4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy
)
127 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
132 static inline int task_has_rt_policy(struct task_struct
*p
)
134 return rt_policy(p
->policy
);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array
{
141 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
142 struct list_head queue
[MAX_RT_PRIO
];
145 struct rt_bandwidth
{
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock
;
150 struct hrtimer rt_period_timer
;
153 static struct rt_bandwidth def_rt_bandwidth
;
155 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
157 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
159 struct rt_bandwidth
*rt_b
=
160 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
166 now
= hrtimer_cb_get_time(timer
);
167 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
172 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
175 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
179 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
181 rt_b
->rt_period
= ns_to_ktime(period
);
182 rt_b
->rt_runtime
= runtime
;
184 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
186 hrtimer_init(&rt_b
->rt_period_timer
,
187 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
188 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime
>= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
200 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
203 if (hrtimer_active(&rt_b
->rt_period_timer
))
206 raw_spin_lock(&rt_b
->rt_runtime_lock
);
211 if (hrtimer_active(&rt_b
->rt_period_timer
))
214 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
215 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
217 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
218 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
219 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
220 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
221 HRTIMER_MODE_ABS_PINNED
, 0);
223 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
229 hrtimer_cancel(&rt_b
->rt_period_timer
);
234 * sched_domains_mutex serializes calls to init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex
);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups
);
247 /* task group related information */
249 struct cgroup_subsys_state css
;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity
**se
;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq
**cfs_rq
;
256 unsigned long shares
;
258 atomic_t load_weight
;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity
**rt_se
;
263 struct rt_rq
**rt_rq
;
265 struct rt_bandwidth rt_bandwidth
;
269 struct list_head list
;
271 struct task_group
*parent
;
272 struct list_head siblings
;
273 struct list_head children
;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup
*autogroup
;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock
);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
295 #define MIN_SHARES (1UL << 1)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group
;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load
;
311 unsigned long nr_running
;
316 u64 min_vruntime_copy
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
322 struct list_head tasks
;
323 struct list_head
*balance_iterator
;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity
*curr
, *next
, *last
, *skip
;
331 #ifdef CONFIG_SCHED_DEBUG
332 unsigned int nr_spread_over
;
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
347 struct list_head leaf_cfs_rq_list
;
348 struct task_group
*tg
; /* group that "owns" this runqueue */
352 * the part of load.weight contributed by tasks
354 unsigned long task_weight
;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
362 unsigned long h_load
;
365 * Maintaining per-cpu shares distribution for group scheduling
367 * load_stamp is the last time we updated the load average
368 * load_last is the last time we updated the load average and saw load
369 * load_unacc_exec_time is currently unaccounted execution time
373 u64 load_stamp
, load_last
, load_unacc_exec_time
;
375 unsigned long load_contribution
;
380 /* Real-Time classes' related field in a runqueue: */
382 struct rt_prio_array active
;
383 unsigned long rt_nr_running
;
384 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
386 int curr
; /* highest queued rt task prio */
388 int next
; /* next highest */
393 unsigned long rt_nr_migratory
;
394 unsigned long rt_nr_total
;
396 struct plist_head pushable_tasks
;
401 /* Nests inside the rq lock: */
402 raw_spinlock_t rt_runtime_lock
;
404 #ifdef CONFIG_RT_GROUP_SCHED
405 unsigned long rt_nr_boosted
;
408 struct list_head leaf_rt_rq_list
;
409 struct task_group
*tg
;
416 * We add the notion of a root-domain which will be used to define per-domain
417 * variables. Each exclusive cpuset essentially defines an island domain by
418 * fully partitioning the member cpus from any other cpuset. Whenever a new
419 * exclusive cpuset is created, we also create and attach a new root-domain
427 cpumask_var_t online
;
430 * The "RT overload" flag: it gets set if a CPU has more than
431 * one runnable RT task.
433 cpumask_var_t rto_mask
;
435 struct cpupri cpupri
;
439 * By default the system creates a single root-domain with all cpus as
440 * members (mimicking the global state we have today).
442 static struct root_domain def_root_domain
;
444 #endif /* CONFIG_SMP */
447 * This is the main, per-CPU runqueue data structure.
449 * Locking rule: those places that want to lock multiple runqueues
450 * (such as the load balancing or the thread migration code), lock
451 * acquire operations must be ordered by ascending &runqueue.
458 * nr_running and cpu_load should be in the same cacheline because
459 * remote CPUs use both these fields when doing load calculation.
461 unsigned long nr_running
;
462 #define CPU_LOAD_IDX_MAX 5
463 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
464 unsigned long last_load_update_tick
;
467 unsigned char nohz_balance_kick
;
469 int skip_clock_update
;
471 /* capture load from *all* tasks on this cpu: */
472 struct load_weight load
;
473 unsigned long nr_load_updates
;
479 #ifdef CONFIG_FAIR_GROUP_SCHED
480 /* list of leaf cfs_rq on this cpu: */
481 struct list_head leaf_cfs_rq_list
;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 struct list_head leaf_rt_rq_list
;
488 * This is part of a global counter where only the total sum
489 * over all CPUs matters. A task can increase this counter on
490 * one CPU and if it got migrated afterwards it may decrease
491 * it on another CPU. Always updated under the runqueue lock:
493 unsigned long nr_uninterruptible
;
495 struct task_struct
*curr
, *idle
, *stop
;
496 unsigned long next_balance
;
497 struct mm_struct
*prev_mm
;
505 struct root_domain
*rd
;
506 struct sched_domain
*sd
;
508 unsigned long cpu_power
;
510 unsigned char idle_at_tick
;
511 /* For active balancing */
515 struct cpu_stop_work active_balance_work
;
516 /* cpu of this runqueue: */
520 unsigned long avg_load_per_task
;
528 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 /* calc_load related fields */
533 unsigned long calc_load_update
;
534 long calc_load_active
;
536 #ifdef CONFIG_SCHED_HRTICK
538 int hrtick_csd_pending
;
539 struct call_single_data hrtick_csd
;
541 struct hrtimer hrtick_timer
;
544 #ifdef CONFIG_SCHEDSTATS
546 struct sched_info rq_sched_info
;
547 unsigned long long rq_cpu_time
;
548 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
550 /* sys_sched_yield() stats */
551 unsigned int yld_count
;
553 /* schedule() stats */
554 unsigned int sched_switch
;
555 unsigned int sched_count
;
556 unsigned int sched_goidle
;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count
;
560 unsigned int ttwu_local
;
564 struct task_struct
*wake_list
;
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
571 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
573 static inline int cpu_of(struct rq
*rq
)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification with
609 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
610 * task it moves into the cgroup. Therefore by holding either of those locks,
611 * we pin the task to the current cgroup.
613 static inline struct task_group
*task_group(struct task_struct
*p
)
615 struct task_group
*tg
;
616 struct cgroup_subsys_state
*css
;
618 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
619 lockdep_is_held(&p
->pi_lock
) ||
620 lockdep_is_held(&task_rq(p
)->lock
));
621 tg
= container_of(css
, struct task_group
, css
);
623 return autogroup_task_group(p
, tg
);
626 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
627 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
629 #ifdef CONFIG_FAIR_GROUP_SCHED
630 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
631 p
->se
.parent
= task_group(p
)->se
[cpu
];
634 #ifdef CONFIG_RT_GROUP_SCHED
635 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
636 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
640 #else /* CONFIG_CGROUP_SCHED */
642 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
643 static inline struct task_group
*task_group(struct task_struct
*p
)
648 #endif /* CONFIG_CGROUP_SCHED */
650 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
652 static void update_rq_clock(struct rq
*rq
)
656 if (rq
->skip_clock_update
> 0)
659 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
661 update_rq_clock_task(rq
, delta
);
665 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
667 #ifdef CONFIG_SCHED_DEBUG
668 # define const_debug __read_mostly
670 # define const_debug static const
674 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
675 * @cpu: the processor in question.
677 * This interface allows printk to be called with the runqueue lock
678 * held and know whether or not it is OK to wake up the klogd.
680 int runqueue_is_locked(int cpu
)
682 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
686 * Debugging: various feature bits
689 #define SCHED_FEAT(name, enabled) \
690 __SCHED_FEAT_##name ,
693 #include "sched_features.h"
698 #define SCHED_FEAT(name, enabled) \
699 (1UL << __SCHED_FEAT_##name) * enabled |
701 const_debug
unsigned int sysctl_sched_features
=
702 #include "sched_features.h"
707 #ifdef CONFIG_SCHED_DEBUG
708 #define SCHED_FEAT(name, enabled) \
711 static __read_mostly
char *sched_feat_names
[] = {
712 #include "sched_features.h"
718 static int sched_feat_show(struct seq_file
*m
, void *v
)
722 for (i
= 0; sched_feat_names
[i
]; i
++) {
723 if (!(sysctl_sched_features
& (1UL << i
)))
725 seq_printf(m
, "%s ", sched_feat_names
[i
]);
733 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
734 size_t cnt
, loff_t
*ppos
)
744 if (copy_from_user(&buf
, ubuf
, cnt
))
750 if (strncmp(cmp
, "NO_", 3) == 0) {
755 for (i
= 0; sched_feat_names
[i
]; i
++) {
756 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
758 sysctl_sched_features
&= ~(1UL << i
);
760 sysctl_sched_features
|= (1UL << i
);
765 if (!sched_feat_names
[i
])
773 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
775 return single_open(filp
, sched_feat_show
, NULL
);
778 static const struct file_operations sched_feat_fops
= {
779 .open
= sched_feat_open
,
780 .write
= sched_feat_write
,
783 .release
= single_release
,
786 static __init
int sched_init_debug(void)
788 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
793 late_initcall(sched_init_debug
);
797 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
800 * Number of tasks to iterate in a single balance run.
801 * Limited because this is done with IRQs disabled.
803 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
806 * period over which we average the RT time consumption, measured
811 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
814 * period over which we measure -rt task cpu usage in us.
817 unsigned int sysctl_sched_rt_period
= 1000000;
819 static __read_mostly
int scheduler_running
;
822 * part of the period that we allow rt tasks to run in us.
825 int sysctl_sched_rt_runtime
= 950000;
827 static inline u64
global_rt_period(void)
829 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
832 static inline u64
global_rt_runtime(void)
834 if (sysctl_sched_rt_runtime
< 0)
837 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
840 #ifndef prepare_arch_switch
841 # define prepare_arch_switch(next) do { } while (0)
843 #ifndef finish_arch_switch
844 # define finish_arch_switch(prev) do { } while (0)
847 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
849 return rq
->curr
== p
;
852 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
857 return task_current(rq
, p
);
861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
862 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
866 * We can optimise this out completely for !SMP, because the
867 * SMP rebalancing from interrupt is the only thing that cares
874 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
878 * After ->on_cpu is cleared, the task can be moved to a different CPU.
879 * We must ensure this doesn't happen until the switch is completely
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq
->lock
.owner
= current
;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
896 raw_spin_unlock_irq(&rq
->lock
);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
904 * We can optimise this out completely for !SMP, because the
905 * SMP rebalancing from interrupt is the only thing that cares
910 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
911 raw_spin_unlock_irq(&rq
->lock
);
913 raw_spin_unlock(&rq
->lock
);
917 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
921 * After ->on_cpu is cleared, the task can be moved to a different CPU.
922 * We must ensure this doesn't happen until the switch is completely
928 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
932 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
935 * __task_rq_lock - lock the rq @p resides on.
937 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
942 lockdep_assert_held(&p
->pi_lock
);
946 raw_spin_lock(&rq
->lock
);
947 if (likely(rq
== task_rq(p
)))
949 raw_spin_unlock(&rq
->lock
);
954 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
956 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
957 __acquires(p
->pi_lock
)
963 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
965 raw_spin_lock(&rq
->lock
);
966 if (likely(rq
== task_rq(p
)))
968 raw_spin_unlock(&rq
->lock
);
969 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
973 static void __task_rq_unlock(struct rq
*rq
)
976 raw_spin_unlock(&rq
->lock
);
980 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
982 __releases(p
->pi_lock
)
984 raw_spin_unlock(&rq
->lock
);
985 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
989 * this_rq_lock - lock this runqueue and disable interrupts.
991 static struct rq
*this_rq_lock(void)
998 raw_spin_lock(&rq
->lock
);
1003 #ifdef CONFIG_SCHED_HRTICK
1005 * Use HR-timers to deliver accurate preemption points.
1007 * Its all a bit involved since we cannot program an hrt while holding the
1008 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1011 * When we get rescheduled we reprogram the hrtick_timer outside of the
1017 * - enabled by features
1018 * - hrtimer is actually high res
1020 static inline int hrtick_enabled(struct rq
*rq
)
1022 if (!sched_feat(HRTICK
))
1024 if (!cpu_active(cpu_of(rq
)))
1026 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1029 static void hrtick_clear(struct rq
*rq
)
1031 if (hrtimer_active(&rq
->hrtick_timer
))
1032 hrtimer_cancel(&rq
->hrtick_timer
);
1036 * High-resolution timer tick.
1037 * Runs from hardirq context with interrupts disabled.
1039 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1041 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1043 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1045 raw_spin_lock(&rq
->lock
);
1046 update_rq_clock(rq
);
1047 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1048 raw_spin_unlock(&rq
->lock
);
1050 return HRTIMER_NORESTART
;
1055 * called from hardirq (IPI) context
1057 static void __hrtick_start(void *arg
)
1059 struct rq
*rq
= arg
;
1061 raw_spin_lock(&rq
->lock
);
1062 hrtimer_restart(&rq
->hrtick_timer
);
1063 rq
->hrtick_csd_pending
= 0;
1064 raw_spin_unlock(&rq
->lock
);
1068 * Called to set the hrtick timer state.
1070 * called with rq->lock held and irqs disabled
1072 static void hrtick_start(struct rq
*rq
, u64 delay
)
1074 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1075 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1077 hrtimer_set_expires(timer
, time
);
1079 if (rq
== this_rq()) {
1080 hrtimer_restart(timer
);
1081 } else if (!rq
->hrtick_csd_pending
) {
1082 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1083 rq
->hrtick_csd_pending
= 1;
1088 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1090 int cpu
= (int)(long)hcpu
;
1093 case CPU_UP_CANCELED
:
1094 case CPU_UP_CANCELED_FROZEN
:
1095 case CPU_DOWN_PREPARE
:
1096 case CPU_DOWN_PREPARE_FROZEN
:
1098 case CPU_DEAD_FROZEN
:
1099 hrtick_clear(cpu_rq(cpu
));
1106 static __init
void init_hrtick(void)
1108 hotcpu_notifier(hotplug_hrtick
, 0);
1112 * Called to set the hrtick timer state.
1114 * called with rq->lock held and irqs disabled
1116 static void hrtick_start(struct rq
*rq
, u64 delay
)
1118 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1119 HRTIMER_MODE_REL_PINNED
, 0);
1122 static inline void init_hrtick(void)
1125 #endif /* CONFIG_SMP */
1127 static void init_rq_hrtick(struct rq
*rq
)
1130 rq
->hrtick_csd_pending
= 0;
1132 rq
->hrtick_csd
.flags
= 0;
1133 rq
->hrtick_csd
.func
= __hrtick_start
;
1134 rq
->hrtick_csd
.info
= rq
;
1137 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1138 rq
->hrtick_timer
.function
= hrtick
;
1140 #else /* CONFIG_SCHED_HRTICK */
1141 static inline void hrtick_clear(struct rq
*rq
)
1145 static inline void init_rq_hrtick(struct rq
*rq
)
1149 static inline void init_hrtick(void)
1152 #endif /* CONFIG_SCHED_HRTICK */
1155 * resched_task - mark a task 'to be rescheduled now'.
1157 * On UP this means the setting of the need_resched flag, on SMP it
1158 * might also involve a cross-CPU call to trigger the scheduler on
1163 #ifndef tsk_is_polling
1164 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1167 static void resched_task(struct task_struct
*p
)
1171 assert_raw_spin_locked(&task_rq(p
)->lock
);
1173 if (test_tsk_need_resched(p
))
1176 set_tsk_need_resched(p
);
1179 if (cpu
== smp_processor_id())
1182 /* NEED_RESCHED must be visible before we test polling */
1184 if (!tsk_is_polling(p
))
1185 smp_send_reschedule(cpu
);
1188 static void resched_cpu(int cpu
)
1190 struct rq
*rq
= cpu_rq(cpu
);
1191 unsigned long flags
;
1193 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1195 resched_task(cpu_curr(cpu
));
1196 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1201 * In the semi idle case, use the nearest busy cpu for migrating timers
1202 * from an idle cpu. This is good for power-savings.
1204 * We don't do similar optimization for completely idle system, as
1205 * selecting an idle cpu will add more delays to the timers than intended
1206 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1208 int get_nohz_timer_target(void)
1210 int cpu
= smp_processor_id();
1212 struct sched_domain
*sd
;
1215 for_each_domain(cpu
, sd
) {
1216 for_each_cpu(i
, sched_domain_span(sd
)) {
1228 * When add_timer_on() enqueues a timer into the timer wheel of an
1229 * idle CPU then this timer might expire before the next timer event
1230 * which is scheduled to wake up that CPU. In case of a completely
1231 * idle system the next event might even be infinite time into the
1232 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1233 * leaves the inner idle loop so the newly added timer is taken into
1234 * account when the CPU goes back to idle and evaluates the timer
1235 * wheel for the next timer event.
1237 void wake_up_idle_cpu(int cpu
)
1239 struct rq
*rq
= cpu_rq(cpu
);
1241 if (cpu
== smp_processor_id())
1245 * This is safe, as this function is called with the timer
1246 * wheel base lock of (cpu) held. When the CPU is on the way
1247 * to idle and has not yet set rq->curr to idle then it will
1248 * be serialized on the timer wheel base lock and take the new
1249 * timer into account automatically.
1251 if (rq
->curr
!= rq
->idle
)
1255 * We can set TIF_RESCHED on the idle task of the other CPU
1256 * lockless. The worst case is that the other CPU runs the
1257 * idle task through an additional NOOP schedule()
1259 set_tsk_need_resched(rq
->idle
);
1261 /* NEED_RESCHED must be visible before we test polling */
1263 if (!tsk_is_polling(rq
->idle
))
1264 smp_send_reschedule(cpu
);
1267 #endif /* CONFIG_NO_HZ */
1269 static u64
sched_avg_period(void)
1271 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1274 static void sched_avg_update(struct rq
*rq
)
1276 s64 period
= sched_avg_period();
1278 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1280 * Inline assembly required to prevent the compiler
1281 * optimising this loop into a divmod call.
1282 * See __iter_div_u64_rem() for another example of this.
1284 asm("" : "+rm" (rq
->age_stamp
));
1285 rq
->age_stamp
+= period
;
1290 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1292 rq
->rt_avg
+= rt_delta
;
1293 sched_avg_update(rq
);
1296 #else /* !CONFIG_SMP */
1297 static void resched_task(struct task_struct
*p
)
1299 assert_raw_spin_locked(&task_rq(p
)->lock
);
1300 set_tsk_need_resched(p
);
1303 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1307 static void sched_avg_update(struct rq
*rq
)
1310 #endif /* CONFIG_SMP */
1312 #if BITS_PER_LONG == 32
1313 # define WMULT_CONST (~0UL)
1315 # define WMULT_CONST (1UL << 32)
1318 #define WMULT_SHIFT 32
1321 * Shift right and round:
1323 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1326 * delta *= weight / lw
1328 static unsigned long
1329 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1330 struct load_weight
*lw
)
1335 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1336 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1337 * 2^SCHED_LOAD_RESOLUTION.
1339 if (likely(weight
> (1UL << SCHED_LOAD_RESOLUTION
)))
1340 tmp
= (u64
)delta_exec
* scale_load_down(weight
);
1342 tmp
= (u64
)delta_exec
;
1344 if (!lw
->inv_weight
) {
1345 unsigned long w
= scale_load_down(lw
->weight
);
1347 if (BITS_PER_LONG
> 32 && unlikely(w
>= WMULT_CONST
))
1349 else if (unlikely(!w
))
1350 lw
->inv_weight
= WMULT_CONST
;
1352 lw
->inv_weight
= WMULT_CONST
/ w
;
1356 * Check whether we'd overflow the 64-bit multiplication:
1358 if (unlikely(tmp
> WMULT_CONST
))
1359 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1362 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1364 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1367 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1373 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1379 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
1386 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1387 * of tasks with abnormal "nice" values across CPUs the contribution that
1388 * each task makes to its run queue's load is weighted according to its
1389 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1390 * scaled version of the new time slice allocation that they receive on time
1394 #define WEIGHT_IDLEPRIO 3
1395 #define WMULT_IDLEPRIO 1431655765
1398 * Nice levels are multiplicative, with a gentle 10% change for every
1399 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1400 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1401 * that remained on nice 0.
1403 * The "10% effect" is relative and cumulative: from _any_ nice level,
1404 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1405 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1406 * If a task goes up by ~10% and another task goes down by ~10% then
1407 * the relative distance between them is ~25%.)
1409 static const int prio_to_weight
[40] = {
1410 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1411 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1412 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1413 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1414 /* 0 */ 1024, 820, 655, 526, 423,
1415 /* 5 */ 335, 272, 215, 172, 137,
1416 /* 10 */ 110, 87, 70, 56, 45,
1417 /* 15 */ 36, 29, 23, 18, 15,
1421 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1423 * In cases where the weight does not change often, we can use the
1424 * precalculated inverse to speed up arithmetics by turning divisions
1425 * into multiplications:
1427 static const u32 prio_to_wmult
[40] = {
1428 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1429 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1430 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1431 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1432 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1433 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1434 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1435 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1438 /* Time spent by the tasks of the cpu accounting group executing in ... */
1439 enum cpuacct_stat_index
{
1440 CPUACCT_STAT_USER
, /* ... user mode */
1441 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1443 CPUACCT_STAT_NSTATS
,
1446 #ifdef CONFIG_CGROUP_CPUACCT
1447 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1448 static void cpuacct_update_stats(struct task_struct
*tsk
,
1449 enum cpuacct_stat_index idx
, cputime_t val
);
1451 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1452 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1453 enum cpuacct_stat_index idx
, cputime_t val
) {}
1456 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1458 update_load_add(&rq
->load
, load
);
1461 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1463 update_load_sub(&rq
->load
, load
);
1466 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1467 typedef int (*tg_visitor
)(struct task_group
*, void *);
1470 * Iterate the full tree, calling @down when first entering a node and @up when
1471 * leaving it for the final time.
1473 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1475 struct task_group
*parent
, *child
;
1479 parent
= &root_task_group
;
1481 ret
= (*down
)(parent
, data
);
1484 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1491 ret
= (*up
)(parent
, data
);
1496 parent
= parent
->parent
;
1505 static int tg_nop(struct task_group
*tg
, void *data
)
1512 /* Used instead of source_load when we know the type == 0 */
1513 static unsigned long weighted_cpuload(const int cpu
)
1515 return cpu_rq(cpu
)->load
.weight
;
1519 * Return a low guess at the load of a migration-source cpu weighted
1520 * according to the scheduling class and "nice" value.
1522 * We want to under-estimate the load of migration sources, to
1523 * balance conservatively.
1525 static unsigned long source_load(int cpu
, int type
)
1527 struct rq
*rq
= cpu_rq(cpu
);
1528 unsigned long total
= weighted_cpuload(cpu
);
1530 if (type
== 0 || !sched_feat(LB_BIAS
))
1533 return min(rq
->cpu_load
[type
-1], total
);
1537 * Return a high guess at the load of a migration-target cpu weighted
1538 * according to the scheduling class and "nice" value.
1540 static unsigned long target_load(int cpu
, int type
)
1542 struct rq
*rq
= cpu_rq(cpu
);
1543 unsigned long total
= weighted_cpuload(cpu
);
1545 if (type
== 0 || !sched_feat(LB_BIAS
))
1548 return max(rq
->cpu_load
[type
-1], total
);
1551 static unsigned long power_of(int cpu
)
1553 return cpu_rq(cpu
)->cpu_power
;
1556 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1558 static unsigned long cpu_avg_load_per_task(int cpu
)
1560 struct rq
*rq
= cpu_rq(cpu
);
1561 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1564 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1566 rq
->avg_load_per_task
= 0;
1568 return rq
->avg_load_per_task
;
1571 #ifdef CONFIG_FAIR_GROUP_SCHED
1574 * Compute the cpu's hierarchical load factor for each task group.
1575 * This needs to be done in a top-down fashion because the load of a child
1576 * group is a fraction of its parents load.
1578 static int tg_load_down(struct task_group
*tg
, void *data
)
1581 long cpu
= (long)data
;
1584 load
= cpu_rq(cpu
)->load
.weight
;
1586 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1587 load
*= tg
->se
[cpu
]->load
.weight
;
1588 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1591 tg
->cfs_rq
[cpu
]->h_load
= load
;
1596 static void update_h_load(long cpu
)
1598 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1603 #ifdef CONFIG_PREEMPT
1605 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1608 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1609 * way at the expense of forcing extra atomic operations in all
1610 * invocations. This assures that the double_lock is acquired using the
1611 * same underlying policy as the spinlock_t on this architecture, which
1612 * reduces latency compared to the unfair variant below. However, it
1613 * also adds more overhead and therefore may reduce throughput.
1615 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1616 __releases(this_rq
->lock
)
1617 __acquires(busiest
->lock
)
1618 __acquires(this_rq
->lock
)
1620 raw_spin_unlock(&this_rq
->lock
);
1621 double_rq_lock(this_rq
, busiest
);
1628 * Unfair double_lock_balance: Optimizes throughput at the expense of
1629 * latency by eliminating extra atomic operations when the locks are
1630 * already in proper order on entry. This favors lower cpu-ids and will
1631 * grant the double lock to lower cpus over higher ids under contention,
1632 * regardless of entry order into the function.
1634 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1635 __releases(this_rq
->lock
)
1636 __acquires(busiest
->lock
)
1637 __acquires(this_rq
->lock
)
1641 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1642 if (busiest
< this_rq
) {
1643 raw_spin_unlock(&this_rq
->lock
);
1644 raw_spin_lock(&busiest
->lock
);
1645 raw_spin_lock_nested(&this_rq
->lock
,
1646 SINGLE_DEPTH_NESTING
);
1649 raw_spin_lock_nested(&busiest
->lock
,
1650 SINGLE_DEPTH_NESTING
);
1655 #endif /* CONFIG_PREEMPT */
1658 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1660 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1662 if (unlikely(!irqs_disabled())) {
1663 /* printk() doesn't work good under rq->lock */
1664 raw_spin_unlock(&this_rq
->lock
);
1668 return _double_lock_balance(this_rq
, busiest
);
1671 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1672 __releases(busiest
->lock
)
1674 raw_spin_unlock(&busiest
->lock
);
1675 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1679 * double_rq_lock - safely lock two runqueues
1681 * Note this does not disable interrupts like task_rq_lock,
1682 * you need to do so manually before calling.
1684 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1685 __acquires(rq1
->lock
)
1686 __acquires(rq2
->lock
)
1688 BUG_ON(!irqs_disabled());
1690 raw_spin_lock(&rq1
->lock
);
1691 __acquire(rq2
->lock
); /* Fake it out ;) */
1694 raw_spin_lock(&rq1
->lock
);
1695 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1697 raw_spin_lock(&rq2
->lock
);
1698 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1704 * double_rq_unlock - safely unlock two runqueues
1706 * Note this does not restore interrupts like task_rq_unlock,
1707 * you need to do so manually after calling.
1709 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1710 __releases(rq1
->lock
)
1711 __releases(rq2
->lock
)
1713 raw_spin_unlock(&rq1
->lock
);
1715 raw_spin_unlock(&rq2
->lock
);
1717 __release(rq2
->lock
);
1720 #else /* CONFIG_SMP */
1723 * double_rq_lock - safely lock two runqueues
1725 * Note this does not disable interrupts like task_rq_lock,
1726 * you need to do so manually before calling.
1728 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1729 __acquires(rq1
->lock
)
1730 __acquires(rq2
->lock
)
1732 BUG_ON(!irqs_disabled());
1734 raw_spin_lock(&rq1
->lock
);
1735 __acquire(rq2
->lock
); /* Fake it out ;) */
1739 * double_rq_unlock - safely unlock two runqueues
1741 * Note this does not restore interrupts like task_rq_unlock,
1742 * you need to do so manually after calling.
1744 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1745 __releases(rq1
->lock
)
1746 __releases(rq2
->lock
)
1749 raw_spin_unlock(&rq1
->lock
);
1750 __release(rq2
->lock
);
1755 static void calc_load_account_idle(struct rq
*this_rq
);
1756 static void update_sysctl(void);
1757 static int get_update_sysctl_factor(void);
1758 static void update_cpu_load(struct rq
*this_rq
);
1760 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1762 set_task_rq(p
, cpu
);
1765 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1766 * successfuly executed on another CPU. We must ensure that updates of
1767 * per-task data have been completed by this moment.
1770 task_thread_info(p
)->cpu
= cpu
;
1774 static const struct sched_class rt_sched_class
;
1776 #define sched_class_highest (&stop_sched_class)
1777 #define for_each_class(class) \
1778 for (class = sched_class_highest; class; class = class->next)
1780 #include "sched_stats.h"
1782 static void inc_nr_running(struct rq
*rq
)
1787 static void dec_nr_running(struct rq
*rq
)
1792 static void set_load_weight(struct task_struct
*p
)
1794 int prio
= p
->static_prio
- MAX_RT_PRIO
;
1795 struct load_weight
*load
= &p
->se
.load
;
1798 * SCHED_IDLE tasks get minimal weight:
1800 if (p
->policy
== SCHED_IDLE
) {
1801 load
->weight
= scale_load(WEIGHT_IDLEPRIO
);
1802 load
->inv_weight
= WMULT_IDLEPRIO
;
1806 load
->weight
= scale_load(prio_to_weight
[prio
]);
1807 load
->inv_weight
= prio_to_wmult
[prio
];
1810 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1812 update_rq_clock(rq
);
1813 sched_info_queued(p
);
1814 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1817 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1819 update_rq_clock(rq
);
1820 sched_info_dequeued(p
);
1821 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1825 * activate_task - move a task to the runqueue.
1827 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1829 if (task_contributes_to_load(p
))
1830 rq
->nr_uninterruptible
--;
1832 enqueue_task(rq
, p
, flags
);
1837 * deactivate_task - remove a task from the runqueue.
1839 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1841 if (task_contributes_to_load(p
))
1842 rq
->nr_uninterruptible
++;
1844 dequeue_task(rq
, p
, flags
);
1848 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1851 * There are no locks covering percpu hardirq/softirq time.
1852 * They are only modified in account_system_vtime, on corresponding CPU
1853 * with interrupts disabled. So, writes are safe.
1854 * They are read and saved off onto struct rq in update_rq_clock().
1855 * This may result in other CPU reading this CPU's irq time and can
1856 * race with irq/account_system_vtime on this CPU. We would either get old
1857 * or new value with a side effect of accounting a slice of irq time to wrong
1858 * task when irq is in progress while we read rq->clock. That is a worthy
1859 * compromise in place of having locks on each irq in account_system_time.
1861 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1862 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1864 static DEFINE_PER_CPU(u64
, irq_start_time
);
1865 static int sched_clock_irqtime
;
1867 void enable_sched_clock_irqtime(void)
1869 sched_clock_irqtime
= 1;
1872 void disable_sched_clock_irqtime(void)
1874 sched_clock_irqtime
= 0;
1877 #ifndef CONFIG_64BIT
1878 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
1880 static inline void irq_time_write_begin(void)
1882 __this_cpu_inc(irq_time_seq
.sequence
);
1886 static inline void irq_time_write_end(void)
1889 __this_cpu_inc(irq_time_seq
.sequence
);
1892 static inline u64
irq_time_read(int cpu
)
1898 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1899 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1900 per_cpu(cpu_hardirq_time
, cpu
);
1901 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1905 #else /* CONFIG_64BIT */
1906 static inline void irq_time_write_begin(void)
1910 static inline void irq_time_write_end(void)
1914 static inline u64
irq_time_read(int cpu
)
1916 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1918 #endif /* CONFIG_64BIT */
1921 * Called before incrementing preempt_count on {soft,}irq_enter
1922 * and before decrementing preempt_count on {soft,}irq_exit.
1924 void account_system_vtime(struct task_struct
*curr
)
1926 unsigned long flags
;
1930 if (!sched_clock_irqtime
)
1933 local_irq_save(flags
);
1935 cpu
= smp_processor_id();
1936 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
1937 __this_cpu_add(irq_start_time
, delta
);
1939 irq_time_write_begin();
1941 * We do not account for softirq time from ksoftirqd here.
1942 * We want to continue accounting softirq time to ksoftirqd thread
1943 * in that case, so as not to confuse scheduler with a special task
1944 * that do not consume any time, but still wants to run.
1946 if (hardirq_count())
1947 __this_cpu_add(cpu_hardirq_time
, delta
);
1948 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
1949 __this_cpu_add(cpu_softirq_time
, delta
);
1951 irq_time_write_end();
1952 local_irq_restore(flags
);
1954 EXPORT_SYMBOL_GPL(account_system_vtime
);
1956 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1960 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
1963 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1964 * this case when a previous update_rq_clock() happened inside a
1965 * {soft,}irq region.
1967 * When this happens, we stop ->clock_task and only update the
1968 * prev_irq_time stamp to account for the part that fit, so that a next
1969 * update will consume the rest. This ensures ->clock_task is
1972 * It does however cause some slight miss-attribution of {soft,}irq
1973 * time, a more accurate solution would be to update the irq_time using
1974 * the current rq->clock timestamp, except that would require using
1977 if (irq_delta
> delta
)
1980 rq
->prev_irq_time
+= irq_delta
;
1982 rq
->clock_task
+= delta
;
1984 if (irq_delta
&& sched_feat(NONIRQ_POWER
))
1985 sched_rt_avg_update(rq
, irq_delta
);
1988 static int irqtime_account_hi_update(void)
1990 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
1991 unsigned long flags
;
1995 local_irq_save(flags
);
1996 latest_ns
= this_cpu_read(cpu_hardirq_time
);
1997 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->irq
))
1999 local_irq_restore(flags
);
2003 static int irqtime_account_si_update(void)
2005 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
2006 unsigned long flags
;
2010 local_irq_save(flags
);
2011 latest_ns
= this_cpu_read(cpu_softirq_time
);
2012 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->softirq
))
2014 local_irq_restore(flags
);
2018 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2020 #define sched_clock_irqtime (0)
2022 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
2024 rq
->clock_task
+= delta
;
2027 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2029 #include "sched_idletask.c"
2030 #include "sched_fair.c"
2031 #include "sched_rt.c"
2032 #include "sched_autogroup.c"
2033 #include "sched_stoptask.c"
2034 #ifdef CONFIG_SCHED_DEBUG
2035 # include "sched_debug.c"
2038 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2040 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2041 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2045 * Make it appear like a SCHED_FIFO task, its something
2046 * userspace knows about and won't get confused about.
2048 * Also, it will make PI more or less work without too
2049 * much confusion -- but then, stop work should not
2050 * rely on PI working anyway.
2052 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2054 stop
->sched_class
= &stop_sched_class
;
2057 cpu_rq(cpu
)->stop
= stop
;
2061 * Reset it back to a normal scheduling class so that
2062 * it can die in pieces.
2064 old_stop
->sched_class
= &rt_sched_class
;
2069 * __normal_prio - return the priority that is based on the static prio
2071 static inline int __normal_prio(struct task_struct
*p
)
2073 return p
->static_prio
;
2077 * Calculate the expected normal priority: i.e. priority
2078 * without taking RT-inheritance into account. Might be
2079 * boosted by interactivity modifiers. Changes upon fork,
2080 * setprio syscalls, and whenever the interactivity
2081 * estimator recalculates.
2083 static inline int normal_prio(struct task_struct
*p
)
2087 if (task_has_rt_policy(p
))
2088 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2090 prio
= __normal_prio(p
);
2095 * Calculate the current priority, i.e. the priority
2096 * taken into account by the scheduler. This value might
2097 * be boosted by RT tasks, or might be boosted by
2098 * interactivity modifiers. Will be RT if the task got
2099 * RT-boosted. If not then it returns p->normal_prio.
2101 static int effective_prio(struct task_struct
*p
)
2103 p
->normal_prio
= normal_prio(p
);
2105 * If we are RT tasks or we were boosted to RT priority,
2106 * keep the priority unchanged. Otherwise, update priority
2107 * to the normal priority:
2109 if (!rt_prio(p
->prio
))
2110 return p
->normal_prio
;
2115 * task_curr - is this task currently executing on a CPU?
2116 * @p: the task in question.
2118 inline int task_curr(const struct task_struct
*p
)
2120 return cpu_curr(task_cpu(p
)) == p
;
2123 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2124 const struct sched_class
*prev_class
,
2127 if (prev_class
!= p
->sched_class
) {
2128 if (prev_class
->switched_from
)
2129 prev_class
->switched_from(rq
, p
);
2130 p
->sched_class
->switched_to(rq
, p
);
2131 } else if (oldprio
!= p
->prio
)
2132 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2135 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2137 const struct sched_class
*class;
2139 if (p
->sched_class
== rq
->curr
->sched_class
) {
2140 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2142 for_each_class(class) {
2143 if (class == rq
->curr
->sched_class
)
2145 if (class == p
->sched_class
) {
2146 resched_task(rq
->curr
);
2153 * A queue event has occurred, and we're going to schedule. In
2154 * this case, we can save a useless back to back clock update.
2156 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
2157 rq
->skip_clock_update
= 1;
2162 * Is this task likely cache-hot:
2165 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2169 if (p
->sched_class
!= &fair_sched_class
)
2172 if (unlikely(p
->policy
== SCHED_IDLE
))
2176 * Buddy candidates are cache hot:
2178 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2179 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2180 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2183 if (sysctl_sched_migration_cost
== -1)
2185 if (sysctl_sched_migration_cost
== 0)
2188 delta
= now
- p
->se
.exec_start
;
2190 return delta
< (s64
)sysctl_sched_migration_cost
;
2193 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2195 #ifdef CONFIG_SCHED_DEBUG
2197 * We should never call set_task_cpu() on a blocked task,
2198 * ttwu() will sort out the placement.
2200 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2201 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2203 #ifdef CONFIG_LOCKDEP
2205 * The caller should hold either p->pi_lock or rq->lock, when changing
2206 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2208 * sched_move_task() holds both and thus holding either pins the cgroup,
2209 * see set_task_rq().
2211 * Furthermore, all task_rq users should acquire both locks, see
2214 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2215 lockdep_is_held(&task_rq(p
)->lock
)));
2219 trace_sched_migrate_task(p
, new_cpu
);
2221 if (task_cpu(p
) != new_cpu
) {
2222 p
->se
.nr_migrations
++;
2223 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2226 __set_task_cpu(p
, new_cpu
);
2229 struct migration_arg
{
2230 struct task_struct
*task
;
2234 static int migration_cpu_stop(void *data
);
2237 * wait_task_inactive - wait for a thread to unschedule.
2239 * If @match_state is nonzero, it's the @p->state value just checked and
2240 * not expected to change. If it changes, i.e. @p might have woken up,
2241 * then return zero. When we succeed in waiting for @p to be off its CPU,
2242 * we return a positive number (its total switch count). If a second call
2243 * a short while later returns the same number, the caller can be sure that
2244 * @p has remained unscheduled the whole time.
2246 * The caller must ensure that the task *will* unschedule sometime soon,
2247 * else this function might spin for a *long* time. This function can't
2248 * be called with interrupts off, or it may introduce deadlock with
2249 * smp_call_function() if an IPI is sent by the same process we are
2250 * waiting to become inactive.
2252 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2254 unsigned long flags
;
2261 * We do the initial early heuristics without holding
2262 * any task-queue locks at all. We'll only try to get
2263 * the runqueue lock when things look like they will
2269 * If the task is actively running on another CPU
2270 * still, just relax and busy-wait without holding
2273 * NOTE! Since we don't hold any locks, it's not
2274 * even sure that "rq" stays as the right runqueue!
2275 * But we don't care, since "task_running()" will
2276 * return false if the runqueue has changed and p
2277 * is actually now running somewhere else!
2279 while (task_running(rq
, p
)) {
2280 if (match_state
&& unlikely(p
->state
!= match_state
))
2286 * Ok, time to look more closely! We need the rq
2287 * lock now, to be *sure*. If we're wrong, we'll
2288 * just go back and repeat.
2290 rq
= task_rq_lock(p
, &flags
);
2291 trace_sched_wait_task(p
);
2292 running
= task_running(rq
, p
);
2295 if (!match_state
|| p
->state
== match_state
)
2296 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2297 task_rq_unlock(rq
, p
, &flags
);
2300 * If it changed from the expected state, bail out now.
2302 if (unlikely(!ncsw
))
2306 * Was it really running after all now that we
2307 * checked with the proper locks actually held?
2309 * Oops. Go back and try again..
2311 if (unlikely(running
)) {
2317 * It's not enough that it's not actively running,
2318 * it must be off the runqueue _entirely_, and not
2321 * So if it was still runnable (but just not actively
2322 * running right now), it's preempted, and we should
2323 * yield - it could be a while.
2325 if (unlikely(on_rq
)) {
2326 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
2328 set_current_state(TASK_UNINTERRUPTIBLE
);
2329 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2334 * Ahh, all good. It wasn't running, and it wasn't
2335 * runnable, which means that it will never become
2336 * running in the future either. We're all done!
2345 * kick_process - kick a running thread to enter/exit the kernel
2346 * @p: the to-be-kicked thread
2348 * Cause a process which is running on another CPU to enter
2349 * kernel-mode, without any delay. (to get signals handled.)
2351 * NOTE: this function doesn't have to take the runqueue lock,
2352 * because all it wants to ensure is that the remote task enters
2353 * the kernel. If the IPI races and the task has been migrated
2354 * to another CPU then no harm is done and the purpose has been
2357 void kick_process(struct task_struct
*p
)
2363 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2364 smp_send_reschedule(cpu
);
2367 EXPORT_SYMBOL_GPL(kick_process
);
2368 #endif /* CONFIG_SMP */
2372 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2374 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2377 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2379 /* Look for allowed, online CPU in same node. */
2380 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2381 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2384 /* Any allowed, online CPU? */
2385 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2386 if (dest_cpu
< nr_cpu_ids
)
2389 /* No more Mr. Nice Guy. */
2390 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2392 * Don't tell them about moving exiting tasks or
2393 * kernel threads (both mm NULL), since they never
2396 if (p
->mm
&& printk_ratelimit()) {
2397 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2398 task_pid_nr(p
), p
->comm
, cpu
);
2405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2408 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2410 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2413 * In order not to call set_task_cpu() on a blocking task we need
2414 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2417 * Since this is common to all placement strategies, this lives here.
2419 * [ this allows ->select_task() to simply return task_cpu(p) and
2420 * not worry about this generic constraint ]
2422 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2424 cpu
= select_fallback_rq(task_cpu(p
), p
);
2429 static void update_avg(u64
*avg
, u64 sample
)
2431 s64 diff
= sample
- *avg
;
2437 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2439 #ifdef CONFIG_SCHEDSTATS
2440 struct rq
*rq
= this_rq();
2443 int this_cpu
= smp_processor_id();
2445 if (cpu
== this_cpu
) {
2446 schedstat_inc(rq
, ttwu_local
);
2447 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2449 struct sched_domain
*sd
;
2451 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2453 for_each_domain(this_cpu
, sd
) {
2454 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2455 schedstat_inc(sd
, ttwu_wake_remote
);
2462 if (wake_flags
& WF_MIGRATED
)
2463 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2465 #endif /* CONFIG_SMP */
2467 schedstat_inc(rq
, ttwu_count
);
2468 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2470 if (wake_flags
& WF_SYNC
)
2471 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2473 #endif /* CONFIG_SCHEDSTATS */
2476 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
2478 activate_task(rq
, p
, en_flags
);
2481 /* if a worker is waking up, notify workqueue */
2482 if (p
->flags
& PF_WQ_WORKER
)
2483 wq_worker_waking_up(p
, cpu_of(rq
));
2487 * Mark the task runnable and perform wakeup-preemption.
2490 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2492 trace_sched_wakeup(p
, true);
2493 check_preempt_curr(rq
, p
, wake_flags
);
2495 p
->state
= TASK_RUNNING
;
2497 if (p
->sched_class
->task_woken
)
2498 p
->sched_class
->task_woken(rq
, p
);
2500 if (unlikely(rq
->idle_stamp
)) {
2501 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2502 u64 max
= 2*sysctl_sched_migration_cost
;
2507 update_avg(&rq
->avg_idle
, delta
);
2514 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2517 if (p
->sched_contributes_to_load
)
2518 rq
->nr_uninterruptible
--;
2521 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
2522 ttwu_do_wakeup(rq
, p
, wake_flags
);
2526 * Called in case the task @p isn't fully descheduled from its runqueue,
2527 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2528 * since all we need to do is flip p->state to TASK_RUNNING, since
2529 * the task is still ->on_rq.
2531 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2536 rq
= __task_rq_lock(p
);
2538 ttwu_do_wakeup(rq
, p
, wake_flags
);
2541 __task_rq_unlock(rq
);
2547 static void sched_ttwu_do_pending(struct task_struct
*list
)
2549 struct rq
*rq
= this_rq();
2551 raw_spin_lock(&rq
->lock
);
2554 struct task_struct
*p
= list
;
2555 list
= list
->wake_entry
;
2556 ttwu_do_activate(rq
, p
, 0);
2559 raw_spin_unlock(&rq
->lock
);
2562 #ifdef CONFIG_HOTPLUG_CPU
2564 static void sched_ttwu_pending(void)
2566 struct rq
*rq
= this_rq();
2567 struct task_struct
*list
= xchg(&rq
->wake_list
, NULL
);
2572 sched_ttwu_do_pending(list
);
2575 #endif /* CONFIG_HOTPLUG_CPU */
2577 void scheduler_ipi(void)
2579 struct rq
*rq
= this_rq();
2580 struct task_struct
*list
= xchg(&rq
->wake_list
, NULL
);
2586 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2587 * traditionally all their work was done from the interrupt return
2588 * path. Now that we actually do some work, we need to make sure
2591 * Some archs already do call them, luckily irq_enter/exit nest
2594 * Arguably we should visit all archs and update all handlers,
2595 * however a fair share of IPIs are still resched only so this would
2596 * somewhat pessimize the simple resched case.
2599 sched_ttwu_do_pending(list
);
2603 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
2605 struct rq
*rq
= cpu_rq(cpu
);
2606 struct task_struct
*next
= rq
->wake_list
;
2609 struct task_struct
*old
= next
;
2611 p
->wake_entry
= next
;
2612 next
= cmpxchg(&rq
->wake_list
, old
, p
);
2618 smp_send_reschedule(cpu
);
2621 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2622 static int ttwu_activate_remote(struct task_struct
*p
, int wake_flags
)
2627 rq
= __task_rq_lock(p
);
2629 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2630 ttwu_do_wakeup(rq
, p
, wake_flags
);
2633 __task_rq_unlock(rq
);
2638 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2639 #endif /* CONFIG_SMP */
2641 static void ttwu_queue(struct task_struct
*p
, int cpu
)
2643 struct rq
*rq
= cpu_rq(cpu
);
2645 #if defined(CONFIG_SMP)
2646 if (sched_feat(TTWU_QUEUE
) && cpu
!= smp_processor_id()) {
2647 sched_clock_cpu(cpu
); /* sync clocks x-cpu */
2648 ttwu_queue_remote(p
, cpu
);
2653 raw_spin_lock(&rq
->lock
);
2654 ttwu_do_activate(rq
, p
, 0);
2655 raw_spin_unlock(&rq
->lock
);
2659 * try_to_wake_up - wake up a thread
2660 * @p: the thread to be awakened
2661 * @state: the mask of task states that can be woken
2662 * @wake_flags: wake modifier flags (WF_*)
2664 * Put it on the run-queue if it's not already there. The "current"
2665 * thread is always on the run-queue (except when the actual
2666 * re-schedule is in progress), and as such you're allowed to do
2667 * the simpler "current->state = TASK_RUNNING" to mark yourself
2668 * runnable without the overhead of this.
2670 * Returns %true if @p was woken up, %false if it was already running
2671 * or @state didn't match @p's state.
2674 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2676 unsigned long flags
;
2677 int cpu
, success
= 0;
2680 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2681 if (!(p
->state
& state
))
2684 success
= 1; /* we're going to change ->state */
2687 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2692 * If the owning (remote) cpu is still in the middle of schedule() with
2693 * this task as prev, wait until its done referencing the task.
2696 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2698 * In case the architecture enables interrupts in
2699 * context_switch(), we cannot busy wait, since that
2700 * would lead to deadlocks when an interrupt hits and
2701 * tries to wake up @prev. So bail and do a complete
2704 if (ttwu_activate_remote(p
, wake_flags
))
2711 * Pairs with the smp_wmb() in finish_lock_switch().
2715 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2716 p
->state
= TASK_WAKING
;
2718 if (p
->sched_class
->task_waking
)
2719 p
->sched_class
->task_waking(p
);
2721 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2722 if (task_cpu(p
) != cpu
) {
2723 wake_flags
|= WF_MIGRATED
;
2724 set_task_cpu(p
, cpu
);
2726 #endif /* CONFIG_SMP */
2730 ttwu_stat(p
, cpu
, wake_flags
);
2732 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2738 * try_to_wake_up_local - try to wake up a local task with rq lock held
2739 * @p: the thread to be awakened
2741 * Put @p on the run-queue if it's not already there. The caller must
2742 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2745 static void try_to_wake_up_local(struct task_struct
*p
)
2747 struct rq
*rq
= task_rq(p
);
2749 BUG_ON(rq
!= this_rq());
2750 BUG_ON(p
== current
);
2751 lockdep_assert_held(&rq
->lock
);
2753 if (!raw_spin_trylock(&p
->pi_lock
)) {
2754 raw_spin_unlock(&rq
->lock
);
2755 raw_spin_lock(&p
->pi_lock
);
2756 raw_spin_lock(&rq
->lock
);
2759 if (!(p
->state
& TASK_NORMAL
))
2763 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2765 ttwu_do_wakeup(rq
, p
, 0);
2766 ttwu_stat(p
, smp_processor_id(), 0);
2768 raw_spin_unlock(&p
->pi_lock
);
2772 * wake_up_process - Wake up a specific process
2773 * @p: The process to be woken up.
2775 * Attempt to wake up the nominated process and move it to the set of runnable
2776 * processes. Returns 1 if the process was woken up, 0 if it was already
2779 * It may be assumed that this function implies a write memory barrier before
2780 * changing the task state if and only if any tasks are woken up.
2782 int wake_up_process(struct task_struct
*p
)
2784 return try_to_wake_up(p
, TASK_ALL
, 0);
2786 EXPORT_SYMBOL(wake_up_process
);
2788 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2790 return try_to_wake_up(p
, state
, 0);
2794 * Perform scheduler related setup for a newly forked process p.
2795 * p is forked by current.
2797 * __sched_fork() is basic setup used by init_idle() too:
2799 static void __sched_fork(struct task_struct
*p
)
2804 p
->se
.exec_start
= 0;
2805 p
->se
.sum_exec_runtime
= 0;
2806 p
->se
.prev_sum_exec_runtime
= 0;
2807 p
->se
.nr_migrations
= 0;
2809 INIT_LIST_HEAD(&p
->se
.group_node
);
2811 #ifdef CONFIG_SCHEDSTATS
2812 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2815 INIT_LIST_HEAD(&p
->rt
.run_list
);
2817 #ifdef CONFIG_PREEMPT_NOTIFIERS
2818 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2823 * fork()/clone()-time setup:
2825 void sched_fork(struct task_struct
*p
)
2827 unsigned long flags
;
2828 int cpu
= get_cpu();
2832 * We mark the process as running here. This guarantees that
2833 * nobody will actually run it, and a signal or other external
2834 * event cannot wake it up and insert it on the runqueue either.
2836 p
->state
= TASK_RUNNING
;
2839 * Revert to default priority/policy on fork if requested.
2841 if (unlikely(p
->sched_reset_on_fork
)) {
2842 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2843 p
->policy
= SCHED_NORMAL
;
2844 p
->normal_prio
= p
->static_prio
;
2847 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2848 p
->static_prio
= NICE_TO_PRIO(0);
2849 p
->normal_prio
= p
->static_prio
;
2854 * We don't need the reset flag anymore after the fork. It has
2855 * fulfilled its duty:
2857 p
->sched_reset_on_fork
= 0;
2861 * Make sure we do not leak PI boosting priority to the child.
2863 p
->prio
= current
->normal_prio
;
2865 if (!rt_prio(p
->prio
))
2866 p
->sched_class
= &fair_sched_class
;
2868 if (p
->sched_class
->task_fork
)
2869 p
->sched_class
->task_fork(p
);
2872 * The child is not yet in the pid-hash so no cgroup attach races,
2873 * and the cgroup is pinned to this child due to cgroup_fork()
2874 * is ran before sched_fork().
2876 * Silence PROVE_RCU.
2878 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2879 set_task_cpu(p
, cpu
);
2880 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2882 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2883 if (likely(sched_info_on()))
2884 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2886 #if defined(CONFIG_SMP)
2889 #ifdef CONFIG_PREEMPT
2890 /* Want to start with kernel preemption disabled. */
2891 task_thread_info(p
)->preempt_count
= 1;
2894 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2901 * wake_up_new_task - wake up a newly created task for the first time.
2903 * This function will do some initial scheduler statistics housekeeping
2904 * that must be done for every newly created context, then puts the task
2905 * on the runqueue and wakes it.
2907 void wake_up_new_task(struct task_struct
*p
)
2909 unsigned long flags
;
2912 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2915 * Fork balancing, do it here and not earlier because:
2916 * - cpus_allowed can change in the fork path
2917 * - any previously selected cpu might disappear through hotplug
2919 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
2922 rq
= __task_rq_lock(p
);
2923 activate_task(rq
, p
, 0);
2925 trace_sched_wakeup_new(p
, true);
2926 check_preempt_curr(rq
, p
, WF_FORK
);
2928 if (p
->sched_class
->task_woken
)
2929 p
->sched_class
->task_woken(rq
, p
);
2931 task_rq_unlock(rq
, p
, &flags
);
2934 #ifdef CONFIG_PREEMPT_NOTIFIERS
2937 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2938 * @notifier: notifier struct to register
2940 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2942 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2944 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2947 * preempt_notifier_unregister - no longer interested in preemption notifications
2948 * @notifier: notifier struct to unregister
2950 * This is safe to call from within a preemption notifier.
2952 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2954 hlist_del(¬ifier
->link
);
2956 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2958 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2960 struct preempt_notifier
*notifier
;
2961 struct hlist_node
*node
;
2963 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2964 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2968 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2969 struct task_struct
*next
)
2971 struct preempt_notifier
*notifier
;
2972 struct hlist_node
*node
;
2974 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2975 notifier
->ops
->sched_out(notifier
, next
);
2978 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2980 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2985 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2986 struct task_struct
*next
)
2990 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2993 * prepare_task_switch - prepare to switch tasks
2994 * @rq: the runqueue preparing to switch
2995 * @prev: the current task that is being switched out
2996 * @next: the task we are going to switch to.
2998 * This is called with the rq lock held and interrupts off. It must
2999 * be paired with a subsequent finish_task_switch after the context
3002 * prepare_task_switch sets up locking and calls architecture specific
3006 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
3007 struct task_struct
*next
)
3009 sched_info_switch(prev
, next
);
3010 perf_event_task_sched_out(prev
, next
);
3011 fire_sched_out_preempt_notifiers(prev
, next
);
3012 prepare_lock_switch(rq
, next
);
3013 prepare_arch_switch(next
);
3014 trace_sched_switch(prev
, next
);
3018 * finish_task_switch - clean up after a task-switch
3019 * @rq: runqueue associated with task-switch
3020 * @prev: the thread we just switched away from.
3022 * finish_task_switch must be called after the context switch, paired
3023 * with a prepare_task_switch call before the context switch.
3024 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3025 * and do any other architecture-specific cleanup actions.
3027 * Note that we may have delayed dropping an mm in context_switch(). If
3028 * so, we finish that here outside of the runqueue lock. (Doing it
3029 * with the lock held can cause deadlocks; see schedule() for
3032 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
3033 __releases(rq
->lock
)
3035 struct mm_struct
*mm
= rq
->prev_mm
;
3041 * A task struct has one reference for the use as "current".
3042 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3043 * schedule one last time. The schedule call will never return, and
3044 * the scheduled task must drop that reference.
3045 * The test for TASK_DEAD must occur while the runqueue locks are
3046 * still held, otherwise prev could be scheduled on another cpu, die
3047 * there before we look at prev->state, and then the reference would
3049 * Manfred Spraul <manfred@colorfullife.com>
3051 prev_state
= prev
->state
;
3052 finish_arch_switch(prev
);
3053 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3054 local_irq_disable();
3055 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3056 perf_event_task_sched_in(current
);
3057 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3059 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3060 finish_lock_switch(rq
, prev
);
3062 fire_sched_in_preempt_notifiers(current
);
3065 if (unlikely(prev_state
== TASK_DEAD
)) {
3067 * Remove function-return probe instances associated with this
3068 * task and put them back on the free list.
3070 kprobe_flush_task(prev
);
3071 put_task_struct(prev
);
3077 /* assumes rq->lock is held */
3078 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
3080 if (prev
->sched_class
->pre_schedule
)
3081 prev
->sched_class
->pre_schedule(rq
, prev
);
3084 /* rq->lock is NOT held, but preemption is disabled */
3085 static inline void post_schedule(struct rq
*rq
)
3087 if (rq
->post_schedule
) {
3088 unsigned long flags
;
3090 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3091 if (rq
->curr
->sched_class
->post_schedule
)
3092 rq
->curr
->sched_class
->post_schedule(rq
);
3093 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3095 rq
->post_schedule
= 0;
3101 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
3105 static inline void post_schedule(struct rq
*rq
)
3112 * schedule_tail - first thing a freshly forked thread must call.
3113 * @prev: the thread we just switched away from.
3115 asmlinkage
void schedule_tail(struct task_struct
*prev
)
3116 __releases(rq
->lock
)
3118 struct rq
*rq
= this_rq();
3120 finish_task_switch(rq
, prev
);
3123 * FIXME: do we need to worry about rq being invalidated by the
3128 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3129 /* In this case, finish_task_switch does not reenable preemption */
3132 if (current
->set_child_tid
)
3133 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3137 * context_switch - switch to the new MM and the new
3138 * thread's register state.
3141 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3142 struct task_struct
*next
)
3144 struct mm_struct
*mm
, *oldmm
;
3146 prepare_task_switch(rq
, prev
, next
);
3149 oldmm
= prev
->active_mm
;
3151 * For paravirt, this is coupled with an exit in switch_to to
3152 * combine the page table reload and the switch backend into
3155 arch_start_context_switch(prev
);
3158 next
->active_mm
= oldmm
;
3159 atomic_inc(&oldmm
->mm_count
);
3160 enter_lazy_tlb(oldmm
, next
);
3162 switch_mm(oldmm
, mm
, next
);
3165 prev
->active_mm
= NULL
;
3166 rq
->prev_mm
= oldmm
;
3169 * Since the runqueue lock will be released by the next
3170 * task (which is an invalid locking op but in the case
3171 * of the scheduler it's an obvious special-case), so we
3172 * do an early lockdep release here:
3174 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3175 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3178 /* Here we just switch the register state and the stack. */
3179 switch_to(prev
, next
, prev
);
3183 * this_rq must be evaluated again because prev may have moved
3184 * CPUs since it called schedule(), thus the 'rq' on its stack
3185 * frame will be invalid.
3187 finish_task_switch(this_rq(), prev
);
3191 * nr_running, nr_uninterruptible and nr_context_switches:
3193 * externally visible scheduler statistics: current number of runnable
3194 * threads, current number of uninterruptible-sleeping threads, total
3195 * number of context switches performed since bootup.
3197 unsigned long nr_running(void)
3199 unsigned long i
, sum
= 0;
3201 for_each_online_cpu(i
)
3202 sum
+= cpu_rq(i
)->nr_running
;
3207 unsigned long nr_uninterruptible(void)
3209 unsigned long i
, sum
= 0;
3211 for_each_possible_cpu(i
)
3212 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3215 * Since we read the counters lockless, it might be slightly
3216 * inaccurate. Do not allow it to go below zero though:
3218 if (unlikely((long)sum
< 0))
3224 unsigned long long nr_context_switches(void)
3227 unsigned long long sum
= 0;
3229 for_each_possible_cpu(i
)
3230 sum
+= cpu_rq(i
)->nr_switches
;
3235 unsigned long nr_iowait(void)
3237 unsigned long i
, sum
= 0;
3239 for_each_possible_cpu(i
)
3240 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3245 unsigned long nr_iowait_cpu(int cpu
)
3247 struct rq
*this = cpu_rq(cpu
);
3248 return atomic_read(&this->nr_iowait
);
3251 unsigned long this_cpu_load(void)
3253 struct rq
*this = this_rq();
3254 return this->cpu_load
[0];
3258 /* Variables and functions for calc_load */
3259 static atomic_long_t calc_load_tasks
;
3260 static unsigned long calc_load_update
;
3261 unsigned long avenrun
[3];
3262 EXPORT_SYMBOL(avenrun
);
3264 static long calc_load_fold_active(struct rq
*this_rq
)
3266 long nr_active
, delta
= 0;
3268 nr_active
= this_rq
->nr_running
;
3269 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3271 if (nr_active
!= this_rq
->calc_load_active
) {
3272 delta
= nr_active
- this_rq
->calc_load_active
;
3273 this_rq
->calc_load_active
= nr_active
;
3279 static unsigned long
3280 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3283 load
+= active
* (FIXED_1
- exp
);
3284 load
+= 1UL << (FSHIFT
- 1);
3285 return load
>> FSHIFT
;
3290 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3292 * When making the ILB scale, we should try to pull this in as well.
3294 static atomic_long_t calc_load_tasks_idle
;
3296 static void calc_load_account_idle(struct rq
*this_rq
)
3300 delta
= calc_load_fold_active(this_rq
);
3302 atomic_long_add(delta
, &calc_load_tasks_idle
);
3305 static long calc_load_fold_idle(void)
3310 * Its got a race, we don't care...
3312 if (atomic_long_read(&calc_load_tasks_idle
))
3313 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3319 * fixed_power_int - compute: x^n, in O(log n) time
3321 * @x: base of the power
3322 * @frac_bits: fractional bits of @x
3323 * @n: power to raise @x to.
3325 * By exploiting the relation between the definition of the natural power
3326 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3327 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3328 * (where: n_i \elem {0, 1}, the binary vector representing n),
3329 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3330 * of course trivially computable in O(log_2 n), the length of our binary
3333 static unsigned long
3334 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3336 unsigned long result
= 1UL << frac_bits
;
3341 result
+= 1UL << (frac_bits
- 1);
3342 result
>>= frac_bits
;
3348 x
+= 1UL << (frac_bits
- 1);
3356 * a1 = a0 * e + a * (1 - e)
3358 * a2 = a1 * e + a * (1 - e)
3359 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3360 * = a0 * e^2 + a * (1 - e) * (1 + e)
3362 * a3 = a2 * e + a * (1 - e)
3363 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3364 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3368 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3369 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3370 * = a0 * e^n + a * (1 - e^n)
3372 * [1] application of the geometric series:
3375 * S_n := \Sum x^i = -------------
3378 static unsigned long
3379 calc_load_n(unsigned long load
, unsigned long exp
,
3380 unsigned long active
, unsigned int n
)
3383 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3387 * NO_HZ can leave us missing all per-cpu ticks calling
3388 * calc_load_account_active(), but since an idle CPU folds its delta into
3389 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3390 * in the pending idle delta if our idle period crossed a load cycle boundary.
3392 * Once we've updated the global active value, we need to apply the exponential
3393 * weights adjusted to the number of cycles missed.
3395 static void calc_global_nohz(unsigned long ticks
)
3397 long delta
, active
, n
;
3399 if (time_before(jiffies
, calc_load_update
))
3403 * If we crossed a calc_load_update boundary, make sure to fold
3404 * any pending idle changes, the respective CPUs might have
3405 * missed the tick driven calc_load_account_active() update
3408 delta
= calc_load_fold_idle();
3410 atomic_long_add(delta
, &calc_load_tasks
);
3413 * If we were idle for multiple load cycles, apply them.
3415 if (ticks
>= LOAD_FREQ
) {
3416 n
= ticks
/ LOAD_FREQ
;
3418 active
= atomic_long_read(&calc_load_tasks
);
3419 active
= active
> 0 ? active
* FIXED_1
: 0;
3421 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3422 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3423 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3425 calc_load_update
+= n
* LOAD_FREQ
;
3429 * Its possible the remainder of the above division also crosses
3430 * a LOAD_FREQ period, the regular check in calc_global_load()
3431 * which comes after this will take care of that.
3433 * Consider us being 11 ticks before a cycle completion, and us
3434 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3435 * age us 4 cycles, and the test in calc_global_load() will
3436 * pick up the final one.
3440 static void calc_load_account_idle(struct rq
*this_rq
)
3444 static inline long calc_load_fold_idle(void)
3449 static void calc_global_nohz(unsigned long ticks
)
3455 * get_avenrun - get the load average array
3456 * @loads: pointer to dest load array
3457 * @offset: offset to add
3458 * @shift: shift count to shift the result left
3460 * These values are estimates at best, so no need for locking.
3462 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3464 loads
[0] = (avenrun
[0] + offset
) << shift
;
3465 loads
[1] = (avenrun
[1] + offset
) << shift
;
3466 loads
[2] = (avenrun
[2] + offset
) << shift
;
3470 * calc_load - update the avenrun load estimates 10 ticks after the
3471 * CPUs have updated calc_load_tasks.
3473 void calc_global_load(unsigned long ticks
)
3477 calc_global_nohz(ticks
);
3479 if (time_before(jiffies
, calc_load_update
+ 10))
3482 active
= atomic_long_read(&calc_load_tasks
);
3483 active
= active
> 0 ? active
* FIXED_1
: 0;
3485 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3486 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3487 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3489 calc_load_update
+= LOAD_FREQ
;
3493 * Called from update_cpu_load() to periodically update this CPU's
3496 static void calc_load_account_active(struct rq
*this_rq
)
3500 if (time_before(jiffies
, this_rq
->calc_load_update
))
3503 delta
= calc_load_fold_active(this_rq
);
3504 delta
+= calc_load_fold_idle();
3506 atomic_long_add(delta
, &calc_load_tasks
);
3508 this_rq
->calc_load_update
+= LOAD_FREQ
;
3512 * The exact cpuload at various idx values, calculated at every tick would be
3513 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3515 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3516 * on nth tick when cpu may be busy, then we have:
3517 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3518 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3520 * decay_load_missed() below does efficient calculation of
3521 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3522 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3524 * The calculation is approximated on a 128 point scale.
3525 * degrade_zero_ticks is the number of ticks after which load at any
3526 * particular idx is approximated to be zero.
3527 * degrade_factor is a precomputed table, a row for each load idx.
3528 * Each column corresponds to degradation factor for a power of two ticks,
3529 * based on 128 point scale.
3531 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3532 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3534 * With this power of 2 load factors, we can degrade the load n times
3535 * by looking at 1 bits in n and doing as many mult/shift instead of
3536 * n mult/shifts needed by the exact degradation.
3538 #define DEGRADE_SHIFT 7
3539 static const unsigned char
3540 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3541 static const unsigned char
3542 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3543 {0, 0, 0, 0, 0, 0, 0, 0},
3544 {64, 32, 8, 0, 0, 0, 0, 0},
3545 {96, 72, 40, 12, 1, 0, 0},
3546 {112, 98, 75, 43, 15, 1, 0},
3547 {120, 112, 98, 76, 45, 16, 2} };
3550 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3551 * would be when CPU is idle and so we just decay the old load without
3552 * adding any new load.
3554 static unsigned long
3555 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3559 if (!missed_updates
)
3562 if (missed_updates
>= degrade_zero_ticks
[idx
])
3566 return load
>> missed_updates
;
3568 while (missed_updates
) {
3569 if (missed_updates
% 2)
3570 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3572 missed_updates
>>= 1;
3579 * Update rq->cpu_load[] statistics. This function is usually called every
3580 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3581 * every tick. We fix it up based on jiffies.
3583 static void update_cpu_load(struct rq
*this_rq
)
3585 unsigned long this_load
= this_rq
->load
.weight
;
3586 unsigned long curr_jiffies
= jiffies
;
3587 unsigned long pending_updates
;
3590 this_rq
->nr_load_updates
++;
3592 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3593 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3596 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3597 this_rq
->last_load_update_tick
= curr_jiffies
;
3599 /* Update our load: */
3600 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3601 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3602 unsigned long old_load
, new_load
;
3604 /* scale is effectively 1 << i now, and >> i divides by scale */
3606 old_load
= this_rq
->cpu_load
[i
];
3607 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3608 new_load
= this_load
;
3610 * Round up the averaging division if load is increasing. This
3611 * prevents us from getting stuck on 9 if the load is 10, for
3614 if (new_load
> old_load
)
3615 new_load
+= scale
- 1;
3617 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3620 sched_avg_update(this_rq
);
3623 static void update_cpu_load_active(struct rq
*this_rq
)
3625 update_cpu_load(this_rq
);
3627 calc_load_account_active(this_rq
);
3633 * sched_exec - execve() is a valuable balancing opportunity, because at
3634 * this point the task has the smallest effective memory and cache footprint.
3636 void sched_exec(void)
3638 struct task_struct
*p
= current
;
3639 unsigned long flags
;
3642 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3643 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3644 if (dest_cpu
== smp_processor_id())
3647 if (likely(cpu_active(dest_cpu
))) {
3648 struct migration_arg arg
= { p
, dest_cpu
};
3650 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3651 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3655 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3660 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3662 EXPORT_PER_CPU_SYMBOL(kstat
);
3665 * Return any ns on the sched_clock that have not yet been accounted in
3666 * @p in case that task is currently running.
3668 * Called with task_rq_lock() held on @rq.
3670 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3674 if (task_current(rq
, p
)) {
3675 update_rq_clock(rq
);
3676 ns
= rq
->clock_task
- p
->se
.exec_start
;
3684 unsigned long long task_delta_exec(struct task_struct
*p
)
3686 unsigned long flags
;
3690 rq
= task_rq_lock(p
, &flags
);
3691 ns
= do_task_delta_exec(p
, rq
);
3692 task_rq_unlock(rq
, p
, &flags
);
3698 * Return accounted runtime for the task.
3699 * In case the task is currently running, return the runtime plus current's
3700 * pending runtime that have not been accounted yet.
3702 unsigned long long task_sched_runtime(struct task_struct
*p
)
3704 unsigned long flags
;
3708 rq
= task_rq_lock(p
, &flags
);
3709 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3710 task_rq_unlock(rq
, p
, &flags
);
3716 * Return sum_exec_runtime for the thread group.
3717 * In case the task is currently running, return the sum plus current's
3718 * pending runtime that have not been accounted yet.
3720 * Note that the thread group might have other running tasks as well,
3721 * so the return value not includes other pending runtime that other
3722 * running tasks might have.
3724 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3726 struct task_cputime totals
;
3727 unsigned long flags
;
3731 rq
= task_rq_lock(p
, &flags
);
3732 thread_group_cputime(p
, &totals
);
3733 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3734 task_rq_unlock(rq
, p
, &flags
);
3740 * Account user cpu time to a process.
3741 * @p: the process that the cpu time gets accounted to
3742 * @cputime: the cpu time spent in user space since the last update
3743 * @cputime_scaled: cputime scaled by cpu frequency
3745 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3746 cputime_t cputime_scaled
)
3748 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3751 /* Add user time to process. */
3752 p
->utime
= cputime_add(p
->utime
, cputime
);
3753 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3754 account_group_user_time(p
, cputime
);
3756 /* Add user time to cpustat. */
3757 tmp
= cputime_to_cputime64(cputime
);
3758 if (TASK_NICE(p
) > 0)
3759 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3761 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3763 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3764 /* Account for user time used */
3765 acct_update_integrals(p
);
3769 * Account guest cpu time to a process.
3770 * @p: the process that the cpu time gets accounted to
3771 * @cputime: the cpu time spent in virtual machine since the last update
3772 * @cputime_scaled: cputime scaled by cpu frequency
3774 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3775 cputime_t cputime_scaled
)
3778 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3780 tmp
= cputime_to_cputime64(cputime
);
3782 /* Add guest time to process. */
3783 p
->utime
= cputime_add(p
->utime
, cputime
);
3784 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3785 account_group_user_time(p
, cputime
);
3786 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3788 /* Add guest time to cpustat. */
3789 if (TASK_NICE(p
) > 0) {
3790 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3791 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3793 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3794 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3799 * Account system cpu time to a process and desired cpustat field
3800 * @p: the process that the cpu time gets accounted to
3801 * @cputime: the cpu time spent in kernel space since the last update
3802 * @cputime_scaled: cputime scaled by cpu frequency
3803 * @target_cputime64: pointer to cpustat field that has to be updated
3806 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
3807 cputime_t cputime_scaled
, cputime64_t
*target_cputime64
)
3809 cputime64_t tmp
= cputime_to_cputime64(cputime
);
3811 /* Add system time to process. */
3812 p
->stime
= cputime_add(p
->stime
, cputime
);
3813 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3814 account_group_system_time(p
, cputime
);
3816 /* Add system time to cpustat. */
3817 *target_cputime64
= cputime64_add(*target_cputime64
, tmp
);
3818 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3820 /* Account for system time used */
3821 acct_update_integrals(p
);
3825 * Account system cpu time to a process.
3826 * @p: the process that the cpu time gets accounted to
3827 * @hardirq_offset: the offset to subtract from hardirq_count()
3828 * @cputime: the cpu time spent in kernel space since the last update
3829 * @cputime_scaled: cputime scaled by cpu frequency
3831 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3832 cputime_t cputime
, cputime_t cputime_scaled
)
3834 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3835 cputime64_t
*target_cputime64
;
3837 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3838 account_guest_time(p
, cputime
, cputime_scaled
);
3842 if (hardirq_count() - hardirq_offset
)
3843 target_cputime64
= &cpustat
->irq
;
3844 else if (in_serving_softirq())
3845 target_cputime64
= &cpustat
->softirq
;
3847 target_cputime64
= &cpustat
->system
;
3849 __account_system_time(p
, cputime
, cputime_scaled
, target_cputime64
);
3853 * Account for involuntary wait time.
3854 * @cputime: the cpu time spent in involuntary wait
3856 void account_steal_time(cputime_t cputime
)
3858 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3859 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3861 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3865 * Account for idle time.
3866 * @cputime: the cpu time spent in idle wait
3868 void account_idle_time(cputime_t cputime
)
3870 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3871 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3872 struct rq
*rq
= this_rq();
3874 if (atomic_read(&rq
->nr_iowait
) > 0)
3875 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3877 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3880 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3882 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3884 * Account a tick to a process and cpustat
3885 * @p: the process that the cpu time gets accounted to
3886 * @user_tick: is the tick from userspace
3887 * @rq: the pointer to rq
3889 * Tick demultiplexing follows the order
3890 * - pending hardirq update
3891 * - pending softirq update
3895 * - check for guest_time
3896 * - else account as system_time
3898 * Check for hardirq is done both for system and user time as there is
3899 * no timer going off while we are on hardirq and hence we may never get an
3900 * opportunity to update it solely in system time.
3901 * p->stime and friends are only updated on system time and not on irq
3902 * softirq as those do not count in task exec_runtime any more.
3904 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3907 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3908 cputime64_t tmp
= cputime_to_cputime64(cputime_one_jiffy
);
3909 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3911 if (irqtime_account_hi_update()) {
3912 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3913 } else if (irqtime_account_si_update()) {
3914 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3915 } else if (this_cpu_ksoftirqd() == p
) {
3917 * ksoftirqd time do not get accounted in cpu_softirq_time.
3918 * So, we have to handle it separately here.
3919 * Also, p->stime needs to be updated for ksoftirqd.
3921 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3923 } else if (user_tick
) {
3924 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3925 } else if (p
== rq
->idle
) {
3926 account_idle_time(cputime_one_jiffy
);
3927 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
3928 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3930 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3935 static void irqtime_account_idle_ticks(int ticks
)
3938 struct rq
*rq
= this_rq();
3940 for (i
= 0; i
< ticks
; i
++)
3941 irqtime_account_process_tick(current
, 0, rq
);
3943 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3944 static void irqtime_account_idle_ticks(int ticks
) {}
3945 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3947 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3950 * Account a single tick of cpu time.
3951 * @p: the process that the cpu time gets accounted to
3952 * @user_tick: indicates if the tick is a user or a system tick
3954 void account_process_tick(struct task_struct
*p
, int user_tick
)
3956 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3957 struct rq
*rq
= this_rq();
3959 if (sched_clock_irqtime
) {
3960 irqtime_account_process_tick(p
, user_tick
, rq
);
3965 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3966 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3967 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3970 account_idle_time(cputime_one_jiffy
);
3974 * Account multiple ticks of steal time.
3975 * @p: the process from which the cpu time has been stolen
3976 * @ticks: number of stolen ticks
3978 void account_steal_ticks(unsigned long ticks
)
3980 account_steal_time(jiffies_to_cputime(ticks
));
3984 * Account multiple ticks of idle time.
3985 * @ticks: number of stolen ticks
3987 void account_idle_ticks(unsigned long ticks
)
3990 if (sched_clock_irqtime
) {
3991 irqtime_account_idle_ticks(ticks
);
3995 account_idle_time(jiffies_to_cputime(ticks
));
4001 * Use precise platform statistics if available:
4003 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4004 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4010 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4012 struct task_cputime cputime
;
4014 thread_group_cputime(p
, &cputime
);
4016 *ut
= cputime
.utime
;
4017 *st
= cputime
.stime
;
4021 #ifndef nsecs_to_cputime
4022 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4025 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4027 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
4030 * Use CFS's precise accounting:
4032 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
4038 do_div(temp
, total
);
4039 utime
= (cputime_t
)temp
;
4044 * Compare with previous values, to keep monotonicity:
4046 p
->prev_utime
= max(p
->prev_utime
, utime
);
4047 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
4049 *ut
= p
->prev_utime
;
4050 *st
= p
->prev_stime
;
4054 * Must be called with siglock held.
4056 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
4058 struct signal_struct
*sig
= p
->signal
;
4059 struct task_cputime cputime
;
4060 cputime_t rtime
, utime
, total
;
4062 thread_group_cputime(p
, &cputime
);
4064 total
= cputime_add(cputime
.utime
, cputime
.stime
);
4065 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
4070 temp
*= cputime
.utime
;
4071 do_div(temp
, total
);
4072 utime
= (cputime_t
)temp
;
4076 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
4077 sig
->prev_stime
= max(sig
->prev_stime
,
4078 cputime_sub(rtime
, sig
->prev_utime
));
4080 *ut
= sig
->prev_utime
;
4081 *st
= sig
->prev_stime
;
4086 * This function gets called by the timer code, with HZ frequency.
4087 * We call it with interrupts disabled.
4089 void scheduler_tick(void)
4091 int cpu
= smp_processor_id();
4092 struct rq
*rq
= cpu_rq(cpu
);
4093 struct task_struct
*curr
= rq
->curr
;
4097 raw_spin_lock(&rq
->lock
);
4098 update_rq_clock(rq
);
4099 update_cpu_load_active(rq
);
4100 curr
->sched_class
->task_tick(rq
, curr
, 0);
4101 raw_spin_unlock(&rq
->lock
);
4103 perf_event_task_tick();
4106 rq
->idle_at_tick
= idle_cpu(cpu
);
4107 trigger_load_balance(rq
, cpu
);
4111 notrace
unsigned long get_parent_ip(unsigned long addr
)
4113 if (in_lock_functions(addr
)) {
4114 addr
= CALLER_ADDR2
;
4115 if (in_lock_functions(addr
))
4116 addr
= CALLER_ADDR3
;
4121 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4122 defined(CONFIG_PREEMPT_TRACER))
4124 void __kprobes
add_preempt_count(int val
)
4126 #ifdef CONFIG_DEBUG_PREEMPT
4130 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4133 preempt_count() += val
;
4134 #ifdef CONFIG_DEBUG_PREEMPT
4136 * Spinlock count overflowing soon?
4138 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4141 if (preempt_count() == val
)
4142 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4144 EXPORT_SYMBOL(add_preempt_count
);
4146 void __kprobes
sub_preempt_count(int val
)
4148 #ifdef CONFIG_DEBUG_PREEMPT
4152 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4155 * Is the spinlock portion underflowing?
4157 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4158 !(preempt_count() & PREEMPT_MASK
)))
4162 if (preempt_count() == val
)
4163 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4164 preempt_count() -= val
;
4166 EXPORT_SYMBOL(sub_preempt_count
);
4171 * Print scheduling while atomic bug:
4173 static noinline
void __schedule_bug(struct task_struct
*prev
)
4175 struct pt_regs
*regs
= get_irq_regs();
4177 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4178 prev
->comm
, prev
->pid
, preempt_count());
4180 debug_show_held_locks(prev
);
4182 if (irqs_disabled())
4183 print_irqtrace_events(prev
);
4192 * Various schedule()-time debugging checks and statistics:
4194 static inline void schedule_debug(struct task_struct
*prev
)
4197 * Test if we are atomic. Since do_exit() needs to call into
4198 * schedule() atomically, we ignore that path for now.
4199 * Otherwise, whine if we are scheduling when we should not be.
4201 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4202 __schedule_bug(prev
);
4204 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4206 schedstat_inc(this_rq(), sched_count
);
4209 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4211 if (prev
->on_rq
|| rq
->skip_clock_update
< 0)
4212 update_rq_clock(rq
);
4213 prev
->sched_class
->put_prev_task(rq
, prev
);
4217 * Pick up the highest-prio task:
4219 static inline struct task_struct
*
4220 pick_next_task(struct rq
*rq
)
4222 const struct sched_class
*class;
4223 struct task_struct
*p
;
4226 * Optimization: we know that if all tasks are in
4227 * the fair class we can call that function directly:
4229 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4230 p
= fair_sched_class
.pick_next_task(rq
);
4235 for_each_class(class) {
4236 p
= class->pick_next_task(rq
);
4241 BUG(); /* the idle class will always have a runnable task */
4245 * __schedule() is the main scheduler function.
4247 static void __sched
__schedule(void)
4249 struct task_struct
*prev
, *next
;
4250 unsigned long *switch_count
;
4256 cpu
= smp_processor_id();
4258 rcu_note_context_switch(cpu
);
4261 schedule_debug(prev
);
4263 if (sched_feat(HRTICK
))
4266 raw_spin_lock_irq(&rq
->lock
);
4268 switch_count
= &prev
->nivcsw
;
4269 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4270 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
4271 prev
->state
= TASK_RUNNING
;
4273 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
4277 * If a worker went to sleep, notify and ask workqueue
4278 * whether it wants to wake up a task to maintain
4281 if (prev
->flags
& PF_WQ_WORKER
) {
4282 struct task_struct
*to_wakeup
;
4284 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
4286 try_to_wake_up_local(to_wakeup
);
4289 switch_count
= &prev
->nvcsw
;
4292 pre_schedule(rq
, prev
);
4294 if (unlikely(!rq
->nr_running
))
4295 idle_balance(cpu
, rq
);
4297 put_prev_task(rq
, prev
);
4298 next
= pick_next_task(rq
);
4299 clear_tsk_need_resched(prev
);
4300 rq
->skip_clock_update
= 0;
4302 if (likely(prev
!= next
)) {
4307 context_switch(rq
, prev
, next
); /* unlocks the rq */
4309 * The context switch have flipped the stack from under us
4310 * and restored the local variables which were saved when
4311 * this task called schedule() in the past. prev == current
4312 * is still correct, but it can be moved to another cpu/rq.
4314 cpu
= smp_processor_id();
4317 raw_spin_unlock_irq(&rq
->lock
);
4321 preempt_enable_no_resched();
4326 static inline void sched_submit_work(struct task_struct
*tsk
)
4331 * If we are going to sleep and we have plugged IO queued,
4332 * make sure to submit it to avoid deadlocks.
4334 if (blk_needs_flush_plug(tsk
))
4335 blk_schedule_flush_plug(tsk
);
4338 asmlinkage
void schedule(void)
4340 struct task_struct
*tsk
= current
;
4342 sched_submit_work(tsk
);
4345 EXPORT_SYMBOL(schedule
);
4347 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4349 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
4354 if (lock
->owner
!= owner
)
4358 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4359 * lock->owner still matches owner, if that fails, owner might
4360 * point to free()d memory, if it still matches, the rcu_read_lock()
4361 * ensures the memory stays valid.
4365 ret
= owner
->on_cpu
;
4373 * Look out! "owner" is an entirely speculative pointer
4374 * access and not reliable.
4376 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
4378 if (!sched_feat(OWNER_SPIN
))
4381 while (owner_running(lock
, owner
)) {
4385 arch_mutex_cpu_relax();
4389 * If the owner changed to another task there is likely
4390 * heavy contention, stop spinning.
4399 #ifdef CONFIG_PREEMPT
4401 * this is the entry point to schedule() from in-kernel preemption
4402 * off of preempt_enable. Kernel preemptions off return from interrupt
4403 * occur there and call schedule directly.
4405 asmlinkage
void __sched notrace
preempt_schedule(void)
4407 struct thread_info
*ti
= current_thread_info();
4410 * If there is a non-zero preempt_count or interrupts are disabled,
4411 * we do not want to preempt the current task. Just return..
4413 if (likely(ti
->preempt_count
|| irqs_disabled()))
4417 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4419 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4422 * Check again in case we missed a preemption opportunity
4423 * between schedule and now.
4426 } while (need_resched());
4428 EXPORT_SYMBOL(preempt_schedule
);
4431 * this is the entry point to schedule() from kernel preemption
4432 * off of irq context.
4433 * Note, that this is called and return with irqs disabled. This will
4434 * protect us against recursive calling from irq.
4436 asmlinkage
void __sched
preempt_schedule_irq(void)
4438 struct thread_info
*ti
= current_thread_info();
4440 /* Catch callers which need to be fixed */
4441 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4444 add_preempt_count(PREEMPT_ACTIVE
);
4447 local_irq_disable();
4448 sub_preempt_count(PREEMPT_ACTIVE
);
4451 * Check again in case we missed a preemption opportunity
4452 * between schedule and now.
4455 } while (need_resched());
4458 #endif /* CONFIG_PREEMPT */
4460 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4463 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4465 EXPORT_SYMBOL(default_wake_function
);
4468 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4469 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4470 * number) then we wake all the non-exclusive tasks and one exclusive task.
4472 * There are circumstances in which we can try to wake a task which has already
4473 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4474 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4476 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4477 int nr_exclusive
, int wake_flags
, void *key
)
4479 wait_queue_t
*curr
, *next
;
4481 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4482 unsigned flags
= curr
->flags
;
4484 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4485 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4491 * __wake_up - wake up threads blocked on a waitqueue.
4493 * @mode: which threads
4494 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4495 * @key: is directly passed to the wakeup function
4497 * It may be assumed that this function implies a write memory barrier before
4498 * changing the task state if and only if any tasks are woken up.
4500 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4501 int nr_exclusive
, void *key
)
4503 unsigned long flags
;
4505 spin_lock_irqsave(&q
->lock
, flags
);
4506 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4507 spin_unlock_irqrestore(&q
->lock
, flags
);
4509 EXPORT_SYMBOL(__wake_up
);
4512 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4514 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4516 __wake_up_common(q
, mode
, 1, 0, NULL
);
4518 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4520 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4522 __wake_up_common(q
, mode
, 1, 0, key
);
4524 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
4527 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4529 * @mode: which threads
4530 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4531 * @key: opaque value to be passed to wakeup targets
4533 * The sync wakeup differs that the waker knows that it will schedule
4534 * away soon, so while the target thread will be woken up, it will not
4535 * be migrated to another CPU - ie. the two threads are 'synchronized'
4536 * with each other. This can prevent needless bouncing between CPUs.
4538 * On UP it can prevent extra preemption.
4540 * It may be assumed that this function implies a write memory barrier before
4541 * changing the task state if and only if any tasks are woken up.
4543 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4544 int nr_exclusive
, void *key
)
4546 unsigned long flags
;
4547 int wake_flags
= WF_SYNC
;
4552 if (unlikely(!nr_exclusive
))
4555 spin_lock_irqsave(&q
->lock
, flags
);
4556 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4557 spin_unlock_irqrestore(&q
->lock
, flags
);
4559 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4562 * __wake_up_sync - see __wake_up_sync_key()
4564 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4566 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4568 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4571 * complete: - signals a single thread waiting on this completion
4572 * @x: holds the state of this particular completion
4574 * This will wake up a single thread waiting on this completion. Threads will be
4575 * awakened in the same order in which they were queued.
4577 * See also complete_all(), wait_for_completion() and related routines.
4579 * It may be assumed that this function implies a write memory barrier before
4580 * changing the task state if and only if any tasks are woken up.
4582 void complete(struct completion
*x
)
4584 unsigned long flags
;
4586 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4588 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4589 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4591 EXPORT_SYMBOL(complete
);
4594 * complete_all: - signals all threads waiting on this completion
4595 * @x: holds the state of this particular completion
4597 * This will wake up all threads waiting on this particular completion event.
4599 * It may be assumed that this function implies a write memory barrier before
4600 * changing the task state if and only if any tasks are woken up.
4602 void complete_all(struct completion
*x
)
4604 unsigned long flags
;
4606 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4607 x
->done
+= UINT_MAX
/2;
4608 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4609 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4611 EXPORT_SYMBOL(complete_all
);
4613 static inline long __sched
4614 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4617 DECLARE_WAITQUEUE(wait
, current
);
4619 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4621 if (signal_pending_state(state
, current
)) {
4622 timeout
= -ERESTARTSYS
;
4625 __set_current_state(state
);
4626 spin_unlock_irq(&x
->wait
.lock
);
4627 timeout
= schedule_timeout(timeout
);
4628 spin_lock_irq(&x
->wait
.lock
);
4629 } while (!x
->done
&& timeout
);
4630 __remove_wait_queue(&x
->wait
, &wait
);
4635 return timeout
?: 1;
4639 wait_for_common(struct completion
*x
, long timeout
, int state
)
4643 spin_lock_irq(&x
->wait
.lock
);
4644 timeout
= do_wait_for_common(x
, timeout
, state
);
4645 spin_unlock_irq(&x
->wait
.lock
);
4650 * wait_for_completion: - waits for completion of a task
4651 * @x: holds the state of this particular completion
4653 * This waits to be signaled for completion of a specific task. It is NOT
4654 * interruptible and there is no timeout.
4656 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4657 * and interrupt capability. Also see complete().
4659 void __sched
wait_for_completion(struct completion
*x
)
4661 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4663 EXPORT_SYMBOL(wait_for_completion
);
4666 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4667 * @x: holds the state of this particular completion
4668 * @timeout: timeout value in jiffies
4670 * This waits for either a completion of a specific task to be signaled or for a
4671 * specified timeout to expire. The timeout is in jiffies. It is not
4674 unsigned long __sched
4675 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4677 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4679 EXPORT_SYMBOL(wait_for_completion_timeout
);
4682 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4683 * @x: holds the state of this particular completion
4685 * This waits for completion of a specific task to be signaled. It is
4688 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4690 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4691 if (t
== -ERESTARTSYS
)
4695 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4698 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4699 * @x: holds the state of this particular completion
4700 * @timeout: timeout value in jiffies
4702 * This waits for either a completion of a specific task to be signaled or for a
4703 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4706 wait_for_completion_interruptible_timeout(struct completion
*x
,
4707 unsigned long timeout
)
4709 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4711 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4714 * wait_for_completion_killable: - waits for completion of a task (killable)
4715 * @x: holds the state of this particular completion
4717 * This waits to be signaled for completion of a specific task. It can be
4718 * interrupted by a kill signal.
4720 int __sched
wait_for_completion_killable(struct completion
*x
)
4722 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4723 if (t
== -ERESTARTSYS
)
4727 EXPORT_SYMBOL(wait_for_completion_killable
);
4730 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4731 * @x: holds the state of this particular completion
4732 * @timeout: timeout value in jiffies
4734 * This waits for either a completion of a specific task to be
4735 * signaled or for a specified timeout to expire. It can be
4736 * interrupted by a kill signal. The timeout is in jiffies.
4739 wait_for_completion_killable_timeout(struct completion
*x
,
4740 unsigned long timeout
)
4742 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4744 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4747 * try_wait_for_completion - try to decrement a completion without blocking
4748 * @x: completion structure
4750 * Returns: 0 if a decrement cannot be done without blocking
4751 * 1 if a decrement succeeded.
4753 * If a completion is being used as a counting completion,
4754 * attempt to decrement the counter without blocking. This
4755 * enables us to avoid waiting if the resource the completion
4756 * is protecting is not available.
4758 bool try_wait_for_completion(struct completion
*x
)
4760 unsigned long flags
;
4763 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4768 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4771 EXPORT_SYMBOL(try_wait_for_completion
);
4774 * completion_done - Test to see if a completion has any waiters
4775 * @x: completion structure
4777 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4778 * 1 if there are no waiters.
4781 bool completion_done(struct completion
*x
)
4783 unsigned long flags
;
4786 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4789 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4792 EXPORT_SYMBOL(completion_done
);
4795 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4797 unsigned long flags
;
4800 init_waitqueue_entry(&wait
, current
);
4802 __set_current_state(state
);
4804 spin_lock_irqsave(&q
->lock
, flags
);
4805 __add_wait_queue(q
, &wait
);
4806 spin_unlock(&q
->lock
);
4807 timeout
= schedule_timeout(timeout
);
4808 spin_lock_irq(&q
->lock
);
4809 __remove_wait_queue(q
, &wait
);
4810 spin_unlock_irqrestore(&q
->lock
, flags
);
4815 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4817 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4819 EXPORT_SYMBOL(interruptible_sleep_on
);
4822 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4824 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4826 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4828 void __sched
sleep_on(wait_queue_head_t
*q
)
4830 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4832 EXPORT_SYMBOL(sleep_on
);
4834 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4836 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4838 EXPORT_SYMBOL(sleep_on_timeout
);
4840 #ifdef CONFIG_RT_MUTEXES
4843 * rt_mutex_setprio - set the current priority of a task
4845 * @prio: prio value (kernel-internal form)
4847 * This function changes the 'effective' priority of a task. It does
4848 * not touch ->normal_prio like __setscheduler().
4850 * Used by the rt_mutex code to implement priority inheritance logic.
4852 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4854 int oldprio
, on_rq
, running
;
4856 const struct sched_class
*prev_class
;
4858 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4860 rq
= __task_rq_lock(p
);
4862 trace_sched_pi_setprio(p
, prio
);
4864 prev_class
= p
->sched_class
;
4866 running
= task_current(rq
, p
);
4868 dequeue_task(rq
, p
, 0);
4870 p
->sched_class
->put_prev_task(rq
, p
);
4873 p
->sched_class
= &rt_sched_class
;
4875 p
->sched_class
= &fair_sched_class
;
4880 p
->sched_class
->set_curr_task(rq
);
4882 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4884 check_class_changed(rq
, p
, prev_class
, oldprio
);
4885 __task_rq_unlock(rq
);
4890 void set_user_nice(struct task_struct
*p
, long nice
)
4892 int old_prio
, delta
, on_rq
;
4893 unsigned long flags
;
4896 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4899 * We have to be careful, if called from sys_setpriority(),
4900 * the task might be in the middle of scheduling on another CPU.
4902 rq
= task_rq_lock(p
, &flags
);
4904 * The RT priorities are set via sched_setscheduler(), but we still
4905 * allow the 'normal' nice value to be set - but as expected
4906 * it wont have any effect on scheduling until the task is
4907 * SCHED_FIFO/SCHED_RR:
4909 if (task_has_rt_policy(p
)) {
4910 p
->static_prio
= NICE_TO_PRIO(nice
);
4915 dequeue_task(rq
, p
, 0);
4917 p
->static_prio
= NICE_TO_PRIO(nice
);
4920 p
->prio
= effective_prio(p
);
4921 delta
= p
->prio
- old_prio
;
4924 enqueue_task(rq
, p
, 0);
4926 * If the task increased its priority or is running and
4927 * lowered its priority, then reschedule its CPU:
4929 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4930 resched_task(rq
->curr
);
4933 task_rq_unlock(rq
, p
, &flags
);
4935 EXPORT_SYMBOL(set_user_nice
);
4938 * can_nice - check if a task can reduce its nice value
4942 int can_nice(const struct task_struct
*p
, const int nice
)
4944 /* convert nice value [19,-20] to rlimit style value [1,40] */
4945 int nice_rlim
= 20 - nice
;
4947 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4948 capable(CAP_SYS_NICE
));
4951 #ifdef __ARCH_WANT_SYS_NICE
4954 * sys_nice - change the priority of the current process.
4955 * @increment: priority increment
4957 * sys_setpriority is a more generic, but much slower function that
4958 * does similar things.
4960 SYSCALL_DEFINE1(nice
, int, increment
)
4965 * Setpriority might change our priority at the same moment.
4966 * We don't have to worry. Conceptually one call occurs first
4967 * and we have a single winner.
4969 if (increment
< -40)
4974 nice
= TASK_NICE(current
) + increment
;
4980 if (increment
< 0 && !can_nice(current
, nice
))
4983 retval
= security_task_setnice(current
, nice
);
4987 set_user_nice(current
, nice
);
4994 * task_prio - return the priority value of a given task.
4995 * @p: the task in question.
4997 * This is the priority value as seen by users in /proc.
4998 * RT tasks are offset by -200. Normal tasks are centered
4999 * around 0, value goes from -16 to +15.
5001 int task_prio(const struct task_struct
*p
)
5003 return p
->prio
- MAX_RT_PRIO
;
5007 * task_nice - return the nice value of a given task.
5008 * @p: the task in question.
5010 int task_nice(const struct task_struct
*p
)
5012 return TASK_NICE(p
);
5014 EXPORT_SYMBOL(task_nice
);
5017 * idle_cpu - is a given cpu idle currently?
5018 * @cpu: the processor in question.
5020 int idle_cpu(int cpu
)
5022 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5026 * idle_task - return the idle task for a given cpu.
5027 * @cpu: the processor in question.
5029 struct task_struct
*idle_task(int cpu
)
5031 return cpu_rq(cpu
)->idle
;
5035 * find_process_by_pid - find a process with a matching PID value.
5036 * @pid: the pid in question.
5038 static struct task_struct
*find_process_by_pid(pid_t pid
)
5040 return pid
? find_task_by_vpid(pid
) : current
;
5043 /* Actually do priority change: must hold rq lock. */
5045 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5048 p
->rt_priority
= prio
;
5049 p
->normal_prio
= normal_prio(p
);
5050 /* we are holding p->pi_lock already */
5051 p
->prio
= rt_mutex_getprio(p
);
5052 if (rt_prio(p
->prio
))
5053 p
->sched_class
= &rt_sched_class
;
5055 p
->sched_class
= &fair_sched_class
;
5060 * check the target process has a UID that matches the current process's
5062 static bool check_same_owner(struct task_struct
*p
)
5064 const struct cred
*cred
= current_cred(), *pcred
;
5068 pcred
= __task_cred(p
);
5069 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
5070 match
= (cred
->euid
== pcred
->euid
||
5071 cred
->euid
== pcred
->uid
);
5078 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5079 const struct sched_param
*param
, bool user
)
5081 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5082 unsigned long flags
;
5083 const struct sched_class
*prev_class
;
5087 /* may grab non-irq protected spin_locks */
5088 BUG_ON(in_interrupt());
5090 /* double check policy once rq lock held */
5092 reset_on_fork
= p
->sched_reset_on_fork
;
5093 policy
= oldpolicy
= p
->policy
;
5095 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
5096 policy
&= ~SCHED_RESET_ON_FORK
;
5098 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5099 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5100 policy
!= SCHED_IDLE
)
5105 * Valid priorities for SCHED_FIFO and SCHED_RR are
5106 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5107 * SCHED_BATCH and SCHED_IDLE is 0.
5109 if (param
->sched_priority
< 0 ||
5110 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5111 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5113 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5117 * Allow unprivileged RT tasks to decrease priority:
5119 if (user
&& !capable(CAP_SYS_NICE
)) {
5120 if (rt_policy(policy
)) {
5121 unsigned long rlim_rtprio
=
5122 task_rlimit(p
, RLIMIT_RTPRIO
);
5124 /* can't set/change the rt policy */
5125 if (policy
!= p
->policy
&& !rlim_rtprio
)
5128 /* can't increase priority */
5129 if (param
->sched_priority
> p
->rt_priority
&&
5130 param
->sched_priority
> rlim_rtprio
)
5135 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5136 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5138 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
5139 if (!can_nice(p
, TASK_NICE(p
)))
5143 /* can't change other user's priorities */
5144 if (!check_same_owner(p
))
5147 /* Normal users shall not reset the sched_reset_on_fork flag */
5148 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5153 retval
= security_task_setscheduler(p
);
5159 * make sure no PI-waiters arrive (or leave) while we are
5160 * changing the priority of the task:
5162 * To be able to change p->policy safely, the appropriate
5163 * runqueue lock must be held.
5165 rq
= task_rq_lock(p
, &flags
);
5168 * Changing the policy of the stop threads its a very bad idea
5170 if (p
== rq
->stop
) {
5171 task_rq_unlock(rq
, p
, &flags
);
5176 * If not changing anything there's no need to proceed further:
5178 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
5179 param
->sched_priority
== p
->rt_priority
))) {
5181 __task_rq_unlock(rq
);
5182 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5186 #ifdef CONFIG_RT_GROUP_SCHED
5189 * Do not allow realtime tasks into groups that have no runtime
5192 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5193 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5194 !task_group_is_autogroup(task_group(p
))) {
5195 task_rq_unlock(rq
, p
, &flags
);
5201 /* recheck policy now with rq lock held */
5202 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5203 policy
= oldpolicy
= -1;
5204 task_rq_unlock(rq
, p
, &flags
);
5208 running
= task_current(rq
, p
);
5210 deactivate_task(rq
, p
, 0);
5212 p
->sched_class
->put_prev_task(rq
, p
);
5214 p
->sched_reset_on_fork
= reset_on_fork
;
5217 prev_class
= p
->sched_class
;
5218 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5221 p
->sched_class
->set_curr_task(rq
);
5223 activate_task(rq
, p
, 0);
5225 check_class_changed(rq
, p
, prev_class
, oldprio
);
5226 task_rq_unlock(rq
, p
, &flags
);
5228 rt_mutex_adjust_pi(p
);
5234 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5235 * @p: the task in question.
5236 * @policy: new policy.
5237 * @param: structure containing the new RT priority.
5239 * NOTE that the task may be already dead.
5241 int sched_setscheduler(struct task_struct
*p
, int policy
,
5242 const struct sched_param
*param
)
5244 return __sched_setscheduler(p
, policy
, param
, true);
5246 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5249 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5250 * @p: the task in question.
5251 * @policy: new policy.
5252 * @param: structure containing the new RT priority.
5254 * Just like sched_setscheduler, only don't bother checking if the
5255 * current context has permission. For example, this is needed in
5256 * stop_machine(): we create temporary high priority worker threads,
5257 * but our caller might not have that capability.
5259 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5260 const struct sched_param
*param
)
5262 return __sched_setscheduler(p
, policy
, param
, false);
5266 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5268 struct sched_param lparam
;
5269 struct task_struct
*p
;
5272 if (!param
|| pid
< 0)
5274 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5279 p
= find_process_by_pid(pid
);
5281 retval
= sched_setscheduler(p
, policy
, &lparam
);
5288 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5289 * @pid: the pid in question.
5290 * @policy: new policy.
5291 * @param: structure containing the new RT priority.
5293 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5294 struct sched_param __user
*, param
)
5296 /* negative values for policy are not valid */
5300 return do_sched_setscheduler(pid
, policy
, param
);
5304 * sys_sched_setparam - set/change the RT priority of a thread
5305 * @pid: the pid in question.
5306 * @param: structure containing the new RT priority.
5308 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5310 return do_sched_setscheduler(pid
, -1, param
);
5314 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5315 * @pid: the pid in question.
5317 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5319 struct task_struct
*p
;
5327 p
= find_process_by_pid(pid
);
5329 retval
= security_task_getscheduler(p
);
5332 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5339 * sys_sched_getparam - get the RT priority of a thread
5340 * @pid: the pid in question.
5341 * @param: structure containing the RT priority.
5343 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5345 struct sched_param lp
;
5346 struct task_struct
*p
;
5349 if (!param
|| pid
< 0)
5353 p
= find_process_by_pid(pid
);
5358 retval
= security_task_getscheduler(p
);
5362 lp
.sched_priority
= p
->rt_priority
;
5366 * This one might sleep, we cannot do it with a spinlock held ...
5368 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5377 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5379 cpumask_var_t cpus_allowed
, new_mask
;
5380 struct task_struct
*p
;
5386 p
= find_process_by_pid(pid
);
5393 /* Prevent p going away */
5397 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5401 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5403 goto out_free_cpus_allowed
;
5406 if (!check_same_owner(p
) && !task_ns_capable(p
, CAP_SYS_NICE
))
5409 retval
= security_task_setscheduler(p
);
5413 cpuset_cpus_allowed(p
, cpus_allowed
);
5414 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5416 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5419 cpuset_cpus_allowed(p
, cpus_allowed
);
5420 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5422 * We must have raced with a concurrent cpuset
5423 * update. Just reset the cpus_allowed to the
5424 * cpuset's cpus_allowed
5426 cpumask_copy(new_mask
, cpus_allowed
);
5431 free_cpumask_var(new_mask
);
5432 out_free_cpus_allowed
:
5433 free_cpumask_var(cpus_allowed
);
5440 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5441 struct cpumask
*new_mask
)
5443 if (len
< cpumask_size())
5444 cpumask_clear(new_mask
);
5445 else if (len
> cpumask_size())
5446 len
= cpumask_size();
5448 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5452 * sys_sched_setaffinity - set the cpu affinity of a process
5453 * @pid: pid of the process
5454 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5455 * @user_mask_ptr: user-space pointer to the new cpu mask
5457 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5458 unsigned long __user
*, user_mask_ptr
)
5460 cpumask_var_t new_mask
;
5463 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5466 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5468 retval
= sched_setaffinity(pid
, new_mask
);
5469 free_cpumask_var(new_mask
);
5473 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5475 struct task_struct
*p
;
5476 unsigned long flags
;
5483 p
= find_process_by_pid(pid
);
5487 retval
= security_task_getscheduler(p
);
5491 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5492 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5493 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5503 * sys_sched_getaffinity - get the cpu affinity of a process
5504 * @pid: pid of the process
5505 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5506 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5508 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5509 unsigned long __user
*, user_mask_ptr
)
5514 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5516 if (len
& (sizeof(unsigned long)-1))
5519 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5522 ret
= sched_getaffinity(pid
, mask
);
5524 size_t retlen
= min_t(size_t, len
, cpumask_size());
5526 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5531 free_cpumask_var(mask
);
5537 * sys_sched_yield - yield the current processor to other threads.
5539 * This function yields the current CPU to other tasks. If there are no
5540 * other threads running on this CPU then this function will return.
5542 SYSCALL_DEFINE0(sched_yield
)
5544 struct rq
*rq
= this_rq_lock();
5546 schedstat_inc(rq
, yld_count
);
5547 current
->sched_class
->yield_task(rq
);
5550 * Since we are going to call schedule() anyway, there's
5551 * no need to preempt or enable interrupts:
5553 __release(rq
->lock
);
5554 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5555 do_raw_spin_unlock(&rq
->lock
);
5556 preempt_enable_no_resched();
5563 static inline int should_resched(void)
5565 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5568 static void __cond_resched(void)
5570 add_preempt_count(PREEMPT_ACTIVE
);
5572 sub_preempt_count(PREEMPT_ACTIVE
);
5575 int __sched
_cond_resched(void)
5577 if (should_resched()) {
5583 EXPORT_SYMBOL(_cond_resched
);
5586 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5587 * call schedule, and on return reacquire the lock.
5589 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5590 * operations here to prevent schedule() from being called twice (once via
5591 * spin_unlock(), once by hand).
5593 int __cond_resched_lock(spinlock_t
*lock
)
5595 int resched
= should_resched();
5598 lockdep_assert_held(lock
);
5600 if (spin_needbreak(lock
) || resched
) {
5611 EXPORT_SYMBOL(__cond_resched_lock
);
5613 int __sched
__cond_resched_softirq(void)
5615 BUG_ON(!in_softirq());
5617 if (should_resched()) {
5625 EXPORT_SYMBOL(__cond_resched_softirq
);
5628 * yield - yield the current processor to other threads.
5630 * This is a shortcut for kernel-space yielding - it marks the
5631 * thread runnable and calls sys_sched_yield().
5633 void __sched
yield(void)
5635 set_current_state(TASK_RUNNING
);
5638 EXPORT_SYMBOL(yield
);
5641 * yield_to - yield the current processor to another thread in
5642 * your thread group, or accelerate that thread toward the
5643 * processor it's on.
5645 * @preempt: whether task preemption is allowed or not
5647 * It's the caller's job to ensure that the target task struct
5648 * can't go away on us before we can do any checks.
5650 * Returns true if we indeed boosted the target task.
5652 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
5654 struct task_struct
*curr
= current
;
5655 struct rq
*rq
, *p_rq
;
5656 unsigned long flags
;
5659 local_irq_save(flags
);
5664 double_rq_lock(rq
, p_rq
);
5665 while (task_rq(p
) != p_rq
) {
5666 double_rq_unlock(rq
, p_rq
);
5670 if (!curr
->sched_class
->yield_to_task
)
5673 if (curr
->sched_class
!= p
->sched_class
)
5676 if (task_running(p_rq
, p
) || p
->state
)
5679 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5681 schedstat_inc(rq
, yld_count
);
5683 * Make p's CPU reschedule; pick_next_entity takes care of
5686 if (preempt
&& rq
!= p_rq
)
5687 resched_task(p_rq
->curr
);
5691 double_rq_unlock(rq
, p_rq
);
5692 local_irq_restore(flags
);
5699 EXPORT_SYMBOL_GPL(yield_to
);
5702 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5703 * that process accounting knows that this is a task in IO wait state.
5705 void __sched
io_schedule(void)
5707 struct rq
*rq
= raw_rq();
5709 delayacct_blkio_start();
5710 atomic_inc(&rq
->nr_iowait
);
5711 blk_flush_plug(current
);
5712 current
->in_iowait
= 1;
5714 current
->in_iowait
= 0;
5715 atomic_dec(&rq
->nr_iowait
);
5716 delayacct_blkio_end();
5718 EXPORT_SYMBOL(io_schedule
);
5720 long __sched
io_schedule_timeout(long timeout
)
5722 struct rq
*rq
= raw_rq();
5725 delayacct_blkio_start();
5726 atomic_inc(&rq
->nr_iowait
);
5727 blk_flush_plug(current
);
5728 current
->in_iowait
= 1;
5729 ret
= schedule_timeout(timeout
);
5730 current
->in_iowait
= 0;
5731 atomic_dec(&rq
->nr_iowait
);
5732 delayacct_blkio_end();
5737 * sys_sched_get_priority_max - return maximum RT priority.
5738 * @policy: scheduling class.
5740 * this syscall returns the maximum rt_priority that can be used
5741 * by a given scheduling class.
5743 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5750 ret
= MAX_USER_RT_PRIO
-1;
5762 * sys_sched_get_priority_min - return minimum RT priority.
5763 * @policy: scheduling class.
5765 * this syscall returns the minimum rt_priority that can be used
5766 * by a given scheduling class.
5768 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5786 * sys_sched_rr_get_interval - return the default timeslice of a process.
5787 * @pid: pid of the process.
5788 * @interval: userspace pointer to the timeslice value.
5790 * this syscall writes the default timeslice value of a given process
5791 * into the user-space timespec buffer. A value of '0' means infinity.
5793 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5794 struct timespec __user
*, interval
)
5796 struct task_struct
*p
;
5797 unsigned int time_slice
;
5798 unsigned long flags
;
5808 p
= find_process_by_pid(pid
);
5812 retval
= security_task_getscheduler(p
);
5816 rq
= task_rq_lock(p
, &flags
);
5817 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5818 task_rq_unlock(rq
, p
, &flags
);
5821 jiffies_to_timespec(time_slice
, &t
);
5822 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5830 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5832 void sched_show_task(struct task_struct
*p
)
5834 unsigned long free
= 0;
5837 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5838 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5839 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5840 #if BITS_PER_LONG == 32
5841 if (state
== TASK_RUNNING
)
5842 printk(KERN_CONT
" running ");
5844 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5846 if (state
== TASK_RUNNING
)
5847 printk(KERN_CONT
" running task ");
5849 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5851 #ifdef CONFIG_DEBUG_STACK_USAGE
5852 free
= stack_not_used(p
);
5854 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5855 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5856 (unsigned long)task_thread_info(p
)->flags
);
5858 show_stack(p
, NULL
);
5861 void show_state_filter(unsigned long state_filter
)
5863 struct task_struct
*g
, *p
;
5865 #if BITS_PER_LONG == 32
5867 " task PC stack pid father\n");
5870 " task PC stack pid father\n");
5872 read_lock(&tasklist_lock
);
5873 do_each_thread(g
, p
) {
5875 * reset the NMI-timeout, listing all files on a slow
5876 * console might take a lot of time:
5878 touch_nmi_watchdog();
5879 if (!state_filter
|| (p
->state
& state_filter
))
5881 } while_each_thread(g
, p
);
5883 touch_all_softlockup_watchdogs();
5885 #ifdef CONFIG_SCHED_DEBUG
5886 sysrq_sched_debug_show();
5888 read_unlock(&tasklist_lock
);
5890 * Only show locks if all tasks are dumped:
5893 debug_show_all_locks();
5896 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5898 idle
->sched_class
= &idle_sched_class
;
5902 * init_idle - set up an idle thread for a given CPU
5903 * @idle: task in question
5904 * @cpu: cpu the idle task belongs to
5906 * NOTE: this function does not set the idle thread's NEED_RESCHED
5907 * flag, to make booting more robust.
5909 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5911 struct rq
*rq
= cpu_rq(cpu
);
5912 unsigned long flags
;
5914 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5917 idle
->state
= TASK_RUNNING
;
5918 idle
->se
.exec_start
= sched_clock();
5920 do_set_cpus_allowed(idle
, cpumask_of(cpu
));
5922 * We're having a chicken and egg problem, even though we are
5923 * holding rq->lock, the cpu isn't yet set to this cpu so the
5924 * lockdep check in task_group() will fail.
5926 * Similar case to sched_fork(). / Alternatively we could
5927 * use task_rq_lock() here and obtain the other rq->lock.
5932 __set_task_cpu(idle
, cpu
);
5935 rq
->curr
= rq
->idle
= idle
;
5936 #if defined(CONFIG_SMP)
5939 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5941 /* Set the preempt count _outside_ the spinlocks! */
5942 task_thread_info(idle
)->preempt_count
= 0;
5945 * The idle tasks have their own, simple scheduling class:
5947 idle
->sched_class
= &idle_sched_class
;
5948 ftrace_graph_init_idle_task(idle
, cpu
);
5952 * In a system that switches off the HZ timer nohz_cpu_mask
5953 * indicates which cpus entered this state. This is used
5954 * in the rcu update to wait only for active cpus. For system
5955 * which do not switch off the HZ timer nohz_cpu_mask should
5956 * always be CPU_BITS_NONE.
5958 cpumask_var_t nohz_cpu_mask
;
5961 * Increase the granularity value when there are more CPUs,
5962 * because with more CPUs the 'effective latency' as visible
5963 * to users decreases. But the relationship is not linear,
5964 * so pick a second-best guess by going with the log2 of the
5967 * This idea comes from the SD scheduler of Con Kolivas:
5969 static int get_update_sysctl_factor(void)
5971 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5972 unsigned int factor
;
5974 switch (sysctl_sched_tunable_scaling
) {
5975 case SCHED_TUNABLESCALING_NONE
:
5978 case SCHED_TUNABLESCALING_LINEAR
:
5981 case SCHED_TUNABLESCALING_LOG
:
5983 factor
= 1 + ilog2(cpus
);
5990 static void update_sysctl(void)
5992 unsigned int factor
= get_update_sysctl_factor();
5994 #define SET_SYSCTL(name) \
5995 (sysctl_##name = (factor) * normalized_sysctl_##name)
5996 SET_SYSCTL(sched_min_granularity
);
5997 SET_SYSCTL(sched_latency
);
5998 SET_SYSCTL(sched_wakeup_granularity
);
6002 static inline void sched_init_granularity(void)
6008 void do_set_cpus_allowed(struct task_struct
*p
, const struct cpumask
*new_mask
)
6010 if (p
->sched_class
&& p
->sched_class
->set_cpus_allowed
)
6011 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6013 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6014 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6019 * This is how migration works:
6021 * 1) we invoke migration_cpu_stop() on the target CPU using
6023 * 2) stopper starts to run (implicitly forcing the migrated thread
6025 * 3) it checks whether the migrated task is still in the wrong runqueue.
6026 * 4) if it's in the wrong runqueue then the migration thread removes
6027 * it and puts it into the right queue.
6028 * 5) stopper completes and stop_one_cpu() returns and the migration
6033 * Change a given task's CPU affinity. Migrate the thread to a
6034 * proper CPU and schedule it away if the CPU it's executing on
6035 * is removed from the allowed bitmask.
6037 * NOTE: the caller must have a valid reference to the task, the
6038 * task must not exit() & deallocate itself prematurely. The
6039 * call is not atomic; no spinlocks may be held.
6041 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6043 unsigned long flags
;
6045 unsigned int dest_cpu
;
6048 rq
= task_rq_lock(p
, &flags
);
6050 if (cpumask_equal(&p
->cpus_allowed
, new_mask
))
6053 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
6058 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
)) {
6063 do_set_cpus_allowed(p
, new_mask
);
6065 /* Can the task run on the task's current CPU? If so, we're done */
6066 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6069 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
6071 struct migration_arg arg
= { p
, dest_cpu
};
6072 /* Need help from migration thread: drop lock and wait. */
6073 task_rq_unlock(rq
, p
, &flags
);
6074 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
6075 tlb_migrate_finish(p
->mm
);
6079 task_rq_unlock(rq
, p
, &flags
);
6083 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6086 * Move (not current) task off this cpu, onto dest cpu. We're doing
6087 * this because either it can't run here any more (set_cpus_allowed()
6088 * away from this CPU, or CPU going down), or because we're
6089 * attempting to rebalance this task on exec (sched_exec).
6091 * So we race with normal scheduler movements, but that's OK, as long
6092 * as the task is no longer on this CPU.
6094 * Returns non-zero if task was successfully migrated.
6096 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6098 struct rq
*rq_dest
, *rq_src
;
6101 if (unlikely(!cpu_active(dest_cpu
)))
6104 rq_src
= cpu_rq(src_cpu
);
6105 rq_dest
= cpu_rq(dest_cpu
);
6107 raw_spin_lock(&p
->pi_lock
);
6108 double_rq_lock(rq_src
, rq_dest
);
6109 /* Already moved. */
6110 if (task_cpu(p
) != src_cpu
)
6112 /* Affinity changed (again). */
6113 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6117 * If we're not on a rq, the next wake-up will ensure we're
6121 deactivate_task(rq_src
, p
, 0);
6122 set_task_cpu(p
, dest_cpu
);
6123 activate_task(rq_dest
, p
, 0);
6124 check_preempt_curr(rq_dest
, p
, 0);
6129 double_rq_unlock(rq_src
, rq_dest
);
6130 raw_spin_unlock(&p
->pi_lock
);
6135 * migration_cpu_stop - this will be executed by a highprio stopper thread
6136 * and performs thread migration by bumping thread off CPU then
6137 * 'pushing' onto another runqueue.
6139 static int migration_cpu_stop(void *data
)
6141 struct migration_arg
*arg
= data
;
6144 * The original target cpu might have gone down and we might
6145 * be on another cpu but it doesn't matter.
6147 local_irq_disable();
6148 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
6153 #ifdef CONFIG_HOTPLUG_CPU
6156 * Ensures that the idle task is using init_mm right before its cpu goes
6159 void idle_task_exit(void)
6161 struct mm_struct
*mm
= current
->active_mm
;
6163 BUG_ON(cpu_online(smp_processor_id()));
6166 switch_mm(mm
, &init_mm
, current
);
6171 * While a dead CPU has no uninterruptible tasks queued at this point,
6172 * it might still have a nonzero ->nr_uninterruptible counter, because
6173 * for performance reasons the counter is not stricly tracking tasks to
6174 * their home CPUs. So we just add the counter to another CPU's counter,
6175 * to keep the global sum constant after CPU-down:
6177 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6179 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
6181 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6182 rq_src
->nr_uninterruptible
= 0;
6186 * remove the tasks which were accounted by rq from calc_load_tasks.
6188 static void calc_global_load_remove(struct rq
*rq
)
6190 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
6191 rq
->calc_load_active
= 0;
6195 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6196 * try_to_wake_up()->select_task_rq().
6198 * Called with rq->lock held even though we'er in stop_machine() and
6199 * there's no concurrency possible, we hold the required locks anyway
6200 * because of lock validation efforts.
6202 static void migrate_tasks(unsigned int dead_cpu
)
6204 struct rq
*rq
= cpu_rq(dead_cpu
);
6205 struct task_struct
*next
, *stop
= rq
->stop
;
6209 * Fudge the rq selection such that the below task selection loop
6210 * doesn't get stuck on the currently eligible stop task.
6212 * We're currently inside stop_machine() and the rq is either stuck
6213 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6214 * either way we should never end up calling schedule() until we're
6221 * There's this thread running, bail when that's the only
6224 if (rq
->nr_running
== 1)
6227 next
= pick_next_task(rq
);
6229 next
->sched_class
->put_prev_task(rq
, next
);
6231 /* Find suitable destination for @next, with force if needed. */
6232 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
6233 raw_spin_unlock(&rq
->lock
);
6235 __migrate_task(next
, dead_cpu
, dest_cpu
);
6237 raw_spin_lock(&rq
->lock
);
6243 #endif /* CONFIG_HOTPLUG_CPU */
6245 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6247 static struct ctl_table sd_ctl_dir
[] = {
6249 .procname
= "sched_domain",
6255 static struct ctl_table sd_ctl_root
[] = {
6257 .procname
= "kernel",
6259 .child
= sd_ctl_dir
,
6264 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6266 struct ctl_table
*entry
=
6267 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6272 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6274 struct ctl_table
*entry
;
6277 * In the intermediate directories, both the child directory and
6278 * procname are dynamically allocated and could fail but the mode
6279 * will always be set. In the lowest directory the names are
6280 * static strings and all have proc handlers.
6282 for (entry
= *tablep
; entry
->mode
; entry
++) {
6284 sd_free_ctl_entry(&entry
->child
);
6285 if (entry
->proc_handler
== NULL
)
6286 kfree(entry
->procname
);
6294 set_table_entry(struct ctl_table
*entry
,
6295 const char *procname
, void *data
, int maxlen
,
6296 mode_t mode
, proc_handler
*proc_handler
)
6298 entry
->procname
= procname
;
6300 entry
->maxlen
= maxlen
;
6302 entry
->proc_handler
= proc_handler
;
6305 static struct ctl_table
*
6306 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6308 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6313 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6314 sizeof(long), 0644, proc_doulongvec_minmax
);
6315 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6316 sizeof(long), 0644, proc_doulongvec_minmax
);
6317 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6318 sizeof(int), 0644, proc_dointvec_minmax
);
6319 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6320 sizeof(int), 0644, proc_dointvec_minmax
);
6321 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6322 sizeof(int), 0644, proc_dointvec_minmax
);
6323 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6324 sizeof(int), 0644, proc_dointvec_minmax
);
6325 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6326 sizeof(int), 0644, proc_dointvec_minmax
);
6327 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6328 sizeof(int), 0644, proc_dointvec_minmax
);
6329 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6330 sizeof(int), 0644, proc_dointvec_minmax
);
6331 set_table_entry(&table
[9], "cache_nice_tries",
6332 &sd
->cache_nice_tries
,
6333 sizeof(int), 0644, proc_dointvec_minmax
);
6334 set_table_entry(&table
[10], "flags", &sd
->flags
,
6335 sizeof(int), 0644, proc_dointvec_minmax
);
6336 set_table_entry(&table
[11], "name", sd
->name
,
6337 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6338 /* &table[12] is terminator */
6343 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6345 struct ctl_table
*entry
, *table
;
6346 struct sched_domain
*sd
;
6347 int domain_num
= 0, i
;
6350 for_each_domain(cpu
, sd
)
6352 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6357 for_each_domain(cpu
, sd
) {
6358 snprintf(buf
, 32, "domain%d", i
);
6359 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6361 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6368 static struct ctl_table_header
*sd_sysctl_header
;
6369 static void register_sched_domain_sysctl(void)
6371 int i
, cpu_num
= num_possible_cpus();
6372 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6375 WARN_ON(sd_ctl_dir
[0].child
);
6376 sd_ctl_dir
[0].child
= entry
;
6381 for_each_possible_cpu(i
) {
6382 snprintf(buf
, 32, "cpu%d", i
);
6383 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6385 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6389 WARN_ON(sd_sysctl_header
);
6390 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6393 /* may be called multiple times per register */
6394 static void unregister_sched_domain_sysctl(void)
6396 if (sd_sysctl_header
)
6397 unregister_sysctl_table(sd_sysctl_header
);
6398 sd_sysctl_header
= NULL
;
6399 if (sd_ctl_dir
[0].child
)
6400 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6403 static void register_sched_domain_sysctl(void)
6406 static void unregister_sched_domain_sysctl(void)
6411 static void set_rq_online(struct rq
*rq
)
6414 const struct sched_class
*class;
6416 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6419 for_each_class(class) {
6420 if (class->rq_online
)
6421 class->rq_online(rq
);
6426 static void set_rq_offline(struct rq
*rq
)
6429 const struct sched_class
*class;
6431 for_each_class(class) {
6432 if (class->rq_offline
)
6433 class->rq_offline(rq
);
6436 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6442 * migration_call - callback that gets triggered when a CPU is added.
6443 * Here we can start up the necessary migration thread for the new CPU.
6445 static int __cpuinit
6446 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6448 int cpu
= (long)hcpu
;
6449 unsigned long flags
;
6450 struct rq
*rq
= cpu_rq(cpu
);
6452 switch (action
& ~CPU_TASKS_FROZEN
) {
6454 case CPU_UP_PREPARE
:
6455 rq
->calc_load_update
= calc_load_update
;
6459 /* Update our root-domain */
6460 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6462 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6466 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6469 #ifdef CONFIG_HOTPLUG_CPU
6471 sched_ttwu_pending();
6472 /* Update our root-domain */
6473 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6475 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6479 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6480 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6482 migrate_nr_uninterruptible(rq
);
6483 calc_global_load_remove(rq
);
6488 update_max_interval();
6494 * Register at high priority so that task migration (migrate_all_tasks)
6495 * happens before everything else. This has to be lower priority than
6496 * the notifier in the perf_event subsystem, though.
6498 static struct notifier_block __cpuinitdata migration_notifier
= {
6499 .notifier_call
= migration_call
,
6500 .priority
= CPU_PRI_MIGRATION
,
6503 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6504 unsigned long action
, void *hcpu
)
6506 switch (action
& ~CPU_TASKS_FROZEN
) {
6508 case CPU_DOWN_FAILED
:
6509 set_cpu_active((long)hcpu
, true);
6516 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6517 unsigned long action
, void *hcpu
)
6519 switch (action
& ~CPU_TASKS_FROZEN
) {
6520 case CPU_DOWN_PREPARE
:
6521 set_cpu_active((long)hcpu
, false);
6528 static int __init
migration_init(void)
6530 void *cpu
= (void *)(long)smp_processor_id();
6533 /* Initialize migration for the boot CPU */
6534 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6535 BUG_ON(err
== NOTIFY_BAD
);
6536 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6537 register_cpu_notifier(&migration_notifier
);
6539 /* Register cpu active notifiers */
6540 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6541 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6545 early_initcall(migration_init
);
6550 static cpumask_var_t sched_domains_tmpmask
; /* sched_domains_mutex */
6552 #ifdef CONFIG_SCHED_DEBUG
6554 static __read_mostly
int sched_domain_debug_enabled
;
6556 static int __init
sched_domain_debug_setup(char *str
)
6558 sched_domain_debug_enabled
= 1;
6562 early_param("sched_debug", sched_domain_debug_setup
);
6564 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6565 struct cpumask
*groupmask
)
6567 struct sched_group
*group
= sd
->groups
;
6570 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6571 cpumask_clear(groupmask
);
6573 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6575 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6576 printk("does not load-balance\n");
6578 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6583 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6585 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6586 printk(KERN_ERR
"ERROR: domain->span does not contain "
6589 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6590 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6594 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6598 printk(KERN_ERR
"ERROR: group is NULL\n");
6602 if (!group
->sgp
->power
) {
6603 printk(KERN_CONT
"\n");
6604 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6609 if (!cpumask_weight(sched_group_cpus(group
))) {
6610 printk(KERN_CONT
"\n");
6611 printk(KERN_ERR
"ERROR: empty group\n");
6615 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6616 printk(KERN_CONT
"\n");
6617 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6621 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6623 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6625 printk(KERN_CONT
" %s", str
);
6626 if (group
->sgp
->power
!= SCHED_POWER_SCALE
) {
6627 printk(KERN_CONT
" (cpu_power = %d)",
6631 group
= group
->next
;
6632 } while (group
!= sd
->groups
);
6633 printk(KERN_CONT
"\n");
6635 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6636 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6639 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6640 printk(KERN_ERR
"ERROR: parent span is not a superset "
6641 "of domain->span\n");
6645 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6649 if (!sched_domain_debug_enabled
)
6653 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6657 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6660 if (sched_domain_debug_one(sd
, cpu
, level
, sched_domains_tmpmask
))
6668 #else /* !CONFIG_SCHED_DEBUG */
6669 # define sched_domain_debug(sd, cpu) do { } while (0)
6670 #endif /* CONFIG_SCHED_DEBUG */
6672 static int sd_degenerate(struct sched_domain
*sd
)
6674 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6677 /* Following flags need at least 2 groups */
6678 if (sd
->flags
& (SD_LOAD_BALANCE
|
6679 SD_BALANCE_NEWIDLE
|
6683 SD_SHARE_PKG_RESOURCES
)) {
6684 if (sd
->groups
!= sd
->groups
->next
)
6688 /* Following flags don't use groups */
6689 if (sd
->flags
& (SD_WAKE_AFFINE
))
6696 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6698 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6700 if (sd_degenerate(parent
))
6703 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6706 /* Flags needing groups don't count if only 1 group in parent */
6707 if (parent
->groups
== parent
->groups
->next
) {
6708 pflags
&= ~(SD_LOAD_BALANCE
|
6709 SD_BALANCE_NEWIDLE
|
6713 SD_SHARE_PKG_RESOURCES
);
6714 if (nr_node_ids
== 1)
6715 pflags
&= ~SD_SERIALIZE
;
6717 if (~cflags
& pflags
)
6723 static void free_rootdomain(struct rcu_head
*rcu
)
6725 struct root_domain
*rd
= container_of(rcu
, struct root_domain
, rcu
);
6727 cpupri_cleanup(&rd
->cpupri
);
6728 free_cpumask_var(rd
->rto_mask
);
6729 free_cpumask_var(rd
->online
);
6730 free_cpumask_var(rd
->span
);
6734 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6736 struct root_domain
*old_rd
= NULL
;
6737 unsigned long flags
;
6739 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6744 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6747 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6750 * If we dont want to free the old_rt yet then
6751 * set old_rd to NULL to skip the freeing later
6754 if (!atomic_dec_and_test(&old_rd
->refcount
))
6758 atomic_inc(&rd
->refcount
);
6761 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6762 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6765 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6768 call_rcu_sched(&old_rd
->rcu
, free_rootdomain
);
6771 static int init_rootdomain(struct root_domain
*rd
)
6773 memset(rd
, 0, sizeof(*rd
));
6775 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6777 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6779 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6782 if (cpupri_init(&rd
->cpupri
) != 0)
6787 free_cpumask_var(rd
->rto_mask
);
6789 free_cpumask_var(rd
->online
);
6791 free_cpumask_var(rd
->span
);
6796 static void init_defrootdomain(void)
6798 init_rootdomain(&def_root_domain
);
6800 atomic_set(&def_root_domain
.refcount
, 1);
6803 static struct root_domain
*alloc_rootdomain(void)
6805 struct root_domain
*rd
;
6807 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6811 if (init_rootdomain(rd
) != 0) {
6819 static void free_sched_groups(struct sched_group
*sg
, int free_sgp
)
6821 struct sched_group
*tmp
, *first
;
6830 if (free_sgp
&& atomic_dec_and_test(&sg
->sgp
->ref
))
6835 } while (sg
!= first
);
6838 static void free_sched_domain(struct rcu_head
*rcu
)
6840 struct sched_domain
*sd
= container_of(rcu
, struct sched_domain
, rcu
);
6843 * If its an overlapping domain it has private groups, iterate and
6846 if (sd
->flags
& SD_OVERLAP
) {
6847 free_sched_groups(sd
->groups
, 1);
6848 } else if (atomic_dec_and_test(&sd
->groups
->ref
)) {
6849 kfree(sd
->groups
->sgp
);
6855 static void destroy_sched_domain(struct sched_domain
*sd
, int cpu
)
6857 call_rcu(&sd
->rcu
, free_sched_domain
);
6860 static void destroy_sched_domains(struct sched_domain
*sd
, int cpu
)
6862 for (; sd
; sd
= sd
->parent
)
6863 destroy_sched_domain(sd
, cpu
);
6867 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6868 * hold the hotplug lock.
6871 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6873 struct rq
*rq
= cpu_rq(cpu
);
6874 struct sched_domain
*tmp
;
6876 /* Remove the sched domains which do not contribute to scheduling. */
6877 for (tmp
= sd
; tmp
; ) {
6878 struct sched_domain
*parent
= tmp
->parent
;
6882 if (sd_parent_degenerate(tmp
, parent
)) {
6883 tmp
->parent
= parent
->parent
;
6885 parent
->parent
->child
= tmp
;
6886 destroy_sched_domain(parent
, cpu
);
6891 if (sd
&& sd_degenerate(sd
)) {
6894 destroy_sched_domain(tmp
, cpu
);
6899 sched_domain_debug(sd
, cpu
);
6901 rq_attach_root(rq
, rd
);
6903 rcu_assign_pointer(rq
->sd
, sd
);
6904 destroy_sched_domains(tmp
, cpu
);
6907 /* cpus with isolated domains */
6908 static cpumask_var_t cpu_isolated_map
;
6910 /* Setup the mask of cpus configured for isolated domains */
6911 static int __init
isolated_cpu_setup(char *str
)
6913 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6914 cpulist_parse(str
, cpu_isolated_map
);
6918 __setup("isolcpus=", isolated_cpu_setup
);
6920 #define SD_NODES_PER_DOMAIN 16
6925 * find_next_best_node - find the next node to include in a sched_domain
6926 * @node: node whose sched_domain we're building
6927 * @used_nodes: nodes already in the sched_domain
6929 * Find the next node to include in a given scheduling domain. Simply
6930 * finds the closest node not already in the @used_nodes map.
6932 * Should use nodemask_t.
6934 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6936 int i
, n
, val
, min_val
, best_node
= -1;
6940 for (i
= 0; i
< nr_node_ids
; i
++) {
6941 /* Start at @node */
6942 n
= (node
+ i
) % nr_node_ids
;
6944 if (!nr_cpus_node(n
))
6947 /* Skip already used nodes */
6948 if (node_isset(n
, *used_nodes
))
6951 /* Simple min distance search */
6952 val
= node_distance(node
, n
);
6954 if (val
< min_val
) {
6960 if (best_node
!= -1)
6961 node_set(best_node
, *used_nodes
);
6966 * sched_domain_node_span - get a cpumask for a node's sched_domain
6967 * @node: node whose cpumask we're constructing
6968 * @span: resulting cpumask
6970 * Given a node, construct a good cpumask for its sched_domain to span. It
6971 * should be one that prevents unnecessary balancing, but also spreads tasks
6974 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6976 nodemask_t used_nodes
;
6979 cpumask_clear(span
);
6980 nodes_clear(used_nodes
);
6982 cpumask_or(span
, span
, cpumask_of_node(node
));
6983 node_set(node
, used_nodes
);
6985 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6986 int next_node
= find_next_best_node(node
, &used_nodes
);
6989 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6993 static const struct cpumask
*cpu_node_mask(int cpu
)
6995 lockdep_assert_held(&sched_domains_mutex
);
6997 sched_domain_node_span(cpu_to_node(cpu
), sched_domains_tmpmask
);
6999 return sched_domains_tmpmask
;
7002 static const struct cpumask
*cpu_allnodes_mask(int cpu
)
7004 return cpu_possible_mask
;
7006 #endif /* CONFIG_NUMA */
7008 static const struct cpumask
*cpu_cpu_mask(int cpu
)
7010 return cpumask_of_node(cpu_to_node(cpu
));
7013 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7016 struct sched_domain
**__percpu sd
;
7017 struct sched_group
**__percpu sg
;
7018 struct sched_group_power
**__percpu sgp
;
7022 struct sched_domain
** __percpu sd
;
7023 struct root_domain
*rd
;
7033 struct sched_domain_topology_level
;
7035 typedef struct sched_domain
*(*sched_domain_init_f
)(struct sched_domain_topology_level
*tl
, int cpu
);
7036 typedef const struct cpumask
*(*sched_domain_mask_f
)(int cpu
);
7038 #define SDTL_OVERLAP 0x01
7040 struct sched_domain_topology_level
{
7041 sched_domain_init_f init
;
7042 sched_domain_mask_f mask
;
7044 struct sd_data data
;
7048 build_overlap_sched_groups(struct sched_domain
*sd
, int cpu
)
7050 struct sched_group
*first
= NULL
, *last
= NULL
, *groups
= NULL
, *sg
;
7051 const struct cpumask
*span
= sched_domain_span(sd
);
7052 struct cpumask
*covered
= sched_domains_tmpmask
;
7053 struct sd_data
*sdd
= sd
->private;
7054 struct sched_domain
*child
;
7057 cpumask_clear(covered
);
7059 for_each_cpu(i
, span
) {
7060 struct cpumask
*sg_span
;
7062 if (cpumask_test_cpu(i
, covered
))
7065 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7066 GFP_KERNEL
, cpu_to_node(i
));
7071 sg_span
= sched_group_cpus(sg
);
7073 child
= *per_cpu_ptr(sdd
->sd
, i
);
7075 child
= child
->child
;
7076 cpumask_copy(sg_span
, sched_domain_span(child
));
7078 cpumask_set_cpu(i
, sg_span
);
7080 cpumask_or(covered
, covered
, sg_span
);
7082 sg
->sgp
= *per_cpu_ptr(sdd
->sgp
, cpumask_first(sg_span
));
7083 atomic_inc(&sg
->sgp
->ref
);
7085 if (cpumask_test_cpu(cpu
, sg_span
))
7095 sd
->groups
= groups
;
7100 free_sched_groups(first
, 0);
7105 static int get_group(int cpu
, struct sd_data
*sdd
, struct sched_group
**sg
)
7107 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, cpu
);
7108 struct sched_domain
*child
= sd
->child
;
7111 cpu
= cpumask_first(sched_domain_span(child
));
7114 *sg
= *per_cpu_ptr(sdd
->sg
, cpu
);
7115 (*sg
)->sgp
= *per_cpu_ptr(sdd
->sgp
, cpu
);
7116 atomic_set(&(*sg
)->sgp
->ref
, 1); /* for claim_allocations */
7123 * build_sched_groups will build a circular linked list of the groups
7124 * covered by the given span, and will set each group's ->cpumask correctly,
7125 * and ->cpu_power to 0.
7127 * Assumes the sched_domain tree is fully constructed
7130 build_sched_groups(struct sched_domain
*sd
, int cpu
)
7132 struct sched_group
*first
= NULL
, *last
= NULL
;
7133 struct sd_data
*sdd
= sd
->private;
7134 const struct cpumask
*span
= sched_domain_span(sd
);
7135 struct cpumask
*covered
;
7138 get_group(cpu
, sdd
, &sd
->groups
);
7139 atomic_inc(&sd
->groups
->ref
);
7141 if (cpu
!= cpumask_first(sched_domain_span(sd
)))
7144 lockdep_assert_held(&sched_domains_mutex
);
7145 covered
= sched_domains_tmpmask
;
7147 cpumask_clear(covered
);
7149 for_each_cpu(i
, span
) {
7150 struct sched_group
*sg
;
7151 int group
= get_group(i
, sdd
, &sg
);
7154 if (cpumask_test_cpu(i
, covered
))
7157 cpumask_clear(sched_group_cpus(sg
));
7160 for_each_cpu(j
, span
) {
7161 if (get_group(j
, sdd
, NULL
) != group
)
7164 cpumask_set_cpu(j
, covered
);
7165 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7180 * Initialize sched groups cpu_power.
7182 * cpu_power indicates the capacity of sched group, which is used while
7183 * distributing the load between different sched groups in a sched domain.
7184 * Typically cpu_power for all the groups in a sched domain will be same unless
7185 * there are asymmetries in the topology. If there are asymmetries, group
7186 * having more cpu_power will pickup more load compared to the group having
7189 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7191 struct sched_group
*sg
= sd
->groups
;
7193 WARN_ON(!sd
|| !sg
);
7196 sg
->group_weight
= cpumask_weight(sched_group_cpus(sg
));
7198 } while (sg
!= sd
->groups
);
7200 if (cpu
!= group_first_cpu(sg
))
7203 update_group_power(sd
, cpu
);
7207 * Initializers for schedule domains
7208 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7211 #ifdef CONFIG_SCHED_DEBUG
7212 # define SD_INIT_NAME(sd, type) sd->name = #type
7214 # define SD_INIT_NAME(sd, type) do { } while (0)
7217 #define SD_INIT_FUNC(type) \
7218 static noinline struct sched_domain * \
7219 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7221 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7222 *sd = SD_##type##_INIT; \
7223 SD_INIT_NAME(sd, type); \
7224 sd->private = &tl->data; \
7230 SD_INIT_FUNC(ALLNODES
)
7233 #ifdef CONFIG_SCHED_SMT
7234 SD_INIT_FUNC(SIBLING
)
7236 #ifdef CONFIG_SCHED_MC
7239 #ifdef CONFIG_SCHED_BOOK
7243 static int default_relax_domain_level
= -1;
7244 int sched_domain_level_max
;
7246 static int __init
setup_relax_domain_level(char *str
)
7250 val
= simple_strtoul(str
, NULL
, 0);
7251 if (val
< sched_domain_level_max
)
7252 default_relax_domain_level
= val
;
7256 __setup("relax_domain_level=", setup_relax_domain_level
);
7258 static void set_domain_attribute(struct sched_domain
*sd
,
7259 struct sched_domain_attr
*attr
)
7263 if (!attr
|| attr
->relax_domain_level
< 0) {
7264 if (default_relax_domain_level
< 0)
7267 request
= default_relax_domain_level
;
7269 request
= attr
->relax_domain_level
;
7270 if (request
< sd
->level
) {
7271 /* turn off idle balance on this domain */
7272 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7274 /* turn on idle balance on this domain */
7275 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7279 static void __sdt_free(const struct cpumask
*cpu_map
);
7280 static int __sdt_alloc(const struct cpumask
*cpu_map
);
7282 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7283 const struct cpumask
*cpu_map
)
7287 if (!atomic_read(&d
->rd
->refcount
))
7288 free_rootdomain(&d
->rd
->rcu
); /* fall through */
7290 free_percpu(d
->sd
); /* fall through */
7292 __sdt_free(cpu_map
); /* fall through */
7298 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7299 const struct cpumask
*cpu_map
)
7301 memset(d
, 0, sizeof(*d
));
7303 if (__sdt_alloc(cpu_map
))
7304 return sa_sd_storage
;
7305 d
->sd
= alloc_percpu(struct sched_domain
*);
7307 return sa_sd_storage
;
7308 d
->rd
= alloc_rootdomain();
7311 return sa_rootdomain
;
7315 * NULL the sd_data elements we've used to build the sched_domain and
7316 * sched_group structure so that the subsequent __free_domain_allocs()
7317 * will not free the data we're using.
7319 static void claim_allocations(int cpu
, struct sched_domain
*sd
)
7321 struct sd_data
*sdd
= sd
->private;
7323 WARN_ON_ONCE(*per_cpu_ptr(sdd
->sd
, cpu
) != sd
);
7324 *per_cpu_ptr(sdd
->sd
, cpu
) = NULL
;
7326 if (atomic_read(&(*per_cpu_ptr(sdd
->sg
, cpu
))->ref
))
7327 *per_cpu_ptr(sdd
->sg
, cpu
) = NULL
;
7329 if (atomic_read(&(*per_cpu_ptr(sdd
->sgp
, cpu
))->ref
))
7330 *per_cpu_ptr(sdd
->sgp
, cpu
) = NULL
;
7333 #ifdef CONFIG_SCHED_SMT
7334 static const struct cpumask
*cpu_smt_mask(int cpu
)
7336 return topology_thread_cpumask(cpu
);
7341 * Topology list, bottom-up.
7343 static struct sched_domain_topology_level default_topology
[] = {
7344 #ifdef CONFIG_SCHED_SMT
7345 { sd_init_SIBLING
, cpu_smt_mask
, },
7347 #ifdef CONFIG_SCHED_MC
7348 { sd_init_MC
, cpu_coregroup_mask
, },
7350 #ifdef CONFIG_SCHED_BOOK
7351 { sd_init_BOOK
, cpu_book_mask
, },
7353 { sd_init_CPU
, cpu_cpu_mask
, },
7355 { sd_init_NODE
, cpu_node_mask
, SDTL_OVERLAP
, },
7356 { sd_init_ALLNODES
, cpu_allnodes_mask
, },
7361 static struct sched_domain_topology_level
*sched_domain_topology
= default_topology
;
7363 static int __sdt_alloc(const struct cpumask
*cpu_map
)
7365 struct sched_domain_topology_level
*tl
;
7368 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7369 struct sd_data
*sdd
= &tl
->data
;
7371 sdd
->sd
= alloc_percpu(struct sched_domain
*);
7375 sdd
->sg
= alloc_percpu(struct sched_group
*);
7379 sdd
->sgp
= alloc_percpu(struct sched_group_power
*);
7383 for_each_cpu(j
, cpu_map
) {
7384 struct sched_domain
*sd
;
7385 struct sched_group
*sg
;
7386 struct sched_group_power
*sgp
;
7388 sd
= kzalloc_node(sizeof(struct sched_domain
) + cpumask_size(),
7389 GFP_KERNEL
, cpu_to_node(j
));
7393 *per_cpu_ptr(sdd
->sd
, j
) = sd
;
7395 sg
= kzalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7396 GFP_KERNEL
, cpu_to_node(j
));
7400 *per_cpu_ptr(sdd
->sg
, j
) = sg
;
7402 sgp
= kzalloc_node(sizeof(struct sched_group_power
),
7403 GFP_KERNEL
, cpu_to_node(j
));
7407 *per_cpu_ptr(sdd
->sgp
, j
) = sgp
;
7414 static void __sdt_free(const struct cpumask
*cpu_map
)
7416 struct sched_domain_topology_level
*tl
;
7419 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7420 struct sd_data
*sdd
= &tl
->data
;
7422 for_each_cpu(j
, cpu_map
) {
7423 struct sched_domain
*sd
= *per_cpu_ptr(sdd
->sd
, j
);
7424 if (sd
&& (sd
->flags
& SD_OVERLAP
))
7425 free_sched_groups(sd
->groups
, 0);
7426 kfree(*per_cpu_ptr(sdd
->sd
, j
));
7427 kfree(*per_cpu_ptr(sdd
->sg
, j
));
7428 kfree(*per_cpu_ptr(sdd
->sgp
, j
));
7430 free_percpu(sdd
->sd
);
7431 free_percpu(sdd
->sg
);
7432 free_percpu(sdd
->sgp
);
7436 struct sched_domain
*build_sched_domain(struct sched_domain_topology_level
*tl
,
7437 struct s_data
*d
, const struct cpumask
*cpu_map
,
7438 struct sched_domain_attr
*attr
, struct sched_domain
*child
,
7441 struct sched_domain
*sd
= tl
->init(tl
, cpu
);
7445 set_domain_attribute(sd
, attr
);
7446 cpumask_and(sched_domain_span(sd
), cpu_map
, tl
->mask(cpu
));
7448 sd
->level
= child
->level
+ 1;
7449 sched_domain_level_max
= max(sched_domain_level_max
, sd
->level
);
7458 * Build sched domains for a given set of cpus and attach the sched domains
7459 * to the individual cpus
7461 static int build_sched_domains(const struct cpumask
*cpu_map
,
7462 struct sched_domain_attr
*attr
)
7464 enum s_alloc alloc_state
= sa_none
;
7465 struct sched_domain
*sd
;
7467 int i
, ret
= -ENOMEM
;
7469 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7470 if (alloc_state
!= sa_rootdomain
)
7473 /* Set up domains for cpus specified by the cpu_map. */
7474 for_each_cpu(i
, cpu_map
) {
7475 struct sched_domain_topology_level
*tl
;
7478 for (tl
= sched_domain_topology
; tl
->init
; tl
++) {
7479 sd
= build_sched_domain(tl
, &d
, cpu_map
, attr
, sd
, i
);
7480 if (tl
->flags
& SDTL_OVERLAP
|| sched_feat(FORCE_SD_OVERLAP
))
7481 sd
->flags
|= SD_OVERLAP
;
7482 if (cpumask_equal(cpu_map
, sched_domain_span(sd
)))
7489 *per_cpu_ptr(d
.sd
, i
) = sd
;
7492 /* Build the groups for the domains */
7493 for_each_cpu(i
, cpu_map
) {
7494 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7495 sd
->span_weight
= cpumask_weight(sched_domain_span(sd
));
7496 if (sd
->flags
& SD_OVERLAP
) {
7497 if (build_overlap_sched_groups(sd
, i
))
7500 if (build_sched_groups(sd
, i
))
7506 /* Calculate CPU power for physical packages and nodes */
7507 for (i
= nr_cpumask_bits
-1; i
>= 0; i
--) {
7508 if (!cpumask_test_cpu(i
, cpu_map
))
7511 for (sd
= *per_cpu_ptr(d
.sd
, i
); sd
; sd
= sd
->parent
) {
7512 claim_allocations(i
, sd
);
7513 init_sched_groups_power(i
, sd
);
7517 /* Attach the domains */
7519 for_each_cpu(i
, cpu_map
) {
7520 sd
= *per_cpu_ptr(d
.sd
, i
);
7521 cpu_attach_domain(sd
, d
.rd
, i
);
7527 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7531 static cpumask_var_t
*doms_cur
; /* current sched domains */
7532 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7533 static struct sched_domain_attr
*dattr_cur
;
7534 /* attribues of custom domains in 'doms_cur' */
7537 * Special case: If a kmalloc of a doms_cur partition (array of
7538 * cpumask) fails, then fallback to a single sched domain,
7539 * as determined by the single cpumask fallback_doms.
7541 static cpumask_var_t fallback_doms
;
7544 * arch_update_cpu_topology lets virtualized architectures update the
7545 * cpu core maps. It is supposed to return 1 if the topology changed
7546 * or 0 if it stayed the same.
7548 int __attribute__((weak
)) arch_update_cpu_topology(void)
7553 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7556 cpumask_var_t
*doms
;
7558 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7561 for (i
= 0; i
< ndoms
; i
++) {
7562 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7563 free_sched_domains(doms
, i
);
7570 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7573 for (i
= 0; i
< ndoms
; i
++)
7574 free_cpumask_var(doms
[i
]);
7579 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7580 * For now this just excludes isolated cpus, but could be used to
7581 * exclude other special cases in the future.
7583 static int init_sched_domains(const struct cpumask
*cpu_map
)
7587 arch_update_cpu_topology();
7589 doms_cur
= alloc_sched_domains(ndoms_cur
);
7591 doms_cur
= &fallback_doms
;
7592 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7594 err
= build_sched_domains(doms_cur
[0], NULL
);
7595 register_sched_domain_sysctl();
7601 * Detach sched domains from a group of cpus specified in cpu_map
7602 * These cpus will now be attached to the NULL domain
7604 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7609 for_each_cpu(i
, cpu_map
)
7610 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7614 /* handle null as "default" */
7615 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7616 struct sched_domain_attr
*new, int idx_new
)
7618 struct sched_domain_attr tmp
;
7625 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7626 new ? (new + idx_new
) : &tmp
,
7627 sizeof(struct sched_domain_attr
));
7631 * Partition sched domains as specified by the 'ndoms_new'
7632 * cpumasks in the array doms_new[] of cpumasks. This compares
7633 * doms_new[] to the current sched domain partitioning, doms_cur[].
7634 * It destroys each deleted domain and builds each new domain.
7636 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7637 * The masks don't intersect (don't overlap.) We should setup one
7638 * sched domain for each mask. CPUs not in any of the cpumasks will
7639 * not be load balanced. If the same cpumask appears both in the
7640 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7643 * The passed in 'doms_new' should be allocated using
7644 * alloc_sched_domains. This routine takes ownership of it and will
7645 * free_sched_domains it when done with it. If the caller failed the
7646 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7647 * and partition_sched_domains() will fallback to the single partition
7648 * 'fallback_doms', it also forces the domains to be rebuilt.
7650 * If doms_new == NULL it will be replaced with cpu_online_mask.
7651 * ndoms_new == 0 is a special case for destroying existing domains,
7652 * and it will not create the default domain.
7654 * Call with hotplug lock held
7656 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7657 struct sched_domain_attr
*dattr_new
)
7662 mutex_lock(&sched_domains_mutex
);
7664 /* always unregister in case we don't destroy any domains */
7665 unregister_sched_domain_sysctl();
7667 /* Let architecture update cpu core mappings. */
7668 new_topology
= arch_update_cpu_topology();
7670 n
= doms_new
? ndoms_new
: 0;
7672 /* Destroy deleted domains */
7673 for (i
= 0; i
< ndoms_cur
; i
++) {
7674 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7675 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7676 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7679 /* no match - a current sched domain not in new doms_new[] */
7680 detach_destroy_domains(doms_cur
[i
]);
7685 if (doms_new
== NULL
) {
7687 doms_new
= &fallback_doms
;
7688 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7689 WARN_ON_ONCE(dattr_new
);
7692 /* Build new domains */
7693 for (i
= 0; i
< ndoms_new
; i
++) {
7694 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7695 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7696 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7699 /* no match - add a new doms_new */
7700 build_sched_domains(doms_new
[i
], dattr_new
? dattr_new
+ i
: NULL
);
7705 /* Remember the new sched domains */
7706 if (doms_cur
!= &fallback_doms
)
7707 free_sched_domains(doms_cur
, ndoms_cur
);
7708 kfree(dattr_cur
); /* kfree(NULL) is safe */
7709 doms_cur
= doms_new
;
7710 dattr_cur
= dattr_new
;
7711 ndoms_cur
= ndoms_new
;
7713 register_sched_domain_sysctl();
7715 mutex_unlock(&sched_domains_mutex
);
7718 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7719 static void reinit_sched_domains(void)
7723 /* Destroy domains first to force the rebuild */
7724 partition_sched_domains(0, NULL
, NULL
);
7726 rebuild_sched_domains();
7730 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7732 unsigned int level
= 0;
7734 if (sscanf(buf
, "%u", &level
) != 1)
7738 * level is always be positive so don't check for
7739 * level < POWERSAVINGS_BALANCE_NONE which is 0
7740 * What happens on 0 or 1 byte write,
7741 * need to check for count as well?
7744 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7748 sched_smt_power_savings
= level
;
7750 sched_mc_power_savings
= level
;
7752 reinit_sched_domains();
7757 #ifdef CONFIG_SCHED_MC
7758 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7759 struct sysdev_class_attribute
*attr
,
7762 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7764 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7765 struct sysdev_class_attribute
*attr
,
7766 const char *buf
, size_t count
)
7768 return sched_power_savings_store(buf
, count
, 0);
7770 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7771 sched_mc_power_savings_show
,
7772 sched_mc_power_savings_store
);
7775 #ifdef CONFIG_SCHED_SMT
7776 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7777 struct sysdev_class_attribute
*attr
,
7780 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7782 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7783 struct sysdev_class_attribute
*attr
,
7784 const char *buf
, size_t count
)
7786 return sched_power_savings_store(buf
, count
, 1);
7788 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7789 sched_smt_power_savings_show
,
7790 sched_smt_power_savings_store
);
7793 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7797 #ifdef CONFIG_SCHED_SMT
7799 err
= sysfs_create_file(&cls
->kset
.kobj
,
7800 &attr_sched_smt_power_savings
.attr
);
7802 #ifdef CONFIG_SCHED_MC
7803 if (!err
&& mc_capable())
7804 err
= sysfs_create_file(&cls
->kset
.kobj
,
7805 &attr_sched_mc_power_savings
.attr
);
7809 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7812 * Update cpusets according to cpu_active mask. If cpusets are
7813 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7814 * around partition_sched_domains().
7816 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7819 switch (action
& ~CPU_TASKS_FROZEN
) {
7821 case CPU_DOWN_FAILED
:
7822 cpuset_update_active_cpus();
7829 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7832 switch (action
& ~CPU_TASKS_FROZEN
) {
7833 case CPU_DOWN_PREPARE
:
7834 cpuset_update_active_cpus();
7841 static int update_runtime(struct notifier_block
*nfb
,
7842 unsigned long action
, void *hcpu
)
7844 int cpu
= (int)(long)hcpu
;
7847 case CPU_DOWN_PREPARE
:
7848 case CPU_DOWN_PREPARE_FROZEN
:
7849 disable_runtime(cpu_rq(cpu
));
7852 case CPU_DOWN_FAILED
:
7853 case CPU_DOWN_FAILED_FROZEN
:
7855 case CPU_ONLINE_FROZEN
:
7856 enable_runtime(cpu_rq(cpu
));
7864 void __init
sched_init_smp(void)
7866 cpumask_var_t non_isolated_cpus
;
7868 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7869 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7872 mutex_lock(&sched_domains_mutex
);
7873 init_sched_domains(cpu_active_mask
);
7874 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7875 if (cpumask_empty(non_isolated_cpus
))
7876 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7877 mutex_unlock(&sched_domains_mutex
);
7880 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7881 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7883 /* RT runtime code needs to handle some hotplug events */
7884 hotcpu_notifier(update_runtime
, 0);
7888 /* Move init over to a non-isolated CPU */
7889 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7891 sched_init_granularity();
7892 free_cpumask_var(non_isolated_cpus
);
7894 init_sched_rt_class();
7897 void __init
sched_init_smp(void)
7899 sched_init_granularity();
7901 #endif /* CONFIG_SMP */
7903 const_debug
unsigned int sysctl_timer_migration
= 1;
7905 int in_sched_functions(unsigned long addr
)
7907 return in_lock_functions(addr
) ||
7908 (addr
>= (unsigned long)__sched_text_start
7909 && addr
< (unsigned long)__sched_text_end
);
7912 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7914 cfs_rq
->tasks_timeline
= RB_ROOT
;
7915 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7916 #ifdef CONFIG_FAIR_GROUP_SCHED
7918 /* allow initial update_cfs_load() to truncate */
7920 cfs_rq
->load_stamp
= 1;
7923 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7924 #ifndef CONFIG_64BIT
7925 cfs_rq
->min_vruntime_copy
= cfs_rq
->min_vruntime
;
7929 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7931 struct rt_prio_array
*array
;
7934 array
= &rt_rq
->active
;
7935 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7936 INIT_LIST_HEAD(array
->queue
+ i
);
7937 __clear_bit(i
, array
->bitmap
);
7939 /* delimiter for bitsearch: */
7940 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7942 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7943 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7945 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7949 rt_rq
->rt_nr_migratory
= 0;
7950 rt_rq
->overloaded
= 0;
7951 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7955 rt_rq
->rt_throttled
= 0;
7956 rt_rq
->rt_runtime
= 0;
7957 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7959 #ifdef CONFIG_RT_GROUP_SCHED
7960 rt_rq
->rt_nr_boosted
= 0;
7965 #ifdef CONFIG_FAIR_GROUP_SCHED
7966 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7967 struct sched_entity
*se
, int cpu
,
7968 struct sched_entity
*parent
)
7970 struct rq
*rq
= cpu_rq(cpu
);
7971 tg
->cfs_rq
[cpu
] = cfs_rq
;
7972 init_cfs_rq(cfs_rq
, rq
);
7976 /* se could be NULL for root_task_group */
7981 se
->cfs_rq
= &rq
->cfs
;
7983 se
->cfs_rq
= parent
->my_q
;
7986 update_load_set(&se
->load
, 0);
7987 se
->parent
= parent
;
7991 #ifdef CONFIG_RT_GROUP_SCHED
7992 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7993 struct sched_rt_entity
*rt_se
, int cpu
,
7994 struct sched_rt_entity
*parent
)
7996 struct rq
*rq
= cpu_rq(cpu
);
7998 tg
->rt_rq
[cpu
] = rt_rq
;
7999 init_rt_rq(rt_rq
, rq
);
8001 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8003 tg
->rt_se
[cpu
] = rt_se
;
8008 rt_se
->rt_rq
= &rq
->rt
;
8010 rt_se
->rt_rq
= parent
->my_q
;
8012 rt_se
->my_q
= rt_rq
;
8013 rt_se
->parent
= parent
;
8014 INIT_LIST_HEAD(&rt_se
->run_list
);
8018 void __init
sched_init(void)
8021 unsigned long alloc_size
= 0, ptr
;
8023 #ifdef CONFIG_FAIR_GROUP_SCHED
8024 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8026 #ifdef CONFIG_RT_GROUP_SCHED
8027 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8029 #ifdef CONFIG_CPUMASK_OFFSTACK
8030 alloc_size
+= num_possible_cpus() * cpumask_size();
8033 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
8035 #ifdef CONFIG_FAIR_GROUP_SCHED
8036 root_task_group
.se
= (struct sched_entity
**)ptr
;
8037 ptr
+= nr_cpu_ids
* sizeof(void **);
8039 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8040 ptr
+= nr_cpu_ids
* sizeof(void **);
8042 #endif /* CONFIG_FAIR_GROUP_SCHED */
8043 #ifdef CONFIG_RT_GROUP_SCHED
8044 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8045 ptr
+= nr_cpu_ids
* sizeof(void **);
8047 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8048 ptr
+= nr_cpu_ids
* sizeof(void **);
8050 #endif /* CONFIG_RT_GROUP_SCHED */
8051 #ifdef CONFIG_CPUMASK_OFFSTACK
8052 for_each_possible_cpu(i
) {
8053 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8054 ptr
+= cpumask_size();
8056 #endif /* CONFIG_CPUMASK_OFFSTACK */
8060 init_defrootdomain();
8063 init_rt_bandwidth(&def_rt_bandwidth
,
8064 global_rt_period(), global_rt_runtime());
8066 #ifdef CONFIG_RT_GROUP_SCHED
8067 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8068 global_rt_period(), global_rt_runtime());
8069 #endif /* CONFIG_RT_GROUP_SCHED */
8071 #ifdef CONFIG_CGROUP_SCHED
8072 list_add(&root_task_group
.list
, &task_groups
);
8073 INIT_LIST_HEAD(&root_task_group
.children
);
8074 autogroup_init(&init_task
);
8075 #endif /* CONFIG_CGROUP_SCHED */
8077 for_each_possible_cpu(i
) {
8081 raw_spin_lock_init(&rq
->lock
);
8083 rq
->calc_load_active
= 0;
8084 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8085 init_cfs_rq(&rq
->cfs
, rq
);
8086 init_rt_rq(&rq
->rt
, rq
);
8087 #ifdef CONFIG_FAIR_GROUP_SCHED
8088 root_task_group
.shares
= root_task_group_load
;
8089 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8091 * How much cpu bandwidth does root_task_group get?
8093 * In case of task-groups formed thr' the cgroup filesystem, it
8094 * gets 100% of the cpu resources in the system. This overall
8095 * system cpu resource is divided among the tasks of
8096 * root_task_group and its child task-groups in a fair manner,
8097 * based on each entity's (task or task-group's) weight
8098 * (se->load.weight).
8100 * In other words, if root_task_group has 10 tasks of weight
8101 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8102 * then A0's share of the cpu resource is:
8104 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8106 * We achieve this by letting root_task_group's tasks sit
8107 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8109 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8110 #endif /* CONFIG_FAIR_GROUP_SCHED */
8112 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8113 #ifdef CONFIG_RT_GROUP_SCHED
8114 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8115 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8118 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8119 rq
->cpu_load
[j
] = 0;
8121 rq
->last_load_update_tick
= jiffies
;
8126 rq
->cpu_power
= SCHED_POWER_SCALE
;
8127 rq
->post_schedule
= 0;
8128 rq
->active_balance
= 0;
8129 rq
->next_balance
= jiffies
;
8134 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8135 rq_attach_root(rq
, &def_root_domain
);
8137 rq
->nohz_balance_kick
= 0;
8138 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
8142 atomic_set(&rq
->nr_iowait
, 0);
8145 set_load_weight(&init_task
);
8147 #ifdef CONFIG_PREEMPT_NOTIFIERS
8148 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8152 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8155 #ifdef CONFIG_RT_MUTEXES
8156 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8160 * The boot idle thread does lazy MMU switching as well:
8162 atomic_inc(&init_mm
.mm_count
);
8163 enter_lazy_tlb(&init_mm
, current
);
8166 * Make us the idle thread. Technically, schedule() should not be
8167 * called from this thread, however somewhere below it might be,
8168 * but because we are the idle thread, we just pick up running again
8169 * when this runqueue becomes "idle".
8171 init_idle(current
, smp_processor_id());
8173 calc_load_update
= jiffies
+ LOAD_FREQ
;
8176 * During early bootup we pretend to be a normal task:
8178 current
->sched_class
= &fair_sched_class
;
8180 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8181 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
8183 zalloc_cpumask_var(&sched_domains_tmpmask
, GFP_NOWAIT
);
8185 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8186 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8187 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8188 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8189 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8191 /* May be allocated at isolcpus cmdline parse time */
8192 if (cpu_isolated_map
== NULL
)
8193 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8196 scheduler_running
= 1;
8199 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8200 static inline int preempt_count_equals(int preempt_offset
)
8202 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8204 return (nested
== preempt_offset
);
8207 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8210 static unsigned long prev_jiffy
; /* ratelimiting */
8212 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8213 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8215 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8217 prev_jiffy
= jiffies
;
8220 "BUG: sleeping function called from invalid context at %s:%d\n",
8223 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8224 in_atomic(), irqs_disabled(),
8225 current
->pid
, current
->comm
);
8227 debug_show_held_locks(current
);
8228 if (irqs_disabled())
8229 print_irqtrace_events(current
);
8233 EXPORT_SYMBOL(__might_sleep
);
8236 #ifdef CONFIG_MAGIC_SYSRQ
8237 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8239 const struct sched_class
*prev_class
= p
->sched_class
;
8240 int old_prio
= p
->prio
;
8245 deactivate_task(rq
, p
, 0);
8246 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8248 activate_task(rq
, p
, 0);
8249 resched_task(rq
->curr
);
8252 check_class_changed(rq
, p
, prev_class
, old_prio
);
8255 void normalize_rt_tasks(void)
8257 struct task_struct
*g
, *p
;
8258 unsigned long flags
;
8261 read_lock_irqsave(&tasklist_lock
, flags
);
8262 do_each_thread(g
, p
) {
8264 * Only normalize user tasks:
8269 p
->se
.exec_start
= 0;
8270 #ifdef CONFIG_SCHEDSTATS
8271 p
->se
.statistics
.wait_start
= 0;
8272 p
->se
.statistics
.sleep_start
= 0;
8273 p
->se
.statistics
.block_start
= 0;
8278 * Renice negative nice level userspace
8281 if (TASK_NICE(p
) < 0 && p
->mm
)
8282 set_user_nice(p
, 0);
8286 raw_spin_lock(&p
->pi_lock
);
8287 rq
= __task_rq_lock(p
);
8289 normalize_task(rq
, p
);
8291 __task_rq_unlock(rq
);
8292 raw_spin_unlock(&p
->pi_lock
);
8293 } while_each_thread(g
, p
);
8295 read_unlock_irqrestore(&tasklist_lock
, flags
);
8298 #endif /* CONFIG_MAGIC_SYSRQ */
8300 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8302 * These functions are only useful for the IA64 MCA handling, or kdb.
8304 * They can only be called when the whole system has been
8305 * stopped - every CPU needs to be quiescent, and no scheduling
8306 * activity can take place. Using them for anything else would
8307 * be a serious bug, and as a result, they aren't even visible
8308 * under any other configuration.
8312 * curr_task - return the current task for a given cpu.
8313 * @cpu: the processor in question.
8315 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8317 struct task_struct
*curr_task(int cpu
)
8319 return cpu_curr(cpu
);
8322 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8326 * set_curr_task - set the current task for a given cpu.
8327 * @cpu: the processor in question.
8328 * @p: the task pointer to set.
8330 * Description: This function must only be used when non-maskable interrupts
8331 * are serviced on a separate stack. It allows the architecture to switch the
8332 * notion of the current task on a cpu in a non-blocking manner. This function
8333 * must be called with all CPU's synchronized, and interrupts disabled, the
8334 * and caller must save the original value of the current task (see
8335 * curr_task() above) and restore that value before reenabling interrupts and
8336 * re-starting the system.
8338 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8340 void set_curr_task(int cpu
, struct task_struct
*p
)
8347 #ifdef CONFIG_FAIR_GROUP_SCHED
8348 static void free_fair_sched_group(struct task_group
*tg
)
8352 for_each_possible_cpu(i
) {
8354 kfree(tg
->cfs_rq
[i
]);
8364 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8366 struct cfs_rq
*cfs_rq
;
8367 struct sched_entity
*se
;
8370 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8373 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8377 tg
->shares
= NICE_0_LOAD
;
8379 for_each_possible_cpu(i
) {
8380 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8381 GFP_KERNEL
, cpu_to_node(i
));
8385 se
= kzalloc_node(sizeof(struct sched_entity
),
8386 GFP_KERNEL
, cpu_to_node(i
));
8390 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8401 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8403 struct rq
*rq
= cpu_rq(cpu
);
8404 unsigned long flags
;
8407 * Only empty task groups can be destroyed; so we can speculatively
8408 * check on_list without danger of it being re-added.
8410 if (!tg
->cfs_rq
[cpu
]->on_list
)
8413 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8414 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8415 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8417 #else /* !CONFG_FAIR_GROUP_SCHED */
8418 static inline void free_fair_sched_group(struct task_group
*tg
)
8423 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8428 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8431 #endif /* CONFIG_FAIR_GROUP_SCHED */
8433 #ifdef CONFIG_RT_GROUP_SCHED
8434 static void free_rt_sched_group(struct task_group
*tg
)
8438 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8440 for_each_possible_cpu(i
) {
8442 kfree(tg
->rt_rq
[i
]);
8444 kfree(tg
->rt_se
[i
]);
8452 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8454 struct rt_rq
*rt_rq
;
8455 struct sched_rt_entity
*rt_se
;
8458 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8461 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8465 init_rt_bandwidth(&tg
->rt_bandwidth
,
8466 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8468 for_each_possible_cpu(i
) {
8469 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8470 GFP_KERNEL
, cpu_to_node(i
));
8474 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8475 GFP_KERNEL
, cpu_to_node(i
));
8479 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8489 #else /* !CONFIG_RT_GROUP_SCHED */
8490 static inline void free_rt_sched_group(struct task_group
*tg
)
8495 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8499 #endif /* CONFIG_RT_GROUP_SCHED */
8501 #ifdef CONFIG_CGROUP_SCHED
8502 static void free_sched_group(struct task_group
*tg
)
8504 free_fair_sched_group(tg
);
8505 free_rt_sched_group(tg
);
8510 /* allocate runqueue etc for a new task group */
8511 struct task_group
*sched_create_group(struct task_group
*parent
)
8513 struct task_group
*tg
;
8514 unsigned long flags
;
8516 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8518 return ERR_PTR(-ENOMEM
);
8520 if (!alloc_fair_sched_group(tg
, parent
))
8523 if (!alloc_rt_sched_group(tg
, parent
))
8526 spin_lock_irqsave(&task_group_lock
, flags
);
8527 list_add_rcu(&tg
->list
, &task_groups
);
8529 WARN_ON(!parent
); /* root should already exist */
8531 tg
->parent
= parent
;
8532 INIT_LIST_HEAD(&tg
->children
);
8533 list_add_rcu(&tg
->siblings
, &parent
->children
);
8534 spin_unlock_irqrestore(&task_group_lock
, flags
);
8539 free_sched_group(tg
);
8540 return ERR_PTR(-ENOMEM
);
8543 /* rcu callback to free various structures associated with a task group */
8544 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8546 /* now it should be safe to free those cfs_rqs */
8547 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8550 /* Destroy runqueue etc associated with a task group */
8551 void sched_destroy_group(struct task_group
*tg
)
8553 unsigned long flags
;
8556 /* end participation in shares distribution */
8557 for_each_possible_cpu(i
)
8558 unregister_fair_sched_group(tg
, i
);
8560 spin_lock_irqsave(&task_group_lock
, flags
);
8561 list_del_rcu(&tg
->list
);
8562 list_del_rcu(&tg
->siblings
);
8563 spin_unlock_irqrestore(&task_group_lock
, flags
);
8565 /* wait for possible concurrent references to cfs_rqs complete */
8566 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8569 /* change task's runqueue when it moves between groups.
8570 * The caller of this function should have put the task in its new group
8571 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8572 * reflect its new group.
8574 void sched_move_task(struct task_struct
*tsk
)
8577 unsigned long flags
;
8580 rq
= task_rq_lock(tsk
, &flags
);
8582 running
= task_current(rq
, tsk
);
8586 dequeue_task(rq
, tsk
, 0);
8587 if (unlikely(running
))
8588 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8590 #ifdef CONFIG_FAIR_GROUP_SCHED
8591 if (tsk
->sched_class
->task_move_group
)
8592 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8595 set_task_rq(tsk
, task_cpu(tsk
));
8597 if (unlikely(running
))
8598 tsk
->sched_class
->set_curr_task(rq
);
8600 enqueue_task(rq
, tsk
, 0);
8602 task_rq_unlock(rq
, tsk
, &flags
);
8604 #endif /* CONFIG_CGROUP_SCHED */
8606 #ifdef CONFIG_FAIR_GROUP_SCHED
8607 static DEFINE_MUTEX(shares_mutex
);
8609 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8612 unsigned long flags
;
8615 * We can't change the weight of the root cgroup.
8620 shares
= clamp(shares
, scale_load(MIN_SHARES
), scale_load(MAX_SHARES
));
8622 mutex_lock(&shares_mutex
);
8623 if (tg
->shares
== shares
)
8626 tg
->shares
= shares
;
8627 for_each_possible_cpu(i
) {
8628 struct rq
*rq
= cpu_rq(i
);
8629 struct sched_entity
*se
;
8632 /* Propagate contribution to hierarchy */
8633 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8634 for_each_sched_entity(se
)
8635 update_cfs_shares(group_cfs_rq(se
));
8636 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8640 mutex_unlock(&shares_mutex
);
8644 unsigned long sched_group_shares(struct task_group
*tg
)
8650 #ifdef CONFIG_RT_GROUP_SCHED
8652 * Ensure that the real time constraints are schedulable.
8654 static DEFINE_MUTEX(rt_constraints_mutex
);
8656 static unsigned long to_ratio(u64 period
, u64 runtime
)
8658 if (runtime
== RUNTIME_INF
)
8661 return div64_u64(runtime
<< 20, period
);
8664 /* Must be called with tasklist_lock held */
8665 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8667 struct task_struct
*g
, *p
;
8669 do_each_thread(g
, p
) {
8670 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8672 } while_each_thread(g
, p
);
8677 struct rt_schedulable_data
{
8678 struct task_group
*tg
;
8683 static int tg_schedulable(struct task_group
*tg
, void *data
)
8685 struct rt_schedulable_data
*d
= data
;
8686 struct task_group
*child
;
8687 unsigned long total
, sum
= 0;
8688 u64 period
, runtime
;
8690 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8691 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8694 period
= d
->rt_period
;
8695 runtime
= d
->rt_runtime
;
8699 * Cannot have more runtime than the period.
8701 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8705 * Ensure we don't starve existing RT tasks.
8707 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8710 total
= to_ratio(period
, runtime
);
8713 * Nobody can have more than the global setting allows.
8715 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8719 * The sum of our children's runtime should not exceed our own.
8721 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8722 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8723 runtime
= child
->rt_bandwidth
.rt_runtime
;
8725 if (child
== d
->tg
) {
8726 period
= d
->rt_period
;
8727 runtime
= d
->rt_runtime
;
8730 sum
+= to_ratio(period
, runtime
);
8739 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8741 struct rt_schedulable_data data
= {
8743 .rt_period
= period
,
8744 .rt_runtime
= runtime
,
8747 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8750 static int tg_set_bandwidth(struct task_group
*tg
,
8751 u64 rt_period
, u64 rt_runtime
)
8755 mutex_lock(&rt_constraints_mutex
);
8756 read_lock(&tasklist_lock
);
8757 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8761 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8762 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8763 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8765 for_each_possible_cpu(i
) {
8766 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8768 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8769 rt_rq
->rt_runtime
= rt_runtime
;
8770 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8772 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8774 read_unlock(&tasklist_lock
);
8775 mutex_unlock(&rt_constraints_mutex
);
8780 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8782 u64 rt_runtime
, rt_period
;
8784 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8785 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8786 if (rt_runtime_us
< 0)
8787 rt_runtime
= RUNTIME_INF
;
8789 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8792 long sched_group_rt_runtime(struct task_group
*tg
)
8796 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8799 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8800 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8801 return rt_runtime_us
;
8804 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8806 u64 rt_runtime
, rt_period
;
8808 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8809 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8814 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8817 long sched_group_rt_period(struct task_group
*tg
)
8821 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8822 do_div(rt_period_us
, NSEC_PER_USEC
);
8823 return rt_period_us
;
8826 static int sched_rt_global_constraints(void)
8828 u64 runtime
, period
;
8831 if (sysctl_sched_rt_period
<= 0)
8834 runtime
= global_rt_runtime();
8835 period
= global_rt_period();
8838 * Sanity check on the sysctl variables.
8840 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8843 mutex_lock(&rt_constraints_mutex
);
8844 read_lock(&tasklist_lock
);
8845 ret
= __rt_schedulable(NULL
, 0, 0);
8846 read_unlock(&tasklist_lock
);
8847 mutex_unlock(&rt_constraints_mutex
);
8852 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8854 /* Don't accept realtime tasks when there is no way for them to run */
8855 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8861 #else /* !CONFIG_RT_GROUP_SCHED */
8862 static int sched_rt_global_constraints(void)
8864 unsigned long flags
;
8867 if (sysctl_sched_rt_period
<= 0)
8871 * There's always some RT tasks in the root group
8872 * -- migration, kstopmachine etc..
8874 if (sysctl_sched_rt_runtime
== 0)
8877 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8878 for_each_possible_cpu(i
) {
8879 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8881 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8882 rt_rq
->rt_runtime
= global_rt_runtime();
8883 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8885 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8889 #endif /* CONFIG_RT_GROUP_SCHED */
8891 int sched_rt_handler(struct ctl_table
*table
, int write
,
8892 void __user
*buffer
, size_t *lenp
,
8896 int old_period
, old_runtime
;
8897 static DEFINE_MUTEX(mutex
);
8900 old_period
= sysctl_sched_rt_period
;
8901 old_runtime
= sysctl_sched_rt_runtime
;
8903 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8905 if (!ret
&& write
) {
8906 ret
= sched_rt_global_constraints();
8908 sysctl_sched_rt_period
= old_period
;
8909 sysctl_sched_rt_runtime
= old_runtime
;
8911 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8912 def_rt_bandwidth
.rt_period
=
8913 ns_to_ktime(global_rt_period());
8916 mutex_unlock(&mutex
);
8921 #ifdef CONFIG_CGROUP_SCHED
8923 /* return corresponding task_group object of a cgroup */
8924 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8926 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8927 struct task_group
, css
);
8930 static struct cgroup_subsys_state
*
8931 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8933 struct task_group
*tg
, *parent
;
8935 if (!cgrp
->parent
) {
8936 /* This is early initialization for the top cgroup */
8937 return &root_task_group
.css
;
8940 parent
= cgroup_tg(cgrp
->parent
);
8941 tg
= sched_create_group(parent
);
8943 return ERR_PTR(-ENOMEM
);
8949 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8951 struct task_group
*tg
= cgroup_tg(cgrp
);
8953 sched_destroy_group(tg
);
8957 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8959 #ifdef CONFIG_RT_GROUP_SCHED
8960 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8963 /* We don't support RT-tasks being in separate groups */
8964 if (tsk
->sched_class
!= &fair_sched_class
)
8971 cpu_cgroup_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8973 sched_move_task(tsk
);
8977 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8978 struct cgroup
*old_cgrp
, struct task_struct
*task
)
8981 * cgroup_exit() is called in the copy_process() failure path.
8982 * Ignore this case since the task hasn't ran yet, this avoids
8983 * trying to poke a half freed task state from generic code.
8985 if (!(task
->flags
& PF_EXITING
))
8988 sched_move_task(task
);
8991 #ifdef CONFIG_FAIR_GROUP_SCHED
8992 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8995 return sched_group_set_shares(cgroup_tg(cgrp
), scale_load(shareval
));
8998 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9000 struct task_group
*tg
= cgroup_tg(cgrp
);
9002 return (u64
) scale_load_down(tg
->shares
);
9004 #endif /* CONFIG_FAIR_GROUP_SCHED */
9006 #ifdef CONFIG_RT_GROUP_SCHED
9007 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9010 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9013 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9015 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9018 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9021 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9024 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9026 return sched_group_rt_period(cgroup_tg(cgrp
));
9028 #endif /* CONFIG_RT_GROUP_SCHED */
9030 static struct cftype cpu_files
[] = {
9031 #ifdef CONFIG_FAIR_GROUP_SCHED
9034 .read_u64
= cpu_shares_read_u64
,
9035 .write_u64
= cpu_shares_write_u64
,
9038 #ifdef CONFIG_RT_GROUP_SCHED
9040 .name
= "rt_runtime_us",
9041 .read_s64
= cpu_rt_runtime_read
,
9042 .write_s64
= cpu_rt_runtime_write
,
9045 .name
= "rt_period_us",
9046 .read_u64
= cpu_rt_period_read_uint
,
9047 .write_u64
= cpu_rt_period_write_uint
,
9052 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9054 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9057 struct cgroup_subsys cpu_cgroup_subsys
= {
9059 .create
= cpu_cgroup_create
,
9060 .destroy
= cpu_cgroup_destroy
,
9061 .can_attach_task
= cpu_cgroup_can_attach_task
,
9062 .attach_task
= cpu_cgroup_attach_task
,
9063 .exit
= cpu_cgroup_exit
,
9064 .populate
= cpu_cgroup_populate
,
9065 .subsys_id
= cpu_cgroup_subsys_id
,
9069 #endif /* CONFIG_CGROUP_SCHED */
9071 #ifdef CONFIG_CGROUP_CPUACCT
9074 * CPU accounting code for task groups.
9076 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9077 * (balbir@in.ibm.com).
9080 /* track cpu usage of a group of tasks and its child groups */
9082 struct cgroup_subsys_state css
;
9083 /* cpuusage holds pointer to a u64-type object on every cpu */
9084 u64 __percpu
*cpuusage
;
9085 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9086 struct cpuacct
*parent
;
9089 struct cgroup_subsys cpuacct_subsys
;
9091 /* return cpu accounting group corresponding to this container */
9092 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9094 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9095 struct cpuacct
, css
);
9098 /* return cpu accounting group to which this task belongs */
9099 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9101 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9102 struct cpuacct
, css
);
9105 /* create a new cpu accounting group */
9106 static struct cgroup_subsys_state
*cpuacct_create(
9107 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9109 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9115 ca
->cpuusage
= alloc_percpu(u64
);
9119 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9120 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9121 goto out_free_counters
;
9124 ca
->parent
= cgroup_ca(cgrp
->parent
);
9130 percpu_counter_destroy(&ca
->cpustat
[i
]);
9131 free_percpu(ca
->cpuusage
);
9135 return ERR_PTR(-ENOMEM
);
9138 /* destroy an existing cpu accounting group */
9140 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9142 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9145 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9146 percpu_counter_destroy(&ca
->cpustat
[i
]);
9147 free_percpu(ca
->cpuusage
);
9151 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9153 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9156 #ifndef CONFIG_64BIT
9158 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9160 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9162 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9170 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9172 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9174 #ifndef CONFIG_64BIT
9176 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9178 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9180 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9186 /* return total cpu usage (in nanoseconds) of a group */
9187 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9189 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9190 u64 totalcpuusage
= 0;
9193 for_each_present_cpu(i
)
9194 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9196 return totalcpuusage
;
9199 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9202 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9211 for_each_present_cpu(i
)
9212 cpuacct_cpuusage_write(ca
, i
, 0);
9218 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9221 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9225 for_each_present_cpu(i
) {
9226 percpu
= cpuacct_cpuusage_read(ca
, i
);
9227 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9229 seq_printf(m
, "\n");
9233 static const char *cpuacct_stat_desc
[] = {
9234 [CPUACCT_STAT_USER
] = "user",
9235 [CPUACCT_STAT_SYSTEM
] = "system",
9238 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9239 struct cgroup_map_cb
*cb
)
9241 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9244 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9245 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9246 val
= cputime64_to_clock_t(val
);
9247 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9252 static struct cftype files
[] = {
9255 .read_u64
= cpuusage_read
,
9256 .write_u64
= cpuusage_write
,
9259 .name
= "usage_percpu",
9260 .read_seq_string
= cpuacct_percpu_seq_read
,
9264 .read_map
= cpuacct_stats_show
,
9268 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9270 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9274 * charge this task's execution time to its accounting group.
9276 * called with rq->lock held.
9278 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9283 if (unlikely(!cpuacct_subsys
.active
))
9286 cpu
= task_cpu(tsk
);
9292 for (; ca
; ca
= ca
->parent
) {
9293 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9294 *cpuusage
+= cputime
;
9301 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9302 * in cputime_t units. As a result, cpuacct_update_stats calls
9303 * percpu_counter_add with values large enough to always overflow the
9304 * per cpu batch limit causing bad SMP scalability.
9306 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9307 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9308 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9311 #define CPUACCT_BATCH \
9312 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9314 #define CPUACCT_BATCH 0
9318 * Charge the system/user time to the task's accounting group.
9320 static void cpuacct_update_stats(struct task_struct
*tsk
,
9321 enum cpuacct_stat_index idx
, cputime_t val
)
9324 int batch
= CPUACCT_BATCH
;
9326 if (unlikely(!cpuacct_subsys
.active
))
9333 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9339 struct cgroup_subsys cpuacct_subsys
= {
9341 .create
= cpuacct_create
,
9342 .destroy
= cpuacct_destroy
,
9343 .populate
= cpuacct_populate
,
9344 .subsys_id
= cpuacct_subsys_id
,
9346 #endif /* CONFIG_CGROUP_CPUACCT */