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
, *stop
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
498 struct root_domain
*rd
;
499 struct sched_domain
*sd
;
501 unsigned long cpu_power
;
503 unsigned char idle_at_tick
;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work
;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task
;
521 /* calc_load related fields */
522 unsigned long calc_load_update
;
523 long calc_load_active
;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending
;
528 struct call_single_data hrtick_csd
;
530 struct hrtimer hrtick_timer
;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info
;
536 unsigned long long rq_cpu_time
;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count
;
542 /* schedule() stats */
543 unsigned int sched_switch
;
544 unsigned int sched_count
;
545 unsigned int sched_goidle
;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count
;
549 unsigned int ttwu_local
;
552 unsigned int bkl_count
;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
559 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
561 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p
))
568 rq
->skip_clock_update
= 1;
571 static inline int cpu_of(struct rq
*rq
)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group
*task_group(struct task_struct
*p
)
613 struct cgroup_subsys_state
*css
;
615 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
616 lockdep_is_held(&task_rq(p
)->lock
));
617 return container_of(css
, struct task_group
, css
);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
625 p
->se
.parent
= task_group(p
)->se
[cpu
];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
630 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
637 static inline struct task_group
*task_group(struct task_struct
*p
)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq
*rq
)
646 if (!rq
->skip_clock_update
)
647 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu
)
669 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug
unsigned int sysctl_sched_features
=
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly
char *sched_feat_names
[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file
*m
, void *v
)
709 for (i
= 0; sched_feat_names
[i
]; i
++) {
710 if (!(sysctl_sched_features
& (1UL << i
)))
712 seq_printf(m
, "%s ", sched_feat_names
[i
]);
720 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
721 size_t cnt
, loff_t
*ppos
)
731 if (copy_from_user(&buf
, ubuf
, cnt
))
737 if (strncmp(buf
, "NO_", 3) == 0) {
742 for (i
= 0; sched_feat_names
[i
]; i
++) {
743 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
745 sysctl_sched_features
&= ~(1UL << i
);
747 sysctl_sched_features
|= (1UL << i
);
752 if (!sched_feat_names
[i
])
760 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
762 return single_open(filp
, sched_feat_show
, NULL
);
765 static const struct file_operations sched_feat_fops
= {
766 .open
= sched_feat_open
,
767 .write
= sched_feat_write
,
770 .release
= single_release
,
773 static __init
int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
780 late_initcall(sched_init_debug
);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit
= 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh
= 4;
807 * period over which we average the RT time consumption, measured
812 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period
= 1000000;
820 static __read_mostly
int scheduler_running
;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime
= 950000;
828 static inline u64
global_rt_period(void)
830 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
833 static inline u64
global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime
< 0)
838 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
850 return rq
->curr
== p
;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
856 return task_current(rq
, p
);
859 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
863 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq
->lock
.owner
= current
;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
876 raw_spin_unlock_irq(&rq
->lock
);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
885 return task_current(rq
, p
);
889 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq
->lock
);
902 raw_spin_unlock(&rq
->lock
);
906 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct
*p
)
929 return unlikely(p
->state
== TASK_WAKING
);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
943 raw_spin_lock(&rq
->lock
);
944 if (likely(rq
== task_rq(p
)))
946 raw_spin_unlock(&rq
->lock
);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
961 local_irq_save(*flags
);
963 raw_spin_lock(&rq
->lock
);
964 if (likely(rq
== task_rq(p
)))
966 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
970 static void __task_rq_unlock(struct rq
*rq
)
973 raw_spin_unlock(&rq
->lock
);
976 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
979 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq
*this_rq_lock(void)
992 raw_spin_lock(&rq
->lock
);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq
*rq
)
1016 if (!sched_feat(HRTICK
))
1018 if (!cpu_active(cpu_of(rq
)))
1020 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1023 static void hrtick_clear(struct rq
*rq
)
1025 if (hrtimer_active(&rq
->hrtick_timer
))
1026 hrtimer_cancel(&rq
->hrtick_timer
);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1035 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1037 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1039 raw_spin_lock(&rq
->lock
);
1040 update_rq_clock(rq
);
1041 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1042 raw_spin_unlock(&rq
->lock
);
1044 return HRTIMER_NORESTART
;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg
)
1053 struct rq
*rq
= arg
;
1055 raw_spin_lock(&rq
->lock
);
1056 hrtimer_restart(&rq
->hrtick_timer
);
1057 rq
->hrtick_csd_pending
= 0;
1058 raw_spin_unlock(&rq
->lock
);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq
*rq
, u64 delay
)
1068 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1069 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1071 hrtimer_set_expires(timer
, time
);
1073 if (rq
== this_rq()) {
1074 hrtimer_restart(timer
);
1075 } else if (!rq
->hrtick_csd_pending
) {
1076 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1077 rq
->hrtick_csd_pending
= 1;
1082 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1084 int cpu
= (int)(long)hcpu
;
1087 case CPU_UP_CANCELED
:
1088 case CPU_UP_CANCELED_FROZEN
:
1089 case CPU_DOWN_PREPARE
:
1090 case CPU_DOWN_PREPARE_FROZEN
:
1092 case CPU_DEAD_FROZEN
:
1093 hrtick_clear(cpu_rq(cpu
));
1100 static __init
void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick
, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq
*rq
, u64 delay
)
1112 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1113 HRTIMER_MODE_REL_PINNED
, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq
*rq
)
1124 rq
->hrtick_csd_pending
= 0;
1126 rq
->hrtick_csd
.flags
= 0;
1127 rq
->hrtick_csd
.func
= __hrtick_start
;
1128 rq
->hrtick_csd
.info
= rq
;
1131 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1132 rq
->hrtick_timer
.function
= hrtick
;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq
*rq
)
1139 static inline void init_rq_hrtick(struct rq
*rq
)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct
*p
)
1165 assert_raw_spin_locked(&task_rq(p
)->lock
);
1167 if (test_tsk_need_resched(p
))
1170 set_tsk_need_resched(p
);
1173 if (cpu
== smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p
))
1179 smp_send_reschedule(cpu
);
1182 static void resched_cpu(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long flags
;
1187 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1189 resched_task(cpu_curr(cpu
));
1190 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu
= smp_processor_id();
1206 struct sched_domain
*sd
;
1208 for_each_domain(cpu
, sd
) {
1209 for_each_cpu(i
, sched_domain_span(sd
))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu
)
1227 struct rq
*rq
= cpu_rq(cpu
);
1229 if (cpu
== smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq
->curr
!= rq
->idle
)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq
->idle
);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq
->idle
))
1252 smp_send_reschedule(cpu
);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64
sched_avg_period(void)
1259 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1262 static void sched_avg_update(struct rq
*rq
)
1264 s64 period
= sched_avg_period();
1266 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq
->age_stamp
));
1273 rq
->age_stamp
+= period
;
1278 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1280 rq
->rt_avg
+= rt_delta
;
1281 sched_avg_update(rq
);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct
*p
)
1287 assert_raw_spin_locked(&task_rq(p
)->lock
);
1288 set_tsk_need_resched(p
);
1291 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1295 static void sched_avg_update(struct rq
*rq
)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1318 struct load_weight
*lw
)
1322 if (!lw
->inv_weight
) {
1323 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1326 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1330 tmp
= (u64
)delta_exec
* weight
;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp
> WMULT_CONST
))
1335 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1338 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1340 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1343 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1349 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1356 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1357 * of tasks with abnormal "nice" values across CPUs the contribution that
1358 * each task makes to its run queue's load is weighted according to its
1359 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1360 * scaled version of the new time slice allocation that they receive on time
1364 #define WEIGHT_IDLEPRIO 3
1365 #define WMULT_IDLEPRIO 1431655765
1368 * Nice levels are multiplicative, with a gentle 10% change for every
1369 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1370 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1371 * that remained on nice 0.
1373 * The "10% effect" is relative and cumulative: from _any_ nice level,
1374 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1375 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1376 * If a task goes up by ~10% and another task goes down by ~10% then
1377 * the relative distance between them is ~25%.)
1379 static const int prio_to_weight
[40] = {
1380 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1381 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1382 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1383 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1384 /* 0 */ 1024, 820, 655, 526, 423,
1385 /* 5 */ 335, 272, 215, 172, 137,
1386 /* 10 */ 110, 87, 70, 56, 45,
1387 /* 15 */ 36, 29, 23, 18, 15,
1391 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1393 * In cases where the weight does not change often, we can use the
1394 * precalculated inverse to speed up arithmetics by turning divisions
1395 * into multiplications:
1397 static const u32 prio_to_wmult
[40] = {
1398 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1399 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1400 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1401 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1402 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1403 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1404 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1405 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1408 /* Time spent by the tasks of the cpu accounting group executing in ... */
1409 enum cpuacct_stat_index
{
1410 CPUACCT_STAT_USER
, /* ... user mode */
1411 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1413 CPUACCT_STAT_NSTATS
,
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1418 static void cpuacct_update_stats(struct task_struct
*tsk
,
1419 enum cpuacct_stat_index idx
, cputime_t val
);
1421 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1422 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1423 enum cpuacct_stat_index idx
, cputime_t val
) {}
1426 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1428 update_load_add(&rq
->load
, load
);
1431 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1433 update_load_sub(&rq
->load
, load
);
1436 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1437 typedef int (*tg_visitor
)(struct task_group
*, void *);
1440 * Iterate the full tree, calling @down when first entering a node and @up when
1441 * leaving it for the final time.
1443 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1445 struct task_group
*parent
, *child
;
1449 parent
= &root_task_group
;
1451 ret
= (*down
)(parent
, data
);
1454 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1461 ret
= (*up
)(parent
, data
);
1466 parent
= parent
->parent
;
1475 static int tg_nop(struct task_group
*tg
, void *data
)
1482 /* Used instead of source_load when we know the type == 0 */
1483 static unsigned long weighted_cpuload(const int cpu
)
1485 return cpu_rq(cpu
)->load
.weight
;
1489 * Return a low guess at the load of a migration-source cpu weighted
1490 * according to the scheduling class and "nice" value.
1492 * We want to under-estimate the load of migration sources, to
1493 * balance conservatively.
1495 static unsigned long source_load(int cpu
, int type
)
1497 struct rq
*rq
= cpu_rq(cpu
);
1498 unsigned long total
= weighted_cpuload(cpu
);
1500 if (type
== 0 || !sched_feat(LB_BIAS
))
1503 return min(rq
->cpu_load
[type
-1], total
);
1507 * Return a high guess at the load of a migration-target cpu weighted
1508 * according to the scheduling class and "nice" value.
1510 static unsigned long target_load(int cpu
, int type
)
1512 struct rq
*rq
= cpu_rq(cpu
);
1513 unsigned long total
= weighted_cpuload(cpu
);
1515 if (type
== 0 || !sched_feat(LB_BIAS
))
1518 return max(rq
->cpu_load
[type
-1], total
);
1521 static unsigned long power_of(int cpu
)
1523 return cpu_rq(cpu
)->cpu_power
;
1526 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1528 static unsigned long cpu_avg_load_per_task(int cpu
)
1530 struct rq
*rq
= cpu_rq(cpu
);
1531 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1534 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1536 rq
->avg_load_per_task
= 0;
1538 return rq
->avg_load_per_task
;
1541 #ifdef CONFIG_FAIR_GROUP_SCHED
1543 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1545 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1548 * Calculate and set the cpu's group shares.
1550 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1551 unsigned long sd_shares
,
1552 unsigned long sd_rq_weight
,
1553 unsigned long *usd_rq_weight
)
1555 unsigned long shares
, rq_weight
;
1558 rq_weight
= usd_rq_weight
[cpu
];
1561 rq_weight
= NICE_0_LOAD
;
1565 * \Sum_j shares_j * rq_weight_i
1566 * shares_i = -----------------------------
1567 * \Sum_j rq_weight_j
1569 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1570 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1572 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1573 sysctl_sched_shares_thresh
) {
1574 struct rq
*rq
= cpu_rq(cpu
);
1575 unsigned long flags
;
1577 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1578 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1579 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1580 __set_se_shares(tg
->se
[cpu
], shares
);
1581 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1586 * Re-compute the task group their per cpu shares over the given domain.
1587 * This needs to be done in a bottom-up fashion because the rq weight of a
1588 * parent group depends on the shares of its child groups.
1590 static int tg_shares_up(struct task_group
*tg
, void *data
)
1592 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1593 unsigned long *usd_rq_weight
;
1594 struct sched_domain
*sd
= data
;
1595 unsigned long flags
;
1601 local_irq_save(flags
);
1602 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1604 for_each_cpu(i
, sched_domain_span(sd
)) {
1605 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1606 usd_rq_weight
[i
] = weight
;
1608 rq_weight
+= weight
;
1610 * If there are currently no tasks on the cpu pretend there
1611 * is one of average load so that when a new task gets to
1612 * run here it will not get delayed by group starvation.
1615 weight
= NICE_0_LOAD
;
1617 sum_weight
+= weight
;
1618 shares
+= tg
->cfs_rq
[i
]->shares
;
1622 rq_weight
= sum_weight
;
1624 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1625 shares
= tg
->shares
;
1627 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1628 shares
= tg
->shares
;
1630 for_each_cpu(i
, sched_domain_span(sd
))
1631 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1633 local_irq_restore(flags
);
1639 * Compute the cpu's hierarchical load factor for each task group.
1640 * This needs to be done in a top-down fashion because the load of a child
1641 * group is a fraction of its parents load.
1643 static int tg_load_down(struct task_group
*tg
, void *data
)
1646 long cpu
= (long)data
;
1649 load
= cpu_rq(cpu
)->load
.weight
;
1651 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1652 load
*= tg
->cfs_rq
[cpu
]->shares
;
1653 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1656 tg
->cfs_rq
[cpu
]->h_load
= load
;
1661 static void update_shares(struct sched_domain
*sd
)
1666 if (root_task_group_empty())
1669 now
= local_clock();
1670 elapsed
= now
- sd
->last_update
;
1672 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1673 sd
->last_update
= now
;
1674 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1678 static void update_h_load(long cpu
)
1680 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1685 static inline void update_shares(struct sched_domain
*sd
)
1691 #ifdef CONFIG_PREEMPT
1693 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1696 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1697 * way at the expense of forcing extra atomic operations in all
1698 * invocations. This assures that the double_lock is acquired using the
1699 * same underlying policy as the spinlock_t on this architecture, which
1700 * reduces latency compared to the unfair variant below. However, it
1701 * also adds more overhead and therefore may reduce throughput.
1703 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1704 __releases(this_rq
->lock
)
1705 __acquires(busiest
->lock
)
1706 __acquires(this_rq
->lock
)
1708 raw_spin_unlock(&this_rq
->lock
);
1709 double_rq_lock(this_rq
, busiest
);
1716 * Unfair double_lock_balance: Optimizes throughput at the expense of
1717 * latency by eliminating extra atomic operations when the locks are
1718 * already in proper order on entry. This favors lower cpu-ids and will
1719 * grant the double lock to lower cpus over higher ids under contention,
1720 * regardless of entry order into the function.
1722 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1723 __releases(this_rq
->lock
)
1724 __acquires(busiest
->lock
)
1725 __acquires(this_rq
->lock
)
1729 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1730 if (busiest
< this_rq
) {
1731 raw_spin_unlock(&this_rq
->lock
);
1732 raw_spin_lock(&busiest
->lock
);
1733 raw_spin_lock_nested(&this_rq
->lock
,
1734 SINGLE_DEPTH_NESTING
);
1737 raw_spin_lock_nested(&busiest
->lock
,
1738 SINGLE_DEPTH_NESTING
);
1743 #endif /* CONFIG_PREEMPT */
1746 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1748 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1750 if (unlikely(!irqs_disabled())) {
1751 /* printk() doesn't work good under rq->lock */
1752 raw_spin_unlock(&this_rq
->lock
);
1756 return _double_lock_balance(this_rq
, busiest
);
1759 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1760 __releases(busiest
->lock
)
1762 raw_spin_unlock(&busiest
->lock
);
1763 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1767 * double_rq_lock - safely lock two runqueues
1769 * Note this does not disable interrupts like task_rq_lock,
1770 * you need to do so manually before calling.
1772 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1773 __acquires(rq1
->lock
)
1774 __acquires(rq2
->lock
)
1776 BUG_ON(!irqs_disabled());
1778 raw_spin_lock(&rq1
->lock
);
1779 __acquire(rq2
->lock
); /* Fake it out ;) */
1782 raw_spin_lock(&rq1
->lock
);
1783 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1785 raw_spin_lock(&rq2
->lock
);
1786 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1792 * double_rq_unlock - safely unlock two runqueues
1794 * Note this does not restore interrupts like task_rq_unlock,
1795 * you need to do so manually after calling.
1797 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1798 __releases(rq1
->lock
)
1799 __releases(rq2
->lock
)
1801 raw_spin_unlock(&rq1
->lock
);
1803 raw_spin_unlock(&rq2
->lock
);
1805 __release(rq2
->lock
);
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1814 cfs_rq
->shares
= shares
;
1819 static void calc_load_account_idle(struct rq
*this_rq
);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1822 static void update_cpu_load(struct rq
*this_rq
);
1824 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1826 set_task_rq(p
, cpu
);
1829 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1830 * successfuly executed on another CPU. We must ensure that updates of
1831 * per-task data have been completed by this moment.
1834 task_thread_info(p
)->cpu
= cpu
;
1838 static const struct sched_class rt_sched_class
;
1840 #define sched_class_highest (&stop_sched_class)
1841 #define for_each_class(class) \
1842 for (class = sched_class_highest; class; class = class->next)
1844 #include "sched_stats.h"
1846 static void inc_nr_running(struct rq
*rq
)
1851 static void dec_nr_running(struct rq
*rq
)
1856 static void set_load_weight(struct task_struct
*p
)
1859 * SCHED_IDLE tasks get minimal weight:
1861 if (p
->policy
== SCHED_IDLE
) {
1862 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1863 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1867 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1868 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1871 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1873 update_rq_clock(rq
);
1874 sched_info_queued(p
);
1875 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1879 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1881 update_rq_clock(rq
);
1882 sched_info_dequeued(p
);
1883 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1888 * activate_task - move a task to the runqueue.
1890 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1892 if (task_contributes_to_load(p
))
1893 rq
->nr_uninterruptible
--;
1895 enqueue_task(rq
, p
, flags
);
1900 * deactivate_task - remove a task from the runqueue.
1902 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1904 if (task_contributes_to_load(p
))
1905 rq
->nr_uninterruptible
++;
1907 dequeue_task(rq
, p
, flags
);
1911 #include "sched_idletask.c"
1912 #include "sched_fair.c"
1913 #include "sched_rt.c"
1914 #include "sched_stoptask.c"
1915 #ifdef CONFIG_SCHED_DEBUG
1916 # include "sched_debug.c"
1919 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1921 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1922 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1926 * Make it appear like a SCHED_FIFO task, its something
1927 * userspace knows about and won't get confused about.
1929 * Also, it will make PI more or less work without too
1930 * much confusion -- but then, stop work should not
1931 * rely on PI working anyway.
1933 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1935 stop
->sched_class
= &stop_sched_class
;
1938 cpu_rq(cpu
)->stop
= stop
;
1942 * Reset it back to a normal scheduling class so that
1943 * it can die in pieces.
1945 old_stop
->sched_class
= &rt_sched_class
;
1950 * __normal_prio - return the priority that is based on the static prio
1952 static inline int __normal_prio(struct task_struct
*p
)
1954 return p
->static_prio
;
1958 * Calculate the expected normal priority: i.e. priority
1959 * without taking RT-inheritance into account. Might be
1960 * boosted by interactivity modifiers. Changes upon fork,
1961 * setprio syscalls, and whenever the interactivity
1962 * estimator recalculates.
1964 static inline int normal_prio(struct task_struct
*p
)
1968 if (task_has_rt_policy(p
))
1969 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1971 prio
= __normal_prio(p
);
1976 * Calculate the current priority, i.e. the priority
1977 * taken into account by the scheduler. This value might
1978 * be boosted by RT tasks, or might be boosted by
1979 * interactivity modifiers. Will be RT if the task got
1980 * RT-boosted. If not then it returns p->normal_prio.
1982 static int effective_prio(struct task_struct
*p
)
1984 p
->normal_prio
= normal_prio(p
);
1986 * If we are RT tasks or we were boosted to RT priority,
1987 * keep the priority unchanged. Otherwise, update priority
1988 * to the normal priority:
1990 if (!rt_prio(p
->prio
))
1991 return p
->normal_prio
;
1996 * task_curr - is this task currently executing on a CPU?
1997 * @p: the task in question.
1999 inline int task_curr(const struct task_struct
*p
)
2001 return cpu_curr(task_cpu(p
)) == p
;
2004 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2005 const struct sched_class
*prev_class
,
2006 int oldprio
, int running
)
2008 if (prev_class
!= p
->sched_class
) {
2009 if (prev_class
->switched_from
)
2010 prev_class
->switched_from(rq
, p
, running
);
2011 p
->sched_class
->switched_to(rq
, p
, running
);
2013 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2018 * Is this task likely cache-hot:
2021 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2025 if (p
->sched_class
!= &fair_sched_class
)
2028 if (unlikely(p
->policy
== SCHED_IDLE
))
2032 * Buddy candidates are cache hot:
2034 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2035 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2036 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2039 if (sysctl_sched_migration_cost
== -1)
2041 if (sysctl_sched_migration_cost
== 0)
2044 delta
= now
- p
->se
.exec_start
;
2046 return delta
< (s64
)sysctl_sched_migration_cost
;
2049 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2051 #ifdef CONFIG_SCHED_DEBUG
2053 * We should never call set_task_cpu() on a blocked task,
2054 * ttwu() will sort out the placement.
2056 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2057 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2060 trace_sched_migrate_task(p
, new_cpu
);
2062 if (task_cpu(p
) != new_cpu
) {
2063 p
->se
.nr_migrations
++;
2064 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2067 __set_task_cpu(p
, new_cpu
);
2070 struct migration_arg
{
2071 struct task_struct
*task
;
2075 static int migration_cpu_stop(void *data
);
2078 * The task's runqueue lock must be held.
2079 * Returns true if you have to wait for migration thread.
2081 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2083 struct rq
*rq
= task_rq(p
);
2086 * If the task is not on a runqueue (and not running), then
2087 * the next wake-up will properly place the task.
2089 return p
->se
.on_rq
|| task_running(rq
, p
);
2093 * wait_task_inactive - wait for a thread to unschedule.
2095 * If @match_state is nonzero, it's the @p->state value just checked and
2096 * not expected to change. If it changes, i.e. @p might have woken up,
2097 * then return zero. When we succeed in waiting for @p to be off its CPU,
2098 * we return a positive number (its total switch count). If a second call
2099 * a short while later returns the same number, the caller can be sure that
2100 * @p has remained unscheduled the whole time.
2102 * The caller must ensure that the task *will* unschedule sometime soon,
2103 * else this function might spin for a *long* time. This function can't
2104 * be called with interrupts off, or it may introduce deadlock with
2105 * smp_call_function() if an IPI is sent by the same process we are
2106 * waiting to become inactive.
2108 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2110 unsigned long flags
;
2117 * We do the initial early heuristics without holding
2118 * any task-queue locks at all. We'll only try to get
2119 * the runqueue lock when things look like they will
2125 * If the task is actively running on another CPU
2126 * still, just relax and busy-wait without holding
2129 * NOTE! Since we don't hold any locks, it's not
2130 * even sure that "rq" stays as the right runqueue!
2131 * But we don't care, since "task_running()" will
2132 * return false if the runqueue has changed and p
2133 * is actually now running somewhere else!
2135 while (task_running(rq
, p
)) {
2136 if (match_state
&& unlikely(p
->state
!= match_state
))
2142 * Ok, time to look more closely! We need the rq
2143 * lock now, to be *sure*. If we're wrong, we'll
2144 * just go back and repeat.
2146 rq
= task_rq_lock(p
, &flags
);
2147 trace_sched_wait_task(p
);
2148 running
= task_running(rq
, p
);
2149 on_rq
= p
->se
.on_rq
;
2151 if (!match_state
|| p
->state
== match_state
)
2152 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2153 task_rq_unlock(rq
, &flags
);
2156 * If it changed from the expected state, bail out now.
2158 if (unlikely(!ncsw
))
2162 * Was it really running after all now that we
2163 * checked with the proper locks actually held?
2165 * Oops. Go back and try again..
2167 if (unlikely(running
)) {
2173 * It's not enough that it's not actively running,
2174 * it must be off the runqueue _entirely_, and not
2177 * So if it was still runnable (but just not actively
2178 * running right now), it's preempted, and we should
2179 * yield - it could be a while.
2181 if (unlikely(on_rq
)) {
2182 schedule_timeout_uninterruptible(1);
2187 * Ahh, all good. It wasn't running, and it wasn't
2188 * runnable, which means that it will never become
2189 * running in the future either. We're all done!
2198 * kick_process - kick a running thread to enter/exit the kernel
2199 * @p: the to-be-kicked thread
2201 * Cause a process which is running on another CPU to enter
2202 * kernel-mode, without any delay. (to get signals handled.)
2204 * NOTE: this function doesnt have to take the runqueue lock,
2205 * because all it wants to ensure is that the remote task enters
2206 * the kernel. If the IPI races and the task has been migrated
2207 * to another CPU then no harm is done and the purpose has been
2210 void kick_process(struct task_struct
*p
)
2216 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2217 smp_send_reschedule(cpu
);
2220 EXPORT_SYMBOL_GPL(kick_process
);
2221 #endif /* CONFIG_SMP */
2224 * task_oncpu_function_call - call a function on the cpu on which a task runs
2225 * @p: the task to evaluate
2226 * @func: the function to be called
2227 * @info: the function call argument
2229 * Calls the function @func when the task is currently running. This might
2230 * be on the current CPU, which just calls the function directly
2232 void task_oncpu_function_call(struct task_struct
*p
,
2233 void (*func
) (void *info
), void *info
)
2240 smp_call_function_single(cpu
, func
, info
, 1);
2246 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2248 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2251 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2253 /* Look for allowed, online CPU in same node. */
2254 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2255 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2258 /* Any allowed, online CPU? */
2259 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2260 if (dest_cpu
< nr_cpu_ids
)
2263 /* No more Mr. Nice Guy. */
2264 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2265 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2267 * Don't tell them about moving exiting tasks or
2268 * kernel threads (both mm NULL), since they never
2271 if (p
->mm
&& printk_ratelimit()) {
2272 printk(KERN_INFO
"process %d (%s) no "
2273 "longer affine to cpu%d\n",
2274 task_pid_nr(p
), p
->comm
, cpu
);
2282 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2285 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2287 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2290 * In order not to call set_task_cpu() on a blocking task we need
2291 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2294 * Since this is common to all placement strategies, this lives here.
2296 * [ this allows ->select_task() to simply return task_cpu(p) and
2297 * not worry about this generic constraint ]
2299 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2301 cpu
= select_fallback_rq(task_cpu(p
), p
);
2306 static void update_avg(u64
*avg
, u64 sample
)
2308 s64 diff
= sample
- *avg
;
2313 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2314 bool is_sync
, bool is_migrate
, bool is_local
,
2315 unsigned long en_flags
)
2317 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2319 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2321 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2323 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2325 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2327 activate_task(rq
, p
, en_flags
);
2330 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2331 int wake_flags
, bool success
)
2333 trace_sched_wakeup(p
, success
);
2334 check_preempt_curr(rq
, p
, wake_flags
);
2336 p
->state
= TASK_RUNNING
;
2338 if (p
->sched_class
->task_woken
)
2339 p
->sched_class
->task_woken(rq
, p
);
2341 if (unlikely(rq
->idle_stamp
)) {
2342 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2343 u64 max
= 2*sysctl_sched_migration_cost
;
2348 update_avg(&rq
->avg_idle
, delta
);
2352 /* if a worker is waking up, notify workqueue */
2353 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2354 wq_worker_waking_up(p
, cpu_of(rq
));
2358 * try_to_wake_up - wake up a thread
2359 * @p: the thread to be awakened
2360 * @state: the mask of task states that can be woken
2361 * @wake_flags: wake modifier flags (WF_*)
2363 * Put it on the run-queue if it's not already there. The "current"
2364 * thread is always on the run-queue (except when the actual
2365 * re-schedule is in progress), and as such you're allowed to do
2366 * the simpler "current->state = TASK_RUNNING" to mark yourself
2367 * runnable without the overhead of this.
2369 * Returns %true if @p was woken up, %false if it was already running
2370 * or @state didn't match @p's state.
2372 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2375 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2376 unsigned long flags
;
2377 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2380 this_cpu
= get_cpu();
2383 rq
= task_rq_lock(p
, &flags
);
2384 if (!(p
->state
& state
))
2394 if (unlikely(task_running(rq
, p
)))
2398 * In order to handle concurrent wakeups and release the rq->lock
2399 * we put the task in TASK_WAKING state.
2401 * First fix up the nr_uninterruptible count:
2403 if (task_contributes_to_load(p
)) {
2404 if (likely(cpu_online(orig_cpu
)))
2405 rq
->nr_uninterruptible
--;
2407 this_rq()->nr_uninterruptible
--;
2409 p
->state
= TASK_WAKING
;
2411 if (p
->sched_class
->task_waking
) {
2412 p
->sched_class
->task_waking(rq
, p
);
2413 en_flags
|= ENQUEUE_WAKING
;
2416 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2417 if (cpu
!= orig_cpu
)
2418 set_task_cpu(p
, cpu
);
2419 __task_rq_unlock(rq
);
2422 raw_spin_lock(&rq
->lock
);
2425 * We migrated the task without holding either rq->lock, however
2426 * since the task is not on the task list itself, nobody else
2427 * will try and migrate the task, hence the rq should match the
2428 * cpu we just moved it to.
2430 WARN_ON(task_cpu(p
) != cpu
);
2431 WARN_ON(p
->state
!= TASK_WAKING
);
2433 #ifdef CONFIG_SCHEDSTATS
2434 schedstat_inc(rq
, ttwu_count
);
2435 if (cpu
== this_cpu
)
2436 schedstat_inc(rq
, ttwu_local
);
2438 struct sched_domain
*sd
;
2439 for_each_domain(this_cpu
, sd
) {
2440 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2441 schedstat_inc(sd
, ttwu_wake_remote
);
2446 #endif /* CONFIG_SCHEDSTATS */
2449 #endif /* CONFIG_SMP */
2450 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2451 cpu
== this_cpu
, en_flags
);
2454 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2456 task_rq_unlock(rq
, &flags
);
2463 * try_to_wake_up_local - try to wake up a local task with rq lock held
2464 * @p: the thread to be awakened
2466 * Put @p on the run-queue if it's not alredy there. The caller must
2467 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2468 * the current task. this_rq() stays locked over invocation.
2470 static void try_to_wake_up_local(struct task_struct
*p
)
2472 struct rq
*rq
= task_rq(p
);
2473 bool success
= false;
2475 BUG_ON(rq
!= this_rq());
2476 BUG_ON(p
== current
);
2477 lockdep_assert_held(&rq
->lock
);
2479 if (!(p
->state
& TASK_NORMAL
))
2483 if (likely(!task_running(rq
, p
))) {
2484 schedstat_inc(rq
, ttwu_count
);
2485 schedstat_inc(rq
, ttwu_local
);
2487 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2490 ttwu_post_activation(p
, rq
, 0, success
);
2494 * wake_up_process - Wake up a specific process
2495 * @p: The process to be woken up.
2497 * Attempt to wake up the nominated process and move it to the set of runnable
2498 * processes. Returns 1 if the process was woken up, 0 if it was already
2501 * It may be assumed that this function implies a write memory barrier before
2502 * changing the task state if and only if any tasks are woken up.
2504 int wake_up_process(struct task_struct
*p
)
2506 return try_to_wake_up(p
, TASK_ALL
, 0);
2508 EXPORT_SYMBOL(wake_up_process
);
2510 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2512 return try_to_wake_up(p
, state
, 0);
2516 * Perform scheduler related setup for a newly forked process p.
2517 * p is forked by current.
2519 * __sched_fork() is basic setup used by init_idle() too:
2521 static void __sched_fork(struct task_struct
*p
)
2523 p
->se
.exec_start
= 0;
2524 p
->se
.sum_exec_runtime
= 0;
2525 p
->se
.prev_sum_exec_runtime
= 0;
2526 p
->se
.nr_migrations
= 0;
2528 #ifdef CONFIG_SCHEDSTATS
2529 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2532 INIT_LIST_HEAD(&p
->rt
.run_list
);
2534 INIT_LIST_HEAD(&p
->se
.group_node
);
2536 #ifdef CONFIG_PREEMPT_NOTIFIERS
2537 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2542 * fork()/clone()-time setup:
2544 void sched_fork(struct task_struct
*p
, int clone_flags
)
2546 int cpu
= get_cpu();
2550 * We mark the process as running here. This guarantees that
2551 * nobody will actually run it, and a signal or other external
2552 * event cannot wake it up and insert it on the runqueue either.
2554 p
->state
= TASK_RUNNING
;
2557 * Revert to default priority/policy on fork if requested.
2559 if (unlikely(p
->sched_reset_on_fork
)) {
2560 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2561 p
->policy
= SCHED_NORMAL
;
2562 p
->normal_prio
= p
->static_prio
;
2565 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2566 p
->static_prio
= NICE_TO_PRIO(0);
2567 p
->normal_prio
= p
->static_prio
;
2572 * We don't need the reset flag anymore after the fork. It has
2573 * fulfilled its duty:
2575 p
->sched_reset_on_fork
= 0;
2579 * Make sure we do not leak PI boosting priority to the child.
2581 p
->prio
= current
->normal_prio
;
2583 if (!rt_prio(p
->prio
))
2584 p
->sched_class
= &fair_sched_class
;
2586 if (p
->sched_class
->task_fork
)
2587 p
->sched_class
->task_fork(p
);
2590 * The child is not yet in the pid-hash so no cgroup attach races,
2591 * and the cgroup is pinned to this child due to cgroup_fork()
2592 * is ran before sched_fork().
2594 * Silence PROVE_RCU.
2597 set_task_cpu(p
, cpu
);
2600 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2601 if (likely(sched_info_on()))
2602 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2604 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2607 #ifdef CONFIG_PREEMPT
2608 /* Want to start with kernel preemption disabled. */
2609 task_thread_info(p
)->preempt_count
= 1;
2611 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2617 * wake_up_new_task - wake up a newly created task for the first time.
2619 * This function will do some initial scheduler statistics housekeeping
2620 * that must be done for every newly created context, then puts the task
2621 * on the runqueue and wakes it.
2623 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2625 unsigned long flags
;
2627 int cpu __maybe_unused
= get_cpu();
2630 rq
= task_rq_lock(p
, &flags
);
2631 p
->state
= TASK_WAKING
;
2634 * Fork balancing, do it here and not earlier because:
2635 * - cpus_allowed can change in the fork path
2636 * - any previously selected cpu might disappear through hotplug
2638 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2639 * without people poking at ->cpus_allowed.
2641 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2642 set_task_cpu(p
, cpu
);
2644 p
->state
= TASK_RUNNING
;
2645 task_rq_unlock(rq
, &flags
);
2648 rq
= task_rq_lock(p
, &flags
);
2649 activate_task(rq
, p
, 0);
2650 trace_sched_wakeup_new(p
, 1);
2651 check_preempt_curr(rq
, p
, WF_FORK
);
2653 if (p
->sched_class
->task_woken
)
2654 p
->sched_class
->task_woken(rq
, p
);
2656 task_rq_unlock(rq
, &flags
);
2660 #ifdef CONFIG_PREEMPT_NOTIFIERS
2663 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2664 * @notifier: notifier struct to register
2666 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2668 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2670 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2673 * preempt_notifier_unregister - no longer interested in preemption notifications
2674 * @notifier: notifier struct to unregister
2676 * This is safe to call from within a preemption notifier.
2678 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2680 hlist_del(¬ifier
->link
);
2682 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2684 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2686 struct preempt_notifier
*notifier
;
2687 struct hlist_node
*node
;
2689 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2690 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2694 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2695 struct task_struct
*next
)
2697 struct preempt_notifier
*notifier
;
2698 struct hlist_node
*node
;
2700 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2701 notifier
->ops
->sched_out(notifier
, next
);
2704 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2706 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2711 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2712 struct task_struct
*next
)
2716 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2719 * prepare_task_switch - prepare to switch tasks
2720 * @rq: the runqueue preparing to switch
2721 * @prev: the current task that is being switched out
2722 * @next: the task we are going to switch to.
2724 * This is called with the rq lock held and interrupts off. It must
2725 * be paired with a subsequent finish_task_switch after the context
2728 * prepare_task_switch sets up locking and calls architecture specific
2732 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2733 struct task_struct
*next
)
2735 fire_sched_out_preempt_notifiers(prev
, next
);
2736 prepare_lock_switch(rq
, next
);
2737 prepare_arch_switch(next
);
2741 * finish_task_switch - clean up after a task-switch
2742 * @rq: runqueue associated with task-switch
2743 * @prev: the thread we just switched away from.
2745 * finish_task_switch must be called after the context switch, paired
2746 * with a prepare_task_switch call before the context switch.
2747 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2748 * and do any other architecture-specific cleanup actions.
2750 * Note that we may have delayed dropping an mm in context_switch(). If
2751 * so, we finish that here outside of the runqueue lock. (Doing it
2752 * with the lock held can cause deadlocks; see schedule() for
2755 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2756 __releases(rq
->lock
)
2758 struct mm_struct
*mm
= rq
->prev_mm
;
2764 * A task struct has one reference for the use as "current".
2765 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2766 * schedule one last time. The schedule call will never return, and
2767 * the scheduled task must drop that reference.
2768 * The test for TASK_DEAD must occur while the runqueue locks are
2769 * still held, otherwise prev could be scheduled on another cpu, die
2770 * there before we look at prev->state, and then the reference would
2772 * Manfred Spraul <manfred@colorfullife.com>
2774 prev_state
= prev
->state
;
2775 finish_arch_switch(prev
);
2776 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2777 local_irq_disable();
2778 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2779 perf_event_task_sched_in(current
);
2780 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2782 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2783 finish_lock_switch(rq
, prev
);
2785 fire_sched_in_preempt_notifiers(current
);
2788 if (unlikely(prev_state
== TASK_DEAD
)) {
2790 * Remove function-return probe instances associated with this
2791 * task and put them back on the free list.
2793 kprobe_flush_task(prev
);
2794 put_task_struct(prev
);
2800 /* assumes rq->lock is held */
2801 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2803 if (prev
->sched_class
->pre_schedule
)
2804 prev
->sched_class
->pre_schedule(rq
, prev
);
2807 /* rq->lock is NOT held, but preemption is disabled */
2808 static inline void post_schedule(struct rq
*rq
)
2810 if (rq
->post_schedule
) {
2811 unsigned long flags
;
2813 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2814 if (rq
->curr
->sched_class
->post_schedule
)
2815 rq
->curr
->sched_class
->post_schedule(rq
);
2816 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2818 rq
->post_schedule
= 0;
2824 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2828 static inline void post_schedule(struct rq
*rq
)
2835 * schedule_tail - first thing a freshly forked thread must call.
2836 * @prev: the thread we just switched away from.
2838 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2839 __releases(rq
->lock
)
2841 struct rq
*rq
= this_rq();
2843 finish_task_switch(rq
, prev
);
2846 * FIXME: do we need to worry about rq being invalidated by the
2851 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2852 /* In this case, finish_task_switch does not reenable preemption */
2855 if (current
->set_child_tid
)
2856 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2860 * context_switch - switch to the new MM and the new
2861 * thread's register state.
2864 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2865 struct task_struct
*next
)
2867 struct mm_struct
*mm
, *oldmm
;
2869 prepare_task_switch(rq
, prev
, next
);
2870 trace_sched_switch(prev
, next
);
2872 oldmm
= prev
->active_mm
;
2874 * For paravirt, this is coupled with an exit in switch_to to
2875 * combine the page table reload and the switch backend into
2878 arch_start_context_switch(prev
);
2881 next
->active_mm
= oldmm
;
2882 atomic_inc(&oldmm
->mm_count
);
2883 enter_lazy_tlb(oldmm
, next
);
2885 switch_mm(oldmm
, mm
, next
);
2888 prev
->active_mm
= NULL
;
2889 rq
->prev_mm
= oldmm
;
2892 * Since the runqueue lock will be released by the next
2893 * task (which is an invalid locking op but in the case
2894 * of the scheduler it's an obvious special-case), so we
2895 * do an early lockdep release here:
2897 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2898 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2901 /* Here we just switch the register state and the stack. */
2902 switch_to(prev
, next
, prev
);
2906 * this_rq must be evaluated again because prev may have moved
2907 * CPUs since it called schedule(), thus the 'rq' on its stack
2908 * frame will be invalid.
2910 finish_task_switch(this_rq(), prev
);
2914 * nr_running, nr_uninterruptible and nr_context_switches:
2916 * externally visible scheduler statistics: current number of runnable
2917 * threads, current number of uninterruptible-sleeping threads, total
2918 * number of context switches performed since bootup.
2920 unsigned long nr_running(void)
2922 unsigned long i
, sum
= 0;
2924 for_each_online_cpu(i
)
2925 sum
+= cpu_rq(i
)->nr_running
;
2930 unsigned long nr_uninterruptible(void)
2932 unsigned long i
, sum
= 0;
2934 for_each_possible_cpu(i
)
2935 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2938 * Since we read the counters lockless, it might be slightly
2939 * inaccurate. Do not allow it to go below zero though:
2941 if (unlikely((long)sum
< 0))
2947 unsigned long long nr_context_switches(void)
2950 unsigned long long sum
= 0;
2952 for_each_possible_cpu(i
)
2953 sum
+= cpu_rq(i
)->nr_switches
;
2958 unsigned long nr_iowait(void)
2960 unsigned long i
, sum
= 0;
2962 for_each_possible_cpu(i
)
2963 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2968 unsigned long nr_iowait_cpu(int cpu
)
2970 struct rq
*this = cpu_rq(cpu
);
2971 return atomic_read(&this->nr_iowait
);
2974 unsigned long this_cpu_load(void)
2976 struct rq
*this = this_rq();
2977 return this->cpu_load
[0];
2981 /* Variables and functions for calc_load */
2982 static atomic_long_t calc_load_tasks
;
2983 static unsigned long calc_load_update
;
2984 unsigned long avenrun
[3];
2985 EXPORT_SYMBOL(avenrun
);
2987 static long calc_load_fold_active(struct rq
*this_rq
)
2989 long nr_active
, delta
= 0;
2991 nr_active
= this_rq
->nr_running
;
2992 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2994 if (nr_active
!= this_rq
->calc_load_active
) {
2995 delta
= nr_active
- this_rq
->calc_load_active
;
2996 this_rq
->calc_load_active
= nr_active
;
3004 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3006 * When making the ILB scale, we should try to pull this in as well.
3008 static atomic_long_t calc_load_tasks_idle
;
3010 static void calc_load_account_idle(struct rq
*this_rq
)
3014 delta
= calc_load_fold_active(this_rq
);
3016 atomic_long_add(delta
, &calc_load_tasks_idle
);
3019 static long calc_load_fold_idle(void)
3024 * Its got a race, we don't care...
3026 if (atomic_long_read(&calc_load_tasks_idle
))
3027 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3032 static void calc_load_account_idle(struct rq
*this_rq
)
3036 static inline long calc_load_fold_idle(void)
3043 * get_avenrun - get the load average array
3044 * @loads: pointer to dest load array
3045 * @offset: offset to add
3046 * @shift: shift count to shift the result left
3048 * These values are estimates at best, so no need for locking.
3050 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3052 loads
[0] = (avenrun
[0] + offset
) << shift
;
3053 loads
[1] = (avenrun
[1] + offset
) << shift
;
3054 loads
[2] = (avenrun
[2] + offset
) << shift
;
3057 static unsigned long
3058 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3061 load
+= active
* (FIXED_1
- exp
);
3062 return load
>> FSHIFT
;
3066 * calc_load - update the avenrun load estimates 10 ticks after the
3067 * CPUs have updated calc_load_tasks.
3069 void calc_global_load(void)
3071 unsigned long upd
= calc_load_update
+ 10;
3074 if (time_before(jiffies
, upd
))
3077 active
= atomic_long_read(&calc_load_tasks
);
3078 active
= active
> 0 ? active
* FIXED_1
: 0;
3080 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3081 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3082 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3084 calc_load_update
+= LOAD_FREQ
;
3088 * Called from update_cpu_load() to periodically update this CPU's
3091 static void calc_load_account_active(struct rq
*this_rq
)
3095 if (time_before(jiffies
, this_rq
->calc_load_update
))
3098 delta
= calc_load_fold_active(this_rq
);
3099 delta
+= calc_load_fold_idle();
3101 atomic_long_add(delta
, &calc_load_tasks
);
3103 this_rq
->calc_load_update
+= LOAD_FREQ
;
3107 * The exact cpuload at various idx values, calculated at every tick would be
3108 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3110 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3111 * on nth tick when cpu may be busy, then we have:
3112 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3113 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3115 * decay_load_missed() below does efficient calculation of
3116 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3117 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3119 * The calculation is approximated on a 128 point scale.
3120 * degrade_zero_ticks is the number of ticks after which load at any
3121 * particular idx is approximated to be zero.
3122 * degrade_factor is a precomputed table, a row for each load idx.
3123 * Each column corresponds to degradation factor for a power of two ticks,
3124 * based on 128 point scale.
3126 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3127 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3129 * With this power of 2 load factors, we can degrade the load n times
3130 * by looking at 1 bits in n and doing as many mult/shift instead of
3131 * n mult/shifts needed by the exact degradation.
3133 #define DEGRADE_SHIFT 7
3134 static const unsigned char
3135 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3136 static const unsigned char
3137 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3138 {0, 0, 0, 0, 0, 0, 0, 0},
3139 {64, 32, 8, 0, 0, 0, 0, 0},
3140 {96, 72, 40, 12, 1, 0, 0},
3141 {112, 98, 75, 43, 15, 1, 0},
3142 {120, 112, 98, 76, 45, 16, 2} };
3145 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3146 * would be when CPU is idle and so we just decay the old load without
3147 * adding any new load.
3149 static unsigned long
3150 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3154 if (!missed_updates
)
3157 if (missed_updates
>= degrade_zero_ticks
[idx
])
3161 return load
>> missed_updates
;
3163 while (missed_updates
) {
3164 if (missed_updates
% 2)
3165 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3167 missed_updates
>>= 1;
3174 * Update rq->cpu_load[] statistics. This function is usually called every
3175 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3176 * every tick. We fix it up based on jiffies.
3178 static void update_cpu_load(struct rq
*this_rq
)
3180 unsigned long this_load
= this_rq
->load
.weight
;
3181 unsigned long curr_jiffies
= jiffies
;
3182 unsigned long pending_updates
;
3185 this_rq
->nr_load_updates
++;
3187 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3188 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3191 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3192 this_rq
->last_load_update_tick
= curr_jiffies
;
3194 /* Update our load: */
3195 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3196 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3197 unsigned long old_load
, new_load
;
3199 /* scale is effectively 1 << i now, and >> i divides by scale */
3201 old_load
= this_rq
->cpu_load
[i
];
3202 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3203 new_load
= this_load
;
3205 * Round up the averaging division if load is increasing. This
3206 * prevents us from getting stuck on 9 if the load is 10, for
3209 if (new_load
> old_load
)
3210 new_load
+= scale
- 1;
3212 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3215 sched_avg_update(this_rq
);
3218 static void update_cpu_load_active(struct rq
*this_rq
)
3220 update_cpu_load(this_rq
);
3222 calc_load_account_active(this_rq
);
3228 * sched_exec - execve() is a valuable balancing opportunity, because at
3229 * this point the task has the smallest effective memory and cache footprint.
3231 void sched_exec(void)
3233 struct task_struct
*p
= current
;
3234 unsigned long flags
;
3238 rq
= task_rq_lock(p
, &flags
);
3239 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3240 if (dest_cpu
== smp_processor_id())
3244 * select_task_rq() can race against ->cpus_allowed
3246 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3247 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3248 struct migration_arg arg
= { p
, dest_cpu
};
3250 task_rq_unlock(rq
, &flags
);
3251 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3255 task_rq_unlock(rq
, &flags
);
3260 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3262 EXPORT_PER_CPU_SYMBOL(kstat
);
3265 * Return any ns on the sched_clock that have not yet been accounted in
3266 * @p in case that task is currently running.
3268 * Called with task_rq_lock() held on @rq.
3270 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3274 if (task_current(rq
, p
)) {
3275 update_rq_clock(rq
);
3276 ns
= rq
->clock
- p
->se
.exec_start
;
3284 unsigned long long task_delta_exec(struct task_struct
*p
)
3286 unsigned long flags
;
3290 rq
= task_rq_lock(p
, &flags
);
3291 ns
= do_task_delta_exec(p
, rq
);
3292 task_rq_unlock(rq
, &flags
);
3298 * Return accounted runtime for the task.
3299 * In case the task is currently running, return the runtime plus current's
3300 * pending runtime that have not been accounted yet.
3302 unsigned long long task_sched_runtime(struct task_struct
*p
)
3304 unsigned long flags
;
3308 rq
= task_rq_lock(p
, &flags
);
3309 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3310 task_rq_unlock(rq
, &flags
);
3316 * Return sum_exec_runtime for the thread group.
3317 * In case the task is currently running, return the sum plus current's
3318 * pending runtime that have not been accounted yet.
3320 * Note that the thread group might have other running tasks as well,
3321 * so the return value not includes other pending runtime that other
3322 * running tasks might have.
3324 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3326 struct task_cputime totals
;
3327 unsigned long flags
;
3331 rq
= task_rq_lock(p
, &flags
);
3332 thread_group_cputime(p
, &totals
);
3333 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3334 task_rq_unlock(rq
, &flags
);
3340 * Account user cpu time to a process.
3341 * @p: the process that the cpu time gets accounted to
3342 * @cputime: the cpu time spent in user space since the last update
3343 * @cputime_scaled: cputime scaled by cpu frequency
3345 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3346 cputime_t cputime_scaled
)
3348 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3351 /* Add user time to process. */
3352 p
->utime
= cputime_add(p
->utime
, cputime
);
3353 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3354 account_group_user_time(p
, cputime
);
3356 /* Add user time to cpustat. */
3357 tmp
= cputime_to_cputime64(cputime
);
3358 if (TASK_NICE(p
) > 0)
3359 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3361 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3363 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3364 /* Account for user time used */
3365 acct_update_integrals(p
);
3369 * Account guest cpu time to a process.
3370 * @p: the process that the cpu time gets accounted to
3371 * @cputime: the cpu time spent in virtual machine since the last update
3372 * @cputime_scaled: cputime scaled by cpu frequency
3374 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3375 cputime_t cputime_scaled
)
3378 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3380 tmp
= cputime_to_cputime64(cputime
);
3382 /* Add guest time to process. */
3383 p
->utime
= cputime_add(p
->utime
, cputime
);
3384 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3385 account_group_user_time(p
, cputime
);
3386 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3388 /* Add guest time to cpustat. */
3389 if (TASK_NICE(p
) > 0) {
3390 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3391 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3393 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3394 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3399 * Account system cpu time to a process.
3400 * @p: the process that the cpu time gets accounted to
3401 * @hardirq_offset: the offset to subtract from hardirq_count()
3402 * @cputime: the cpu time spent in kernel space since the last update
3403 * @cputime_scaled: cputime scaled by cpu frequency
3405 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3406 cputime_t cputime
, cputime_t cputime_scaled
)
3408 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3411 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3412 account_guest_time(p
, cputime
, cputime_scaled
);
3416 /* Add system time to process. */
3417 p
->stime
= cputime_add(p
->stime
, cputime
);
3418 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3419 account_group_system_time(p
, cputime
);
3421 /* Add system time to cpustat. */
3422 tmp
= cputime_to_cputime64(cputime
);
3423 if (hardirq_count() - hardirq_offset
)
3424 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3425 else if (in_serving_softirq())
3426 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3428 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3430 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3432 /* Account for system time used */
3433 acct_update_integrals(p
);
3437 * Account for involuntary wait time.
3438 * @steal: the cpu time spent in involuntary wait
3440 void account_steal_time(cputime_t cputime
)
3442 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3443 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3445 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3449 * Account for idle time.
3450 * @cputime: the cpu time spent in idle wait
3452 void account_idle_time(cputime_t cputime
)
3454 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3455 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3456 struct rq
*rq
= this_rq();
3458 if (atomic_read(&rq
->nr_iowait
) > 0)
3459 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3461 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3464 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3467 * Account a single tick of cpu time.
3468 * @p: the process that the cpu time gets accounted to
3469 * @user_tick: indicates if the tick is a user or a system tick
3471 void account_process_tick(struct task_struct
*p
, int user_tick
)
3473 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3474 struct rq
*rq
= this_rq();
3477 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3478 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3479 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3482 account_idle_time(cputime_one_jiffy
);
3486 * Account multiple ticks of steal time.
3487 * @p: the process from which the cpu time has been stolen
3488 * @ticks: number of stolen ticks
3490 void account_steal_ticks(unsigned long ticks
)
3492 account_steal_time(jiffies_to_cputime(ticks
));
3496 * Account multiple ticks of idle time.
3497 * @ticks: number of stolen ticks
3499 void account_idle_ticks(unsigned long ticks
)
3501 account_idle_time(jiffies_to_cputime(ticks
));
3507 * Use precise platform statistics if available:
3509 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3510 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3516 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3518 struct task_cputime cputime
;
3520 thread_group_cputime(p
, &cputime
);
3522 *ut
= cputime
.utime
;
3523 *st
= cputime
.stime
;
3527 #ifndef nsecs_to_cputime
3528 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3531 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3533 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3536 * Use CFS's precise accounting:
3538 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3544 do_div(temp
, total
);
3545 utime
= (cputime_t
)temp
;
3550 * Compare with previous values, to keep monotonicity:
3552 p
->prev_utime
= max(p
->prev_utime
, utime
);
3553 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3555 *ut
= p
->prev_utime
;
3556 *st
= p
->prev_stime
;
3560 * Must be called with siglock held.
3562 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3564 struct signal_struct
*sig
= p
->signal
;
3565 struct task_cputime cputime
;
3566 cputime_t rtime
, utime
, total
;
3568 thread_group_cputime(p
, &cputime
);
3570 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3571 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3576 temp
*= cputime
.utime
;
3577 do_div(temp
, total
);
3578 utime
= (cputime_t
)temp
;
3582 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3583 sig
->prev_stime
= max(sig
->prev_stime
,
3584 cputime_sub(rtime
, sig
->prev_utime
));
3586 *ut
= sig
->prev_utime
;
3587 *st
= sig
->prev_stime
;
3592 * This function gets called by the timer code, with HZ frequency.
3593 * We call it with interrupts disabled.
3595 * It also gets called by the fork code, when changing the parent's
3598 void scheduler_tick(void)
3600 int cpu
= smp_processor_id();
3601 struct rq
*rq
= cpu_rq(cpu
);
3602 struct task_struct
*curr
= rq
->curr
;
3606 raw_spin_lock(&rq
->lock
);
3607 update_rq_clock(rq
);
3608 update_cpu_load_active(rq
);
3609 curr
->sched_class
->task_tick(rq
, curr
, 0);
3610 raw_spin_unlock(&rq
->lock
);
3612 perf_event_task_tick(curr
);
3615 rq
->idle_at_tick
= idle_cpu(cpu
);
3616 trigger_load_balance(rq
, cpu
);
3620 notrace
unsigned long get_parent_ip(unsigned long addr
)
3622 if (in_lock_functions(addr
)) {
3623 addr
= CALLER_ADDR2
;
3624 if (in_lock_functions(addr
))
3625 addr
= CALLER_ADDR3
;
3630 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3631 defined(CONFIG_PREEMPT_TRACER))
3633 void __kprobes
add_preempt_count(int val
)
3635 #ifdef CONFIG_DEBUG_PREEMPT
3639 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3642 preempt_count() += val
;
3643 #ifdef CONFIG_DEBUG_PREEMPT
3645 * Spinlock count overflowing soon?
3647 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3650 if (preempt_count() == val
)
3651 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3653 EXPORT_SYMBOL(add_preempt_count
);
3655 void __kprobes
sub_preempt_count(int val
)
3657 #ifdef CONFIG_DEBUG_PREEMPT
3661 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3664 * Is the spinlock portion underflowing?
3666 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3667 !(preempt_count() & PREEMPT_MASK
)))
3671 if (preempt_count() == val
)
3672 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3673 preempt_count() -= val
;
3675 EXPORT_SYMBOL(sub_preempt_count
);
3680 * Print scheduling while atomic bug:
3682 static noinline
void __schedule_bug(struct task_struct
*prev
)
3684 struct pt_regs
*regs
= get_irq_regs();
3686 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3687 prev
->comm
, prev
->pid
, preempt_count());
3689 debug_show_held_locks(prev
);
3691 if (irqs_disabled())
3692 print_irqtrace_events(prev
);
3701 * Various schedule()-time debugging checks and statistics:
3703 static inline void schedule_debug(struct task_struct
*prev
)
3706 * Test if we are atomic. Since do_exit() needs to call into
3707 * schedule() atomically, we ignore that path for now.
3708 * Otherwise, whine if we are scheduling when we should not be.
3710 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3711 __schedule_bug(prev
);
3713 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3715 schedstat_inc(this_rq(), sched_count
);
3716 #ifdef CONFIG_SCHEDSTATS
3717 if (unlikely(prev
->lock_depth
>= 0)) {
3718 schedstat_inc(this_rq(), bkl_count
);
3719 schedstat_inc(prev
, sched_info
.bkl_count
);
3724 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3727 update_rq_clock(rq
);
3728 rq
->skip_clock_update
= 0;
3729 prev
->sched_class
->put_prev_task(rq
, prev
);
3733 * Pick up the highest-prio task:
3735 static inline struct task_struct
*
3736 pick_next_task(struct rq
*rq
)
3738 const struct sched_class
*class;
3739 struct task_struct
*p
;
3742 * Optimization: we know that if all tasks are in
3743 * the fair class we can call that function directly:
3745 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3746 p
= fair_sched_class
.pick_next_task(rq
);
3751 for_each_class(class) {
3752 p
= class->pick_next_task(rq
);
3757 BUG(); /* the idle class will always have a runnable task */
3761 * schedule() is the main scheduler function.
3763 asmlinkage
void __sched
schedule(void)
3765 struct task_struct
*prev
, *next
;
3766 unsigned long *switch_count
;
3772 cpu
= smp_processor_id();
3774 rcu_note_context_switch(cpu
);
3777 release_kernel_lock(prev
);
3778 need_resched_nonpreemptible
:
3780 schedule_debug(prev
);
3782 if (sched_feat(HRTICK
))
3785 raw_spin_lock_irq(&rq
->lock
);
3786 clear_tsk_need_resched(prev
);
3788 switch_count
= &prev
->nivcsw
;
3789 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3790 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3791 prev
->state
= TASK_RUNNING
;
3794 * If a worker is going to sleep, notify and
3795 * ask workqueue whether it wants to wake up a
3796 * task to maintain concurrency. If so, wake
3799 if (prev
->flags
& PF_WQ_WORKER
) {
3800 struct task_struct
*to_wakeup
;
3802 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3804 try_to_wake_up_local(to_wakeup
);
3806 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3808 switch_count
= &prev
->nvcsw
;
3811 pre_schedule(rq
, prev
);
3813 if (unlikely(!rq
->nr_running
))
3814 idle_balance(cpu
, rq
);
3816 put_prev_task(rq
, prev
);
3817 next
= pick_next_task(rq
);
3819 if (likely(prev
!= next
)) {
3820 sched_info_switch(prev
, next
);
3821 perf_event_task_sched_out(prev
, next
);
3827 context_switch(rq
, prev
, next
); /* unlocks the rq */
3829 * The context switch have flipped the stack from under us
3830 * and restored the local variables which were saved when
3831 * this task called schedule() in the past. prev == current
3832 * is still correct, but it can be moved to another cpu/rq.
3834 cpu
= smp_processor_id();
3837 raw_spin_unlock_irq(&rq
->lock
);
3841 if (unlikely(reacquire_kernel_lock(prev
)))
3842 goto need_resched_nonpreemptible
;
3844 preempt_enable_no_resched();
3848 EXPORT_SYMBOL(schedule
);
3850 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3852 * Look out! "owner" is an entirely speculative pointer
3853 * access and not reliable.
3855 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3860 if (!sched_feat(OWNER_SPIN
))
3863 #ifdef CONFIG_DEBUG_PAGEALLOC
3865 * Need to access the cpu field knowing that
3866 * DEBUG_PAGEALLOC could have unmapped it if
3867 * the mutex owner just released it and exited.
3869 if (probe_kernel_address(&owner
->cpu
, cpu
))
3876 * Even if the access succeeded (likely case),
3877 * the cpu field may no longer be valid.
3879 if (cpu
>= nr_cpumask_bits
)
3883 * We need to validate that we can do a
3884 * get_cpu() and that we have the percpu area.
3886 if (!cpu_online(cpu
))
3893 * Owner changed, break to re-assess state.
3895 if (lock
->owner
!= owner
) {
3897 * If the lock has switched to a different owner,
3898 * we likely have heavy contention. Return 0 to quit
3899 * optimistic spinning and not contend further:
3907 * Is that owner really running on that cpu?
3909 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3919 #ifdef CONFIG_PREEMPT
3921 * this is the entry point to schedule() from in-kernel preemption
3922 * off of preempt_enable. Kernel preemptions off return from interrupt
3923 * occur there and call schedule directly.
3925 asmlinkage
void __sched notrace
preempt_schedule(void)
3927 struct thread_info
*ti
= current_thread_info();
3930 * If there is a non-zero preempt_count or interrupts are disabled,
3931 * we do not want to preempt the current task. Just return..
3933 if (likely(ti
->preempt_count
|| irqs_disabled()))
3937 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3939 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3942 * Check again in case we missed a preemption opportunity
3943 * between schedule and now.
3946 } while (need_resched());
3948 EXPORT_SYMBOL(preempt_schedule
);
3951 * this is the entry point to schedule() from kernel preemption
3952 * off of irq context.
3953 * Note, that this is called and return with irqs disabled. This will
3954 * protect us against recursive calling from irq.
3956 asmlinkage
void __sched
preempt_schedule_irq(void)
3958 struct thread_info
*ti
= current_thread_info();
3960 /* Catch callers which need to be fixed */
3961 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3964 add_preempt_count(PREEMPT_ACTIVE
);
3967 local_irq_disable();
3968 sub_preempt_count(PREEMPT_ACTIVE
);
3971 * Check again in case we missed a preemption opportunity
3972 * between schedule and now.
3975 } while (need_resched());
3978 #endif /* CONFIG_PREEMPT */
3980 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3983 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3985 EXPORT_SYMBOL(default_wake_function
);
3988 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3989 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3990 * number) then we wake all the non-exclusive tasks and one exclusive task.
3992 * There are circumstances in which we can try to wake a task which has already
3993 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3994 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3996 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3997 int nr_exclusive
, int wake_flags
, void *key
)
3999 wait_queue_t
*curr
, *next
;
4001 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4002 unsigned flags
= curr
->flags
;
4004 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4005 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4011 * __wake_up - wake up threads blocked on a waitqueue.
4013 * @mode: which threads
4014 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4015 * @key: is directly passed to the wakeup function
4017 * It may be assumed that this function implies a write memory barrier before
4018 * changing the task state if and only if any tasks are woken up.
4020 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4021 int nr_exclusive
, void *key
)
4023 unsigned long flags
;
4025 spin_lock_irqsave(&q
->lock
, flags
);
4026 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4027 spin_unlock_irqrestore(&q
->lock
, flags
);
4029 EXPORT_SYMBOL(__wake_up
);
4032 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4034 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4036 __wake_up_common(q
, mode
, 1, 0, NULL
);
4038 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4040 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4042 __wake_up_common(q
, mode
, 1, 0, key
);
4046 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4048 * @mode: which threads
4049 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4050 * @key: opaque value to be passed to wakeup targets
4052 * The sync wakeup differs that the waker knows that it will schedule
4053 * away soon, so while the target thread will be woken up, it will not
4054 * be migrated to another CPU - ie. the two threads are 'synchronized'
4055 * with each other. This can prevent needless bouncing between CPUs.
4057 * On UP it can prevent extra preemption.
4059 * It may be assumed that this function implies a write memory barrier before
4060 * changing the task state if and only if any tasks are woken up.
4062 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4063 int nr_exclusive
, void *key
)
4065 unsigned long flags
;
4066 int wake_flags
= WF_SYNC
;
4071 if (unlikely(!nr_exclusive
))
4074 spin_lock_irqsave(&q
->lock
, flags
);
4075 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4076 spin_unlock_irqrestore(&q
->lock
, flags
);
4078 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4081 * __wake_up_sync - see __wake_up_sync_key()
4083 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4085 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4087 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4090 * complete: - signals a single thread waiting on this completion
4091 * @x: holds the state of this particular completion
4093 * This will wake up a single thread waiting on this completion. Threads will be
4094 * awakened in the same order in which they were queued.
4096 * See also complete_all(), wait_for_completion() and related routines.
4098 * It may be assumed that this function implies a write memory barrier before
4099 * changing the task state if and only if any tasks are woken up.
4101 void complete(struct completion
*x
)
4103 unsigned long flags
;
4105 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4107 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4108 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4110 EXPORT_SYMBOL(complete
);
4113 * complete_all: - signals all threads waiting on this completion
4114 * @x: holds the state of this particular completion
4116 * This will wake up all threads waiting on this particular completion event.
4118 * It may be assumed that this function implies a write memory barrier before
4119 * changing the task state if and only if any tasks are woken up.
4121 void complete_all(struct completion
*x
)
4123 unsigned long flags
;
4125 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4126 x
->done
+= UINT_MAX
/2;
4127 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4128 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4130 EXPORT_SYMBOL(complete_all
);
4132 static inline long __sched
4133 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4136 DECLARE_WAITQUEUE(wait
, current
);
4138 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4140 if (signal_pending_state(state
, current
)) {
4141 timeout
= -ERESTARTSYS
;
4144 __set_current_state(state
);
4145 spin_unlock_irq(&x
->wait
.lock
);
4146 timeout
= schedule_timeout(timeout
);
4147 spin_lock_irq(&x
->wait
.lock
);
4148 } while (!x
->done
&& timeout
);
4149 __remove_wait_queue(&x
->wait
, &wait
);
4154 return timeout
?: 1;
4158 wait_for_common(struct completion
*x
, long timeout
, int state
)
4162 spin_lock_irq(&x
->wait
.lock
);
4163 timeout
= do_wait_for_common(x
, timeout
, state
);
4164 spin_unlock_irq(&x
->wait
.lock
);
4169 * wait_for_completion: - waits for completion of a task
4170 * @x: holds the state of this particular completion
4172 * This waits to be signaled for completion of a specific task. It is NOT
4173 * interruptible and there is no timeout.
4175 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4176 * and interrupt capability. Also see complete().
4178 void __sched
wait_for_completion(struct completion
*x
)
4180 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4182 EXPORT_SYMBOL(wait_for_completion
);
4185 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4186 * @x: holds the state of this particular completion
4187 * @timeout: timeout value in jiffies
4189 * This waits for either a completion of a specific task to be signaled or for a
4190 * specified timeout to expire. The timeout is in jiffies. It is not
4193 unsigned long __sched
4194 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4196 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4198 EXPORT_SYMBOL(wait_for_completion_timeout
);
4201 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4202 * @x: holds the state of this particular completion
4204 * This waits for completion of a specific task to be signaled. It is
4207 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4209 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4210 if (t
== -ERESTARTSYS
)
4214 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4217 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4218 * @x: holds the state of this particular completion
4219 * @timeout: timeout value in jiffies
4221 * This waits for either a completion of a specific task to be signaled or for a
4222 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4224 unsigned long __sched
4225 wait_for_completion_interruptible_timeout(struct completion
*x
,
4226 unsigned long timeout
)
4228 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4230 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4233 * wait_for_completion_killable: - waits for completion of a task (killable)
4234 * @x: holds the state of this particular completion
4236 * This waits to be signaled for completion of a specific task. It can be
4237 * interrupted by a kill signal.
4239 int __sched
wait_for_completion_killable(struct completion
*x
)
4241 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4242 if (t
== -ERESTARTSYS
)
4246 EXPORT_SYMBOL(wait_for_completion_killable
);
4249 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4250 * @x: holds the state of this particular completion
4251 * @timeout: timeout value in jiffies
4253 * This waits for either a completion of a specific task to be
4254 * signaled or for a specified timeout to expire. It can be
4255 * interrupted by a kill signal. The timeout is in jiffies.
4257 unsigned long __sched
4258 wait_for_completion_killable_timeout(struct completion
*x
,
4259 unsigned long timeout
)
4261 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4263 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4266 * try_wait_for_completion - try to decrement a completion without blocking
4267 * @x: completion structure
4269 * Returns: 0 if a decrement cannot be done without blocking
4270 * 1 if a decrement succeeded.
4272 * If a completion is being used as a counting completion,
4273 * attempt to decrement the counter without blocking. This
4274 * enables us to avoid waiting if the resource the completion
4275 * is protecting is not available.
4277 bool try_wait_for_completion(struct completion
*x
)
4279 unsigned long flags
;
4282 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4287 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4290 EXPORT_SYMBOL(try_wait_for_completion
);
4293 * completion_done - Test to see if a completion has any waiters
4294 * @x: completion structure
4296 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4297 * 1 if there are no waiters.
4300 bool completion_done(struct completion
*x
)
4302 unsigned long flags
;
4305 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4308 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4311 EXPORT_SYMBOL(completion_done
);
4314 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4316 unsigned long flags
;
4319 init_waitqueue_entry(&wait
, current
);
4321 __set_current_state(state
);
4323 spin_lock_irqsave(&q
->lock
, flags
);
4324 __add_wait_queue(q
, &wait
);
4325 spin_unlock(&q
->lock
);
4326 timeout
= schedule_timeout(timeout
);
4327 spin_lock_irq(&q
->lock
);
4328 __remove_wait_queue(q
, &wait
);
4329 spin_unlock_irqrestore(&q
->lock
, flags
);
4334 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4336 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4338 EXPORT_SYMBOL(interruptible_sleep_on
);
4341 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4343 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4345 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4347 void __sched
sleep_on(wait_queue_head_t
*q
)
4349 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4351 EXPORT_SYMBOL(sleep_on
);
4353 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4355 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4357 EXPORT_SYMBOL(sleep_on_timeout
);
4359 #ifdef CONFIG_RT_MUTEXES
4362 * rt_mutex_setprio - set the current priority of a task
4364 * @prio: prio value (kernel-internal form)
4366 * This function changes the 'effective' priority of a task. It does
4367 * not touch ->normal_prio like __setscheduler().
4369 * Used by the rt_mutex code to implement priority inheritance logic.
4371 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4373 unsigned long flags
;
4374 int oldprio
, on_rq
, running
;
4376 const struct sched_class
*prev_class
;
4378 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4380 rq
= task_rq_lock(p
, &flags
);
4382 trace_sched_pi_setprio(p
, prio
);
4384 prev_class
= p
->sched_class
;
4385 on_rq
= p
->se
.on_rq
;
4386 running
= task_current(rq
, p
);
4388 dequeue_task(rq
, p
, 0);
4390 p
->sched_class
->put_prev_task(rq
, p
);
4393 p
->sched_class
= &rt_sched_class
;
4395 p
->sched_class
= &fair_sched_class
;
4400 p
->sched_class
->set_curr_task(rq
);
4402 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4404 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4406 task_rq_unlock(rq
, &flags
);
4411 void set_user_nice(struct task_struct
*p
, long nice
)
4413 int old_prio
, delta
, on_rq
;
4414 unsigned long flags
;
4417 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4420 * We have to be careful, if called from sys_setpriority(),
4421 * the task might be in the middle of scheduling on another CPU.
4423 rq
= task_rq_lock(p
, &flags
);
4425 * The RT priorities are set via sched_setscheduler(), but we still
4426 * allow the 'normal' nice value to be set - but as expected
4427 * it wont have any effect on scheduling until the task is
4428 * SCHED_FIFO/SCHED_RR:
4430 if (task_has_rt_policy(p
)) {
4431 p
->static_prio
= NICE_TO_PRIO(nice
);
4434 on_rq
= p
->se
.on_rq
;
4436 dequeue_task(rq
, p
, 0);
4438 p
->static_prio
= NICE_TO_PRIO(nice
);
4441 p
->prio
= effective_prio(p
);
4442 delta
= p
->prio
- old_prio
;
4445 enqueue_task(rq
, p
, 0);
4447 * If the task increased its priority or is running and
4448 * lowered its priority, then reschedule its CPU:
4450 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4451 resched_task(rq
->curr
);
4454 task_rq_unlock(rq
, &flags
);
4456 EXPORT_SYMBOL(set_user_nice
);
4459 * can_nice - check if a task can reduce its nice value
4463 int can_nice(const struct task_struct
*p
, const int nice
)
4465 /* convert nice value [19,-20] to rlimit style value [1,40] */
4466 int nice_rlim
= 20 - nice
;
4468 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4469 capable(CAP_SYS_NICE
));
4472 #ifdef __ARCH_WANT_SYS_NICE
4475 * sys_nice - change the priority of the current process.
4476 * @increment: priority increment
4478 * sys_setpriority is a more generic, but much slower function that
4479 * does similar things.
4481 SYSCALL_DEFINE1(nice
, int, increment
)
4486 * Setpriority might change our priority at the same moment.
4487 * We don't have to worry. Conceptually one call occurs first
4488 * and we have a single winner.
4490 if (increment
< -40)
4495 nice
= TASK_NICE(current
) + increment
;
4501 if (increment
< 0 && !can_nice(current
, nice
))
4504 retval
= security_task_setnice(current
, nice
);
4508 set_user_nice(current
, nice
);
4515 * task_prio - return the priority value of a given task.
4516 * @p: the task in question.
4518 * This is the priority value as seen by users in /proc.
4519 * RT tasks are offset by -200. Normal tasks are centered
4520 * around 0, value goes from -16 to +15.
4522 int task_prio(const struct task_struct
*p
)
4524 return p
->prio
- MAX_RT_PRIO
;
4528 * task_nice - return the nice value of a given task.
4529 * @p: the task in question.
4531 int task_nice(const struct task_struct
*p
)
4533 return TASK_NICE(p
);
4535 EXPORT_SYMBOL(task_nice
);
4538 * idle_cpu - is a given cpu idle currently?
4539 * @cpu: the processor in question.
4541 int idle_cpu(int cpu
)
4543 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4547 * idle_task - return the idle task for a given cpu.
4548 * @cpu: the processor in question.
4550 struct task_struct
*idle_task(int cpu
)
4552 return cpu_rq(cpu
)->idle
;
4556 * find_process_by_pid - find a process with a matching PID value.
4557 * @pid: the pid in question.
4559 static struct task_struct
*find_process_by_pid(pid_t pid
)
4561 return pid
? find_task_by_vpid(pid
) : current
;
4564 /* Actually do priority change: must hold rq lock. */
4566 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4568 BUG_ON(p
->se
.on_rq
);
4571 p
->rt_priority
= prio
;
4572 p
->normal_prio
= normal_prio(p
);
4573 /* we are holding p->pi_lock already */
4574 p
->prio
= rt_mutex_getprio(p
);
4575 if (rt_prio(p
->prio
))
4576 p
->sched_class
= &rt_sched_class
;
4578 p
->sched_class
= &fair_sched_class
;
4583 * check the target process has a UID that matches the current process's
4585 static bool check_same_owner(struct task_struct
*p
)
4587 const struct cred
*cred
= current_cred(), *pcred
;
4591 pcred
= __task_cred(p
);
4592 match
= (cred
->euid
== pcred
->euid
||
4593 cred
->euid
== pcred
->uid
);
4598 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4599 struct sched_param
*param
, bool user
)
4601 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4602 unsigned long flags
;
4603 const struct sched_class
*prev_class
;
4607 /* may grab non-irq protected spin_locks */
4608 BUG_ON(in_interrupt());
4610 /* double check policy once rq lock held */
4612 reset_on_fork
= p
->sched_reset_on_fork
;
4613 policy
= oldpolicy
= p
->policy
;
4615 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4616 policy
&= ~SCHED_RESET_ON_FORK
;
4618 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4619 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4620 policy
!= SCHED_IDLE
)
4625 * Valid priorities for SCHED_FIFO and SCHED_RR are
4626 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4627 * SCHED_BATCH and SCHED_IDLE is 0.
4629 if (param
->sched_priority
< 0 ||
4630 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4631 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4633 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4637 * Allow unprivileged RT tasks to decrease priority:
4639 if (user
&& !capable(CAP_SYS_NICE
)) {
4640 if (rt_policy(policy
)) {
4641 unsigned long rlim_rtprio
=
4642 task_rlimit(p
, RLIMIT_RTPRIO
);
4644 /* can't set/change the rt policy */
4645 if (policy
!= p
->policy
&& !rlim_rtprio
)
4648 /* can't increase priority */
4649 if (param
->sched_priority
> p
->rt_priority
&&
4650 param
->sched_priority
> rlim_rtprio
)
4654 * Like positive nice levels, dont allow tasks to
4655 * move out of SCHED_IDLE either:
4657 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4660 /* can't change other user's priorities */
4661 if (!check_same_owner(p
))
4664 /* Normal users shall not reset the sched_reset_on_fork flag */
4665 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4670 retval
= security_task_setscheduler(p
, policy
, param
);
4676 * make sure no PI-waiters arrive (or leave) while we are
4677 * changing the priority of the task:
4679 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4681 * To be able to change p->policy safely, the apropriate
4682 * runqueue lock must be held.
4684 rq
= __task_rq_lock(p
);
4687 * Changing the policy of the stop threads its a very bad idea
4689 if (p
== rq
->stop
) {
4690 __task_rq_unlock(rq
);
4691 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4695 #ifdef CONFIG_RT_GROUP_SCHED
4698 * Do not allow realtime tasks into groups that have no runtime
4701 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4702 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4703 __task_rq_unlock(rq
);
4704 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4710 /* recheck policy now with rq lock held */
4711 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4712 policy
= oldpolicy
= -1;
4713 __task_rq_unlock(rq
);
4714 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4717 on_rq
= p
->se
.on_rq
;
4718 running
= task_current(rq
, p
);
4720 deactivate_task(rq
, p
, 0);
4722 p
->sched_class
->put_prev_task(rq
, p
);
4724 p
->sched_reset_on_fork
= reset_on_fork
;
4727 prev_class
= p
->sched_class
;
4728 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4731 p
->sched_class
->set_curr_task(rq
);
4733 activate_task(rq
, p
, 0);
4735 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4737 __task_rq_unlock(rq
);
4738 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4740 rt_mutex_adjust_pi(p
);
4746 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4747 * @p: the task in question.
4748 * @policy: new policy.
4749 * @param: structure containing the new RT priority.
4751 * NOTE that the task may be already dead.
4753 int sched_setscheduler(struct task_struct
*p
, int policy
,
4754 struct sched_param
*param
)
4756 return __sched_setscheduler(p
, policy
, param
, true);
4758 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4761 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4762 * @p: the task in question.
4763 * @policy: new policy.
4764 * @param: structure containing the new RT priority.
4766 * Just like sched_setscheduler, only don't bother checking if the
4767 * current context has permission. For example, this is needed in
4768 * stop_machine(): we create temporary high priority worker threads,
4769 * but our caller might not have that capability.
4771 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4772 struct sched_param
*param
)
4774 return __sched_setscheduler(p
, policy
, param
, false);
4778 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4780 struct sched_param lparam
;
4781 struct task_struct
*p
;
4784 if (!param
|| pid
< 0)
4786 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4791 p
= find_process_by_pid(pid
);
4793 retval
= sched_setscheduler(p
, policy
, &lparam
);
4800 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4801 * @pid: the pid in question.
4802 * @policy: new policy.
4803 * @param: structure containing the new RT priority.
4805 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4806 struct sched_param __user
*, param
)
4808 /* negative values for policy are not valid */
4812 return do_sched_setscheduler(pid
, policy
, param
);
4816 * sys_sched_setparam - set/change the RT priority of a thread
4817 * @pid: the pid in question.
4818 * @param: structure containing the new RT priority.
4820 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4822 return do_sched_setscheduler(pid
, -1, param
);
4826 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4827 * @pid: the pid in question.
4829 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4831 struct task_struct
*p
;
4839 p
= find_process_by_pid(pid
);
4841 retval
= security_task_getscheduler(p
);
4844 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4851 * sys_sched_getparam - get the RT priority of a thread
4852 * @pid: the pid in question.
4853 * @param: structure containing the RT priority.
4855 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4857 struct sched_param lp
;
4858 struct task_struct
*p
;
4861 if (!param
|| pid
< 0)
4865 p
= find_process_by_pid(pid
);
4870 retval
= security_task_getscheduler(p
);
4874 lp
.sched_priority
= p
->rt_priority
;
4878 * This one might sleep, we cannot do it with a spinlock held ...
4880 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4889 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4891 cpumask_var_t cpus_allowed
, new_mask
;
4892 struct task_struct
*p
;
4898 p
= find_process_by_pid(pid
);
4905 /* Prevent p going away */
4909 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4913 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4915 goto out_free_cpus_allowed
;
4918 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4921 retval
= security_task_setscheduler(p
, 0, NULL
);
4925 cpuset_cpus_allowed(p
, cpus_allowed
);
4926 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4928 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4931 cpuset_cpus_allowed(p
, cpus_allowed
);
4932 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4934 * We must have raced with a concurrent cpuset
4935 * update. Just reset the cpus_allowed to the
4936 * cpuset's cpus_allowed
4938 cpumask_copy(new_mask
, cpus_allowed
);
4943 free_cpumask_var(new_mask
);
4944 out_free_cpus_allowed
:
4945 free_cpumask_var(cpus_allowed
);
4952 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4953 struct cpumask
*new_mask
)
4955 if (len
< cpumask_size())
4956 cpumask_clear(new_mask
);
4957 else if (len
> cpumask_size())
4958 len
= cpumask_size();
4960 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4964 * sys_sched_setaffinity - set the cpu affinity of a process
4965 * @pid: pid of the process
4966 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4967 * @user_mask_ptr: user-space pointer to the new cpu mask
4969 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4970 unsigned long __user
*, user_mask_ptr
)
4972 cpumask_var_t new_mask
;
4975 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4978 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4980 retval
= sched_setaffinity(pid
, new_mask
);
4981 free_cpumask_var(new_mask
);
4985 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4987 struct task_struct
*p
;
4988 unsigned long flags
;
4996 p
= find_process_by_pid(pid
);
5000 retval
= security_task_getscheduler(p
);
5004 rq
= task_rq_lock(p
, &flags
);
5005 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5006 task_rq_unlock(rq
, &flags
);
5016 * sys_sched_getaffinity - get the cpu affinity of a process
5017 * @pid: pid of the process
5018 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5019 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5021 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5022 unsigned long __user
*, user_mask_ptr
)
5027 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5029 if (len
& (sizeof(unsigned long)-1))
5032 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5035 ret
= sched_getaffinity(pid
, mask
);
5037 size_t retlen
= min_t(size_t, len
, cpumask_size());
5039 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5044 free_cpumask_var(mask
);
5050 * sys_sched_yield - yield the current processor to other threads.
5052 * This function yields the current CPU to other tasks. If there are no
5053 * other threads running on this CPU then this function will return.
5055 SYSCALL_DEFINE0(sched_yield
)
5057 struct rq
*rq
= this_rq_lock();
5059 schedstat_inc(rq
, yld_count
);
5060 current
->sched_class
->yield_task(rq
);
5063 * Since we are going to call schedule() anyway, there's
5064 * no need to preempt or enable interrupts:
5066 __release(rq
->lock
);
5067 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5068 do_raw_spin_unlock(&rq
->lock
);
5069 preempt_enable_no_resched();
5076 static inline int should_resched(void)
5078 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5081 static void __cond_resched(void)
5083 add_preempt_count(PREEMPT_ACTIVE
);
5085 sub_preempt_count(PREEMPT_ACTIVE
);
5088 int __sched
_cond_resched(void)
5090 if (should_resched()) {
5096 EXPORT_SYMBOL(_cond_resched
);
5099 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5100 * call schedule, and on return reacquire the lock.
5102 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5103 * operations here to prevent schedule() from being called twice (once via
5104 * spin_unlock(), once by hand).
5106 int __cond_resched_lock(spinlock_t
*lock
)
5108 int resched
= should_resched();
5111 lockdep_assert_held(lock
);
5113 if (spin_needbreak(lock
) || resched
) {
5124 EXPORT_SYMBOL(__cond_resched_lock
);
5126 int __sched
__cond_resched_softirq(void)
5128 BUG_ON(!in_softirq());
5130 if (should_resched()) {
5138 EXPORT_SYMBOL(__cond_resched_softirq
);
5141 * yield - yield the current processor to other threads.
5143 * This is a shortcut for kernel-space yielding - it marks the
5144 * thread runnable and calls sys_sched_yield().
5146 void __sched
yield(void)
5148 set_current_state(TASK_RUNNING
);
5151 EXPORT_SYMBOL(yield
);
5154 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5155 * that process accounting knows that this is a task in IO wait state.
5157 void __sched
io_schedule(void)
5159 struct rq
*rq
= raw_rq();
5161 delayacct_blkio_start();
5162 atomic_inc(&rq
->nr_iowait
);
5163 current
->in_iowait
= 1;
5165 current
->in_iowait
= 0;
5166 atomic_dec(&rq
->nr_iowait
);
5167 delayacct_blkio_end();
5169 EXPORT_SYMBOL(io_schedule
);
5171 long __sched
io_schedule_timeout(long timeout
)
5173 struct rq
*rq
= raw_rq();
5176 delayacct_blkio_start();
5177 atomic_inc(&rq
->nr_iowait
);
5178 current
->in_iowait
= 1;
5179 ret
= schedule_timeout(timeout
);
5180 current
->in_iowait
= 0;
5181 atomic_dec(&rq
->nr_iowait
);
5182 delayacct_blkio_end();
5187 * sys_sched_get_priority_max - return maximum RT priority.
5188 * @policy: scheduling class.
5190 * this syscall returns the maximum rt_priority that can be used
5191 * by a given scheduling class.
5193 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5200 ret
= MAX_USER_RT_PRIO
-1;
5212 * sys_sched_get_priority_min - return minimum RT priority.
5213 * @policy: scheduling class.
5215 * this syscall returns the minimum rt_priority that can be used
5216 * by a given scheduling class.
5218 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5236 * sys_sched_rr_get_interval - return the default timeslice of a process.
5237 * @pid: pid of the process.
5238 * @interval: userspace pointer to the timeslice value.
5240 * this syscall writes the default timeslice value of a given process
5241 * into the user-space timespec buffer. A value of '0' means infinity.
5243 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5244 struct timespec __user
*, interval
)
5246 struct task_struct
*p
;
5247 unsigned int time_slice
;
5248 unsigned long flags
;
5258 p
= find_process_by_pid(pid
);
5262 retval
= security_task_getscheduler(p
);
5266 rq
= task_rq_lock(p
, &flags
);
5267 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5268 task_rq_unlock(rq
, &flags
);
5271 jiffies_to_timespec(time_slice
, &t
);
5272 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5280 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5282 void sched_show_task(struct task_struct
*p
)
5284 unsigned long free
= 0;
5287 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5288 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5289 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5290 #if BITS_PER_LONG == 32
5291 if (state
== TASK_RUNNING
)
5292 printk(KERN_CONT
" running ");
5294 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5296 if (state
== TASK_RUNNING
)
5297 printk(KERN_CONT
" running task ");
5299 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5301 #ifdef CONFIG_DEBUG_STACK_USAGE
5302 free
= stack_not_used(p
);
5304 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5305 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5306 (unsigned long)task_thread_info(p
)->flags
);
5308 show_stack(p
, NULL
);
5311 void show_state_filter(unsigned long state_filter
)
5313 struct task_struct
*g
, *p
;
5315 #if BITS_PER_LONG == 32
5317 " task PC stack pid father\n");
5320 " task PC stack pid father\n");
5322 read_lock(&tasklist_lock
);
5323 do_each_thread(g
, p
) {
5325 * reset the NMI-timeout, listing all files on a slow
5326 * console might take alot of time:
5328 touch_nmi_watchdog();
5329 if (!state_filter
|| (p
->state
& state_filter
))
5331 } while_each_thread(g
, p
);
5333 touch_all_softlockup_watchdogs();
5335 #ifdef CONFIG_SCHED_DEBUG
5336 sysrq_sched_debug_show();
5338 read_unlock(&tasklist_lock
);
5340 * Only show locks if all tasks are dumped:
5343 debug_show_all_locks();
5346 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5348 idle
->sched_class
= &idle_sched_class
;
5352 * init_idle - set up an idle thread for a given CPU
5353 * @idle: task in question
5354 * @cpu: cpu the idle task belongs to
5356 * NOTE: this function does not set the idle thread's NEED_RESCHED
5357 * flag, to make booting more robust.
5359 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5361 struct rq
*rq
= cpu_rq(cpu
);
5362 unsigned long flags
;
5364 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5367 idle
->state
= TASK_RUNNING
;
5368 idle
->se
.exec_start
= sched_clock();
5370 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5371 __set_task_cpu(idle
, cpu
);
5373 rq
->curr
= rq
->idle
= idle
;
5374 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5377 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5379 /* Set the preempt count _outside_ the spinlocks! */
5380 #if defined(CONFIG_PREEMPT)
5381 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5383 task_thread_info(idle
)->preempt_count
= 0;
5386 * The idle tasks have their own, simple scheduling class:
5388 idle
->sched_class
= &idle_sched_class
;
5389 ftrace_graph_init_task(idle
);
5393 * In a system that switches off the HZ timer nohz_cpu_mask
5394 * indicates which cpus entered this state. This is used
5395 * in the rcu update to wait only for active cpus. For system
5396 * which do not switch off the HZ timer nohz_cpu_mask should
5397 * always be CPU_BITS_NONE.
5399 cpumask_var_t nohz_cpu_mask
;
5402 * Increase the granularity value when there are more CPUs,
5403 * because with more CPUs the 'effective latency' as visible
5404 * to users decreases. But the relationship is not linear,
5405 * so pick a second-best guess by going with the log2 of the
5408 * This idea comes from the SD scheduler of Con Kolivas:
5410 static int get_update_sysctl_factor(void)
5412 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5413 unsigned int factor
;
5415 switch (sysctl_sched_tunable_scaling
) {
5416 case SCHED_TUNABLESCALING_NONE
:
5419 case SCHED_TUNABLESCALING_LINEAR
:
5422 case SCHED_TUNABLESCALING_LOG
:
5424 factor
= 1 + ilog2(cpus
);
5431 static void update_sysctl(void)
5433 unsigned int factor
= get_update_sysctl_factor();
5435 #define SET_SYSCTL(name) \
5436 (sysctl_##name = (factor) * normalized_sysctl_##name)
5437 SET_SYSCTL(sched_min_granularity
);
5438 SET_SYSCTL(sched_latency
);
5439 SET_SYSCTL(sched_wakeup_granularity
);
5440 SET_SYSCTL(sched_shares_ratelimit
);
5444 static inline void sched_init_granularity(void)
5451 * This is how migration works:
5453 * 1) we invoke migration_cpu_stop() on the target CPU using
5455 * 2) stopper starts to run (implicitly forcing the migrated thread
5457 * 3) it checks whether the migrated task is still in the wrong runqueue.
5458 * 4) if it's in the wrong runqueue then the migration thread removes
5459 * it and puts it into the right queue.
5460 * 5) stopper completes and stop_one_cpu() returns and the migration
5465 * Change a given task's CPU affinity. Migrate the thread to a
5466 * proper CPU and schedule it away if the CPU it's executing on
5467 * is removed from the allowed bitmask.
5469 * NOTE: the caller must have a valid reference to the task, the
5470 * task must not exit() & deallocate itself prematurely. The
5471 * call is not atomic; no spinlocks may be held.
5473 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5475 unsigned long flags
;
5477 unsigned int dest_cpu
;
5481 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5482 * drop the rq->lock and still rely on ->cpus_allowed.
5485 while (task_is_waking(p
))
5487 rq
= task_rq_lock(p
, &flags
);
5488 if (task_is_waking(p
)) {
5489 task_rq_unlock(rq
, &flags
);
5493 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5498 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5499 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5504 if (p
->sched_class
->set_cpus_allowed
)
5505 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5507 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5508 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5511 /* Can the task run on the task's current CPU? If so, we're done */
5512 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5515 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5516 if (migrate_task(p
, dest_cpu
)) {
5517 struct migration_arg arg
= { p
, dest_cpu
};
5518 /* Need help from migration thread: drop lock and wait. */
5519 task_rq_unlock(rq
, &flags
);
5520 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5521 tlb_migrate_finish(p
->mm
);
5525 task_rq_unlock(rq
, &flags
);
5529 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5532 * Move (not current) task off this cpu, onto dest cpu. We're doing
5533 * this because either it can't run here any more (set_cpus_allowed()
5534 * away from this CPU, or CPU going down), or because we're
5535 * attempting to rebalance this task on exec (sched_exec).
5537 * So we race with normal scheduler movements, but that's OK, as long
5538 * as the task is no longer on this CPU.
5540 * Returns non-zero if task was successfully migrated.
5542 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5544 struct rq
*rq_dest
, *rq_src
;
5547 if (unlikely(!cpu_active(dest_cpu
)))
5550 rq_src
= cpu_rq(src_cpu
);
5551 rq_dest
= cpu_rq(dest_cpu
);
5553 double_rq_lock(rq_src
, rq_dest
);
5554 /* Already moved. */
5555 if (task_cpu(p
) != src_cpu
)
5557 /* Affinity changed (again). */
5558 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5562 * If we're not on a rq, the next wake-up will ensure we're
5566 deactivate_task(rq_src
, p
, 0);
5567 set_task_cpu(p
, dest_cpu
);
5568 activate_task(rq_dest
, p
, 0);
5569 check_preempt_curr(rq_dest
, p
, 0);
5574 double_rq_unlock(rq_src
, rq_dest
);
5579 * migration_cpu_stop - this will be executed by a highprio stopper thread
5580 * and performs thread migration by bumping thread off CPU then
5581 * 'pushing' onto another runqueue.
5583 static int migration_cpu_stop(void *data
)
5585 struct migration_arg
*arg
= data
;
5588 * The original target cpu might have gone down and we might
5589 * be on another cpu but it doesn't matter.
5591 local_irq_disable();
5592 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5597 #ifdef CONFIG_HOTPLUG_CPU
5599 * Figure out where task on dead CPU should go, use force if necessary.
5601 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5603 struct rq
*rq
= cpu_rq(dead_cpu
);
5604 int needs_cpu
, uninitialized_var(dest_cpu
);
5605 unsigned long flags
;
5607 local_irq_save(flags
);
5609 raw_spin_lock(&rq
->lock
);
5610 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5612 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5613 raw_spin_unlock(&rq
->lock
);
5615 * It can only fail if we race with set_cpus_allowed(),
5616 * in the racer should migrate the task anyway.
5619 __migrate_task(p
, dead_cpu
, dest_cpu
);
5620 local_irq_restore(flags
);
5624 * While a dead CPU has no uninterruptible tasks queued at this point,
5625 * it might still have a nonzero ->nr_uninterruptible counter, because
5626 * for performance reasons the counter is not stricly tracking tasks to
5627 * their home CPUs. So we just add the counter to another CPU's counter,
5628 * to keep the global sum constant after CPU-down:
5630 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5632 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5633 unsigned long flags
;
5635 local_irq_save(flags
);
5636 double_rq_lock(rq_src
, rq_dest
);
5637 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5638 rq_src
->nr_uninterruptible
= 0;
5639 double_rq_unlock(rq_src
, rq_dest
);
5640 local_irq_restore(flags
);
5643 /* Run through task list and migrate tasks from the dead cpu. */
5644 static void migrate_live_tasks(int src_cpu
)
5646 struct task_struct
*p
, *t
;
5648 read_lock(&tasklist_lock
);
5650 do_each_thread(t
, p
) {
5654 if (task_cpu(p
) == src_cpu
)
5655 move_task_off_dead_cpu(src_cpu
, p
);
5656 } while_each_thread(t
, p
);
5658 read_unlock(&tasklist_lock
);
5662 * Schedules idle task to be the next runnable task on current CPU.
5663 * It does so by boosting its priority to highest possible.
5664 * Used by CPU offline code.
5666 void sched_idle_next(void)
5668 int this_cpu
= smp_processor_id();
5669 struct rq
*rq
= cpu_rq(this_cpu
);
5670 struct task_struct
*p
= rq
->idle
;
5671 unsigned long flags
;
5673 /* cpu has to be offline */
5674 BUG_ON(cpu_online(this_cpu
));
5677 * Strictly not necessary since rest of the CPUs are stopped by now
5678 * and interrupts disabled on the current cpu.
5680 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5682 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5684 activate_task(rq
, p
, 0);
5686 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5690 * Ensures that the idle task is using init_mm right before its cpu goes
5693 void idle_task_exit(void)
5695 struct mm_struct
*mm
= current
->active_mm
;
5697 BUG_ON(cpu_online(smp_processor_id()));
5700 switch_mm(mm
, &init_mm
, current
);
5704 /* called under rq->lock with disabled interrupts */
5705 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5707 struct rq
*rq
= cpu_rq(dead_cpu
);
5709 /* Must be exiting, otherwise would be on tasklist. */
5710 BUG_ON(!p
->exit_state
);
5712 /* Cannot have done final schedule yet: would have vanished. */
5713 BUG_ON(p
->state
== TASK_DEAD
);
5718 * Drop lock around migration; if someone else moves it,
5719 * that's OK. No task can be added to this CPU, so iteration is
5722 raw_spin_unlock_irq(&rq
->lock
);
5723 move_task_off_dead_cpu(dead_cpu
, p
);
5724 raw_spin_lock_irq(&rq
->lock
);
5729 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5730 static void migrate_dead_tasks(unsigned int dead_cpu
)
5732 struct rq
*rq
= cpu_rq(dead_cpu
);
5733 struct task_struct
*next
;
5736 if (!rq
->nr_running
)
5738 next
= pick_next_task(rq
);
5741 next
->sched_class
->put_prev_task(rq
, next
);
5742 migrate_dead(dead_cpu
, next
);
5748 * remove the tasks which were accounted by rq from calc_load_tasks.
5750 static void calc_global_load_remove(struct rq
*rq
)
5752 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5753 rq
->calc_load_active
= 0;
5755 #endif /* CONFIG_HOTPLUG_CPU */
5757 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5759 static struct ctl_table sd_ctl_dir
[] = {
5761 .procname
= "sched_domain",
5767 static struct ctl_table sd_ctl_root
[] = {
5769 .procname
= "kernel",
5771 .child
= sd_ctl_dir
,
5776 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5778 struct ctl_table
*entry
=
5779 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5784 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5786 struct ctl_table
*entry
;
5789 * In the intermediate directories, both the child directory and
5790 * procname are dynamically allocated and could fail but the mode
5791 * will always be set. In the lowest directory the names are
5792 * static strings and all have proc handlers.
5794 for (entry
= *tablep
; entry
->mode
; entry
++) {
5796 sd_free_ctl_entry(&entry
->child
);
5797 if (entry
->proc_handler
== NULL
)
5798 kfree(entry
->procname
);
5806 set_table_entry(struct ctl_table
*entry
,
5807 const char *procname
, void *data
, int maxlen
,
5808 mode_t mode
, proc_handler
*proc_handler
)
5810 entry
->procname
= procname
;
5812 entry
->maxlen
= maxlen
;
5814 entry
->proc_handler
= proc_handler
;
5817 static struct ctl_table
*
5818 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5820 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5825 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5826 sizeof(long), 0644, proc_doulongvec_minmax
);
5827 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5828 sizeof(long), 0644, proc_doulongvec_minmax
);
5829 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5830 sizeof(int), 0644, proc_dointvec_minmax
);
5831 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5832 sizeof(int), 0644, proc_dointvec_minmax
);
5833 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5834 sizeof(int), 0644, proc_dointvec_minmax
);
5835 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5836 sizeof(int), 0644, proc_dointvec_minmax
);
5837 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5838 sizeof(int), 0644, proc_dointvec_minmax
);
5839 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5840 sizeof(int), 0644, proc_dointvec_minmax
);
5841 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5842 sizeof(int), 0644, proc_dointvec_minmax
);
5843 set_table_entry(&table
[9], "cache_nice_tries",
5844 &sd
->cache_nice_tries
,
5845 sizeof(int), 0644, proc_dointvec_minmax
);
5846 set_table_entry(&table
[10], "flags", &sd
->flags
,
5847 sizeof(int), 0644, proc_dointvec_minmax
);
5848 set_table_entry(&table
[11], "name", sd
->name
,
5849 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5850 /* &table[12] is terminator */
5855 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5857 struct ctl_table
*entry
, *table
;
5858 struct sched_domain
*sd
;
5859 int domain_num
= 0, i
;
5862 for_each_domain(cpu
, sd
)
5864 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5869 for_each_domain(cpu
, sd
) {
5870 snprintf(buf
, 32, "domain%d", i
);
5871 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5873 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5880 static struct ctl_table_header
*sd_sysctl_header
;
5881 static void register_sched_domain_sysctl(void)
5883 int i
, cpu_num
= num_possible_cpus();
5884 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5887 WARN_ON(sd_ctl_dir
[0].child
);
5888 sd_ctl_dir
[0].child
= entry
;
5893 for_each_possible_cpu(i
) {
5894 snprintf(buf
, 32, "cpu%d", i
);
5895 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5897 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5901 WARN_ON(sd_sysctl_header
);
5902 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5905 /* may be called multiple times per register */
5906 static void unregister_sched_domain_sysctl(void)
5908 if (sd_sysctl_header
)
5909 unregister_sysctl_table(sd_sysctl_header
);
5910 sd_sysctl_header
= NULL
;
5911 if (sd_ctl_dir
[0].child
)
5912 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5915 static void register_sched_domain_sysctl(void)
5918 static void unregister_sched_domain_sysctl(void)
5923 static void set_rq_online(struct rq
*rq
)
5926 const struct sched_class
*class;
5928 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5931 for_each_class(class) {
5932 if (class->rq_online
)
5933 class->rq_online(rq
);
5938 static void set_rq_offline(struct rq
*rq
)
5941 const struct sched_class
*class;
5943 for_each_class(class) {
5944 if (class->rq_offline
)
5945 class->rq_offline(rq
);
5948 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5954 * migration_call - callback that gets triggered when a CPU is added.
5955 * Here we can start up the necessary migration thread for the new CPU.
5957 static int __cpuinit
5958 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5960 int cpu
= (long)hcpu
;
5961 unsigned long flags
;
5962 struct rq
*rq
= cpu_rq(cpu
);
5966 case CPU_UP_PREPARE
:
5967 case CPU_UP_PREPARE_FROZEN
:
5968 rq
->calc_load_update
= calc_load_update
;
5972 case CPU_ONLINE_FROZEN
:
5973 /* Update our root-domain */
5974 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5976 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5980 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5983 #ifdef CONFIG_HOTPLUG_CPU
5985 case CPU_DEAD_FROZEN
:
5986 migrate_live_tasks(cpu
);
5987 /* Idle task back to normal (off runqueue, low prio) */
5988 raw_spin_lock_irq(&rq
->lock
);
5989 deactivate_task(rq
, rq
->idle
, 0);
5990 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5991 rq
->idle
->sched_class
= &idle_sched_class
;
5992 migrate_dead_tasks(cpu
);
5993 raw_spin_unlock_irq(&rq
->lock
);
5994 migrate_nr_uninterruptible(rq
);
5995 BUG_ON(rq
->nr_running
!= 0);
5996 calc_global_load_remove(rq
);
6000 case CPU_DYING_FROZEN
:
6001 /* Update our root-domain */
6002 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6004 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6007 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6015 * Register at high priority so that task migration (migrate_all_tasks)
6016 * happens before everything else. This has to be lower priority than
6017 * the notifier in the perf_event subsystem, though.
6019 static struct notifier_block __cpuinitdata migration_notifier
= {
6020 .notifier_call
= migration_call
,
6021 .priority
= CPU_PRI_MIGRATION
,
6024 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6025 unsigned long action
, void *hcpu
)
6027 switch (action
& ~CPU_TASKS_FROZEN
) {
6029 case CPU_DOWN_FAILED
:
6030 set_cpu_active((long)hcpu
, true);
6037 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6038 unsigned long action
, void *hcpu
)
6040 switch (action
& ~CPU_TASKS_FROZEN
) {
6041 case CPU_DOWN_PREPARE
:
6042 set_cpu_active((long)hcpu
, false);
6049 static int __init
migration_init(void)
6051 void *cpu
= (void *)(long)smp_processor_id();
6054 /* Initialize migration for the boot CPU */
6055 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6056 BUG_ON(err
== NOTIFY_BAD
);
6057 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6058 register_cpu_notifier(&migration_notifier
);
6060 /* Register cpu active notifiers */
6061 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6062 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6066 early_initcall(migration_init
);
6071 #ifdef CONFIG_SCHED_DEBUG
6073 static __read_mostly
int sched_domain_debug_enabled
;
6075 static int __init
sched_domain_debug_setup(char *str
)
6077 sched_domain_debug_enabled
= 1;
6081 early_param("sched_debug", sched_domain_debug_setup
);
6083 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6084 struct cpumask
*groupmask
)
6086 struct sched_group
*group
= sd
->groups
;
6089 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6090 cpumask_clear(groupmask
);
6092 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6094 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6095 printk("does not load-balance\n");
6097 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6102 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6104 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6105 printk(KERN_ERR
"ERROR: domain->span does not contain "
6108 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6109 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6113 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6117 printk(KERN_ERR
"ERROR: group is NULL\n");
6121 if (!group
->cpu_power
) {
6122 printk(KERN_CONT
"\n");
6123 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6128 if (!cpumask_weight(sched_group_cpus(group
))) {
6129 printk(KERN_CONT
"\n");
6130 printk(KERN_ERR
"ERROR: empty group\n");
6134 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6135 printk(KERN_CONT
"\n");
6136 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6140 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6142 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6144 printk(KERN_CONT
" %s", str
);
6145 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6146 printk(KERN_CONT
" (cpu_power = %d)",
6150 group
= group
->next
;
6151 } while (group
!= sd
->groups
);
6152 printk(KERN_CONT
"\n");
6154 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6155 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6158 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6159 printk(KERN_ERR
"ERROR: parent span is not a superset "
6160 "of domain->span\n");
6164 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6166 cpumask_var_t groupmask
;
6169 if (!sched_domain_debug_enabled
)
6173 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6177 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6179 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6180 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6185 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6192 free_cpumask_var(groupmask
);
6194 #else /* !CONFIG_SCHED_DEBUG */
6195 # define sched_domain_debug(sd, cpu) do { } while (0)
6196 #endif /* CONFIG_SCHED_DEBUG */
6198 static int sd_degenerate(struct sched_domain
*sd
)
6200 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6203 /* Following flags need at least 2 groups */
6204 if (sd
->flags
& (SD_LOAD_BALANCE
|
6205 SD_BALANCE_NEWIDLE
|
6209 SD_SHARE_PKG_RESOURCES
)) {
6210 if (sd
->groups
!= sd
->groups
->next
)
6214 /* Following flags don't use groups */
6215 if (sd
->flags
& (SD_WAKE_AFFINE
))
6222 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6224 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6226 if (sd_degenerate(parent
))
6229 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6232 /* Flags needing groups don't count if only 1 group in parent */
6233 if (parent
->groups
== parent
->groups
->next
) {
6234 pflags
&= ~(SD_LOAD_BALANCE
|
6235 SD_BALANCE_NEWIDLE
|
6239 SD_SHARE_PKG_RESOURCES
);
6240 if (nr_node_ids
== 1)
6241 pflags
&= ~SD_SERIALIZE
;
6243 if (~cflags
& pflags
)
6249 static void free_rootdomain(struct root_domain
*rd
)
6251 synchronize_sched();
6253 cpupri_cleanup(&rd
->cpupri
);
6255 free_cpumask_var(rd
->rto_mask
);
6256 free_cpumask_var(rd
->online
);
6257 free_cpumask_var(rd
->span
);
6261 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6263 struct root_domain
*old_rd
= NULL
;
6264 unsigned long flags
;
6266 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6271 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6274 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6277 * If we dont want to free the old_rt yet then
6278 * set old_rd to NULL to skip the freeing later
6281 if (!atomic_dec_and_test(&old_rd
->refcount
))
6285 atomic_inc(&rd
->refcount
);
6288 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6289 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6292 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6295 free_rootdomain(old_rd
);
6298 static int init_rootdomain(struct root_domain
*rd
)
6300 memset(rd
, 0, sizeof(*rd
));
6302 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6304 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6306 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6309 if (cpupri_init(&rd
->cpupri
) != 0)
6314 free_cpumask_var(rd
->rto_mask
);
6316 free_cpumask_var(rd
->online
);
6318 free_cpumask_var(rd
->span
);
6323 static void init_defrootdomain(void)
6325 init_rootdomain(&def_root_domain
);
6327 atomic_set(&def_root_domain
.refcount
, 1);
6330 static struct root_domain
*alloc_rootdomain(void)
6332 struct root_domain
*rd
;
6334 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6338 if (init_rootdomain(rd
) != 0) {
6347 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6348 * hold the hotplug lock.
6351 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6353 struct rq
*rq
= cpu_rq(cpu
);
6354 struct sched_domain
*tmp
;
6356 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6357 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6359 /* Remove the sched domains which do not contribute to scheduling. */
6360 for (tmp
= sd
; tmp
; ) {
6361 struct sched_domain
*parent
= tmp
->parent
;
6365 if (sd_parent_degenerate(tmp
, parent
)) {
6366 tmp
->parent
= parent
->parent
;
6368 parent
->parent
->child
= tmp
;
6373 if (sd
&& sd_degenerate(sd
)) {
6379 sched_domain_debug(sd
, cpu
);
6381 rq_attach_root(rq
, rd
);
6382 rcu_assign_pointer(rq
->sd
, sd
);
6385 /* cpus with isolated domains */
6386 static cpumask_var_t cpu_isolated_map
;
6388 /* Setup the mask of cpus configured for isolated domains */
6389 static int __init
isolated_cpu_setup(char *str
)
6391 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6392 cpulist_parse(str
, cpu_isolated_map
);
6396 __setup("isolcpus=", isolated_cpu_setup
);
6399 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6400 * to a function which identifies what group(along with sched group) a CPU
6401 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6402 * (due to the fact that we keep track of groups covered with a struct cpumask).
6404 * init_sched_build_groups will build a circular linked list of the groups
6405 * covered by the given span, and will set each group's ->cpumask correctly,
6406 * and ->cpu_power to 0.
6409 init_sched_build_groups(const struct cpumask
*span
,
6410 const struct cpumask
*cpu_map
,
6411 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6412 struct sched_group
**sg
,
6413 struct cpumask
*tmpmask
),
6414 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6416 struct sched_group
*first
= NULL
, *last
= NULL
;
6419 cpumask_clear(covered
);
6421 for_each_cpu(i
, span
) {
6422 struct sched_group
*sg
;
6423 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6426 if (cpumask_test_cpu(i
, covered
))
6429 cpumask_clear(sched_group_cpus(sg
));
6432 for_each_cpu(j
, span
) {
6433 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6436 cpumask_set_cpu(j
, covered
);
6437 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6448 #define SD_NODES_PER_DOMAIN 16
6453 * find_next_best_node - find the next node to include in a sched_domain
6454 * @node: node whose sched_domain we're building
6455 * @used_nodes: nodes already in the sched_domain
6457 * Find the next node to include in a given scheduling domain. Simply
6458 * finds the closest node not already in the @used_nodes map.
6460 * Should use nodemask_t.
6462 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6464 int i
, n
, val
, min_val
, best_node
= 0;
6468 for (i
= 0; i
< nr_node_ids
; i
++) {
6469 /* Start at @node */
6470 n
= (node
+ i
) % nr_node_ids
;
6472 if (!nr_cpus_node(n
))
6475 /* Skip already used nodes */
6476 if (node_isset(n
, *used_nodes
))
6479 /* Simple min distance search */
6480 val
= node_distance(node
, n
);
6482 if (val
< min_val
) {
6488 node_set(best_node
, *used_nodes
);
6493 * sched_domain_node_span - get a cpumask for a node's sched_domain
6494 * @node: node whose cpumask we're constructing
6495 * @span: resulting cpumask
6497 * Given a node, construct a good cpumask for its sched_domain to span. It
6498 * should be one that prevents unnecessary balancing, but also spreads tasks
6501 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6503 nodemask_t used_nodes
;
6506 cpumask_clear(span
);
6507 nodes_clear(used_nodes
);
6509 cpumask_or(span
, span
, cpumask_of_node(node
));
6510 node_set(node
, used_nodes
);
6512 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6513 int next_node
= find_next_best_node(node
, &used_nodes
);
6515 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6518 #endif /* CONFIG_NUMA */
6520 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6523 * The cpus mask in sched_group and sched_domain hangs off the end.
6525 * ( See the the comments in include/linux/sched.h:struct sched_group
6526 * and struct sched_domain. )
6528 struct static_sched_group
{
6529 struct sched_group sg
;
6530 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6533 struct static_sched_domain
{
6534 struct sched_domain sd
;
6535 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6541 cpumask_var_t domainspan
;
6542 cpumask_var_t covered
;
6543 cpumask_var_t notcovered
;
6545 cpumask_var_t nodemask
;
6546 cpumask_var_t this_sibling_map
;
6547 cpumask_var_t this_core_map
;
6548 cpumask_var_t this_book_map
;
6549 cpumask_var_t send_covered
;
6550 cpumask_var_t tmpmask
;
6551 struct sched_group
**sched_group_nodes
;
6552 struct root_domain
*rd
;
6556 sa_sched_groups
= 0,
6562 sa_this_sibling_map
,
6564 sa_sched_group_nodes
,
6574 * SMT sched-domains:
6576 #ifdef CONFIG_SCHED_SMT
6577 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6578 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6581 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6582 struct sched_group
**sg
, struct cpumask
*unused
)
6585 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6588 #endif /* CONFIG_SCHED_SMT */
6591 * multi-core sched-domains:
6593 #ifdef CONFIG_SCHED_MC
6594 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6595 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6598 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6599 struct sched_group
**sg
, struct cpumask
*mask
)
6602 #ifdef CONFIG_SCHED_SMT
6603 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6604 group
= cpumask_first(mask
);
6609 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6612 #endif /* CONFIG_SCHED_MC */
6615 * book sched-domains:
6617 #ifdef CONFIG_SCHED_BOOK
6618 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
6619 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
6622 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
6623 struct sched_group
**sg
, struct cpumask
*mask
)
6626 #ifdef CONFIG_SCHED_MC
6627 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6628 group
= cpumask_first(mask
);
6629 #elif defined(CONFIG_SCHED_SMT)
6630 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6631 group
= cpumask_first(mask
);
6634 *sg
= &per_cpu(sched_group_book
, group
).sg
;
6637 #endif /* CONFIG_SCHED_BOOK */
6639 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6640 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6643 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6644 struct sched_group
**sg
, struct cpumask
*mask
)
6647 #ifdef CONFIG_SCHED_BOOK
6648 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
6649 group
= cpumask_first(mask
);
6650 #elif defined(CONFIG_SCHED_MC)
6651 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6652 group
= cpumask_first(mask
);
6653 #elif defined(CONFIG_SCHED_SMT)
6654 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6655 group
= cpumask_first(mask
);
6660 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6666 * The init_sched_build_groups can't handle what we want to do with node
6667 * groups, so roll our own. Now each node has its own list of groups which
6668 * gets dynamically allocated.
6670 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6671 static struct sched_group
***sched_group_nodes_bycpu
;
6673 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6674 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6676 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6677 struct sched_group
**sg
,
6678 struct cpumask
*nodemask
)
6682 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6683 group
= cpumask_first(nodemask
);
6686 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6690 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6692 struct sched_group
*sg
= group_head
;
6698 for_each_cpu(j
, sched_group_cpus(sg
)) {
6699 struct sched_domain
*sd
;
6701 sd
= &per_cpu(phys_domains
, j
).sd
;
6702 if (j
!= group_first_cpu(sd
->groups
)) {
6704 * Only add "power" once for each
6710 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6713 } while (sg
!= group_head
);
6716 static int build_numa_sched_groups(struct s_data
*d
,
6717 const struct cpumask
*cpu_map
, int num
)
6719 struct sched_domain
*sd
;
6720 struct sched_group
*sg
, *prev
;
6723 cpumask_clear(d
->covered
);
6724 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6725 if (cpumask_empty(d
->nodemask
)) {
6726 d
->sched_group_nodes
[num
] = NULL
;
6730 sched_domain_node_span(num
, d
->domainspan
);
6731 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6733 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6736 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6740 d
->sched_group_nodes
[num
] = sg
;
6742 for_each_cpu(j
, d
->nodemask
) {
6743 sd
= &per_cpu(node_domains
, j
).sd
;
6748 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6750 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6753 for (j
= 0; j
< nr_node_ids
; j
++) {
6754 n
= (num
+ j
) % nr_node_ids
;
6755 cpumask_complement(d
->notcovered
, d
->covered
);
6756 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6757 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6758 if (cpumask_empty(d
->tmpmask
))
6760 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6761 if (cpumask_empty(d
->tmpmask
))
6763 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6767 "Can not alloc domain group for node %d\n", j
);
6771 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6772 sg
->next
= prev
->next
;
6773 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6780 #endif /* CONFIG_NUMA */
6783 /* Free memory allocated for various sched_group structures */
6784 static void free_sched_groups(const struct cpumask
*cpu_map
,
6785 struct cpumask
*nodemask
)
6789 for_each_cpu(cpu
, cpu_map
) {
6790 struct sched_group
**sched_group_nodes
6791 = sched_group_nodes_bycpu
[cpu
];
6793 if (!sched_group_nodes
)
6796 for (i
= 0; i
< nr_node_ids
; i
++) {
6797 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6799 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6800 if (cpumask_empty(nodemask
))
6810 if (oldsg
!= sched_group_nodes
[i
])
6813 kfree(sched_group_nodes
);
6814 sched_group_nodes_bycpu
[cpu
] = NULL
;
6817 #else /* !CONFIG_NUMA */
6818 static void free_sched_groups(const struct cpumask
*cpu_map
,
6819 struct cpumask
*nodemask
)
6822 #endif /* CONFIG_NUMA */
6825 * Initialize sched groups cpu_power.
6827 * cpu_power indicates the capacity of sched group, which is used while
6828 * distributing the load between different sched groups in a sched domain.
6829 * Typically cpu_power for all the groups in a sched domain will be same unless
6830 * there are asymmetries in the topology. If there are asymmetries, group
6831 * having more cpu_power will pickup more load compared to the group having
6834 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6836 struct sched_domain
*child
;
6837 struct sched_group
*group
;
6841 WARN_ON(!sd
|| !sd
->groups
);
6843 if (cpu
!= group_first_cpu(sd
->groups
))
6848 sd
->groups
->cpu_power
= 0;
6851 power
= SCHED_LOAD_SCALE
;
6852 weight
= cpumask_weight(sched_domain_span(sd
));
6854 * SMT siblings share the power of a single core.
6855 * Usually multiple threads get a better yield out of
6856 * that one core than a single thread would have,
6857 * reflect that in sd->smt_gain.
6859 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6860 power
*= sd
->smt_gain
;
6862 power
>>= SCHED_LOAD_SHIFT
;
6864 sd
->groups
->cpu_power
+= power
;
6869 * Add cpu_power of each child group to this groups cpu_power.
6871 group
= child
->groups
;
6873 sd
->groups
->cpu_power
+= group
->cpu_power
;
6874 group
= group
->next
;
6875 } while (group
!= child
->groups
);
6879 * Initializers for schedule domains
6880 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6883 #ifdef CONFIG_SCHED_DEBUG
6884 # define SD_INIT_NAME(sd, type) sd->name = #type
6886 # define SD_INIT_NAME(sd, type) do { } while (0)
6889 #define SD_INIT(sd, type) sd_init_##type(sd)
6891 #define SD_INIT_FUNC(type) \
6892 static noinline void sd_init_##type(struct sched_domain *sd) \
6894 memset(sd, 0, sizeof(*sd)); \
6895 *sd = SD_##type##_INIT; \
6896 sd->level = SD_LV_##type; \
6897 SD_INIT_NAME(sd, type); \
6902 SD_INIT_FUNC(ALLNODES
)
6905 #ifdef CONFIG_SCHED_SMT
6906 SD_INIT_FUNC(SIBLING
)
6908 #ifdef CONFIG_SCHED_MC
6911 #ifdef CONFIG_SCHED_BOOK
6915 static int default_relax_domain_level
= -1;
6917 static int __init
setup_relax_domain_level(char *str
)
6921 val
= simple_strtoul(str
, NULL
, 0);
6922 if (val
< SD_LV_MAX
)
6923 default_relax_domain_level
= val
;
6927 __setup("relax_domain_level=", setup_relax_domain_level
);
6929 static void set_domain_attribute(struct sched_domain
*sd
,
6930 struct sched_domain_attr
*attr
)
6934 if (!attr
|| attr
->relax_domain_level
< 0) {
6935 if (default_relax_domain_level
< 0)
6938 request
= default_relax_domain_level
;
6940 request
= attr
->relax_domain_level
;
6941 if (request
< sd
->level
) {
6942 /* turn off idle balance on this domain */
6943 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6945 /* turn on idle balance on this domain */
6946 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6950 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6951 const struct cpumask
*cpu_map
)
6954 case sa_sched_groups
:
6955 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6956 d
->sched_group_nodes
= NULL
;
6958 free_rootdomain(d
->rd
); /* fall through */
6960 free_cpumask_var(d
->tmpmask
); /* fall through */
6961 case sa_send_covered
:
6962 free_cpumask_var(d
->send_covered
); /* fall through */
6963 case sa_this_book_map
:
6964 free_cpumask_var(d
->this_book_map
); /* fall through */
6965 case sa_this_core_map
:
6966 free_cpumask_var(d
->this_core_map
); /* fall through */
6967 case sa_this_sibling_map
:
6968 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6970 free_cpumask_var(d
->nodemask
); /* fall through */
6971 case sa_sched_group_nodes
:
6973 kfree(d
->sched_group_nodes
); /* fall through */
6975 free_cpumask_var(d
->notcovered
); /* fall through */
6977 free_cpumask_var(d
->covered
); /* fall through */
6979 free_cpumask_var(d
->domainspan
); /* fall through */
6986 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6987 const struct cpumask
*cpu_map
)
6990 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6992 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6993 return sa_domainspan
;
6994 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6996 /* Allocate the per-node list of sched groups */
6997 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6998 sizeof(struct sched_group
*), GFP_KERNEL
);
6999 if (!d
->sched_group_nodes
) {
7000 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7001 return sa_notcovered
;
7003 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
7005 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
7006 return sa_sched_group_nodes
;
7007 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
7009 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
7010 return sa_this_sibling_map
;
7011 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
7012 return sa_this_core_map
;
7013 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
7014 return sa_this_book_map
;
7015 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
7016 return sa_send_covered
;
7017 d
->rd
= alloc_rootdomain();
7019 printk(KERN_WARNING
"Cannot alloc root domain\n");
7022 return sa_rootdomain
;
7025 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
7026 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
7028 struct sched_domain
*sd
= NULL
;
7030 struct sched_domain
*parent
;
7033 if (cpumask_weight(cpu_map
) >
7034 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7035 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7036 SD_INIT(sd
, ALLNODES
);
7037 set_domain_attribute(sd
, attr
);
7038 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7039 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7044 sd
= &per_cpu(node_domains
, i
).sd
;
7046 set_domain_attribute(sd
, attr
);
7047 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7048 sd
->parent
= parent
;
7051 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7056 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7057 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7058 struct sched_domain
*parent
, int i
)
7060 struct sched_domain
*sd
;
7061 sd
= &per_cpu(phys_domains
, i
).sd
;
7063 set_domain_attribute(sd
, attr
);
7064 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7065 sd
->parent
= parent
;
7068 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7072 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7073 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7074 struct sched_domain
*parent
, int i
)
7076 struct sched_domain
*sd
= parent
;
7077 #ifdef CONFIG_SCHED_BOOK
7078 sd
= &per_cpu(book_domains
, i
).sd
;
7080 set_domain_attribute(sd
, attr
);
7081 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7082 sd
->parent
= parent
;
7084 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7089 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7090 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7091 struct sched_domain
*parent
, int i
)
7093 struct sched_domain
*sd
= parent
;
7094 #ifdef CONFIG_SCHED_MC
7095 sd
= &per_cpu(core_domains
, i
).sd
;
7097 set_domain_attribute(sd
, attr
);
7098 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7099 sd
->parent
= parent
;
7101 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7106 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7107 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7108 struct sched_domain
*parent
, int i
)
7110 struct sched_domain
*sd
= parent
;
7111 #ifdef CONFIG_SCHED_SMT
7112 sd
= &per_cpu(cpu_domains
, i
).sd
;
7113 SD_INIT(sd
, SIBLING
);
7114 set_domain_attribute(sd
, attr
);
7115 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7116 sd
->parent
= parent
;
7118 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7123 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7124 const struct cpumask
*cpu_map
, int cpu
)
7127 #ifdef CONFIG_SCHED_SMT
7128 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7129 cpumask_and(d
->this_sibling_map
, cpu_map
,
7130 topology_thread_cpumask(cpu
));
7131 if (cpu
== cpumask_first(d
->this_sibling_map
))
7132 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7134 d
->send_covered
, d
->tmpmask
);
7137 #ifdef CONFIG_SCHED_MC
7138 case SD_LV_MC
: /* set up multi-core groups */
7139 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7140 if (cpu
== cpumask_first(d
->this_core_map
))
7141 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7143 d
->send_covered
, d
->tmpmask
);
7146 #ifdef CONFIG_SCHED_BOOK
7147 case SD_LV_BOOK
: /* set up book groups */
7148 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7149 if (cpu
== cpumask_first(d
->this_book_map
))
7150 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7152 d
->send_covered
, d
->tmpmask
);
7155 case SD_LV_CPU
: /* set up physical groups */
7156 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7157 if (!cpumask_empty(d
->nodemask
))
7158 init_sched_build_groups(d
->nodemask
, cpu_map
,
7160 d
->send_covered
, d
->tmpmask
);
7163 case SD_LV_ALLNODES
:
7164 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7165 d
->send_covered
, d
->tmpmask
);
7174 * Build sched domains for a given set of cpus and attach the sched domains
7175 * to the individual cpus
7177 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7178 struct sched_domain_attr
*attr
)
7180 enum s_alloc alloc_state
= sa_none
;
7182 struct sched_domain
*sd
;
7188 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7189 if (alloc_state
!= sa_rootdomain
)
7191 alloc_state
= sa_sched_groups
;
7194 * Set up domains for cpus specified by the cpu_map.
7196 for_each_cpu(i
, cpu_map
) {
7197 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7200 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7201 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7202 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7203 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7204 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7207 for_each_cpu(i
, cpu_map
) {
7208 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7209 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7210 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7213 /* Set up physical groups */
7214 for (i
= 0; i
< nr_node_ids
; i
++)
7215 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7218 /* Set up node groups */
7220 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7222 for (i
= 0; i
< nr_node_ids
; i
++)
7223 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7227 /* Calculate CPU power for physical packages and nodes */
7228 #ifdef CONFIG_SCHED_SMT
7229 for_each_cpu(i
, cpu_map
) {
7230 sd
= &per_cpu(cpu_domains
, i
).sd
;
7231 init_sched_groups_power(i
, sd
);
7234 #ifdef CONFIG_SCHED_MC
7235 for_each_cpu(i
, cpu_map
) {
7236 sd
= &per_cpu(core_domains
, i
).sd
;
7237 init_sched_groups_power(i
, sd
);
7240 #ifdef CONFIG_SCHED_BOOK
7241 for_each_cpu(i
, cpu_map
) {
7242 sd
= &per_cpu(book_domains
, i
).sd
;
7243 init_sched_groups_power(i
, sd
);
7247 for_each_cpu(i
, cpu_map
) {
7248 sd
= &per_cpu(phys_domains
, i
).sd
;
7249 init_sched_groups_power(i
, sd
);
7253 for (i
= 0; i
< nr_node_ids
; i
++)
7254 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7256 if (d
.sd_allnodes
) {
7257 struct sched_group
*sg
;
7259 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7261 init_numa_sched_groups_power(sg
);
7265 /* Attach the domains */
7266 for_each_cpu(i
, cpu_map
) {
7267 #ifdef CONFIG_SCHED_SMT
7268 sd
= &per_cpu(cpu_domains
, i
).sd
;
7269 #elif defined(CONFIG_SCHED_MC)
7270 sd
= &per_cpu(core_domains
, i
).sd
;
7271 #elif defined(CONFIG_SCHED_BOOK)
7272 sd
= &per_cpu(book_domains
, i
).sd
;
7274 sd
= &per_cpu(phys_domains
, i
).sd
;
7276 cpu_attach_domain(sd
, d
.rd
, i
);
7279 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7280 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7284 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7288 static int build_sched_domains(const struct cpumask
*cpu_map
)
7290 return __build_sched_domains(cpu_map
, NULL
);
7293 static cpumask_var_t
*doms_cur
; /* current sched domains */
7294 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7295 static struct sched_domain_attr
*dattr_cur
;
7296 /* attribues of custom domains in 'doms_cur' */
7299 * Special case: If a kmalloc of a doms_cur partition (array of
7300 * cpumask) fails, then fallback to a single sched domain,
7301 * as determined by the single cpumask fallback_doms.
7303 static cpumask_var_t fallback_doms
;
7306 * arch_update_cpu_topology lets virtualized architectures update the
7307 * cpu core maps. It is supposed to return 1 if the topology changed
7308 * or 0 if it stayed the same.
7310 int __attribute__((weak
)) arch_update_cpu_topology(void)
7315 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7318 cpumask_var_t
*doms
;
7320 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7323 for (i
= 0; i
< ndoms
; i
++) {
7324 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7325 free_sched_domains(doms
, i
);
7332 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7335 for (i
= 0; i
< ndoms
; i
++)
7336 free_cpumask_var(doms
[i
]);
7341 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7342 * For now this just excludes isolated cpus, but could be used to
7343 * exclude other special cases in the future.
7345 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7349 arch_update_cpu_topology();
7351 doms_cur
= alloc_sched_domains(ndoms_cur
);
7353 doms_cur
= &fallback_doms
;
7354 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7356 err
= build_sched_domains(doms_cur
[0]);
7357 register_sched_domain_sysctl();
7362 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7363 struct cpumask
*tmpmask
)
7365 free_sched_groups(cpu_map
, tmpmask
);
7369 * Detach sched domains from a group of cpus specified in cpu_map
7370 * These cpus will now be attached to the NULL domain
7372 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7374 /* Save because hotplug lock held. */
7375 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7378 for_each_cpu(i
, cpu_map
)
7379 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7380 synchronize_sched();
7381 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7384 /* handle null as "default" */
7385 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7386 struct sched_domain_attr
*new, int idx_new
)
7388 struct sched_domain_attr tmp
;
7395 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7396 new ? (new + idx_new
) : &tmp
,
7397 sizeof(struct sched_domain_attr
));
7401 * Partition sched domains as specified by the 'ndoms_new'
7402 * cpumasks in the array doms_new[] of cpumasks. This compares
7403 * doms_new[] to the current sched domain partitioning, doms_cur[].
7404 * It destroys each deleted domain and builds each new domain.
7406 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7407 * The masks don't intersect (don't overlap.) We should setup one
7408 * sched domain for each mask. CPUs not in any of the cpumasks will
7409 * not be load balanced. If the same cpumask appears both in the
7410 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7413 * The passed in 'doms_new' should be allocated using
7414 * alloc_sched_domains. This routine takes ownership of it and will
7415 * free_sched_domains it when done with it. If the caller failed the
7416 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7417 * and partition_sched_domains() will fallback to the single partition
7418 * 'fallback_doms', it also forces the domains to be rebuilt.
7420 * If doms_new == NULL it will be replaced with cpu_online_mask.
7421 * ndoms_new == 0 is a special case for destroying existing domains,
7422 * and it will not create the default domain.
7424 * Call with hotplug lock held
7426 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7427 struct sched_domain_attr
*dattr_new
)
7432 mutex_lock(&sched_domains_mutex
);
7434 /* always unregister in case we don't destroy any domains */
7435 unregister_sched_domain_sysctl();
7437 /* Let architecture update cpu core mappings. */
7438 new_topology
= arch_update_cpu_topology();
7440 n
= doms_new
? ndoms_new
: 0;
7442 /* Destroy deleted domains */
7443 for (i
= 0; i
< ndoms_cur
; i
++) {
7444 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7445 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7446 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7449 /* no match - a current sched domain not in new doms_new[] */
7450 detach_destroy_domains(doms_cur
[i
]);
7455 if (doms_new
== NULL
) {
7457 doms_new
= &fallback_doms
;
7458 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7459 WARN_ON_ONCE(dattr_new
);
7462 /* Build new domains */
7463 for (i
= 0; i
< ndoms_new
; i
++) {
7464 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7465 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7466 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7469 /* no match - add a new doms_new */
7470 __build_sched_domains(doms_new
[i
],
7471 dattr_new
? dattr_new
+ i
: NULL
);
7476 /* Remember the new sched domains */
7477 if (doms_cur
!= &fallback_doms
)
7478 free_sched_domains(doms_cur
, ndoms_cur
);
7479 kfree(dattr_cur
); /* kfree(NULL) is safe */
7480 doms_cur
= doms_new
;
7481 dattr_cur
= dattr_new
;
7482 ndoms_cur
= ndoms_new
;
7484 register_sched_domain_sysctl();
7486 mutex_unlock(&sched_domains_mutex
);
7489 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7490 static void arch_reinit_sched_domains(void)
7494 /* Destroy domains first to force the rebuild */
7495 partition_sched_domains(0, NULL
, NULL
);
7497 rebuild_sched_domains();
7501 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7503 unsigned int level
= 0;
7505 if (sscanf(buf
, "%u", &level
) != 1)
7509 * level is always be positive so don't check for
7510 * level < POWERSAVINGS_BALANCE_NONE which is 0
7511 * What happens on 0 or 1 byte write,
7512 * need to check for count as well?
7515 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7519 sched_smt_power_savings
= level
;
7521 sched_mc_power_savings
= level
;
7523 arch_reinit_sched_domains();
7528 #ifdef CONFIG_SCHED_MC
7529 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7530 struct sysdev_class_attribute
*attr
,
7533 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7535 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7536 struct sysdev_class_attribute
*attr
,
7537 const char *buf
, size_t count
)
7539 return sched_power_savings_store(buf
, count
, 0);
7541 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7542 sched_mc_power_savings_show
,
7543 sched_mc_power_savings_store
);
7546 #ifdef CONFIG_SCHED_SMT
7547 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7548 struct sysdev_class_attribute
*attr
,
7551 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7553 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7554 struct sysdev_class_attribute
*attr
,
7555 const char *buf
, size_t count
)
7557 return sched_power_savings_store(buf
, count
, 1);
7559 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7560 sched_smt_power_savings_show
,
7561 sched_smt_power_savings_store
);
7564 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7568 #ifdef CONFIG_SCHED_SMT
7570 err
= sysfs_create_file(&cls
->kset
.kobj
,
7571 &attr_sched_smt_power_savings
.attr
);
7573 #ifdef CONFIG_SCHED_MC
7574 if (!err
&& mc_capable())
7575 err
= sysfs_create_file(&cls
->kset
.kobj
,
7576 &attr_sched_mc_power_savings
.attr
);
7580 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7583 * Update cpusets according to cpu_active mask. If cpusets are
7584 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7585 * around partition_sched_domains().
7587 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7590 switch (action
& ~CPU_TASKS_FROZEN
) {
7592 case CPU_DOWN_FAILED
:
7593 cpuset_update_active_cpus();
7600 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7603 switch (action
& ~CPU_TASKS_FROZEN
) {
7604 case CPU_DOWN_PREPARE
:
7605 cpuset_update_active_cpus();
7612 static int update_runtime(struct notifier_block
*nfb
,
7613 unsigned long action
, void *hcpu
)
7615 int cpu
= (int)(long)hcpu
;
7618 case CPU_DOWN_PREPARE
:
7619 case CPU_DOWN_PREPARE_FROZEN
:
7620 disable_runtime(cpu_rq(cpu
));
7623 case CPU_DOWN_FAILED
:
7624 case CPU_DOWN_FAILED_FROZEN
:
7626 case CPU_ONLINE_FROZEN
:
7627 enable_runtime(cpu_rq(cpu
));
7635 void __init
sched_init_smp(void)
7637 cpumask_var_t non_isolated_cpus
;
7639 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7640 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7642 #if defined(CONFIG_NUMA)
7643 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7645 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7648 mutex_lock(&sched_domains_mutex
);
7649 arch_init_sched_domains(cpu_active_mask
);
7650 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7651 if (cpumask_empty(non_isolated_cpus
))
7652 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7653 mutex_unlock(&sched_domains_mutex
);
7656 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7657 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7659 /* RT runtime code needs to handle some hotplug events */
7660 hotcpu_notifier(update_runtime
, 0);
7664 /* Move init over to a non-isolated CPU */
7665 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7667 sched_init_granularity();
7668 free_cpumask_var(non_isolated_cpus
);
7670 init_sched_rt_class();
7673 void __init
sched_init_smp(void)
7675 sched_init_granularity();
7677 #endif /* CONFIG_SMP */
7679 const_debug
unsigned int sysctl_timer_migration
= 1;
7681 int in_sched_functions(unsigned long addr
)
7683 return in_lock_functions(addr
) ||
7684 (addr
>= (unsigned long)__sched_text_start
7685 && addr
< (unsigned long)__sched_text_end
);
7688 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7690 cfs_rq
->tasks_timeline
= RB_ROOT
;
7691 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7692 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7698 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7700 struct rt_prio_array
*array
;
7703 array
= &rt_rq
->active
;
7704 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7705 INIT_LIST_HEAD(array
->queue
+ i
);
7706 __clear_bit(i
, array
->bitmap
);
7708 /* delimiter for bitsearch: */
7709 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7711 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7712 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7714 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7718 rt_rq
->rt_nr_migratory
= 0;
7719 rt_rq
->overloaded
= 0;
7720 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7724 rt_rq
->rt_throttled
= 0;
7725 rt_rq
->rt_runtime
= 0;
7726 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7728 #ifdef CONFIG_RT_GROUP_SCHED
7729 rt_rq
->rt_nr_boosted
= 0;
7734 #ifdef CONFIG_FAIR_GROUP_SCHED
7735 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7736 struct sched_entity
*se
, int cpu
, int add
,
7737 struct sched_entity
*parent
)
7739 struct rq
*rq
= cpu_rq(cpu
);
7740 tg
->cfs_rq
[cpu
] = cfs_rq
;
7741 init_cfs_rq(cfs_rq
, rq
);
7744 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7747 /* se could be NULL for init_task_group */
7752 se
->cfs_rq
= &rq
->cfs
;
7754 se
->cfs_rq
= parent
->my_q
;
7757 se
->load
.weight
= tg
->shares
;
7758 se
->load
.inv_weight
= 0;
7759 se
->parent
= parent
;
7763 #ifdef CONFIG_RT_GROUP_SCHED
7764 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7765 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7766 struct sched_rt_entity
*parent
)
7768 struct rq
*rq
= cpu_rq(cpu
);
7770 tg
->rt_rq
[cpu
] = rt_rq
;
7771 init_rt_rq(rt_rq
, rq
);
7773 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7775 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7777 tg
->rt_se
[cpu
] = rt_se
;
7782 rt_se
->rt_rq
= &rq
->rt
;
7784 rt_se
->rt_rq
= parent
->my_q
;
7786 rt_se
->my_q
= rt_rq
;
7787 rt_se
->parent
= parent
;
7788 INIT_LIST_HEAD(&rt_se
->run_list
);
7792 void __init
sched_init(void)
7795 unsigned long alloc_size
= 0, ptr
;
7797 #ifdef CONFIG_FAIR_GROUP_SCHED
7798 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7800 #ifdef CONFIG_RT_GROUP_SCHED
7801 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7803 #ifdef CONFIG_CPUMASK_OFFSTACK
7804 alloc_size
+= num_possible_cpus() * cpumask_size();
7807 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7809 #ifdef CONFIG_FAIR_GROUP_SCHED
7810 init_task_group
.se
= (struct sched_entity
**)ptr
;
7811 ptr
+= nr_cpu_ids
* sizeof(void **);
7813 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7814 ptr
+= nr_cpu_ids
* sizeof(void **);
7816 #endif /* CONFIG_FAIR_GROUP_SCHED */
7817 #ifdef CONFIG_RT_GROUP_SCHED
7818 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7819 ptr
+= nr_cpu_ids
* sizeof(void **);
7821 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7822 ptr
+= nr_cpu_ids
* sizeof(void **);
7824 #endif /* CONFIG_RT_GROUP_SCHED */
7825 #ifdef CONFIG_CPUMASK_OFFSTACK
7826 for_each_possible_cpu(i
) {
7827 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7828 ptr
+= cpumask_size();
7830 #endif /* CONFIG_CPUMASK_OFFSTACK */
7834 init_defrootdomain();
7837 init_rt_bandwidth(&def_rt_bandwidth
,
7838 global_rt_period(), global_rt_runtime());
7840 #ifdef CONFIG_RT_GROUP_SCHED
7841 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7842 global_rt_period(), global_rt_runtime());
7843 #endif /* CONFIG_RT_GROUP_SCHED */
7845 #ifdef CONFIG_CGROUP_SCHED
7846 list_add(&init_task_group
.list
, &task_groups
);
7847 INIT_LIST_HEAD(&init_task_group
.children
);
7849 #endif /* CONFIG_CGROUP_SCHED */
7851 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7852 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7853 __alignof__(unsigned long));
7855 for_each_possible_cpu(i
) {
7859 raw_spin_lock_init(&rq
->lock
);
7861 rq
->calc_load_active
= 0;
7862 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7863 init_cfs_rq(&rq
->cfs
, rq
);
7864 init_rt_rq(&rq
->rt
, rq
);
7865 #ifdef CONFIG_FAIR_GROUP_SCHED
7866 init_task_group
.shares
= init_task_group_load
;
7867 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7868 #ifdef CONFIG_CGROUP_SCHED
7870 * How much cpu bandwidth does init_task_group get?
7872 * In case of task-groups formed thr' the cgroup filesystem, it
7873 * gets 100% of the cpu resources in the system. This overall
7874 * system cpu resource is divided among the tasks of
7875 * init_task_group and its child task-groups in a fair manner,
7876 * based on each entity's (task or task-group's) weight
7877 * (se->load.weight).
7879 * In other words, if init_task_group has 10 tasks of weight
7880 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7881 * then A0's share of the cpu resource is:
7883 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7885 * We achieve this by letting init_task_group's tasks sit
7886 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7888 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7890 #endif /* CONFIG_FAIR_GROUP_SCHED */
7892 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7893 #ifdef CONFIG_RT_GROUP_SCHED
7894 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7895 #ifdef CONFIG_CGROUP_SCHED
7896 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7900 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7901 rq
->cpu_load
[j
] = 0;
7903 rq
->last_load_update_tick
= jiffies
;
7908 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7909 rq
->post_schedule
= 0;
7910 rq
->active_balance
= 0;
7911 rq
->next_balance
= jiffies
;
7916 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7917 rq_attach_root(rq
, &def_root_domain
);
7919 rq
->nohz_balance_kick
= 0;
7920 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
7924 atomic_set(&rq
->nr_iowait
, 0);
7927 set_load_weight(&init_task
);
7929 #ifdef CONFIG_PREEMPT_NOTIFIERS
7930 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7934 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7937 #ifdef CONFIG_RT_MUTEXES
7938 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7942 * The boot idle thread does lazy MMU switching as well:
7944 atomic_inc(&init_mm
.mm_count
);
7945 enter_lazy_tlb(&init_mm
, current
);
7948 * Make us the idle thread. Technically, schedule() should not be
7949 * called from this thread, however somewhere below it might be,
7950 * but because we are the idle thread, we just pick up running again
7951 * when this runqueue becomes "idle".
7953 init_idle(current
, smp_processor_id());
7955 calc_load_update
= jiffies
+ LOAD_FREQ
;
7958 * During early bootup we pretend to be a normal task:
7960 current
->sched_class
= &fair_sched_class
;
7962 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7963 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7966 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7967 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
7968 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
7969 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
7970 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
7972 /* May be allocated at isolcpus cmdline parse time */
7973 if (cpu_isolated_map
== NULL
)
7974 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7979 scheduler_running
= 1;
7982 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7983 static inline int preempt_count_equals(int preempt_offset
)
7985 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7987 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7990 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7993 static unsigned long prev_jiffy
; /* ratelimiting */
7995 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7996 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7998 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8000 prev_jiffy
= jiffies
;
8003 "BUG: sleeping function called from invalid context at %s:%d\n",
8006 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8007 in_atomic(), irqs_disabled(),
8008 current
->pid
, current
->comm
);
8010 debug_show_held_locks(current
);
8011 if (irqs_disabled())
8012 print_irqtrace_events(current
);
8016 EXPORT_SYMBOL(__might_sleep
);
8019 #ifdef CONFIG_MAGIC_SYSRQ
8020 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8024 on_rq
= p
->se
.on_rq
;
8026 deactivate_task(rq
, p
, 0);
8027 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8029 activate_task(rq
, p
, 0);
8030 resched_task(rq
->curr
);
8034 void normalize_rt_tasks(void)
8036 struct task_struct
*g
, *p
;
8037 unsigned long flags
;
8040 read_lock_irqsave(&tasklist_lock
, flags
);
8041 do_each_thread(g
, p
) {
8043 * Only normalize user tasks:
8048 p
->se
.exec_start
= 0;
8049 #ifdef CONFIG_SCHEDSTATS
8050 p
->se
.statistics
.wait_start
= 0;
8051 p
->se
.statistics
.sleep_start
= 0;
8052 p
->se
.statistics
.block_start
= 0;
8057 * Renice negative nice level userspace
8060 if (TASK_NICE(p
) < 0 && p
->mm
)
8061 set_user_nice(p
, 0);
8065 raw_spin_lock(&p
->pi_lock
);
8066 rq
= __task_rq_lock(p
);
8068 normalize_task(rq
, p
);
8070 __task_rq_unlock(rq
);
8071 raw_spin_unlock(&p
->pi_lock
);
8072 } while_each_thread(g
, p
);
8074 read_unlock_irqrestore(&tasklist_lock
, flags
);
8077 #endif /* CONFIG_MAGIC_SYSRQ */
8079 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8081 * These functions are only useful for the IA64 MCA handling, or kdb.
8083 * They can only be called when the whole system has been
8084 * stopped - every CPU needs to be quiescent, and no scheduling
8085 * activity can take place. Using them for anything else would
8086 * be a serious bug, and as a result, they aren't even visible
8087 * under any other configuration.
8091 * curr_task - return the current task for a given cpu.
8092 * @cpu: the processor in question.
8094 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8096 struct task_struct
*curr_task(int cpu
)
8098 return cpu_curr(cpu
);
8101 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8105 * set_curr_task - set the current task for a given cpu.
8106 * @cpu: the processor in question.
8107 * @p: the task pointer to set.
8109 * Description: This function must only be used when non-maskable interrupts
8110 * are serviced on a separate stack. It allows the architecture to switch the
8111 * notion of the current task on a cpu in a non-blocking manner. This function
8112 * must be called with all CPU's synchronized, and interrupts disabled, the
8113 * and caller must save the original value of the current task (see
8114 * curr_task() above) and restore that value before reenabling interrupts and
8115 * re-starting the system.
8117 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8119 void set_curr_task(int cpu
, struct task_struct
*p
)
8126 #ifdef CONFIG_FAIR_GROUP_SCHED
8127 static void free_fair_sched_group(struct task_group
*tg
)
8131 for_each_possible_cpu(i
) {
8133 kfree(tg
->cfs_rq
[i
]);
8143 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8145 struct cfs_rq
*cfs_rq
;
8146 struct sched_entity
*se
;
8150 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8153 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8157 tg
->shares
= NICE_0_LOAD
;
8159 for_each_possible_cpu(i
) {
8162 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8163 GFP_KERNEL
, cpu_to_node(i
));
8167 se
= kzalloc_node(sizeof(struct sched_entity
),
8168 GFP_KERNEL
, cpu_to_node(i
));
8172 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8183 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8185 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8186 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8189 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8191 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8193 #else /* !CONFG_FAIR_GROUP_SCHED */
8194 static inline void free_fair_sched_group(struct task_group
*tg
)
8199 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8204 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8208 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8211 #endif /* CONFIG_FAIR_GROUP_SCHED */
8213 #ifdef CONFIG_RT_GROUP_SCHED
8214 static void free_rt_sched_group(struct task_group
*tg
)
8218 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8220 for_each_possible_cpu(i
) {
8222 kfree(tg
->rt_rq
[i
]);
8224 kfree(tg
->rt_se
[i
]);
8232 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8234 struct rt_rq
*rt_rq
;
8235 struct sched_rt_entity
*rt_se
;
8239 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8242 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8246 init_rt_bandwidth(&tg
->rt_bandwidth
,
8247 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8249 for_each_possible_cpu(i
) {
8252 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8253 GFP_KERNEL
, cpu_to_node(i
));
8257 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8258 GFP_KERNEL
, cpu_to_node(i
));
8262 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8273 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8275 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8276 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8279 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8281 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8283 #else /* !CONFIG_RT_GROUP_SCHED */
8284 static inline void free_rt_sched_group(struct task_group
*tg
)
8289 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8294 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8298 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8301 #endif /* CONFIG_RT_GROUP_SCHED */
8303 #ifdef CONFIG_CGROUP_SCHED
8304 static void free_sched_group(struct task_group
*tg
)
8306 free_fair_sched_group(tg
);
8307 free_rt_sched_group(tg
);
8311 /* allocate runqueue etc for a new task group */
8312 struct task_group
*sched_create_group(struct task_group
*parent
)
8314 struct task_group
*tg
;
8315 unsigned long flags
;
8318 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8320 return ERR_PTR(-ENOMEM
);
8322 if (!alloc_fair_sched_group(tg
, parent
))
8325 if (!alloc_rt_sched_group(tg
, parent
))
8328 spin_lock_irqsave(&task_group_lock
, flags
);
8329 for_each_possible_cpu(i
) {
8330 register_fair_sched_group(tg
, i
);
8331 register_rt_sched_group(tg
, i
);
8333 list_add_rcu(&tg
->list
, &task_groups
);
8335 WARN_ON(!parent
); /* root should already exist */
8337 tg
->parent
= parent
;
8338 INIT_LIST_HEAD(&tg
->children
);
8339 list_add_rcu(&tg
->siblings
, &parent
->children
);
8340 spin_unlock_irqrestore(&task_group_lock
, flags
);
8345 free_sched_group(tg
);
8346 return ERR_PTR(-ENOMEM
);
8349 /* rcu callback to free various structures associated with a task group */
8350 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8352 /* now it should be safe to free those cfs_rqs */
8353 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8356 /* Destroy runqueue etc associated with a task group */
8357 void sched_destroy_group(struct task_group
*tg
)
8359 unsigned long flags
;
8362 spin_lock_irqsave(&task_group_lock
, flags
);
8363 for_each_possible_cpu(i
) {
8364 unregister_fair_sched_group(tg
, i
);
8365 unregister_rt_sched_group(tg
, i
);
8367 list_del_rcu(&tg
->list
);
8368 list_del_rcu(&tg
->siblings
);
8369 spin_unlock_irqrestore(&task_group_lock
, flags
);
8371 /* wait for possible concurrent references to cfs_rqs complete */
8372 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8375 /* change task's runqueue when it moves between groups.
8376 * The caller of this function should have put the task in its new group
8377 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8378 * reflect its new group.
8380 void sched_move_task(struct task_struct
*tsk
)
8383 unsigned long flags
;
8386 rq
= task_rq_lock(tsk
, &flags
);
8388 running
= task_current(rq
, tsk
);
8389 on_rq
= tsk
->se
.on_rq
;
8392 dequeue_task(rq
, tsk
, 0);
8393 if (unlikely(running
))
8394 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8396 set_task_rq(tsk
, task_cpu(tsk
));
8398 #ifdef CONFIG_FAIR_GROUP_SCHED
8399 if (tsk
->sched_class
->moved_group
)
8400 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8403 if (unlikely(running
))
8404 tsk
->sched_class
->set_curr_task(rq
);
8406 enqueue_task(rq
, tsk
, 0);
8408 task_rq_unlock(rq
, &flags
);
8410 #endif /* CONFIG_CGROUP_SCHED */
8412 #ifdef CONFIG_FAIR_GROUP_SCHED
8413 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8415 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8420 dequeue_entity(cfs_rq
, se
, 0);
8422 se
->load
.weight
= shares
;
8423 se
->load
.inv_weight
= 0;
8426 enqueue_entity(cfs_rq
, se
, 0);
8429 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8431 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8432 struct rq
*rq
= cfs_rq
->rq
;
8433 unsigned long flags
;
8435 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8436 __set_se_shares(se
, shares
);
8437 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8440 static DEFINE_MUTEX(shares_mutex
);
8442 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8445 unsigned long flags
;
8448 * We can't change the weight of the root cgroup.
8453 if (shares
< MIN_SHARES
)
8454 shares
= MIN_SHARES
;
8455 else if (shares
> MAX_SHARES
)
8456 shares
= MAX_SHARES
;
8458 mutex_lock(&shares_mutex
);
8459 if (tg
->shares
== shares
)
8462 spin_lock_irqsave(&task_group_lock
, flags
);
8463 for_each_possible_cpu(i
)
8464 unregister_fair_sched_group(tg
, i
);
8465 list_del_rcu(&tg
->siblings
);
8466 spin_unlock_irqrestore(&task_group_lock
, flags
);
8468 /* wait for any ongoing reference to this group to finish */
8469 synchronize_sched();
8472 * Now we are free to modify the group's share on each cpu
8473 * w/o tripping rebalance_share or load_balance_fair.
8475 tg
->shares
= shares
;
8476 for_each_possible_cpu(i
) {
8480 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8481 set_se_shares(tg
->se
[i
], shares
);
8485 * Enable load balance activity on this group, by inserting it back on
8486 * each cpu's rq->leaf_cfs_rq_list.
8488 spin_lock_irqsave(&task_group_lock
, flags
);
8489 for_each_possible_cpu(i
)
8490 register_fair_sched_group(tg
, i
);
8491 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8492 spin_unlock_irqrestore(&task_group_lock
, flags
);
8494 mutex_unlock(&shares_mutex
);
8498 unsigned long sched_group_shares(struct task_group
*tg
)
8504 #ifdef CONFIG_RT_GROUP_SCHED
8506 * Ensure that the real time constraints are schedulable.
8508 static DEFINE_MUTEX(rt_constraints_mutex
);
8510 static unsigned long to_ratio(u64 period
, u64 runtime
)
8512 if (runtime
== RUNTIME_INF
)
8515 return div64_u64(runtime
<< 20, period
);
8518 /* Must be called with tasklist_lock held */
8519 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8521 struct task_struct
*g
, *p
;
8523 do_each_thread(g
, p
) {
8524 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8526 } while_each_thread(g
, p
);
8531 struct rt_schedulable_data
{
8532 struct task_group
*tg
;
8537 static int tg_schedulable(struct task_group
*tg
, void *data
)
8539 struct rt_schedulable_data
*d
= data
;
8540 struct task_group
*child
;
8541 unsigned long total
, sum
= 0;
8542 u64 period
, runtime
;
8544 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8545 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8548 period
= d
->rt_period
;
8549 runtime
= d
->rt_runtime
;
8553 * Cannot have more runtime than the period.
8555 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8559 * Ensure we don't starve existing RT tasks.
8561 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8564 total
= to_ratio(period
, runtime
);
8567 * Nobody can have more than the global setting allows.
8569 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8573 * The sum of our children's runtime should not exceed our own.
8575 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8576 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8577 runtime
= child
->rt_bandwidth
.rt_runtime
;
8579 if (child
== d
->tg
) {
8580 period
= d
->rt_period
;
8581 runtime
= d
->rt_runtime
;
8584 sum
+= to_ratio(period
, runtime
);
8593 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8595 struct rt_schedulable_data data
= {
8597 .rt_period
= period
,
8598 .rt_runtime
= runtime
,
8601 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8604 static int tg_set_bandwidth(struct task_group
*tg
,
8605 u64 rt_period
, u64 rt_runtime
)
8609 mutex_lock(&rt_constraints_mutex
);
8610 read_lock(&tasklist_lock
);
8611 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8615 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8616 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8617 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8619 for_each_possible_cpu(i
) {
8620 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8622 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8623 rt_rq
->rt_runtime
= rt_runtime
;
8624 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8626 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8628 read_unlock(&tasklist_lock
);
8629 mutex_unlock(&rt_constraints_mutex
);
8634 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8636 u64 rt_runtime
, rt_period
;
8638 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8639 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8640 if (rt_runtime_us
< 0)
8641 rt_runtime
= RUNTIME_INF
;
8643 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8646 long sched_group_rt_runtime(struct task_group
*tg
)
8650 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8653 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8654 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8655 return rt_runtime_us
;
8658 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8660 u64 rt_runtime
, rt_period
;
8662 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8663 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8668 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8671 long sched_group_rt_period(struct task_group
*tg
)
8675 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8676 do_div(rt_period_us
, NSEC_PER_USEC
);
8677 return rt_period_us
;
8680 static int sched_rt_global_constraints(void)
8682 u64 runtime
, period
;
8685 if (sysctl_sched_rt_period
<= 0)
8688 runtime
= global_rt_runtime();
8689 period
= global_rt_period();
8692 * Sanity check on the sysctl variables.
8694 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8697 mutex_lock(&rt_constraints_mutex
);
8698 read_lock(&tasklist_lock
);
8699 ret
= __rt_schedulable(NULL
, 0, 0);
8700 read_unlock(&tasklist_lock
);
8701 mutex_unlock(&rt_constraints_mutex
);
8706 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8708 /* Don't accept realtime tasks when there is no way for them to run */
8709 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8715 #else /* !CONFIG_RT_GROUP_SCHED */
8716 static int sched_rt_global_constraints(void)
8718 unsigned long flags
;
8721 if (sysctl_sched_rt_period
<= 0)
8725 * There's always some RT tasks in the root group
8726 * -- migration, kstopmachine etc..
8728 if (sysctl_sched_rt_runtime
== 0)
8731 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8732 for_each_possible_cpu(i
) {
8733 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8735 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8736 rt_rq
->rt_runtime
= global_rt_runtime();
8737 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8739 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8743 #endif /* CONFIG_RT_GROUP_SCHED */
8745 int sched_rt_handler(struct ctl_table
*table
, int write
,
8746 void __user
*buffer
, size_t *lenp
,
8750 int old_period
, old_runtime
;
8751 static DEFINE_MUTEX(mutex
);
8754 old_period
= sysctl_sched_rt_period
;
8755 old_runtime
= sysctl_sched_rt_runtime
;
8757 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8759 if (!ret
&& write
) {
8760 ret
= sched_rt_global_constraints();
8762 sysctl_sched_rt_period
= old_period
;
8763 sysctl_sched_rt_runtime
= old_runtime
;
8765 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8766 def_rt_bandwidth
.rt_period
=
8767 ns_to_ktime(global_rt_period());
8770 mutex_unlock(&mutex
);
8775 #ifdef CONFIG_CGROUP_SCHED
8777 /* return corresponding task_group object of a cgroup */
8778 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8780 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8781 struct task_group
, css
);
8784 static struct cgroup_subsys_state
*
8785 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8787 struct task_group
*tg
, *parent
;
8789 if (!cgrp
->parent
) {
8790 /* This is early initialization for the top cgroup */
8791 return &init_task_group
.css
;
8794 parent
= cgroup_tg(cgrp
->parent
);
8795 tg
= sched_create_group(parent
);
8797 return ERR_PTR(-ENOMEM
);
8803 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8805 struct task_group
*tg
= cgroup_tg(cgrp
);
8807 sched_destroy_group(tg
);
8811 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8813 #ifdef CONFIG_RT_GROUP_SCHED
8814 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8817 /* We don't support RT-tasks being in separate groups */
8818 if (tsk
->sched_class
!= &fair_sched_class
)
8825 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8826 struct task_struct
*tsk
, bool threadgroup
)
8828 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8832 struct task_struct
*c
;
8834 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8835 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8847 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8848 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8851 sched_move_task(tsk
);
8853 struct task_struct
*c
;
8855 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8862 #ifdef CONFIG_FAIR_GROUP_SCHED
8863 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8866 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8869 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8871 struct task_group
*tg
= cgroup_tg(cgrp
);
8873 return (u64
) tg
->shares
;
8875 #endif /* CONFIG_FAIR_GROUP_SCHED */
8877 #ifdef CONFIG_RT_GROUP_SCHED
8878 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8881 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8884 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8886 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8889 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8892 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8895 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8897 return sched_group_rt_period(cgroup_tg(cgrp
));
8899 #endif /* CONFIG_RT_GROUP_SCHED */
8901 static struct cftype cpu_files
[] = {
8902 #ifdef CONFIG_FAIR_GROUP_SCHED
8905 .read_u64
= cpu_shares_read_u64
,
8906 .write_u64
= cpu_shares_write_u64
,
8909 #ifdef CONFIG_RT_GROUP_SCHED
8911 .name
= "rt_runtime_us",
8912 .read_s64
= cpu_rt_runtime_read
,
8913 .write_s64
= cpu_rt_runtime_write
,
8916 .name
= "rt_period_us",
8917 .read_u64
= cpu_rt_period_read_uint
,
8918 .write_u64
= cpu_rt_period_write_uint
,
8923 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8925 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8928 struct cgroup_subsys cpu_cgroup_subsys
= {
8930 .create
= cpu_cgroup_create
,
8931 .destroy
= cpu_cgroup_destroy
,
8932 .can_attach
= cpu_cgroup_can_attach
,
8933 .attach
= cpu_cgroup_attach
,
8934 .populate
= cpu_cgroup_populate
,
8935 .subsys_id
= cpu_cgroup_subsys_id
,
8939 #endif /* CONFIG_CGROUP_SCHED */
8941 #ifdef CONFIG_CGROUP_CPUACCT
8944 * CPU accounting code for task groups.
8946 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8947 * (balbir@in.ibm.com).
8950 /* track cpu usage of a group of tasks and its child groups */
8952 struct cgroup_subsys_state css
;
8953 /* cpuusage holds pointer to a u64-type object on every cpu */
8954 u64 __percpu
*cpuusage
;
8955 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8956 struct cpuacct
*parent
;
8959 struct cgroup_subsys cpuacct_subsys
;
8961 /* return cpu accounting group corresponding to this container */
8962 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8964 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8965 struct cpuacct
, css
);
8968 /* return cpu accounting group to which this task belongs */
8969 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8971 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8972 struct cpuacct
, css
);
8975 /* create a new cpu accounting group */
8976 static struct cgroup_subsys_state
*cpuacct_create(
8977 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8979 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8985 ca
->cpuusage
= alloc_percpu(u64
);
8989 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8990 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8991 goto out_free_counters
;
8994 ca
->parent
= cgroup_ca(cgrp
->parent
);
9000 percpu_counter_destroy(&ca
->cpustat
[i
]);
9001 free_percpu(ca
->cpuusage
);
9005 return ERR_PTR(-ENOMEM
);
9008 /* destroy an existing cpu accounting group */
9010 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9012 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9015 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9016 percpu_counter_destroy(&ca
->cpustat
[i
]);
9017 free_percpu(ca
->cpuusage
);
9021 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9023 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9026 #ifndef CONFIG_64BIT
9028 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9030 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9032 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9040 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9042 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9044 #ifndef CONFIG_64BIT
9046 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9048 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9050 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9056 /* return total cpu usage (in nanoseconds) of a group */
9057 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9059 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9060 u64 totalcpuusage
= 0;
9063 for_each_present_cpu(i
)
9064 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9066 return totalcpuusage
;
9069 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9072 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9081 for_each_present_cpu(i
)
9082 cpuacct_cpuusage_write(ca
, i
, 0);
9088 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9091 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9095 for_each_present_cpu(i
) {
9096 percpu
= cpuacct_cpuusage_read(ca
, i
);
9097 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9099 seq_printf(m
, "\n");
9103 static const char *cpuacct_stat_desc
[] = {
9104 [CPUACCT_STAT_USER
] = "user",
9105 [CPUACCT_STAT_SYSTEM
] = "system",
9108 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9109 struct cgroup_map_cb
*cb
)
9111 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9114 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9115 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9116 val
= cputime64_to_clock_t(val
);
9117 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9122 static struct cftype files
[] = {
9125 .read_u64
= cpuusage_read
,
9126 .write_u64
= cpuusage_write
,
9129 .name
= "usage_percpu",
9130 .read_seq_string
= cpuacct_percpu_seq_read
,
9134 .read_map
= cpuacct_stats_show
,
9138 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9140 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9144 * charge this task's execution time to its accounting group.
9146 * called with rq->lock held.
9148 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9153 if (unlikely(!cpuacct_subsys
.active
))
9156 cpu
= task_cpu(tsk
);
9162 for (; ca
; ca
= ca
->parent
) {
9163 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9164 *cpuusage
+= cputime
;
9171 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9172 * in cputime_t units. As a result, cpuacct_update_stats calls
9173 * percpu_counter_add with values large enough to always overflow the
9174 * per cpu batch limit causing bad SMP scalability.
9176 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9177 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9178 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9181 #define CPUACCT_BATCH \
9182 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9184 #define CPUACCT_BATCH 0
9188 * Charge the system/user time to the task's accounting group.
9190 static void cpuacct_update_stats(struct task_struct
*tsk
,
9191 enum cpuacct_stat_index idx
, cputime_t val
)
9194 int batch
= CPUACCT_BATCH
;
9196 if (unlikely(!cpuacct_subsys
.active
))
9203 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9209 struct cgroup_subsys cpuacct_subsys
= {
9211 .create
= cpuacct_create
,
9212 .destroy
= cpuacct_destroy
,
9213 .populate
= cpuacct_populate
,
9214 .subsys_id
= cpuacct_subsys_id
,
9216 #endif /* CONFIG_CGROUP_CPUACCT */
9220 void synchronize_sched_expedited(void)
9224 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9226 #else /* #ifndef CONFIG_SMP */
9228 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9230 static int synchronize_sched_expedited_cpu_stop(void *data
)
9233 * There must be a full memory barrier on each affected CPU
9234 * between the time that try_stop_cpus() is called and the
9235 * time that it returns.
9237 * In the current initial implementation of cpu_stop, the
9238 * above condition is already met when the control reaches
9239 * this point and the following smp_mb() is not strictly
9240 * necessary. Do smp_mb() anyway for documentation and
9241 * robustness against future implementation changes.
9243 smp_mb(); /* See above comment block. */
9248 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9249 * approach to force grace period to end quickly. This consumes
9250 * significant time on all CPUs, and is thus not recommended for
9251 * any sort of common-case code.
9253 * Note that it is illegal to call this function while holding any
9254 * lock that is acquired by a CPU-hotplug notifier. Failing to
9255 * observe this restriction will result in deadlock.
9257 void synchronize_sched_expedited(void)
9259 int snap
, trycount
= 0;
9261 smp_mb(); /* ensure prior mod happens before capturing snap. */
9262 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9264 while (try_stop_cpus(cpu_online_mask
,
9265 synchronize_sched_expedited_cpu_stop
,
9268 if (trycount
++ < 10)
9269 udelay(trycount
* num_online_cpus());
9271 synchronize_sched();
9274 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9275 smp_mb(); /* ensure test happens before caller kfree */
9280 atomic_inc(&synchronize_sched_expedited_count
);
9281 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9284 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9286 #endif /* #else #ifndef CONFIG_SMP */