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 <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy
)
126 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
131 static inline int task_has_rt_policy(struct task_struct
*p
)
133 return rt_policy(p
->policy
);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array
{
140 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
141 struct list_head queue
[MAX_RT_PRIO
];
144 struct rt_bandwidth
{
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock
;
149 struct hrtimer rt_period_timer
;
152 static struct rt_bandwidth def_rt_bandwidth
;
154 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
156 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
158 struct rt_bandwidth
*rt_b
=
159 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
165 now
= hrtimer_cb_get_time(timer
);
166 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
171 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
174 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
178 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
180 rt_b
->rt_period
= ns_to_ktime(period
);
181 rt_b
->rt_runtime
= runtime
;
183 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
185 hrtimer_init(&rt_b
->rt_period_timer
,
186 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
187 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime
>= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
199 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
202 if (hrtimer_active(&rt_b
->rt_period_timer
))
205 raw_spin_lock(&rt_b
->rt_runtime_lock
);
210 if (hrtimer_active(&rt_b
->rt_period_timer
))
213 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
214 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
216 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
217 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
218 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
219 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
220 HRTIMER_MODE_ABS_PINNED
, 0);
222 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
228 hrtimer_cancel(&rt_b
->rt_period_timer
);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex
);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups
);
246 /* task group related information */
248 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity
**se
;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq
**cfs_rq
;
255 unsigned long shares
;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity
**rt_se
;
260 struct rt_rq
**rt_rq
;
262 struct rt_bandwidth rt_bandwidth
;
266 struct list_head list
;
268 struct task_group
*parent
;
269 struct list_head siblings
;
270 struct list_head children
;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock
);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group
.children
);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group
;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load
;
315 unsigned long nr_running
;
320 struct rb_root tasks_timeline
;
321 struct rb_node
*rb_leftmost
;
323 struct list_head tasks
;
324 struct list_head
*balance_iterator
;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity
*curr
, *next
, *last
;
332 unsigned int nr_spread_over
;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list
;
346 struct task_group
*tg
; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight
;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load
;
363 * this cpu's part of tg->shares
365 unsigned long shares
;
368 * load.weight at the time we set shares
370 unsigned long rq_weight
;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active
;
378 unsigned long rt_nr_running
;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr
; /* highest queued rt task prio */
383 int next
; /* next highest */
388 unsigned long rt_nr_migratory
;
389 unsigned long rt_nr_total
;
391 struct plist_head pushable_tasks
;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock
;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted
;
403 struct list_head leaf_rt_rq_list
;
404 struct task_group
*tg
;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online
;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask
;
429 struct cpupri cpupri
;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain
;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running
;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
458 unsigned long last_load_update_tick
;
461 unsigned char nohz_balance_kick
;
463 unsigned int skip_clock_update
;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load
;
467 unsigned long nr_load_updates
;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list
;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list
;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible
;
489 struct task_struct
*curr
, *idle
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
498 struct root_domain
*rd
;
499 struct sched_domain
*sd
;
501 unsigned long cpu_power
;
503 unsigned char idle_at_tick
;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work
;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task
;
521 /* calc_load related fields */
522 unsigned long calc_load_update
;
523 long calc_load_active
;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending
;
528 struct call_single_data hrtick_csd
;
530 struct hrtimer hrtick_timer
;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info
;
536 unsigned long long rq_cpu_time
;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count
;
542 /* schedule() stats */
543 unsigned int sched_switch
;
544 unsigned int sched_count
;
545 unsigned int sched_goidle
;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count
;
549 unsigned int ttwu_local
;
552 unsigned int bkl_count
;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
559 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
561 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p
))
568 rq
->skip_clock_update
= 1;
571 static inline int cpu_of(struct rq
*rq
)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group
*task_group(struct task_struct
*p
)
613 struct cgroup_subsys_state
*css
;
615 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
616 lockdep_is_held(&task_rq(p
)->lock
));
617 return container_of(css
, struct task_group
, css
);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
625 p
->se
.parent
= task_group(p
)->se
[cpu
];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
630 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
637 static inline struct task_group
*task_group(struct task_struct
*p
)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq
*rq
)
646 if (!rq
->skip_clock_update
)
647 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu
)
669 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug
unsigned int sysctl_sched_features
=
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly
char *sched_feat_names
[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file
*m
, void *v
)
709 for (i
= 0; sched_feat_names
[i
]; i
++) {
710 if (!(sysctl_sched_features
& (1UL << i
)))
712 seq_printf(m
, "%s ", sched_feat_names
[i
]);
720 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
721 size_t cnt
, loff_t
*ppos
)
731 if (copy_from_user(&buf
, ubuf
, cnt
))
737 if (strncmp(buf
, "NO_", 3) == 0) {
742 for (i
= 0; sched_feat_names
[i
]; i
++) {
743 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
745 sysctl_sched_features
&= ~(1UL << i
);
747 sysctl_sched_features
|= (1UL << i
);
752 if (!sched_feat_names
[i
])
760 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
762 return single_open(filp
, sched_feat_show
, NULL
);
765 static const struct file_operations sched_feat_fops
= {
766 .open
= sched_feat_open
,
767 .write
= sched_feat_write
,
770 .release
= single_release
,
773 static __init
int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
780 late_initcall(sched_init_debug
);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit
= 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh
= 4;
807 * period over which we average the RT time consumption, measured
812 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period
= 1000000;
820 static __read_mostly
int scheduler_running
;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime
= 950000;
828 static inline u64
global_rt_period(void)
830 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
833 static inline u64
global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime
< 0)
838 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
850 return rq
->curr
== p
;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
856 return task_current(rq
, p
);
859 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
863 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq
->lock
.owner
= current
;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
876 raw_spin_unlock_irq(&rq
->lock
);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
885 return task_current(rq
, p
);
889 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq
->lock
);
902 raw_spin_unlock(&rq
->lock
);
906 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct
*p
)
929 return unlikely(p
->state
== TASK_WAKING
);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
943 raw_spin_lock(&rq
->lock
);
944 if (likely(rq
== task_rq(p
)))
946 raw_spin_unlock(&rq
->lock
);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
961 local_irq_save(*flags
);
963 raw_spin_lock(&rq
->lock
);
964 if (likely(rq
== task_rq(p
)))
966 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
970 static void __task_rq_unlock(struct rq
*rq
)
973 raw_spin_unlock(&rq
->lock
);
976 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
979 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq
*this_rq_lock(void)
992 raw_spin_lock(&rq
->lock
);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq
*rq
)
1016 if (!sched_feat(HRTICK
))
1018 if (!cpu_active(cpu_of(rq
)))
1020 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1023 static void hrtick_clear(struct rq
*rq
)
1025 if (hrtimer_active(&rq
->hrtick_timer
))
1026 hrtimer_cancel(&rq
->hrtick_timer
);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1035 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1037 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1039 raw_spin_lock(&rq
->lock
);
1040 update_rq_clock(rq
);
1041 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1042 raw_spin_unlock(&rq
->lock
);
1044 return HRTIMER_NORESTART
;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg
)
1053 struct rq
*rq
= arg
;
1055 raw_spin_lock(&rq
->lock
);
1056 hrtimer_restart(&rq
->hrtick_timer
);
1057 rq
->hrtick_csd_pending
= 0;
1058 raw_spin_unlock(&rq
->lock
);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq
*rq
, u64 delay
)
1068 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1069 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1071 hrtimer_set_expires(timer
, time
);
1073 if (rq
== this_rq()) {
1074 hrtimer_restart(timer
);
1075 } else if (!rq
->hrtick_csd_pending
) {
1076 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1077 rq
->hrtick_csd_pending
= 1;
1082 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1084 int cpu
= (int)(long)hcpu
;
1087 case CPU_UP_CANCELED
:
1088 case CPU_UP_CANCELED_FROZEN
:
1089 case CPU_DOWN_PREPARE
:
1090 case CPU_DOWN_PREPARE_FROZEN
:
1092 case CPU_DEAD_FROZEN
:
1093 hrtick_clear(cpu_rq(cpu
));
1100 static __init
void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick
, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq
*rq
, u64 delay
)
1112 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1113 HRTIMER_MODE_REL_PINNED
, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq
*rq
)
1124 rq
->hrtick_csd_pending
= 0;
1126 rq
->hrtick_csd
.flags
= 0;
1127 rq
->hrtick_csd
.func
= __hrtick_start
;
1128 rq
->hrtick_csd
.info
= rq
;
1131 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1132 rq
->hrtick_timer
.function
= hrtick
;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq
*rq
)
1139 static inline void init_rq_hrtick(struct rq
*rq
)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct
*p
)
1165 assert_raw_spin_locked(&task_rq(p
)->lock
);
1167 if (test_tsk_need_resched(p
))
1170 set_tsk_need_resched(p
);
1173 if (cpu
== smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p
))
1179 smp_send_reschedule(cpu
);
1182 static void resched_cpu(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long flags
;
1187 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1189 resched_task(cpu_curr(cpu
));
1190 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu
= smp_processor_id();
1206 struct sched_domain
*sd
;
1208 for_each_domain(cpu
, sd
) {
1209 for_each_cpu(i
, sched_domain_span(sd
))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu
)
1227 struct rq
*rq
= cpu_rq(cpu
);
1229 if (cpu
== smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq
->curr
!= rq
->idle
)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq
->idle
);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq
->idle
))
1252 smp_send_reschedule(cpu
);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64
sched_avg_period(void)
1259 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1262 static void sched_avg_update(struct rq
*rq
)
1264 s64 period
= sched_avg_period();
1266 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq
->age_stamp
));
1273 rq
->age_stamp
+= period
;
1278 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1280 rq
->rt_avg
+= rt_delta
;
1281 sched_avg_update(rq
);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct
*p
)
1287 assert_raw_spin_locked(&task_rq(p
)->lock
);
1288 set_tsk_need_resched(p
);
1291 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1295 static void sched_avg_update(struct rq
*rq
)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1318 struct load_weight
*lw
)
1322 if (!lw
->inv_weight
) {
1323 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1326 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1330 tmp
= (u64
)delta_exec
* weight
;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp
> WMULT_CONST
))
1335 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1338 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1340 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1343 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1349 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1356 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1357 * of tasks with abnormal "nice" values across CPUs the contribution that
1358 * each task makes to its run queue's load is weighted according to its
1359 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1360 * scaled version of the new time slice allocation that they receive on time
1364 #define WEIGHT_IDLEPRIO 3
1365 #define WMULT_IDLEPRIO 1431655765
1368 * Nice levels are multiplicative, with a gentle 10% change for every
1369 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1370 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1371 * that remained on nice 0.
1373 * The "10% effect" is relative and cumulative: from _any_ nice level,
1374 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1375 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1376 * If a task goes up by ~10% and another task goes down by ~10% then
1377 * the relative distance between them is ~25%.)
1379 static const int prio_to_weight
[40] = {
1380 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1381 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1382 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1383 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1384 /* 0 */ 1024, 820, 655, 526, 423,
1385 /* 5 */ 335, 272, 215, 172, 137,
1386 /* 10 */ 110, 87, 70, 56, 45,
1387 /* 15 */ 36, 29, 23, 18, 15,
1391 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1393 * In cases where the weight does not change often, we can use the
1394 * precalculated inverse to speed up arithmetics by turning divisions
1395 * into multiplications:
1397 static const u32 prio_to_wmult
[40] = {
1398 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1399 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1400 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1401 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1402 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1403 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1404 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1405 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1408 /* Time spent by the tasks of the cpu accounting group executing in ... */
1409 enum cpuacct_stat_index
{
1410 CPUACCT_STAT_USER
, /* ... user mode */
1411 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1413 CPUACCT_STAT_NSTATS
,
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1418 static void cpuacct_update_stats(struct task_struct
*tsk
,
1419 enum cpuacct_stat_index idx
, cputime_t val
);
1421 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1422 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1423 enum cpuacct_stat_index idx
, cputime_t val
) {}
1426 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1428 update_load_add(&rq
->load
, load
);
1431 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1433 update_load_sub(&rq
->load
, load
);
1436 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1437 typedef int (*tg_visitor
)(struct task_group
*, void *);
1440 * Iterate the full tree, calling @down when first entering a node and @up when
1441 * leaving it for the final time.
1443 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1445 struct task_group
*parent
, *child
;
1449 parent
= &root_task_group
;
1451 ret
= (*down
)(parent
, data
);
1454 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1461 ret
= (*up
)(parent
, data
);
1466 parent
= parent
->parent
;
1475 static int tg_nop(struct task_group
*tg
, void *data
)
1482 /* Used instead of source_load when we know the type == 0 */
1483 static unsigned long weighted_cpuload(const int cpu
)
1485 return cpu_rq(cpu
)->load
.weight
;
1489 * Return a low guess at the load of a migration-source cpu weighted
1490 * according to the scheduling class and "nice" value.
1492 * We want to under-estimate the load of migration sources, to
1493 * balance conservatively.
1495 static unsigned long source_load(int cpu
, int type
)
1497 struct rq
*rq
= cpu_rq(cpu
);
1498 unsigned long total
= weighted_cpuload(cpu
);
1500 if (type
== 0 || !sched_feat(LB_BIAS
))
1503 return min(rq
->cpu_load
[type
-1], total
);
1507 * Return a high guess at the load of a migration-target cpu weighted
1508 * according to the scheduling class and "nice" value.
1510 static unsigned long target_load(int cpu
, int type
)
1512 struct rq
*rq
= cpu_rq(cpu
);
1513 unsigned long total
= weighted_cpuload(cpu
);
1515 if (type
== 0 || !sched_feat(LB_BIAS
))
1518 return max(rq
->cpu_load
[type
-1], total
);
1521 static unsigned long power_of(int cpu
)
1523 return cpu_rq(cpu
)->cpu_power
;
1526 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1528 static unsigned long cpu_avg_load_per_task(int cpu
)
1530 struct rq
*rq
= cpu_rq(cpu
);
1531 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1534 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1536 rq
->avg_load_per_task
= 0;
1538 return rq
->avg_load_per_task
;
1541 #ifdef CONFIG_FAIR_GROUP_SCHED
1543 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1545 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1548 * Calculate and set the cpu's group shares.
1550 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1551 unsigned long sd_shares
,
1552 unsigned long sd_rq_weight
,
1553 unsigned long *usd_rq_weight
)
1555 unsigned long shares
, rq_weight
;
1558 rq_weight
= usd_rq_weight
[cpu
];
1561 rq_weight
= NICE_0_LOAD
;
1565 * \Sum_j shares_j * rq_weight_i
1566 * shares_i = -----------------------------
1567 * \Sum_j rq_weight_j
1569 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1570 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1572 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1573 sysctl_sched_shares_thresh
) {
1574 struct rq
*rq
= cpu_rq(cpu
);
1575 unsigned long flags
;
1577 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1578 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1579 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1580 __set_se_shares(tg
->se
[cpu
], shares
);
1581 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1586 * Re-compute the task group their per cpu shares over the given domain.
1587 * This needs to be done in a bottom-up fashion because the rq weight of a
1588 * parent group depends on the shares of its child groups.
1590 static int tg_shares_up(struct task_group
*tg
, void *data
)
1592 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1593 unsigned long *usd_rq_weight
;
1594 struct sched_domain
*sd
= data
;
1595 unsigned long flags
;
1601 local_irq_save(flags
);
1602 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1604 for_each_cpu(i
, sched_domain_span(sd
)) {
1605 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1606 usd_rq_weight
[i
] = weight
;
1608 rq_weight
+= weight
;
1610 * If there are currently no tasks on the cpu pretend there
1611 * is one of average load so that when a new task gets to
1612 * run here it will not get delayed by group starvation.
1615 weight
= NICE_0_LOAD
;
1617 sum_weight
+= weight
;
1618 shares
+= tg
->cfs_rq
[i
]->shares
;
1622 rq_weight
= sum_weight
;
1624 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1625 shares
= tg
->shares
;
1627 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1628 shares
= tg
->shares
;
1630 for_each_cpu(i
, sched_domain_span(sd
))
1631 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1633 local_irq_restore(flags
);
1639 * Compute the cpu's hierarchical load factor for each task group.
1640 * This needs to be done in a top-down fashion because the load of a child
1641 * group is a fraction of its parents load.
1643 static int tg_load_down(struct task_group
*tg
, void *data
)
1646 long cpu
= (long)data
;
1649 load
= cpu_rq(cpu
)->load
.weight
;
1651 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1652 load
*= tg
->cfs_rq
[cpu
]->shares
;
1653 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1656 tg
->cfs_rq
[cpu
]->h_load
= load
;
1661 static void update_shares(struct sched_domain
*sd
)
1666 if (root_task_group_empty())
1669 now
= local_clock();
1670 elapsed
= now
- sd
->last_update
;
1672 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1673 sd
->last_update
= now
;
1674 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1678 static void update_h_load(long cpu
)
1680 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1685 static inline void update_shares(struct sched_domain
*sd
)
1691 #ifdef CONFIG_PREEMPT
1693 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1696 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1697 * way at the expense of forcing extra atomic operations in all
1698 * invocations. This assures that the double_lock is acquired using the
1699 * same underlying policy as the spinlock_t on this architecture, which
1700 * reduces latency compared to the unfair variant below. However, it
1701 * also adds more overhead and therefore may reduce throughput.
1703 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1704 __releases(this_rq
->lock
)
1705 __acquires(busiest
->lock
)
1706 __acquires(this_rq
->lock
)
1708 raw_spin_unlock(&this_rq
->lock
);
1709 double_rq_lock(this_rq
, busiest
);
1716 * Unfair double_lock_balance: Optimizes throughput at the expense of
1717 * latency by eliminating extra atomic operations when the locks are
1718 * already in proper order on entry. This favors lower cpu-ids and will
1719 * grant the double lock to lower cpus over higher ids under contention,
1720 * regardless of entry order into the function.
1722 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1723 __releases(this_rq
->lock
)
1724 __acquires(busiest
->lock
)
1725 __acquires(this_rq
->lock
)
1729 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1730 if (busiest
< this_rq
) {
1731 raw_spin_unlock(&this_rq
->lock
);
1732 raw_spin_lock(&busiest
->lock
);
1733 raw_spin_lock_nested(&this_rq
->lock
,
1734 SINGLE_DEPTH_NESTING
);
1737 raw_spin_lock_nested(&busiest
->lock
,
1738 SINGLE_DEPTH_NESTING
);
1743 #endif /* CONFIG_PREEMPT */
1746 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1748 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1750 if (unlikely(!irqs_disabled())) {
1751 /* printk() doesn't work good under rq->lock */
1752 raw_spin_unlock(&this_rq
->lock
);
1756 return _double_lock_balance(this_rq
, busiest
);
1759 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1760 __releases(busiest
->lock
)
1762 raw_spin_unlock(&busiest
->lock
);
1763 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1767 * double_rq_lock - safely lock two runqueues
1769 * Note this does not disable interrupts like task_rq_lock,
1770 * you need to do so manually before calling.
1772 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1773 __acquires(rq1
->lock
)
1774 __acquires(rq2
->lock
)
1776 BUG_ON(!irqs_disabled());
1778 raw_spin_lock(&rq1
->lock
);
1779 __acquire(rq2
->lock
); /* Fake it out ;) */
1782 raw_spin_lock(&rq1
->lock
);
1783 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1785 raw_spin_lock(&rq2
->lock
);
1786 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1792 * double_rq_unlock - safely unlock two runqueues
1794 * Note this does not restore interrupts like task_rq_unlock,
1795 * you need to do so manually after calling.
1797 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1798 __releases(rq1
->lock
)
1799 __releases(rq2
->lock
)
1801 raw_spin_unlock(&rq1
->lock
);
1803 raw_spin_unlock(&rq2
->lock
);
1805 __release(rq2
->lock
);
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1814 cfs_rq
->shares
= shares
;
1819 static void calc_load_account_idle(struct rq
*this_rq
);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1822 static void update_cpu_load(struct rq
*this_rq
);
1824 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1826 set_task_rq(p
, cpu
);
1829 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1830 * successfuly executed on another CPU. We must ensure that updates of
1831 * per-task data have been completed by this moment.
1834 task_thread_info(p
)->cpu
= cpu
;
1838 static const struct sched_class rt_sched_class
;
1840 #define sched_class_highest (&rt_sched_class)
1841 #define for_each_class(class) \
1842 for (class = sched_class_highest; class; class = class->next)
1844 #include "sched_stats.h"
1846 static void inc_nr_running(struct rq
*rq
)
1851 static void dec_nr_running(struct rq
*rq
)
1856 static void set_load_weight(struct task_struct
*p
)
1858 if (task_has_rt_policy(p
)) {
1859 p
->se
.load
.weight
= 0;
1860 p
->se
.load
.inv_weight
= WMULT_CONST
;
1865 * SCHED_IDLE tasks get minimal weight:
1867 if (p
->policy
== SCHED_IDLE
) {
1868 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1869 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1873 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1874 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1877 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1879 update_rq_clock(rq
);
1880 sched_info_queued(p
);
1881 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1885 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1887 update_rq_clock(rq
);
1888 sched_info_dequeued(p
);
1889 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1894 * activate_task - move a task to the runqueue.
1896 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1898 if (task_contributes_to_load(p
))
1899 rq
->nr_uninterruptible
--;
1901 enqueue_task(rq
, p
, flags
);
1906 * deactivate_task - remove a task from the runqueue.
1908 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1910 if (task_contributes_to_load(p
))
1911 rq
->nr_uninterruptible
++;
1913 dequeue_task(rq
, p
, flags
);
1917 #include "sched_idletask.c"
1918 #include "sched_fair.c"
1919 #include "sched_rt.c"
1920 #ifdef CONFIG_SCHED_DEBUG
1921 # include "sched_debug.c"
1925 * __normal_prio - return the priority that is based on the static prio
1927 static inline int __normal_prio(struct task_struct
*p
)
1929 return p
->static_prio
;
1933 * Calculate the expected normal priority: i.e. priority
1934 * without taking RT-inheritance into account. Might be
1935 * boosted by interactivity modifiers. Changes upon fork,
1936 * setprio syscalls, and whenever the interactivity
1937 * estimator recalculates.
1939 static inline int normal_prio(struct task_struct
*p
)
1943 if (task_has_rt_policy(p
))
1944 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1946 prio
= __normal_prio(p
);
1951 * Calculate the current priority, i.e. the priority
1952 * taken into account by the scheduler. This value might
1953 * be boosted by RT tasks, or might be boosted by
1954 * interactivity modifiers. Will be RT if the task got
1955 * RT-boosted. If not then it returns p->normal_prio.
1957 static int effective_prio(struct task_struct
*p
)
1959 p
->normal_prio
= normal_prio(p
);
1961 * If we are RT tasks or we were boosted to RT priority,
1962 * keep the priority unchanged. Otherwise, update priority
1963 * to the normal priority:
1965 if (!rt_prio(p
->prio
))
1966 return p
->normal_prio
;
1971 * task_curr - is this task currently executing on a CPU?
1972 * @p: the task in question.
1974 inline int task_curr(const struct task_struct
*p
)
1976 return cpu_curr(task_cpu(p
)) == p
;
1979 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1980 const struct sched_class
*prev_class
,
1981 int oldprio
, int running
)
1983 if (prev_class
!= p
->sched_class
) {
1984 if (prev_class
->switched_from
)
1985 prev_class
->switched_from(rq
, p
, running
);
1986 p
->sched_class
->switched_to(rq
, p
, running
);
1988 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1993 * Is this task likely cache-hot:
1996 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2000 if (p
->sched_class
!= &fair_sched_class
)
2004 * Buddy candidates are cache hot:
2006 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2007 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2008 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2011 if (sysctl_sched_migration_cost
== -1)
2013 if (sysctl_sched_migration_cost
== 0)
2016 delta
= now
- p
->se
.exec_start
;
2018 return delta
< (s64
)sysctl_sched_migration_cost
;
2021 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2023 #ifdef CONFIG_SCHED_DEBUG
2025 * We should never call set_task_cpu() on a blocked task,
2026 * ttwu() will sort out the placement.
2028 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2029 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2032 trace_sched_migrate_task(p
, new_cpu
);
2034 if (task_cpu(p
) != new_cpu
) {
2035 p
->se
.nr_migrations
++;
2036 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2039 __set_task_cpu(p
, new_cpu
);
2042 struct migration_arg
{
2043 struct task_struct
*task
;
2047 static int migration_cpu_stop(void *data
);
2050 * The task's runqueue lock must be held.
2051 * Returns true if you have to wait for migration thread.
2053 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2055 struct rq
*rq
= task_rq(p
);
2058 * If the task is not on a runqueue (and not running), then
2059 * the next wake-up will properly place the task.
2061 return p
->se
.on_rq
|| task_running(rq
, p
);
2065 * wait_task_inactive - wait for a thread to unschedule.
2067 * If @match_state is nonzero, it's the @p->state value just checked and
2068 * not expected to change. If it changes, i.e. @p might have woken up,
2069 * then return zero. When we succeed in waiting for @p to be off its CPU,
2070 * we return a positive number (its total switch count). If a second call
2071 * a short while later returns the same number, the caller can be sure that
2072 * @p has remained unscheduled the whole time.
2074 * The caller must ensure that the task *will* unschedule sometime soon,
2075 * else this function might spin for a *long* time. This function can't
2076 * be called with interrupts off, or it may introduce deadlock with
2077 * smp_call_function() if an IPI is sent by the same process we are
2078 * waiting to become inactive.
2080 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2082 unsigned long flags
;
2089 * We do the initial early heuristics without holding
2090 * any task-queue locks at all. We'll only try to get
2091 * the runqueue lock when things look like they will
2097 * If the task is actively running on another CPU
2098 * still, just relax and busy-wait without holding
2101 * NOTE! Since we don't hold any locks, it's not
2102 * even sure that "rq" stays as the right runqueue!
2103 * But we don't care, since "task_running()" will
2104 * return false if the runqueue has changed and p
2105 * is actually now running somewhere else!
2107 while (task_running(rq
, p
)) {
2108 if (match_state
&& unlikely(p
->state
!= match_state
))
2114 * Ok, time to look more closely! We need the rq
2115 * lock now, to be *sure*. If we're wrong, we'll
2116 * just go back and repeat.
2118 rq
= task_rq_lock(p
, &flags
);
2119 trace_sched_wait_task(p
);
2120 running
= task_running(rq
, p
);
2121 on_rq
= p
->se
.on_rq
;
2123 if (!match_state
|| p
->state
== match_state
)
2124 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2125 task_rq_unlock(rq
, &flags
);
2128 * If it changed from the expected state, bail out now.
2130 if (unlikely(!ncsw
))
2134 * Was it really running after all now that we
2135 * checked with the proper locks actually held?
2137 * Oops. Go back and try again..
2139 if (unlikely(running
)) {
2145 * It's not enough that it's not actively running,
2146 * it must be off the runqueue _entirely_, and not
2149 * So if it was still runnable (but just not actively
2150 * running right now), it's preempted, and we should
2151 * yield - it could be a while.
2153 if (unlikely(on_rq
)) {
2154 schedule_timeout_uninterruptible(1);
2159 * Ahh, all good. It wasn't running, and it wasn't
2160 * runnable, which means that it will never become
2161 * running in the future either. We're all done!
2170 * kick_process - kick a running thread to enter/exit the kernel
2171 * @p: the to-be-kicked thread
2173 * Cause a process which is running on another CPU to enter
2174 * kernel-mode, without any delay. (to get signals handled.)
2176 * NOTE: this function doesnt have to take the runqueue lock,
2177 * because all it wants to ensure is that the remote task enters
2178 * the kernel. If the IPI races and the task has been migrated
2179 * to another CPU then no harm is done and the purpose has been
2182 void kick_process(struct task_struct
*p
)
2188 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2189 smp_send_reschedule(cpu
);
2192 EXPORT_SYMBOL_GPL(kick_process
);
2193 #endif /* CONFIG_SMP */
2196 * task_oncpu_function_call - call a function on the cpu on which a task runs
2197 * @p: the task to evaluate
2198 * @func: the function to be called
2199 * @info: the function call argument
2201 * Calls the function @func when the task is currently running. This might
2202 * be on the current CPU, which just calls the function directly
2204 void task_oncpu_function_call(struct task_struct
*p
,
2205 void (*func
) (void *info
), void *info
)
2212 smp_call_function_single(cpu
, func
, info
, 1);
2218 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2220 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2223 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2225 /* Look for allowed, online CPU in same node. */
2226 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2227 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2230 /* Any allowed, online CPU? */
2231 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2232 if (dest_cpu
< nr_cpu_ids
)
2235 /* No more Mr. Nice Guy. */
2236 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2237 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2239 * Don't tell them about moving exiting tasks or
2240 * kernel threads (both mm NULL), since they never
2243 if (p
->mm
&& printk_ratelimit()) {
2244 printk(KERN_INFO
"process %d (%s) no "
2245 "longer affine to cpu%d\n",
2246 task_pid_nr(p
), p
->comm
, cpu
);
2254 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2257 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2259 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2262 * In order not to call set_task_cpu() on a blocking task we need
2263 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2266 * Since this is common to all placement strategies, this lives here.
2268 * [ this allows ->select_task() to simply return task_cpu(p) and
2269 * not worry about this generic constraint ]
2271 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2273 cpu
= select_fallback_rq(task_cpu(p
), p
);
2278 static void update_avg(u64
*avg
, u64 sample
)
2280 s64 diff
= sample
- *avg
;
2285 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2286 bool is_sync
, bool is_migrate
, bool is_local
,
2287 unsigned long en_flags
)
2289 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2291 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2293 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2295 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2297 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2299 activate_task(rq
, p
, en_flags
);
2302 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2303 int wake_flags
, bool success
)
2305 trace_sched_wakeup(p
, success
);
2306 check_preempt_curr(rq
, p
, wake_flags
);
2308 p
->state
= TASK_RUNNING
;
2310 if (p
->sched_class
->task_woken
)
2311 p
->sched_class
->task_woken(rq
, p
);
2313 if (unlikely(rq
->idle_stamp
)) {
2314 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2315 u64 max
= 2*sysctl_sched_migration_cost
;
2320 update_avg(&rq
->avg_idle
, delta
);
2324 /* if a worker is waking up, notify workqueue */
2325 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2326 wq_worker_waking_up(p
, cpu_of(rq
));
2330 * try_to_wake_up - wake up a thread
2331 * @p: the thread to be awakened
2332 * @state: the mask of task states that can be woken
2333 * @wake_flags: wake modifier flags (WF_*)
2335 * Put it on the run-queue if it's not already there. The "current"
2336 * thread is always on the run-queue (except when the actual
2337 * re-schedule is in progress), and as such you're allowed to do
2338 * the simpler "current->state = TASK_RUNNING" to mark yourself
2339 * runnable without the overhead of this.
2341 * Returns %true if @p was woken up, %false if it was already running
2342 * or @state didn't match @p's state.
2344 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2347 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2348 unsigned long flags
;
2349 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2352 this_cpu
= get_cpu();
2355 rq
= task_rq_lock(p
, &flags
);
2356 if (!(p
->state
& state
))
2366 if (unlikely(task_running(rq
, p
)))
2370 * In order to handle concurrent wakeups and release the rq->lock
2371 * we put the task in TASK_WAKING state.
2373 * First fix up the nr_uninterruptible count:
2375 if (task_contributes_to_load(p
)) {
2376 if (likely(cpu_online(orig_cpu
)))
2377 rq
->nr_uninterruptible
--;
2379 this_rq()->nr_uninterruptible
--;
2381 p
->state
= TASK_WAKING
;
2383 if (p
->sched_class
->task_waking
) {
2384 p
->sched_class
->task_waking(rq
, p
);
2385 en_flags
|= ENQUEUE_WAKING
;
2388 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2389 if (cpu
!= orig_cpu
)
2390 set_task_cpu(p
, cpu
);
2391 __task_rq_unlock(rq
);
2394 raw_spin_lock(&rq
->lock
);
2397 * We migrated the task without holding either rq->lock, however
2398 * since the task is not on the task list itself, nobody else
2399 * will try and migrate the task, hence the rq should match the
2400 * cpu we just moved it to.
2402 WARN_ON(task_cpu(p
) != cpu
);
2403 WARN_ON(p
->state
!= TASK_WAKING
);
2405 #ifdef CONFIG_SCHEDSTATS
2406 schedstat_inc(rq
, ttwu_count
);
2407 if (cpu
== this_cpu
)
2408 schedstat_inc(rq
, ttwu_local
);
2410 struct sched_domain
*sd
;
2411 for_each_domain(this_cpu
, sd
) {
2412 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2413 schedstat_inc(sd
, ttwu_wake_remote
);
2418 #endif /* CONFIG_SCHEDSTATS */
2421 #endif /* CONFIG_SMP */
2422 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2423 cpu
== this_cpu
, en_flags
);
2426 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2428 task_rq_unlock(rq
, &flags
);
2435 * try_to_wake_up_local - try to wake up a local task with rq lock held
2436 * @p: the thread to be awakened
2438 * Put @p on the run-queue if it's not alredy there. The caller must
2439 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2440 * the current task. this_rq() stays locked over invocation.
2442 static void try_to_wake_up_local(struct task_struct
*p
)
2444 struct rq
*rq
= task_rq(p
);
2445 bool success
= false;
2447 BUG_ON(rq
!= this_rq());
2448 BUG_ON(p
== current
);
2449 lockdep_assert_held(&rq
->lock
);
2451 if (!(p
->state
& TASK_NORMAL
))
2455 if (likely(!task_running(rq
, p
))) {
2456 schedstat_inc(rq
, ttwu_count
);
2457 schedstat_inc(rq
, ttwu_local
);
2459 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2462 ttwu_post_activation(p
, rq
, 0, success
);
2466 * wake_up_process - Wake up a specific process
2467 * @p: The process to be woken up.
2469 * Attempt to wake up the nominated process and move it to the set of runnable
2470 * processes. Returns 1 if the process was woken up, 0 if it was already
2473 * It may be assumed that this function implies a write memory barrier before
2474 * changing the task state if and only if any tasks are woken up.
2476 int wake_up_process(struct task_struct
*p
)
2478 return try_to_wake_up(p
, TASK_ALL
, 0);
2480 EXPORT_SYMBOL(wake_up_process
);
2482 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2484 return try_to_wake_up(p
, state
, 0);
2488 * Perform scheduler related setup for a newly forked process p.
2489 * p is forked by current.
2491 * __sched_fork() is basic setup used by init_idle() too:
2493 static void __sched_fork(struct task_struct
*p
)
2495 p
->se
.exec_start
= 0;
2496 p
->se
.sum_exec_runtime
= 0;
2497 p
->se
.prev_sum_exec_runtime
= 0;
2498 p
->se
.nr_migrations
= 0;
2500 #ifdef CONFIG_SCHEDSTATS
2501 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2504 INIT_LIST_HEAD(&p
->rt
.run_list
);
2506 INIT_LIST_HEAD(&p
->se
.group_node
);
2508 #ifdef CONFIG_PREEMPT_NOTIFIERS
2509 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2514 * fork()/clone()-time setup:
2516 void sched_fork(struct task_struct
*p
, int clone_flags
)
2518 int cpu
= get_cpu();
2522 * We mark the process as running here. This guarantees that
2523 * nobody will actually run it, and a signal or other external
2524 * event cannot wake it up and insert it on the runqueue either.
2526 p
->state
= TASK_RUNNING
;
2529 * Revert to default priority/policy on fork if requested.
2531 if (unlikely(p
->sched_reset_on_fork
)) {
2532 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2533 p
->policy
= SCHED_NORMAL
;
2534 p
->normal_prio
= p
->static_prio
;
2537 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2538 p
->static_prio
= NICE_TO_PRIO(0);
2539 p
->normal_prio
= p
->static_prio
;
2544 * We don't need the reset flag anymore after the fork. It has
2545 * fulfilled its duty:
2547 p
->sched_reset_on_fork
= 0;
2551 * Make sure we do not leak PI boosting priority to the child.
2553 p
->prio
= current
->normal_prio
;
2555 if (!rt_prio(p
->prio
))
2556 p
->sched_class
= &fair_sched_class
;
2558 if (p
->sched_class
->task_fork
)
2559 p
->sched_class
->task_fork(p
);
2562 * The child is not yet in the pid-hash so no cgroup attach races,
2563 * and the cgroup is pinned to this child due to cgroup_fork()
2564 * is ran before sched_fork().
2566 * Silence PROVE_RCU.
2569 set_task_cpu(p
, cpu
);
2572 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2573 if (likely(sched_info_on()))
2574 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2576 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2579 #ifdef CONFIG_PREEMPT
2580 /* Want to start with kernel preemption disabled. */
2581 task_thread_info(p
)->preempt_count
= 1;
2583 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2589 * wake_up_new_task - wake up a newly created task for the first time.
2591 * This function will do some initial scheduler statistics housekeeping
2592 * that must be done for every newly created context, then puts the task
2593 * on the runqueue and wakes it.
2595 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2597 unsigned long flags
;
2599 int cpu __maybe_unused
= get_cpu();
2602 rq
= task_rq_lock(p
, &flags
);
2603 p
->state
= TASK_WAKING
;
2606 * Fork balancing, do it here and not earlier because:
2607 * - cpus_allowed can change in the fork path
2608 * - any previously selected cpu might disappear through hotplug
2610 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2611 * without people poking at ->cpus_allowed.
2613 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2614 set_task_cpu(p
, cpu
);
2616 p
->state
= TASK_RUNNING
;
2617 task_rq_unlock(rq
, &flags
);
2620 rq
= task_rq_lock(p
, &flags
);
2621 activate_task(rq
, p
, 0);
2622 trace_sched_wakeup_new(p
, 1);
2623 check_preempt_curr(rq
, p
, WF_FORK
);
2625 if (p
->sched_class
->task_woken
)
2626 p
->sched_class
->task_woken(rq
, p
);
2628 task_rq_unlock(rq
, &flags
);
2632 #ifdef CONFIG_PREEMPT_NOTIFIERS
2635 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2636 * @notifier: notifier struct to register
2638 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2640 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2642 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2645 * preempt_notifier_unregister - no longer interested in preemption notifications
2646 * @notifier: notifier struct to unregister
2648 * This is safe to call from within a preemption notifier.
2650 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2652 hlist_del(¬ifier
->link
);
2654 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2656 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2658 struct preempt_notifier
*notifier
;
2659 struct hlist_node
*node
;
2661 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2662 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2666 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2667 struct task_struct
*next
)
2669 struct preempt_notifier
*notifier
;
2670 struct hlist_node
*node
;
2672 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2673 notifier
->ops
->sched_out(notifier
, next
);
2676 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2678 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2683 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2684 struct task_struct
*next
)
2688 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2691 * prepare_task_switch - prepare to switch tasks
2692 * @rq: the runqueue preparing to switch
2693 * @prev: the current task that is being switched out
2694 * @next: the task we are going to switch to.
2696 * This is called with the rq lock held and interrupts off. It must
2697 * be paired with a subsequent finish_task_switch after the context
2700 * prepare_task_switch sets up locking and calls architecture specific
2704 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2705 struct task_struct
*next
)
2707 fire_sched_out_preempt_notifiers(prev
, next
);
2708 prepare_lock_switch(rq
, next
);
2709 prepare_arch_switch(next
);
2713 * finish_task_switch - clean up after a task-switch
2714 * @rq: runqueue associated with task-switch
2715 * @prev: the thread we just switched away from.
2717 * finish_task_switch must be called after the context switch, paired
2718 * with a prepare_task_switch call before the context switch.
2719 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2720 * and do any other architecture-specific cleanup actions.
2722 * Note that we may have delayed dropping an mm in context_switch(). If
2723 * so, we finish that here outside of the runqueue lock. (Doing it
2724 * with the lock held can cause deadlocks; see schedule() for
2727 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2728 __releases(rq
->lock
)
2730 struct mm_struct
*mm
= rq
->prev_mm
;
2736 * A task struct has one reference for the use as "current".
2737 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2738 * schedule one last time. The schedule call will never return, and
2739 * the scheduled task must drop that reference.
2740 * The test for TASK_DEAD must occur while the runqueue locks are
2741 * still held, otherwise prev could be scheduled on another cpu, die
2742 * there before we look at prev->state, and then the reference would
2744 * Manfred Spraul <manfred@colorfullife.com>
2746 prev_state
= prev
->state
;
2747 finish_arch_switch(prev
);
2748 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2749 local_irq_disable();
2750 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2751 perf_event_task_sched_in(current
);
2752 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2754 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2755 finish_lock_switch(rq
, prev
);
2757 fire_sched_in_preempt_notifiers(current
);
2760 if (unlikely(prev_state
== TASK_DEAD
)) {
2762 * Remove function-return probe instances associated with this
2763 * task and put them back on the free list.
2765 kprobe_flush_task(prev
);
2766 put_task_struct(prev
);
2772 /* assumes rq->lock is held */
2773 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2775 if (prev
->sched_class
->pre_schedule
)
2776 prev
->sched_class
->pre_schedule(rq
, prev
);
2779 /* rq->lock is NOT held, but preemption is disabled */
2780 static inline void post_schedule(struct rq
*rq
)
2782 if (rq
->post_schedule
) {
2783 unsigned long flags
;
2785 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2786 if (rq
->curr
->sched_class
->post_schedule
)
2787 rq
->curr
->sched_class
->post_schedule(rq
);
2788 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2790 rq
->post_schedule
= 0;
2796 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2800 static inline void post_schedule(struct rq
*rq
)
2807 * schedule_tail - first thing a freshly forked thread must call.
2808 * @prev: the thread we just switched away from.
2810 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2811 __releases(rq
->lock
)
2813 struct rq
*rq
= this_rq();
2815 finish_task_switch(rq
, prev
);
2818 * FIXME: do we need to worry about rq being invalidated by the
2823 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2824 /* In this case, finish_task_switch does not reenable preemption */
2827 if (current
->set_child_tid
)
2828 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2832 * context_switch - switch to the new MM and the new
2833 * thread's register state.
2836 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2837 struct task_struct
*next
)
2839 struct mm_struct
*mm
, *oldmm
;
2841 prepare_task_switch(rq
, prev
, next
);
2842 trace_sched_switch(prev
, next
);
2844 oldmm
= prev
->active_mm
;
2846 * For paravirt, this is coupled with an exit in switch_to to
2847 * combine the page table reload and the switch backend into
2850 arch_start_context_switch(prev
);
2853 next
->active_mm
= oldmm
;
2854 atomic_inc(&oldmm
->mm_count
);
2855 enter_lazy_tlb(oldmm
, next
);
2857 switch_mm(oldmm
, mm
, next
);
2860 prev
->active_mm
= NULL
;
2861 rq
->prev_mm
= oldmm
;
2864 * Since the runqueue lock will be released by the next
2865 * task (which is an invalid locking op but in the case
2866 * of the scheduler it's an obvious special-case), so we
2867 * do an early lockdep release here:
2869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2870 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2873 /* Here we just switch the register state and the stack. */
2874 switch_to(prev
, next
, prev
);
2878 * this_rq must be evaluated again because prev may have moved
2879 * CPUs since it called schedule(), thus the 'rq' on its stack
2880 * frame will be invalid.
2882 finish_task_switch(this_rq(), prev
);
2886 * nr_running, nr_uninterruptible and nr_context_switches:
2888 * externally visible scheduler statistics: current number of runnable
2889 * threads, current number of uninterruptible-sleeping threads, total
2890 * number of context switches performed since bootup.
2892 unsigned long nr_running(void)
2894 unsigned long i
, sum
= 0;
2896 for_each_online_cpu(i
)
2897 sum
+= cpu_rq(i
)->nr_running
;
2902 unsigned long nr_uninterruptible(void)
2904 unsigned long i
, sum
= 0;
2906 for_each_possible_cpu(i
)
2907 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2910 * Since we read the counters lockless, it might be slightly
2911 * inaccurate. Do not allow it to go below zero though:
2913 if (unlikely((long)sum
< 0))
2919 unsigned long long nr_context_switches(void)
2922 unsigned long long sum
= 0;
2924 for_each_possible_cpu(i
)
2925 sum
+= cpu_rq(i
)->nr_switches
;
2930 unsigned long nr_iowait(void)
2932 unsigned long i
, sum
= 0;
2934 for_each_possible_cpu(i
)
2935 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2940 unsigned long nr_iowait_cpu(int cpu
)
2942 struct rq
*this = cpu_rq(cpu
);
2943 return atomic_read(&this->nr_iowait
);
2946 unsigned long this_cpu_load(void)
2948 struct rq
*this = this_rq();
2949 return this->cpu_load
[0];
2953 /* Variables and functions for calc_load */
2954 static atomic_long_t calc_load_tasks
;
2955 static unsigned long calc_load_update
;
2956 unsigned long avenrun
[3];
2957 EXPORT_SYMBOL(avenrun
);
2959 static long calc_load_fold_active(struct rq
*this_rq
)
2961 long nr_active
, delta
= 0;
2963 nr_active
= this_rq
->nr_running
;
2964 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2966 if (nr_active
!= this_rq
->calc_load_active
) {
2967 delta
= nr_active
- this_rq
->calc_load_active
;
2968 this_rq
->calc_load_active
= nr_active
;
2976 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2978 * When making the ILB scale, we should try to pull this in as well.
2980 static atomic_long_t calc_load_tasks_idle
;
2982 static void calc_load_account_idle(struct rq
*this_rq
)
2986 delta
= calc_load_fold_active(this_rq
);
2988 atomic_long_add(delta
, &calc_load_tasks_idle
);
2991 static long calc_load_fold_idle(void)
2996 * Its got a race, we don't care...
2998 if (atomic_long_read(&calc_load_tasks_idle
))
2999 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3004 static void calc_load_account_idle(struct rq
*this_rq
)
3008 static inline long calc_load_fold_idle(void)
3015 * get_avenrun - get the load average array
3016 * @loads: pointer to dest load array
3017 * @offset: offset to add
3018 * @shift: shift count to shift the result left
3020 * These values are estimates at best, so no need for locking.
3022 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3024 loads
[0] = (avenrun
[0] + offset
) << shift
;
3025 loads
[1] = (avenrun
[1] + offset
) << shift
;
3026 loads
[2] = (avenrun
[2] + offset
) << shift
;
3029 static unsigned long
3030 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3033 load
+= active
* (FIXED_1
- exp
);
3034 return load
>> FSHIFT
;
3038 * calc_load - update the avenrun load estimates 10 ticks after the
3039 * CPUs have updated calc_load_tasks.
3041 void calc_global_load(void)
3043 unsigned long upd
= calc_load_update
+ 10;
3046 if (time_before(jiffies
, upd
))
3049 active
= atomic_long_read(&calc_load_tasks
);
3050 active
= active
> 0 ? active
* FIXED_1
: 0;
3052 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3053 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3054 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3056 calc_load_update
+= LOAD_FREQ
;
3060 * Called from update_cpu_load() to periodically update this CPU's
3063 static void calc_load_account_active(struct rq
*this_rq
)
3067 if (time_before(jiffies
, this_rq
->calc_load_update
))
3070 delta
= calc_load_fold_active(this_rq
);
3071 delta
+= calc_load_fold_idle();
3073 atomic_long_add(delta
, &calc_load_tasks
);
3075 this_rq
->calc_load_update
+= LOAD_FREQ
;
3079 * The exact cpuload at various idx values, calculated at every tick would be
3080 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3082 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3083 * on nth tick when cpu may be busy, then we have:
3084 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3085 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3087 * decay_load_missed() below does efficient calculation of
3088 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3089 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3091 * The calculation is approximated on a 128 point scale.
3092 * degrade_zero_ticks is the number of ticks after which load at any
3093 * particular idx is approximated to be zero.
3094 * degrade_factor is a precomputed table, a row for each load idx.
3095 * Each column corresponds to degradation factor for a power of two ticks,
3096 * based on 128 point scale.
3098 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3099 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3101 * With this power of 2 load factors, we can degrade the load n times
3102 * by looking at 1 bits in n and doing as many mult/shift instead of
3103 * n mult/shifts needed by the exact degradation.
3105 #define DEGRADE_SHIFT 7
3106 static const unsigned char
3107 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3108 static const unsigned char
3109 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3110 {0, 0, 0, 0, 0, 0, 0, 0},
3111 {64, 32, 8, 0, 0, 0, 0, 0},
3112 {96, 72, 40, 12, 1, 0, 0},
3113 {112, 98, 75, 43, 15, 1, 0},
3114 {120, 112, 98, 76, 45, 16, 2} };
3117 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3118 * would be when CPU is idle and so we just decay the old load without
3119 * adding any new load.
3121 static unsigned long
3122 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3126 if (!missed_updates
)
3129 if (missed_updates
>= degrade_zero_ticks
[idx
])
3133 return load
>> missed_updates
;
3135 while (missed_updates
) {
3136 if (missed_updates
% 2)
3137 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3139 missed_updates
>>= 1;
3146 * Update rq->cpu_load[] statistics. This function is usually called every
3147 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3148 * every tick. We fix it up based on jiffies.
3150 static void update_cpu_load(struct rq
*this_rq
)
3152 unsigned long this_load
= this_rq
->load
.weight
;
3153 unsigned long curr_jiffies
= jiffies
;
3154 unsigned long pending_updates
;
3157 this_rq
->nr_load_updates
++;
3159 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3160 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3163 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3164 this_rq
->last_load_update_tick
= curr_jiffies
;
3166 /* Update our load: */
3167 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3168 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3169 unsigned long old_load
, new_load
;
3171 /* scale is effectively 1 << i now, and >> i divides by scale */
3173 old_load
= this_rq
->cpu_load
[i
];
3174 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3175 new_load
= this_load
;
3177 * Round up the averaging division if load is increasing. This
3178 * prevents us from getting stuck on 9 if the load is 10, for
3181 if (new_load
> old_load
)
3182 new_load
+= scale
- 1;
3184 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3187 sched_avg_update(this_rq
);
3190 static void update_cpu_load_active(struct rq
*this_rq
)
3192 update_cpu_load(this_rq
);
3194 calc_load_account_active(this_rq
);
3200 * sched_exec - execve() is a valuable balancing opportunity, because at
3201 * this point the task has the smallest effective memory and cache footprint.
3203 void sched_exec(void)
3205 struct task_struct
*p
= current
;
3206 unsigned long flags
;
3210 rq
= task_rq_lock(p
, &flags
);
3211 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3212 if (dest_cpu
== smp_processor_id())
3216 * select_task_rq() can race against ->cpus_allowed
3218 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3219 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3220 struct migration_arg arg
= { p
, dest_cpu
};
3222 task_rq_unlock(rq
, &flags
);
3223 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3227 task_rq_unlock(rq
, &flags
);
3232 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3234 EXPORT_PER_CPU_SYMBOL(kstat
);
3237 * Return any ns on the sched_clock that have not yet been accounted in
3238 * @p in case that task is currently running.
3240 * Called with task_rq_lock() held on @rq.
3242 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3246 if (task_current(rq
, p
)) {
3247 update_rq_clock(rq
);
3248 ns
= rq
->clock
- p
->se
.exec_start
;
3256 unsigned long long task_delta_exec(struct task_struct
*p
)
3258 unsigned long flags
;
3262 rq
= task_rq_lock(p
, &flags
);
3263 ns
= do_task_delta_exec(p
, rq
);
3264 task_rq_unlock(rq
, &flags
);
3270 * Return accounted runtime for the task.
3271 * In case the task is currently running, return the runtime plus current's
3272 * pending runtime that have not been accounted yet.
3274 unsigned long long task_sched_runtime(struct task_struct
*p
)
3276 unsigned long flags
;
3280 rq
= task_rq_lock(p
, &flags
);
3281 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3282 task_rq_unlock(rq
, &flags
);
3288 * Return sum_exec_runtime for the thread group.
3289 * In case the task is currently running, return the sum plus current's
3290 * pending runtime that have not been accounted yet.
3292 * Note that the thread group might have other running tasks as well,
3293 * so the return value not includes other pending runtime that other
3294 * running tasks might have.
3296 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3298 struct task_cputime totals
;
3299 unsigned long flags
;
3303 rq
= task_rq_lock(p
, &flags
);
3304 thread_group_cputime(p
, &totals
);
3305 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3306 task_rq_unlock(rq
, &flags
);
3312 * Account user cpu time to a process.
3313 * @p: the process that the cpu time gets accounted to
3314 * @cputime: the cpu time spent in user space since the last update
3315 * @cputime_scaled: cputime scaled by cpu frequency
3317 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3318 cputime_t cputime_scaled
)
3320 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3323 /* Add user time to process. */
3324 p
->utime
= cputime_add(p
->utime
, cputime
);
3325 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3326 account_group_user_time(p
, cputime
);
3328 /* Add user time to cpustat. */
3329 tmp
= cputime_to_cputime64(cputime
);
3330 if (TASK_NICE(p
) > 0)
3331 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3333 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3335 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3336 /* Account for user time used */
3337 acct_update_integrals(p
);
3341 * Account guest cpu time to a process.
3342 * @p: the process that the cpu time gets accounted to
3343 * @cputime: the cpu time spent in virtual machine since the last update
3344 * @cputime_scaled: cputime scaled by cpu frequency
3346 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3347 cputime_t cputime_scaled
)
3350 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3352 tmp
= cputime_to_cputime64(cputime
);
3354 /* Add guest time to process. */
3355 p
->utime
= cputime_add(p
->utime
, cputime
);
3356 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3357 account_group_user_time(p
, cputime
);
3358 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3360 /* Add guest time to cpustat. */
3361 if (TASK_NICE(p
) > 0) {
3362 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3363 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3365 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3366 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3371 * Account system cpu time to a process.
3372 * @p: the process that the cpu time gets accounted to
3373 * @hardirq_offset: the offset to subtract from hardirq_count()
3374 * @cputime: the cpu time spent in kernel space since the last update
3375 * @cputime_scaled: cputime scaled by cpu frequency
3377 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3378 cputime_t cputime
, cputime_t cputime_scaled
)
3380 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3383 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3384 account_guest_time(p
, cputime
, cputime_scaled
);
3388 /* Add system time to process. */
3389 p
->stime
= cputime_add(p
->stime
, cputime
);
3390 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3391 account_group_system_time(p
, cputime
);
3393 /* Add system time to cpustat. */
3394 tmp
= cputime_to_cputime64(cputime
);
3395 if (hardirq_count() - hardirq_offset
)
3396 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3397 else if (softirq_count())
3398 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3400 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3402 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3404 /* Account for system time used */
3405 acct_update_integrals(p
);
3409 * Account for involuntary wait time.
3410 * @steal: the cpu time spent in involuntary wait
3412 void account_steal_time(cputime_t cputime
)
3414 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3415 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3417 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3421 * Account for idle time.
3422 * @cputime: the cpu time spent in idle wait
3424 void account_idle_time(cputime_t cputime
)
3426 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3427 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3428 struct rq
*rq
= this_rq();
3430 if (atomic_read(&rq
->nr_iowait
) > 0)
3431 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3433 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3436 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3439 * Account a single tick of cpu time.
3440 * @p: the process that the cpu time gets accounted to
3441 * @user_tick: indicates if the tick is a user or a system tick
3443 void account_process_tick(struct task_struct
*p
, int user_tick
)
3445 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3446 struct rq
*rq
= this_rq();
3449 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3450 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3451 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3454 account_idle_time(cputime_one_jiffy
);
3458 * Account multiple ticks of steal time.
3459 * @p: the process from which the cpu time has been stolen
3460 * @ticks: number of stolen ticks
3462 void account_steal_ticks(unsigned long ticks
)
3464 account_steal_time(jiffies_to_cputime(ticks
));
3468 * Account multiple ticks of idle time.
3469 * @ticks: number of stolen ticks
3471 void account_idle_ticks(unsigned long ticks
)
3473 account_idle_time(jiffies_to_cputime(ticks
));
3479 * Use precise platform statistics if available:
3481 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3482 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3488 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3490 struct task_cputime cputime
;
3492 thread_group_cputime(p
, &cputime
);
3494 *ut
= cputime
.utime
;
3495 *st
= cputime
.stime
;
3499 #ifndef nsecs_to_cputime
3500 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3503 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3505 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3508 * Use CFS's precise accounting:
3510 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3515 temp
= (u64
)(rtime
* utime
);
3516 do_div(temp
, total
);
3517 utime
= (cputime_t
)temp
;
3522 * Compare with previous values, to keep monotonicity:
3524 p
->prev_utime
= max(p
->prev_utime
, utime
);
3525 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3527 *ut
= p
->prev_utime
;
3528 *st
= p
->prev_stime
;
3532 * Must be called with siglock held.
3534 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3536 struct signal_struct
*sig
= p
->signal
;
3537 struct task_cputime cputime
;
3538 cputime_t rtime
, utime
, total
;
3540 thread_group_cputime(p
, &cputime
);
3542 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3543 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3548 temp
= (u64
)(rtime
* cputime
.utime
);
3549 do_div(temp
, total
);
3550 utime
= (cputime_t
)temp
;
3554 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3555 sig
->prev_stime
= max(sig
->prev_stime
,
3556 cputime_sub(rtime
, sig
->prev_utime
));
3558 *ut
= sig
->prev_utime
;
3559 *st
= sig
->prev_stime
;
3564 * This function gets called by the timer code, with HZ frequency.
3565 * We call it with interrupts disabled.
3567 * It also gets called by the fork code, when changing the parent's
3570 void scheduler_tick(void)
3572 int cpu
= smp_processor_id();
3573 struct rq
*rq
= cpu_rq(cpu
);
3574 struct task_struct
*curr
= rq
->curr
;
3578 raw_spin_lock(&rq
->lock
);
3579 update_rq_clock(rq
);
3580 update_cpu_load_active(rq
);
3581 curr
->sched_class
->task_tick(rq
, curr
, 0);
3582 raw_spin_unlock(&rq
->lock
);
3584 perf_event_task_tick(curr
);
3587 rq
->idle_at_tick
= idle_cpu(cpu
);
3588 trigger_load_balance(rq
, cpu
);
3592 notrace
unsigned long get_parent_ip(unsigned long addr
)
3594 if (in_lock_functions(addr
)) {
3595 addr
= CALLER_ADDR2
;
3596 if (in_lock_functions(addr
))
3597 addr
= CALLER_ADDR3
;
3602 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3603 defined(CONFIG_PREEMPT_TRACER))
3605 void __kprobes
add_preempt_count(int val
)
3607 #ifdef CONFIG_DEBUG_PREEMPT
3611 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3614 preempt_count() += val
;
3615 #ifdef CONFIG_DEBUG_PREEMPT
3617 * Spinlock count overflowing soon?
3619 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3622 if (preempt_count() == val
)
3623 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3625 EXPORT_SYMBOL(add_preempt_count
);
3627 void __kprobes
sub_preempt_count(int val
)
3629 #ifdef CONFIG_DEBUG_PREEMPT
3633 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3636 * Is the spinlock portion underflowing?
3638 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3639 !(preempt_count() & PREEMPT_MASK
)))
3643 if (preempt_count() == val
)
3644 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3645 preempt_count() -= val
;
3647 EXPORT_SYMBOL(sub_preempt_count
);
3652 * Print scheduling while atomic bug:
3654 static noinline
void __schedule_bug(struct task_struct
*prev
)
3656 struct pt_regs
*regs
= get_irq_regs();
3658 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3659 prev
->comm
, prev
->pid
, preempt_count());
3661 debug_show_held_locks(prev
);
3663 if (irqs_disabled())
3664 print_irqtrace_events(prev
);
3673 * Various schedule()-time debugging checks and statistics:
3675 static inline void schedule_debug(struct task_struct
*prev
)
3678 * Test if we are atomic. Since do_exit() needs to call into
3679 * schedule() atomically, we ignore that path for now.
3680 * Otherwise, whine if we are scheduling when we should not be.
3682 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3683 __schedule_bug(prev
);
3685 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3687 schedstat_inc(this_rq(), sched_count
);
3688 #ifdef CONFIG_SCHEDSTATS
3689 if (unlikely(prev
->lock_depth
>= 0)) {
3690 schedstat_inc(this_rq(), bkl_count
);
3691 schedstat_inc(prev
, sched_info
.bkl_count
);
3696 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3699 update_rq_clock(rq
);
3700 rq
->skip_clock_update
= 0;
3701 prev
->sched_class
->put_prev_task(rq
, prev
);
3705 * Pick up the highest-prio task:
3707 static inline struct task_struct
*
3708 pick_next_task(struct rq
*rq
)
3710 const struct sched_class
*class;
3711 struct task_struct
*p
;
3714 * Optimization: we know that if all tasks are in
3715 * the fair class we can call that function directly:
3717 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3718 p
= fair_sched_class
.pick_next_task(rq
);
3723 class = sched_class_highest
;
3725 p
= class->pick_next_task(rq
);
3729 * Will never be NULL as the idle class always
3730 * returns a non-NULL p:
3732 class = class->next
;
3737 * schedule() is the main scheduler function.
3739 asmlinkage
void __sched
schedule(void)
3741 struct task_struct
*prev
, *next
;
3742 unsigned long *switch_count
;
3748 cpu
= smp_processor_id();
3750 rcu_note_context_switch(cpu
);
3753 release_kernel_lock(prev
);
3754 need_resched_nonpreemptible
:
3756 schedule_debug(prev
);
3758 if (sched_feat(HRTICK
))
3761 raw_spin_lock_irq(&rq
->lock
);
3762 clear_tsk_need_resched(prev
);
3764 switch_count
= &prev
->nivcsw
;
3765 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3766 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3767 prev
->state
= TASK_RUNNING
;
3770 * If a worker is going to sleep, notify and
3771 * ask workqueue whether it wants to wake up a
3772 * task to maintain concurrency. If so, wake
3775 if (prev
->flags
& PF_WQ_WORKER
) {
3776 struct task_struct
*to_wakeup
;
3778 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3780 try_to_wake_up_local(to_wakeup
);
3782 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3784 switch_count
= &prev
->nvcsw
;
3787 pre_schedule(rq
, prev
);
3789 if (unlikely(!rq
->nr_running
))
3790 idle_balance(cpu
, rq
);
3792 put_prev_task(rq
, prev
);
3793 next
= pick_next_task(rq
);
3795 if (likely(prev
!= next
)) {
3796 sched_info_switch(prev
, next
);
3797 perf_event_task_sched_out(prev
, next
);
3803 context_switch(rq
, prev
, next
); /* unlocks the rq */
3805 * The context switch have flipped the stack from under us
3806 * and restored the local variables which were saved when
3807 * this task called schedule() in the past. prev == current
3808 * is still correct, but it can be moved to another cpu/rq.
3810 cpu
= smp_processor_id();
3813 raw_spin_unlock_irq(&rq
->lock
);
3817 if (unlikely(reacquire_kernel_lock(prev
)))
3818 goto need_resched_nonpreemptible
;
3820 preempt_enable_no_resched();
3824 EXPORT_SYMBOL(schedule
);
3826 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3828 * Look out! "owner" is an entirely speculative pointer
3829 * access and not reliable.
3831 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3836 if (!sched_feat(OWNER_SPIN
))
3839 #ifdef CONFIG_DEBUG_PAGEALLOC
3841 * Need to access the cpu field knowing that
3842 * DEBUG_PAGEALLOC could have unmapped it if
3843 * the mutex owner just released it and exited.
3845 if (probe_kernel_address(&owner
->cpu
, cpu
))
3852 * Even if the access succeeded (likely case),
3853 * the cpu field may no longer be valid.
3855 if (cpu
>= nr_cpumask_bits
)
3859 * We need to validate that we can do a
3860 * get_cpu() and that we have the percpu area.
3862 if (!cpu_online(cpu
))
3869 * Owner changed, break to re-assess state.
3871 if (lock
->owner
!= owner
) {
3873 * If the lock has switched to a different owner,
3874 * we likely have heavy contention. Return 0 to quit
3875 * optimistic spinning and not contend further:
3883 * Is that owner really running on that cpu?
3885 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3895 #ifdef CONFIG_PREEMPT
3897 * this is the entry point to schedule() from in-kernel preemption
3898 * off of preempt_enable. Kernel preemptions off return from interrupt
3899 * occur there and call schedule directly.
3901 asmlinkage
void __sched notrace
preempt_schedule(void)
3903 struct thread_info
*ti
= current_thread_info();
3906 * If there is a non-zero preempt_count or interrupts are disabled,
3907 * we do not want to preempt the current task. Just return..
3909 if (likely(ti
->preempt_count
|| irqs_disabled()))
3913 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3915 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3918 * Check again in case we missed a preemption opportunity
3919 * between schedule and now.
3922 } while (need_resched());
3924 EXPORT_SYMBOL(preempt_schedule
);
3927 * this is the entry point to schedule() from kernel preemption
3928 * off of irq context.
3929 * Note, that this is called and return with irqs disabled. This will
3930 * protect us against recursive calling from irq.
3932 asmlinkage
void __sched
preempt_schedule_irq(void)
3934 struct thread_info
*ti
= current_thread_info();
3936 /* Catch callers which need to be fixed */
3937 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3940 add_preempt_count(PREEMPT_ACTIVE
);
3943 local_irq_disable();
3944 sub_preempt_count(PREEMPT_ACTIVE
);
3947 * Check again in case we missed a preemption opportunity
3948 * between schedule and now.
3951 } while (need_resched());
3954 #endif /* CONFIG_PREEMPT */
3956 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3959 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3961 EXPORT_SYMBOL(default_wake_function
);
3964 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3965 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3966 * number) then we wake all the non-exclusive tasks and one exclusive task.
3968 * There are circumstances in which we can try to wake a task which has already
3969 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3970 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3972 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3973 int nr_exclusive
, int wake_flags
, void *key
)
3975 wait_queue_t
*curr
, *next
;
3977 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3978 unsigned flags
= curr
->flags
;
3980 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3981 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3987 * __wake_up - wake up threads blocked on a waitqueue.
3989 * @mode: which threads
3990 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3991 * @key: is directly passed to the wakeup function
3993 * It may be assumed that this function implies a write memory barrier before
3994 * changing the task state if and only if any tasks are woken up.
3996 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3997 int nr_exclusive
, void *key
)
3999 unsigned long flags
;
4001 spin_lock_irqsave(&q
->lock
, flags
);
4002 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4003 spin_unlock_irqrestore(&q
->lock
, flags
);
4005 EXPORT_SYMBOL(__wake_up
);
4008 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4010 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4012 __wake_up_common(q
, mode
, 1, 0, NULL
);
4014 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4016 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4018 __wake_up_common(q
, mode
, 1, 0, key
);
4022 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4024 * @mode: which threads
4025 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4026 * @key: opaque value to be passed to wakeup targets
4028 * The sync wakeup differs that the waker knows that it will schedule
4029 * away soon, so while the target thread will be woken up, it will not
4030 * be migrated to another CPU - ie. the two threads are 'synchronized'
4031 * with each other. This can prevent needless bouncing between CPUs.
4033 * On UP it can prevent extra preemption.
4035 * It may be assumed that this function implies a write memory barrier before
4036 * changing the task state if and only if any tasks are woken up.
4038 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4039 int nr_exclusive
, void *key
)
4041 unsigned long flags
;
4042 int wake_flags
= WF_SYNC
;
4047 if (unlikely(!nr_exclusive
))
4050 spin_lock_irqsave(&q
->lock
, flags
);
4051 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4052 spin_unlock_irqrestore(&q
->lock
, flags
);
4054 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4057 * __wake_up_sync - see __wake_up_sync_key()
4059 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4061 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4063 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4066 * complete: - signals a single thread waiting on this completion
4067 * @x: holds the state of this particular completion
4069 * This will wake up a single thread waiting on this completion. Threads will be
4070 * awakened in the same order in which they were queued.
4072 * See also complete_all(), wait_for_completion() and related routines.
4074 * It may be assumed that this function implies a write memory barrier before
4075 * changing the task state if and only if any tasks are woken up.
4077 void complete(struct completion
*x
)
4079 unsigned long flags
;
4081 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4083 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4084 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4086 EXPORT_SYMBOL(complete
);
4089 * complete_all: - signals all threads waiting on this completion
4090 * @x: holds the state of this particular completion
4092 * This will wake up all threads waiting on this particular completion event.
4094 * It may be assumed that this function implies a write memory barrier before
4095 * changing the task state if and only if any tasks are woken up.
4097 void complete_all(struct completion
*x
)
4099 unsigned long flags
;
4101 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4102 x
->done
+= UINT_MAX
/2;
4103 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4104 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4106 EXPORT_SYMBOL(complete_all
);
4108 static inline long __sched
4109 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4112 DECLARE_WAITQUEUE(wait
, current
);
4114 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4116 if (signal_pending_state(state
, current
)) {
4117 timeout
= -ERESTARTSYS
;
4120 __set_current_state(state
);
4121 spin_unlock_irq(&x
->wait
.lock
);
4122 timeout
= schedule_timeout(timeout
);
4123 spin_lock_irq(&x
->wait
.lock
);
4124 } while (!x
->done
&& timeout
);
4125 __remove_wait_queue(&x
->wait
, &wait
);
4130 return timeout
?: 1;
4134 wait_for_common(struct completion
*x
, long timeout
, int state
)
4138 spin_lock_irq(&x
->wait
.lock
);
4139 timeout
= do_wait_for_common(x
, timeout
, state
);
4140 spin_unlock_irq(&x
->wait
.lock
);
4145 * wait_for_completion: - waits for completion of a task
4146 * @x: holds the state of this particular completion
4148 * This waits to be signaled for completion of a specific task. It is NOT
4149 * interruptible and there is no timeout.
4151 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4152 * and interrupt capability. Also see complete().
4154 void __sched
wait_for_completion(struct completion
*x
)
4156 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4158 EXPORT_SYMBOL(wait_for_completion
);
4161 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4162 * @x: holds the state of this particular completion
4163 * @timeout: timeout value in jiffies
4165 * This waits for either a completion of a specific task to be signaled or for a
4166 * specified timeout to expire. The timeout is in jiffies. It is not
4169 unsigned long __sched
4170 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4172 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4174 EXPORT_SYMBOL(wait_for_completion_timeout
);
4177 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4178 * @x: holds the state of this particular completion
4180 * This waits for completion of a specific task to be signaled. It is
4183 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4185 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4186 if (t
== -ERESTARTSYS
)
4190 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4193 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4194 * @x: holds the state of this particular completion
4195 * @timeout: timeout value in jiffies
4197 * This waits for either a completion of a specific task to be signaled or for a
4198 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4200 unsigned long __sched
4201 wait_for_completion_interruptible_timeout(struct completion
*x
,
4202 unsigned long timeout
)
4204 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4206 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4209 * wait_for_completion_killable: - waits for completion of a task (killable)
4210 * @x: holds the state of this particular completion
4212 * This waits to be signaled for completion of a specific task. It can be
4213 * interrupted by a kill signal.
4215 int __sched
wait_for_completion_killable(struct completion
*x
)
4217 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4218 if (t
== -ERESTARTSYS
)
4222 EXPORT_SYMBOL(wait_for_completion_killable
);
4225 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4226 * @x: holds the state of this particular completion
4227 * @timeout: timeout value in jiffies
4229 * This waits for either a completion of a specific task to be
4230 * signaled or for a specified timeout to expire. It can be
4231 * interrupted by a kill signal. The timeout is in jiffies.
4233 unsigned long __sched
4234 wait_for_completion_killable_timeout(struct completion
*x
,
4235 unsigned long timeout
)
4237 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4239 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4242 * try_wait_for_completion - try to decrement a completion without blocking
4243 * @x: completion structure
4245 * Returns: 0 if a decrement cannot be done without blocking
4246 * 1 if a decrement succeeded.
4248 * If a completion is being used as a counting completion,
4249 * attempt to decrement the counter without blocking. This
4250 * enables us to avoid waiting if the resource the completion
4251 * is protecting is not available.
4253 bool try_wait_for_completion(struct completion
*x
)
4255 unsigned long flags
;
4258 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4263 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4266 EXPORT_SYMBOL(try_wait_for_completion
);
4269 * completion_done - Test to see if a completion has any waiters
4270 * @x: completion structure
4272 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4273 * 1 if there are no waiters.
4276 bool completion_done(struct completion
*x
)
4278 unsigned long flags
;
4281 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4284 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4287 EXPORT_SYMBOL(completion_done
);
4290 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4292 unsigned long flags
;
4295 init_waitqueue_entry(&wait
, current
);
4297 __set_current_state(state
);
4299 spin_lock_irqsave(&q
->lock
, flags
);
4300 __add_wait_queue(q
, &wait
);
4301 spin_unlock(&q
->lock
);
4302 timeout
= schedule_timeout(timeout
);
4303 spin_lock_irq(&q
->lock
);
4304 __remove_wait_queue(q
, &wait
);
4305 spin_unlock_irqrestore(&q
->lock
, flags
);
4310 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4312 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4314 EXPORT_SYMBOL(interruptible_sleep_on
);
4317 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4319 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4321 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4323 void __sched
sleep_on(wait_queue_head_t
*q
)
4325 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4327 EXPORT_SYMBOL(sleep_on
);
4329 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4331 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4333 EXPORT_SYMBOL(sleep_on_timeout
);
4335 #ifdef CONFIG_RT_MUTEXES
4338 * rt_mutex_setprio - set the current priority of a task
4340 * @prio: prio value (kernel-internal form)
4342 * This function changes the 'effective' priority of a task. It does
4343 * not touch ->normal_prio like __setscheduler().
4345 * Used by the rt_mutex code to implement priority inheritance logic.
4347 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4349 unsigned long flags
;
4350 int oldprio
, on_rq
, running
;
4352 const struct sched_class
*prev_class
;
4354 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4356 rq
= task_rq_lock(p
, &flags
);
4359 prev_class
= p
->sched_class
;
4360 on_rq
= p
->se
.on_rq
;
4361 running
= task_current(rq
, p
);
4363 dequeue_task(rq
, p
, 0);
4365 p
->sched_class
->put_prev_task(rq
, p
);
4368 p
->sched_class
= &rt_sched_class
;
4370 p
->sched_class
= &fair_sched_class
;
4375 p
->sched_class
->set_curr_task(rq
);
4377 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4379 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4381 task_rq_unlock(rq
, &flags
);
4386 void set_user_nice(struct task_struct
*p
, long nice
)
4388 int old_prio
, delta
, on_rq
;
4389 unsigned long flags
;
4392 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4395 * We have to be careful, if called from sys_setpriority(),
4396 * the task might be in the middle of scheduling on another CPU.
4398 rq
= task_rq_lock(p
, &flags
);
4400 * The RT priorities are set via sched_setscheduler(), but we still
4401 * allow the 'normal' nice value to be set - but as expected
4402 * it wont have any effect on scheduling until the task is
4403 * SCHED_FIFO/SCHED_RR:
4405 if (task_has_rt_policy(p
)) {
4406 p
->static_prio
= NICE_TO_PRIO(nice
);
4409 on_rq
= p
->se
.on_rq
;
4411 dequeue_task(rq
, p
, 0);
4413 p
->static_prio
= NICE_TO_PRIO(nice
);
4416 p
->prio
= effective_prio(p
);
4417 delta
= p
->prio
- old_prio
;
4420 enqueue_task(rq
, p
, 0);
4422 * If the task increased its priority or is running and
4423 * lowered its priority, then reschedule its CPU:
4425 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4426 resched_task(rq
->curr
);
4429 task_rq_unlock(rq
, &flags
);
4431 EXPORT_SYMBOL(set_user_nice
);
4434 * can_nice - check if a task can reduce its nice value
4438 int can_nice(const struct task_struct
*p
, const int nice
)
4440 /* convert nice value [19,-20] to rlimit style value [1,40] */
4441 int nice_rlim
= 20 - nice
;
4443 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4444 capable(CAP_SYS_NICE
));
4447 #ifdef __ARCH_WANT_SYS_NICE
4450 * sys_nice - change the priority of the current process.
4451 * @increment: priority increment
4453 * sys_setpriority is a more generic, but much slower function that
4454 * does similar things.
4456 SYSCALL_DEFINE1(nice
, int, increment
)
4461 * Setpriority might change our priority at the same moment.
4462 * We don't have to worry. Conceptually one call occurs first
4463 * and we have a single winner.
4465 if (increment
< -40)
4470 nice
= TASK_NICE(current
) + increment
;
4476 if (increment
< 0 && !can_nice(current
, nice
))
4479 retval
= security_task_setnice(current
, nice
);
4483 set_user_nice(current
, nice
);
4490 * task_prio - return the priority value of a given task.
4491 * @p: the task in question.
4493 * This is the priority value as seen by users in /proc.
4494 * RT tasks are offset by -200. Normal tasks are centered
4495 * around 0, value goes from -16 to +15.
4497 int task_prio(const struct task_struct
*p
)
4499 return p
->prio
- MAX_RT_PRIO
;
4503 * task_nice - return the nice value of a given task.
4504 * @p: the task in question.
4506 int task_nice(const struct task_struct
*p
)
4508 return TASK_NICE(p
);
4510 EXPORT_SYMBOL(task_nice
);
4513 * idle_cpu - is a given cpu idle currently?
4514 * @cpu: the processor in question.
4516 int idle_cpu(int cpu
)
4518 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4522 * idle_task - return the idle task for a given cpu.
4523 * @cpu: the processor in question.
4525 struct task_struct
*idle_task(int cpu
)
4527 return cpu_rq(cpu
)->idle
;
4531 * find_process_by_pid - find a process with a matching PID value.
4532 * @pid: the pid in question.
4534 static struct task_struct
*find_process_by_pid(pid_t pid
)
4536 return pid
? find_task_by_vpid(pid
) : current
;
4539 /* Actually do priority change: must hold rq lock. */
4541 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4543 BUG_ON(p
->se
.on_rq
);
4546 p
->rt_priority
= prio
;
4547 p
->normal_prio
= normal_prio(p
);
4548 /* we are holding p->pi_lock already */
4549 p
->prio
= rt_mutex_getprio(p
);
4550 if (rt_prio(p
->prio
))
4551 p
->sched_class
= &rt_sched_class
;
4553 p
->sched_class
= &fair_sched_class
;
4558 * check the target process has a UID that matches the current process's
4560 static bool check_same_owner(struct task_struct
*p
)
4562 const struct cred
*cred
= current_cred(), *pcred
;
4566 pcred
= __task_cred(p
);
4567 match
= (cred
->euid
== pcred
->euid
||
4568 cred
->euid
== pcred
->uid
);
4573 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4574 struct sched_param
*param
, bool user
)
4576 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4577 unsigned long flags
;
4578 const struct sched_class
*prev_class
;
4582 /* may grab non-irq protected spin_locks */
4583 BUG_ON(in_interrupt());
4585 /* double check policy once rq lock held */
4587 reset_on_fork
= p
->sched_reset_on_fork
;
4588 policy
= oldpolicy
= p
->policy
;
4590 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4591 policy
&= ~SCHED_RESET_ON_FORK
;
4593 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4594 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4595 policy
!= SCHED_IDLE
)
4600 * Valid priorities for SCHED_FIFO and SCHED_RR are
4601 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4602 * SCHED_BATCH and SCHED_IDLE is 0.
4604 if (param
->sched_priority
< 0 ||
4605 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4606 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4608 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4612 * Allow unprivileged RT tasks to decrease priority:
4614 if (user
&& !capable(CAP_SYS_NICE
)) {
4615 if (rt_policy(policy
)) {
4616 unsigned long rlim_rtprio
=
4617 task_rlimit(p
, RLIMIT_RTPRIO
);
4619 /* can't set/change the rt policy */
4620 if (policy
!= p
->policy
&& !rlim_rtprio
)
4623 /* can't increase priority */
4624 if (param
->sched_priority
> p
->rt_priority
&&
4625 param
->sched_priority
> rlim_rtprio
)
4629 * Like positive nice levels, dont allow tasks to
4630 * move out of SCHED_IDLE either:
4632 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4635 /* can't change other user's priorities */
4636 if (!check_same_owner(p
))
4639 /* Normal users shall not reset the sched_reset_on_fork flag */
4640 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4645 retval
= security_task_setscheduler(p
, policy
, param
);
4651 * make sure no PI-waiters arrive (or leave) while we are
4652 * changing the priority of the task:
4654 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4656 * To be able to change p->policy safely, the apropriate
4657 * runqueue lock must be held.
4659 rq
= __task_rq_lock(p
);
4661 #ifdef CONFIG_RT_GROUP_SCHED
4664 * Do not allow realtime tasks into groups that have no runtime
4667 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4668 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4669 __task_rq_unlock(rq
);
4670 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4676 /* recheck policy now with rq lock held */
4677 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4678 policy
= oldpolicy
= -1;
4679 __task_rq_unlock(rq
);
4680 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4683 on_rq
= p
->se
.on_rq
;
4684 running
= task_current(rq
, p
);
4686 deactivate_task(rq
, p
, 0);
4688 p
->sched_class
->put_prev_task(rq
, p
);
4690 p
->sched_reset_on_fork
= reset_on_fork
;
4693 prev_class
= p
->sched_class
;
4694 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4697 p
->sched_class
->set_curr_task(rq
);
4699 activate_task(rq
, p
, 0);
4701 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4703 __task_rq_unlock(rq
);
4704 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4706 rt_mutex_adjust_pi(p
);
4712 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4713 * @p: the task in question.
4714 * @policy: new policy.
4715 * @param: structure containing the new RT priority.
4717 * NOTE that the task may be already dead.
4719 int sched_setscheduler(struct task_struct
*p
, int policy
,
4720 struct sched_param
*param
)
4722 return __sched_setscheduler(p
, policy
, param
, true);
4724 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4727 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4728 * @p: the task in question.
4729 * @policy: new policy.
4730 * @param: structure containing the new RT priority.
4732 * Just like sched_setscheduler, only don't bother checking if the
4733 * current context has permission. For example, this is needed in
4734 * stop_machine(): we create temporary high priority worker threads,
4735 * but our caller might not have that capability.
4737 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4738 struct sched_param
*param
)
4740 return __sched_setscheduler(p
, policy
, param
, false);
4744 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4746 struct sched_param lparam
;
4747 struct task_struct
*p
;
4750 if (!param
|| pid
< 0)
4752 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4757 p
= find_process_by_pid(pid
);
4759 retval
= sched_setscheduler(p
, policy
, &lparam
);
4766 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4767 * @pid: the pid in question.
4768 * @policy: new policy.
4769 * @param: structure containing the new RT priority.
4771 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4772 struct sched_param __user
*, param
)
4774 /* negative values for policy are not valid */
4778 return do_sched_setscheduler(pid
, policy
, param
);
4782 * sys_sched_setparam - set/change the RT priority of a thread
4783 * @pid: the pid in question.
4784 * @param: structure containing the new RT priority.
4786 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4788 return do_sched_setscheduler(pid
, -1, param
);
4792 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4793 * @pid: the pid in question.
4795 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4797 struct task_struct
*p
;
4805 p
= find_process_by_pid(pid
);
4807 retval
= security_task_getscheduler(p
);
4810 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4817 * sys_sched_getparam - get the RT priority of a thread
4818 * @pid: the pid in question.
4819 * @param: structure containing the RT priority.
4821 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4823 struct sched_param lp
;
4824 struct task_struct
*p
;
4827 if (!param
|| pid
< 0)
4831 p
= find_process_by_pid(pid
);
4836 retval
= security_task_getscheduler(p
);
4840 lp
.sched_priority
= p
->rt_priority
;
4844 * This one might sleep, we cannot do it with a spinlock held ...
4846 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4855 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4857 cpumask_var_t cpus_allowed
, new_mask
;
4858 struct task_struct
*p
;
4864 p
= find_process_by_pid(pid
);
4871 /* Prevent p going away */
4875 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4879 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4881 goto out_free_cpus_allowed
;
4884 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4887 retval
= security_task_setscheduler(p
, 0, NULL
);
4891 cpuset_cpus_allowed(p
, cpus_allowed
);
4892 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4894 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4897 cpuset_cpus_allowed(p
, cpus_allowed
);
4898 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4900 * We must have raced with a concurrent cpuset
4901 * update. Just reset the cpus_allowed to the
4902 * cpuset's cpus_allowed
4904 cpumask_copy(new_mask
, cpus_allowed
);
4909 free_cpumask_var(new_mask
);
4910 out_free_cpus_allowed
:
4911 free_cpumask_var(cpus_allowed
);
4918 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4919 struct cpumask
*new_mask
)
4921 if (len
< cpumask_size())
4922 cpumask_clear(new_mask
);
4923 else if (len
> cpumask_size())
4924 len
= cpumask_size();
4926 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4930 * sys_sched_setaffinity - set the cpu affinity of a process
4931 * @pid: pid of the process
4932 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4933 * @user_mask_ptr: user-space pointer to the new cpu mask
4935 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4936 unsigned long __user
*, user_mask_ptr
)
4938 cpumask_var_t new_mask
;
4941 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4944 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4946 retval
= sched_setaffinity(pid
, new_mask
);
4947 free_cpumask_var(new_mask
);
4951 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4953 struct task_struct
*p
;
4954 unsigned long flags
;
4962 p
= find_process_by_pid(pid
);
4966 retval
= security_task_getscheduler(p
);
4970 rq
= task_rq_lock(p
, &flags
);
4971 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4972 task_rq_unlock(rq
, &flags
);
4982 * sys_sched_getaffinity - get the cpu affinity of a process
4983 * @pid: pid of the process
4984 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4985 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4987 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4988 unsigned long __user
*, user_mask_ptr
)
4993 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4995 if (len
& (sizeof(unsigned long)-1))
4998 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5001 ret
= sched_getaffinity(pid
, mask
);
5003 size_t retlen
= min_t(size_t, len
, cpumask_size());
5005 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5010 free_cpumask_var(mask
);
5016 * sys_sched_yield - yield the current processor to other threads.
5018 * This function yields the current CPU to other tasks. If there are no
5019 * other threads running on this CPU then this function will return.
5021 SYSCALL_DEFINE0(sched_yield
)
5023 struct rq
*rq
= this_rq_lock();
5025 schedstat_inc(rq
, yld_count
);
5026 current
->sched_class
->yield_task(rq
);
5029 * Since we are going to call schedule() anyway, there's
5030 * no need to preempt or enable interrupts:
5032 __release(rq
->lock
);
5033 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5034 do_raw_spin_unlock(&rq
->lock
);
5035 preempt_enable_no_resched();
5042 static inline int should_resched(void)
5044 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5047 static void __cond_resched(void)
5049 add_preempt_count(PREEMPT_ACTIVE
);
5051 sub_preempt_count(PREEMPT_ACTIVE
);
5054 int __sched
_cond_resched(void)
5056 if (should_resched()) {
5062 EXPORT_SYMBOL(_cond_resched
);
5065 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5066 * call schedule, and on return reacquire the lock.
5068 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5069 * operations here to prevent schedule() from being called twice (once via
5070 * spin_unlock(), once by hand).
5072 int __cond_resched_lock(spinlock_t
*lock
)
5074 int resched
= should_resched();
5077 lockdep_assert_held(lock
);
5079 if (spin_needbreak(lock
) || resched
) {
5090 EXPORT_SYMBOL(__cond_resched_lock
);
5092 int __sched
__cond_resched_softirq(void)
5094 BUG_ON(!in_softirq());
5096 if (should_resched()) {
5104 EXPORT_SYMBOL(__cond_resched_softirq
);
5107 * yield - yield the current processor to other threads.
5109 * This is a shortcut for kernel-space yielding - it marks the
5110 * thread runnable and calls sys_sched_yield().
5112 void __sched
yield(void)
5114 set_current_state(TASK_RUNNING
);
5117 EXPORT_SYMBOL(yield
);
5120 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5121 * that process accounting knows that this is a task in IO wait state.
5123 void __sched
io_schedule(void)
5125 struct rq
*rq
= raw_rq();
5127 delayacct_blkio_start();
5128 atomic_inc(&rq
->nr_iowait
);
5129 current
->in_iowait
= 1;
5131 current
->in_iowait
= 0;
5132 atomic_dec(&rq
->nr_iowait
);
5133 delayacct_blkio_end();
5135 EXPORT_SYMBOL(io_schedule
);
5137 long __sched
io_schedule_timeout(long timeout
)
5139 struct rq
*rq
= raw_rq();
5142 delayacct_blkio_start();
5143 atomic_inc(&rq
->nr_iowait
);
5144 current
->in_iowait
= 1;
5145 ret
= schedule_timeout(timeout
);
5146 current
->in_iowait
= 0;
5147 atomic_dec(&rq
->nr_iowait
);
5148 delayacct_blkio_end();
5153 * sys_sched_get_priority_max - return maximum RT priority.
5154 * @policy: scheduling class.
5156 * this syscall returns the maximum rt_priority that can be used
5157 * by a given scheduling class.
5159 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5166 ret
= MAX_USER_RT_PRIO
-1;
5178 * sys_sched_get_priority_min - return minimum RT priority.
5179 * @policy: scheduling class.
5181 * this syscall returns the minimum rt_priority that can be used
5182 * by a given scheduling class.
5184 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5202 * sys_sched_rr_get_interval - return the default timeslice of a process.
5203 * @pid: pid of the process.
5204 * @interval: userspace pointer to the timeslice value.
5206 * this syscall writes the default timeslice value of a given process
5207 * into the user-space timespec buffer. A value of '0' means infinity.
5209 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5210 struct timespec __user
*, interval
)
5212 struct task_struct
*p
;
5213 unsigned int time_slice
;
5214 unsigned long flags
;
5224 p
= find_process_by_pid(pid
);
5228 retval
= security_task_getscheduler(p
);
5232 rq
= task_rq_lock(p
, &flags
);
5233 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5234 task_rq_unlock(rq
, &flags
);
5237 jiffies_to_timespec(time_slice
, &t
);
5238 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5246 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5248 void sched_show_task(struct task_struct
*p
)
5250 unsigned long free
= 0;
5253 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5254 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5255 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5256 #if BITS_PER_LONG == 32
5257 if (state
== TASK_RUNNING
)
5258 printk(KERN_CONT
" running ");
5260 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5262 if (state
== TASK_RUNNING
)
5263 printk(KERN_CONT
" running task ");
5265 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5267 #ifdef CONFIG_DEBUG_STACK_USAGE
5268 free
= stack_not_used(p
);
5270 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5271 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5272 (unsigned long)task_thread_info(p
)->flags
);
5274 show_stack(p
, NULL
);
5277 void show_state_filter(unsigned long state_filter
)
5279 struct task_struct
*g
, *p
;
5281 #if BITS_PER_LONG == 32
5283 " task PC stack pid father\n");
5286 " task PC stack pid father\n");
5288 read_lock(&tasklist_lock
);
5289 do_each_thread(g
, p
) {
5291 * reset the NMI-timeout, listing all files on a slow
5292 * console might take alot of time:
5294 touch_nmi_watchdog();
5295 if (!state_filter
|| (p
->state
& state_filter
))
5297 } while_each_thread(g
, p
);
5299 touch_all_softlockup_watchdogs();
5301 #ifdef CONFIG_SCHED_DEBUG
5302 sysrq_sched_debug_show();
5304 read_unlock(&tasklist_lock
);
5306 * Only show locks if all tasks are dumped:
5309 debug_show_all_locks();
5312 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5314 idle
->sched_class
= &idle_sched_class
;
5318 * init_idle - set up an idle thread for a given CPU
5319 * @idle: task in question
5320 * @cpu: cpu the idle task belongs to
5322 * NOTE: this function does not set the idle thread's NEED_RESCHED
5323 * flag, to make booting more robust.
5325 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5327 struct rq
*rq
= cpu_rq(cpu
);
5328 unsigned long flags
;
5330 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5333 idle
->state
= TASK_RUNNING
;
5334 idle
->se
.exec_start
= sched_clock();
5336 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5337 __set_task_cpu(idle
, cpu
);
5339 rq
->curr
= rq
->idle
= idle
;
5340 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5343 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5345 /* Set the preempt count _outside_ the spinlocks! */
5346 #if defined(CONFIG_PREEMPT)
5347 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5349 task_thread_info(idle
)->preempt_count
= 0;
5352 * The idle tasks have their own, simple scheduling class:
5354 idle
->sched_class
= &idle_sched_class
;
5355 ftrace_graph_init_task(idle
);
5359 * In a system that switches off the HZ timer nohz_cpu_mask
5360 * indicates which cpus entered this state. This is used
5361 * in the rcu update to wait only for active cpus. For system
5362 * which do not switch off the HZ timer nohz_cpu_mask should
5363 * always be CPU_BITS_NONE.
5365 cpumask_var_t nohz_cpu_mask
;
5368 * Increase the granularity value when there are more CPUs,
5369 * because with more CPUs the 'effective latency' as visible
5370 * to users decreases. But the relationship is not linear,
5371 * so pick a second-best guess by going with the log2 of the
5374 * This idea comes from the SD scheduler of Con Kolivas:
5376 static int get_update_sysctl_factor(void)
5378 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5379 unsigned int factor
;
5381 switch (sysctl_sched_tunable_scaling
) {
5382 case SCHED_TUNABLESCALING_NONE
:
5385 case SCHED_TUNABLESCALING_LINEAR
:
5388 case SCHED_TUNABLESCALING_LOG
:
5390 factor
= 1 + ilog2(cpus
);
5397 static void update_sysctl(void)
5399 unsigned int factor
= get_update_sysctl_factor();
5401 #define SET_SYSCTL(name) \
5402 (sysctl_##name = (factor) * normalized_sysctl_##name)
5403 SET_SYSCTL(sched_min_granularity
);
5404 SET_SYSCTL(sched_latency
);
5405 SET_SYSCTL(sched_wakeup_granularity
);
5406 SET_SYSCTL(sched_shares_ratelimit
);
5410 static inline void sched_init_granularity(void)
5417 * This is how migration works:
5419 * 1) we invoke migration_cpu_stop() on the target CPU using
5421 * 2) stopper starts to run (implicitly forcing the migrated thread
5423 * 3) it checks whether the migrated task is still in the wrong runqueue.
5424 * 4) if it's in the wrong runqueue then the migration thread removes
5425 * it and puts it into the right queue.
5426 * 5) stopper completes and stop_one_cpu() returns and the migration
5431 * Change a given task's CPU affinity. Migrate the thread to a
5432 * proper CPU and schedule it away if the CPU it's executing on
5433 * is removed from the allowed bitmask.
5435 * NOTE: the caller must have a valid reference to the task, the
5436 * task must not exit() & deallocate itself prematurely. The
5437 * call is not atomic; no spinlocks may be held.
5439 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5441 unsigned long flags
;
5443 unsigned int dest_cpu
;
5447 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5448 * drop the rq->lock and still rely on ->cpus_allowed.
5451 while (task_is_waking(p
))
5453 rq
= task_rq_lock(p
, &flags
);
5454 if (task_is_waking(p
)) {
5455 task_rq_unlock(rq
, &flags
);
5459 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5464 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5465 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5470 if (p
->sched_class
->set_cpus_allowed
)
5471 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5473 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5474 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5477 /* Can the task run on the task's current CPU? If so, we're done */
5478 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5481 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5482 if (migrate_task(p
, dest_cpu
)) {
5483 struct migration_arg arg
= { p
, dest_cpu
};
5484 /* Need help from migration thread: drop lock and wait. */
5485 task_rq_unlock(rq
, &flags
);
5486 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5487 tlb_migrate_finish(p
->mm
);
5491 task_rq_unlock(rq
, &flags
);
5495 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5498 * Move (not current) task off this cpu, onto dest cpu. We're doing
5499 * this because either it can't run here any more (set_cpus_allowed()
5500 * away from this CPU, or CPU going down), or because we're
5501 * attempting to rebalance this task on exec (sched_exec).
5503 * So we race with normal scheduler movements, but that's OK, as long
5504 * as the task is no longer on this CPU.
5506 * Returns non-zero if task was successfully migrated.
5508 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5510 struct rq
*rq_dest
, *rq_src
;
5513 if (unlikely(!cpu_active(dest_cpu
)))
5516 rq_src
= cpu_rq(src_cpu
);
5517 rq_dest
= cpu_rq(dest_cpu
);
5519 double_rq_lock(rq_src
, rq_dest
);
5520 /* Already moved. */
5521 if (task_cpu(p
) != src_cpu
)
5523 /* Affinity changed (again). */
5524 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5528 * If we're not on a rq, the next wake-up will ensure we're
5532 deactivate_task(rq_src
, p
, 0);
5533 set_task_cpu(p
, dest_cpu
);
5534 activate_task(rq_dest
, p
, 0);
5535 check_preempt_curr(rq_dest
, p
, 0);
5540 double_rq_unlock(rq_src
, rq_dest
);
5545 * migration_cpu_stop - this will be executed by a highprio stopper thread
5546 * and performs thread migration by bumping thread off CPU then
5547 * 'pushing' onto another runqueue.
5549 static int migration_cpu_stop(void *data
)
5551 struct migration_arg
*arg
= data
;
5554 * The original target cpu might have gone down and we might
5555 * be on another cpu but it doesn't matter.
5557 local_irq_disable();
5558 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5563 #ifdef CONFIG_HOTPLUG_CPU
5565 * Figure out where task on dead CPU should go, use force if necessary.
5567 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5569 struct rq
*rq
= cpu_rq(dead_cpu
);
5570 int needs_cpu
, uninitialized_var(dest_cpu
);
5571 unsigned long flags
;
5573 local_irq_save(flags
);
5575 raw_spin_lock(&rq
->lock
);
5576 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5578 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5579 raw_spin_unlock(&rq
->lock
);
5581 * It can only fail if we race with set_cpus_allowed(),
5582 * in the racer should migrate the task anyway.
5585 __migrate_task(p
, dead_cpu
, dest_cpu
);
5586 local_irq_restore(flags
);
5590 * While a dead CPU has no uninterruptible tasks queued at this point,
5591 * it might still have a nonzero ->nr_uninterruptible counter, because
5592 * for performance reasons the counter is not stricly tracking tasks to
5593 * their home CPUs. So we just add the counter to another CPU's counter,
5594 * to keep the global sum constant after CPU-down:
5596 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5598 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5599 unsigned long flags
;
5601 local_irq_save(flags
);
5602 double_rq_lock(rq_src
, rq_dest
);
5603 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5604 rq_src
->nr_uninterruptible
= 0;
5605 double_rq_unlock(rq_src
, rq_dest
);
5606 local_irq_restore(flags
);
5609 /* Run through task list and migrate tasks from the dead cpu. */
5610 static void migrate_live_tasks(int src_cpu
)
5612 struct task_struct
*p
, *t
;
5614 read_lock(&tasklist_lock
);
5616 do_each_thread(t
, p
) {
5620 if (task_cpu(p
) == src_cpu
)
5621 move_task_off_dead_cpu(src_cpu
, p
);
5622 } while_each_thread(t
, p
);
5624 read_unlock(&tasklist_lock
);
5628 * Schedules idle task to be the next runnable task on current CPU.
5629 * It does so by boosting its priority to highest possible.
5630 * Used by CPU offline code.
5632 void sched_idle_next(void)
5634 int this_cpu
= smp_processor_id();
5635 struct rq
*rq
= cpu_rq(this_cpu
);
5636 struct task_struct
*p
= rq
->idle
;
5637 unsigned long flags
;
5639 /* cpu has to be offline */
5640 BUG_ON(cpu_online(this_cpu
));
5643 * Strictly not necessary since rest of the CPUs are stopped by now
5644 * and interrupts disabled on the current cpu.
5646 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5648 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5650 activate_task(rq
, p
, 0);
5652 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5656 * Ensures that the idle task is using init_mm right before its cpu goes
5659 void idle_task_exit(void)
5661 struct mm_struct
*mm
= current
->active_mm
;
5663 BUG_ON(cpu_online(smp_processor_id()));
5666 switch_mm(mm
, &init_mm
, current
);
5670 /* called under rq->lock with disabled interrupts */
5671 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5673 struct rq
*rq
= cpu_rq(dead_cpu
);
5675 /* Must be exiting, otherwise would be on tasklist. */
5676 BUG_ON(!p
->exit_state
);
5678 /* Cannot have done final schedule yet: would have vanished. */
5679 BUG_ON(p
->state
== TASK_DEAD
);
5684 * Drop lock around migration; if someone else moves it,
5685 * that's OK. No task can be added to this CPU, so iteration is
5688 raw_spin_unlock_irq(&rq
->lock
);
5689 move_task_off_dead_cpu(dead_cpu
, p
);
5690 raw_spin_lock_irq(&rq
->lock
);
5695 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5696 static void migrate_dead_tasks(unsigned int dead_cpu
)
5698 struct rq
*rq
= cpu_rq(dead_cpu
);
5699 struct task_struct
*next
;
5702 if (!rq
->nr_running
)
5704 next
= pick_next_task(rq
);
5707 next
->sched_class
->put_prev_task(rq
, next
);
5708 migrate_dead(dead_cpu
, next
);
5714 * remove the tasks which were accounted by rq from calc_load_tasks.
5716 static void calc_global_load_remove(struct rq
*rq
)
5718 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5719 rq
->calc_load_active
= 0;
5721 #endif /* CONFIG_HOTPLUG_CPU */
5723 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5725 static struct ctl_table sd_ctl_dir
[] = {
5727 .procname
= "sched_domain",
5733 static struct ctl_table sd_ctl_root
[] = {
5735 .procname
= "kernel",
5737 .child
= sd_ctl_dir
,
5742 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5744 struct ctl_table
*entry
=
5745 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5750 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5752 struct ctl_table
*entry
;
5755 * In the intermediate directories, both the child directory and
5756 * procname are dynamically allocated and could fail but the mode
5757 * will always be set. In the lowest directory the names are
5758 * static strings and all have proc handlers.
5760 for (entry
= *tablep
; entry
->mode
; entry
++) {
5762 sd_free_ctl_entry(&entry
->child
);
5763 if (entry
->proc_handler
== NULL
)
5764 kfree(entry
->procname
);
5772 set_table_entry(struct ctl_table
*entry
,
5773 const char *procname
, void *data
, int maxlen
,
5774 mode_t mode
, proc_handler
*proc_handler
)
5776 entry
->procname
= procname
;
5778 entry
->maxlen
= maxlen
;
5780 entry
->proc_handler
= proc_handler
;
5783 static struct ctl_table
*
5784 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5786 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5791 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5792 sizeof(long), 0644, proc_doulongvec_minmax
);
5793 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5794 sizeof(long), 0644, proc_doulongvec_minmax
);
5795 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5796 sizeof(int), 0644, proc_dointvec_minmax
);
5797 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5798 sizeof(int), 0644, proc_dointvec_minmax
);
5799 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5800 sizeof(int), 0644, proc_dointvec_minmax
);
5801 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5802 sizeof(int), 0644, proc_dointvec_minmax
);
5803 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5804 sizeof(int), 0644, proc_dointvec_minmax
);
5805 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5806 sizeof(int), 0644, proc_dointvec_minmax
);
5807 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5808 sizeof(int), 0644, proc_dointvec_minmax
);
5809 set_table_entry(&table
[9], "cache_nice_tries",
5810 &sd
->cache_nice_tries
,
5811 sizeof(int), 0644, proc_dointvec_minmax
);
5812 set_table_entry(&table
[10], "flags", &sd
->flags
,
5813 sizeof(int), 0644, proc_dointvec_minmax
);
5814 set_table_entry(&table
[11], "name", sd
->name
,
5815 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5816 /* &table[12] is terminator */
5821 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5823 struct ctl_table
*entry
, *table
;
5824 struct sched_domain
*sd
;
5825 int domain_num
= 0, i
;
5828 for_each_domain(cpu
, sd
)
5830 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5835 for_each_domain(cpu
, sd
) {
5836 snprintf(buf
, 32, "domain%d", i
);
5837 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5839 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5846 static struct ctl_table_header
*sd_sysctl_header
;
5847 static void register_sched_domain_sysctl(void)
5849 int i
, cpu_num
= num_possible_cpus();
5850 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5853 WARN_ON(sd_ctl_dir
[0].child
);
5854 sd_ctl_dir
[0].child
= entry
;
5859 for_each_possible_cpu(i
) {
5860 snprintf(buf
, 32, "cpu%d", i
);
5861 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5863 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5867 WARN_ON(sd_sysctl_header
);
5868 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5871 /* may be called multiple times per register */
5872 static void unregister_sched_domain_sysctl(void)
5874 if (sd_sysctl_header
)
5875 unregister_sysctl_table(sd_sysctl_header
);
5876 sd_sysctl_header
= NULL
;
5877 if (sd_ctl_dir
[0].child
)
5878 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5881 static void register_sched_domain_sysctl(void)
5884 static void unregister_sched_domain_sysctl(void)
5889 static void set_rq_online(struct rq
*rq
)
5892 const struct sched_class
*class;
5894 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5897 for_each_class(class) {
5898 if (class->rq_online
)
5899 class->rq_online(rq
);
5904 static void set_rq_offline(struct rq
*rq
)
5907 const struct sched_class
*class;
5909 for_each_class(class) {
5910 if (class->rq_offline
)
5911 class->rq_offline(rq
);
5914 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5920 * migration_call - callback that gets triggered when a CPU is added.
5921 * Here we can start up the necessary migration thread for the new CPU.
5923 static int __cpuinit
5924 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5926 int cpu
= (long)hcpu
;
5927 unsigned long flags
;
5928 struct rq
*rq
= cpu_rq(cpu
);
5932 case CPU_UP_PREPARE
:
5933 case CPU_UP_PREPARE_FROZEN
:
5934 rq
->calc_load_update
= calc_load_update
;
5938 case CPU_ONLINE_FROZEN
:
5939 /* Update our root-domain */
5940 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5942 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5946 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5949 #ifdef CONFIG_HOTPLUG_CPU
5951 case CPU_DEAD_FROZEN
:
5952 migrate_live_tasks(cpu
);
5953 /* Idle task back to normal (off runqueue, low prio) */
5954 raw_spin_lock_irq(&rq
->lock
);
5955 deactivate_task(rq
, rq
->idle
, 0);
5956 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5957 rq
->idle
->sched_class
= &idle_sched_class
;
5958 migrate_dead_tasks(cpu
);
5959 raw_spin_unlock_irq(&rq
->lock
);
5960 migrate_nr_uninterruptible(rq
);
5961 BUG_ON(rq
->nr_running
!= 0);
5962 calc_global_load_remove(rq
);
5966 case CPU_DYING_FROZEN
:
5967 /* Update our root-domain */
5968 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5970 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5973 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5981 * Register at high priority so that task migration (migrate_all_tasks)
5982 * happens before everything else. This has to be lower priority than
5983 * the notifier in the perf_event subsystem, though.
5985 static struct notifier_block __cpuinitdata migration_notifier
= {
5986 .notifier_call
= migration_call
,
5987 .priority
= CPU_PRI_MIGRATION
,
5990 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5991 unsigned long action
, void *hcpu
)
5993 switch (action
& ~CPU_TASKS_FROZEN
) {
5995 case CPU_DOWN_FAILED
:
5996 set_cpu_active((long)hcpu
, true);
6003 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6004 unsigned long action
, void *hcpu
)
6006 switch (action
& ~CPU_TASKS_FROZEN
) {
6007 case CPU_DOWN_PREPARE
:
6008 set_cpu_active((long)hcpu
, false);
6015 static int __init
migration_init(void)
6017 void *cpu
= (void *)(long)smp_processor_id();
6020 /* Initialize migration for the boot CPU */
6021 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6022 BUG_ON(err
== NOTIFY_BAD
);
6023 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6024 register_cpu_notifier(&migration_notifier
);
6026 /* Register cpu active notifiers */
6027 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6028 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6032 early_initcall(migration_init
);
6037 #ifdef CONFIG_SCHED_DEBUG
6039 static __read_mostly
int sched_domain_debug_enabled
;
6041 static int __init
sched_domain_debug_setup(char *str
)
6043 sched_domain_debug_enabled
= 1;
6047 early_param("sched_debug", sched_domain_debug_setup
);
6049 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6050 struct cpumask
*groupmask
)
6052 struct sched_group
*group
= sd
->groups
;
6055 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6056 cpumask_clear(groupmask
);
6058 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6060 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6061 printk("does not load-balance\n");
6063 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6068 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6070 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6071 printk(KERN_ERR
"ERROR: domain->span does not contain "
6074 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6075 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6079 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6083 printk(KERN_ERR
"ERROR: group is NULL\n");
6087 if (!group
->cpu_power
) {
6088 printk(KERN_CONT
"\n");
6089 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6094 if (!cpumask_weight(sched_group_cpus(group
))) {
6095 printk(KERN_CONT
"\n");
6096 printk(KERN_ERR
"ERROR: empty group\n");
6100 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6101 printk(KERN_CONT
"\n");
6102 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6106 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6108 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6110 printk(KERN_CONT
" %s", str
);
6111 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6112 printk(KERN_CONT
" (cpu_power = %d)",
6116 group
= group
->next
;
6117 } while (group
!= sd
->groups
);
6118 printk(KERN_CONT
"\n");
6120 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6121 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6124 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6125 printk(KERN_ERR
"ERROR: parent span is not a superset "
6126 "of domain->span\n");
6130 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6132 cpumask_var_t groupmask
;
6135 if (!sched_domain_debug_enabled
)
6139 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6143 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6145 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6146 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6151 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6158 free_cpumask_var(groupmask
);
6160 #else /* !CONFIG_SCHED_DEBUG */
6161 # define sched_domain_debug(sd, cpu) do { } while (0)
6162 #endif /* CONFIG_SCHED_DEBUG */
6164 static int sd_degenerate(struct sched_domain
*sd
)
6166 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6169 /* Following flags need at least 2 groups */
6170 if (sd
->flags
& (SD_LOAD_BALANCE
|
6171 SD_BALANCE_NEWIDLE
|
6175 SD_SHARE_PKG_RESOURCES
)) {
6176 if (sd
->groups
!= sd
->groups
->next
)
6180 /* Following flags don't use groups */
6181 if (sd
->flags
& (SD_WAKE_AFFINE
))
6188 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6190 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6192 if (sd_degenerate(parent
))
6195 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6198 /* Flags needing groups don't count if only 1 group in parent */
6199 if (parent
->groups
== parent
->groups
->next
) {
6200 pflags
&= ~(SD_LOAD_BALANCE
|
6201 SD_BALANCE_NEWIDLE
|
6205 SD_SHARE_PKG_RESOURCES
);
6206 if (nr_node_ids
== 1)
6207 pflags
&= ~SD_SERIALIZE
;
6209 if (~cflags
& pflags
)
6215 static void free_rootdomain(struct root_domain
*rd
)
6217 synchronize_sched();
6219 cpupri_cleanup(&rd
->cpupri
);
6221 free_cpumask_var(rd
->rto_mask
);
6222 free_cpumask_var(rd
->online
);
6223 free_cpumask_var(rd
->span
);
6227 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6229 struct root_domain
*old_rd
= NULL
;
6230 unsigned long flags
;
6232 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6237 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6240 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6243 * If we dont want to free the old_rt yet then
6244 * set old_rd to NULL to skip the freeing later
6247 if (!atomic_dec_and_test(&old_rd
->refcount
))
6251 atomic_inc(&rd
->refcount
);
6254 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6255 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6258 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6261 free_rootdomain(old_rd
);
6264 static int init_rootdomain(struct root_domain
*rd
)
6266 memset(rd
, 0, sizeof(*rd
));
6268 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6270 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6272 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6275 if (cpupri_init(&rd
->cpupri
) != 0)
6280 free_cpumask_var(rd
->rto_mask
);
6282 free_cpumask_var(rd
->online
);
6284 free_cpumask_var(rd
->span
);
6289 static void init_defrootdomain(void)
6291 init_rootdomain(&def_root_domain
);
6293 atomic_set(&def_root_domain
.refcount
, 1);
6296 static struct root_domain
*alloc_rootdomain(void)
6298 struct root_domain
*rd
;
6300 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6304 if (init_rootdomain(rd
) != 0) {
6313 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6314 * hold the hotplug lock.
6317 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6319 struct rq
*rq
= cpu_rq(cpu
);
6320 struct sched_domain
*tmp
;
6322 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6323 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6325 /* Remove the sched domains which do not contribute to scheduling. */
6326 for (tmp
= sd
; tmp
; ) {
6327 struct sched_domain
*parent
= tmp
->parent
;
6331 if (sd_parent_degenerate(tmp
, parent
)) {
6332 tmp
->parent
= parent
->parent
;
6334 parent
->parent
->child
= tmp
;
6339 if (sd
&& sd_degenerate(sd
)) {
6345 sched_domain_debug(sd
, cpu
);
6347 rq_attach_root(rq
, rd
);
6348 rcu_assign_pointer(rq
->sd
, sd
);
6351 /* cpus with isolated domains */
6352 static cpumask_var_t cpu_isolated_map
;
6354 /* Setup the mask of cpus configured for isolated domains */
6355 static int __init
isolated_cpu_setup(char *str
)
6357 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6358 cpulist_parse(str
, cpu_isolated_map
);
6362 __setup("isolcpus=", isolated_cpu_setup
);
6365 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6366 * to a function which identifies what group(along with sched group) a CPU
6367 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6368 * (due to the fact that we keep track of groups covered with a struct cpumask).
6370 * init_sched_build_groups will build a circular linked list of the groups
6371 * covered by the given span, and will set each group's ->cpumask correctly,
6372 * and ->cpu_power to 0.
6375 init_sched_build_groups(const struct cpumask
*span
,
6376 const struct cpumask
*cpu_map
,
6377 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6378 struct sched_group
**sg
,
6379 struct cpumask
*tmpmask
),
6380 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6382 struct sched_group
*first
= NULL
, *last
= NULL
;
6385 cpumask_clear(covered
);
6387 for_each_cpu(i
, span
) {
6388 struct sched_group
*sg
;
6389 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6392 if (cpumask_test_cpu(i
, covered
))
6395 cpumask_clear(sched_group_cpus(sg
));
6398 for_each_cpu(j
, span
) {
6399 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6402 cpumask_set_cpu(j
, covered
);
6403 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6414 #define SD_NODES_PER_DOMAIN 16
6419 * find_next_best_node - find the next node to include in a sched_domain
6420 * @node: node whose sched_domain we're building
6421 * @used_nodes: nodes already in the sched_domain
6423 * Find the next node to include in a given scheduling domain. Simply
6424 * finds the closest node not already in the @used_nodes map.
6426 * Should use nodemask_t.
6428 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6430 int i
, n
, val
, min_val
, best_node
= 0;
6434 for (i
= 0; i
< nr_node_ids
; i
++) {
6435 /* Start at @node */
6436 n
= (node
+ i
) % nr_node_ids
;
6438 if (!nr_cpus_node(n
))
6441 /* Skip already used nodes */
6442 if (node_isset(n
, *used_nodes
))
6445 /* Simple min distance search */
6446 val
= node_distance(node
, n
);
6448 if (val
< min_val
) {
6454 node_set(best_node
, *used_nodes
);
6459 * sched_domain_node_span - get a cpumask for a node's sched_domain
6460 * @node: node whose cpumask we're constructing
6461 * @span: resulting cpumask
6463 * Given a node, construct a good cpumask for its sched_domain to span. It
6464 * should be one that prevents unnecessary balancing, but also spreads tasks
6467 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6469 nodemask_t used_nodes
;
6472 cpumask_clear(span
);
6473 nodes_clear(used_nodes
);
6475 cpumask_or(span
, span
, cpumask_of_node(node
));
6476 node_set(node
, used_nodes
);
6478 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6479 int next_node
= find_next_best_node(node
, &used_nodes
);
6481 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6484 #endif /* CONFIG_NUMA */
6486 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6489 * The cpus mask in sched_group and sched_domain hangs off the end.
6491 * ( See the the comments in include/linux/sched.h:struct sched_group
6492 * and struct sched_domain. )
6494 struct static_sched_group
{
6495 struct sched_group sg
;
6496 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6499 struct static_sched_domain
{
6500 struct sched_domain sd
;
6501 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6507 cpumask_var_t domainspan
;
6508 cpumask_var_t covered
;
6509 cpumask_var_t notcovered
;
6511 cpumask_var_t nodemask
;
6512 cpumask_var_t this_sibling_map
;
6513 cpumask_var_t this_core_map
;
6514 cpumask_var_t this_book_map
;
6515 cpumask_var_t send_covered
;
6516 cpumask_var_t tmpmask
;
6517 struct sched_group
**sched_group_nodes
;
6518 struct root_domain
*rd
;
6522 sa_sched_groups
= 0,
6528 sa_this_sibling_map
,
6530 sa_sched_group_nodes
,
6540 * SMT sched-domains:
6542 #ifdef CONFIG_SCHED_SMT
6543 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6544 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6547 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6548 struct sched_group
**sg
, struct cpumask
*unused
)
6551 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6554 #endif /* CONFIG_SCHED_SMT */
6557 * multi-core sched-domains:
6559 #ifdef CONFIG_SCHED_MC
6560 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6561 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6564 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6565 struct sched_group
**sg
, struct cpumask
*mask
)
6568 #ifdef CONFIG_SCHED_SMT
6569 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6570 group
= cpumask_first(mask
);
6575 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6578 #endif /* CONFIG_SCHED_MC */
6581 * book sched-domains:
6583 #ifdef CONFIG_SCHED_BOOK
6584 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
6585 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
6588 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
6589 struct sched_group
**sg
, struct cpumask
*mask
)
6592 #ifdef CONFIG_SCHED_MC
6593 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6594 group
= cpumask_first(mask
);
6595 #elif defined(CONFIG_SCHED_SMT)
6596 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6597 group
= cpumask_first(mask
);
6600 *sg
= &per_cpu(sched_group_book
, group
).sg
;
6603 #endif /* CONFIG_SCHED_BOOK */
6605 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6606 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6609 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6610 struct sched_group
**sg
, struct cpumask
*mask
)
6613 #ifdef CONFIG_SCHED_BOOK
6614 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
6615 group
= cpumask_first(mask
);
6616 #elif defined(CONFIG_SCHED_MC)
6617 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6618 group
= cpumask_first(mask
);
6619 #elif defined(CONFIG_SCHED_SMT)
6620 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6621 group
= cpumask_first(mask
);
6626 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6632 * The init_sched_build_groups can't handle what we want to do with node
6633 * groups, so roll our own. Now each node has its own list of groups which
6634 * gets dynamically allocated.
6636 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6637 static struct sched_group
***sched_group_nodes_bycpu
;
6639 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6640 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6642 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6643 struct sched_group
**sg
,
6644 struct cpumask
*nodemask
)
6648 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6649 group
= cpumask_first(nodemask
);
6652 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6656 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6658 struct sched_group
*sg
= group_head
;
6664 for_each_cpu(j
, sched_group_cpus(sg
)) {
6665 struct sched_domain
*sd
;
6667 sd
= &per_cpu(phys_domains
, j
).sd
;
6668 if (j
!= group_first_cpu(sd
->groups
)) {
6670 * Only add "power" once for each
6676 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6679 } while (sg
!= group_head
);
6682 static int build_numa_sched_groups(struct s_data
*d
,
6683 const struct cpumask
*cpu_map
, int num
)
6685 struct sched_domain
*sd
;
6686 struct sched_group
*sg
, *prev
;
6689 cpumask_clear(d
->covered
);
6690 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6691 if (cpumask_empty(d
->nodemask
)) {
6692 d
->sched_group_nodes
[num
] = NULL
;
6696 sched_domain_node_span(num
, d
->domainspan
);
6697 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6699 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6702 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6706 d
->sched_group_nodes
[num
] = sg
;
6708 for_each_cpu(j
, d
->nodemask
) {
6709 sd
= &per_cpu(node_domains
, j
).sd
;
6714 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6716 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6719 for (j
= 0; j
< nr_node_ids
; j
++) {
6720 n
= (num
+ j
) % nr_node_ids
;
6721 cpumask_complement(d
->notcovered
, d
->covered
);
6722 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6723 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6724 if (cpumask_empty(d
->tmpmask
))
6726 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6727 if (cpumask_empty(d
->tmpmask
))
6729 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6733 "Can not alloc domain group for node %d\n", j
);
6737 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6738 sg
->next
= prev
->next
;
6739 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6746 #endif /* CONFIG_NUMA */
6749 /* Free memory allocated for various sched_group structures */
6750 static void free_sched_groups(const struct cpumask
*cpu_map
,
6751 struct cpumask
*nodemask
)
6755 for_each_cpu(cpu
, cpu_map
) {
6756 struct sched_group
**sched_group_nodes
6757 = sched_group_nodes_bycpu
[cpu
];
6759 if (!sched_group_nodes
)
6762 for (i
= 0; i
< nr_node_ids
; i
++) {
6763 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6765 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6766 if (cpumask_empty(nodemask
))
6776 if (oldsg
!= sched_group_nodes
[i
])
6779 kfree(sched_group_nodes
);
6780 sched_group_nodes_bycpu
[cpu
] = NULL
;
6783 #else /* !CONFIG_NUMA */
6784 static void free_sched_groups(const struct cpumask
*cpu_map
,
6785 struct cpumask
*nodemask
)
6788 #endif /* CONFIG_NUMA */
6791 * Initialize sched groups cpu_power.
6793 * cpu_power indicates the capacity of sched group, which is used while
6794 * distributing the load between different sched groups in a sched domain.
6795 * Typically cpu_power for all the groups in a sched domain will be same unless
6796 * there are asymmetries in the topology. If there are asymmetries, group
6797 * having more cpu_power will pickup more load compared to the group having
6800 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6802 struct sched_domain
*child
;
6803 struct sched_group
*group
;
6807 WARN_ON(!sd
|| !sd
->groups
);
6809 if (cpu
!= group_first_cpu(sd
->groups
))
6814 sd
->groups
->cpu_power
= 0;
6817 power
= SCHED_LOAD_SCALE
;
6818 weight
= cpumask_weight(sched_domain_span(sd
));
6820 * SMT siblings share the power of a single core.
6821 * Usually multiple threads get a better yield out of
6822 * that one core than a single thread would have,
6823 * reflect that in sd->smt_gain.
6825 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6826 power
*= sd
->smt_gain
;
6828 power
>>= SCHED_LOAD_SHIFT
;
6830 sd
->groups
->cpu_power
+= power
;
6835 * Add cpu_power of each child group to this groups cpu_power.
6837 group
= child
->groups
;
6839 sd
->groups
->cpu_power
+= group
->cpu_power
;
6840 group
= group
->next
;
6841 } while (group
!= child
->groups
);
6845 * Initializers for schedule domains
6846 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6849 #ifdef CONFIG_SCHED_DEBUG
6850 # define SD_INIT_NAME(sd, type) sd->name = #type
6852 # define SD_INIT_NAME(sd, type) do { } while (0)
6855 #define SD_INIT(sd, type) sd_init_##type(sd)
6857 #define SD_INIT_FUNC(type) \
6858 static noinline void sd_init_##type(struct sched_domain *sd) \
6860 memset(sd, 0, sizeof(*sd)); \
6861 *sd = SD_##type##_INIT; \
6862 sd->level = SD_LV_##type; \
6863 SD_INIT_NAME(sd, type); \
6868 SD_INIT_FUNC(ALLNODES
)
6871 #ifdef CONFIG_SCHED_SMT
6872 SD_INIT_FUNC(SIBLING
)
6874 #ifdef CONFIG_SCHED_MC
6877 #ifdef CONFIG_SCHED_BOOK
6881 static int default_relax_domain_level
= -1;
6883 static int __init
setup_relax_domain_level(char *str
)
6887 val
= simple_strtoul(str
, NULL
, 0);
6888 if (val
< SD_LV_MAX
)
6889 default_relax_domain_level
= val
;
6893 __setup("relax_domain_level=", setup_relax_domain_level
);
6895 static void set_domain_attribute(struct sched_domain
*sd
,
6896 struct sched_domain_attr
*attr
)
6900 if (!attr
|| attr
->relax_domain_level
< 0) {
6901 if (default_relax_domain_level
< 0)
6904 request
= default_relax_domain_level
;
6906 request
= attr
->relax_domain_level
;
6907 if (request
< sd
->level
) {
6908 /* turn off idle balance on this domain */
6909 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6911 /* turn on idle balance on this domain */
6912 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6916 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6917 const struct cpumask
*cpu_map
)
6920 case sa_sched_groups
:
6921 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6922 d
->sched_group_nodes
= NULL
;
6924 free_rootdomain(d
->rd
); /* fall through */
6926 free_cpumask_var(d
->tmpmask
); /* fall through */
6927 case sa_send_covered
:
6928 free_cpumask_var(d
->send_covered
); /* fall through */
6929 case sa_this_book_map
:
6930 free_cpumask_var(d
->this_book_map
); /* fall through */
6931 case sa_this_core_map
:
6932 free_cpumask_var(d
->this_core_map
); /* fall through */
6933 case sa_this_sibling_map
:
6934 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6936 free_cpumask_var(d
->nodemask
); /* fall through */
6937 case sa_sched_group_nodes
:
6939 kfree(d
->sched_group_nodes
); /* fall through */
6941 free_cpumask_var(d
->notcovered
); /* fall through */
6943 free_cpumask_var(d
->covered
); /* fall through */
6945 free_cpumask_var(d
->domainspan
); /* fall through */
6952 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6953 const struct cpumask
*cpu_map
)
6956 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6958 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6959 return sa_domainspan
;
6960 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6962 /* Allocate the per-node list of sched groups */
6963 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6964 sizeof(struct sched_group
*), GFP_KERNEL
);
6965 if (!d
->sched_group_nodes
) {
6966 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6967 return sa_notcovered
;
6969 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6971 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6972 return sa_sched_group_nodes
;
6973 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6975 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6976 return sa_this_sibling_map
;
6977 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
6978 return sa_this_core_map
;
6979 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6980 return sa_this_book_map
;
6981 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6982 return sa_send_covered
;
6983 d
->rd
= alloc_rootdomain();
6985 printk(KERN_WARNING
"Cannot alloc root domain\n");
6988 return sa_rootdomain
;
6991 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6992 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6994 struct sched_domain
*sd
= NULL
;
6996 struct sched_domain
*parent
;
6999 if (cpumask_weight(cpu_map
) >
7000 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7001 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7002 SD_INIT(sd
, ALLNODES
);
7003 set_domain_attribute(sd
, attr
);
7004 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7005 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7010 sd
= &per_cpu(node_domains
, i
).sd
;
7012 set_domain_attribute(sd
, attr
);
7013 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7014 sd
->parent
= parent
;
7017 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7022 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7023 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7024 struct sched_domain
*parent
, int i
)
7026 struct sched_domain
*sd
;
7027 sd
= &per_cpu(phys_domains
, i
).sd
;
7029 set_domain_attribute(sd
, attr
);
7030 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7031 sd
->parent
= parent
;
7034 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7038 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7039 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7040 struct sched_domain
*parent
, int i
)
7042 struct sched_domain
*sd
= parent
;
7043 #ifdef CONFIG_SCHED_BOOK
7044 sd
= &per_cpu(book_domains
, i
).sd
;
7046 set_domain_attribute(sd
, attr
);
7047 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7048 sd
->parent
= parent
;
7050 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7055 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7056 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7057 struct sched_domain
*parent
, int i
)
7059 struct sched_domain
*sd
= parent
;
7060 #ifdef CONFIG_SCHED_MC
7061 sd
= &per_cpu(core_domains
, i
).sd
;
7063 set_domain_attribute(sd
, attr
);
7064 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7065 sd
->parent
= parent
;
7067 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7072 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7073 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7074 struct sched_domain
*parent
, int i
)
7076 struct sched_domain
*sd
= parent
;
7077 #ifdef CONFIG_SCHED_SMT
7078 sd
= &per_cpu(cpu_domains
, i
).sd
;
7079 SD_INIT(sd
, SIBLING
);
7080 set_domain_attribute(sd
, attr
);
7081 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7082 sd
->parent
= parent
;
7084 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7089 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7090 const struct cpumask
*cpu_map
, int cpu
)
7093 #ifdef CONFIG_SCHED_SMT
7094 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7095 cpumask_and(d
->this_sibling_map
, cpu_map
,
7096 topology_thread_cpumask(cpu
));
7097 if (cpu
== cpumask_first(d
->this_sibling_map
))
7098 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7100 d
->send_covered
, d
->tmpmask
);
7103 #ifdef CONFIG_SCHED_MC
7104 case SD_LV_MC
: /* set up multi-core groups */
7105 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7106 if (cpu
== cpumask_first(d
->this_core_map
))
7107 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7109 d
->send_covered
, d
->tmpmask
);
7112 #ifdef CONFIG_SCHED_BOOK
7113 case SD_LV_BOOK
: /* set up book groups */
7114 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7115 if (cpu
== cpumask_first(d
->this_book_map
))
7116 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7118 d
->send_covered
, d
->tmpmask
);
7121 case SD_LV_CPU
: /* set up physical groups */
7122 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7123 if (!cpumask_empty(d
->nodemask
))
7124 init_sched_build_groups(d
->nodemask
, cpu_map
,
7126 d
->send_covered
, d
->tmpmask
);
7129 case SD_LV_ALLNODES
:
7130 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7131 d
->send_covered
, d
->tmpmask
);
7140 * Build sched domains for a given set of cpus and attach the sched domains
7141 * to the individual cpus
7143 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7144 struct sched_domain_attr
*attr
)
7146 enum s_alloc alloc_state
= sa_none
;
7148 struct sched_domain
*sd
;
7154 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7155 if (alloc_state
!= sa_rootdomain
)
7157 alloc_state
= sa_sched_groups
;
7160 * Set up domains for cpus specified by the cpu_map.
7162 for_each_cpu(i
, cpu_map
) {
7163 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7166 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7167 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7168 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7169 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7170 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7173 for_each_cpu(i
, cpu_map
) {
7174 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7175 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7176 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7179 /* Set up physical groups */
7180 for (i
= 0; i
< nr_node_ids
; i
++)
7181 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7184 /* Set up node groups */
7186 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7188 for (i
= 0; i
< nr_node_ids
; i
++)
7189 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7193 /* Calculate CPU power for physical packages and nodes */
7194 #ifdef CONFIG_SCHED_SMT
7195 for_each_cpu(i
, cpu_map
) {
7196 sd
= &per_cpu(cpu_domains
, i
).sd
;
7197 init_sched_groups_power(i
, sd
);
7200 #ifdef CONFIG_SCHED_MC
7201 for_each_cpu(i
, cpu_map
) {
7202 sd
= &per_cpu(core_domains
, i
).sd
;
7203 init_sched_groups_power(i
, sd
);
7206 #ifdef CONFIG_SCHED_BOOK
7207 for_each_cpu(i
, cpu_map
) {
7208 sd
= &per_cpu(book_domains
, i
).sd
;
7209 init_sched_groups_power(i
, sd
);
7213 for_each_cpu(i
, cpu_map
) {
7214 sd
= &per_cpu(phys_domains
, i
).sd
;
7215 init_sched_groups_power(i
, sd
);
7219 for (i
= 0; i
< nr_node_ids
; i
++)
7220 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7222 if (d
.sd_allnodes
) {
7223 struct sched_group
*sg
;
7225 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7227 init_numa_sched_groups_power(sg
);
7231 /* Attach the domains */
7232 for_each_cpu(i
, cpu_map
) {
7233 #ifdef CONFIG_SCHED_SMT
7234 sd
= &per_cpu(cpu_domains
, i
).sd
;
7235 #elif defined(CONFIG_SCHED_MC)
7236 sd
= &per_cpu(core_domains
, i
).sd
;
7237 #elif defined(CONFIG_SCHED_BOOK)
7238 sd
= &per_cpu(book_domains
, i
).sd
;
7240 sd
= &per_cpu(phys_domains
, i
).sd
;
7242 cpu_attach_domain(sd
, d
.rd
, i
);
7245 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7246 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7250 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7254 static int build_sched_domains(const struct cpumask
*cpu_map
)
7256 return __build_sched_domains(cpu_map
, NULL
);
7259 static cpumask_var_t
*doms_cur
; /* current sched domains */
7260 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7261 static struct sched_domain_attr
*dattr_cur
;
7262 /* attribues of custom domains in 'doms_cur' */
7265 * Special case: If a kmalloc of a doms_cur partition (array of
7266 * cpumask) fails, then fallback to a single sched domain,
7267 * as determined by the single cpumask fallback_doms.
7269 static cpumask_var_t fallback_doms
;
7272 * arch_update_cpu_topology lets virtualized architectures update the
7273 * cpu core maps. It is supposed to return 1 if the topology changed
7274 * or 0 if it stayed the same.
7276 int __attribute__((weak
)) arch_update_cpu_topology(void)
7281 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7284 cpumask_var_t
*doms
;
7286 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7289 for (i
= 0; i
< ndoms
; i
++) {
7290 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7291 free_sched_domains(doms
, i
);
7298 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7301 for (i
= 0; i
< ndoms
; i
++)
7302 free_cpumask_var(doms
[i
]);
7307 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7308 * For now this just excludes isolated cpus, but could be used to
7309 * exclude other special cases in the future.
7311 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7315 arch_update_cpu_topology();
7317 doms_cur
= alloc_sched_domains(ndoms_cur
);
7319 doms_cur
= &fallback_doms
;
7320 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7322 err
= build_sched_domains(doms_cur
[0]);
7323 register_sched_domain_sysctl();
7328 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7329 struct cpumask
*tmpmask
)
7331 free_sched_groups(cpu_map
, tmpmask
);
7335 * Detach sched domains from a group of cpus specified in cpu_map
7336 * These cpus will now be attached to the NULL domain
7338 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7340 /* Save because hotplug lock held. */
7341 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7344 for_each_cpu(i
, cpu_map
)
7345 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7346 synchronize_sched();
7347 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7350 /* handle null as "default" */
7351 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7352 struct sched_domain_attr
*new, int idx_new
)
7354 struct sched_domain_attr tmp
;
7361 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7362 new ? (new + idx_new
) : &tmp
,
7363 sizeof(struct sched_domain_attr
));
7367 * Partition sched domains as specified by the 'ndoms_new'
7368 * cpumasks in the array doms_new[] of cpumasks. This compares
7369 * doms_new[] to the current sched domain partitioning, doms_cur[].
7370 * It destroys each deleted domain and builds each new domain.
7372 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7373 * The masks don't intersect (don't overlap.) We should setup one
7374 * sched domain for each mask. CPUs not in any of the cpumasks will
7375 * not be load balanced. If the same cpumask appears both in the
7376 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7379 * The passed in 'doms_new' should be allocated using
7380 * alloc_sched_domains. This routine takes ownership of it and will
7381 * free_sched_domains it when done with it. If the caller failed the
7382 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7383 * and partition_sched_domains() will fallback to the single partition
7384 * 'fallback_doms', it also forces the domains to be rebuilt.
7386 * If doms_new == NULL it will be replaced with cpu_online_mask.
7387 * ndoms_new == 0 is a special case for destroying existing domains,
7388 * and it will not create the default domain.
7390 * Call with hotplug lock held
7392 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7393 struct sched_domain_attr
*dattr_new
)
7398 mutex_lock(&sched_domains_mutex
);
7400 /* always unregister in case we don't destroy any domains */
7401 unregister_sched_domain_sysctl();
7403 /* Let architecture update cpu core mappings. */
7404 new_topology
= arch_update_cpu_topology();
7406 n
= doms_new
? ndoms_new
: 0;
7408 /* Destroy deleted domains */
7409 for (i
= 0; i
< ndoms_cur
; i
++) {
7410 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7411 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7412 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7415 /* no match - a current sched domain not in new doms_new[] */
7416 detach_destroy_domains(doms_cur
[i
]);
7421 if (doms_new
== NULL
) {
7423 doms_new
= &fallback_doms
;
7424 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7425 WARN_ON_ONCE(dattr_new
);
7428 /* Build new domains */
7429 for (i
= 0; i
< ndoms_new
; i
++) {
7430 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7431 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7432 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7435 /* no match - add a new doms_new */
7436 __build_sched_domains(doms_new
[i
],
7437 dattr_new
? dattr_new
+ i
: NULL
);
7442 /* Remember the new sched domains */
7443 if (doms_cur
!= &fallback_doms
)
7444 free_sched_domains(doms_cur
, ndoms_cur
);
7445 kfree(dattr_cur
); /* kfree(NULL) is safe */
7446 doms_cur
= doms_new
;
7447 dattr_cur
= dattr_new
;
7448 ndoms_cur
= ndoms_new
;
7450 register_sched_domain_sysctl();
7452 mutex_unlock(&sched_domains_mutex
);
7455 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7456 static void arch_reinit_sched_domains(void)
7460 /* Destroy domains first to force the rebuild */
7461 partition_sched_domains(0, NULL
, NULL
);
7463 rebuild_sched_domains();
7467 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7469 unsigned int level
= 0;
7471 if (sscanf(buf
, "%u", &level
) != 1)
7475 * level is always be positive so don't check for
7476 * level < POWERSAVINGS_BALANCE_NONE which is 0
7477 * What happens on 0 or 1 byte write,
7478 * need to check for count as well?
7481 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7485 sched_smt_power_savings
= level
;
7487 sched_mc_power_savings
= level
;
7489 arch_reinit_sched_domains();
7494 #ifdef CONFIG_SCHED_MC
7495 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7496 struct sysdev_class_attribute
*attr
,
7499 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7501 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7502 struct sysdev_class_attribute
*attr
,
7503 const char *buf
, size_t count
)
7505 return sched_power_savings_store(buf
, count
, 0);
7507 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7508 sched_mc_power_savings_show
,
7509 sched_mc_power_savings_store
);
7512 #ifdef CONFIG_SCHED_SMT
7513 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7514 struct sysdev_class_attribute
*attr
,
7517 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7519 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7520 struct sysdev_class_attribute
*attr
,
7521 const char *buf
, size_t count
)
7523 return sched_power_savings_store(buf
, count
, 1);
7525 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7526 sched_smt_power_savings_show
,
7527 sched_smt_power_savings_store
);
7530 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7534 #ifdef CONFIG_SCHED_SMT
7536 err
= sysfs_create_file(&cls
->kset
.kobj
,
7537 &attr_sched_smt_power_savings
.attr
);
7539 #ifdef CONFIG_SCHED_MC
7540 if (!err
&& mc_capable())
7541 err
= sysfs_create_file(&cls
->kset
.kobj
,
7542 &attr_sched_mc_power_savings
.attr
);
7546 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7549 * Update cpusets according to cpu_active mask. If cpusets are
7550 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7551 * around partition_sched_domains().
7553 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7556 switch (action
& ~CPU_TASKS_FROZEN
) {
7558 case CPU_DOWN_FAILED
:
7559 cpuset_update_active_cpus();
7566 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7569 switch (action
& ~CPU_TASKS_FROZEN
) {
7570 case CPU_DOWN_PREPARE
:
7571 cpuset_update_active_cpus();
7578 static int update_runtime(struct notifier_block
*nfb
,
7579 unsigned long action
, void *hcpu
)
7581 int cpu
= (int)(long)hcpu
;
7584 case CPU_DOWN_PREPARE
:
7585 case CPU_DOWN_PREPARE_FROZEN
:
7586 disable_runtime(cpu_rq(cpu
));
7589 case CPU_DOWN_FAILED
:
7590 case CPU_DOWN_FAILED_FROZEN
:
7592 case CPU_ONLINE_FROZEN
:
7593 enable_runtime(cpu_rq(cpu
));
7601 void __init
sched_init_smp(void)
7603 cpumask_var_t non_isolated_cpus
;
7605 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7606 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7608 #if defined(CONFIG_NUMA)
7609 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7611 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7614 mutex_lock(&sched_domains_mutex
);
7615 arch_init_sched_domains(cpu_active_mask
);
7616 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7617 if (cpumask_empty(non_isolated_cpus
))
7618 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7619 mutex_unlock(&sched_domains_mutex
);
7622 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7623 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7625 /* RT runtime code needs to handle some hotplug events */
7626 hotcpu_notifier(update_runtime
, 0);
7630 /* Move init over to a non-isolated CPU */
7631 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7633 sched_init_granularity();
7634 free_cpumask_var(non_isolated_cpus
);
7636 init_sched_rt_class();
7639 void __init
sched_init_smp(void)
7641 sched_init_granularity();
7643 #endif /* CONFIG_SMP */
7645 const_debug
unsigned int sysctl_timer_migration
= 1;
7647 int in_sched_functions(unsigned long addr
)
7649 return in_lock_functions(addr
) ||
7650 (addr
>= (unsigned long)__sched_text_start
7651 && addr
< (unsigned long)__sched_text_end
);
7654 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7656 cfs_rq
->tasks_timeline
= RB_ROOT
;
7657 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7658 #ifdef CONFIG_FAIR_GROUP_SCHED
7661 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7664 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7666 struct rt_prio_array
*array
;
7669 array
= &rt_rq
->active
;
7670 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7671 INIT_LIST_HEAD(array
->queue
+ i
);
7672 __clear_bit(i
, array
->bitmap
);
7674 /* delimiter for bitsearch: */
7675 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7677 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7678 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7680 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7684 rt_rq
->rt_nr_migratory
= 0;
7685 rt_rq
->overloaded
= 0;
7686 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7690 rt_rq
->rt_throttled
= 0;
7691 rt_rq
->rt_runtime
= 0;
7692 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7694 #ifdef CONFIG_RT_GROUP_SCHED
7695 rt_rq
->rt_nr_boosted
= 0;
7700 #ifdef CONFIG_FAIR_GROUP_SCHED
7701 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7702 struct sched_entity
*se
, int cpu
, int add
,
7703 struct sched_entity
*parent
)
7705 struct rq
*rq
= cpu_rq(cpu
);
7706 tg
->cfs_rq
[cpu
] = cfs_rq
;
7707 init_cfs_rq(cfs_rq
, rq
);
7710 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7713 /* se could be NULL for init_task_group */
7718 se
->cfs_rq
= &rq
->cfs
;
7720 se
->cfs_rq
= parent
->my_q
;
7723 se
->load
.weight
= tg
->shares
;
7724 se
->load
.inv_weight
= 0;
7725 se
->parent
= parent
;
7729 #ifdef CONFIG_RT_GROUP_SCHED
7730 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7731 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7732 struct sched_rt_entity
*parent
)
7734 struct rq
*rq
= cpu_rq(cpu
);
7736 tg
->rt_rq
[cpu
] = rt_rq
;
7737 init_rt_rq(rt_rq
, rq
);
7739 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7741 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7743 tg
->rt_se
[cpu
] = rt_se
;
7748 rt_se
->rt_rq
= &rq
->rt
;
7750 rt_se
->rt_rq
= parent
->my_q
;
7752 rt_se
->my_q
= rt_rq
;
7753 rt_se
->parent
= parent
;
7754 INIT_LIST_HEAD(&rt_se
->run_list
);
7758 void __init
sched_init(void)
7761 unsigned long alloc_size
= 0, ptr
;
7763 #ifdef CONFIG_FAIR_GROUP_SCHED
7764 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7766 #ifdef CONFIG_RT_GROUP_SCHED
7767 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7769 #ifdef CONFIG_CPUMASK_OFFSTACK
7770 alloc_size
+= num_possible_cpus() * cpumask_size();
7773 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7775 #ifdef CONFIG_FAIR_GROUP_SCHED
7776 init_task_group
.se
= (struct sched_entity
**)ptr
;
7777 ptr
+= nr_cpu_ids
* sizeof(void **);
7779 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7780 ptr
+= nr_cpu_ids
* sizeof(void **);
7782 #endif /* CONFIG_FAIR_GROUP_SCHED */
7783 #ifdef CONFIG_RT_GROUP_SCHED
7784 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7785 ptr
+= nr_cpu_ids
* sizeof(void **);
7787 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7788 ptr
+= nr_cpu_ids
* sizeof(void **);
7790 #endif /* CONFIG_RT_GROUP_SCHED */
7791 #ifdef CONFIG_CPUMASK_OFFSTACK
7792 for_each_possible_cpu(i
) {
7793 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7794 ptr
+= cpumask_size();
7796 #endif /* CONFIG_CPUMASK_OFFSTACK */
7800 init_defrootdomain();
7803 init_rt_bandwidth(&def_rt_bandwidth
,
7804 global_rt_period(), global_rt_runtime());
7806 #ifdef CONFIG_RT_GROUP_SCHED
7807 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7808 global_rt_period(), global_rt_runtime());
7809 #endif /* CONFIG_RT_GROUP_SCHED */
7811 #ifdef CONFIG_CGROUP_SCHED
7812 list_add(&init_task_group
.list
, &task_groups
);
7813 INIT_LIST_HEAD(&init_task_group
.children
);
7815 #endif /* CONFIG_CGROUP_SCHED */
7817 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7818 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7819 __alignof__(unsigned long));
7821 for_each_possible_cpu(i
) {
7825 raw_spin_lock_init(&rq
->lock
);
7827 rq
->calc_load_active
= 0;
7828 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7829 init_cfs_rq(&rq
->cfs
, rq
);
7830 init_rt_rq(&rq
->rt
, rq
);
7831 #ifdef CONFIG_FAIR_GROUP_SCHED
7832 init_task_group
.shares
= init_task_group_load
;
7833 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7834 #ifdef CONFIG_CGROUP_SCHED
7836 * How much cpu bandwidth does init_task_group get?
7838 * In case of task-groups formed thr' the cgroup filesystem, it
7839 * gets 100% of the cpu resources in the system. This overall
7840 * system cpu resource is divided among the tasks of
7841 * init_task_group and its child task-groups in a fair manner,
7842 * based on each entity's (task or task-group's) weight
7843 * (se->load.weight).
7845 * In other words, if init_task_group has 10 tasks of weight
7846 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7847 * then A0's share of the cpu resource is:
7849 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7851 * We achieve this by letting init_task_group's tasks sit
7852 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7854 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7856 #endif /* CONFIG_FAIR_GROUP_SCHED */
7858 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7859 #ifdef CONFIG_RT_GROUP_SCHED
7860 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7861 #ifdef CONFIG_CGROUP_SCHED
7862 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7866 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7867 rq
->cpu_load
[j
] = 0;
7869 rq
->last_load_update_tick
= jiffies
;
7874 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7875 rq
->post_schedule
= 0;
7876 rq
->active_balance
= 0;
7877 rq
->next_balance
= jiffies
;
7882 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7883 rq_attach_root(rq
, &def_root_domain
);
7885 rq
->nohz_balance_kick
= 0;
7886 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
7890 atomic_set(&rq
->nr_iowait
, 0);
7893 set_load_weight(&init_task
);
7895 #ifdef CONFIG_PREEMPT_NOTIFIERS
7896 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7900 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7903 #ifdef CONFIG_RT_MUTEXES
7904 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7908 * The boot idle thread does lazy MMU switching as well:
7910 atomic_inc(&init_mm
.mm_count
);
7911 enter_lazy_tlb(&init_mm
, current
);
7914 * Make us the idle thread. Technically, schedule() should not be
7915 * called from this thread, however somewhere below it might be,
7916 * but because we are the idle thread, we just pick up running again
7917 * when this runqueue becomes "idle".
7919 init_idle(current
, smp_processor_id());
7921 calc_load_update
= jiffies
+ LOAD_FREQ
;
7924 * During early bootup we pretend to be a normal task:
7926 current
->sched_class
= &fair_sched_class
;
7928 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7929 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7932 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7933 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
7934 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
7935 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
7936 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
7938 /* May be allocated at isolcpus cmdline parse time */
7939 if (cpu_isolated_map
== NULL
)
7940 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7945 scheduler_running
= 1;
7948 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7949 static inline int preempt_count_equals(int preempt_offset
)
7951 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7953 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7956 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7959 static unsigned long prev_jiffy
; /* ratelimiting */
7961 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7962 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7964 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7966 prev_jiffy
= jiffies
;
7969 "BUG: sleeping function called from invalid context at %s:%d\n",
7972 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7973 in_atomic(), irqs_disabled(),
7974 current
->pid
, current
->comm
);
7976 debug_show_held_locks(current
);
7977 if (irqs_disabled())
7978 print_irqtrace_events(current
);
7982 EXPORT_SYMBOL(__might_sleep
);
7985 #ifdef CONFIG_MAGIC_SYSRQ
7986 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7990 on_rq
= p
->se
.on_rq
;
7992 deactivate_task(rq
, p
, 0);
7993 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7995 activate_task(rq
, p
, 0);
7996 resched_task(rq
->curr
);
8000 void normalize_rt_tasks(void)
8002 struct task_struct
*g
, *p
;
8003 unsigned long flags
;
8006 read_lock_irqsave(&tasklist_lock
, flags
);
8007 do_each_thread(g
, p
) {
8009 * Only normalize user tasks:
8014 p
->se
.exec_start
= 0;
8015 #ifdef CONFIG_SCHEDSTATS
8016 p
->se
.statistics
.wait_start
= 0;
8017 p
->se
.statistics
.sleep_start
= 0;
8018 p
->se
.statistics
.block_start
= 0;
8023 * Renice negative nice level userspace
8026 if (TASK_NICE(p
) < 0 && p
->mm
)
8027 set_user_nice(p
, 0);
8031 raw_spin_lock(&p
->pi_lock
);
8032 rq
= __task_rq_lock(p
);
8034 normalize_task(rq
, p
);
8036 __task_rq_unlock(rq
);
8037 raw_spin_unlock(&p
->pi_lock
);
8038 } while_each_thread(g
, p
);
8040 read_unlock_irqrestore(&tasklist_lock
, flags
);
8043 #endif /* CONFIG_MAGIC_SYSRQ */
8045 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8047 * These functions are only useful for the IA64 MCA handling, or kdb.
8049 * They can only be called when the whole system has been
8050 * stopped - every CPU needs to be quiescent, and no scheduling
8051 * activity can take place. Using them for anything else would
8052 * be a serious bug, and as a result, they aren't even visible
8053 * under any other configuration.
8057 * curr_task - return the current task for a given cpu.
8058 * @cpu: the processor in question.
8060 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8062 struct task_struct
*curr_task(int cpu
)
8064 return cpu_curr(cpu
);
8067 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8071 * set_curr_task - set the current task for a given cpu.
8072 * @cpu: the processor in question.
8073 * @p: the task pointer to set.
8075 * Description: This function must only be used when non-maskable interrupts
8076 * are serviced on a separate stack. It allows the architecture to switch the
8077 * notion of the current task on a cpu in a non-blocking manner. This function
8078 * must be called with all CPU's synchronized, and interrupts disabled, the
8079 * and caller must save the original value of the current task (see
8080 * curr_task() above) and restore that value before reenabling interrupts and
8081 * re-starting the system.
8083 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8085 void set_curr_task(int cpu
, struct task_struct
*p
)
8092 #ifdef CONFIG_FAIR_GROUP_SCHED
8093 static void free_fair_sched_group(struct task_group
*tg
)
8097 for_each_possible_cpu(i
) {
8099 kfree(tg
->cfs_rq
[i
]);
8109 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8111 struct cfs_rq
*cfs_rq
;
8112 struct sched_entity
*se
;
8116 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8119 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8123 tg
->shares
= NICE_0_LOAD
;
8125 for_each_possible_cpu(i
) {
8128 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8129 GFP_KERNEL
, cpu_to_node(i
));
8133 se
= kzalloc_node(sizeof(struct sched_entity
),
8134 GFP_KERNEL
, cpu_to_node(i
));
8138 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8149 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8151 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8152 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8155 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8157 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8159 #else /* !CONFG_FAIR_GROUP_SCHED */
8160 static inline void free_fair_sched_group(struct task_group
*tg
)
8165 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8170 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8174 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8177 #endif /* CONFIG_FAIR_GROUP_SCHED */
8179 #ifdef CONFIG_RT_GROUP_SCHED
8180 static void free_rt_sched_group(struct task_group
*tg
)
8184 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8186 for_each_possible_cpu(i
) {
8188 kfree(tg
->rt_rq
[i
]);
8190 kfree(tg
->rt_se
[i
]);
8198 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8200 struct rt_rq
*rt_rq
;
8201 struct sched_rt_entity
*rt_se
;
8205 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8208 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8212 init_rt_bandwidth(&tg
->rt_bandwidth
,
8213 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8215 for_each_possible_cpu(i
) {
8218 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8219 GFP_KERNEL
, cpu_to_node(i
));
8223 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8224 GFP_KERNEL
, cpu_to_node(i
));
8228 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8239 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8241 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8242 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8245 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8247 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8249 #else /* !CONFIG_RT_GROUP_SCHED */
8250 static inline void free_rt_sched_group(struct task_group
*tg
)
8255 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8260 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8264 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8267 #endif /* CONFIG_RT_GROUP_SCHED */
8269 #ifdef CONFIG_CGROUP_SCHED
8270 static void free_sched_group(struct task_group
*tg
)
8272 free_fair_sched_group(tg
);
8273 free_rt_sched_group(tg
);
8277 /* allocate runqueue etc for a new task group */
8278 struct task_group
*sched_create_group(struct task_group
*parent
)
8280 struct task_group
*tg
;
8281 unsigned long flags
;
8284 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8286 return ERR_PTR(-ENOMEM
);
8288 if (!alloc_fair_sched_group(tg
, parent
))
8291 if (!alloc_rt_sched_group(tg
, parent
))
8294 spin_lock_irqsave(&task_group_lock
, flags
);
8295 for_each_possible_cpu(i
) {
8296 register_fair_sched_group(tg
, i
);
8297 register_rt_sched_group(tg
, i
);
8299 list_add_rcu(&tg
->list
, &task_groups
);
8301 WARN_ON(!parent
); /* root should already exist */
8303 tg
->parent
= parent
;
8304 INIT_LIST_HEAD(&tg
->children
);
8305 list_add_rcu(&tg
->siblings
, &parent
->children
);
8306 spin_unlock_irqrestore(&task_group_lock
, flags
);
8311 free_sched_group(tg
);
8312 return ERR_PTR(-ENOMEM
);
8315 /* rcu callback to free various structures associated with a task group */
8316 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8318 /* now it should be safe to free those cfs_rqs */
8319 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8322 /* Destroy runqueue etc associated with a task group */
8323 void sched_destroy_group(struct task_group
*tg
)
8325 unsigned long flags
;
8328 spin_lock_irqsave(&task_group_lock
, flags
);
8329 for_each_possible_cpu(i
) {
8330 unregister_fair_sched_group(tg
, i
);
8331 unregister_rt_sched_group(tg
, i
);
8333 list_del_rcu(&tg
->list
);
8334 list_del_rcu(&tg
->siblings
);
8335 spin_unlock_irqrestore(&task_group_lock
, flags
);
8337 /* wait for possible concurrent references to cfs_rqs complete */
8338 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8341 /* change task's runqueue when it moves between groups.
8342 * The caller of this function should have put the task in its new group
8343 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8344 * reflect its new group.
8346 void sched_move_task(struct task_struct
*tsk
)
8349 unsigned long flags
;
8352 rq
= task_rq_lock(tsk
, &flags
);
8354 running
= task_current(rq
, tsk
);
8355 on_rq
= tsk
->se
.on_rq
;
8358 dequeue_task(rq
, tsk
, 0);
8359 if (unlikely(running
))
8360 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8362 set_task_rq(tsk
, task_cpu(tsk
));
8364 #ifdef CONFIG_FAIR_GROUP_SCHED
8365 if (tsk
->sched_class
->moved_group
)
8366 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8369 if (unlikely(running
))
8370 tsk
->sched_class
->set_curr_task(rq
);
8372 enqueue_task(rq
, tsk
, 0);
8374 task_rq_unlock(rq
, &flags
);
8376 #endif /* CONFIG_CGROUP_SCHED */
8378 #ifdef CONFIG_FAIR_GROUP_SCHED
8379 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8381 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8386 dequeue_entity(cfs_rq
, se
, 0);
8388 se
->load
.weight
= shares
;
8389 se
->load
.inv_weight
= 0;
8392 enqueue_entity(cfs_rq
, se
, 0);
8395 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8397 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8398 struct rq
*rq
= cfs_rq
->rq
;
8399 unsigned long flags
;
8401 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8402 __set_se_shares(se
, shares
);
8403 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8406 static DEFINE_MUTEX(shares_mutex
);
8408 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8411 unsigned long flags
;
8414 * We can't change the weight of the root cgroup.
8419 if (shares
< MIN_SHARES
)
8420 shares
= MIN_SHARES
;
8421 else if (shares
> MAX_SHARES
)
8422 shares
= MAX_SHARES
;
8424 mutex_lock(&shares_mutex
);
8425 if (tg
->shares
== shares
)
8428 spin_lock_irqsave(&task_group_lock
, flags
);
8429 for_each_possible_cpu(i
)
8430 unregister_fair_sched_group(tg
, i
);
8431 list_del_rcu(&tg
->siblings
);
8432 spin_unlock_irqrestore(&task_group_lock
, flags
);
8434 /* wait for any ongoing reference to this group to finish */
8435 synchronize_sched();
8438 * Now we are free to modify the group's share on each cpu
8439 * w/o tripping rebalance_share or load_balance_fair.
8441 tg
->shares
= shares
;
8442 for_each_possible_cpu(i
) {
8446 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8447 set_se_shares(tg
->se
[i
], shares
);
8451 * Enable load balance activity on this group, by inserting it back on
8452 * each cpu's rq->leaf_cfs_rq_list.
8454 spin_lock_irqsave(&task_group_lock
, flags
);
8455 for_each_possible_cpu(i
)
8456 register_fair_sched_group(tg
, i
);
8457 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8458 spin_unlock_irqrestore(&task_group_lock
, flags
);
8460 mutex_unlock(&shares_mutex
);
8464 unsigned long sched_group_shares(struct task_group
*tg
)
8470 #ifdef CONFIG_RT_GROUP_SCHED
8472 * Ensure that the real time constraints are schedulable.
8474 static DEFINE_MUTEX(rt_constraints_mutex
);
8476 static unsigned long to_ratio(u64 period
, u64 runtime
)
8478 if (runtime
== RUNTIME_INF
)
8481 return div64_u64(runtime
<< 20, period
);
8484 /* Must be called with tasklist_lock held */
8485 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8487 struct task_struct
*g
, *p
;
8489 do_each_thread(g
, p
) {
8490 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8492 } while_each_thread(g
, p
);
8497 struct rt_schedulable_data
{
8498 struct task_group
*tg
;
8503 static int tg_schedulable(struct task_group
*tg
, void *data
)
8505 struct rt_schedulable_data
*d
= data
;
8506 struct task_group
*child
;
8507 unsigned long total
, sum
= 0;
8508 u64 period
, runtime
;
8510 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8511 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8514 period
= d
->rt_period
;
8515 runtime
= d
->rt_runtime
;
8519 * Cannot have more runtime than the period.
8521 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8525 * Ensure we don't starve existing RT tasks.
8527 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8530 total
= to_ratio(period
, runtime
);
8533 * Nobody can have more than the global setting allows.
8535 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8539 * The sum of our children's runtime should not exceed our own.
8541 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8542 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8543 runtime
= child
->rt_bandwidth
.rt_runtime
;
8545 if (child
== d
->tg
) {
8546 period
= d
->rt_period
;
8547 runtime
= d
->rt_runtime
;
8550 sum
+= to_ratio(period
, runtime
);
8559 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8561 struct rt_schedulable_data data
= {
8563 .rt_period
= period
,
8564 .rt_runtime
= runtime
,
8567 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8570 static int tg_set_bandwidth(struct task_group
*tg
,
8571 u64 rt_period
, u64 rt_runtime
)
8575 mutex_lock(&rt_constraints_mutex
);
8576 read_lock(&tasklist_lock
);
8577 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8581 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8582 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8583 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8585 for_each_possible_cpu(i
) {
8586 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8588 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8589 rt_rq
->rt_runtime
= rt_runtime
;
8590 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8592 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8594 read_unlock(&tasklist_lock
);
8595 mutex_unlock(&rt_constraints_mutex
);
8600 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8602 u64 rt_runtime
, rt_period
;
8604 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8605 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8606 if (rt_runtime_us
< 0)
8607 rt_runtime
= RUNTIME_INF
;
8609 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8612 long sched_group_rt_runtime(struct task_group
*tg
)
8616 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8619 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8620 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8621 return rt_runtime_us
;
8624 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8626 u64 rt_runtime
, rt_period
;
8628 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8629 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8634 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8637 long sched_group_rt_period(struct task_group
*tg
)
8641 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8642 do_div(rt_period_us
, NSEC_PER_USEC
);
8643 return rt_period_us
;
8646 static int sched_rt_global_constraints(void)
8648 u64 runtime
, period
;
8651 if (sysctl_sched_rt_period
<= 0)
8654 runtime
= global_rt_runtime();
8655 period
= global_rt_period();
8658 * Sanity check on the sysctl variables.
8660 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8663 mutex_lock(&rt_constraints_mutex
);
8664 read_lock(&tasklist_lock
);
8665 ret
= __rt_schedulable(NULL
, 0, 0);
8666 read_unlock(&tasklist_lock
);
8667 mutex_unlock(&rt_constraints_mutex
);
8672 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8674 /* Don't accept realtime tasks when there is no way for them to run */
8675 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8681 #else /* !CONFIG_RT_GROUP_SCHED */
8682 static int sched_rt_global_constraints(void)
8684 unsigned long flags
;
8687 if (sysctl_sched_rt_period
<= 0)
8691 * There's always some RT tasks in the root group
8692 * -- migration, kstopmachine etc..
8694 if (sysctl_sched_rt_runtime
== 0)
8697 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8698 for_each_possible_cpu(i
) {
8699 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8701 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8702 rt_rq
->rt_runtime
= global_rt_runtime();
8703 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8705 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8709 #endif /* CONFIG_RT_GROUP_SCHED */
8711 int sched_rt_handler(struct ctl_table
*table
, int write
,
8712 void __user
*buffer
, size_t *lenp
,
8716 int old_period
, old_runtime
;
8717 static DEFINE_MUTEX(mutex
);
8720 old_period
= sysctl_sched_rt_period
;
8721 old_runtime
= sysctl_sched_rt_runtime
;
8723 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8725 if (!ret
&& write
) {
8726 ret
= sched_rt_global_constraints();
8728 sysctl_sched_rt_period
= old_period
;
8729 sysctl_sched_rt_runtime
= old_runtime
;
8731 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8732 def_rt_bandwidth
.rt_period
=
8733 ns_to_ktime(global_rt_period());
8736 mutex_unlock(&mutex
);
8741 #ifdef CONFIG_CGROUP_SCHED
8743 /* return corresponding task_group object of a cgroup */
8744 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8746 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8747 struct task_group
, css
);
8750 static struct cgroup_subsys_state
*
8751 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8753 struct task_group
*tg
, *parent
;
8755 if (!cgrp
->parent
) {
8756 /* This is early initialization for the top cgroup */
8757 return &init_task_group
.css
;
8760 parent
= cgroup_tg(cgrp
->parent
);
8761 tg
= sched_create_group(parent
);
8763 return ERR_PTR(-ENOMEM
);
8769 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8771 struct task_group
*tg
= cgroup_tg(cgrp
);
8773 sched_destroy_group(tg
);
8777 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8779 #ifdef CONFIG_RT_GROUP_SCHED
8780 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8783 /* We don't support RT-tasks being in separate groups */
8784 if (tsk
->sched_class
!= &fair_sched_class
)
8791 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8792 struct task_struct
*tsk
, bool threadgroup
)
8794 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8798 struct task_struct
*c
;
8800 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8801 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8813 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8814 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8817 sched_move_task(tsk
);
8819 struct task_struct
*c
;
8821 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8828 #ifdef CONFIG_FAIR_GROUP_SCHED
8829 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8832 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8835 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8837 struct task_group
*tg
= cgroup_tg(cgrp
);
8839 return (u64
) tg
->shares
;
8841 #endif /* CONFIG_FAIR_GROUP_SCHED */
8843 #ifdef CONFIG_RT_GROUP_SCHED
8844 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8847 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8850 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8852 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8855 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8858 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8861 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8863 return sched_group_rt_period(cgroup_tg(cgrp
));
8865 #endif /* CONFIG_RT_GROUP_SCHED */
8867 static struct cftype cpu_files
[] = {
8868 #ifdef CONFIG_FAIR_GROUP_SCHED
8871 .read_u64
= cpu_shares_read_u64
,
8872 .write_u64
= cpu_shares_write_u64
,
8875 #ifdef CONFIG_RT_GROUP_SCHED
8877 .name
= "rt_runtime_us",
8878 .read_s64
= cpu_rt_runtime_read
,
8879 .write_s64
= cpu_rt_runtime_write
,
8882 .name
= "rt_period_us",
8883 .read_u64
= cpu_rt_period_read_uint
,
8884 .write_u64
= cpu_rt_period_write_uint
,
8889 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8891 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8894 struct cgroup_subsys cpu_cgroup_subsys
= {
8896 .create
= cpu_cgroup_create
,
8897 .destroy
= cpu_cgroup_destroy
,
8898 .can_attach
= cpu_cgroup_can_attach
,
8899 .attach
= cpu_cgroup_attach
,
8900 .populate
= cpu_cgroup_populate
,
8901 .subsys_id
= cpu_cgroup_subsys_id
,
8905 #endif /* CONFIG_CGROUP_SCHED */
8907 #ifdef CONFIG_CGROUP_CPUACCT
8910 * CPU accounting code for task groups.
8912 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8913 * (balbir@in.ibm.com).
8916 /* track cpu usage of a group of tasks and its child groups */
8918 struct cgroup_subsys_state css
;
8919 /* cpuusage holds pointer to a u64-type object on every cpu */
8920 u64 __percpu
*cpuusage
;
8921 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8922 struct cpuacct
*parent
;
8925 struct cgroup_subsys cpuacct_subsys
;
8927 /* return cpu accounting group corresponding to this container */
8928 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8930 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8931 struct cpuacct
, css
);
8934 /* return cpu accounting group to which this task belongs */
8935 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8937 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8938 struct cpuacct
, css
);
8941 /* create a new cpu accounting group */
8942 static struct cgroup_subsys_state
*cpuacct_create(
8943 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8945 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8951 ca
->cpuusage
= alloc_percpu(u64
);
8955 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8956 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8957 goto out_free_counters
;
8960 ca
->parent
= cgroup_ca(cgrp
->parent
);
8966 percpu_counter_destroy(&ca
->cpustat
[i
]);
8967 free_percpu(ca
->cpuusage
);
8971 return ERR_PTR(-ENOMEM
);
8974 /* destroy an existing cpu accounting group */
8976 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8978 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8981 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8982 percpu_counter_destroy(&ca
->cpustat
[i
]);
8983 free_percpu(ca
->cpuusage
);
8987 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8989 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8992 #ifndef CONFIG_64BIT
8994 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8996 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8998 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9006 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9008 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9010 #ifndef CONFIG_64BIT
9012 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9014 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9016 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9022 /* return total cpu usage (in nanoseconds) of a group */
9023 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9025 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9026 u64 totalcpuusage
= 0;
9029 for_each_present_cpu(i
)
9030 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9032 return totalcpuusage
;
9035 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9038 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9047 for_each_present_cpu(i
)
9048 cpuacct_cpuusage_write(ca
, i
, 0);
9054 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9057 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9061 for_each_present_cpu(i
) {
9062 percpu
= cpuacct_cpuusage_read(ca
, i
);
9063 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9065 seq_printf(m
, "\n");
9069 static const char *cpuacct_stat_desc
[] = {
9070 [CPUACCT_STAT_USER
] = "user",
9071 [CPUACCT_STAT_SYSTEM
] = "system",
9074 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9075 struct cgroup_map_cb
*cb
)
9077 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9080 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9081 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9082 val
= cputime64_to_clock_t(val
);
9083 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9088 static struct cftype files
[] = {
9091 .read_u64
= cpuusage_read
,
9092 .write_u64
= cpuusage_write
,
9095 .name
= "usage_percpu",
9096 .read_seq_string
= cpuacct_percpu_seq_read
,
9100 .read_map
= cpuacct_stats_show
,
9104 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9106 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9110 * charge this task's execution time to its accounting group.
9112 * called with rq->lock held.
9114 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9119 if (unlikely(!cpuacct_subsys
.active
))
9122 cpu
= task_cpu(tsk
);
9128 for (; ca
; ca
= ca
->parent
) {
9129 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9130 *cpuusage
+= cputime
;
9137 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9138 * in cputime_t units. As a result, cpuacct_update_stats calls
9139 * percpu_counter_add with values large enough to always overflow the
9140 * per cpu batch limit causing bad SMP scalability.
9142 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9143 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9144 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9147 #define CPUACCT_BATCH \
9148 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9150 #define CPUACCT_BATCH 0
9154 * Charge the system/user time to the task's accounting group.
9156 static void cpuacct_update_stats(struct task_struct
*tsk
,
9157 enum cpuacct_stat_index idx
, cputime_t val
)
9160 int batch
= CPUACCT_BATCH
;
9162 if (unlikely(!cpuacct_subsys
.active
))
9169 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9175 struct cgroup_subsys cpuacct_subsys
= {
9177 .create
= cpuacct_create
,
9178 .destroy
= cpuacct_destroy
,
9179 .populate
= cpuacct_populate
,
9180 .subsys_id
= cpuacct_subsys_id
,
9182 #endif /* CONFIG_CGROUP_CPUACCT */
9186 void synchronize_sched_expedited(void)
9190 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9192 #else /* #ifndef CONFIG_SMP */
9194 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9196 static int synchronize_sched_expedited_cpu_stop(void *data
)
9199 * There must be a full memory barrier on each affected CPU
9200 * between the time that try_stop_cpus() is called and the
9201 * time that it returns.
9203 * In the current initial implementation of cpu_stop, the
9204 * above condition is already met when the control reaches
9205 * this point and the following smp_mb() is not strictly
9206 * necessary. Do smp_mb() anyway for documentation and
9207 * robustness against future implementation changes.
9209 smp_mb(); /* See above comment block. */
9214 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9215 * approach to force grace period to end quickly. This consumes
9216 * significant time on all CPUs, and is thus not recommended for
9217 * any sort of common-case code.
9219 * Note that it is illegal to call this function while holding any
9220 * lock that is acquired by a CPU-hotplug notifier. Failing to
9221 * observe this restriction will result in deadlock.
9223 void synchronize_sched_expedited(void)
9225 int snap
, trycount
= 0;
9227 smp_mb(); /* ensure prior mod happens before capturing snap. */
9228 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9230 while (try_stop_cpus(cpu_online_mask
,
9231 synchronize_sched_expedited_cpu_stop
,
9234 if (trycount
++ < 10)
9235 udelay(trycount
* num_online_cpus());
9237 synchronize_sched();
9240 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9241 smp_mb(); /* ensure test happens before caller kfree */
9246 atomic_inc(&synchronize_sched_expedited_count
);
9247 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9250 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9252 #endif /* #else #ifndef CONFIG_SMP */