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
))
736 if (strncmp(buf
, "NO_", 3) == 0) {
741 for (i
= 0; sched_feat_names
[i
]; i
++) {
742 int len
= strlen(sched_feat_names
[i
]);
744 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
746 sysctl_sched_features
&= ~(1UL << i
);
748 sysctl_sched_features
|= (1UL << i
);
753 if (!sched_feat_names
[i
])
761 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
763 return single_open(filp
, sched_feat_show
, NULL
);
766 static const struct file_operations sched_feat_fops
= {
767 .open
= sched_feat_open
,
768 .write
= sched_feat_write
,
771 .release
= single_release
,
774 static __init
int sched_init_debug(void)
776 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
781 late_initcall(sched_init_debug
);
785 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
788 * Number of tasks to iterate in a single balance run.
789 * Limited because this is done with IRQs disabled.
791 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
794 * ratelimit for updating the group shares.
797 unsigned int sysctl_sched_shares_ratelimit
= 250000;
798 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
801 * Inject some fuzzyness into changing the per-cpu group shares
802 * this avoids remote rq-locks at the expense of fairness.
805 unsigned int sysctl_sched_shares_thresh
= 4;
808 * period over which we average the RT time consumption, measured
813 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period
= 1000000;
821 static __read_mostly
int scheduler_running
;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime
= 950000;
829 static inline u64
global_rt_period(void)
831 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
834 static inline u64
global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime
< 0)
839 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
851 return rq
->curr
== p
;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
857 return task_current(rq
, p
);
860 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
864 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq
->lock
.owner
= current
;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
877 raw_spin_unlock_irq(&rq
->lock
);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
886 return task_current(rq
, p
);
890 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 raw_spin_unlock_irq(&rq
->lock
);
903 raw_spin_unlock(&rq
->lock
);
907 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
928 static inline int task_is_waking(struct task_struct
*p
)
930 return unlikely(p
->state
== TASK_WAKING
);
934 * __task_rq_lock - lock the runqueue a given task resides on.
935 * Must be called interrupts disabled.
937 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
944 raw_spin_lock(&rq
->lock
);
945 if (likely(rq
== task_rq(p
)))
947 raw_spin_unlock(&rq
->lock
);
952 * task_rq_lock - lock the runqueue a given task resides on and disable
953 * interrupts. Note the ordering: we can safely lookup the task_rq without
954 * explicitly disabling preemption.
956 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
962 local_irq_save(*flags
);
964 raw_spin_lock(&rq
->lock
);
965 if (likely(rq
== task_rq(p
)))
967 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
971 static void __task_rq_unlock(struct rq
*rq
)
974 raw_spin_unlock(&rq
->lock
);
977 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
980 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
984 * this_rq_lock - lock this runqueue and disable interrupts.
986 static struct rq
*this_rq_lock(void)
993 raw_spin_lock(&rq
->lock
);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * - enabled by features
1013 * - hrtimer is actually high res
1015 static inline int hrtick_enabled(struct rq
*rq
)
1017 if (!sched_feat(HRTICK
))
1019 if (!cpu_active(cpu_of(rq
)))
1021 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1024 static void hrtick_clear(struct rq
*rq
)
1026 if (hrtimer_active(&rq
->hrtick_timer
))
1027 hrtimer_cancel(&rq
->hrtick_timer
);
1031 * High-resolution timer tick.
1032 * Runs from hardirq context with interrupts disabled.
1034 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1036 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1038 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1040 raw_spin_lock(&rq
->lock
);
1041 update_rq_clock(rq
);
1042 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1043 raw_spin_unlock(&rq
->lock
);
1045 return HRTIMER_NORESTART
;
1050 * called from hardirq (IPI) context
1052 static void __hrtick_start(void *arg
)
1054 struct rq
*rq
= arg
;
1056 raw_spin_lock(&rq
->lock
);
1057 hrtimer_restart(&rq
->hrtick_timer
);
1058 rq
->hrtick_csd_pending
= 0;
1059 raw_spin_unlock(&rq
->lock
);
1063 * Called to set the hrtick timer state.
1065 * called with rq->lock held and irqs disabled
1067 static void hrtick_start(struct rq
*rq
, u64 delay
)
1069 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1070 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1072 hrtimer_set_expires(timer
, time
);
1074 if (rq
== this_rq()) {
1075 hrtimer_restart(timer
);
1076 } else if (!rq
->hrtick_csd_pending
) {
1077 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1078 rq
->hrtick_csd_pending
= 1;
1083 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1085 int cpu
= (int)(long)hcpu
;
1088 case CPU_UP_CANCELED
:
1089 case CPU_UP_CANCELED_FROZEN
:
1090 case CPU_DOWN_PREPARE
:
1091 case CPU_DOWN_PREPARE_FROZEN
:
1093 case CPU_DEAD_FROZEN
:
1094 hrtick_clear(cpu_rq(cpu
));
1101 static __init
void init_hrtick(void)
1103 hotcpu_notifier(hotplug_hrtick
, 0);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq
*rq
, u64 delay
)
1113 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1114 HRTIMER_MODE_REL_PINNED
, 0);
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SMP */
1122 static void init_rq_hrtick(struct rq
*rq
)
1125 rq
->hrtick_csd_pending
= 0;
1127 rq
->hrtick_csd
.flags
= 0;
1128 rq
->hrtick_csd
.func
= __hrtick_start
;
1129 rq
->hrtick_csd
.info
= rq
;
1132 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1133 rq
->hrtick_timer
.function
= hrtick
;
1135 #else /* CONFIG_SCHED_HRTICK */
1136 static inline void hrtick_clear(struct rq
*rq
)
1140 static inline void init_rq_hrtick(struct rq
*rq
)
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SCHED_HRTICK */
1150 * resched_task - mark a task 'to be rescheduled now'.
1152 * On UP this means the setting of the need_resched flag, on SMP it
1153 * might also involve a cross-CPU call to trigger the scheduler on
1158 #ifndef tsk_is_polling
1159 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1162 static void resched_task(struct task_struct
*p
)
1166 assert_raw_spin_locked(&task_rq(p
)->lock
);
1168 if (test_tsk_need_resched(p
))
1171 set_tsk_need_resched(p
);
1174 if (cpu
== smp_processor_id())
1177 /* NEED_RESCHED must be visible before we test polling */
1179 if (!tsk_is_polling(p
))
1180 smp_send_reschedule(cpu
);
1183 static void resched_cpu(int cpu
)
1185 struct rq
*rq
= cpu_rq(cpu
);
1186 unsigned long flags
;
1188 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1190 resched_task(cpu_curr(cpu
));
1191 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1196 * In the semi idle case, use the nearest busy cpu for migrating timers
1197 * from an idle cpu. This is good for power-savings.
1199 * We don't do similar optimization for completely idle system, as
1200 * selecting an idle cpu will add more delays to the timers than intended
1201 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1203 int get_nohz_timer_target(void)
1205 int cpu
= smp_processor_id();
1207 struct sched_domain
*sd
;
1209 for_each_domain(cpu
, sd
) {
1210 for_each_cpu(i
, sched_domain_span(sd
))
1217 * When add_timer_on() enqueues a timer into the timer wheel of an
1218 * idle CPU then this timer might expire before the next timer event
1219 * which is scheduled to wake up that CPU. In case of a completely
1220 * idle system the next event might even be infinite time into the
1221 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1222 * leaves the inner idle loop so the newly added timer is taken into
1223 * account when the CPU goes back to idle and evaluates the timer
1224 * wheel for the next timer event.
1226 void wake_up_idle_cpu(int cpu
)
1228 struct rq
*rq
= cpu_rq(cpu
);
1230 if (cpu
== smp_processor_id())
1234 * This is safe, as this function is called with the timer
1235 * wheel base lock of (cpu) held. When the CPU is on the way
1236 * to idle and has not yet set rq->curr to idle then it will
1237 * be serialized on the timer wheel base lock and take the new
1238 * timer into account automatically.
1240 if (rq
->curr
!= rq
->idle
)
1244 * We can set TIF_RESCHED on the idle task of the other CPU
1245 * lockless. The worst case is that the other CPU runs the
1246 * idle task through an additional NOOP schedule()
1248 set_tsk_need_resched(rq
->idle
);
1250 /* NEED_RESCHED must be visible before we test polling */
1252 if (!tsk_is_polling(rq
->idle
))
1253 smp_send_reschedule(cpu
);
1256 #endif /* CONFIG_NO_HZ */
1258 static u64
sched_avg_period(void)
1260 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1263 static void sched_avg_update(struct rq
*rq
)
1265 s64 period
= sched_avg_period();
1267 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1269 * Inline assembly required to prevent the compiler
1270 * optimising this loop into a divmod call.
1271 * See __iter_div_u64_rem() for another example of this.
1273 asm("" : "+rm" (rq
->age_stamp
));
1274 rq
->age_stamp
+= period
;
1279 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1281 rq
->rt_avg
+= rt_delta
;
1282 sched_avg_update(rq
);
1285 #else /* !CONFIG_SMP */
1286 static void resched_task(struct task_struct
*p
)
1288 assert_raw_spin_locked(&task_rq(p
)->lock
);
1289 set_tsk_need_resched(p
);
1292 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1295 #endif /* CONFIG_SMP */
1297 #if BITS_PER_LONG == 32
1298 # define WMULT_CONST (~0UL)
1300 # define WMULT_CONST (1UL << 32)
1303 #define WMULT_SHIFT 32
1306 * Shift right and round:
1308 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1311 * delta *= weight / lw
1313 static unsigned long
1314 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1315 struct load_weight
*lw
)
1319 if (!lw
->inv_weight
) {
1320 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1323 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1327 tmp
= (u64
)delta_exec
* weight
;
1329 * Check whether we'd overflow the 64-bit multiplication:
1331 if (unlikely(tmp
> WMULT_CONST
))
1332 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1335 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1337 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1340 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1346 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1353 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1354 * of tasks with abnormal "nice" values across CPUs the contribution that
1355 * each task makes to its run queue's load is weighted according to its
1356 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1357 * scaled version of the new time slice allocation that they receive on time
1361 #define WEIGHT_IDLEPRIO 3
1362 #define WMULT_IDLEPRIO 1431655765
1365 * Nice levels are multiplicative, with a gentle 10% change for every
1366 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1367 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1368 * that remained on nice 0.
1370 * The "10% effect" is relative and cumulative: from _any_ nice level,
1371 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1372 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1373 * If a task goes up by ~10% and another task goes down by ~10% then
1374 * the relative distance between them is ~25%.)
1376 static const int prio_to_weight
[40] = {
1377 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1378 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1379 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1380 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1381 /* 0 */ 1024, 820, 655, 526, 423,
1382 /* 5 */ 335, 272, 215, 172, 137,
1383 /* 10 */ 110, 87, 70, 56, 45,
1384 /* 15 */ 36, 29, 23, 18, 15,
1388 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1390 * In cases where the weight does not change often, we can use the
1391 * precalculated inverse to speed up arithmetics by turning divisions
1392 * into multiplications:
1394 static const u32 prio_to_wmult
[40] = {
1395 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1396 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1397 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1398 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1399 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1400 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1401 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1402 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1405 /* Time spent by the tasks of the cpu accounting group executing in ... */
1406 enum cpuacct_stat_index
{
1407 CPUACCT_STAT_USER
, /* ... user mode */
1408 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1410 CPUACCT_STAT_NSTATS
,
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1415 static void cpuacct_update_stats(struct task_struct
*tsk
,
1416 enum cpuacct_stat_index idx
, cputime_t val
);
1418 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1419 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1420 enum cpuacct_stat_index idx
, cputime_t val
) {}
1423 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1425 update_load_add(&rq
->load
, load
);
1428 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1430 update_load_sub(&rq
->load
, load
);
1433 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1434 typedef int (*tg_visitor
)(struct task_group
*, void *);
1437 * Iterate the full tree, calling @down when first entering a node and @up when
1438 * leaving it for the final time.
1440 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1442 struct task_group
*parent
, *child
;
1446 parent
= &root_task_group
;
1448 ret
= (*down
)(parent
, data
);
1451 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1458 ret
= (*up
)(parent
, data
);
1463 parent
= parent
->parent
;
1472 static int tg_nop(struct task_group
*tg
, void *data
)
1479 /* Used instead of source_load when we know the type == 0 */
1480 static unsigned long weighted_cpuload(const int cpu
)
1482 return cpu_rq(cpu
)->load
.weight
;
1486 * Return a low guess at the load of a migration-source cpu weighted
1487 * according to the scheduling class and "nice" value.
1489 * We want to under-estimate the load of migration sources, to
1490 * balance conservatively.
1492 static unsigned long source_load(int cpu
, int type
)
1494 struct rq
*rq
= cpu_rq(cpu
);
1495 unsigned long total
= weighted_cpuload(cpu
);
1497 if (type
== 0 || !sched_feat(LB_BIAS
))
1500 return min(rq
->cpu_load
[type
-1], total
);
1504 * Return a high guess at the load of a migration-target cpu weighted
1505 * according to the scheduling class and "nice" value.
1507 static unsigned long target_load(int cpu
, int type
)
1509 struct rq
*rq
= cpu_rq(cpu
);
1510 unsigned long total
= weighted_cpuload(cpu
);
1512 if (type
== 0 || !sched_feat(LB_BIAS
))
1515 return max(rq
->cpu_load
[type
-1], total
);
1518 static unsigned long power_of(int cpu
)
1520 return cpu_rq(cpu
)->cpu_power
;
1523 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1525 static unsigned long cpu_avg_load_per_task(int cpu
)
1527 struct rq
*rq
= cpu_rq(cpu
);
1528 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1531 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1533 rq
->avg_load_per_task
= 0;
1535 return rq
->avg_load_per_task
;
1538 #ifdef CONFIG_FAIR_GROUP_SCHED
1540 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1542 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1545 * Calculate and set the cpu's group shares.
1547 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1548 unsigned long sd_shares
,
1549 unsigned long sd_rq_weight
,
1550 unsigned long *usd_rq_weight
)
1552 unsigned long shares
, rq_weight
;
1555 rq_weight
= usd_rq_weight
[cpu
];
1558 rq_weight
= NICE_0_LOAD
;
1562 * \Sum_j shares_j * rq_weight_i
1563 * shares_i = -----------------------------
1564 * \Sum_j rq_weight_j
1566 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1567 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1569 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1570 sysctl_sched_shares_thresh
) {
1571 struct rq
*rq
= cpu_rq(cpu
);
1572 unsigned long flags
;
1574 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1575 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1576 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1577 __set_se_shares(tg
->se
[cpu
], shares
);
1578 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1583 * Re-compute the task group their per cpu shares over the given domain.
1584 * This needs to be done in a bottom-up fashion because the rq weight of a
1585 * parent group depends on the shares of its child groups.
1587 static int tg_shares_up(struct task_group
*tg
, void *data
)
1589 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1590 unsigned long *usd_rq_weight
;
1591 struct sched_domain
*sd
= data
;
1592 unsigned long flags
;
1598 local_irq_save(flags
);
1599 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1601 for_each_cpu(i
, sched_domain_span(sd
)) {
1602 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1603 usd_rq_weight
[i
] = weight
;
1605 rq_weight
+= weight
;
1607 * If there are currently no tasks on the cpu pretend there
1608 * is one of average load so that when a new task gets to
1609 * run here it will not get delayed by group starvation.
1612 weight
= NICE_0_LOAD
;
1614 sum_weight
+= weight
;
1615 shares
+= tg
->cfs_rq
[i
]->shares
;
1619 rq_weight
= sum_weight
;
1621 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1622 shares
= tg
->shares
;
1624 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1625 shares
= tg
->shares
;
1627 for_each_cpu(i
, sched_domain_span(sd
))
1628 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1630 local_irq_restore(flags
);
1636 * Compute the cpu's hierarchical load factor for each task group.
1637 * This needs to be done in a top-down fashion because the load of a child
1638 * group is a fraction of its parents load.
1640 static int tg_load_down(struct task_group
*tg
, void *data
)
1643 long cpu
= (long)data
;
1646 load
= cpu_rq(cpu
)->load
.weight
;
1648 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1649 load
*= tg
->cfs_rq
[cpu
]->shares
;
1650 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1653 tg
->cfs_rq
[cpu
]->h_load
= load
;
1658 static void update_shares(struct sched_domain
*sd
)
1663 if (root_task_group_empty())
1666 now
= local_clock();
1667 elapsed
= now
- sd
->last_update
;
1669 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1670 sd
->last_update
= now
;
1671 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1675 static void update_h_load(long cpu
)
1677 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1682 static inline void update_shares(struct sched_domain
*sd
)
1688 #ifdef CONFIG_PREEMPT
1690 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1693 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1694 * way at the expense of forcing extra atomic operations in all
1695 * invocations. This assures that the double_lock is acquired using the
1696 * same underlying policy as the spinlock_t on this architecture, which
1697 * reduces latency compared to the unfair variant below. However, it
1698 * also adds more overhead and therefore may reduce throughput.
1700 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1701 __releases(this_rq
->lock
)
1702 __acquires(busiest
->lock
)
1703 __acquires(this_rq
->lock
)
1705 raw_spin_unlock(&this_rq
->lock
);
1706 double_rq_lock(this_rq
, busiest
);
1713 * Unfair double_lock_balance: Optimizes throughput at the expense of
1714 * latency by eliminating extra atomic operations when the locks are
1715 * already in proper order on entry. This favors lower cpu-ids and will
1716 * grant the double lock to lower cpus over higher ids under contention,
1717 * regardless of entry order into the function.
1719 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1720 __releases(this_rq
->lock
)
1721 __acquires(busiest
->lock
)
1722 __acquires(this_rq
->lock
)
1726 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1727 if (busiest
< this_rq
) {
1728 raw_spin_unlock(&this_rq
->lock
);
1729 raw_spin_lock(&busiest
->lock
);
1730 raw_spin_lock_nested(&this_rq
->lock
,
1731 SINGLE_DEPTH_NESTING
);
1734 raw_spin_lock_nested(&busiest
->lock
,
1735 SINGLE_DEPTH_NESTING
);
1740 #endif /* CONFIG_PREEMPT */
1743 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1745 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1747 if (unlikely(!irqs_disabled())) {
1748 /* printk() doesn't work good under rq->lock */
1749 raw_spin_unlock(&this_rq
->lock
);
1753 return _double_lock_balance(this_rq
, busiest
);
1756 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1757 __releases(busiest
->lock
)
1759 raw_spin_unlock(&busiest
->lock
);
1760 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1764 * double_rq_lock - safely lock two runqueues
1766 * Note this does not disable interrupts like task_rq_lock,
1767 * you need to do so manually before calling.
1769 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1770 __acquires(rq1
->lock
)
1771 __acquires(rq2
->lock
)
1773 BUG_ON(!irqs_disabled());
1775 raw_spin_lock(&rq1
->lock
);
1776 __acquire(rq2
->lock
); /* Fake it out ;) */
1779 raw_spin_lock(&rq1
->lock
);
1780 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1782 raw_spin_lock(&rq2
->lock
);
1783 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1789 * double_rq_unlock - safely unlock two runqueues
1791 * Note this does not restore interrupts like task_rq_unlock,
1792 * you need to do so manually after calling.
1794 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1795 __releases(rq1
->lock
)
1796 __releases(rq2
->lock
)
1798 raw_spin_unlock(&rq1
->lock
);
1800 raw_spin_unlock(&rq2
->lock
);
1802 __release(rq2
->lock
);
1807 #ifdef CONFIG_FAIR_GROUP_SCHED
1808 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1811 cfs_rq
->shares
= shares
;
1816 static void calc_load_account_idle(struct rq
*this_rq
);
1817 static void update_sysctl(void);
1818 static int get_update_sysctl_factor(void);
1819 static void update_cpu_load(struct rq
*this_rq
);
1821 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1823 set_task_rq(p
, cpu
);
1826 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1827 * successfuly executed on another CPU. We must ensure that updates of
1828 * per-task data have been completed by this moment.
1831 task_thread_info(p
)->cpu
= cpu
;
1835 static const struct sched_class rt_sched_class
;
1837 #define sched_class_highest (&rt_sched_class)
1838 #define for_each_class(class) \
1839 for (class = sched_class_highest; class; class = class->next)
1841 #include "sched_stats.h"
1843 static void inc_nr_running(struct rq
*rq
)
1848 static void dec_nr_running(struct rq
*rq
)
1853 static void set_load_weight(struct task_struct
*p
)
1855 if (task_has_rt_policy(p
)) {
1856 p
->se
.load
.weight
= 0;
1857 p
->se
.load
.inv_weight
= WMULT_CONST
;
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p
->policy
== SCHED_IDLE
) {
1865 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1866 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1870 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1871 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1874 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1876 update_rq_clock(rq
);
1877 sched_info_queued(p
);
1878 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1882 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1884 update_rq_clock(rq
);
1885 sched_info_dequeued(p
);
1886 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1891 * activate_task - move a task to the runqueue.
1893 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1895 if (task_contributes_to_load(p
))
1896 rq
->nr_uninterruptible
--;
1898 enqueue_task(rq
, p
, flags
);
1903 * deactivate_task - remove a task from the runqueue.
1905 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1907 if (task_contributes_to_load(p
))
1908 rq
->nr_uninterruptible
++;
1910 dequeue_task(rq
, p
, flags
);
1914 #include "sched_idletask.c"
1915 #include "sched_fair.c"
1916 #include "sched_rt.c"
1917 #ifdef CONFIG_SCHED_DEBUG
1918 # include "sched_debug.c"
1922 * __normal_prio - return the priority that is based on the static prio
1924 static inline int __normal_prio(struct task_struct
*p
)
1926 return p
->static_prio
;
1930 * Calculate the expected normal priority: i.e. priority
1931 * without taking RT-inheritance into account. Might be
1932 * boosted by interactivity modifiers. Changes upon fork,
1933 * setprio syscalls, and whenever the interactivity
1934 * estimator recalculates.
1936 static inline int normal_prio(struct task_struct
*p
)
1940 if (task_has_rt_policy(p
))
1941 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1943 prio
= __normal_prio(p
);
1948 * Calculate the current priority, i.e. the priority
1949 * taken into account by the scheduler. This value might
1950 * be boosted by RT tasks, or might be boosted by
1951 * interactivity modifiers. Will be RT if the task got
1952 * RT-boosted. If not then it returns p->normal_prio.
1954 static int effective_prio(struct task_struct
*p
)
1956 p
->normal_prio
= normal_prio(p
);
1958 * If we are RT tasks or we were boosted to RT priority,
1959 * keep the priority unchanged. Otherwise, update priority
1960 * to the normal priority:
1962 if (!rt_prio(p
->prio
))
1963 return p
->normal_prio
;
1968 * task_curr - is this task currently executing on a CPU?
1969 * @p: the task in question.
1971 inline int task_curr(const struct task_struct
*p
)
1973 return cpu_curr(task_cpu(p
)) == p
;
1976 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1977 const struct sched_class
*prev_class
,
1978 int oldprio
, int running
)
1980 if (prev_class
!= p
->sched_class
) {
1981 if (prev_class
->switched_from
)
1982 prev_class
->switched_from(rq
, p
, running
);
1983 p
->sched_class
->switched_to(rq
, p
, running
);
1985 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1990 * Is this task likely cache-hot:
1993 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1997 if (p
->sched_class
!= &fair_sched_class
)
2001 * Buddy candidates are cache hot:
2003 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2004 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2005 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2008 if (sysctl_sched_migration_cost
== -1)
2010 if (sysctl_sched_migration_cost
== 0)
2013 delta
= now
- p
->se
.exec_start
;
2015 return delta
< (s64
)sysctl_sched_migration_cost
;
2018 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2020 #ifdef CONFIG_SCHED_DEBUG
2022 * We should never call set_task_cpu() on a blocked task,
2023 * ttwu() will sort out the placement.
2025 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2026 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2029 trace_sched_migrate_task(p
, new_cpu
);
2031 if (task_cpu(p
) != new_cpu
) {
2032 p
->se
.nr_migrations
++;
2033 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2036 __set_task_cpu(p
, new_cpu
);
2039 struct migration_arg
{
2040 struct task_struct
*task
;
2044 static int migration_cpu_stop(void *data
);
2047 * The task's runqueue lock must be held.
2048 * Returns true if you have to wait for migration thread.
2050 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2052 struct rq
*rq
= task_rq(p
);
2055 * If the task is not on a runqueue (and not running), then
2056 * the next wake-up will properly place the task.
2058 return p
->se
.on_rq
|| task_running(rq
, p
);
2062 * wait_task_inactive - wait for a thread to unschedule.
2064 * If @match_state is nonzero, it's the @p->state value just checked and
2065 * not expected to change. If it changes, i.e. @p might have woken up,
2066 * then return zero. When we succeed in waiting for @p to be off its CPU,
2067 * we return a positive number (its total switch count). If a second call
2068 * a short while later returns the same number, the caller can be sure that
2069 * @p has remained unscheduled the whole time.
2071 * The caller must ensure that the task *will* unschedule sometime soon,
2072 * else this function might spin for a *long* time. This function can't
2073 * be called with interrupts off, or it may introduce deadlock with
2074 * smp_call_function() if an IPI is sent by the same process we are
2075 * waiting to become inactive.
2077 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2079 unsigned long flags
;
2086 * We do the initial early heuristics without holding
2087 * any task-queue locks at all. We'll only try to get
2088 * the runqueue lock when things look like they will
2094 * If the task is actively running on another CPU
2095 * still, just relax and busy-wait without holding
2098 * NOTE! Since we don't hold any locks, it's not
2099 * even sure that "rq" stays as the right runqueue!
2100 * But we don't care, since "task_running()" will
2101 * return false if the runqueue has changed and p
2102 * is actually now running somewhere else!
2104 while (task_running(rq
, p
)) {
2105 if (match_state
&& unlikely(p
->state
!= match_state
))
2111 * Ok, time to look more closely! We need the rq
2112 * lock now, to be *sure*. If we're wrong, we'll
2113 * just go back and repeat.
2115 rq
= task_rq_lock(p
, &flags
);
2116 trace_sched_wait_task(p
);
2117 running
= task_running(rq
, p
);
2118 on_rq
= p
->se
.on_rq
;
2120 if (!match_state
|| p
->state
== match_state
)
2121 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2122 task_rq_unlock(rq
, &flags
);
2125 * If it changed from the expected state, bail out now.
2127 if (unlikely(!ncsw
))
2131 * Was it really running after all now that we
2132 * checked with the proper locks actually held?
2134 * Oops. Go back and try again..
2136 if (unlikely(running
)) {
2142 * It's not enough that it's not actively running,
2143 * it must be off the runqueue _entirely_, and not
2146 * So if it was still runnable (but just not actively
2147 * running right now), it's preempted, and we should
2148 * yield - it could be a while.
2150 if (unlikely(on_rq
)) {
2151 schedule_timeout_uninterruptible(1);
2156 * Ahh, all good. It wasn't running, and it wasn't
2157 * runnable, which means that it will never become
2158 * running in the future either. We're all done!
2167 * kick_process - kick a running thread to enter/exit the kernel
2168 * @p: the to-be-kicked thread
2170 * Cause a process which is running on another CPU to enter
2171 * kernel-mode, without any delay. (to get signals handled.)
2173 * NOTE: this function doesnt have to take the runqueue lock,
2174 * because all it wants to ensure is that the remote task enters
2175 * the kernel. If the IPI races and the task has been migrated
2176 * to another CPU then no harm is done and the purpose has been
2179 void kick_process(struct task_struct
*p
)
2185 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2186 smp_send_reschedule(cpu
);
2189 EXPORT_SYMBOL_GPL(kick_process
);
2190 #endif /* CONFIG_SMP */
2193 * task_oncpu_function_call - call a function on the cpu on which a task runs
2194 * @p: the task to evaluate
2195 * @func: the function to be called
2196 * @info: the function call argument
2198 * Calls the function @func when the task is currently running. This might
2199 * be on the current CPU, which just calls the function directly
2201 void task_oncpu_function_call(struct task_struct
*p
,
2202 void (*func
) (void *info
), void *info
)
2209 smp_call_function_single(cpu
, func
, info
, 1);
2215 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2217 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2220 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2222 /* Look for allowed, online CPU in same node. */
2223 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2224 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2227 /* Any allowed, online CPU? */
2228 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2229 if (dest_cpu
< nr_cpu_ids
)
2232 /* No more Mr. Nice Guy. */
2233 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2234 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2236 * Don't tell them about moving exiting tasks or
2237 * kernel threads (both mm NULL), since they never
2240 if (p
->mm
&& printk_ratelimit()) {
2241 printk(KERN_INFO
"process %d (%s) no "
2242 "longer affine to cpu%d\n",
2243 task_pid_nr(p
), p
->comm
, cpu
);
2251 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2254 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2256 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2259 * In order not to call set_task_cpu() on a blocking task we need
2260 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2263 * Since this is common to all placement strategies, this lives here.
2265 * [ this allows ->select_task() to simply return task_cpu(p) and
2266 * not worry about this generic constraint ]
2268 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2270 cpu
= select_fallback_rq(task_cpu(p
), p
);
2275 static void update_avg(u64
*avg
, u64 sample
)
2277 s64 diff
= sample
- *avg
;
2282 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2283 bool is_sync
, bool is_migrate
, bool is_local
,
2284 unsigned long en_flags
)
2286 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2288 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2290 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2292 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2294 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2296 activate_task(rq
, p
, en_flags
);
2299 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2300 int wake_flags
, bool success
)
2302 trace_sched_wakeup(p
, success
);
2303 check_preempt_curr(rq
, p
, wake_flags
);
2305 p
->state
= TASK_RUNNING
;
2307 if (p
->sched_class
->task_woken
)
2308 p
->sched_class
->task_woken(rq
, p
);
2310 if (unlikely(rq
->idle_stamp
)) {
2311 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2312 u64 max
= 2*sysctl_sched_migration_cost
;
2317 update_avg(&rq
->avg_idle
, delta
);
2321 /* if a worker is waking up, notify workqueue */
2322 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2323 wq_worker_waking_up(p
, cpu_of(rq
));
2327 * try_to_wake_up - wake up a thread
2328 * @p: the thread to be awakened
2329 * @state: the mask of task states that can be woken
2330 * @wake_flags: wake modifier flags (WF_*)
2332 * Put it on the run-queue if it's not already there. The "current"
2333 * thread is always on the run-queue (except when the actual
2334 * re-schedule is in progress), and as such you're allowed to do
2335 * the simpler "current->state = TASK_RUNNING" to mark yourself
2336 * runnable without the overhead of this.
2338 * Returns %true if @p was woken up, %false if it was already running
2339 * or @state didn't match @p's state.
2341 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2344 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2345 unsigned long flags
;
2346 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2349 this_cpu
= get_cpu();
2352 rq
= task_rq_lock(p
, &flags
);
2353 if (!(p
->state
& state
))
2363 if (unlikely(task_running(rq
, p
)))
2367 * In order to handle concurrent wakeups and release the rq->lock
2368 * we put the task in TASK_WAKING state.
2370 * First fix up the nr_uninterruptible count:
2372 if (task_contributes_to_load(p
)) {
2373 if (likely(cpu_online(orig_cpu
)))
2374 rq
->nr_uninterruptible
--;
2376 this_rq()->nr_uninterruptible
--;
2378 p
->state
= TASK_WAKING
;
2380 if (p
->sched_class
->task_waking
) {
2381 p
->sched_class
->task_waking(rq
, p
);
2382 en_flags
|= ENQUEUE_WAKING
;
2385 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2386 if (cpu
!= orig_cpu
)
2387 set_task_cpu(p
, cpu
);
2388 __task_rq_unlock(rq
);
2391 raw_spin_lock(&rq
->lock
);
2394 * We migrated the task without holding either rq->lock, however
2395 * since the task is not on the task list itself, nobody else
2396 * will try and migrate the task, hence the rq should match the
2397 * cpu we just moved it to.
2399 WARN_ON(task_cpu(p
) != cpu
);
2400 WARN_ON(p
->state
!= TASK_WAKING
);
2402 #ifdef CONFIG_SCHEDSTATS
2403 schedstat_inc(rq
, ttwu_count
);
2404 if (cpu
== this_cpu
)
2405 schedstat_inc(rq
, ttwu_local
);
2407 struct sched_domain
*sd
;
2408 for_each_domain(this_cpu
, sd
) {
2409 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2410 schedstat_inc(sd
, ttwu_wake_remote
);
2415 #endif /* CONFIG_SCHEDSTATS */
2418 #endif /* CONFIG_SMP */
2419 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2420 cpu
== this_cpu
, en_flags
);
2423 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2425 task_rq_unlock(rq
, &flags
);
2432 * try_to_wake_up_local - try to wake up a local task with rq lock held
2433 * @p: the thread to be awakened
2435 * Put @p on the run-queue if it's not alredy there. The caller must
2436 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2437 * the current task. this_rq() stays locked over invocation.
2439 static void try_to_wake_up_local(struct task_struct
*p
)
2441 struct rq
*rq
= task_rq(p
);
2442 bool success
= false;
2444 BUG_ON(rq
!= this_rq());
2445 BUG_ON(p
== current
);
2446 lockdep_assert_held(&rq
->lock
);
2448 if (!(p
->state
& TASK_NORMAL
))
2452 if (likely(!task_running(rq
, p
))) {
2453 schedstat_inc(rq
, ttwu_count
);
2454 schedstat_inc(rq
, ttwu_local
);
2456 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2459 ttwu_post_activation(p
, rq
, 0, success
);
2463 * wake_up_process - Wake up a specific process
2464 * @p: The process to be woken up.
2466 * Attempt to wake up the nominated process and move it to the set of runnable
2467 * processes. Returns 1 if the process was woken up, 0 if it was already
2470 * It may be assumed that this function implies a write memory barrier before
2471 * changing the task state if and only if any tasks are woken up.
2473 int wake_up_process(struct task_struct
*p
)
2475 return try_to_wake_up(p
, TASK_ALL
, 0);
2477 EXPORT_SYMBOL(wake_up_process
);
2479 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2481 return try_to_wake_up(p
, state
, 0);
2485 * Perform scheduler related setup for a newly forked process p.
2486 * p is forked by current.
2488 * __sched_fork() is basic setup used by init_idle() too:
2490 static void __sched_fork(struct task_struct
*p
)
2492 p
->se
.exec_start
= 0;
2493 p
->se
.sum_exec_runtime
= 0;
2494 p
->se
.prev_sum_exec_runtime
= 0;
2495 p
->se
.nr_migrations
= 0;
2497 #ifdef CONFIG_SCHEDSTATS
2498 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2501 INIT_LIST_HEAD(&p
->rt
.run_list
);
2503 INIT_LIST_HEAD(&p
->se
.group_node
);
2505 #ifdef CONFIG_PREEMPT_NOTIFIERS
2506 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2511 * fork()/clone()-time setup:
2513 void sched_fork(struct task_struct
*p
, int clone_flags
)
2515 int cpu
= get_cpu();
2519 * We mark the process as running here. This guarantees that
2520 * nobody will actually run it, and a signal or other external
2521 * event cannot wake it up and insert it on the runqueue either.
2523 p
->state
= TASK_RUNNING
;
2526 * Revert to default priority/policy on fork if requested.
2528 if (unlikely(p
->sched_reset_on_fork
)) {
2529 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2530 p
->policy
= SCHED_NORMAL
;
2531 p
->normal_prio
= p
->static_prio
;
2534 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2535 p
->static_prio
= NICE_TO_PRIO(0);
2536 p
->normal_prio
= p
->static_prio
;
2541 * We don't need the reset flag anymore after the fork. It has
2542 * fulfilled its duty:
2544 p
->sched_reset_on_fork
= 0;
2548 * Make sure we do not leak PI boosting priority to the child.
2550 p
->prio
= current
->normal_prio
;
2552 if (!rt_prio(p
->prio
))
2553 p
->sched_class
= &fair_sched_class
;
2555 if (p
->sched_class
->task_fork
)
2556 p
->sched_class
->task_fork(p
);
2559 * The child is not yet in the pid-hash so no cgroup attach races,
2560 * and the cgroup is pinned to this child due to cgroup_fork()
2561 * is ran before sched_fork().
2563 * Silence PROVE_RCU.
2566 set_task_cpu(p
, cpu
);
2569 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2570 if (likely(sched_info_on()))
2571 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2573 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2576 #ifdef CONFIG_PREEMPT
2577 /* Want to start with kernel preemption disabled. */
2578 task_thread_info(p
)->preempt_count
= 1;
2580 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2586 * wake_up_new_task - wake up a newly created task for the first time.
2588 * This function will do some initial scheduler statistics housekeeping
2589 * that must be done for every newly created context, then puts the task
2590 * on the runqueue and wakes it.
2592 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2594 unsigned long flags
;
2596 int cpu __maybe_unused
= get_cpu();
2599 rq
= task_rq_lock(p
, &flags
);
2600 p
->state
= TASK_WAKING
;
2603 * Fork balancing, do it here and not earlier because:
2604 * - cpus_allowed can change in the fork path
2605 * - any previously selected cpu might disappear through hotplug
2607 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2608 * without people poking at ->cpus_allowed.
2610 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2611 set_task_cpu(p
, cpu
);
2613 p
->state
= TASK_RUNNING
;
2614 task_rq_unlock(rq
, &flags
);
2617 rq
= task_rq_lock(p
, &flags
);
2618 activate_task(rq
, p
, 0);
2619 trace_sched_wakeup_new(p
, 1);
2620 check_preempt_curr(rq
, p
, WF_FORK
);
2622 if (p
->sched_class
->task_woken
)
2623 p
->sched_class
->task_woken(rq
, p
);
2625 task_rq_unlock(rq
, &flags
);
2629 #ifdef CONFIG_PREEMPT_NOTIFIERS
2632 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2633 * @notifier: notifier struct to register
2635 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2637 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2639 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2642 * preempt_notifier_unregister - no longer interested in preemption notifications
2643 * @notifier: notifier struct to unregister
2645 * This is safe to call from within a preemption notifier.
2647 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2649 hlist_del(¬ifier
->link
);
2651 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2653 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2655 struct preempt_notifier
*notifier
;
2656 struct hlist_node
*node
;
2658 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2659 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2663 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2664 struct task_struct
*next
)
2666 struct preempt_notifier
*notifier
;
2667 struct hlist_node
*node
;
2669 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2670 notifier
->ops
->sched_out(notifier
, next
);
2673 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2675 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2680 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2681 struct task_struct
*next
)
2685 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2688 * prepare_task_switch - prepare to switch tasks
2689 * @rq: the runqueue preparing to switch
2690 * @prev: the current task that is being switched out
2691 * @next: the task we are going to switch to.
2693 * This is called with the rq lock held and interrupts off. It must
2694 * be paired with a subsequent finish_task_switch after the context
2697 * prepare_task_switch sets up locking and calls architecture specific
2701 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2702 struct task_struct
*next
)
2704 fire_sched_out_preempt_notifiers(prev
, next
);
2705 prepare_lock_switch(rq
, next
);
2706 prepare_arch_switch(next
);
2710 * finish_task_switch - clean up after a task-switch
2711 * @rq: runqueue associated with task-switch
2712 * @prev: the thread we just switched away from.
2714 * finish_task_switch must be called after the context switch, paired
2715 * with a prepare_task_switch call before the context switch.
2716 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2717 * and do any other architecture-specific cleanup actions.
2719 * Note that we may have delayed dropping an mm in context_switch(). If
2720 * so, we finish that here outside of the runqueue lock. (Doing it
2721 * with the lock held can cause deadlocks; see schedule() for
2724 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2725 __releases(rq
->lock
)
2727 struct mm_struct
*mm
= rq
->prev_mm
;
2733 * A task struct has one reference for the use as "current".
2734 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2735 * schedule one last time. The schedule call will never return, and
2736 * the scheduled task must drop that reference.
2737 * The test for TASK_DEAD must occur while the runqueue locks are
2738 * still held, otherwise prev could be scheduled on another cpu, die
2739 * there before we look at prev->state, and then the reference would
2741 * Manfred Spraul <manfred@colorfullife.com>
2743 prev_state
= prev
->state
;
2744 finish_arch_switch(prev
);
2745 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2746 local_irq_disable();
2747 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2748 perf_event_task_sched_in(current
);
2749 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2751 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2752 finish_lock_switch(rq
, prev
);
2754 fire_sched_in_preempt_notifiers(current
);
2757 if (unlikely(prev_state
== TASK_DEAD
)) {
2759 * Remove function-return probe instances associated with this
2760 * task and put them back on the free list.
2762 kprobe_flush_task(prev
);
2763 put_task_struct(prev
);
2769 /* assumes rq->lock is held */
2770 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2772 if (prev
->sched_class
->pre_schedule
)
2773 prev
->sched_class
->pre_schedule(rq
, prev
);
2776 /* rq->lock is NOT held, but preemption is disabled */
2777 static inline void post_schedule(struct rq
*rq
)
2779 if (rq
->post_schedule
) {
2780 unsigned long flags
;
2782 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2783 if (rq
->curr
->sched_class
->post_schedule
)
2784 rq
->curr
->sched_class
->post_schedule(rq
);
2785 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2787 rq
->post_schedule
= 0;
2793 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2797 static inline void post_schedule(struct rq
*rq
)
2804 * schedule_tail - first thing a freshly forked thread must call.
2805 * @prev: the thread we just switched away from.
2807 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2808 __releases(rq
->lock
)
2810 struct rq
*rq
= this_rq();
2812 finish_task_switch(rq
, prev
);
2815 * FIXME: do we need to worry about rq being invalidated by the
2820 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2821 /* In this case, finish_task_switch does not reenable preemption */
2824 if (current
->set_child_tid
)
2825 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2829 * context_switch - switch to the new MM and the new
2830 * thread's register state.
2833 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2834 struct task_struct
*next
)
2836 struct mm_struct
*mm
, *oldmm
;
2838 prepare_task_switch(rq
, prev
, next
);
2839 trace_sched_switch(prev
, next
);
2841 oldmm
= prev
->active_mm
;
2843 * For paravirt, this is coupled with an exit in switch_to to
2844 * combine the page table reload and the switch backend into
2847 arch_start_context_switch(prev
);
2850 next
->active_mm
= oldmm
;
2851 atomic_inc(&oldmm
->mm_count
);
2852 enter_lazy_tlb(oldmm
, next
);
2854 switch_mm(oldmm
, mm
, next
);
2856 if (likely(!prev
->mm
)) {
2857 prev
->active_mm
= NULL
;
2858 rq
->prev_mm
= oldmm
;
2861 * Since the runqueue lock will be released by the next
2862 * task (which is an invalid locking op but in the case
2863 * of the scheduler it's an obvious special-case), so we
2864 * do an early lockdep release here:
2866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2867 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2870 /* Here we just switch the register state and the stack. */
2871 switch_to(prev
, next
, prev
);
2875 * this_rq must be evaluated again because prev may have moved
2876 * CPUs since it called schedule(), thus the 'rq' on its stack
2877 * frame will be invalid.
2879 finish_task_switch(this_rq(), prev
);
2883 * nr_running, nr_uninterruptible and nr_context_switches:
2885 * externally visible scheduler statistics: current number of runnable
2886 * threads, current number of uninterruptible-sleeping threads, total
2887 * number of context switches performed since bootup.
2889 unsigned long nr_running(void)
2891 unsigned long i
, sum
= 0;
2893 for_each_online_cpu(i
)
2894 sum
+= cpu_rq(i
)->nr_running
;
2899 unsigned long nr_uninterruptible(void)
2901 unsigned long i
, sum
= 0;
2903 for_each_possible_cpu(i
)
2904 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2907 * Since we read the counters lockless, it might be slightly
2908 * inaccurate. Do not allow it to go below zero though:
2910 if (unlikely((long)sum
< 0))
2916 unsigned long long nr_context_switches(void)
2919 unsigned long long sum
= 0;
2921 for_each_possible_cpu(i
)
2922 sum
+= cpu_rq(i
)->nr_switches
;
2927 unsigned long nr_iowait(void)
2929 unsigned long i
, sum
= 0;
2931 for_each_possible_cpu(i
)
2932 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2937 unsigned long nr_iowait_cpu(int cpu
)
2939 struct rq
*this = cpu_rq(cpu
);
2940 return atomic_read(&this->nr_iowait
);
2943 unsigned long this_cpu_load(void)
2945 struct rq
*this = this_rq();
2946 return this->cpu_load
[0];
2950 /* Variables and functions for calc_load */
2951 static atomic_long_t calc_load_tasks
;
2952 static unsigned long calc_load_update
;
2953 unsigned long avenrun
[3];
2954 EXPORT_SYMBOL(avenrun
);
2956 static long calc_load_fold_active(struct rq
*this_rq
)
2958 long nr_active
, delta
= 0;
2960 nr_active
= this_rq
->nr_running
;
2961 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2963 if (nr_active
!= this_rq
->calc_load_active
) {
2964 delta
= nr_active
- this_rq
->calc_load_active
;
2965 this_rq
->calc_load_active
= nr_active
;
2973 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2975 * When making the ILB scale, we should try to pull this in as well.
2977 static atomic_long_t calc_load_tasks_idle
;
2979 static void calc_load_account_idle(struct rq
*this_rq
)
2983 delta
= calc_load_fold_active(this_rq
);
2985 atomic_long_add(delta
, &calc_load_tasks_idle
);
2988 static long calc_load_fold_idle(void)
2993 * Its got a race, we don't care...
2995 if (atomic_long_read(&calc_load_tasks_idle
))
2996 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3001 static void calc_load_account_idle(struct rq
*this_rq
)
3005 static inline long calc_load_fold_idle(void)
3012 * get_avenrun - get the load average array
3013 * @loads: pointer to dest load array
3014 * @offset: offset to add
3015 * @shift: shift count to shift the result left
3017 * These values are estimates at best, so no need for locking.
3019 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3021 loads
[0] = (avenrun
[0] + offset
) << shift
;
3022 loads
[1] = (avenrun
[1] + offset
) << shift
;
3023 loads
[2] = (avenrun
[2] + offset
) << shift
;
3026 static unsigned long
3027 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3030 load
+= active
* (FIXED_1
- exp
);
3031 return load
>> FSHIFT
;
3035 * calc_load - update the avenrun load estimates 10 ticks after the
3036 * CPUs have updated calc_load_tasks.
3038 void calc_global_load(void)
3040 unsigned long upd
= calc_load_update
+ 10;
3043 if (time_before(jiffies
, upd
))
3046 active
= atomic_long_read(&calc_load_tasks
);
3047 active
= active
> 0 ? active
* FIXED_1
: 0;
3049 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3050 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3051 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3053 calc_load_update
+= LOAD_FREQ
;
3057 * Called from update_cpu_load() to periodically update this CPU's
3060 static void calc_load_account_active(struct rq
*this_rq
)
3064 if (time_before(jiffies
, this_rq
->calc_load_update
))
3067 delta
= calc_load_fold_active(this_rq
);
3068 delta
+= calc_load_fold_idle();
3070 atomic_long_add(delta
, &calc_load_tasks
);
3072 this_rq
->calc_load_update
+= LOAD_FREQ
;
3076 * The exact cpuload at various idx values, calculated at every tick would be
3077 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3079 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3080 * on nth tick when cpu may be busy, then we have:
3081 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3082 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3084 * decay_load_missed() below does efficient calculation of
3085 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3086 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3088 * The calculation is approximated on a 128 point scale.
3089 * degrade_zero_ticks is the number of ticks after which load at any
3090 * particular idx is approximated to be zero.
3091 * degrade_factor is a precomputed table, a row for each load idx.
3092 * Each column corresponds to degradation factor for a power of two ticks,
3093 * based on 128 point scale.
3095 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3096 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3098 * With this power of 2 load factors, we can degrade the load n times
3099 * by looking at 1 bits in n and doing as many mult/shift instead of
3100 * n mult/shifts needed by the exact degradation.
3102 #define DEGRADE_SHIFT 7
3103 static const unsigned char
3104 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3105 static const unsigned char
3106 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3107 {0, 0, 0, 0, 0, 0, 0, 0},
3108 {64, 32, 8, 0, 0, 0, 0, 0},
3109 {96, 72, 40, 12, 1, 0, 0},
3110 {112, 98, 75, 43, 15, 1, 0},
3111 {120, 112, 98, 76, 45, 16, 2} };
3114 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3115 * would be when CPU is idle and so we just decay the old load without
3116 * adding any new load.
3118 static unsigned long
3119 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3123 if (!missed_updates
)
3126 if (missed_updates
>= degrade_zero_ticks
[idx
])
3130 return load
>> missed_updates
;
3132 while (missed_updates
) {
3133 if (missed_updates
% 2)
3134 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3136 missed_updates
>>= 1;
3143 * Update rq->cpu_load[] statistics. This function is usually called every
3144 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3145 * every tick. We fix it up based on jiffies.
3147 static void update_cpu_load(struct rq
*this_rq
)
3149 unsigned long this_load
= this_rq
->load
.weight
;
3150 unsigned long curr_jiffies
= jiffies
;
3151 unsigned long pending_updates
;
3154 this_rq
->nr_load_updates
++;
3156 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3157 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3160 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3161 this_rq
->last_load_update_tick
= curr_jiffies
;
3163 /* Update our load: */
3164 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3165 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3166 unsigned long old_load
, new_load
;
3168 /* scale is effectively 1 << i now, and >> i divides by scale */
3170 old_load
= this_rq
->cpu_load
[i
];
3171 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3172 new_load
= this_load
;
3174 * Round up the averaging division if load is increasing. This
3175 * prevents us from getting stuck on 9 if the load is 10, for
3178 if (new_load
> old_load
)
3179 new_load
+= scale
- 1;
3181 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3185 static void update_cpu_load_active(struct rq
*this_rq
)
3187 update_cpu_load(this_rq
);
3189 calc_load_account_active(this_rq
);
3195 * sched_exec - execve() is a valuable balancing opportunity, because at
3196 * this point the task has the smallest effective memory and cache footprint.
3198 void sched_exec(void)
3200 struct task_struct
*p
= current
;
3201 unsigned long flags
;
3205 rq
= task_rq_lock(p
, &flags
);
3206 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3207 if (dest_cpu
== smp_processor_id())
3211 * select_task_rq() can race against ->cpus_allowed
3213 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3214 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3215 struct migration_arg arg
= { p
, dest_cpu
};
3217 task_rq_unlock(rq
, &flags
);
3218 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3222 task_rq_unlock(rq
, &flags
);
3227 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3229 EXPORT_PER_CPU_SYMBOL(kstat
);
3232 * Return any ns on the sched_clock that have not yet been accounted in
3233 * @p in case that task is currently running.
3235 * Called with task_rq_lock() held on @rq.
3237 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3241 if (task_current(rq
, p
)) {
3242 update_rq_clock(rq
);
3243 ns
= rq
->clock
- p
->se
.exec_start
;
3251 unsigned long long task_delta_exec(struct task_struct
*p
)
3253 unsigned long flags
;
3257 rq
= task_rq_lock(p
, &flags
);
3258 ns
= do_task_delta_exec(p
, rq
);
3259 task_rq_unlock(rq
, &flags
);
3265 * Return accounted runtime for the task.
3266 * In case the task is currently running, return the runtime plus current's
3267 * pending runtime that have not been accounted yet.
3269 unsigned long long task_sched_runtime(struct task_struct
*p
)
3271 unsigned long flags
;
3275 rq
= task_rq_lock(p
, &flags
);
3276 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3277 task_rq_unlock(rq
, &flags
);
3283 * Return sum_exec_runtime for the thread group.
3284 * In case the task is currently running, return the sum plus current's
3285 * pending runtime that have not been accounted yet.
3287 * Note that the thread group might have other running tasks as well,
3288 * so the return value not includes other pending runtime that other
3289 * running tasks might have.
3291 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3293 struct task_cputime totals
;
3294 unsigned long flags
;
3298 rq
= task_rq_lock(p
, &flags
);
3299 thread_group_cputime(p
, &totals
);
3300 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3301 task_rq_unlock(rq
, &flags
);
3307 * Account user cpu time to a process.
3308 * @p: the process that the cpu time gets accounted to
3309 * @cputime: the cpu time spent in user space since the last update
3310 * @cputime_scaled: cputime scaled by cpu frequency
3312 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3313 cputime_t cputime_scaled
)
3315 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3318 /* Add user time to process. */
3319 p
->utime
= cputime_add(p
->utime
, cputime
);
3320 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3321 account_group_user_time(p
, cputime
);
3323 /* Add user time to cpustat. */
3324 tmp
= cputime_to_cputime64(cputime
);
3325 if (TASK_NICE(p
) > 0)
3326 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3328 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3330 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3331 /* Account for user time used */
3332 acct_update_integrals(p
);
3336 * Account guest cpu time to a process.
3337 * @p: the process that the cpu time gets accounted to
3338 * @cputime: the cpu time spent in virtual machine since the last update
3339 * @cputime_scaled: cputime scaled by cpu frequency
3341 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3342 cputime_t cputime_scaled
)
3345 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3347 tmp
= cputime_to_cputime64(cputime
);
3349 /* Add guest time to process. */
3350 p
->utime
= cputime_add(p
->utime
, cputime
);
3351 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3352 account_group_user_time(p
, cputime
);
3353 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3355 /* Add guest time to cpustat. */
3356 if (TASK_NICE(p
) > 0) {
3357 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3358 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3360 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3361 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3366 * Account system cpu time to a process.
3367 * @p: the process that the cpu time gets accounted to
3368 * @hardirq_offset: the offset to subtract from hardirq_count()
3369 * @cputime: the cpu time spent in kernel space since the last update
3370 * @cputime_scaled: cputime scaled by cpu frequency
3372 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3373 cputime_t cputime
, cputime_t cputime_scaled
)
3375 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3378 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3379 account_guest_time(p
, cputime
, cputime_scaled
);
3383 /* Add system time to process. */
3384 p
->stime
= cputime_add(p
->stime
, cputime
);
3385 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3386 account_group_system_time(p
, cputime
);
3388 /* Add system time to cpustat. */
3389 tmp
= cputime_to_cputime64(cputime
);
3390 if (hardirq_count() - hardirq_offset
)
3391 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3392 else if (softirq_count())
3393 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3395 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3397 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3399 /* Account for system time used */
3400 acct_update_integrals(p
);
3404 * Account for involuntary wait time.
3405 * @steal: the cpu time spent in involuntary wait
3407 void account_steal_time(cputime_t cputime
)
3409 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3410 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3412 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3416 * Account for idle time.
3417 * @cputime: the cpu time spent in idle wait
3419 void account_idle_time(cputime_t cputime
)
3421 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3422 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3423 struct rq
*rq
= this_rq();
3425 if (atomic_read(&rq
->nr_iowait
) > 0)
3426 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3428 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3431 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3434 * Account a single tick of cpu time.
3435 * @p: the process that the cpu time gets accounted to
3436 * @user_tick: indicates if the tick is a user or a system tick
3438 void account_process_tick(struct task_struct
*p
, int user_tick
)
3440 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3441 struct rq
*rq
= this_rq();
3444 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3445 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3446 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3449 account_idle_time(cputime_one_jiffy
);
3453 * Account multiple ticks of steal time.
3454 * @p: the process from which the cpu time has been stolen
3455 * @ticks: number of stolen ticks
3457 void account_steal_ticks(unsigned long ticks
)
3459 account_steal_time(jiffies_to_cputime(ticks
));
3463 * Account multiple ticks of idle time.
3464 * @ticks: number of stolen ticks
3466 void account_idle_ticks(unsigned long ticks
)
3468 account_idle_time(jiffies_to_cputime(ticks
));
3474 * Use precise platform statistics if available:
3476 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3477 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3483 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3485 struct task_cputime cputime
;
3487 thread_group_cputime(p
, &cputime
);
3489 *ut
= cputime
.utime
;
3490 *st
= cputime
.stime
;
3494 #ifndef nsecs_to_cputime
3495 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3498 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3500 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3503 * Use CFS's precise accounting:
3505 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3510 temp
= (u64
)(rtime
* utime
);
3511 do_div(temp
, total
);
3512 utime
= (cputime_t
)temp
;
3517 * Compare with previous values, to keep monotonicity:
3519 p
->prev_utime
= max(p
->prev_utime
, utime
);
3520 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3522 *ut
= p
->prev_utime
;
3523 *st
= p
->prev_stime
;
3527 * Must be called with siglock held.
3529 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3531 struct signal_struct
*sig
= p
->signal
;
3532 struct task_cputime cputime
;
3533 cputime_t rtime
, utime
, total
;
3535 thread_group_cputime(p
, &cputime
);
3537 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3538 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3543 temp
= (u64
)(rtime
* cputime
.utime
);
3544 do_div(temp
, total
);
3545 utime
= (cputime_t
)temp
;
3549 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3550 sig
->prev_stime
= max(sig
->prev_stime
,
3551 cputime_sub(rtime
, sig
->prev_utime
));
3553 *ut
= sig
->prev_utime
;
3554 *st
= sig
->prev_stime
;
3559 * This function gets called by the timer code, with HZ frequency.
3560 * We call it with interrupts disabled.
3562 * It also gets called by the fork code, when changing the parent's
3565 void scheduler_tick(void)
3567 int cpu
= smp_processor_id();
3568 struct rq
*rq
= cpu_rq(cpu
);
3569 struct task_struct
*curr
= rq
->curr
;
3573 raw_spin_lock(&rq
->lock
);
3574 update_rq_clock(rq
);
3575 update_cpu_load_active(rq
);
3576 curr
->sched_class
->task_tick(rq
, curr
, 0);
3577 raw_spin_unlock(&rq
->lock
);
3579 perf_event_task_tick(curr
);
3582 rq
->idle_at_tick
= idle_cpu(cpu
);
3583 trigger_load_balance(rq
, cpu
);
3587 notrace
unsigned long get_parent_ip(unsigned long addr
)
3589 if (in_lock_functions(addr
)) {
3590 addr
= CALLER_ADDR2
;
3591 if (in_lock_functions(addr
))
3592 addr
= CALLER_ADDR3
;
3597 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3598 defined(CONFIG_PREEMPT_TRACER))
3600 void __kprobes
add_preempt_count(int val
)
3602 #ifdef CONFIG_DEBUG_PREEMPT
3606 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3609 preempt_count() += val
;
3610 #ifdef CONFIG_DEBUG_PREEMPT
3612 * Spinlock count overflowing soon?
3614 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3617 if (preempt_count() == val
)
3618 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3620 EXPORT_SYMBOL(add_preempt_count
);
3622 void __kprobes
sub_preempt_count(int val
)
3624 #ifdef CONFIG_DEBUG_PREEMPT
3628 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3631 * Is the spinlock portion underflowing?
3633 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3634 !(preempt_count() & PREEMPT_MASK
)))
3638 if (preempt_count() == val
)
3639 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3640 preempt_count() -= val
;
3642 EXPORT_SYMBOL(sub_preempt_count
);
3647 * Print scheduling while atomic bug:
3649 static noinline
void __schedule_bug(struct task_struct
*prev
)
3651 struct pt_regs
*regs
= get_irq_regs();
3653 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3654 prev
->comm
, prev
->pid
, preempt_count());
3656 debug_show_held_locks(prev
);
3658 if (irqs_disabled())
3659 print_irqtrace_events(prev
);
3668 * Various schedule()-time debugging checks and statistics:
3670 static inline void schedule_debug(struct task_struct
*prev
)
3673 * Test if we are atomic. Since do_exit() needs to call into
3674 * schedule() atomically, we ignore that path for now.
3675 * Otherwise, whine if we are scheduling when we should not be.
3677 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3678 __schedule_bug(prev
);
3680 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3682 schedstat_inc(this_rq(), sched_count
);
3683 #ifdef CONFIG_SCHEDSTATS
3684 if (unlikely(prev
->lock_depth
>= 0)) {
3685 schedstat_inc(this_rq(), bkl_count
);
3686 schedstat_inc(prev
, sched_info
.bkl_count
);
3691 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3694 update_rq_clock(rq
);
3695 rq
->skip_clock_update
= 0;
3696 prev
->sched_class
->put_prev_task(rq
, prev
);
3700 * Pick up the highest-prio task:
3702 static inline struct task_struct
*
3703 pick_next_task(struct rq
*rq
)
3705 const struct sched_class
*class;
3706 struct task_struct
*p
;
3709 * Optimization: we know that if all tasks are in
3710 * the fair class we can call that function directly:
3712 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3713 p
= fair_sched_class
.pick_next_task(rq
);
3718 class = sched_class_highest
;
3720 p
= class->pick_next_task(rq
);
3724 * Will never be NULL as the idle class always
3725 * returns a non-NULL p:
3727 class = class->next
;
3732 * schedule() is the main scheduler function.
3734 asmlinkage
void __sched
schedule(void)
3736 struct task_struct
*prev
, *next
;
3737 unsigned long *switch_count
;
3743 cpu
= smp_processor_id();
3745 rcu_note_context_switch(cpu
);
3748 release_kernel_lock(prev
);
3749 need_resched_nonpreemptible
:
3751 schedule_debug(prev
);
3753 if (sched_feat(HRTICK
))
3756 raw_spin_lock_irq(&rq
->lock
);
3757 clear_tsk_need_resched(prev
);
3759 switch_count
= &prev
->nivcsw
;
3760 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3761 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3762 prev
->state
= TASK_RUNNING
;
3765 * If a worker is going to sleep, notify and
3766 * ask workqueue whether it wants to wake up a
3767 * task to maintain concurrency. If so, wake
3770 if (prev
->flags
& PF_WQ_WORKER
) {
3771 struct task_struct
*to_wakeup
;
3773 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3775 try_to_wake_up_local(to_wakeup
);
3777 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3779 switch_count
= &prev
->nvcsw
;
3782 pre_schedule(rq
, prev
);
3784 if (unlikely(!rq
->nr_running
))
3785 idle_balance(cpu
, rq
);
3787 put_prev_task(rq
, prev
);
3788 next
= pick_next_task(rq
);
3790 if (likely(prev
!= next
)) {
3791 sched_info_switch(prev
, next
);
3792 perf_event_task_sched_out(prev
, next
);
3798 context_switch(rq
, prev
, next
); /* unlocks the rq */
3800 * The context switch have flipped the stack from under us
3801 * and restored the local variables which were saved when
3802 * this task called schedule() in the past. prev == current
3803 * is still correct, but it can be moved to another cpu/rq.
3805 cpu
= smp_processor_id();
3808 raw_spin_unlock_irq(&rq
->lock
);
3812 if (unlikely(reacquire_kernel_lock(prev
)))
3813 goto need_resched_nonpreemptible
;
3815 preempt_enable_no_resched();
3819 EXPORT_SYMBOL(schedule
);
3821 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3823 * Look out! "owner" is an entirely speculative pointer
3824 * access and not reliable.
3826 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3831 if (!sched_feat(OWNER_SPIN
))
3834 #ifdef CONFIG_DEBUG_PAGEALLOC
3836 * Need to access the cpu field knowing that
3837 * DEBUG_PAGEALLOC could have unmapped it if
3838 * the mutex owner just released it and exited.
3840 if (probe_kernel_address(&owner
->cpu
, cpu
))
3847 * Even if the access succeeded (likely case),
3848 * the cpu field may no longer be valid.
3850 if (cpu
>= nr_cpumask_bits
)
3854 * We need to validate that we can do a
3855 * get_cpu() and that we have the percpu area.
3857 if (!cpu_online(cpu
))
3864 * Owner changed, break to re-assess state.
3866 if (lock
->owner
!= owner
) {
3868 * If the lock has switched to a different owner,
3869 * we likely have heavy contention. Return 0 to quit
3870 * optimistic spinning and not contend further:
3878 * Is that owner really running on that cpu?
3880 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3890 #ifdef CONFIG_PREEMPT
3892 * this is the entry point to schedule() from in-kernel preemption
3893 * off of preempt_enable. Kernel preemptions off return from interrupt
3894 * occur there and call schedule directly.
3896 asmlinkage
void __sched notrace
preempt_schedule(void)
3898 struct thread_info
*ti
= current_thread_info();
3901 * If there is a non-zero preempt_count or interrupts are disabled,
3902 * we do not want to preempt the current task. Just return..
3904 if (likely(ti
->preempt_count
|| irqs_disabled()))
3908 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3910 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3913 * Check again in case we missed a preemption opportunity
3914 * between schedule and now.
3917 } while (need_resched());
3919 EXPORT_SYMBOL(preempt_schedule
);
3922 * this is the entry point to schedule() from kernel preemption
3923 * off of irq context.
3924 * Note, that this is called and return with irqs disabled. This will
3925 * protect us against recursive calling from irq.
3927 asmlinkage
void __sched
preempt_schedule_irq(void)
3929 struct thread_info
*ti
= current_thread_info();
3931 /* Catch callers which need to be fixed */
3932 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3935 add_preempt_count(PREEMPT_ACTIVE
);
3938 local_irq_disable();
3939 sub_preempt_count(PREEMPT_ACTIVE
);
3942 * Check again in case we missed a preemption opportunity
3943 * between schedule and now.
3946 } while (need_resched());
3949 #endif /* CONFIG_PREEMPT */
3951 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3954 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3956 EXPORT_SYMBOL(default_wake_function
);
3959 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3960 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3961 * number) then we wake all the non-exclusive tasks and one exclusive task.
3963 * There are circumstances in which we can try to wake a task which has already
3964 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3965 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3967 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3968 int nr_exclusive
, int wake_flags
, void *key
)
3970 wait_queue_t
*curr
, *next
;
3972 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3973 unsigned flags
= curr
->flags
;
3975 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3976 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3982 * __wake_up - wake up threads blocked on a waitqueue.
3984 * @mode: which threads
3985 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3986 * @key: is directly passed to the wakeup function
3988 * It may be assumed that this function implies a write memory barrier before
3989 * changing the task state if and only if any tasks are woken up.
3991 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3992 int nr_exclusive
, void *key
)
3994 unsigned long flags
;
3996 spin_lock_irqsave(&q
->lock
, flags
);
3997 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3998 spin_unlock_irqrestore(&q
->lock
, flags
);
4000 EXPORT_SYMBOL(__wake_up
);
4003 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4005 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4007 __wake_up_common(q
, mode
, 1, 0, NULL
);
4009 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4011 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4013 __wake_up_common(q
, mode
, 1, 0, key
);
4017 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4019 * @mode: which threads
4020 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4021 * @key: opaque value to be passed to wakeup targets
4023 * The sync wakeup differs that the waker knows that it will schedule
4024 * away soon, so while the target thread will be woken up, it will not
4025 * be migrated to another CPU - ie. the two threads are 'synchronized'
4026 * with each other. This can prevent needless bouncing between CPUs.
4028 * On UP it can prevent extra preemption.
4030 * It may be assumed that this function implies a write memory barrier before
4031 * changing the task state if and only if any tasks are woken up.
4033 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4034 int nr_exclusive
, void *key
)
4036 unsigned long flags
;
4037 int wake_flags
= WF_SYNC
;
4042 if (unlikely(!nr_exclusive
))
4045 spin_lock_irqsave(&q
->lock
, flags
);
4046 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4047 spin_unlock_irqrestore(&q
->lock
, flags
);
4049 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4052 * __wake_up_sync - see __wake_up_sync_key()
4054 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4056 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4058 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4061 * complete: - signals a single thread waiting on this completion
4062 * @x: holds the state of this particular completion
4064 * This will wake up a single thread waiting on this completion. Threads will be
4065 * awakened in the same order in which they were queued.
4067 * See also complete_all(), wait_for_completion() and related routines.
4069 * It may be assumed that this function implies a write memory barrier before
4070 * changing the task state if and only if any tasks are woken up.
4072 void complete(struct completion
*x
)
4074 unsigned long flags
;
4076 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4078 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4079 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4081 EXPORT_SYMBOL(complete
);
4084 * complete_all: - signals all threads waiting on this completion
4085 * @x: holds the state of this particular completion
4087 * This will wake up all threads waiting on this particular completion event.
4089 * It may be assumed that this function implies a write memory barrier before
4090 * changing the task state if and only if any tasks are woken up.
4092 void complete_all(struct completion
*x
)
4094 unsigned long flags
;
4096 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4097 x
->done
+= UINT_MAX
/2;
4098 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4099 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4101 EXPORT_SYMBOL(complete_all
);
4103 static inline long __sched
4104 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4107 DECLARE_WAITQUEUE(wait
, current
);
4109 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4111 if (signal_pending_state(state
, current
)) {
4112 timeout
= -ERESTARTSYS
;
4115 __set_current_state(state
);
4116 spin_unlock_irq(&x
->wait
.lock
);
4117 timeout
= schedule_timeout(timeout
);
4118 spin_lock_irq(&x
->wait
.lock
);
4119 } while (!x
->done
&& timeout
);
4120 __remove_wait_queue(&x
->wait
, &wait
);
4125 return timeout
?: 1;
4129 wait_for_common(struct completion
*x
, long timeout
, int state
)
4133 spin_lock_irq(&x
->wait
.lock
);
4134 timeout
= do_wait_for_common(x
, timeout
, state
);
4135 spin_unlock_irq(&x
->wait
.lock
);
4140 * wait_for_completion: - waits for completion of a task
4141 * @x: holds the state of this particular completion
4143 * This waits to be signaled for completion of a specific task. It is NOT
4144 * interruptible and there is no timeout.
4146 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4147 * and interrupt capability. Also see complete().
4149 void __sched
wait_for_completion(struct completion
*x
)
4151 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4153 EXPORT_SYMBOL(wait_for_completion
);
4156 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4157 * @x: holds the state of this particular completion
4158 * @timeout: timeout value in jiffies
4160 * This waits for either a completion of a specific task to be signaled or for a
4161 * specified timeout to expire. The timeout is in jiffies. It is not
4164 unsigned long __sched
4165 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4167 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4169 EXPORT_SYMBOL(wait_for_completion_timeout
);
4172 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4173 * @x: holds the state of this particular completion
4175 * This waits for completion of a specific task to be signaled. It is
4178 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4180 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4181 if (t
== -ERESTARTSYS
)
4185 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4188 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4189 * @x: holds the state of this particular completion
4190 * @timeout: timeout value in jiffies
4192 * This waits for either a completion of a specific task to be signaled or for a
4193 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4195 unsigned long __sched
4196 wait_for_completion_interruptible_timeout(struct completion
*x
,
4197 unsigned long timeout
)
4199 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4201 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4204 * wait_for_completion_killable: - waits for completion of a task (killable)
4205 * @x: holds the state of this particular completion
4207 * This waits to be signaled for completion of a specific task. It can be
4208 * interrupted by a kill signal.
4210 int __sched
wait_for_completion_killable(struct completion
*x
)
4212 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4213 if (t
== -ERESTARTSYS
)
4217 EXPORT_SYMBOL(wait_for_completion_killable
);
4220 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4221 * @x: holds the state of this particular completion
4222 * @timeout: timeout value in jiffies
4224 * This waits for either a completion of a specific task to be
4225 * signaled or for a specified timeout to expire. It can be
4226 * interrupted by a kill signal. The timeout is in jiffies.
4228 unsigned long __sched
4229 wait_for_completion_killable_timeout(struct completion
*x
,
4230 unsigned long timeout
)
4232 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4234 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4237 * try_wait_for_completion - try to decrement a completion without blocking
4238 * @x: completion structure
4240 * Returns: 0 if a decrement cannot be done without blocking
4241 * 1 if a decrement succeeded.
4243 * If a completion is being used as a counting completion,
4244 * attempt to decrement the counter without blocking. This
4245 * enables us to avoid waiting if the resource the completion
4246 * is protecting is not available.
4248 bool try_wait_for_completion(struct completion
*x
)
4250 unsigned long flags
;
4253 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4258 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4261 EXPORT_SYMBOL(try_wait_for_completion
);
4264 * completion_done - Test to see if a completion has any waiters
4265 * @x: completion structure
4267 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4268 * 1 if there are no waiters.
4271 bool completion_done(struct completion
*x
)
4273 unsigned long flags
;
4276 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4279 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4282 EXPORT_SYMBOL(completion_done
);
4285 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4287 unsigned long flags
;
4290 init_waitqueue_entry(&wait
, current
);
4292 __set_current_state(state
);
4294 spin_lock_irqsave(&q
->lock
, flags
);
4295 __add_wait_queue(q
, &wait
);
4296 spin_unlock(&q
->lock
);
4297 timeout
= schedule_timeout(timeout
);
4298 spin_lock_irq(&q
->lock
);
4299 __remove_wait_queue(q
, &wait
);
4300 spin_unlock_irqrestore(&q
->lock
, flags
);
4305 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4307 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4309 EXPORT_SYMBOL(interruptible_sleep_on
);
4312 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4314 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4316 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4318 void __sched
sleep_on(wait_queue_head_t
*q
)
4320 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4322 EXPORT_SYMBOL(sleep_on
);
4324 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4326 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4328 EXPORT_SYMBOL(sleep_on_timeout
);
4330 #ifdef CONFIG_RT_MUTEXES
4333 * rt_mutex_setprio - set the current priority of a task
4335 * @prio: prio value (kernel-internal form)
4337 * This function changes the 'effective' priority of a task. It does
4338 * not touch ->normal_prio like __setscheduler().
4340 * Used by the rt_mutex code to implement priority inheritance logic.
4342 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4344 unsigned long flags
;
4345 int oldprio
, on_rq
, running
;
4347 const struct sched_class
*prev_class
;
4349 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4351 rq
= task_rq_lock(p
, &flags
);
4354 prev_class
= p
->sched_class
;
4355 on_rq
= p
->se
.on_rq
;
4356 running
= task_current(rq
, p
);
4358 dequeue_task(rq
, p
, 0);
4360 p
->sched_class
->put_prev_task(rq
, p
);
4363 p
->sched_class
= &rt_sched_class
;
4365 p
->sched_class
= &fair_sched_class
;
4370 p
->sched_class
->set_curr_task(rq
);
4372 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4374 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4376 task_rq_unlock(rq
, &flags
);
4381 void set_user_nice(struct task_struct
*p
, long nice
)
4383 int old_prio
, delta
, on_rq
;
4384 unsigned long flags
;
4387 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4390 * We have to be careful, if called from sys_setpriority(),
4391 * the task might be in the middle of scheduling on another CPU.
4393 rq
= task_rq_lock(p
, &flags
);
4395 * The RT priorities are set via sched_setscheduler(), but we still
4396 * allow the 'normal' nice value to be set - but as expected
4397 * it wont have any effect on scheduling until the task is
4398 * SCHED_FIFO/SCHED_RR:
4400 if (task_has_rt_policy(p
)) {
4401 p
->static_prio
= NICE_TO_PRIO(nice
);
4404 on_rq
= p
->se
.on_rq
;
4406 dequeue_task(rq
, p
, 0);
4408 p
->static_prio
= NICE_TO_PRIO(nice
);
4411 p
->prio
= effective_prio(p
);
4412 delta
= p
->prio
- old_prio
;
4415 enqueue_task(rq
, p
, 0);
4417 * If the task increased its priority or is running and
4418 * lowered its priority, then reschedule its CPU:
4420 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4421 resched_task(rq
->curr
);
4424 task_rq_unlock(rq
, &flags
);
4426 EXPORT_SYMBOL(set_user_nice
);
4429 * can_nice - check if a task can reduce its nice value
4433 int can_nice(const struct task_struct
*p
, const int nice
)
4435 /* convert nice value [19,-20] to rlimit style value [1,40] */
4436 int nice_rlim
= 20 - nice
;
4438 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4439 capable(CAP_SYS_NICE
));
4442 #ifdef __ARCH_WANT_SYS_NICE
4445 * sys_nice - change the priority of the current process.
4446 * @increment: priority increment
4448 * sys_setpriority is a more generic, but much slower function that
4449 * does similar things.
4451 SYSCALL_DEFINE1(nice
, int, increment
)
4456 * Setpriority might change our priority at the same moment.
4457 * We don't have to worry. Conceptually one call occurs first
4458 * and we have a single winner.
4460 if (increment
< -40)
4465 nice
= TASK_NICE(current
) + increment
;
4471 if (increment
< 0 && !can_nice(current
, nice
))
4474 retval
= security_task_setnice(current
, nice
);
4478 set_user_nice(current
, nice
);
4485 * task_prio - return the priority value of a given task.
4486 * @p: the task in question.
4488 * This is the priority value as seen by users in /proc.
4489 * RT tasks are offset by -200. Normal tasks are centered
4490 * around 0, value goes from -16 to +15.
4492 int task_prio(const struct task_struct
*p
)
4494 return p
->prio
- MAX_RT_PRIO
;
4498 * task_nice - return the nice value of a given task.
4499 * @p: the task in question.
4501 int task_nice(const struct task_struct
*p
)
4503 return TASK_NICE(p
);
4505 EXPORT_SYMBOL(task_nice
);
4508 * idle_cpu - is a given cpu idle currently?
4509 * @cpu: the processor in question.
4511 int idle_cpu(int cpu
)
4513 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4517 * idle_task - return the idle task for a given cpu.
4518 * @cpu: the processor in question.
4520 struct task_struct
*idle_task(int cpu
)
4522 return cpu_rq(cpu
)->idle
;
4526 * find_process_by_pid - find a process with a matching PID value.
4527 * @pid: the pid in question.
4529 static struct task_struct
*find_process_by_pid(pid_t pid
)
4531 return pid
? find_task_by_vpid(pid
) : current
;
4534 /* Actually do priority change: must hold rq lock. */
4536 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4538 BUG_ON(p
->se
.on_rq
);
4541 p
->rt_priority
= prio
;
4542 p
->normal_prio
= normal_prio(p
);
4543 /* we are holding p->pi_lock already */
4544 p
->prio
= rt_mutex_getprio(p
);
4545 if (rt_prio(p
->prio
))
4546 p
->sched_class
= &rt_sched_class
;
4548 p
->sched_class
= &fair_sched_class
;
4553 * check the target process has a UID that matches the current process's
4555 static bool check_same_owner(struct task_struct
*p
)
4557 const struct cred
*cred
= current_cred(), *pcred
;
4561 pcred
= __task_cred(p
);
4562 match
= (cred
->euid
== pcred
->euid
||
4563 cred
->euid
== pcred
->uid
);
4568 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4569 struct sched_param
*param
, bool user
)
4571 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4572 unsigned long flags
;
4573 const struct sched_class
*prev_class
;
4577 /* may grab non-irq protected spin_locks */
4578 BUG_ON(in_interrupt());
4580 /* double check policy once rq lock held */
4582 reset_on_fork
= p
->sched_reset_on_fork
;
4583 policy
= oldpolicy
= p
->policy
;
4585 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4586 policy
&= ~SCHED_RESET_ON_FORK
;
4588 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4589 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4590 policy
!= SCHED_IDLE
)
4595 * Valid priorities for SCHED_FIFO and SCHED_RR are
4596 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4597 * SCHED_BATCH and SCHED_IDLE is 0.
4599 if (param
->sched_priority
< 0 ||
4600 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4601 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4603 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4607 * Allow unprivileged RT tasks to decrease priority:
4609 if (user
&& !capable(CAP_SYS_NICE
)) {
4610 if (rt_policy(policy
)) {
4611 unsigned long rlim_rtprio
=
4612 task_rlimit(p
, RLIMIT_RTPRIO
);
4614 /* can't set/change the rt policy */
4615 if (policy
!= p
->policy
&& !rlim_rtprio
)
4618 /* can't increase priority */
4619 if (param
->sched_priority
> p
->rt_priority
&&
4620 param
->sched_priority
> rlim_rtprio
)
4624 * Like positive nice levels, dont allow tasks to
4625 * move out of SCHED_IDLE either:
4627 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4630 /* can't change other user's priorities */
4631 if (!check_same_owner(p
))
4634 /* Normal users shall not reset the sched_reset_on_fork flag */
4635 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4640 retval
= security_task_setscheduler(p
, policy
, param
);
4646 * make sure no PI-waiters arrive (or leave) while we are
4647 * changing the priority of the task:
4649 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4651 * To be able to change p->policy safely, the apropriate
4652 * runqueue lock must be held.
4654 rq
= __task_rq_lock(p
);
4656 #ifdef CONFIG_RT_GROUP_SCHED
4659 * Do not allow realtime tasks into groups that have no runtime
4662 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4663 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4664 __task_rq_unlock(rq
);
4665 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4671 /* recheck policy now with rq lock held */
4672 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4673 policy
= oldpolicy
= -1;
4674 __task_rq_unlock(rq
);
4675 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4678 on_rq
= p
->se
.on_rq
;
4679 running
= task_current(rq
, p
);
4681 deactivate_task(rq
, p
, 0);
4683 p
->sched_class
->put_prev_task(rq
, p
);
4685 p
->sched_reset_on_fork
= reset_on_fork
;
4688 prev_class
= p
->sched_class
;
4689 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4692 p
->sched_class
->set_curr_task(rq
);
4694 activate_task(rq
, p
, 0);
4696 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4698 __task_rq_unlock(rq
);
4699 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4701 rt_mutex_adjust_pi(p
);
4707 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4708 * @p: the task in question.
4709 * @policy: new policy.
4710 * @param: structure containing the new RT priority.
4712 * NOTE that the task may be already dead.
4714 int sched_setscheduler(struct task_struct
*p
, int policy
,
4715 struct sched_param
*param
)
4717 return __sched_setscheduler(p
, policy
, param
, true);
4719 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4722 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4723 * @p: the task in question.
4724 * @policy: new policy.
4725 * @param: structure containing the new RT priority.
4727 * Just like sched_setscheduler, only don't bother checking if the
4728 * current context has permission. For example, this is needed in
4729 * stop_machine(): we create temporary high priority worker threads,
4730 * but our caller might not have that capability.
4732 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4733 struct sched_param
*param
)
4735 return __sched_setscheduler(p
, policy
, param
, false);
4739 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4741 struct sched_param lparam
;
4742 struct task_struct
*p
;
4745 if (!param
|| pid
< 0)
4747 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4752 p
= find_process_by_pid(pid
);
4754 retval
= sched_setscheduler(p
, policy
, &lparam
);
4761 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4762 * @pid: the pid in question.
4763 * @policy: new policy.
4764 * @param: structure containing the new RT priority.
4766 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4767 struct sched_param __user
*, param
)
4769 /* negative values for policy are not valid */
4773 return do_sched_setscheduler(pid
, policy
, param
);
4777 * sys_sched_setparam - set/change the RT priority of a thread
4778 * @pid: the pid in question.
4779 * @param: structure containing the new RT priority.
4781 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4783 return do_sched_setscheduler(pid
, -1, param
);
4787 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4788 * @pid: the pid in question.
4790 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4792 struct task_struct
*p
;
4800 p
= find_process_by_pid(pid
);
4802 retval
= security_task_getscheduler(p
);
4805 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4812 * sys_sched_getparam - get the RT priority of a thread
4813 * @pid: the pid in question.
4814 * @param: structure containing the RT priority.
4816 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4818 struct sched_param lp
;
4819 struct task_struct
*p
;
4822 if (!param
|| pid
< 0)
4826 p
= find_process_by_pid(pid
);
4831 retval
= security_task_getscheduler(p
);
4835 lp
.sched_priority
= p
->rt_priority
;
4839 * This one might sleep, we cannot do it with a spinlock held ...
4841 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4850 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4852 cpumask_var_t cpus_allowed
, new_mask
;
4853 struct task_struct
*p
;
4859 p
= find_process_by_pid(pid
);
4866 /* Prevent p going away */
4870 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4874 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4876 goto out_free_cpus_allowed
;
4879 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4882 retval
= security_task_setscheduler(p
, 0, NULL
);
4886 cpuset_cpus_allowed(p
, cpus_allowed
);
4887 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4889 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4892 cpuset_cpus_allowed(p
, cpus_allowed
);
4893 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4895 * We must have raced with a concurrent cpuset
4896 * update. Just reset the cpus_allowed to the
4897 * cpuset's cpus_allowed
4899 cpumask_copy(new_mask
, cpus_allowed
);
4904 free_cpumask_var(new_mask
);
4905 out_free_cpus_allowed
:
4906 free_cpumask_var(cpus_allowed
);
4913 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4914 struct cpumask
*new_mask
)
4916 if (len
< cpumask_size())
4917 cpumask_clear(new_mask
);
4918 else if (len
> cpumask_size())
4919 len
= cpumask_size();
4921 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4925 * sys_sched_setaffinity - set the cpu affinity of a process
4926 * @pid: pid of the process
4927 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4928 * @user_mask_ptr: user-space pointer to the new cpu mask
4930 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4931 unsigned long __user
*, user_mask_ptr
)
4933 cpumask_var_t new_mask
;
4936 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4939 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4941 retval
= sched_setaffinity(pid
, new_mask
);
4942 free_cpumask_var(new_mask
);
4946 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4948 struct task_struct
*p
;
4949 unsigned long flags
;
4957 p
= find_process_by_pid(pid
);
4961 retval
= security_task_getscheduler(p
);
4965 rq
= task_rq_lock(p
, &flags
);
4966 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4967 task_rq_unlock(rq
, &flags
);
4977 * sys_sched_getaffinity - get the cpu affinity of a process
4978 * @pid: pid of the process
4979 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4980 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4982 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4983 unsigned long __user
*, user_mask_ptr
)
4988 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4990 if (len
& (sizeof(unsigned long)-1))
4993 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4996 ret
= sched_getaffinity(pid
, mask
);
4998 size_t retlen
= min_t(size_t, len
, cpumask_size());
5000 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5005 free_cpumask_var(mask
);
5011 * sys_sched_yield - yield the current processor to other threads.
5013 * This function yields the current CPU to other tasks. If there are no
5014 * other threads running on this CPU then this function will return.
5016 SYSCALL_DEFINE0(sched_yield
)
5018 struct rq
*rq
= this_rq_lock();
5020 schedstat_inc(rq
, yld_count
);
5021 current
->sched_class
->yield_task(rq
);
5024 * Since we are going to call schedule() anyway, there's
5025 * no need to preempt or enable interrupts:
5027 __release(rq
->lock
);
5028 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5029 do_raw_spin_unlock(&rq
->lock
);
5030 preempt_enable_no_resched();
5037 static inline int should_resched(void)
5039 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5042 static void __cond_resched(void)
5044 add_preempt_count(PREEMPT_ACTIVE
);
5046 sub_preempt_count(PREEMPT_ACTIVE
);
5049 int __sched
_cond_resched(void)
5051 if (should_resched()) {
5057 EXPORT_SYMBOL(_cond_resched
);
5060 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5061 * call schedule, and on return reacquire the lock.
5063 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5064 * operations here to prevent schedule() from being called twice (once via
5065 * spin_unlock(), once by hand).
5067 int __cond_resched_lock(spinlock_t
*lock
)
5069 int resched
= should_resched();
5072 lockdep_assert_held(lock
);
5074 if (spin_needbreak(lock
) || resched
) {
5085 EXPORT_SYMBOL(__cond_resched_lock
);
5087 int __sched
__cond_resched_softirq(void)
5089 BUG_ON(!in_softirq());
5091 if (should_resched()) {
5099 EXPORT_SYMBOL(__cond_resched_softirq
);
5102 * yield - yield the current processor to other threads.
5104 * This is a shortcut for kernel-space yielding - it marks the
5105 * thread runnable and calls sys_sched_yield().
5107 void __sched
yield(void)
5109 set_current_state(TASK_RUNNING
);
5112 EXPORT_SYMBOL(yield
);
5115 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5116 * that process accounting knows that this is a task in IO wait state.
5118 void __sched
io_schedule(void)
5120 struct rq
*rq
= raw_rq();
5122 delayacct_blkio_start();
5123 atomic_inc(&rq
->nr_iowait
);
5124 current
->in_iowait
= 1;
5126 current
->in_iowait
= 0;
5127 atomic_dec(&rq
->nr_iowait
);
5128 delayacct_blkio_end();
5130 EXPORT_SYMBOL(io_schedule
);
5132 long __sched
io_schedule_timeout(long timeout
)
5134 struct rq
*rq
= raw_rq();
5137 delayacct_blkio_start();
5138 atomic_inc(&rq
->nr_iowait
);
5139 current
->in_iowait
= 1;
5140 ret
= schedule_timeout(timeout
);
5141 current
->in_iowait
= 0;
5142 atomic_dec(&rq
->nr_iowait
);
5143 delayacct_blkio_end();
5148 * sys_sched_get_priority_max - return maximum RT priority.
5149 * @policy: scheduling class.
5151 * this syscall returns the maximum rt_priority that can be used
5152 * by a given scheduling class.
5154 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5161 ret
= MAX_USER_RT_PRIO
-1;
5173 * sys_sched_get_priority_min - return minimum RT priority.
5174 * @policy: scheduling class.
5176 * this syscall returns the minimum rt_priority that can be used
5177 * by a given scheduling class.
5179 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5197 * sys_sched_rr_get_interval - return the default timeslice of a process.
5198 * @pid: pid of the process.
5199 * @interval: userspace pointer to the timeslice value.
5201 * this syscall writes the default timeslice value of a given process
5202 * into the user-space timespec buffer. A value of '0' means infinity.
5204 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5205 struct timespec __user
*, interval
)
5207 struct task_struct
*p
;
5208 unsigned int time_slice
;
5209 unsigned long flags
;
5219 p
= find_process_by_pid(pid
);
5223 retval
= security_task_getscheduler(p
);
5227 rq
= task_rq_lock(p
, &flags
);
5228 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5229 task_rq_unlock(rq
, &flags
);
5232 jiffies_to_timespec(time_slice
, &t
);
5233 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5241 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5243 void sched_show_task(struct task_struct
*p
)
5245 unsigned long free
= 0;
5248 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5249 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5250 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5251 #if BITS_PER_LONG == 32
5252 if (state
== TASK_RUNNING
)
5253 printk(KERN_CONT
" running ");
5255 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5257 if (state
== TASK_RUNNING
)
5258 printk(KERN_CONT
" running task ");
5260 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5262 #ifdef CONFIG_DEBUG_STACK_USAGE
5263 free
= stack_not_used(p
);
5265 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5266 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5267 (unsigned long)task_thread_info(p
)->flags
);
5269 show_stack(p
, NULL
);
5272 void show_state_filter(unsigned long state_filter
)
5274 struct task_struct
*g
, *p
;
5276 #if BITS_PER_LONG == 32
5278 " task PC stack pid father\n");
5281 " task PC stack pid father\n");
5283 read_lock(&tasklist_lock
);
5284 do_each_thread(g
, p
) {
5286 * reset the NMI-timeout, listing all files on a slow
5287 * console might take alot of time:
5289 touch_nmi_watchdog();
5290 if (!state_filter
|| (p
->state
& state_filter
))
5292 } while_each_thread(g
, p
);
5294 touch_all_softlockup_watchdogs();
5296 #ifdef CONFIG_SCHED_DEBUG
5297 sysrq_sched_debug_show();
5299 read_unlock(&tasklist_lock
);
5301 * Only show locks if all tasks are dumped:
5304 debug_show_all_locks();
5307 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5309 idle
->sched_class
= &idle_sched_class
;
5313 * init_idle - set up an idle thread for a given CPU
5314 * @idle: task in question
5315 * @cpu: cpu the idle task belongs to
5317 * NOTE: this function does not set the idle thread's NEED_RESCHED
5318 * flag, to make booting more robust.
5320 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5322 struct rq
*rq
= cpu_rq(cpu
);
5323 unsigned long flags
;
5325 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5328 idle
->state
= TASK_RUNNING
;
5329 idle
->se
.exec_start
= sched_clock();
5331 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5332 __set_task_cpu(idle
, cpu
);
5334 rq
->curr
= rq
->idle
= idle
;
5335 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5338 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5340 /* Set the preempt count _outside_ the spinlocks! */
5341 #if defined(CONFIG_PREEMPT)
5342 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5344 task_thread_info(idle
)->preempt_count
= 0;
5347 * The idle tasks have their own, simple scheduling class:
5349 idle
->sched_class
= &idle_sched_class
;
5350 ftrace_graph_init_task(idle
);
5354 * In a system that switches off the HZ timer nohz_cpu_mask
5355 * indicates which cpus entered this state. This is used
5356 * in the rcu update to wait only for active cpus. For system
5357 * which do not switch off the HZ timer nohz_cpu_mask should
5358 * always be CPU_BITS_NONE.
5360 cpumask_var_t nohz_cpu_mask
;
5363 * Increase the granularity value when there are more CPUs,
5364 * because with more CPUs the 'effective latency' as visible
5365 * to users decreases. But the relationship is not linear,
5366 * so pick a second-best guess by going with the log2 of the
5369 * This idea comes from the SD scheduler of Con Kolivas:
5371 static int get_update_sysctl_factor(void)
5373 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5374 unsigned int factor
;
5376 switch (sysctl_sched_tunable_scaling
) {
5377 case SCHED_TUNABLESCALING_NONE
:
5380 case SCHED_TUNABLESCALING_LINEAR
:
5383 case SCHED_TUNABLESCALING_LOG
:
5385 factor
= 1 + ilog2(cpus
);
5392 static void update_sysctl(void)
5394 unsigned int factor
= get_update_sysctl_factor();
5396 #define SET_SYSCTL(name) \
5397 (sysctl_##name = (factor) * normalized_sysctl_##name)
5398 SET_SYSCTL(sched_min_granularity
);
5399 SET_SYSCTL(sched_latency
);
5400 SET_SYSCTL(sched_wakeup_granularity
);
5401 SET_SYSCTL(sched_shares_ratelimit
);
5405 static inline void sched_init_granularity(void)
5412 * This is how migration works:
5414 * 1) we invoke migration_cpu_stop() on the target CPU using
5416 * 2) stopper starts to run (implicitly forcing the migrated thread
5418 * 3) it checks whether the migrated task is still in the wrong runqueue.
5419 * 4) if it's in the wrong runqueue then the migration thread removes
5420 * it and puts it into the right queue.
5421 * 5) stopper completes and stop_one_cpu() returns and the migration
5426 * Change a given task's CPU affinity. Migrate the thread to a
5427 * proper CPU and schedule it away if the CPU it's executing on
5428 * is removed from the allowed bitmask.
5430 * NOTE: the caller must have a valid reference to the task, the
5431 * task must not exit() & deallocate itself prematurely. The
5432 * call is not atomic; no spinlocks may be held.
5434 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5436 unsigned long flags
;
5438 unsigned int dest_cpu
;
5442 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5443 * drop the rq->lock and still rely on ->cpus_allowed.
5446 while (task_is_waking(p
))
5448 rq
= task_rq_lock(p
, &flags
);
5449 if (task_is_waking(p
)) {
5450 task_rq_unlock(rq
, &flags
);
5454 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5459 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5460 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5465 if (p
->sched_class
->set_cpus_allowed
)
5466 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5468 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5469 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5472 /* Can the task run on the task's current CPU? If so, we're done */
5473 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5476 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5477 if (migrate_task(p
, dest_cpu
)) {
5478 struct migration_arg arg
= { p
, dest_cpu
};
5479 /* Need help from migration thread: drop lock and wait. */
5480 task_rq_unlock(rq
, &flags
);
5481 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5482 tlb_migrate_finish(p
->mm
);
5486 task_rq_unlock(rq
, &flags
);
5490 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5493 * Move (not current) task off this cpu, onto dest cpu. We're doing
5494 * this because either it can't run here any more (set_cpus_allowed()
5495 * away from this CPU, or CPU going down), or because we're
5496 * attempting to rebalance this task on exec (sched_exec).
5498 * So we race with normal scheduler movements, but that's OK, as long
5499 * as the task is no longer on this CPU.
5501 * Returns non-zero if task was successfully migrated.
5503 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5505 struct rq
*rq_dest
, *rq_src
;
5508 if (unlikely(!cpu_active(dest_cpu
)))
5511 rq_src
= cpu_rq(src_cpu
);
5512 rq_dest
= cpu_rq(dest_cpu
);
5514 double_rq_lock(rq_src
, rq_dest
);
5515 /* Already moved. */
5516 if (task_cpu(p
) != src_cpu
)
5518 /* Affinity changed (again). */
5519 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5523 * If we're not on a rq, the next wake-up will ensure we're
5527 deactivate_task(rq_src
, p
, 0);
5528 set_task_cpu(p
, dest_cpu
);
5529 activate_task(rq_dest
, p
, 0);
5530 check_preempt_curr(rq_dest
, p
, 0);
5535 double_rq_unlock(rq_src
, rq_dest
);
5540 * migration_cpu_stop - this will be executed by a highprio stopper thread
5541 * and performs thread migration by bumping thread off CPU then
5542 * 'pushing' onto another runqueue.
5544 static int migration_cpu_stop(void *data
)
5546 struct migration_arg
*arg
= data
;
5549 * The original target cpu might have gone down and we might
5550 * be on another cpu but it doesn't matter.
5552 local_irq_disable();
5553 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5558 #ifdef CONFIG_HOTPLUG_CPU
5560 * Figure out where task on dead CPU should go, use force if necessary.
5562 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5564 struct rq
*rq
= cpu_rq(dead_cpu
);
5565 int needs_cpu
, uninitialized_var(dest_cpu
);
5566 unsigned long flags
;
5568 local_irq_save(flags
);
5570 raw_spin_lock(&rq
->lock
);
5571 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5573 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5574 raw_spin_unlock(&rq
->lock
);
5576 * It can only fail if we race with set_cpus_allowed(),
5577 * in the racer should migrate the task anyway.
5580 __migrate_task(p
, dead_cpu
, dest_cpu
);
5581 local_irq_restore(flags
);
5585 * While a dead CPU has no uninterruptible tasks queued at this point,
5586 * it might still have a nonzero ->nr_uninterruptible counter, because
5587 * for performance reasons the counter is not stricly tracking tasks to
5588 * their home CPUs. So we just add the counter to another CPU's counter,
5589 * to keep the global sum constant after CPU-down:
5591 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5593 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5594 unsigned long flags
;
5596 local_irq_save(flags
);
5597 double_rq_lock(rq_src
, rq_dest
);
5598 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5599 rq_src
->nr_uninterruptible
= 0;
5600 double_rq_unlock(rq_src
, rq_dest
);
5601 local_irq_restore(flags
);
5604 /* Run through task list and migrate tasks from the dead cpu. */
5605 static void migrate_live_tasks(int src_cpu
)
5607 struct task_struct
*p
, *t
;
5609 read_lock(&tasklist_lock
);
5611 do_each_thread(t
, p
) {
5615 if (task_cpu(p
) == src_cpu
)
5616 move_task_off_dead_cpu(src_cpu
, p
);
5617 } while_each_thread(t
, p
);
5619 read_unlock(&tasklist_lock
);
5623 * Schedules idle task to be the next runnable task on current CPU.
5624 * It does so by boosting its priority to highest possible.
5625 * Used by CPU offline code.
5627 void sched_idle_next(void)
5629 int this_cpu
= smp_processor_id();
5630 struct rq
*rq
= cpu_rq(this_cpu
);
5631 struct task_struct
*p
= rq
->idle
;
5632 unsigned long flags
;
5634 /* cpu has to be offline */
5635 BUG_ON(cpu_online(this_cpu
));
5638 * Strictly not necessary since rest of the CPUs are stopped by now
5639 * and interrupts disabled on the current cpu.
5641 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5643 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5645 activate_task(rq
, p
, 0);
5647 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5651 * Ensures that the idle task is using init_mm right before its cpu goes
5654 void idle_task_exit(void)
5656 struct mm_struct
*mm
= current
->active_mm
;
5658 BUG_ON(cpu_online(smp_processor_id()));
5661 switch_mm(mm
, &init_mm
, current
);
5665 /* called under rq->lock with disabled interrupts */
5666 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5668 struct rq
*rq
= cpu_rq(dead_cpu
);
5670 /* Must be exiting, otherwise would be on tasklist. */
5671 BUG_ON(!p
->exit_state
);
5673 /* Cannot have done final schedule yet: would have vanished. */
5674 BUG_ON(p
->state
== TASK_DEAD
);
5679 * Drop lock around migration; if someone else moves it,
5680 * that's OK. No task can be added to this CPU, so iteration is
5683 raw_spin_unlock_irq(&rq
->lock
);
5684 move_task_off_dead_cpu(dead_cpu
, p
);
5685 raw_spin_lock_irq(&rq
->lock
);
5690 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5691 static void migrate_dead_tasks(unsigned int dead_cpu
)
5693 struct rq
*rq
= cpu_rq(dead_cpu
);
5694 struct task_struct
*next
;
5697 if (!rq
->nr_running
)
5699 next
= pick_next_task(rq
);
5702 next
->sched_class
->put_prev_task(rq
, next
);
5703 migrate_dead(dead_cpu
, next
);
5709 * remove the tasks which were accounted by rq from calc_load_tasks.
5711 static void calc_global_load_remove(struct rq
*rq
)
5713 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5714 rq
->calc_load_active
= 0;
5716 #endif /* CONFIG_HOTPLUG_CPU */
5718 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5720 static struct ctl_table sd_ctl_dir
[] = {
5722 .procname
= "sched_domain",
5728 static struct ctl_table sd_ctl_root
[] = {
5730 .procname
= "kernel",
5732 .child
= sd_ctl_dir
,
5737 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5739 struct ctl_table
*entry
=
5740 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5745 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5747 struct ctl_table
*entry
;
5750 * In the intermediate directories, both the child directory and
5751 * procname are dynamically allocated and could fail but the mode
5752 * will always be set. In the lowest directory the names are
5753 * static strings and all have proc handlers.
5755 for (entry
= *tablep
; entry
->mode
; entry
++) {
5757 sd_free_ctl_entry(&entry
->child
);
5758 if (entry
->proc_handler
== NULL
)
5759 kfree(entry
->procname
);
5767 set_table_entry(struct ctl_table
*entry
,
5768 const char *procname
, void *data
, int maxlen
,
5769 mode_t mode
, proc_handler
*proc_handler
)
5771 entry
->procname
= procname
;
5773 entry
->maxlen
= maxlen
;
5775 entry
->proc_handler
= proc_handler
;
5778 static struct ctl_table
*
5779 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5781 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5786 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5787 sizeof(long), 0644, proc_doulongvec_minmax
);
5788 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5789 sizeof(long), 0644, proc_doulongvec_minmax
);
5790 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5791 sizeof(int), 0644, proc_dointvec_minmax
);
5792 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5793 sizeof(int), 0644, proc_dointvec_minmax
);
5794 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5795 sizeof(int), 0644, proc_dointvec_minmax
);
5796 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5797 sizeof(int), 0644, proc_dointvec_minmax
);
5798 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5799 sizeof(int), 0644, proc_dointvec_minmax
);
5800 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5801 sizeof(int), 0644, proc_dointvec_minmax
);
5802 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5803 sizeof(int), 0644, proc_dointvec_minmax
);
5804 set_table_entry(&table
[9], "cache_nice_tries",
5805 &sd
->cache_nice_tries
,
5806 sizeof(int), 0644, proc_dointvec_minmax
);
5807 set_table_entry(&table
[10], "flags", &sd
->flags
,
5808 sizeof(int), 0644, proc_dointvec_minmax
);
5809 set_table_entry(&table
[11], "name", sd
->name
,
5810 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5811 /* &table[12] is terminator */
5816 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5818 struct ctl_table
*entry
, *table
;
5819 struct sched_domain
*sd
;
5820 int domain_num
= 0, i
;
5823 for_each_domain(cpu
, sd
)
5825 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5830 for_each_domain(cpu
, sd
) {
5831 snprintf(buf
, 32, "domain%d", i
);
5832 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5834 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5841 static struct ctl_table_header
*sd_sysctl_header
;
5842 static void register_sched_domain_sysctl(void)
5844 int i
, cpu_num
= num_possible_cpus();
5845 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5848 WARN_ON(sd_ctl_dir
[0].child
);
5849 sd_ctl_dir
[0].child
= entry
;
5854 for_each_possible_cpu(i
) {
5855 snprintf(buf
, 32, "cpu%d", i
);
5856 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5858 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5862 WARN_ON(sd_sysctl_header
);
5863 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5866 /* may be called multiple times per register */
5867 static void unregister_sched_domain_sysctl(void)
5869 if (sd_sysctl_header
)
5870 unregister_sysctl_table(sd_sysctl_header
);
5871 sd_sysctl_header
= NULL
;
5872 if (sd_ctl_dir
[0].child
)
5873 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5876 static void register_sched_domain_sysctl(void)
5879 static void unregister_sched_domain_sysctl(void)
5884 static void set_rq_online(struct rq
*rq
)
5887 const struct sched_class
*class;
5889 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5892 for_each_class(class) {
5893 if (class->rq_online
)
5894 class->rq_online(rq
);
5899 static void set_rq_offline(struct rq
*rq
)
5902 const struct sched_class
*class;
5904 for_each_class(class) {
5905 if (class->rq_offline
)
5906 class->rq_offline(rq
);
5909 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5915 * migration_call - callback that gets triggered when a CPU is added.
5916 * Here we can start up the necessary migration thread for the new CPU.
5918 static int __cpuinit
5919 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5921 int cpu
= (long)hcpu
;
5922 unsigned long flags
;
5923 struct rq
*rq
= cpu_rq(cpu
);
5927 case CPU_UP_PREPARE
:
5928 case CPU_UP_PREPARE_FROZEN
:
5929 rq
->calc_load_update
= calc_load_update
;
5933 case CPU_ONLINE_FROZEN
:
5934 /* Update our root-domain */
5935 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5937 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5941 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5944 #ifdef CONFIG_HOTPLUG_CPU
5946 case CPU_DEAD_FROZEN
:
5947 migrate_live_tasks(cpu
);
5948 /* Idle task back to normal (off runqueue, low prio) */
5949 raw_spin_lock_irq(&rq
->lock
);
5950 deactivate_task(rq
, rq
->idle
, 0);
5951 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5952 rq
->idle
->sched_class
= &idle_sched_class
;
5953 migrate_dead_tasks(cpu
);
5954 raw_spin_unlock_irq(&rq
->lock
);
5955 migrate_nr_uninterruptible(rq
);
5956 BUG_ON(rq
->nr_running
!= 0);
5957 calc_global_load_remove(rq
);
5961 case CPU_DYING_FROZEN
:
5962 /* Update our root-domain */
5963 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5965 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5968 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5976 * Register at high priority so that task migration (migrate_all_tasks)
5977 * happens before everything else. This has to be lower priority than
5978 * the notifier in the perf_event subsystem, though.
5980 static struct notifier_block __cpuinitdata migration_notifier
= {
5981 .notifier_call
= migration_call
,
5982 .priority
= CPU_PRI_MIGRATION
,
5985 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5986 unsigned long action
, void *hcpu
)
5988 switch (action
& ~CPU_TASKS_FROZEN
) {
5990 case CPU_DOWN_FAILED
:
5991 set_cpu_active((long)hcpu
, true);
5998 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
5999 unsigned long action
, void *hcpu
)
6001 switch (action
& ~CPU_TASKS_FROZEN
) {
6002 case CPU_DOWN_PREPARE
:
6003 set_cpu_active((long)hcpu
, false);
6010 static int __init
migration_init(void)
6012 void *cpu
= (void *)(long)smp_processor_id();
6015 /* Initialize migration for the boot CPU */
6016 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6017 BUG_ON(err
== NOTIFY_BAD
);
6018 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6019 register_cpu_notifier(&migration_notifier
);
6021 /* Register cpu active notifiers */
6022 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6023 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6027 early_initcall(migration_init
);
6032 #ifdef CONFIG_SCHED_DEBUG
6034 static __read_mostly
int sched_domain_debug_enabled
;
6036 static int __init
sched_domain_debug_setup(char *str
)
6038 sched_domain_debug_enabled
= 1;
6042 early_param("sched_debug", sched_domain_debug_setup
);
6044 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6045 struct cpumask
*groupmask
)
6047 struct sched_group
*group
= sd
->groups
;
6050 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6051 cpumask_clear(groupmask
);
6053 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6055 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6056 printk("does not load-balance\n");
6058 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6063 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6065 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6066 printk(KERN_ERR
"ERROR: domain->span does not contain "
6069 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6070 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6074 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6078 printk(KERN_ERR
"ERROR: group is NULL\n");
6082 if (!group
->cpu_power
) {
6083 printk(KERN_CONT
"\n");
6084 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6089 if (!cpumask_weight(sched_group_cpus(group
))) {
6090 printk(KERN_CONT
"\n");
6091 printk(KERN_ERR
"ERROR: empty group\n");
6095 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6096 printk(KERN_CONT
"\n");
6097 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6101 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6103 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6105 printk(KERN_CONT
" %s", str
);
6106 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6107 printk(KERN_CONT
" (cpu_power = %d)",
6111 group
= group
->next
;
6112 } while (group
!= sd
->groups
);
6113 printk(KERN_CONT
"\n");
6115 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6116 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6119 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6120 printk(KERN_ERR
"ERROR: parent span is not a superset "
6121 "of domain->span\n");
6125 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6127 cpumask_var_t groupmask
;
6130 if (!sched_domain_debug_enabled
)
6134 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6138 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6140 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6141 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6146 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6153 free_cpumask_var(groupmask
);
6155 #else /* !CONFIG_SCHED_DEBUG */
6156 # define sched_domain_debug(sd, cpu) do { } while (0)
6157 #endif /* CONFIG_SCHED_DEBUG */
6159 static int sd_degenerate(struct sched_domain
*sd
)
6161 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6164 /* Following flags need at least 2 groups */
6165 if (sd
->flags
& (SD_LOAD_BALANCE
|
6166 SD_BALANCE_NEWIDLE
|
6170 SD_SHARE_PKG_RESOURCES
)) {
6171 if (sd
->groups
!= sd
->groups
->next
)
6175 /* Following flags don't use groups */
6176 if (sd
->flags
& (SD_WAKE_AFFINE
))
6183 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6185 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6187 if (sd_degenerate(parent
))
6190 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6193 /* Flags needing groups don't count if only 1 group in parent */
6194 if (parent
->groups
== parent
->groups
->next
) {
6195 pflags
&= ~(SD_LOAD_BALANCE
|
6196 SD_BALANCE_NEWIDLE
|
6200 SD_SHARE_PKG_RESOURCES
);
6201 if (nr_node_ids
== 1)
6202 pflags
&= ~SD_SERIALIZE
;
6204 if (~cflags
& pflags
)
6210 static void free_rootdomain(struct root_domain
*rd
)
6212 synchronize_sched();
6214 cpupri_cleanup(&rd
->cpupri
);
6216 free_cpumask_var(rd
->rto_mask
);
6217 free_cpumask_var(rd
->online
);
6218 free_cpumask_var(rd
->span
);
6222 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6224 struct root_domain
*old_rd
= NULL
;
6225 unsigned long flags
;
6227 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6232 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6235 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6238 * If we dont want to free the old_rt yet then
6239 * set old_rd to NULL to skip the freeing later
6242 if (!atomic_dec_and_test(&old_rd
->refcount
))
6246 atomic_inc(&rd
->refcount
);
6249 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6250 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6253 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6256 free_rootdomain(old_rd
);
6259 static int init_rootdomain(struct root_domain
*rd
)
6261 memset(rd
, 0, sizeof(*rd
));
6263 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6265 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6267 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6270 if (cpupri_init(&rd
->cpupri
) != 0)
6275 free_cpumask_var(rd
->rto_mask
);
6277 free_cpumask_var(rd
->online
);
6279 free_cpumask_var(rd
->span
);
6284 static void init_defrootdomain(void)
6286 init_rootdomain(&def_root_domain
);
6288 atomic_set(&def_root_domain
.refcount
, 1);
6291 static struct root_domain
*alloc_rootdomain(void)
6293 struct root_domain
*rd
;
6295 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6299 if (init_rootdomain(rd
) != 0) {
6308 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6309 * hold the hotplug lock.
6312 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6314 struct rq
*rq
= cpu_rq(cpu
);
6315 struct sched_domain
*tmp
;
6317 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6318 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6320 /* Remove the sched domains which do not contribute to scheduling. */
6321 for (tmp
= sd
; tmp
; ) {
6322 struct sched_domain
*parent
= tmp
->parent
;
6326 if (sd_parent_degenerate(tmp
, parent
)) {
6327 tmp
->parent
= parent
->parent
;
6329 parent
->parent
->child
= tmp
;
6334 if (sd
&& sd_degenerate(sd
)) {
6340 sched_domain_debug(sd
, cpu
);
6342 rq_attach_root(rq
, rd
);
6343 rcu_assign_pointer(rq
->sd
, sd
);
6346 /* cpus with isolated domains */
6347 static cpumask_var_t cpu_isolated_map
;
6349 /* Setup the mask of cpus configured for isolated domains */
6350 static int __init
isolated_cpu_setup(char *str
)
6352 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6353 cpulist_parse(str
, cpu_isolated_map
);
6357 __setup("isolcpus=", isolated_cpu_setup
);
6360 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6361 * to a function which identifies what group(along with sched group) a CPU
6362 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6363 * (due to the fact that we keep track of groups covered with a struct cpumask).
6365 * init_sched_build_groups will build a circular linked list of the groups
6366 * covered by the given span, and will set each group's ->cpumask correctly,
6367 * and ->cpu_power to 0.
6370 init_sched_build_groups(const struct cpumask
*span
,
6371 const struct cpumask
*cpu_map
,
6372 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6373 struct sched_group
**sg
,
6374 struct cpumask
*tmpmask
),
6375 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6377 struct sched_group
*first
= NULL
, *last
= NULL
;
6380 cpumask_clear(covered
);
6382 for_each_cpu(i
, span
) {
6383 struct sched_group
*sg
;
6384 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6387 if (cpumask_test_cpu(i
, covered
))
6390 cpumask_clear(sched_group_cpus(sg
));
6393 for_each_cpu(j
, span
) {
6394 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6397 cpumask_set_cpu(j
, covered
);
6398 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6409 #define SD_NODES_PER_DOMAIN 16
6414 * find_next_best_node - find the next node to include in a sched_domain
6415 * @node: node whose sched_domain we're building
6416 * @used_nodes: nodes already in the sched_domain
6418 * Find the next node to include in a given scheduling domain. Simply
6419 * finds the closest node not already in the @used_nodes map.
6421 * Should use nodemask_t.
6423 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6425 int i
, n
, val
, min_val
, best_node
= 0;
6429 for (i
= 0; i
< nr_node_ids
; i
++) {
6430 /* Start at @node */
6431 n
= (node
+ i
) % nr_node_ids
;
6433 if (!nr_cpus_node(n
))
6436 /* Skip already used nodes */
6437 if (node_isset(n
, *used_nodes
))
6440 /* Simple min distance search */
6441 val
= node_distance(node
, n
);
6443 if (val
< min_val
) {
6449 node_set(best_node
, *used_nodes
);
6454 * sched_domain_node_span - get a cpumask for a node's sched_domain
6455 * @node: node whose cpumask we're constructing
6456 * @span: resulting cpumask
6458 * Given a node, construct a good cpumask for its sched_domain to span. It
6459 * should be one that prevents unnecessary balancing, but also spreads tasks
6462 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6464 nodemask_t used_nodes
;
6467 cpumask_clear(span
);
6468 nodes_clear(used_nodes
);
6470 cpumask_or(span
, span
, cpumask_of_node(node
));
6471 node_set(node
, used_nodes
);
6473 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6474 int next_node
= find_next_best_node(node
, &used_nodes
);
6476 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6479 #endif /* CONFIG_NUMA */
6481 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6484 * The cpus mask in sched_group and sched_domain hangs off the end.
6486 * ( See the the comments in include/linux/sched.h:struct sched_group
6487 * and struct sched_domain. )
6489 struct static_sched_group
{
6490 struct sched_group sg
;
6491 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6494 struct static_sched_domain
{
6495 struct sched_domain sd
;
6496 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6502 cpumask_var_t domainspan
;
6503 cpumask_var_t covered
;
6504 cpumask_var_t notcovered
;
6506 cpumask_var_t nodemask
;
6507 cpumask_var_t this_sibling_map
;
6508 cpumask_var_t this_core_map
;
6509 cpumask_var_t this_book_map
;
6510 cpumask_var_t send_covered
;
6511 cpumask_var_t tmpmask
;
6512 struct sched_group
**sched_group_nodes
;
6513 struct root_domain
*rd
;
6517 sa_sched_groups
= 0,
6523 sa_this_sibling_map
,
6525 sa_sched_group_nodes
,
6535 * SMT sched-domains:
6537 #ifdef CONFIG_SCHED_SMT
6538 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6539 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6542 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6543 struct sched_group
**sg
, struct cpumask
*unused
)
6546 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6549 #endif /* CONFIG_SCHED_SMT */
6552 * multi-core sched-domains:
6554 #ifdef CONFIG_SCHED_MC
6555 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6556 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6559 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6560 struct sched_group
**sg
, struct cpumask
*mask
)
6563 #ifdef CONFIG_SCHED_SMT
6564 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6565 group
= cpumask_first(mask
);
6570 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6573 #endif /* CONFIG_SCHED_MC */
6576 * book sched-domains:
6578 #ifdef CONFIG_SCHED_BOOK
6579 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
6580 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
6583 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
6584 struct sched_group
**sg
, struct cpumask
*mask
)
6587 #ifdef CONFIG_SCHED_MC
6588 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6589 group
= cpumask_first(mask
);
6590 #elif defined(CONFIG_SCHED_SMT)
6591 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6592 group
= cpumask_first(mask
);
6595 *sg
= &per_cpu(sched_group_book
, group
).sg
;
6598 #endif /* CONFIG_SCHED_BOOK */
6600 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6601 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6604 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6605 struct sched_group
**sg
, struct cpumask
*mask
)
6608 #ifdef CONFIG_SCHED_BOOK
6609 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
6610 group
= cpumask_first(mask
);
6611 #elif defined(CONFIG_SCHED_MC)
6612 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6613 group
= cpumask_first(mask
);
6614 #elif defined(CONFIG_SCHED_SMT)
6615 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6616 group
= cpumask_first(mask
);
6621 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6627 * The init_sched_build_groups can't handle what we want to do with node
6628 * groups, so roll our own. Now each node has its own list of groups which
6629 * gets dynamically allocated.
6631 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6632 static struct sched_group
***sched_group_nodes_bycpu
;
6634 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6635 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6637 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6638 struct sched_group
**sg
,
6639 struct cpumask
*nodemask
)
6643 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6644 group
= cpumask_first(nodemask
);
6647 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6651 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6653 struct sched_group
*sg
= group_head
;
6659 for_each_cpu(j
, sched_group_cpus(sg
)) {
6660 struct sched_domain
*sd
;
6662 sd
= &per_cpu(phys_domains
, j
).sd
;
6663 if (j
!= group_first_cpu(sd
->groups
)) {
6665 * Only add "power" once for each
6671 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6674 } while (sg
!= group_head
);
6677 static int build_numa_sched_groups(struct s_data
*d
,
6678 const struct cpumask
*cpu_map
, int num
)
6680 struct sched_domain
*sd
;
6681 struct sched_group
*sg
, *prev
;
6684 cpumask_clear(d
->covered
);
6685 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6686 if (cpumask_empty(d
->nodemask
)) {
6687 d
->sched_group_nodes
[num
] = NULL
;
6691 sched_domain_node_span(num
, d
->domainspan
);
6692 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6694 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6697 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6701 d
->sched_group_nodes
[num
] = sg
;
6703 for_each_cpu(j
, d
->nodemask
) {
6704 sd
= &per_cpu(node_domains
, j
).sd
;
6709 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6711 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6714 for (j
= 0; j
< nr_node_ids
; j
++) {
6715 n
= (num
+ j
) % nr_node_ids
;
6716 cpumask_complement(d
->notcovered
, d
->covered
);
6717 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6718 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6719 if (cpumask_empty(d
->tmpmask
))
6721 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6722 if (cpumask_empty(d
->tmpmask
))
6724 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6728 "Can not alloc domain group for node %d\n", j
);
6732 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6733 sg
->next
= prev
->next
;
6734 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6741 #endif /* CONFIG_NUMA */
6744 /* Free memory allocated for various sched_group structures */
6745 static void free_sched_groups(const struct cpumask
*cpu_map
,
6746 struct cpumask
*nodemask
)
6750 for_each_cpu(cpu
, cpu_map
) {
6751 struct sched_group
**sched_group_nodes
6752 = sched_group_nodes_bycpu
[cpu
];
6754 if (!sched_group_nodes
)
6757 for (i
= 0; i
< nr_node_ids
; i
++) {
6758 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6760 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6761 if (cpumask_empty(nodemask
))
6771 if (oldsg
!= sched_group_nodes
[i
])
6774 kfree(sched_group_nodes
);
6775 sched_group_nodes_bycpu
[cpu
] = NULL
;
6778 #else /* !CONFIG_NUMA */
6779 static void free_sched_groups(const struct cpumask
*cpu_map
,
6780 struct cpumask
*nodemask
)
6783 #endif /* CONFIG_NUMA */
6786 * Initialize sched groups cpu_power.
6788 * cpu_power indicates the capacity of sched group, which is used while
6789 * distributing the load between different sched groups in a sched domain.
6790 * Typically cpu_power for all the groups in a sched domain will be same unless
6791 * there are asymmetries in the topology. If there are asymmetries, group
6792 * having more cpu_power will pickup more load compared to the group having
6795 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6797 struct sched_domain
*child
;
6798 struct sched_group
*group
;
6802 WARN_ON(!sd
|| !sd
->groups
);
6804 if (cpu
!= group_first_cpu(sd
->groups
))
6809 sd
->groups
->cpu_power
= 0;
6812 power
= SCHED_LOAD_SCALE
;
6813 weight
= cpumask_weight(sched_domain_span(sd
));
6815 * SMT siblings share the power of a single core.
6816 * Usually multiple threads get a better yield out of
6817 * that one core than a single thread would have,
6818 * reflect that in sd->smt_gain.
6820 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6821 power
*= sd
->smt_gain
;
6823 power
>>= SCHED_LOAD_SHIFT
;
6825 sd
->groups
->cpu_power
+= power
;
6830 * Add cpu_power of each child group to this groups cpu_power.
6832 group
= child
->groups
;
6834 sd
->groups
->cpu_power
+= group
->cpu_power
;
6835 group
= group
->next
;
6836 } while (group
!= child
->groups
);
6840 * Initializers for schedule domains
6841 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6844 #ifdef CONFIG_SCHED_DEBUG
6845 # define SD_INIT_NAME(sd, type) sd->name = #type
6847 # define SD_INIT_NAME(sd, type) do { } while (0)
6850 #define SD_INIT(sd, type) sd_init_##type(sd)
6852 #define SD_INIT_FUNC(type) \
6853 static noinline void sd_init_##type(struct sched_domain *sd) \
6855 memset(sd, 0, sizeof(*sd)); \
6856 *sd = SD_##type##_INIT; \
6857 sd->level = SD_LV_##type; \
6858 SD_INIT_NAME(sd, type); \
6863 SD_INIT_FUNC(ALLNODES
)
6866 #ifdef CONFIG_SCHED_SMT
6867 SD_INIT_FUNC(SIBLING
)
6869 #ifdef CONFIG_SCHED_MC
6872 #ifdef CONFIG_SCHED_BOOK
6876 static int default_relax_domain_level
= -1;
6878 static int __init
setup_relax_domain_level(char *str
)
6882 val
= simple_strtoul(str
, NULL
, 0);
6883 if (val
< SD_LV_MAX
)
6884 default_relax_domain_level
= val
;
6888 __setup("relax_domain_level=", setup_relax_domain_level
);
6890 static void set_domain_attribute(struct sched_domain
*sd
,
6891 struct sched_domain_attr
*attr
)
6895 if (!attr
|| attr
->relax_domain_level
< 0) {
6896 if (default_relax_domain_level
< 0)
6899 request
= default_relax_domain_level
;
6901 request
= attr
->relax_domain_level
;
6902 if (request
< sd
->level
) {
6903 /* turn off idle balance on this domain */
6904 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6906 /* turn on idle balance on this domain */
6907 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6911 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6912 const struct cpumask
*cpu_map
)
6915 case sa_sched_groups
:
6916 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6917 d
->sched_group_nodes
= NULL
;
6919 free_rootdomain(d
->rd
); /* fall through */
6921 free_cpumask_var(d
->tmpmask
); /* fall through */
6922 case sa_send_covered
:
6923 free_cpumask_var(d
->send_covered
); /* fall through */
6924 case sa_this_book_map
:
6925 free_cpumask_var(d
->this_book_map
); /* fall through */
6926 case sa_this_core_map
:
6927 free_cpumask_var(d
->this_core_map
); /* fall through */
6928 case sa_this_sibling_map
:
6929 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6931 free_cpumask_var(d
->nodemask
); /* fall through */
6932 case sa_sched_group_nodes
:
6934 kfree(d
->sched_group_nodes
); /* fall through */
6936 free_cpumask_var(d
->notcovered
); /* fall through */
6938 free_cpumask_var(d
->covered
); /* fall through */
6940 free_cpumask_var(d
->domainspan
); /* fall through */
6947 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6948 const struct cpumask
*cpu_map
)
6951 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6953 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6954 return sa_domainspan
;
6955 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6957 /* Allocate the per-node list of sched groups */
6958 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6959 sizeof(struct sched_group
*), GFP_KERNEL
);
6960 if (!d
->sched_group_nodes
) {
6961 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6962 return sa_notcovered
;
6964 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6966 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6967 return sa_sched_group_nodes
;
6968 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6970 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6971 return sa_this_sibling_map
;
6972 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
6973 return sa_this_core_map
;
6974 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6975 return sa_this_book_map
;
6976 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6977 return sa_send_covered
;
6978 d
->rd
= alloc_rootdomain();
6980 printk(KERN_WARNING
"Cannot alloc root domain\n");
6983 return sa_rootdomain
;
6986 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6987 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6989 struct sched_domain
*sd
= NULL
;
6991 struct sched_domain
*parent
;
6994 if (cpumask_weight(cpu_map
) >
6995 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6996 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6997 SD_INIT(sd
, ALLNODES
);
6998 set_domain_attribute(sd
, attr
);
6999 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7000 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7005 sd
= &per_cpu(node_domains
, i
).sd
;
7007 set_domain_attribute(sd
, attr
);
7008 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7009 sd
->parent
= parent
;
7012 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7017 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7018 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7019 struct sched_domain
*parent
, int i
)
7021 struct sched_domain
*sd
;
7022 sd
= &per_cpu(phys_domains
, i
).sd
;
7024 set_domain_attribute(sd
, attr
);
7025 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7026 sd
->parent
= parent
;
7029 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7033 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7034 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7035 struct sched_domain
*parent
, int i
)
7037 struct sched_domain
*sd
= parent
;
7038 #ifdef CONFIG_SCHED_BOOK
7039 sd
= &per_cpu(book_domains
, i
).sd
;
7041 set_domain_attribute(sd
, attr
);
7042 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7043 sd
->parent
= parent
;
7045 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7050 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7051 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7052 struct sched_domain
*parent
, int i
)
7054 struct sched_domain
*sd
= parent
;
7055 #ifdef CONFIG_SCHED_MC
7056 sd
= &per_cpu(core_domains
, i
).sd
;
7058 set_domain_attribute(sd
, attr
);
7059 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7060 sd
->parent
= parent
;
7062 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7067 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7068 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7069 struct sched_domain
*parent
, int i
)
7071 struct sched_domain
*sd
= parent
;
7072 #ifdef CONFIG_SCHED_SMT
7073 sd
= &per_cpu(cpu_domains
, i
).sd
;
7074 SD_INIT(sd
, SIBLING
);
7075 set_domain_attribute(sd
, attr
);
7076 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7077 sd
->parent
= parent
;
7079 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7084 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7085 const struct cpumask
*cpu_map
, int cpu
)
7088 #ifdef CONFIG_SCHED_SMT
7089 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7090 cpumask_and(d
->this_sibling_map
, cpu_map
,
7091 topology_thread_cpumask(cpu
));
7092 if (cpu
== cpumask_first(d
->this_sibling_map
))
7093 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7095 d
->send_covered
, d
->tmpmask
);
7098 #ifdef CONFIG_SCHED_MC
7099 case SD_LV_MC
: /* set up multi-core groups */
7100 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7101 if (cpu
== cpumask_first(d
->this_core_map
))
7102 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7104 d
->send_covered
, d
->tmpmask
);
7107 #ifdef CONFIG_SCHED_BOOK
7108 case SD_LV_BOOK
: /* set up book groups */
7109 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7110 if (cpu
== cpumask_first(d
->this_book_map
))
7111 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7113 d
->send_covered
, d
->tmpmask
);
7116 case SD_LV_CPU
: /* set up physical groups */
7117 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7118 if (!cpumask_empty(d
->nodemask
))
7119 init_sched_build_groups(d
->nodemask
, cpu_map
,
7121 d
->send_covered
, d
->tmpmask
);
7124 case SD_LV_ALLNODES
:
7125 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7126 d
->send_covered
, d
->tmpmask
);
7135 * Build sched domains for a given set of cpus and attach the sched domains
7136 * to the individual cpus
7138 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7139 struct sched_domain_attr
*attr
)
7141 enum s_alloc alloc_state
= sa_none
;
7143 struct sched_domain
*sd
;
7149 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7150 if (alloc_state
!= sa_rootdomain
)
7152 alloc_state
= sa_sched_groups
;
7155 * Set up domains for cpus specified by the cpu_map.
7157 for_each_cpu(i
, cpu_map
) {
7158 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7161 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7162 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7163 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7164 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7165 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7168 for_each_cpu(i
, cpu_map
) {
7169 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7170 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7171 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7174 /* Set up physical groups */
7175 for (i
= 0; i
< nr_node_ids
; i
++)
7176 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7179 /* Set up node groups */
7181 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7183 for (i
= 0; i
< nr_node_ids
; i
++)
7184 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7188 /* Calculate CPU power for physical packages and nodes */
7189 #ifdef CONFIG_SCHED_SMT
7190 for_each_cpu(i
, cpu_map
) {
7191 sd
= &per_cpu(cpu_domains
, i
).sd
;
7192 init_sched_groups_power(i
, sd
);
7195 #ifdef CONFIG_SCHED_MC
7196 for_each_cpu(i
, cpu_map
) {
7197 sd
= &per_cpu(core_domains
, i
).sd
;
7198 init_sched_groups_power(i
, sd
);
7201 #ifdef CONFIG_SCHED_BOOK
7202 for_each_cpu(i
, cpu_map
) {
7203 sd
= &per_cpu(book_domains
, i
).sd
;
7204 init_sched_groups_power(i
, sd
);
7208 for_each_cpu(i
, cpu_map
) {
7209 sd
= &per_cpu(phys_domains
, i
).sd
;
7210 init_sched_groups_power(i
, sd
);
7214 for (i
= 0; i
< nr_node_ids
; i
++)
7215 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7217 if (d
.sd_allnodes
) {
7218 struct sched_group
*sg
;
7220 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7222 init_numa_sched_groups_power(sg
);
7226 /* Attach the domains */
7227 for_each_cpu(i
, cpu_map
) {
7228 #ifdef CONFIG_SCHED_SMT
7229 sd
= &per_cpu(cpu_domains
, i
).sd
;
7230 #elif defined(CONFIG_SCHED_MC)
7231 sd
= &per_cpu(core_domains
, i
).sd
;
7232 #elif defined(CONFIG_SCHED_BOOK)
7233 sd
= &per_cpu(book_domains
, i
).sd
;
7235 sd
= &per_cpu(phys_domains
, i
).sd
;
7237 cpu_attach_domain(sd
, d
.rd
, i
);
7240 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7241 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7245 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7249 static int build_sched_domains(const struct cpumask
*cpu_map
)
7251 return __build_sched_domains(cpu_map
, NULL
);
7254 static cpumask_var_t
*doms_cur
; /* current sched domains */
7255 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7256 static struct sched_domain_attr
*dattr_cur
;
7257 /* attribues of custom domains in 'doms_cur' */
7260 * Special case: If a kmalloc of a doms_cur partition (array of
7261 * cpumask) fails, then fallback to a single sched domain,
7262 * as determined by the single cpumask fallback_doms.
7264 static cpumask_var_t fallback_doms
;
7267 * arch_update_cpu_topology lets virtualized architectures update the
7268 * cpu core maps. It is supposed to return 1 if the topology changed
7269 * or 0 if it stayed the same.
7271 int __attribute__((weak
)) arch_update_cpu_topology(void)
7276 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7279 cpumask_var_t
*doms
;
7281 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7284 for (i
= 0; i
< ndoms
; i
++) {
7285 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7286 free_sched_domains(doms
, i
);
7293 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7296 for (i
= 0; i
< ndoms
; i
++)
7297 free_cpumask_var(doms
[i
]);
7302 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7303 * For now this just excludes isolated cpus, but could be used to
7304 * exclude other special cases in the future.
7306 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7310 arch_update_cpu_topology();
7312 doms_cur
= alloc_sched_domains(ndoms_cur
);
7314 doms_cur
= &fallback_doms
;
7315 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7317 err
= build_sched_domains(doms_cur
[0]);
7318 register_sched_domain_sysctl();
7323 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7324 struct cpumask
*tmpmask
)
7326 free_sched_groups(cpu_map
, tmpmask
);
7330 * Detach sched domains from a group of cpus specified in cpu_map
7331 * These cpus will now be attached to the NULL domain
7333 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7335 /* Save because hotplug lock held. */
7336 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7339 for_each_cpu(i
, cpu_map
)
7340 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7341 synchronize_sched();
7342 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7345 /* handle null as "default" */
7346 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7347 struct sched_domain_attr
*new, int idx_new
)
7349 struct sched_domain_attr tmp
;
7356 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7357 new ? (new + idx_new
) : &tmp
,
7358 sizeof(struct sched_domain_attr
));
7362 * Partition sched domains as specified by the 'ndoms_new'
7363 * cpumasks in the array doms_new[] of cpumasks. This compares
7364 * doms_new[] to the current sched domain partitioning, doms_cur[].
7365 * It destroys each deleted domain and builds each new domain.
7367 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7368 * The masks don't intersect (don't overlap.) We should setup one
7369 * sched domain for each mask. CPUs not in any of the cpumasks will
7370 * not be load balanced. If the same cpumask appears both in the
7371 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7374 * The passed in 'doms_new' should be allocated using
7375 * alloc_sched_domains. This routine takes ownership of it and will
7376 * free_sched_domains it when done with it. If the caller failed the
7377 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7378 * and partition_sched_domains() will fallback to the single partition
7379 * 'fallback_doms', it also forces the domains to be rebuilt.
7381 * If doms_new == NULL it will be replaced with cpu_online_mask.
7382 * ndoms_new == 0 is a special case for destroying existing domains,
7383 * and it will not create the default domain.
7385 * Call with hotplug lock held
7387 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7388 struct sched_domain_attr
*dattr_new
)
7393 mutex_lock(&sched_domains_mutex
);
7395 /* always unregister in case we don't destroy any domains */
7396 unregister_sched_domain_sysctl();
7398 /* Let architecture update cpu core mappings. */
7399 new_topology
= arch_update_cpu_topology();
7401 n
= doms_new
? ndoms_new
: 0;
7403 /* Destroy deleted domains */
7404 for (i
= 0; i
< ndoms_cur
; i
++) {
7405 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7406 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7407 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7410 /* no match - a current sched domain not in new doms_new[] */
7411 detach_destroy_domains(doms_cur
[i
]);
7416 if (doms_new
== NULL
) {
7418 doms_new
= &fallback_doms
;
7419 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7420 WARN_ON_ONCE(dattr_new
);
7423 /* Build new domains */
7424 for (i
= 0; i
< ndoms_new
; i
++) {
7425 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7426 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7427 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7430 /* no match - add a new doms_new */
7431 __build_sched_domains(doms_new
[i
],
7432 dattr_new
? dattr_new
+ i
: NULL
);
7437 /* Remember the new sched domains */
7438 if (doms_cur
!= &fallback_doms
)
7439 free_sched_domains(doms_cur
, ndoms_cur
);
7440 kfree(dattr_cur
); /* kfree(NULL) is safe */
7441 doms_cur
= doms_new
;
7442 dattr_cur
= dattr_new
;
7443 ndoms_cur
= ndoms_new
;
7445 register_sched_domain_sysctl();
7447 mutex_unlock(&sched_domains_mutex
);
7450 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7451 static void arch_reinit_sched_domains(void)
7455 /* Destroy domains first to force the rebuild */
7456 partition_sched_domains(0, NULL
, NULL
);
7458 rebuild_sched_domains();
7462 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7464 unsigned int level
= 0;
7466 if (sscanf(buf
, "%u", &level
) != 1)
7470 * level is always be positive so don't check for
7471 * level < POWERSAVINGS_BALANCE_NONE which is 0
7472 * What happens on 0 or 1 byte write,
7473 * need to check for count as well?
7476 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7480 sched_smt_power_savings
= level
;
7482 sched_mc_power_savings
= level
;
7484 arch_reinit_sched_domains();
7489 #ifdef CONFIG_SCHED_MC
7490 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7491 struct sysdev_class_attribute
*attr
,
7494 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7496 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7497 struct sysdev_class_attribute
*attr
,
7498 const char *buf
, size_t count
)
7500 return sched_power_savings_store(buf
, count
, 0);
7502 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7503 sched_mc_power_savings_show
,
7504 sched_mc_power_savings_store
);
7507 #ifdef CONFIG_SCHED_SMT
7508 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7509 struct sysdev_class_attribute
*attr
,
7512 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7514 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7515 struct sysdev_class_attribute
*attr
,
7516 const char *buf
, size_t count
)
7518 return sched_power_savings_store(buf
, count
, 1);
7520 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7521 sched_smt_power_savings_show
,
7522 sched_smt_power_savings_store
);
7525 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7529 #ifdef CONFIG_SCHED_SMT
7531 err
= sysfs_create_file(&cls
->kset
.kobj
,
7532 &attr_sched_smt_power_savings
.attr
);
7534 #ifdef CONFIG_SCHED_MC
7535 if (!err
&& mc_capable())
7536 err
= sysfs_create_file(&cls
->kset
.kobj
,
7537 &attr_sched_mc_power_savings
.attr
);
7541 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7544 * Update cpusets according to cpu_active mask. If cpusets are
7545 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7546 * around partition_sched_domains().
7548 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7551 switch (action
& ~CPU_TASKS_FROZEN
) {
7553 case CPU_DOWN_FAILED
:
7554 cpuset_update_active_cpus();
7561 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7564 switch (action
& ~CPU_TASKS_FROZEN
) {
7565 case CPU_DOWN_PREPARE
:
7566 cpuset_update_active_cpus();
7573 static int update_runtime(struct notifier_block
*nfb
,
7574 unsigned long action
, void *hcpu
)
7576 int cpu
= (int)(long)hcpu
;
7579 case CPU_DOWN_PREPARE
:
7580 case CPU_DOWN_PREPARE_FROZEN
:
7581 disable_runtime(cpu_rq(cpu
));
7584 case CPU_DOWN_FAILED
:
7585 case CPU_DOWN_FAILED_FROZEN
:
7587 case CPU_ONLINE_FROZEN
:
7588 enable_runtime(cpu_rq(cpu
));
7596 void __init
sched_init_smp(void)
7598 cpumask_var_t non_isolated_cpus
;
7600 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7601 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7603 #if defined(CONFIG_NUMA)
7604 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7606 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7609 mutex_lock(&sched_domains_mutex
);
7610 arch_init_sched_domains(cpu_active_mask
);
7611 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7612 if (cpumask_empty(non_isolated_cpus
))
7613 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7614 mutex_unlock(&sched_domains_mutex
);
7617 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7618 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7620 /* RT runtime code needs to handle some hotplug events */
7621 hotcpu_notifier(update_runtime
, 0);
7625 /* Move init over to a non-isolated CPU */
7626 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7628 sched_init_granularity();
7629 free_cpumask_var(non_isolated_cpus
);
7631 init_sched_rt_class();
7634 void __init
sched_init_smp(void)
7636 sched_init_granularity();
7638 #endif /* CONFIG_SMP */
7640 const_debug
unsigned int sysctl_timer_migration
= 1;
7642 int in_sched_functions(unsigned long addr
)
7644 return in_lock_functions(addr
) ||
7645 (addr
>= (unsigned long)__sched_text_start
7646 && addr
< (unsigned long)__sched_text_end
);
7649 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7651 cfs_rq
->tasks_timeline
= RB_ROOT
;
7652 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7653 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7659 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7661 struct rt_prio_array
*array
;
7664 array
= &rt_rq
->active
;
7665 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7666 INIT_LIST_HEAD(array
->queue
+ i
);
7667 __clear_bit(i
, array
->bitmap
);
7669 /* delimiter for bitsearch: */
7670 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7672 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7673 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7675 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7679 rt_rq
->rt_nr_migratory
= 0;
7680 rt_rq
->overloaded
= 0;
7681 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7685 rt_rq
->rt_throttled
= 0;
7686 rt_rq
->rt_runtime
= 0;
7687 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7689 #ifdef CONFIG_RT_GROUP_SCHED
7690 rt_rq
->rt_nr_boosted
= 0;
7695 #ifdef CONFIG_FAIR_GROUP_SCHED
7696 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7697 struct sched_entity
*se
, int cpu
, int add
,
7698 struct sched_entity
*parent
)
7700 struct rq
*rq
= cpu_rq(cpu
);
7701 tg
->cfs_rq
[cpu
] = cfs_rq
;
7702 init_cfs_rq(cfs_rq
, rq
);
7705 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7708 /* se could be NULL for init_task_group */
7713 se
->cfs_rq
= &rq
->cfs
;
7715 se
->cfs_rq
= parent
->my_q
;
7718 se
->load
.weight
= tg
->shares
;
7719 se
->load
.inv_weight
= 0;
7720 se
->parent
= parent
;
7724 #ifdef CONFIG_RT_GROUP_SCHED
7725 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7726 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7727 struct sched_rt_entity
*parent
)
7729 struct rq
*rq
= cpu_rq(cpu
);
7731 tg
->rt_rq
[cpu
] = rt_rq
;
7732 init_rt_rq(rt_rq
, rq
);
7734 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7736 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7738 tg
->rt_se
[cpu
] = rt_se
;
7743 rt_se
->rt_rq
= &rq
->rt
;
7745 rt_se
->rt_rq
= parent
->my_q
;
7747 rt_se
->my_q
= rt_rq
;
7748 rt_se
->parent
= parent
;
7749 INIT_LIST_HEAD(&rt_se
->run_list
);
7753 void __init
sched_init(void)
7756 unsigned long alloc_size
= 0, ptr
;
7758 #ifdef CONFIG_FAIR_GROUP_SCHED
7759 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7761 #ifdef CONFIG_RT_GROUP_SCHED
7762 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7764 #ifdef CONFIG_CPUMASK_OFFSTACK
7765 alloc_size
+= num_possible_cpus() * cpumask_size();
7768 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7770 #ifdef CONFIG_FAIR_GROUP_SCHED
7771 init_task_group
.se
= (struct sched_entity
**)ptr
;
7772 ptr
+= nr_cpu_ids
* sizeof(void **);
7774 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7775 ptr
+= nr_cpu_ids
* sizeof(void **);
7777 #endif /* CONFIG_FAIR_GROUP_SCHED */
7778 #ifdef CONFIG_RT_GROUP_SCHED
7779 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7780 ptr
+= nr_cpu_ids
* sizeof(void **);
7782 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7783 ptr
+= nr_cpu_ids
* sizeof(void **);
7785 #endif /* CONFIG_RT_GROUP_SCHED */
7786 #ifdef CONFIG_CPUMASK_OFFSTACK
7787 for_each_possible_cpu(i
) {
7788 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7789 ptr
+= cpumask_size();
7791 #endif /* CONFIG_CPUMASK_OFFSTACK */
7795 init_defrootdomain();
7798 init_rt_bandwidth(&def_rt_bandwidth
,
7799 global_rt_period(), global_rt_runtime());
7801 #ifdef CONFIG_RT_GROUP_SCHED
7802 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7803 global_rt_period(), global_rt_runtime());
7804 #endif /* CONFIG_RT_GROUP_SCHED */
7806 #ifdef CONFIG_CGROUP_SCHED
7807 list_add(&init_task_group
.list
, &task_groups
);
7808 INIT_LIST_HEAD(&init_task_group
.children
);
7810 #endif /* CONFIG_CGROUP_SCHED */
7812 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7813 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7814 __alignof__(unsigned long));
7816 for_each_possible_cpu(i
) {
7820 raw_spin_lock_init(&rq
->lock
);
7822 rq
->calc_load_active
= 0;
7823 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7824 init_cfs_rq(&rq
->cfs
, rq
);
7825 init_rt_rq(&rq
->rt
, rq
);
7826 #ifdef CONFIG_FAIR_GROUP_SCHED
7827 init_task_group
.shares
= init_task_group_load
;
7828 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7829 #ifdef CONFIG_CGROUP_SCHED
7831 * How much cpu bandwidth does init_task_group get?
7833 * In case of task-groups formed thr' the cgroup filesystem, it
7834 * gets 100% of the cpu resources in the system. This overall
7835 * system cpu resource is divided among the tasks of
7836 * init_task_group and its child task-groups in a fair manner,
7837 * based on each entity's (task or task-group's) weight
7838 * (se->load.weight).
7840 * In other words, if init_task_group has 10 tasks of weight
7841 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7842 * then A0's share of the cpu resource is:
7844 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7846 * We achieve this by letting init_task_group's tasks sit
7847 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7849 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7851 #endif /* CONFIG_FAIR_GROUP_SCHED */
7853 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7854 #ifdef CONFIG_RT_GROUP_SCHED
7855 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7856 #ifdef CONFIG_CGROUP_SCHED
7857 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7861 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7862 rq
->cpu_load
[j
] = 0;
7864 rq
->last_load_update_tick
= jiffies
;
7869 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7870 rq
->post_schedule
= 0;
7871 rq
->active_balance
= 0;
7872 rq
->next_balance
= jiffies
;
7877 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7878 rq_attach_root(rq
, &def_root_domain
);
7880 rq
->nohz_balance_kick
= 0;
7881 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
7885 atomic_set(&rq
->nr_iowait
, 0);
7888 set_load_weight(&init_task
);
7890 #ifdef CONFIG_PREEMPT_NOTIFIERS
7891 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7895 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7898 #ifdef CONFIG_RT_MUTEXES
7899 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7903 * The boot idle thread does lazy MMU switching as well:
7905 atomic_inc(&init_mm
.mm_count
);
7906 enter_lazy_tlb(&init_mm
, current
);
7909 * Make us the idle thread. Technically, schedule() should not be
7910 * called from this thread, however somewhere below it might be,
7911 * but because we are the idle thread, we just pick up running again
7912 * when this runqueue becomes "idle".
7914 init_idle(current
, smp_processor_id());
7916 calc_load_update
= jiffies
+ LOAD_FREQ
;
7919 * During early bootup we pretend to be a normal task:
7921 current
->sched_class
= &fair_sched_class
;
7923 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7924 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7927 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7928 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
7929 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
7930 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
7931 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
7933 /* May be allocated at isolcpus cmdline parse time */
7934 if (cpu_isolated_map
== NULL
)
7935 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7940 scheduler_running
= 1;
7943 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7944 static inline int preempt_count_equals(int preempt_offset
)
7946 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7948 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7951 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7954 static unsigned long prev_jiffy
; /* ratelimiting */
7956 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7957 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7959 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7961 prev_jiffy
= jiffies
;
7964 "BUG: sleeping function called from invalid context at %s:%d\n",
7967 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7968 in_atomic(), irqs_disabled(),
7969 current
->pid
, current
->comm
);
7971 debug_show_held_locks(current
);
7972 if (irqs_disabled())
7973 print_irqtrace_events(current
);
7977 EXPORT_SYMBOL(__might_sleep
);
7980 #ifdef CONFIG_MAGIC_SYSRQ
7981 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7985 on_rq
= p
->se
.on_rq
;
7987 deactivate_task(rq
, p
, 0);
7988 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7990 activate_task(rq
, p
, 0);
7991 resched_task(rq
->curr
);
7995 void normalize_rt_tasks(void)
7997 struct task_struct
*g
, *p
;
7998 unsigned long flags
;
8001 read_lock_irqsave(&tasklist_lock
, flags
);
8002 do_each_thread(g
, p
) {
8004 * Only normalize user tasks:
8009 p
->se
.exec_start
= 0;
8010 #ifdef CONFIG_SCHEDSTATS
8011 p
->se
.statistics
.wait_start
= 0;
8012 p
->se
.statistics
.sleep_start
= 0;
8013 p
->se
.statistics
.block_start
= 0;
8018 * Renice negative nice level userspace
8021 if (TASK_NICE(p
) < 0 && p
->mm
)
8022 set_user_nice(p
, 0);
8026 raw_spin_lock(&p
->pi_lock
);
8027 rq
= __task_rq_lock(p
);
8029 normalize_task(rq
, p
);
8031 __task_rq_unlock(rq
);
8032 raw_spin_unlock(&p
->pi_lock
);
8033 } while_each_thread(g
, p
);
8035 read_unlock_irqrestore(&tasklist_lock
, flags
);
8038 #endif /* CONFIG_MAGIC_SYSRQ */
8040 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8042 * These functions are only useful for the IA64 MCA handling, or kdb.
8044 * They can only be called when the whole system has been
8045 * stopped - every CPU needs to be quiescent, and no scheduling
8046 * activity can take place. Using them for anything else would
8047 * be a serious bug, and as a result, they aren't even visible
8048 * under any other configuration.
8052 * curr_task - return the current task for a given cpu.
8053 * @cpu: the processor in question.
8055 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8057 struct task_struct
*curr_task(int cpu
)
8059 return cpu_curr(cpu
);
8062 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8066 * set_curr_task - set the current task for a given cpu.
8067 * @cpu: the processor in question.
8068 * @p: the task pointer to set.
8070 * Description: This function must only be used when non-maskable interrupts
8071 * are serviced on a separate stack. It allows the architecture to switch the
8072 * notion of the current task on a cpu in a non-blocking manner. This function
8073 * must be called with all CPU's synchronized, and interrupts disabled, the
8074 * and caller must save the original value of the current task (see
8075 * curr_task() above) and restore that value before reenabling interrupts and
8076 * re-starting the system.
8078 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8080 void set_curr_task(int cpu
, struct task_struct
*p
)
8087 #ifdef CONFIG_FAIR_GROUP_SCHED
8088 static void free_fair_sched_group(struct task_group
*tg
)
8092 for_each_possible_cpu(i
) {
8094 kfree(tg
->cfs_rq
[i
]);
8104 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8106 struct cfs_rq
*cfs_rq
;
8107 struct sched_entity
*se
;
8111 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8114 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8118 tg
->shares
= NICE_0_LOAD
;
8120 for_each_possible_cpu(i
) {
8123 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8124 GFP_KERNEL
, cpu_to_node(i
));
8128 se
= kzalloc_node(sizeof(struct sched_entity
),
8129 GFP_KERNEL
, cpu_to_node(i
));
8133 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8144 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8146 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8147 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8150 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8152 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8154 #else /* !CONFG_FAIR_GROUP_SCHED */
8155 static inline void free_fair_sched_group(struct task_group
*tg
)
8160 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8165 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8169 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8172 #endif /* CONFIG_FAIR_GROUP_SCHED */
8174 #ifdef CONFIG_RT_GROUP_SCHED
8175 static void free_rt_sched_group(struct task_group
*tg
)
8179 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8181 for_each_possible_cpu(i
) {
8183 kfree(tg
->rt_rq
[i
]);
8185 kfree(tg
->rt_se
[i
]);
8193 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8195 struct rt_rq
*rt_rq
;
8196 struct sched_rt_entity
*rt_se
;
8200 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8203 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8207 init_rt_bandwidth(&tg
->rt_bandwidth
,
8208 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8210 for_each_possible_cpu(i
) {
8213 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8214 GFP_KERNEL
, cpu_to_node(i
));
8218 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8219 GFP_KERNEL
, cpu_to_node(i
));
8223 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8234 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8236 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8237 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8240 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8242 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8244 #else /* !CONFIG_RT_GROUP_SCHED */
8245 static inline void free_rt_sched_group(struct task_group
*tg
)
8250 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8255 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8259 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8262 #endif /* CONFIG_RT_GROUP_SCHED */
8264 #ifdef CONFIG_CGROUP_SCHED
8265 static void free_sched_group(struct task_group
*tg
)
8267 free_fair_sched_group(tg
);
8268 free_rt_sched_group(tg
);
8272 /* allocate runqueue etc for a new task group */
8273 struct task_group
*sched_create_group(struct task_group
*parent
)
8275 struct task_group
*tg
;
8276 unsigned long flags
;
8279 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8281 return ERR_PTR(-ENOMEM
);
8283 if (!alloc_fair_sched_group(tg
, parent
))
8286 if (!alloc_rt_sched_group(tg
, parent
))
8289 spin_lock_irqsave(&task_group_lock
, flags
);
8290 for_each_possible_cpu(i
) {
8291 register_fair_sched_group(tg
, i
);
8292 register_rt_sched_group(tg
, i
);
8294 list_add_rcu(&tg
->list
, &task_groups
);
8296 WARN_ON(!parent
); /* root should already exist */
8298 tg
->parent
= parent
;
8299 INIT_LIST_HEAD(&tg
->children
);
8300 list_add_rcu(&tg
->siblings
, &parent
->children
);
8301 spin_unlock_irqrestore(&task_group_lock
, flags
);
8306 free_sched_group(tg
);
8307 return ERR_PTR(-ENOMEM
);
8310 /* rcu callback to free various structures associated with a task group */
8311 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8313 /* now it should be safe to free those cfs_rqs */
8314 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8317 /* Destroy runqueue etc associated with a task group */
8318 void sched_destroy_group(struct task_group
*tg
)
8320 unsigned long flags
;
8323 spin_lock_irqsave(&task_group_lock
, flags
);
8324 for_each_possible_cpu(i
) {
8325 unregister_fair_sched_group(tg
, i
);
8326 unregister_rt_sched_group(tg
, i
);
8328 list_del_rcu(&tg
->list
);
8329 list_del_rcu(&tg
->siblings
);
8330 spin_unlock_irqrestore(&task_group_lock
, flags
);
8332 /* wait for possible concurrent references to cfs_rqs complete */
8333 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8336 /* change task's runqueue when it moves between groups.
8337 * The caller of this function should have put the task in its new group
8338 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8339 * reflect its new group.
8341 void sched_move_task(struct task_struct
*tsk
)
8344 unsigned long flags
;
8347 rq
= task_rq_lock(tsk
, &flags
);
8349 running
= task_current(rq
, tsk
);
8350 on_rq
= tsk
->se
.on_rq
;
8353 dequeue_task(rq
, tsk
, 0);
8354 if (unlikely(running
))
8355 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8357 set_task_rq(tsk
, task_cpu(tsk
));
8359 #ifdef CONFIG_FAIR_GROUP_SCHED
8360 if (tsk
->sched_class
->moved_group
)
8361 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8364 if (unlikely(running
))
8365 tsk
->sched_class
->set_curr_task(rq
);
8367 enqueue_task(rq
, tsk
, 0);
8369 task_rq_unlock(rq
, &flags
);
8371 #endif /* CONFIG_CGROUP_SCHED */
8373 #ifdef CONFIG_FAIR_GROUP_SCHED
8374 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8376 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8381 dequeue_entity(cfs_rq
, se
, 0);
8383 se
->load
.weight
= shares
;
8384 se
->load
.inv_weight
= 0;
8387 enqueue_entity(cfs_rq
, se
, 0);
8390 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8392 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8393 struct rq
*rq
= cfs_rq
->rq
;
8394 unsigned long flags
;
8396 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8397 __set_se_shares(se
, shares
);
8398 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8401 static DEFINE_MUTEX(shares_mutex
);
8403 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8406 unsigned long flags
;
8409 * We can't change the weight of the root cgroup.
8414 if (shares
< MIN_SHARES
)
8415 shares
= MIN_SHARES
;
8416 else if (shares
> MAX_SHARES
)
8417 shares
= MAX_SHARES
;
8419 mutex_lock(&shares_mutex
);
8420 if (tg
->shares
== shares
)
8423 spin_lock_irqsave(&task_group_lock
, flags
);
8424 for_each_possible_cpu(i
)
8425 unregister_fair_sched_group(tg
, i
);
8426 list_del_rcu(&tg
->siblings
);
8427 spin_unlock_irqrestore(&task_group_lock
, flags
);
8429 /* wait for any ongoing reference to this group to finish */
8430 synchronize_sched();
8433 * Now we are free to modify the group's share on each cpu
8434 * w/o tripping rebalance_share or load_balance_fair.
8436 tg
->shares
= shares
;
8437 for_each_possible_cpu(i
) {
8441 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8442 set_se_shares(tg
->se
[i
], shares
);
8446 * Enable load balance activity on this group, by inserting it back on
8447 * each cpu's rq->leaf_cfs_rq_list.
8449 spin_lock_irqsave(&task_group_lock
, flags
);
8450 for_each_possible_cpu(i
)
8451 register_fair_sched_group(tg
, i
);
8452 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8453 spin_unlock_irqrestore(&task_group_lock
, flags
);
8455 mutex_unlock(&shares_mutex
);
8459 unsigned long sched_group_shares(struct task_group
*tg
)
8465 #ifdef CONFIG_RT_GROUP_SCHED
8467 * Ensure that the real time constraints are schedulable.
8469 static DEFINE_MUTEX(rt_constraints_mutex
);
8471 static unsigned long to_ratio(u64 period
, u64 runtime
)
8473 if (runtime
== RUNTIME_INF
)
8476 return div64_u64(runtime
<< 20, period
);
8479 /* Must be called with tasklist_lock held */
8480 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8482 struct task_struct
*g
, *p
;
8484 do_each_thread(g
, p
) {
8485 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8487 } while_each_thread(g
, p
);
8492 struct rt_schedulable_data
{
8493 struct task_group
*tg
;
8498 static int tg_schedulable(struct task_group
*tg
, void *data
)
8500 struct rt_schedulable_data
*d
= data
;
8501 struct task_group
*child
;
8502 unsigned long total
, sum
= 0;
8503 u64 period
, runtime
;
8505 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8506 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8509 period
= d
->rt_period
;
8510 runtime
= d
->rt_runtime
;
8514 * Cannot have more runtime than the period.
8516 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8520 * Ensure we don't starve existing RT tasks.
8522 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8525 total
= to_ratio(period
, runtime
);
8528 * Nobody can have more than the global setting allows.
8530 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8534 * The sum of our children's runtime should not exceed our own.
8536 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8537 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8538 runtime
= child
->rt_bandwidth
.rt_runtime
;
8540 if (child
== d
->tg
) {
8541 period
= d
->rt_period
;
8542 runtime
= d
->rt_runtime
;
8545 sum
+= to_ratio(period
, runtime
);
8554 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8556 struct rt_schedulable_data data
= {
8558 .rt_period
= period
,
8559 .rt_runtime
= runtime
,
8562 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8565 static int tg_set_bandwidth(struct task_group
*tg
,
8566 u64 rt_period
, u64 rt_runtime
)
8570 mutex_lock(&rt_constraints_mutex
);
8571 read_lock(&tasklist_lock
);
8572 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8576 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8577 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8578 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8580 for_each_possible_cpu(i
) {
8581 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8583 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8584 rt_rq
->rt_runtime
= rt_runtime
;
8585 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8587 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8589 read_unlock(&tasklist_lock
);
8590 mutex_unlock(&rt_constraints_mutex
);
8595 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8597 u64 rt_runtime
, rt_period
;
8599 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8600 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8601 if (rt_runtime_us
< 0)
8602 rt_runtime
= RUNTIME_INF
;
8604 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8607 long sched_group_rt_runtime(struct task_group
*tg
)
8611 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8614 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8615 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8616 return rt_runtime_us
;
8619 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8621 u64 rt_runtime
, rt_period
;
8623 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8624 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8629 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8632 long sched_group_rt_period(struct task_group
*tg
)
8636 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8637 do_div(rt_period_us
, NSEC_PER_USEC
);
8638 return rt_period_us
;
8641 static int sched_rt_global_constraints(void)
8643 u64 runtime
, period
;
8646 if (sysctl_sched_rt_period
<= 0)
8649 runtime
= global_rt_runtime();
8650 period
= global_rt_period();
8653 * Sanity check on the sysctl variables.
8655 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8658 mutex_lock(&rt_constraints_mutex
);
8659 read_lock(&tasklist_lock
);
8660 ret
= __rt_schedulable(NULL
, 0, 0);
8661 read_unlock(&tasklist_lock
);
8662 mutex_unlock(&rt_constraints_mutex
);
8667 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8669 /* Don't accept realtime tasks when there is no way for them to run */
8670 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8676 #else /* !CONFIG_RT_GROUP_SCHED */
8677 static int sched_rt_global_constraints(void)
8679 unsigned long flags
;
8682 if (sysctl_sched_rt_period
<= 0)
8686 * There's always some RT tasks in the root group
8687 * -- migration, kstopmachine etc..
8689 if (sysctl_sched_rt_runtime
== 0)
8692 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8693 for_each_possible_cpu(i
) {
8694 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8696 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8697 rt_rq
->rt_runtime
= global_rt_runtime();
8698 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8700 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8704 #endif /* CONFIG_RT_GROUP_SCHED */
8706 int sched_rt_handler(struct ctl_table
*table
, int write
,
8707 void __user
*buffer
, size_t *lenp
,
8711 int old_period
, old_runtime
;
8712 static DEFINE_MUTEX(mutex
);
8715 old_period
= sysctl_sched_rt_period
;
8716 old_runtime
= sysctl_sched_rt_runtime
;
8718 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8720 if (!ret
&& write
) {
8721 ret
= sched_rt_global_constraints();
8723 sysctl_sched_rt_period
= old_period
;
8724 sysctl_sched_rt_runtime
= old_runtime
;
8726 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8727 def_rt_bandwidth
.rt_period
=
8728 ns_to_ktime(global_rt_period());
8731 mutex_unlock(&mutex
);
8736 #ifdef CONFIG_CGROUP_SCHED
8738 /* return corresponding task_group object of a cgroup */
8739 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8741 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8742 struct task_group
, css
);
8745 static struct cgroup_subsys_state
*
8746 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8748 struct task_group
*tg
, *parent
;
8750 if (!cgrp
->parent
) {
8751 /* This is early initialization for the top cgroup */
8752 return &init_task_group
.css
;
8755 parent
= cgroup_tg(cgrp
->parent
);
8756 tg
= sched_create_group(parent
);
8758 return ERR_PTR(-ENOMEM
);
8764 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8766 struct task_group
*tg
= cgroup_tg(cgrp
);
8768 sched_destroy_group(tg
);
8772 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8774 #ifdef CONFIG_RT_GROUP_SCHED
8775 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8778 /* We don't support RT-tasks being in separate groups */
8779 if (tsk
->sched_class
!= &fair_sched_class
)
8786 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8787 struct task_struct
*tsk
, bool threadgroup
)
8789 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8793 struct task_struct
*c
;
8795 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8796 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8808 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8809 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8812 sched_move_task(tsk
);
8814 struct task_struct
*c
;
8816 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8823 #ifdef CONFIG_FAIR_GROUP_SCHED
8824 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8827 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8830 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8832 struct task_group
*tg
= cgroup_tg(cgrp
);
8834 return (u64
) tg
->shares
;
8836 #endif /* CONFIG_FAIR_GROUP_SCHED */
8838 #ifdef CONFIG_RT_GROUP_SCHED
8839 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8842 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8845 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8847 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8850 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8853 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8856 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8858 return sched_group_rt_period(cgroup_tg(cgrp
));
8860 #endif /* CONFIG_RT_GROUP_SCHED */
8862 static struct cftype cpu_files
[] = {
8863 #ifdef CONFIG_FAIR_GROUP_SCHED
8866 .read_u64
= cpu_shares_read_u64
,
8867 .write_u64
= cpu_shares_write_u64
,
8870 #ifdef CONFIG_RT_GROUP_SCHED
8872 .name
= "rt_runtime_us",
8873 .read_s64
= cpu_rt_runtime_read
,
8874 .write_s64
= cpu_rt_runtime_write
,
8877 .name
= "rt_period_us",
8878 .read_u64
= cpu_rt_period_read_uint
,
8879 .write_u64
= cpu_rt_period_write_uint
,
8884 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8886 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8889 struct cgroup_subsys cpu_cgroup_subsys
= {
8891 .create
= cpu_cgroup_create
,
8892 .destroy
= cpu_cgroup_destroy
,
8893 .can_attach
= cpu_cgroup_can_attach
,
8894 .attach
= cpu_cgroup_attach
,
8895 .populate
= cpu_cgroup_populate
,
8896 .subsys_id
= cpu_cgroup_subsys_id
,
8900 #endif /* CONFIG_CGROUP_SCHED */
8902 #ifdef CONFIG_CGROUP_CPUACCT
8905 * CPU accounting code for task groups.
8907 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8908 * (balbir@in.ibm.com).
8911 /* track cpu usage of a group of tasks and its child groups */
8913 struct cgroup_subsys_state css
;
8914 /* cpuusage holds pointer to a u64-type object on every cpu */
8915 u64 __percpu
*cpuusage
;
8916 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8917 struct cpuacct
*parent
;
8920 struct cgroup_subsys cpuacct_subsys
;
8922 /* return cpu accounting group corresponding to this container */
8923 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8925 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8926 struct cpuacct
, css
);
8929 /* return cpu accounting group to which this task belongs */
8930 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8932 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8933 struct cpuacct
, css
);
8936 /* create a new cpu accounting group */
8937 static struct cgroup_subsys_state
*cpuacct_create(
8938 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8940 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8946 ca
->cpuusage
= alloc_percpu(u64
);
8950 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8951 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8952 goto out_free_counters
;
8955 ca
->parent
= cgroup_ca(cgrp
->parent
);
8961 percpu_counter_destroy(&ca
->cpustat
[i
]);
8962 free_percpu(ca
->cpuusage
);
8966 return ERR_PTR(-ENOMEM
);
8969 /* destroy an existing cpu accounting group */
8971 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8973 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8976 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8977 percpu_counter_destroy(&ca
->cpustat
[i
]);
8978 free_percpu(ca
->cpuusage
);
8982 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8984 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8987 #ifndef CONFIG_64BIT
8989 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8991 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8993 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9001 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9003 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9005 #ifndef CONFIG_64BIT
9007 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9009 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9011 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9017 /* return total cpu usage (in nanoseconds) of a group */
9018 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9020 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9021 u64 totalcpuusage
= 0;
9024 for_each_present_cpu(i
)
9025 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9027 return totalcpuusage
;
9030 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9033 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9042 for_each_present_cpu(i
)
9043 cpuacct_cpuusage_write(ca
, i
, 0);
9049 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9052 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9056 for_each_present_cpu(i
) {
9057 percpu
= cpuacct_cpuusage_read(ca
, i
);
9058 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9060 seq_printf(m
, "\n");
9064 static const char *cpuacct_stat_desc
[] = {
9065 [CPUACCT_STAT_USER
] = "user",
9066 [CPUACCT_STAT_SYSTEM
] = "system",
9069 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9070 struct cgroup_map_cb
*cb
)
9072 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9075 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9076 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9077 val
= cputime64_to_clock_t(val
);
9078 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9083 static struct cftype files
[] = {
9086 .read_u64
= cpuusage_read
,
9087 .write_u64
= cpuusage_write
,
9090 .name
= "usage_percpu",
9091 .read_seq_string
= cpuacct_percpu_seq_read
,
9095 .read_map
= cpuacct_stats_show
,
9099 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9101 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9105 * charge this task's execution time to its accounting group.
9107 * called with rq->lock held.
9109 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9114 if (unlikely(!cpuacct_subsys
.active
))
9117 cpu
= task_cpu(tsk
);
9123 for (; ca
; ca
= ca
->parent
) {
9124 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9125 *cpuusage
+= cputime
;
9132 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9133 * in cputime_t units. As a result, cpuacct_update_stats calls
9134 * percpu_counter_add with values large enough to always overflow the
9135 * per cpu batch limit causing bad SMP scalability.
9137 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9138 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9139 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9142 #define CPUACCT_BATCH \
9143 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9145 #define CPUACCT_BATCH 0
9149 * Charge the system/user time to the task's accounting group.
9151 static void cpuacct_update_stats(struct task_struct
*tsk
,
9152 enum cpuacct_stat_index idx
, cputime_t val
)
9155 int batch
= CPUACCT_BATCH
;
9157 if (unlikely(!cpuacct_subsys
.active
))
9164 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9170 struct cgroup_subsys cpuacct_subsys
= {
9172 .create
= cpuacct_create
,
9173 .destroy
= cpuacct_destroy
,
9174 .populate
= cpuacct_populate
,
9175 .subsys_id
= cpuacct_subsys_id
,
9177 #endif /* CONFIG_CGROUP_CPUACCT */
9181 void synchronize_sched_expedited(void)
9185 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9187 #else /* #ifndef CONFIG_SMP */
9189 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9191 static int synchronize_sched_expedited_cpu_stop(void *data
)
9194 * There must be a full memory barrier on each affected CPU
9195 * between the time that try_stop_cpus() is called and the
9196 * time that it returns.
9198 * In the current initial implementation of cpu_stop, the
9199 * above condition is already met when the control reaches
9200 * this point and the following smp_mb() is not strictly
9201 * necessary. Do smp_mb() anyway for documentation and
9202 * robustness against future implementation changes.
9204 smp_mb(); /* See above comment block. */
9209 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9210 * approach to force grace period to end quickly. This consumes
9211 * significant time on all CPUs, and is thus not recommended for
9212 * any sort of common-case code.
9214 * Note that it is illegal to call this function while holding any
9215 * lock that is acquired by a CPU-hotplug notifier. Failing to
9216 * observe this restriction will result in deadlock.
9218 void synchronize_sched_expedited(void)
9220 int snap
, trycount
= 0;
9222 smp_mb(); /* ensure prior mod happens before capturing snap. */
9223 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9225 while (try_stop_cpus(cpu_online_mask
,
9226 synchronize_sched_expedited_cpu_stop
,
9229 if (trycount
++ < 10)
9230 udelay(trycount
* num_online_cpus());
9232 synchronize_sched();
9235 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9236 smp_mb(); /* ensure test happens before caller kfree */
9241 atomic_inc(&synchronize_sched_expedited_count
);
9242 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9245 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9247 #endif /* #else #ifndef CONFIG_SMP */