 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
7604 free_percpu(sdd
->sd
);
7605 free_percpu(sdd
->sg
);
7606 free_percpu(sdd
->sgp
);
7610 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
7611 struct s_data
*d
, const struct cpumask
*cpu_map
,
7612 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
7615 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
7619 set_domain_attribute(sd
, attr
);
7620 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
7622 sd
->level
= child
->level
+ 1;
7623 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
7632 * Build sched domains for a given set of cpus and attach the sched domains
7633 * to the individual cpus
7635 static int build_sched_domains(const struct cpumask
*cpu_map
,
7636 struct sched_domain_attr
*attr
)
7638 enum s_alloc alloc_state
= sa_none
;
7639 struct sched_domain
*sd
;
7641 int i
, ret
= -ENOMEM
;
7643 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7644 if (alloc_state
!= sa_rootdomain
)
7647 /* Set up domains for cpus specified by the cpu_map. */
7648 for_each_cpu(i
, cpu_map
) {
7649 struct sched_domain_topology_level
*tl
;
7652 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7653 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
7654 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
7655 sd
->flags
|= SD_OVERLAP
;
7656 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
7663 *per_cpu_ptr(d
.sd
, i
) = sd
;
7666 /* Build the groups for the domains */
7667 for_each_cpu(i
, cpu_map
) {
7668 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7669 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
7670 if (sd
->flags
& SD_OVERLAP
) {
7671 if (build_overlap_sched_groups(sd
, i
))
7674 if (build_sched_groups(sd
, i
))
7680 /* Calculate CPU power for physical packages and nodes */
7681 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
7682 if (!cpumask_test_cpu(i
, cpu_map
))
7685 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7686 claim_allocations(i
, sd
);
7687 init_sched_groups_power(i
, sd
);
7691 /* Attach the domains */
7693 for_each_cpu(i
, cpu_map
) {
7694 sd
= *per_cpu_ptr(d
.sd
, i
);
7695 cpu_attach_domain(sd
, d
.rd
, i
);
7701 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7705 static cpumask_var_t
*doms_cur
; /* current sched domains */
7706 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7707 static struct sched_domain_attr
*dattr_cur
;
7708 /* attribues of custom domains in 'doms_cur' */
7711 * Special case: If a kmalloc of a doms_cur partition (array of
7712 * cpumask) fails, then fallback to a single sched domain,
7713 * as determined by the single cpumask fallback_doms.
7715 static cpumask_var_t fallback_doms
;
7718 * arch_update_cpu_topology lets virtualized architectures update the
7719 * cpu core maps. It is supposed to return 1 if the topology changed
7720 * or 0 if it stayed the same.
7722 int __attribute__((weak
)) arch_update_cpu_topology(void)
7727 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7730 cpumask_var_t
*doms
;
7732 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7735 for (i
= 0; i
< ndoms
; i
++) {
7736 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7737 free_sched_domains(doms
, i
);
7744 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7747 for (i
= 0; i
< ndoms
; i
++)
7748 free_cpumask_var(doms
[i
]);
7753 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7754 * For now this just excludes isolated cpus, but could be used to
7755 * exclude other special cases in the future.
7757 static int init_sched_domains(const struct cpumask
*cpu_map
)
7761 arch_update_cpu_topology();
7763 doms_cur
= alloc_sched_domains(ndoms_cur
);
7765 doms_cur
= &fallback_doms
;
7766 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7768 err
= build_sched_domains(doms_cur
[0], NULL
);
7769 register_sched_domain_sysctl();
7775 * Detach sched domains from a group of cpus specified in cpu_map
7776 * These cpus will now be attached to the NULL domain
7778 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7783 for_each_cpu(i
, cpu_map
)
7784 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7788 /* handle null as "default" */
7789 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7790 struct sched_domain_attr
*new, int idx_new
)
7792 struct sched_domain_attr tmp
;
7799 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7800 new ? (new + idx_new
) : &tmp
,
7801 sizeof(struct sched_domain_attr
));
7805 * Partition sched domains as specified by the 'ndoms_new'
7806 * cpumasks in the array doms_new[] of cpumasks. This compares
7807 * doms_new[] to the current sched domain partitioning, doms_cur[].
7808 * It destroys each deleted domain and builds each new domain.
7810 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7811 * The masks don't intersect (don't overlap.) We should setup one
7812 * sched domain for each mask. CPUs not in any of the cpumasks will
7813 * not be load balanced. If the same cpumask appears both in the
7814 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7817 * The passed in 'doms_new' should be allocated using
7818 * alloc_sched_domains. This routine takes ownership of it and will
7819 * free_sched_domains it when done with it. If the caller failed the
7820 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7821 * and partition_sched_domains() will fallback to the single partition
7822 * 'fallback_doms', it also forces the domains to be rebuilt.
7824 * If doms_new == NULL it will be replaced with cpu_online_mask.
7825 * ndoms_new == 0 is a special case for destroying existing domains,
7826 * and it will not create the default domain.
7828 * Call with hotplug lock held
7830 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7831 struct sched_domain_attr
*dattr_new
)
7836 mutex_lock(&sched_domains_mutex
);
7838 /* always unregister in case we don't destroy any domains */
7839 unregister_sched_domain_sysctl();
7841 /* Let architecture update cpu core mappings. */
7842 new_topology
= arch_update_cpu_topology();
7844 n
= doms_new
? ndoms_new
: 0;
7846 /* Destroy deleted domains */
7847 for (i
= 0; i
< ndoms_cur
; i
++) {
7848 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7849 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7850 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7853 /* no match - a current sched domain not in new doms_new[] */
7854 detach_destroy_domains(doms_cur
[i
]);
7859 if (doms_new
== NULL
) {
7861 doms_new
= &fallback_doms
;
7862 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7863 WARN_ON_ONCE(dattr_new
);
7866 /* Build new domains */
7867 for (i
= 0; i
< ndoms_new
; i
++) {
7868 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7869 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7870 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7873 /* no match - add a new doms_new */
7874 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7879 /* Remember the new sched domains */
7880 if (doms_cur
!= &fallback_doms
)
7881 free_sched_domains(doms_cur
, ndoms_cur
);
7882 kfree(dattr_cur
); /* kfree(NULL) is safe */
7883 doms_cur
= doms_new
;
7884 dattr_cur
= dattr_new
;
7885 ndoms_cur
= ndoms_new
;
7887 register_sched_domain_sysctl();
7889 mutex_unlock(&sched_domains_mutex
);
7892 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7893 static void reinit_sched_domains(void)
7897 /* Destroy domains first to force the rebuild */
7898 partition_sched_domains(0, NULL
, NULL
);
7900 rebuild_sched_domains();
7904 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7906 unsigned int level
= 0;
7908 if (sscanf(buf
, "%u", &level
) != 1)
7912 * level is always be positive so don't check for
7913 * level < POWERSAVINGS_BALANCE_NONE which is 0
7914 * What happens on 0 or 1 byte write,
7915 * need to check for count as well?
7918 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7922 sched_smt_power_savings
= level
;
7924 sched_mc_power_savings
= level
;
7926 reinit_sched_domains();
7931 #ifdef CONFIG_SCHED_MC
7932 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7933 struct sysdev_class_attribute
*attr
,
7936 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7938 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7939 struct sysdev_class_attribute
*attr
,
7940 const char *buf
, size_t count
)
7942 return sched_power_savings_store(buf
, count
, 0);
7944 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7945 sched_mc_power_savings_show
,
7946 sched_mc_power_savings_store
);
7949 #ifdef CONFIG_SCHED_SMT
7950 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7951 struct sysdev_class_attribute
*attr
,
7954 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7956 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7957 struct sysdev_class_attribute
*attr
,
7958 const char *buf
, size_t count
)
7960 return sched_power_savings_store(buf
, count
, 1);
7962 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7963 sched_smt_power_savings_show
,
7964 sched_smt_power_savings_store
);
7967 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7971 #ifdef CONFIG_SCHED_SMT
7973 err
= sysfs_create_file(&cls
->kset
.kobj
,
7974 &attr_sched_smt_power_savings
.attr
);
7976 #ifdef CONFIG_SCHED_MC
7977 if (!err
&& mc_capable())
7978 err
= sysfs_create_file(&cls
->kset
.kobj
,
7979 &attr_sched_mc_power_savings
.attr
);
7983 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7986 * Update cpusets according to cpu_active mask. If cpusets are
7987 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7988 * around partition_sched_domains().
7990 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7993 switch (action
& ~CPU_TASKS_FROZEN
) {
7995 case CPU_DOWN_FAILED
:
7996 cpuset_update_active_cpus();
8003 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
8006 switch (action
& ~CPU_TASKS_FROZEN
) {
8007 case CPU_DOWN_PREPARE
:
8008 cpuset_update_active_cpus();
8015 static int update_runtime(struct notifier_block
*nfb
,
8016 unsigned long action
, void *hcpu
)
8018 int cpu
= (int)(long)hcpu
;
8021 case CPU_DOWN_PREPARE
:
8022 case CPU_DOWN_PREPARE_FROZEN
:
8023 disable_runtime(cpu_rq(cpu
));
8026 case CPU_DOWN_FAILED
:
8027 case CPU_DOWN_FAILED_FROZEN
:
8029 case CPU_ONLINE_FROZEN
:
8030 enable_runtime(cpu_rq(cpu
));
8038 void __init
sched_init_smp(void)
8040 cpumask_var_t non_isolated_cpus
;
8042 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8043 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8046 mutex_lock(&sched_domains_mutex
);
8047 init_sched_domains(cpu_active_mask
);
8048 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8049 if (cpumask_empty(non_isolated_cpus
))
8050 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8051 mutex_unlock(&sched_domains_mutex
);
8054 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
8055 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
8057 /* RT runtime code needs to handle some hotplug events */
8058 hotcpu_notifier(update_runtime
, 0);
8062 /* Move init over to a non-isolated CPU */
8063 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8065 sched_init_granularity();
8066 free_cpumask_var(non_isolated_cpus
);
8068 init_sched_rt_class();
8071 void __init
sched_init_smp(void)
8073 sched_init_granularity();
8075 #endif /* CONFIG_SMP */
8077 const_debug
unsigned int sysctl_timer_migration
= 1;
8079 int in_sched_functions(unsigned long addr
)
8081 return in_lock_functions(addr
) ||
8082 (addr
>= (unsigned long)__sched_text_start
8083 && addr
< (unsigned long)__sched_text_end
);
8086 static void init_cfs_rq(struct cfs_rq
*cfs_rq
)
8088 cfs_rq
->tasks_timeline
= RB_ROOT
;
8089 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8090 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8091 #ifndef CONFIG_64BIT
8092 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
8096 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8098 struct rt_prio_array
*array
;
8101 array
= &rt_rq
->active
;
8102 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8103 INIT_LIST_HEAD(array
->queue
+ i
);
8104 __clear_bit(i
, array
->bitmap
);
8106 /* delimiter for bitsearch: */
8107 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8109 #if defined CONFIG_SMP
8110 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8111 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8112 rt_rq
->rt_nr_migratory
= 0;
8113 rt_rq
->overloaded
= 0;
8114 plist_head_init(&rt_rq
->pushable_tasks
);
8118 rt_rq
->rt_throttled
= 0;
8119 rt_rq
->rt_runtime
= 0;
8120 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
8123 #ifdef CONFIG_FAIR_GROUP_SCHED
8124 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8125 struct sched_entity
*se
, int cpu
,
8126 struct sched_entity
*parent
)
8128 struct rq
*rq
= cpu_rq(cpu
);
8133 /* allow initial update_cfs_load() to truncate */
8134 cfs_rq
->load_stamp
= 1;
8136 init_cfs_rq_runtime(cfs_rq
);
8138 tg
->cfs_rq
[cpu
] = cfs_rq
;
8141 /* se could be NULL for root_task_group */
8146 se
->cfs_rq
= &rq
->cfs
;
8148 se
->cfs_rq
= parent
->my_q
;
8151 update_load_set(&se
->load
, 0);
8152 se
->parent
= parent
;
8156 #ifdef CONFIG_RT_GROUP_SCHED
8157 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8158 struct sched_rt_entity
*rt_se
, int cpu
,
8159 struct sched_rt_entity
*parent
)
8161 struct rq
*rq
= cpu_rq(cpu
);
8163 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8164 rt_rq
->rt_nr_boosted
= 0;
8168 tg
->rt_rq
[cpu
] = rt_rq
;
8169 tg
->rt_se
[cpu
] = rt_se
;
8175 rt_se
->rt_rq
= &rq
->rt
;
8177 rt_se
->rt_rq
= parent
->my_q
;
8179 rt_se
->my_q
= rt_rq
;
8180 rt_se
->parent
= parent
;
8181 INIT_LIST_HEAD(&rt_se
->run_list
);
8185 void __init
sched_init(void)
8188 unsigned long alloc_size
= 0, ptr
;
8190 #ifdef CONFIG_FAIR_GROUP_SCHED
8191 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8193 #ifdef CONFIG_RT_GROUP_SCHED
8194 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8196 #ifdef CONFIG_CPUMASK_OFFSTACK
8197 alloc_size
+= num_possible_cpus() * cpumask_size();
8200 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
8202 #ifdef CONFIG_FAIR_GROUP_SCHED
8203 root_task_group
.se
= (struct sched_entity
**)ptr
;
8204 ptr
+= nr_cpu_ids
* sizeof(void **);
8206 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8207 ptr
+= nr_cpu_ids
* sizeof(void **);
8209 #endif /* CONFIG_FAIR_GROUP_SCHED */
8210 #ifdef CONFIG_RT_GROUP_SCHED
8211 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8212 ptr
+= nr_cpu_ids
* sizeof(void **);
8214 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8215 ptr
+= nr_cpu_ids
* sizeof(void **);
8217 #endif /* CONFIG_RT_GROUP_SCHED */
8218 #ifdef CONFIG_CPUMASK_OFFSTACK
8219 for_each_possible_cpu(i
) {
8220 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8221 ptr
+= cpumask_size();
8223 #endif /* CONFIG_CPUMASK_OFFSTACK */
8227 init_defrootdomain();
8230 init_rt_bandwidth(&def_rt_bandwidth
,
8231 global_rt_period(), global_rt_runtime());
8233 #ifdef CONFIG_RT_GROUP_SCHED
8234 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8235 global_rt_period(), global_rt_runtime());
8236 #endif /* CONFIG_RT_GROUP_SCHED */
8238 #ifdef CONFIG_CGROUP_SCHED
8239 list_add(&root_task_group
.list
, &task_groups
);
8240 INIT_LIST_HEAD(&root_task_group
.children
);
8241 autogroup_init(&init_task
);
8242 #endif /* CONFIG_CGROUP_SCHED */
8244 for_each_possible_cpu(i
) {
8248 raw_spin_lock_init(&rq
->lock
);
8250 rq
->calc_load_active
= 0;
8251 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8252 init_cfs_rq(&rq
->cfs
);
8253 init_rt_rq(&rq
->rt
, rq
);
8254 #ifdef CONFIG_FAIR_GROUP_SCHED
8255 root_task_group
.shares
= root_task_group_load
;
8256 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8258 * How much cpu bandwidth does root_task_group get?
8260 * In case of task-groups formed thr' the cgroup filesystem, it
8261 * gets 100% of the cpu resources in the system. This overall
8262 * system cpu resource is divided among the tasks of
8263 * root_task_group and its child task-groups in a fair manner,
8264 * based on each entity's (task or task-group's) weight
8265 * (se->load.weight).
8267 * In other words, if root_task_group has 10 tasks of weight
8268 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8269 * then A0's share of the cpu resource is:
8271 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8273 * We achieve this by letting root_task_group's tasks sit
8274 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8276 init_cfs_bandwidth(&root_task_group
.cfs_bandwidth
);
8277 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8278 #endif /* CONFIG_FAIR_GROUP_SCHED */
8280 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8281 #ifdef CONFIG_RT_GROUP_SCHED
8282 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8283 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8286 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8287 rq
->cpu_load
[j
] = 0;
8289 rq
->last_load_update_tick
= jiffies
;
8294 rq
->cpu_power
= SCHED_POWER_SCALE
;
8295 rq
->post_schedule
= 0;
8296 rq
->active_balance
= 0;
8297 rq
->next_balance
= jiffies
;
8302 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8303 rq_attach_root(rq
, &def_root_domain
);
8305 rq
->nohz_balance_kick
= 0;
8309 atomic_set(&rq
->nr_iowait
, 0);
8312 set_load_weight(&init_task
);
8314 #ifdef CONFIG_PREEMPT_NOTIFIERS
8315 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8319 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8322 #ifdef CONFIG_RT_MUTEXES
8323 plist_head_init(&init_task
.pi_waiters
);
8327 * The boot idle thread does lazy MMU switching as well:
8329 atomic_inc(&init_mm
.mm_count
);
8330 enter_lazy_tlb(&init_mm
, current
);
8333 * Make us the idle thread. Technically, schedule() should not be
8334 * called from this thread, however somewhere below it might be,
8335 * but because we are the idle thread, we just pick up running again
8336 * when this runqueue becomes "idle".
8338 init_idle(current
, smp_processor_id());
8340 calc_load_update
= jiffies
+ LOAD_FREQ
;
8343 * During early bootup we pretend to be a normal task:
8345 current
->sched_class
= &fair_sched_class
;
8348 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
8350 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8351 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8352 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8353 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8354 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8356 /* May be allocated at isolcpus cmdline parse time */
8357 if (cpu_isolated_map
== NULL
)
8358 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8361 scheduler_running
= 1;
8364 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8365 static inline int preempt_count_equals(int preempt_offset
)
8367 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8369 return (nested
== preempt_offset
);
8372 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8374 static unsigned long prev_jiffy
; /* ratelimiting */
8376 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8377 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8378 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8380 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8382 prev_jiffy
= jiffies
;
8385 "BUG: sleeping function called from invalid context at %s:%d\n",
8388 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8389 in_atomic(), irqs_disabled(),
8390 current
->pid
, current
->comm
);
8392 debug_show_held_locks(current
);
8393 if (irqs_disabled())
8394 print_irqtrace_events(current
);
8397 EXPORT_SYMBOL(__might_sleep
);
8400 #ifdef CONFIG_MAGIC_SYSRQ
8401 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8403 const struct sched_class
*prev_class
= p
->sched_class
;
8404 int old_prio
= p
->prio
;
8409 deactivate_task(rq
, p
, 0);
8410 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8412 activate_task(rq
, p
, 0);
8413 resched_task(rq
->curr
);
8416 check_class_changed(rq
, p
, prev_class
, old_prio
);
8419 void normalize_rt_tasks(void)
8421 struct task_struct
*g
, *p
;
8422 unsigned long flags
;
8425 read_lock_irqsave(&tasklist_lock
, flags
);
8426 do_each_thread(g
, p
) {
8428 * Only normalize user tasks:
8433 p
->se
.exec_start
= 0;
8434 #ifdef CONFIG_SCHEDSTATS
8435 p
->se
.statistics
.wait_start
= 0;
8436 p
->se
.statistics
.sleep_start
= 0;
8437 p
->se
.statistics
.block_start
= 0;
8442 * Renice negative nice level userspace
8445 if (TASK_NICE(p
) < 0 && p
->mm
)
8446 set_user_nice(p
, 0);
8450 raw_spin_lock(&p
->pi_lock
);
8451 rq
= __task_rq_lock(p
);
8453 normalize_task(rq
, p
);
8455 __task_rq_unlock(rq
);
8456 raw_spin_unlock(&p
->pi_lock
);
8457 } while_each_thread(g
, p
);
8459 read_unlock_irqrestore(&tasklist_lock
, flags
);
8462 #endif /* CONFIG_MAGIC_SYSRQ */
8464 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8466 * These functions are only useful for the IA64 MCA handling, or kdb.
8468 * They can only be called when the whole system has been
8469 * stopped - every CPU needs to be quiescent, and no scheduling
8470 * activity can take place. Using them for anything else would
8471 * be a serious bug, and as a result, they aren't even visible
8472 * under any other configuration.
8476 * curr_task - return the current task for a given cpu.
8477 * @cpu: the processor in question.
8479 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8481 struct task_struct
*curr_task(int cpu
)
8483 return cpu_curr(cpu
);
8486 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8490 * set_curr_task - set the current task for a given cpu.
8491 * @cpu: the processor in question.
8492 * @p: the task pointer to set.
8494 * Description: This function must only be used when non-maskable interrupts
8495 * are serviced on a separate stack. It allows the architecture to switch the
8496 * notion of the current task on a cpu in a non-blocking manner. This function
8497 * must be called with all CPU's synchronized, and interrupts disabled, the
8498 * and caller must save the original value of the current task (see
8499 * curr_task() above) and restore that value before reenabling interrupts and
8500 * re-starting the system.
8502 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8504 void set_curr_task(int cpu
, struct task_struct
*p
)
8511 #ifdef CONFIG_FAIR_GROUP_SCHED
8512 static void free_fair_sched_group(struct task_group
*tg
)
8516 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8518 for_each_possible_cpu(i
) {
8520 kfree(tg
->cfs_rq
[i
]);
8530 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8532 struct cfs_rq
*cfs_rq
;
8533 struct sched_entity
*se
;
8536 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8539 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8543 tg
->shares
= NICE_0_LOAD
;
8545 init_cfs_bandwidth(tg_cfs_bandwidth(tg
));
8547 for_each_possible_cpu(i
) {
8548 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8549 GFP_KERNEL
, cpu_to_node(i
));
8553 se
= kzalloc_node(sizeof(struct sched_entity
),
8554 GFP_KERNEL
, cpu_to_node(i
));
8558 init_cfs_rq(cfs_rq
);
8559 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8570 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8572 struct rq
*rq
= cpu_rq(cpu
);
8573 unsigned long flags
;
8576 * Only empty task groups can be destroyed; so we can speculatively
8577 * check on_list without danger of it being re-added.
8579 if (!tg
->cfs_rq
[cpu
]->on_list
)
8582 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8583 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8584 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8586 #else /* !CONFIG_FAIR_GROUP_SCHED */
8587 static inline void free_fair_sched_group(struct task_group
*tg
)
8592 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8597 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8600 #endif /* CONFIG_FAIR_GROUP_SCHED */
8602 #ifdef CONFIG_RT_GROUP_SCHED
8603 static void free_rt_sched_group(struct task_group
*tg
)
8608 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8610 for_each_possible_cpu(i
) {
8612 kfree(tg
->rt_rq
[i
]);
8614 kfree(tg
->rt_se
[i
]);
8622 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8624 struct rt_rq
*rt_rq
;
8625 struct sched_rt_entity
*rt_se
;
8628 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8631 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8635 init_rt_bandwidth(&tg
->rt_bandwidth
,
8636 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8638 for_each_possible_cpu(i
) {
8639 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8640 GFP_KERNEL
, cpu_to_node(i
));
8644 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8645 GFP_KERNEL
, cpu_to_node(i
));
8649 init_rt_rq(rt_rq
, cpu_rq(i
));
8650 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8651 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8661 #else /* !CONFIG_RT_GROUP_SCHED */
8662 static inline void free_rt_sched_group(struct task_group
*tg
)
8667 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8671 #endif /* CONFIG_RT_GROUP_SCHED */
8673 #ifdef CONFIG_CGROUP_SCHED
8674 static void free_sched_group(struct task_group
*tg
)
8676 free_fair_sched_group(tg
);
8677 free_rt_sched_group(tg
);
8682 /* allocate runqueue etc for a new task group */
8683 struct task_group
*sched_create_group(struct task_group
*parent
)
8685 struct task_group
*tg
;
8686 unsigned long flags
;
8688 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8690 return ERR_PTR(-ENOMEM
);
8692 if (!alloc_fair_sched_group(tg
, parent
))
8695 if (!alloc_rt_sched_group(tg
, parent
))
8698 spin_lock_irqsave(&task_group_lock
, flags
);
8699 list_add_rcu(&tg
->list
, &task_groups
);
8701 WARN_ON(!parent
); /* root should already exist */
8703 tg
->parent
= parent
;
8704 INIT_LIST_HEAD(&tg
->children
);
8705 list_add_rcu(&tg
->siblings
, &parent
->children
);
8706 spin_unlock_irqrestore(&task_group_lock
, flags
);
8711 free_sched_group(tg
);
8712 return ERR_PTR(-ENOMEM
);
8715 /* rcu callback to free various structures associated with a task group */
8716 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8718 /* now it should be safe to free those cfs_rqs */
8719 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8722 /* Destroy runqueue etc associated with a task group */
8723 void sched_destroy_group(struct task_group
*tg
)
8725 unsigned long flags
;
8728 /* end participation in shares distribution */
8729 for_each_possible_cpu(i
)
8730 unregister_fair_sched_group(tg
, i
);
8732 spin_lock_irqsave(&task_group_lock
, flags
);
8733 list_del_rcu(&tg
->list
);
8734 list_del_rcu(&tg
->siblings
);
8735 spin_unlock_irqrestore(&task_group_lock
, flags
);
8737 /* wait for possible concurrent references to cfs_rqs complete */
8738 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8741 /* change task's runqueue when it moves between groups.
8742 * The caller of this function should have put the task in its new group
8743 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8744 * reflect its new group.
8746 void sched_move_task(struct task_struct
*tsk
)
8749 unsigned long flags
;
8752 rq
= task_rq_lock(tsk
, &flags
);
8754 running
= task_current(rq
, tsk
);
8758 dequeue_task(rq
, tsk
, 0);
8759 if (unlikely(running
))
8760 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8762 #ifdef CONFIG_FAIR_GROUP_SCHED
8763 if (tsk
->sched_class
->task_move_group
)
8764 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8767 set_task_rq(tsk
, task_cpu(tsk
));
8769 if (unlikely(running
))
8770 tsk
->sched_class
->set_curr_task(rq
);
8772 enqueue_task(rq
, tsk
, 0);
8774 task_rq_unlock(rq
, tsk
, &flags
);
8776 #endif /* CONFIG_CGROUP_SCHED */
8778 #ifdef CONFIG_FAIR_GROUP_SCHED
8779 static DEFINE_MUTEX(shares_mutex
);
8781 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8784 unsigned long flags
;
8787 * We can't change the weight of the root cgroup.
8792 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8794 mutex_lock(&shares_mutex
);
8795 if (tg
->shares
== shares
)
8798 tg
->shares
= shares
;
8799 for_each_possible_cpu(i
) {
8800 struct rq
*rq
= cpu_rq(i
);
8801 struct sched_entity
*se
;
8804 /* Propagate contribution to hierarchy */
8805 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8806 for_each_sched_entity(se
)
8807 update_cfs_shares(group_cfs_rq(se
));
8808 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8812 mutex_unlock(&shares_mutex
);
8816 unsigned long sched_group_shares(struct task_group
*tg
)
8822 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8823 static unsigned long to_ratio(u64 period
, u64 runtime
)
8825 if (runtime
== RUNTIME_INF
)
8828 return div64_u64(runtime
<< 20, period
);
8832 #ifdef CONFIG_RT_GROUP_SCHED
8834 * Ensure that the real time constraints are schedulable.
8836 static DEFINE_MUTEX(rt_constraints_mutex
);
8838 /* Must be called with tasklist_lock held */
8839 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8841 struct task_struct
*g
, *p
;
8843 do_each_thread(g
, p
) {
8844 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8846 } while_each_thread(g
, p
);
8851 struct rt_schedulable_data
{
8852 struct task_group
*tg
;
8857 static int tg_rt_schedulable(struct task_group
*tg
, void *data
)
8859 struct rt_schedulable_data
*d
= data
;
8860 struct task_group
*child
;
8861 unsigned long total
, sum
= 0;
8862 u64 period
, runtime
;
8864 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8865 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8868 period
= d
->rt_period
;
8869 runtime
= d
->rt_runtime
;
8873 * Cannot have more runtime than the period.
8875 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8879 * Ensure we don't starve existing RT tasks.
8881 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8884 total
= to_ratio(period
, runtime
);
8887 * Nobody can have more than the global setting allows.
8889 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8893 * The sum of our children's runtime should not exceed our own.
8895 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8896 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8897 runtime
= child
->rt_bandwidth
.rt_runtime
;
8899 if (child
== d
->tg
) {
8900 period
= d
->rt_period
;
8901 runtime
= d
->rt_runtime
;
8904 sum
+= to_ratio(period
, runtime
);
8913 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8917 struct rt_schedulable_data data
= {
8919 .rt_period
= period
,
8920 .rt_runtime
= runtime
,
8924 ret
= walk_tg_tree(tg_rt_schedulable
, tg_nop
, &data
);
8930 static int tg_set_rt_bandwidth(struct task_group
*tg
,
8931 u64 rt_period
, u64 rt_runtime
)
8935 mutex_lock(&rt_constraints_mutex
);
8936 read_lock(&tasklist_lock
);
8937 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8941 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8942 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8943 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8945 for_each_possible_cpu(i
) {
8946 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8948 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8949 rt_rq
->rt_runtime
= rt_runtime
;
8950 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8952 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8954 read_unlock(&tasklist_lock
);
8955 mutex_unlock(&rt_constraints_mutex
);
8960 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8962 u64 rt_runtime
, rt_period
;
8964 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8965 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8966 if (rt_runtime_us
< 0)
8967 rt_runtime
= RUNTIME_INF
;
8969 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8972 long sched_group_rt_runtime(struct task_group
*tg
)
8976 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8979 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8980 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8981 return rt_runtime_us
;
8984 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8986 u64 rt_runtime
, rt_period
;
8988 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8989 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8994 return tg_set_rt_bandwidth(tg
, rt_period
, rt_runtime
);
8997 long sched_group_rt_period(struct task_group
*tg
)
9001 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9002 do_div(rt_period_us
, NSEC_PER_USEC
);
9003 return rt_period_us
;
9006 static int sched_rt_global_constraints(void)
9008 u64 runtime
, period
;
9011 if (sysctl_sched_rt_period
<= 0)
9014 runtime
= global_rt_runtime();
9015 period
= global_rt_period();
9018 * Sanity check on the sysctl variables.
9020 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9023 mutex_lock(&rt_constraints_mutex
);
9024 read_lock(&tasklist_lock
);
9025 ret
= __rt_schedulable(NULL
, 0, 0);
9026 read_unlock(&tasklist_lock
);
9027 mutex_unlock(&rt_constraints_mutex
);
9032 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9034 /* Don't accept realtime tasks when there is no way for them to run */
9035 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9041 #else /* !CONFIG_RT_GROUP_SCHED */
9042 static int sched_rt_global_constraints(void)
9044 unsigned long flags
;
9047 if (sysctl_sched_rt_period
<= 0)
9051 * There's always some RT tasks in the root group
9052 * -- migration, kstopmachine etc..
9054 if (sysctl_sched_rt_runtime
== 0)
9057 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9058 for_each_possible_cpu(i
) {
9059 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9061 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
9062 rt_rq
->rt_runtime
= global_rt_runtime();
9063 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
9065 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9069 #endif /* CONFIG_RT_GROUP_SCHED */
9071 int sched_rt_handler(struct ctl_table
*table
, int write
,
9072 void __user
*buffer
, size_t *lenp
,
9076 int old_period
, old_runtime
;
9077 static DEFINE_MUTEX(mutex
);
9080 old_period
= sysctl_sched_rt_period
;
9081 old_runtime
= sysctl_sched_rt_runtime
;
9083 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
9085 if (!ret
&& write
) {
9086 ret
= sched_rt_global_constraints();
9088 sysctl_sched_rt_period
= old_period
;
9089 sysctl_sched_rt_runtime
= old_runtime
;
9091 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9092 def_rt_bandwidth
.rt_period
=
9093 ns_to_ktime(global_rt_period());
9096 mutex_unlock(&mutex
);
9101 #ifdef CONFIG_CGROUP_SCHED
9103 /* return corresponding task_group object of a cgroup */
9104 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9106 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9107 struct task_group
, css
);
9110 static struct cgroup_subsys_state
*
9111 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9113 struct task_group
*tg
, *parent
;
9115 if (!cgrp
->parent
) {
9116 /* This is early initialization for the top cgroup */
9117 return &root_task_group
.css
;
9120 parent
= cgroup_tg(cgrp
->parent
);
9121 tg
= sched_create_group(parent
);
9123 return ERR_PTR(-ENOMEM
);
9129 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9131 struct task_group
*tg
= cgroup_tg(cgrp
);
9133 sched_destroy_group(tg
);
9137 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
9139 #ifdef CONFIG_RT_GROUP_SCHED
9140 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9143 /* We don't support RT-tasks being in separate groups */
9144 if (tsk
->sched_class
!= &fair_sched_class
)
9151 cpu_cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
9153 sched_move_task(tsk
);
9157 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9158 struct cgroup
*old_cgrp
, struct task_struct
*task
)
9161 * cgroup_exit() is called in the copy_process() failure path.
9162 * Ignore this case since the task hasn't ran yet, this avoids
9163 * trying to poke a half freed task state from generic code.
9165 if (!(task
->flags
& PF_EXITING
))
9168 sched_move_task(task
);
9171 #ifdef CONFIG_FAIR_GROUP_SCHED
9172 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9175 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
9178 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9180 struct task_group
*tg
= cgroup_tg(cgrp
);
9182 return (u64
) scale_load_down(tg
->shares
);
9185 #ifdef CONFIG_CFS_BANDWIDTH
9186 static DEFINE_MUTEX(cfs_constraints_mutex
);
9188 const u64 max_cfs_quota_period
= 1 * NSEC_PER_SEC
; /* 1s */
9189 const u64 min_cfs_quota_period
= 1 * NSEC_PER_MSEC
; /* 1ms */
9191 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
);
9193 static int tg_set_cfs_bandwidth(struct task_group
*tg
, u64 period
, u64 quota
)
9195 int i
, ret
= 0, runtime_enabled
;
9196 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9198 if (tg
== &root_task_group
)
9202 * Ensure we have at some amount of bandwidth every period. This is
9203 * to prevent reaching a state of large arrears when throttled via
9204 * entity_tick() resulting in prolonged exit starvation.
9206 if (quota
< min_cfs_quota_period
|| period
< min_cfs_quota_period
)
9210 * Likewise, bound things on the otherside by preventing insane quota
9211 * periods. This also allows us to normalize in computing quota
9214 if (period
> max_cfs_quota_period
)
9217 mutex_lock(&cfs_constraints_mutex
);
9218 ret
= __cfs_schedulable(tg
, period
, quota
);
9222 runtime_enabled
= quota
!= RUNTIME_INF
;
9223 raw_spin_lock_irq(&cfs_b
->lock
);
9224 cfs_b
->period
= ns_to_ktime(period
);
9225 cfs_b
->quota
= quota
;
9227 __refill_cfs_bandwidth_runtime(cfs_b
);
9228 /* restart the period timer (if active) to handle new period expiry */
9229 if (runtime_enabled
&& cfs_b
->timer_active
) {
9230 /* force a reprogram */
9231 cfs_b
->timer_active
= 0;
9232 __start_cfs_bandwidth(cfs_b
);
9234 raw_spin_unlock_irq(&cfs_b
->lock
);
9236 for_each_possible_cpu(i
) {
9237 struct cfs_rq
*cfs_rq
= tg
->cfs_rq
[i
];
9238 struct rq
*rq
= rq_of(cfs_rq
);
9240 raw_spin_lock_irq(&rq
->lock
);
9241 cfs_rq
->runtime_enabled
= runtime_enabled
;
9242 cfs_rq
->runtime_remaining
= 0;
9244 if (cfs_rq_throttled(cfs_rq
))
9245 unthrottle_cfs_rq(cfs_rq
);
9246 raw_spin_unlock_irq(&rq
->lock
);
9249 mutex_unlock(&cfs_constraints_mutex
);
9254 int tg_set_cfs_quota(struct task_group
*tg
, long cfs_quota_us
)
9258 period
= ktime_to_ns(tg_cfs_bandwidth(tg
)->period
);
9259 if (cfs_quota_us
< 0)
9260 quota
= RUNTIME_INF
;
9262 quota
= (u64
)cfs_quota_us
* NSEC_PER_USEC
;
9264 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9267 long tg_get_cfs_quota(struct task_group
*tg
)
9271 if (tg_cfs_bandwidth(tg
)->quota
== RUNTIME_INF
)
9274 quota_us
= tg_cfs_bandwidth(tg
)->quota
;
9275 do_div(quota_us
, NSEC_PER_USEC
);
9280 int tg_set_cfs_period(struct task_group
*tg
, long cfs_period_us
)
9284 period
= (u64
)cfs_period_us
* NSEC_PER_USEC
;
9285 quota
= tg_cfs_bandwidth(tg
)->quota
;
9290 return tg_set_cfs_bandwidth(tg
, period
, quota
);
9293 long tg_get_cfs_period(struct task_group
*tg
)
9297 cfs_period_us
= ktime_to_ns(tg_cfs_bandwidth(tg
)->period
);
9298 do_div(cfs_period_us
, NSEC_PER_USEC
);
9300 return cfs_period_us
;
9303 static s64
cpu_cfs_quota_read_s64(struct cgroup
*cgrp
, struct cftype
*cft
)
9305 return tg_get_cfs_quota(cgroup_tg(cgrp
));
9308 static int cpu_cfs_quota_write_s64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9311 return tg_set_cfs_quota(cgroup_tg(cgrp
), cfs_quota_us
);
9314 static u64
cpu_cfs_period_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9316 return tg_get_cfs_period(cgroup_tg(cgrp
));
9319 static int cpu_cfs_period_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9322 return tg_set_cfs_period(cgroup_tg(cgrp
), cfs_period_us
);
9325 struct cfs_schedulable_data
{
9326 struct task_group
*tg
;
9331 * normalize group quota/period to be quota/max_period
9332 * note: units are usecs
9334 static u64
normalize_cfs_quota(struct task_group
*tg
,
9335 struct cfs_schedulable_data
*d
)
9343 period
= tg_get_cfs_period(tg
);
9344 quota
= tg_get_cfs_quota(tg
);
9347 /* note: these should typically be equivalent */
9348 if (quota
== RUNTIME_INF
|| quota
== -1)
9351 return to_ratio(period
, quota
);
9354 static int tg_cfs_schedulable_down(struct task_group
*tg
, void *data
)
9356 struct cfs_schedulable_data
*d
= data
;
9357 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9358 s64 quota
= 0, parent_quota
= -1;
9361 quota
= RUNTIME_INF
;
9363 struct cfs_bandwidth
*parent_b
= tg_cfs_bandwidth(tg
->parent
);
9365 quota
= normalize_cfs_quota(tg
, d
);
9366 parent_quota
= parent_b
->hierarchal_quota
;
9369 * ensure max(child_quota) <= parent_quota, inherit when no
9372 if (quota
== RUNTIME_INF
)
9373 quota
= parent_quota
;
9374 else if (parent_quota
!= RUNTIME_INF
&& quota
> parent_quota
)
9377 cfs_b
->hierarchal_quota
= quota
;
9382 static int __cfs_schedulable(struct task_group
*tg
, u64 period
, u64 quota
)
9385 struct cfs_schedulable_data data
= {
9391 if (quota
!= RUNTIME_INF
) {
9392 do_div(data
.period
, NSEC_PER_USEC
);
9393 do_div(data
.quota
, NSEC_PER_USEC
);
9397 ret
= walk_tg_tree(tg_cfs_schedulable_down
, tg_nop
, &data
);
9403 static int cpu_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9404 struct cgroup_map_cb
*cb
)
9406 struct task_group
*tg
= cgroup_tg(cgrp
);
9407 struct cfs_bandwidth
*cfs_b
= tg_cfs_bandwidth(tg
);
9409 cb
->fill(cb
, "nr_periods", cfs_b
->nr_periods
);
9410 cb
->fill(cb
, "nr_throttled", cfs_b
->nr_throttled
);
9411 cb
->fill(cb
, "throttled_time", cfs_b
->throttled_time
);
9415 #endif /* CONFIG_CFS_BANDWIDTH */
9416 #endif /* CONFIG_FAIR_GROUP_SCHED */
9418 #ifdef CONFIG_RT_GROUP_SCHED
9419 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9422 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9425 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9427 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9430 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9433 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9436 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9438 return sched_group_rt_period(cgroup_tg(cgrp
));
9440 #endif /* CONFIG_RT_GROUP_SCHED */
9442 static struct cftype cpu_files
[] = {
9443 #ifdef CONFIG_FAIR_GROUP_SCHED
9446 .read_u64
= cpu_shares_read_u64
,
9447 .write_u64
= cpu_shares_write_u64
,
9450 #ifdef CONFIG_CFS_BANDWIDTH
9452 .name
= "cfs_quota_us",
9453 .read_s64
= cpu_cfs_quota_read_s64
,
9454 .write_s64
= cpu_cfs_quota_write_s64
,
9457 .name
= "cfs_period_us",
9458 .read_u64
= cpu_cfs_period_read_u64
,
9459 .write_u64
= cpu_cfs_period_write_u64
,
9463 .read_map
= cpu_stats_show
,
9466 #ifdef CONFIG_RT_GROUP_SCHED
9468 .name
= "rt_runtime_us",
9469 .read_s64
= cpu_rt_runtime_read
,
9470 .write_s64
= cpu_rt_runtime_write
,
9473 .name
= "rt_period_us",
9474 .read_u64
= cpu_rt_period_read_uint
,
9475 .write_u64
= cpu_rt_period_write_uint
,
9480 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9482 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9485 struct cgroup_subsys cpu_cgroup_subsys
= {
9487 .create
= cpu_cgroup_create
,
9488 .destroy
= cpu_cgroup_destroy
,
9489 .can_attach_task
= cpu_cgroup_can_attach_task
,
9490 .attach_task
= cpu_cgroup_attach_task
,
9491 .exit
= cpu_cgroup_exit
,
9492 .populate
= cpu_cgroup_populate
,
9493 .subsys_id
= cpu_cgroup_subsys_id
,
9497 #endif /* CONFIG_CGROUP_SCHED */
9499 #ifdef CONFIG_CGROUP_CPUACCT
9502 * CPU accounting code for task groups.
9504 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9505 * (balbir@in.ibm.com).
9508 /* track cpu usage of a group of tasks and its child groups */
9510 struct cgroup_subsys_state css
;
9511 /* cpuusage holds pointer to a u64-type object on every cpu */
9512 u64 __percpu
*cpuusage
;
9513 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9514 struct cpuacct
*parent
;
9517 struct cgroup_subsys cpuacct_subsys
;
9519 /* return cpu accounting group corresponding to this container */
9520 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9522 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9523 struct cpuacct
, css
);
9526 /* return cpu accounting group to which this task belongs */
9527 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9529 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9530 struct cpuacct
, css
);
9533 /* create a new cpu accounting group */
9534 static struct cgroup_subsys_state
*cpuacct_create(
9535 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9537 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9543 ca
->cpuusage
= alloc_percpu(u64
);
9547 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9548 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9549 goto out_free_counters
;
9552 ca
->parent
= cgroup_ca(cgrp
->parent
);
9558 percpu_counter_destroy(&ca
->cpustat
[i
]);
9559 free_percpu(ca
->cpuusage
);
9563 return ERR_PTR(-ENOMEM
);
9566 /* destroy an existing cpu accounting group */
9568 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9570 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9573 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9574 percpu_counter_destroy(&ca
->cpustat
[i
]);
9575 free_percpu(ca
->cpuusage
);
9579 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9581 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9584 #ifndef CONFIG_64BIT
9586 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9588 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9590 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9598 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9600 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9602 #ifndef CONFIG_64BIT
9604 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9606 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9608 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9614 /* return total cpu usage (in nanoseconds) of a group */
9615 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9617 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9618 u64 totalcpuusage
= 0;
9621 for_each_present_cpu(i
)
9622 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9624 return totalcpuusage
;
9627 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9630 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9639 for_each_present_cpu(i
)
9640 cpuacct_cpuusage_write(ca
, i
, 0);
9646 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9649 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9653 for_each_present_cpu(i
) {
9654 percpu
= cpuacct_cpuusage_read(ca
, i
);
9655 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9657 seq_printf(m
, "\n");
9661 static const char *cpuacct_stat_desc
[] = {
9662 [CPUACCT_STAT_USER
] = "user",
9663 [CPUACCT_STAT_SYSTEM
] = "system",
9666 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9667 struct cgroup_map_cb
*cb
)
9669 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9672 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9673 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9674 val
= cputime64_to_clock_t(val
);
9675 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9680 static struct cftype files
[] = {
9683 .read_u64
= cpuusage_read
,
9684 .write_u64
= cpuusage_write
,
9687 .name
= "usage_percpu",
9688 .read_seq_string
= cpuacct_percpu_seq_read
,
9692 .read_map
= cpuacct_stats_show
,
9696 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9698 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9702 * charge this task's execution time to its accounting group.
9704 * called with rq->lock held.
9706 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9711 if (unlikely(!cpuacct_subsys
.active
))
9714 cpu
= task_cpu(tsk
);
9720 for (; ca
; ca
= ca
->parent
) {
9721 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9722 *cpuusage
+= cputime
;
9729 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9730 * in cputime_t units. As a result, cpuacct_update_stats calls
9731 * percpu_counter_add with values large enough to always overflow the
9732 * per cpu batch limit causing bad SMP scalability.
9734 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9735 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9736 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9739 #define CPUACCT_BATCH \
9740 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9742 #define CPUACCT_BATCH 0
9746 * Charge the system/user time to the task's accounting group.
9748 static void cpuacct_update_stats(struct task_struct
*tsk
,
9749 enum cpuacct_stat_index idx
, cputime_t val
)
9752 int batch
= CPUACCT_BATCH
;
9754 if (unlikely(!cpuacct_subsys
.active
))
9761 __percpu_counter_add(&ca
->cpustat
[idx
],
9762 (__force s64
) val
, batch
);
9768 struct cgroup_subsys cpuacct_subsys
= {
9770 .create
= cpuacct_create
,
9771 .destroy
= cpuacct_destroy
,
9772 .populate
= cpuacct_populate
,
9773 .subsys_id
= cpuacct_subsys_id
,
9775 #endif /* CONFIG_CGROUP_CPUACCT */