4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy
)
126 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
131 static inline int task_has_rt_policy(struct task_struct
*p
)
133 return rt_policy(p
->policy
);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array
{
140 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
141 struct list_head queue
[MAX_RT_PRIO
];
144 struct rt_bandwidth
{
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock
;
149 struct hrtimer rt_period_timer
;
152 static struct rt_bandwidth def_rt_bandwidth
;
154 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
156 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
158 struct rt_bandwidth
*rt_b
=
159 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
165 now
= hrtimer_cb_get_time(timer
);
166 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
171 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
174 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
178 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
180 rt_b
->rt_period
= ns_to_ktime(period
);
181 rt_b
->rt_runtime
= runtime
;
183 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
185 hrtimer_init(&rt_b
->rt_period_timer
,
186 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
187 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime
>= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
199 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
202 if (hrtimer_active(&rt_b
->rt_period_timer
))
205 raw_spin_lock(&rt_b
->rt_runtime_lock
);
210 if (hrtimer_active(&rt_b
->rt_period_timer
))
213 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
214 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
216 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
217 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
218 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
219 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
220 HRTIMER_MODE_ABS_PINNED
, 0);
222 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
228 hrtimer_cancel(&rt_b
->rt_period_timer
);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex
);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups
);
246 /* task group related information */
248 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity
**se
;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq
**cfs_rq
;
255 unsigned long shares
;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity
**rt_se
;
260 struct rt_rq
**rt_rq
;
262 struct rt_bandwidth rt_bandwidth
;
266 struct list_head list
;
268 struct task_group
*parent
;
269 struct list_head siblings
;
270 struct list_head children
;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock
);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group
.children
);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group
;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load
;
315 unsigned long nr_running
;
320 struct rb_root tasks_timeline
;
321 struct rb_node
*rb_leftmost
;
323 struct list_head tasks
;
324 struct list_head
*balance_iterator
;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity
*curr
, *next
, *last
;
332 unsigned int nr_spread_over
;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list
;
346 struct task_group
*tg
; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight
;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load
;
363 * this cpu's part of tg->shares
365 unsigned long shares
;
368 * load.weight at the time we set shares
370 unsigned long rq_weight
;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active
;
378 unsigned long rt_nr_running
;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr
; /* highest queued rt task prio */
383 int next
; /* next highest */
388 unsigned long rt_nr_migratory
;
389 unsigned long rt_nr_total
;
391 struct plist_head pushable_tasks
;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock
;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted
;
403 struct list_head leaf_rt_rq_list
;
404 struct task_group
*tg
;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online
;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask
;
430 struct cpupri cpupri
;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain
;
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running
;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
460 unsigned long last_load_update_tick
;
463 unsigned char nohz_balance_kick
;
465 unsigned int skip_clock_update
;
467 /* capture load from *all* tasks on this cpu: */
468 struct load_weight load
;
469 unsigned long nr_load_updates
;
475 #ifdef CONFIG_FAIR_GROUP_SCHED
476 /* list of leaf cfs_rq on this cpu: */
477 struct list_head leaf_cfs_rq_list
;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 struct list_head leaf_rt_rq_list
;
484 * This is part of a global counter where only the total sum
485 * over all CPUs matters. A task can increase this counter on
486 * one CPU and if it got migrated afterwards it may decrease
487 * it on another CPU. Always updated under the runqueue lock:
489 unsigned long nr_uninterruptible
;
491 struct task_struct
*curr
, *idle
;
492 unsigned long next_balance
;
493 struct mm_struct
*prev_mm
;
500 struct root_domain
*rd
;
501 struct sched_domain
*sd
;
503 unsigned long cpu_power
;
505 unsigned char idle_at_tick
;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work
;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task
;
523 /* calc_load related fields */
524 unsigned long calc_load_update
;
525 long calc_load_active
;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending
;
530 struct call_single_data hrtick_csd
;
532 struct hrtimer hrtick_timer
;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info
;
538 unsigned long long rq_cpu_time
;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count
;
544 /* schedule() stats */
545 unsigned int sched_switch
;
546 unsigned int sched_count
;
547 unsigned int sched_goidle
;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count
;
551 unsigned int ttwu_local
;
554 unsigned int bkl_count
;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
561 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
563 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p
))
570 rq
->skip_clock_update
= 1;
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_sched_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
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group
*task_group(struct task_struct
*p
)
615 struct cgroup_subsys_state
*css
;
617 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
618 lockdep_is_held(&task_rq(p
)->lock
));
619 return container_of(css
, struct task_group
, css
);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
627 p
->se
.parent
= task_group(p
)->se
[cpu
];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
632 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
639 static inline struct task_group
*task_group(struct task_struct
*p
)
644 #endif /* CONFIG_CGROUP_SCHED */
646 inline void update_rq_clock(struct rq
*rq
)
648 if (!rq
->skip_clock_update
)
649 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
663 * @cpu: the processor in question.
665 * Returns true if the current cpu runqueue is locked.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu
)
671 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug
unsigned int sysctl_sched_features
=
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly
char *sched_feat_names
[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file
*m
, void *v
)
711 for (i
= 0; sched_feat_names
[i
]; i
++) {
712 if (!(sysctl_sched_features
& (1UL << i
)))
714 seq_printf(m
, "%s ", sched_feat_names
[i
]);
722 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
723 size_t cnt
, loff_t
*ppos
)
733 if (copy_from_user(&buf
, ubuf
, cnt
))
739 if (strncmp(buf
, "NO_", 3) == 0) {
744 for (i
= 0; sched_feat_names
[i
]; i
++) {
745 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
747 sysctl_sched_features
&= ~(1UL << i
);
749 sysctl_sched_features
|= (1UL << i
);
754 if (!sched_feat_names
[i
])
762 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
764 return single_open(filp
, sched_feat_show
, NULL
);
767 static const struct file_operations sched_feat_fops
= {
768 .open
= sched_feat_open
,
769 .write
= sched_feat_write
,
772 .release
= single_release
,
775 static __init
int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
782 late_initcall(sched_init_debug
);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
795 * ratelimit for updating the group shares.
798 unsigned int sysctl_sched_shares_ratelimit
= 250000;
799 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
802 * Inject some fuzzyness into changing the per-cpu group shares
803 * this avoids remote rq-locks at the expense of fairness.
806 unsigned int sysctl_sched_shares_thresh
= 4;
809 * period over which we average the RT time consumption, measured
814 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
817 * period over which we measure -rt task cpu usage in us.
820 unsigned int sysctl_sched_rt_period
= 1000000;
822 static __read_mostly
int scheduler_running
;
825 * part of the period that we allow rt tasks to run in us.
828 int sysctl_sched_rt_runtime
= 950000;
830 static inline u64
global_rt_period(void)
832 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
835 static inline u64
global_rt_runtime(void)
837 if (sysctl_sched_rt_runtime
< 0)
840 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
843 #ifndef prepare_arch_switch
844 # define prepare_arch_switch(next) do { } while (0)
846 #ifndef finish_arch_switch
847 # define finish_arch_switch(prev) do { } while (0)
850 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
852 return rq
->curr
== p
;
855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
856 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
858 return task_current(rq
, p
);
861 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
865 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
867 #ifdef CONFIG_DEBUG_SPINLOCK
868 /* this is a valid case when another task releases the spinlock */
869 rq
->lock
.owner
= current
;
872 * If we are tracking spinlock dependencies then we have to
873 * fix up the runqueue lock - which gets 'carried over' from
876 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
878 raw_spin_unlock_irq(&rq
->lock
);
881 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
882 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
887 return task_current(rq
, p
);
891 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
895 * We can optimise this out completely for !SMP, because the
896 * SMP rebalancing from interrupt is the only thing that cares
901 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 raw_spin_unlock_irq(&rq
->lock
);
904 raw_spin_unlock(&rq
->lock
);
908 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
912 * After ->oncpu is cleared, the task can be moved to a different CPU.
913 * We must ensure this doesn't happen until the switch is completely
919 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
926 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
929 static inline int task_is_waking(struct task_struct
*p
)
931 return unlikely(p
->state
== TASK_WAKING
);
935 * __task_rq_lock - lock the runqueue a given task resides on.
936 * Must be called interrupts disabled.
938 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
945 raw_spin_lock(&rq
->lock
);
946 if (likely(rq
== task_rq(p
)))
948 raw_spin_unlock(&rq
->lock
);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
963 local_irq_save(*flags
);
965 raw_spin_lock(&rq
->lock
);
966 if (likely(rq
== task_rq(p
)))
968 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
972 static void __task_rq_unlock(struct rq
*rq
)
975 raw_spin_unlock(&rq
->lock
);
978 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
981 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq
*this_rq_lock(void)
994 raw_spin_lock(&rq
->lock
);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq
*rq
)
1018 if (!sched_feat(HRTICK
))
1020 if (!cpu_active(cpu_of(rq
)))
1022 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1025 static void hrtick_clear(struct rq
*rq
)
1027 if (hrtimer_active(&rq
->hrtick_timer
))
1028 hrtimer_cancel(&rq
->hrtick_timer
);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1037 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1039 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1041 raw_spin_lock(&rq
->lock
);
1042 update_rq_clock(rq
);
1043 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1044 raw_spin_unlock(&rq
->lock
);
1046 return HRTIMER_NORESTART
;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg
)
1055 struct rq
*rq
= arg
;
1057 raw_spin_lock(&rq
->lock
);
1058 hrtimer_restart(&rq
->hrtick_timer
);
1059 rq
->hrtick_csd_pending
= 0;
1060 raw_spin_unlock(&rq
->lock
);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq
*rq
, u64 delay
)
1070 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1071 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1073 hrtimer_set_expires(timer
, time
);
1075 if (rq
== this_rq()) {
1076 hrtimer_restart(timer
);
1077 } else if (!rq
->hrtick_csd_pending
) {
1078 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1079 rq
->hrtick_csd_pending
= 1;
1084 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1086 int cpu
= (int)(long)hcpu
;
1089 case CPU_UP_CANCELED
:
1090 case CPU_UP_CANCELED_FROZEN
:
1091 case CPU_DOWN_PREPARE
:
1092 case CPU_DOWN_PREPARE_FROZEN
:
1094 case CPU_DEAD_FROZEN
:
1095 hrtick_clear(cpu_rq(cpu
));
1102 static __init
void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick
, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq
*rq
, u64 delay
)
1114 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1115 HRTIMER_MODE_REL_PINNED
, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq
*rq
)
1126 rq
->hrtick_csd_pending
= 0;
1128 rq
->hrtick_csd
.flags
= 0;
1129 rq
->hrtick_csd
.func
= __hrtick_start
;
1130 rq
->hrtick_csd
.info
= rq
;
1133 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1134 rq
->hrtick_timer
.function
= hrtick
;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq
*rq
)
1141 static inline void init_rq_hrtick(struct rq
*rq
)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct
*p
)
1167 assert_raw_spin_locked(&task_rq(p
)->lock
);
1169 if (test_tsk_need_resched(p
))
1172 set_tsk_need_resched(p
);
1175 if (cpu
== smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p
))
1181 smp_send_reschedule(cpu
);
1184 static void resched_cpu(int cpu
)
1186 struct rq
*rq
= cpu_rq(cpu
);
1187 unsigned long flags
;
1189 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1191 resched_task(cpu_curr(cpu
));
1192 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1197 * In the semi idle case, use the nearest busy cpu for migrating timers
1198 * from an idle cpu. This is good for power-savings.
1200 * We don't do similar optimization for completely idle system, as
1201 * selecting an idle cpu will add more delays to the timers than intended
1202 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1204 int get_nohz_timer_target(void)
1206 int cpu
= smp_processor_id();
1208 struct sched_domain
*sd
;
1210 for_each_domain(cpu
, sd
) {
1211 for_each_cpu(i
, sched_domain_span(sd
))
1218 * When add_timer_on() enqueues a timer into the timer wheel of an
1219 * idle CPU then this timer might expire before the next timer event
1220 * which is scheduled to wake up that CPU. In case of a completely
1221 * idle system the next event might even be infinite time into the
1222 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1223 * leaves the inner idle loop so the newly added timer is taken into
1224 * account when the CPU goes back to idle and evaluates the timer
1225 * wheel for the next timer event.
1227 void wake_up_idle_cpu(int cpu
)
1229 struct rq
*rq
= cpu_rq(cpu
);
1231 if (cpu
== smp_processor_id())
1235 * This is safe, as this function is called with the timer
1236 * wheel base lock of (cpu) held. When the CPU is on the way
1237 * to idle and has not yet set rq->curr to idle then it will
1238 * be serialized on the timer wheel base lock and take the new
1239 * timer into account automatically.
1241 if (rq
->curr
!= rq
->idle
)
1245 * We can set TIF_RESCHED on the idle task of the other CPU
1246 * lockless. The worst case is that the other CPU runs the
1247 * idle task through an additional NOOP schedule()
1249 set_tsk_need_resched(rq
->idle
);
1251 /* NEED_RESCHED must be visible before we test polling */
1253 if (!tsk_is_polling(rq
->idle
))
1254 smp_send_reschedule(cpu
);
1257 #endif /* CONFIG_NO_HZ */
1259 static u64
sched_avg_period(void)
1261 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1264 static void sched_avg_update(struct rq
*rq
)
1266 s64 period
= sched_avg_period();
1268 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1270 * Inline assembly required to prevent the compiler
1271 * optimising this loop into a divmod call.
1272 * See __iter_div_u64_rem() for another example of this.
1274 asm("" : "+rm" (rq
->age_stamp
));
1275 rq
->age_stamp
+= period
;
1280 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1282 rq
->rt_avg
+= rt_delta
;
1283 sched_avg_update(rq
);
1286 #else /* !CONFIG_SMP */
1287 static void resched_task(struct task_struct
*p
)
1289 assert_raw_spin_locked(&task_rq(p
)->lock
);
1290 set_tsk_need_resched(p
);
1293 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1297 static void sched_avg_update(struct rq
*rq
)
1300 #endif /* CONFIG_SMP */
1302 #if BITS_PER_LONG == 32
1303 # define WMULT_CONST (~0UL)
1305 # define WMULT_CONST (1UL << 32)
1308 #define WMULT_SHIFT 32
1311 * Shift right and round:
1313 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1316 * delta *= weight / lw
1318 static unsigned long
1319 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1320 struct load_weight
*lw
)
1324 if (!lw
->inv_weight
) {
1325 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1328 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1332 tmp
= (u64
)delta_exec
* weight
;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp
> WMULT_CONST
))
1337 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1340 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1342 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1345 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1351 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1358 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1359 * of tasks with abnormal "nice" values across CPUs the contribution that
1360 * each task makes to its run queue's load is weighted according to its
1361 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1362 * scaled version of the new time slice allocation that they receive on time
1366 #define WEIGHT_IDLEPRIO 3
1367 #define WMULT_IDLEPRIO 1431655765
1370 * Nice levels are multiplicative, with a gentle 10% change for every
1371 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1372 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1373 * that remained on nice 0.
1375 * The "10% effect" is relative and cumulative: from _any_ nice level,
1376 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1377 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1378 * If a task goes up by ~10% and another task goes down by ~10% then
1379 * the relative distance between them is ~25%.)
1381 static const int prio_to_weight
[40] = {
1382 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1383 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1384 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1385 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1386 /* 0 */ 1024, 820, 655, 526, 423,
1387 /* 5 */ 335, 272, 215, 172, 137,
1388 /* 10 */ 110, 87, 70, 56, 45,
1389 /* 15 */ 36, 29, 23, 18, 15,
1393 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1395 * In cases where the weight does not change often, we can use the
1396 * precalculated inverse to speed up arithmetics by turning divisions
1397 * into multiplications:
1399 static const u32 prio_to_wmult
[40] = {
1400 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1401 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1402 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1403 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1404 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1405 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1406 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1407 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1410 /* Time spent by the tasks of the cpu accounting group executing in ... */
1411 enum cpuacct_stat_index
{
1412 CPUACCT_STAT_USER
, /* ... user mode */
1413 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1415 CPUACCT_STAT_NSTATS
,
1418 #ifdef CONFIG_CGROUP_CPUACCT
1419 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1420 static void cpuacct_update_stats(struct task_struct
*tsk
,
1421 enum cpuacct_stat_index idx
, cputime_t val
);
1423 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1424 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1425 enum cpuacct_stat_index idx
, cputime_t val
) {}
1428 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1430 update_load_add(&rq
->load
, load
);
1433 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1435 update_load_sub(&rq
->load
, load
);
1438 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1439 typedef int (*tg_visitor
)(struct task_group
*, void *);
1442 * Iterate the full tree, calling @down when first entering a node and @up when
1443 * leaving it for the final time.
1445 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1447 struct task_group
*parent
, *child
;
1451 parent
= &root_task_group
;
1453 ret
= (*down
)(parent
, data
);
1456 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1463 ret
= (*up
)(parent
, data
);
1468 parent
= parent
->parent
;
1477 static int tg_nop(struct task_group
*tg
, void *data
)
1484 /* Used instead of source_load when we know the type == 0 */
1485 static unsigned long weighted_cpuload(const int cpu
)
1487 return cpu_rq(cpu
)->load
.weight
;
1491 * Return a low guess at the load of a migration-source cpu weighted
1492 * according to the scheduling class and "nice" value.
1494 * We want to under-estimate the load of migration sources, to
1495 * balance conservatively.
1497 static unsigned long source_load(int cpu
, int type
)
1499 struct rq
*rq
= cpu_rq(cpu
);
1500 unsigned long total
= weighted_cpuload(cpu
);
1502 if (type
== 0 || !sched_feat(LB_BIAS
))
1505 return min(rq
->cpu_load
[type
-1], total
);
1509 * Return a high guess at the load of a migration-target cpu weighted
1510 * according to the scheduling class and "nice" value.
1512 static unsigned long target_load(int cpu
, int type
)
1514 struct rq
*rq
= cpu_rq(cpu
);
1515 unsigned long total
= weighted_cpuload(cpu
);
1517 if (type
== 0 || !sched_feat(LB_BIAS
))
1520 return max(rq
->cpu_load
[type
-1], total
);
1523 static unsigned long power_of(int cpu
)
1525 return cpu_rq(cpu
)->cpu_power
;
1528 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1530 static unsigned long cpu_avg_load_per_task(int cpu
)
1532 struct rq
*rq
= cpu_rq(cpu
);
1533 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1536 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1538 rq
->avg_load_per_task
= 0;
1540 return rq
->avg_load_per_task
;
1543 #ifdef CONFIG_FAIR_GROUP_SCHED
1545 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1547 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1550 * Calculate and set the cpu's group shares.
1552 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1553 unsigned long sd_shares
,
1554 unsigned long sd_rq_weight
,
1555 unsigned long *usd_rq_weight
)
1557 unsigned long shares
, rq_weight
;
1560 rq_weight
= usd_rq_weight
[cpu
];
1563 rq_weight
= NICE_0_LOAD
;
1567 * \Sum_j shares_j * rq_weight_i
1568 * shares_i = -----------------------------
1569 * \Sum_j rq_weight_j
1571 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1572 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1574 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1575 sysctl_sched_shares_thresh
) {
1576 struct rq
*rq
= cpu_rq(cpu
);
1577 unsigned long flags
;
1579 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1580 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1581 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1582 __set_se_shares(tg
->se
[cpu
], shares
);
1583 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1588 * Re-compute the task group their per cpu shares over the given domain.
1589 * This needs to be done in a bottom-up fashion because the rq weight of a
1590 * parent group depends on the shares of its child groups.
1592 static int tg_shares_up(struct task_group
*tg
, void *data
)
1594 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1595 unsigned long *usd_rq_weight
;
1596 struct sched_domain
*sd
= data
;
1597 unsigned long flags
;
1603 local_irq_save(flags
);
1604 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1606 for_each_cpu(i
, sched_domain_span(sd
)) {
1607 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1608 usd_rq_weight
[i
] = weight
;
1610 rq_weight
+= weight
;
1612 * If there are currently no tasks on the cpu pretend there
1613 * is one of average load so that when a new task gets to
1614 * run here it will not get delayed by group starvation.
1617 weight
= NICE_0_LOAD
;
1619 sum_weight
+= weight
;
1620 shares
+= tg
->cfs_rq
[i
]->shares
;
1624 rq_weight
= sum_weight
;
1626 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1627 shares
= tg
->shares
;
1629 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1630 shares
= tg
->shares
;
1632 for_each_cpu(i
, sched_domain_span(sd
))
1633 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1635 local_irq_restore(flags
);
1641 * Compute the cpu's hierarchical load factor for each task group.
1642 * This needs to be done in a top-down fashion because the load of a child
1643 * group is a fraction of its parents load.
1645 static int tg_load_down(struct task_group
*tg
, void *data
)
1648 long cpu
= (long)data
;
1651 load
= cpu_rq(cpu
)->load
.weight
;
1653 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1654 load
*= tg
->cfs_rq
[cpu
]->shares
;
1655 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1658 tg
->cfs_rq
[cpu
]->h_load
= load
;
1663 static void update_shares(struct sched_domain
*sd
)
1668 if (root_task_group_empty())
1671 now
= local_clock();
1672 elapsed
= now
- sd
->last_update
;
1674 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1675 sd
->last_update
= now
;
1676 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1680 static void update_h_load(long cpu
)
1682 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1687 static inline void update_shares(struct sched_domain
*sd
)
1693 #ifdef CONFIG_PREEMPT
1695 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1698 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1699 * way at the expense of forcing extra atomic operations in all
1700 * invocations. This assures that the double_lock is acquired using the
1701 * same underlying policy as the spinlock_t on this architecture, which
1702 * reduces latency compared to the unfair variant below. However, it
1703 * also adds more overhead and therefore may reduce throughput.
1705 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1706 __releases(this_rq
->lock
)
1707 __acquires(busiest
->lock
)
1708 __acquires(this_rq
->lock
)
1710 raw_spin_unlock(&this_rq
->lock
);
1711 double_rq_lock(this_rq
, busiest
);
1718 * Unfair double_lock_balance: Optimizes throughput at the expense of
1719 * latency by eliminating extra atomic operations when the locks are
1720 * already in proper order on entry. This favors lower cpu-ids and will
1721 * grant the double lock to lower cpus over higher ids under contention,
1722 * regardless of entry order into the function.
1724 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1725 __releases(this_rq
->lock
)
1726 __acquires(busiest
->lock
)
1727 __acquires(this_rq
->lock
)
1731 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1732 if (busiest
< this_rq
) {
1733 raw_spin_unlock(&this_rq
->lock
);
1734 raw_spin_lock(&busiest
->lock
);
1735 raw_spin_lock_nested(&this_rq
->lock
,
1736 SINGLE_DEPTH_NESTING
);
1739 raw_spin_lock_nested(&busiest
->lock
,
1740 SINGLE_DEPTH_NESTING
);
1745 #endif /* CONFIG_PREEMPT */
1748 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1750 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1752 if (unlikely(!irqs_disabled())) {
1753 /* printk() doesn't work good under rq->lock */
1754 raw_spin_unlock(&this_rq
->lock
);
1758 return _double_lock_balance(this_rq
, busiest
);
1761 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1762 __releases(busiest
->lock
)
1764 raw_spin_unlock(&busiest
->lock
);
1765 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1769 * double_rq_lock - safely lock two runqueues
1771 * Note this does not disable interrupts like task_rq_lock,
1772 * you need to do so manually before calling.
1774 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1775 __acquires(rq1
->lock
)
1776 __acquires(rq2
->lock
)
1778 BUG_ON(!irqs_disabled());
1780 raw_spin_lock(&rq1
->lock
);
1781 __acquire(rq2
->lock
); /* Fake it out ;) */
1784 raw_spin_lock(&rq1
->lock
);
1785 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1787 raw_spin_lock(&rq2
->lock
);
1788 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1794 * double_rq_unlock - safely unlock two runqueues
1796 * Note this does not restore interrupts like task_rq_unlock,
1797 * you need to do so manually after calling.
1799 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1800 __releases(rq1
->lock
)
1801 __releases(rq2
->lock
)
1803 raw_spin_unlock(&rq1
->lock
);
1805 raw_spin_unlock(&rq2
->lock
);
1807 __release(rq2
->lock
);
1812 #ifdef CONFIG_FAIR_GROUP_SCHED
1813 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1816 cfs_rq
->shares
= shares
;
1821 static void calc_load_account_idle(struct rq
*this_rq
);
1822 static void update_sysctl(void);
1823 static int get_update_sysctl_factor(void);
1824 static void update_cpu_load(struct rq
*this_rq
);
1826 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1828 set_task_rq(p
, cpu
);
1831 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1832 * successfuly executed on another CPU. We must ensure that updates of
1833 * per-task data have been completed by this moment.
1836 task_thread_info(p
)->cpu
= cpu
;
1840 static const struct sched_class rt_sched_class
;
1842 #define sched_class_highest (&rt_sched_class)
1843 #define for_each_class(class) \
1844 for (class = sched_class_highest; class; class = class->next)
1846 #include "sched_stats.h"
1848 static void inc_nr_running(struct rq
*rq
)
1853 static void dec_nr_running(struct rq
*rq
)
1858 static void set_load_weight(struct task_struct
*p
)
1861 * SCHED_IDLE tasks get minimal weight:
1863 if (p
->policy
== SCHED_IDLE
) {
1864 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1865 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1869 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1870 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1873 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1875 update_rq_clock(rq
);
1876 sched_info_queued(p
);
1877 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1881 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1883 update_rq_clock(rq
);
1884 sched_info_dequeued(p
);
1885 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1890 * activate_task - move a task to the runqueue.
1892 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1894 if (task_contributes_to_load(p
))
1895 rq
->nr_uninterruptible
--;
1897 enqueue_task(rq
, p
, flags
);
1902 * deactivate_task - remove a task from the runqueue.
1904 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1906 if (task_contributes_to_load(p
))
1907 rq
->nr_uninterruptible
++;
1909 dequeue_task(rq
, p
, flags
);
1913 #include "sched_idletask.c"
1914 #include "sched_fair.c"
1915 #include "sched_rt.c"
1916 #ifdef CONFIG_SCHED_DEBUG
1917 # include "sched_debug.c"
1921 * __normal_prio - return the priority that is based on the static prio
1923 static inline int __normal_prio(struct task_struct
*p
)
1925 return p
->static_prio
;
1929 * Calculate the expected normal priority: i.e. priority
1930 * without taking RT-inheritance into account. Might be
1931 * boosted by interactivity modifiers. Changes upon fork,
1932 * setprio syscalls, and whenever the interactivity
1933 * estimator recalculates.
1935 static inline int normal_prio(struct task_struct
*p
)
1939 if (task_has_rt_policy(p
))
1940 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1942 prio
= __normal_prio(p
);
1947 * Calculate the current priority, i.e. the priority
1948 * taken into account by the scheduler. This value might
1949 * be boosted by RT tasks, or might be boosted by
1950 * interactivity modifiers. Will be RT if the task got
1951 * RT-boosted. If not then it returns p->normal_prio.
1953 static int effective_prio(struct task_struct
*p
)
1955 p
->normal_prio
= normal_prio(p
);
1957 * If we are RT tasks or we were boosted to RT priority,
1958 * keep the priority unchanged. Otherwise, update priority
1959 * to the normal priority:
1961 if (!rt_prio(p
->prio
))
1962 return p
->normal_prio
;
1967 * task_curr - is this task currently executing on a CPU?
1968 * @p: the task in question.
1970 inline int task_curr(const struct task_struct
*p
)
1972 return cpu_curr(task_cpu(p
)) == p
;
1975 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1976 const struct sched_class
*prev_class
,
1977 int oldprio
, int running
)
1979 if (prev_class
!= p
->sched_class
) {
1980 if (prev_class
->switched_from
)
1981 prev_class
->switched_from(rq
, p
, running
);
1982 p
->sched_class
->switched_to(rq
, p
, running
);
1984 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1989 * Is this task likely cache-hot:
1992 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1996 if (p
->sched_class
!= &fair_sched_class
)
2000 * Buddy candidates are cache hot:
2002 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2003 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2004 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2007 if (sysctl_sched_migration_cost
== -1)
2009 if (sysctl_sched_migration_cost
== 0)
2012 delta
= now
- p
->se
.exec_start
;
2014 return delta
< (s64
)sysctl_sched_migration_cost
;
2017 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2019 #ifdef CONFIG_SCHED_DEBUG
2021 * We should never call set_task_cpu() on a blocked task,
2022 * ttwu() will sort out the placement.
2024 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2025 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2028 trace_sched_migrate_task(p
, new_cpu
);
2030 if (task_cpu(p
) != new_cpu
) {
2031 p
->se
.nr_migrations
++;
2032 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2035 __set_task_cpu(p
, new_cpu
);
2038 struct migration_arg
{
2039 struct task_struct
*task
;
2043 static int migration_cpu_stop(void *data
);
2046 * The task's runqueue lock must be held.
2047 * Returns true if you have to wait for migration thread.
2049 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2051 struct rq
*rq
= task_rq(p
);
2054 * If the task is not on a runqueue (and not running), then
2055 * the next wake-up will properly place the task.
2057 return p
->se
.on_rq
|| task_running(rq
, p
);
2061 * wait_task_inactive - wait for a thread to unschedule.
2063 * If @match_state is nonzero, it's the @p->state value just checked and
2064 * not expected to change. If it changes, i.e. @p might have woken up,
2065 * then return zero. When we succeed in waiting for @p to be off its CPU,
2066 * we return a positive number (its total switch count). If a second call
2067 * a short while later returns the same number, the caller can be sure that
2068 * @p has remained unscheduled the whole time.
2070 * The caller must ensure that the task *will* unschedule sometime soon,
2071 * else this function might spin for a *long* time. This function can't
2072 * be called with interrupts off, or it may introduce deadlock with
2073 * smp_call_function() if an IPI is sent by the same process we are
2074 * waiting to become inactive.
2076 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2078 unsigned long flags
;
2085 * We do the initial early heuristics without holding
2086 * any task-queue locks at all. We'll only try to get
2087 * the runqueue lock when things look like they will
2093 * If the task is actively running on another CPU
2094 * still, just relax and busy-wait without holding
2097 * NOTE! Since we don't hold any locks, it's not
2098 * even sure that "rq" stays as the right runqueue!
2099 * But we don't care, since "task_running()" will
2100 * return false if the runqueue has changed and p
2101 * is actually now running somewhere else!
2103 while (task_running(rq
, p
)) {
2104 if (match_state
&& unlikely(p
->state
!= match_state
))
2110 * Ok, time to look more closely! We need the rq
2111 * lock now, to be *sure*. If we're wrong, we'll
2112 * just go back and repeat.
2114 rq
= task_rq_lock(p
, &flags
);
2115 trace_sched_wait_task(p
);
2116 running
= task_running(rq
, p
);
2117 on_rq
= p
->se
.on_rq
;
2119 if (!match_state
|| p
->state
== match_state
)
2120 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2121 task_rq_unlock(rq
, &flags
);
2124 * If it changed from the expected state, bail out now.
2126 if (unlikely(!ncsw
))
2130 * Was it really running after all now that we
2131 * checked with the proper locks actually held?
2133 * Oops. Go back and try again..
2135 if (unlikely(running
)) {
2141 * It's not enough that it's not actively running,
2142 * it must be off the runqueue _entirely_, and not
2145 * So if it was still runnable (but just not actively
2146 * running right now), it's preempted, and we should
2147 * yield - it could be a while.
2149 if (unlikely(on_rq
)) {
2150 schedule_timeout_uninterruptible(1);
2155 * Ahh, all good. It wasn't running, and it wasn't
2156 * runnable, which means that it will never become
2157 * running in the future either. We're all done!
2166 * kick_process - kick a running thread to enter/exit the kernel
2167 * @p: the to-be-kicked thread
2169 * Cause a process which is running on another CPU to enter
2170 * kernel-mode, without any delay. (to get signals handled.)
2172 * NOTE: this function doesnt have to take the runqueue lock,
2173 * because all it wants to ensure is that the remote task enters
2174 * the kernel. If the IPI races and the task has been migrated
2175 * to another CPU then no harm is done and the purpose has been
2178 void kick_process(struct task_struct
*p
)
2184 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2185 smp_send_reschedule(cpu
);
2188 EXPORT_SYMBOL_GPL(kick_process
);
2189 #endif /* CONFIG_SMP */
2192 * task_oncpu_function_call - call a function on the cpu on which a task runs
2193 * @p: the task to evaluate
2194 * @func: the function to be called
2195 * @info: the function call argument
2197 * Calls the function @func when the task is currently running. This might
2198 * be on the current CPU, which just calls the function directly
2200 void task_oncpu_function_call(struct task_struct
*p
,
2201 void (*func
) (void *info
), void *info
)
2208 smp_call_function_single(cpu
, func
, info
, 1);
2214 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2216 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2219 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2221 /* Look for allowed, online CPU in same node. */
2222 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2223 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2226 /* Any allowed, online CPU? */
2227 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2228 if (dest_cpu
< nr_cpu_ids
)
2231 /* No more Mr. Nice Guy. */
2232 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2233 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2235 * Don't tell them about moving exiting tasks or
2236 * kernel threads (both mm NULL), since they never
2239 if (p
->mm
&& printk_ratelimit()) {
2240 printk(KERN_INFO
"process %d (%s) no "
2241 "longer affine to cpu%d\n",
2242 task_pid_nr(p
), p
->comm
, cpu
);
2250 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2253 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2255 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2258 * In order not to call set_task_cpu() on a blocking task we need
2259 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2262 * Since this is common to all placement strategies, this lives here.
2264 * [ this allows ->select_task() to simply return task_cpu(p) and
2265 * not worry about this generic constraint ]
2267 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2269 cpu
= select_fallback_rq(task_cpu(p
), p
);
2274 static void update_avg(u64
*avg
, u64 sample
)
2276 s64 diff
= sample
- *avg
;
2281 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2282 bool is_sync
, bool is_migrate
, bool is_local
,
2283 unsigned long en_flags
)
2285 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2287 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2289 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2291 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2293 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2295 activate_task(rq
, p
, en_flags
);
2298 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2299 int wake_flags
, bool success
)
2301 trace_sched_wakeup(p
, success
);
2302 check_preempt_curr(rq
, p
, wake_flags
);
2304 p
->state
= TASK_RUNNING
;
2306 if (p
->sched_class
->task_woken
)
2307 p
->sched_class
->task_woken(rq
, p
);
2309 if (unlikely(rq
->idle_stamp
)) {
2310 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2311 u64 max
= 2*sysctl_sched_migration_cost
;
2316 update_avg(&rq
->avg_idle
, delta
);
2320 /* if a worker is waking up, notify workqueue */
2321 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2322 wq_worker_waking_up(p
, cpu_of(rq
));
2326 * try_to_wake_up - wake up a thread
2327 * @p: the thread to be awakened
2328 * @state: the mask of task states that can be woken
2329 * @wake_flags: wake modifier flags (WF_*)
2331 * Put it on the run-queue if it's not already there. The "current"
2332 * thread is always on the run-queue (except when the actual
2333 * re-schedule is in progress), and as such you're allowed to do
2334 * the simpler "current->state = TASK_RUNNING" to mark yourself
2335 * runnable without the overhead of this.
2337 * Returns %true if @p was woken up, %false if it was already running
2338 * or @state didn't match @p's state.
2340 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2343 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2344 unsigned long flags
;
2345 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2348 this_cpu
= get_cpu();
2351 rq
= task_rq_lock(p
, &flags
);
2352 if (!(p
->state
& state
))
2362 if (unlikely(task_running(rq
, p
)))
2366 * In order to handle concurrent wakeups and release the rq->lock
2367 * we put the task in TASK_WAKING state.
2369 * First fix up the nr_uninterruptible count:
2371 if (task_contributes_to_load(p
)) {
2372 if (likely(cpu_online(orig_cpu
)))
2373 rq
->nr_uninterruptible
--;
2375 this_rq()->nr_uninterruptible
--;
2377 p
->state
= TASK_WAKING
;
2379 if (p
->sched_class
->task_waking
) {
2380 p
->sched_class
->task_waking(rq
, p
);
2381 en_flags
|= ENQUEUE_WAKING
;
2384 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2385 if (cpu
!= orig_cpu
)
2386 set_task_cpu(p
, cpu
);
2387 __task_rq_unlock(rq
);
2390 raw_spin_lock(&rq
->lock
);
2393 * We migrated the task without holding either rq->lock, however
2394 * since the task is not on the task list itself, nobody else
2395 * will try and migrate the task, hence the rq should match the
2396 * cpu we just moved it to.
2398 WARN_ON(task_cpu(p
) != cpu
);
2399 WARN_ON(p
->state
!= TASK_WAKING
);
2401 #ifdef CONFIG_SCHEDSTATS
2402 schedstat_inc(rq
, ttwu_count
);
2403 if (cpu
== this_cpu
)
2404 schedstat_inc(rq
, ttwu_local
);
2406 struct sched_domain
*sd
;
2407 for_each_domain(this_cpu
, sd
) {
2408 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2409 schedstat_inc(sd
, ttwu_wake_remote
);
2414 #endif /* CONFIG_SCHEDSTATS */
2417 #endif /* CONFIG_SMP */
2418 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2419 cpu
== this_cpu
, en_flags
);
2422 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2424 task_rq_unlock(rq
, &flags
);
2431 * try_to_wake_up_local - try to wake up a local task with rq lock held
2432 * @p: the thread to be awakened
2434 * Put @p on the run-queue if it's not alredy there. The caller must
2435 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2436 * the current task. this_rq() stays locked over invocation.
2438 static void try_to_wake_up_local(struct task_struct
*p
)
2440 struct rq
*rq
= task_rq(p
);
2441 bool success
= false;
2443 BUG_ON(rq
!= this_rq());
2444 BUG_ON(p
== current
);
2445 lockdep_assert_held(&rq
->lock
);
2447 if (!(p
->state
& TASK_NORMAL
))
2451 if (likely(!task_running(rq
, p
))) {
2452 schedstat_inc(rq
, ttwu_count
);
2453 schedstat_inc(rq
, ttwu_local
);
2455 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2458 ttwu_post_activation(p
, rq
, 0, success
);
2462 * wake_up_process - Wake up a specific process
2463 * @p: The process to be woken up.
2465 * Attempt to wake up the nominated process and move it to the set of runnable
2466 * processes. Returns 1 if the process was woken up, 0 if it was already
2469 * It may be assumed that this function implies a write memory barrier before
2470 * changing the task state if and only if any tasks are woken up.
2472 int wake_up_process(struct task_struct
*p
)
2474 return try_to_wake_up(p
, TASK_ALL
, 0);
2476 EXPORT_SYMBOL(wake_up_process
);
2478 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2480 return try_to_wake_up(p
, state
, 0);
2484 * Perform scheduler related setup for a newly forked process p.
2485 * p is forked by current.
2487 * __sched_fork() is basic setup used by init_idle() too:
2489 static void __sched_fork(struct task_struct
*p
)
2491 p
->se
.exec_start
= 0;
2492 p
->se
.sum_exec_runtime
= 0;
2493 p
->se
.prev_sum_exec_runtime
= 0;
2494 p
->se
.nr_migrations
= 0;
2496 #ifdef CONFIG_SCHEDSTATS
2497 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2500 INIT_LIST_HEAD(&p
->rt
.run_list
);
2502 INIT_LIST_HEAD(&p
->se
.group_node
);
2504 #ifdef CONFIG_PREEMPT_NOTIFIERS
2505 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2510 * fork()/clone()-time setup:
2512 void sched_fork(struct task_struct
*p
, int clone_flags
)
2514 int cpu
= get_cpu();
2518 * We mark the process as running here. This guarantees that
2519 * nobody will actually run it, and a signal or other external
2520 * event cannot wake it up and insert it on the runqueue either.
2522 p
->state
= TASK_RUNNING
;
2525 * Revert to default priority/policy on fork if requested.
2527 if (unlikely(p
->sched_reset_on_fork
)) {
2528 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2529 p
->policy
= SCHED_NORMAL
;
2530 p
->normal_prio
= p
->static_prio
;
2533 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2534 p
->static_prio
= NICE_TO_PRIO(0);
2535 p
->normal_prio
= p
->static_prio
;
2540 * We don't need the reset flag anymore after the fork. It has
2541 * fulfilled its duty:
2543 p
->sched_reset_on_fork
= 0;
2547 * Make sure we do not leak PI boosting priority to the child.
2549 p
->prio
= current
->normal_prio
;
2551 if (!rt_prio(p
->prio
))
2552 p
->sched_class
= &fair_sched_class
;
2554 if (p
->sched_class
->task_fork
)
2555 p
->sched_class
->task_fork(p
);
2558 * The child is not yet in the pid-hash so no cgroup attach races,
2559 * and the cgroup is pinned to this child due to cgroup_fork()
2560 * is ran before sched_fork().
2562 * Silence PROVE_RCU.
2565 set_task_cpu(p
, cpu
);
2568 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2569 if (likely(sched_info_on()))
2570 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2572 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2575 #ifdef CONFIG_PREEMPT
2576 /* Want to start with kernel preemption disabled. */
2577 task_thread_info(p
)->preempt_count
= 1;
2579 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2585 * wake_up_new_task - wake up a newly created task for the first time.
2587 * This function will do some initial scheduler statistics housekeeping
2588 * that must be done for every newly created context, then puts the task
2589 * on the runqueue and wakes it.
2591 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2593 unsigned long flags
;
2595 int cpu __maybe_unused
= get_cpu();
2598 rq
= task_rq_lock(p
, &flags
);
2599 p
->state
= TASK_WAKING
;
2602 * Fork balancing, do it here and not earlier because:
2603 * - cpus_allowed can change in the fork path
2604 * - any previously selected cpu might disappear through hotplug
2606 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2607 * without people poking at ->cpus_allowed.
2609 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2610 set_task_cpu(p
, cpu
);
2612 p
->state
= TASK_RUNNING
;
2613 task_rq_unlock(rq
, &flags
);
2616 rq
= task_rq_lock(p
, &flags
);
2617 activate_task(rq
, p
, 0);
2618 trace_sched_wakeup_new(p
, 1);
2619 check_preempt_curr(rq
, p
, WF_FORK
);
2621 if (p
->sched_class
->task_woken
)
2622 p
->sched_class
->task_woken(rq
, p
);
2624 task_rq_unlock(rq
, &flags
);
2628 #ifdef CONFIG_PREEMPT_NOTIFIERS
2631 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2632 * @notifier: notifier struct to register
2634 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2636 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2638 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2641 * preempt_notifier_unregister - no longer interested in preemption notifications
2642 * @notifier: notifier struct to unregister
2644 * This is safe to call from within a preemption notifier.
2646 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2648 hlist_del(¬ifier
->link
);
2650 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2652 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2654 struct preempt_notifier
*notifier
;
2655 struct hlist_node
*node
;
2657 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2658 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2662 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2663 struct task_struct
*next
)
2665 struct preempt_notifier
*notifier
;
2666 struct hlist_node
*node
;
2668 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2669 notifier
->ops
->sched_out(notifier
, next
);
2672 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2674 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2679 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2680 struct task_struct
*next
)
2684 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2687 * prepare_task_switch - prepare to switch tasks
2688 * @rq: the runqueue preparing to switch
2689 * @prev: the current task that is being switched out
2690 * @next: the task we are going to switch to.
2692 * This is called with the rq lock held and interrupts off. It must
2693 * be paired with a subsequent finish_task_switch after the context
2696 * prepare_task_switch sets up locking and calls architecture specific
2700 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2701 struct task_struct
*next
)
2703 fire_sched_out_preempt_notifiers(prev
, next
);
2704 prepare_lock_switch(rq
, next
);
2705 prepare_arch_switch(next
);
2709 * finish_task_switch - clean up after a task-switch
2710 * @rq: runqueue associated with task-switch
2711 * @prev: the thread we just switched away from.
2713 * finish_task_switch must be called after the context switch, paired
2714 * with a prepare_task_switch call before the context switch.
2715 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2716 * and do any other architecture-specific cleanup actions.
2718 * Note that we may have delayed dropping an mm in context_switch(). If
2719 * so, we finish that here outside of the runqueue lock. (Doing it
2720 * with the lock held can cause deadlocks; see schedule() for
2723 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2724 __releases(rq
->lock
)
2726 struct mm_struct
*mm
= rq
->prev_mm
;
2732 * A task struct has one reference for the use as "current".
2733 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2734 * schedule one last time. The schedule call will never return, and
2735 * the scheduled task must drop that reference.
2736 * The test for TASK_DEAD must occur while the runqueue locks are
2737 * still held, otherwise prev could be scheduled on another cpu, die
2738 * there before we look at prev->state, and then the reference would
2740 * Manfred Spraul <manfred@colorfullife.com>
2742 prev_state
= prev
->state
;
2743 finish_arch_switch(prev
);
2744 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2745 local_irq_disable();
2746 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2747 perf_event_task_sched_in(current
);
2748 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2750 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2751 finish_lock_switch(rq
, prev
);
2753 fire_sched_in_preempt_notifiers(current
);
2756 if (unlikely(prev_state
== TASK_DEAD
)) {
2758 * Remove function-return probe instances associated with this
2759 * task and put them back on the free list.
2761 kprobe_flush_task(prev
);
2762 put_task_struct(prev
);
2768 /* assumes rq->lock is held */
2769 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2771 if (prev
->sched_class
->pre_schedule
)
2772 prev
->sched_class
->pre_schedule(rq
, prev
);
2775 /* rq->lock is NOT held, but preemption is disabled */
2776 static inline void post_schedule(struct rq
*rq
)
2778 if (rq
->post_schedule
) {
2779 unsigned long flags
;
2781 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2782 if (rq
->curr
->sched_class
->post_schedule
)
2783 rq
->curr
->sched_class
->post_schedule(rq
);
2784 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2786 rq
->post_schedule
= 0;
2792 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2796 static inline void post_schedule(struct rq
*rq
)
2803 * schedule_tail - first thing a freshly forked thread must call.
2804 * @prev: the thread we just switched away from.
2806 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2807 __releases(rq
->lock
)
2809 struct rq
*rq
= this_rq();
2811 finish_task_switch(rq
, prev
);
2814 * FIXME: do we need to worry about rq being invalidated by the
2819 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2820 /* In this case, finish_task_switch does not reenable preemption */
2823 if (current
->set_child_tid
)
2824 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2828 * context_switch - switch to the new MM and the new
2829 * thread's register state.
2832 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2833 struct task_struct
*next
)
2835 struct mm_struct
*mm
, *oldmm
;
2837 prepare_task_switch(rq
, prev
, next
);
2838 trace_sched_switch(prev
, next
);
2840 oldmm
= prev
->active_mm
;
2842 * For paravirt, this is coupled with an exit in switch_to to
2843 * combine the page table reload and the switch backend into
2846 arch_start_context_switch(prev
);
2849 next
->active_mm
= oldmm
;
2850 atomic_inc(&oldmm
->mm_count
);
2851 enter_lazy_tlb(oldmm
, next
);
2853 switch_mm(oldmm
, mm
, next
);
2855 if (likely(!prev
->mm
)) {
2856 prev
->active_mm
= NULL
;
2857 rq
->prev_mm
= oldmm
;
2860 * Since the runqueue lock will be released by the next
2861 * task (which is an invalid locking op but in the case
2862 * of the scheduler it's an obvious special-case), so we
2863 * do an early lockdep release here:
2865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2866 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2869 /* Here we just switch the register state and the stack. */
2870 switch_to(prev
, next
, prev
);
2874 * this_rq must be evaluated again because prev may have moved
2875 * CPUs since it called schedule(), thus the 'rq' on its stack
2876 * frame will be invalid.
2878 finish_task_switch(this_rq(), prev
);
2882 * nr_running, nr_uninterruptible and nr_context_switches:
2884 * externally visible scheduler statistics: current number of runnable
2885 * threads, current number of uninterruptible-sleeping threads, total
2886 * number of context switches performed since bootup.
2888 unsigned long nr_running(void)
2890 unsigned long i
, sum
= 0;
2892 for_each_online_cpu(i
)
2893 sum
+= cpu_rq(i
)->nr_running
;
2898 unsigned long nr_uninterruptible(void)
2900 unsigned long i
, sum
= 0;
2902 for_each_possible_cpu(i
)
2903 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2906 * Since we read the counters lockless, it might be slightly
2907 * inaccurate. Do not allow it to go below zero though:
2909 if (unlikely((long)sum
< 0))
2915 unsigned long long nr_context_switches(void)
2918 unsigned long long sum
= 0;
2920 for_each_possible_cpu(i
)
2921 sum
+= cpu_rq(i
)->nr_switches
;
2926 unsigned long nr_iowait(void)
2928 unsigned long i
, sum
= 0;
2930 for_each_possible_cpu(i
)
2931 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2936 unsigned long nr_iowait_cpu(int cpu
)
2938 struct rq
*this = cpu_rq(cpu
);
2939 return atomic_read(&this->nr_iowait
);
2942 unsigned long this_cpu_load(void)
2944 struct rq
*this = this_rq();
2945 return this->cpu_load
[0];
2949 /* Variables and functions for calc_load */
2950 static atomic_long_t calc_load_tasks
;
2951 static unsigned long calc_load_update
;
2952 unsigned long avenrun
[3];
2953 EXPORT_SYMBOL(avenrun
);
2955 static long calc_load_fold_active(struct rq
*this_rq
)
2957 long nr_active
, delta
= 0;
2959 nr_active
= this_rq
->nr_running
;
2960 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2962 if (nr_active
!= this_rq
->calc_load_active
) {
2963 delta
= nr_active
- this_rq
->calc_load_active
;
2964 this_rq
->calc_load_active
= nr_active
;
2972 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2974 * When making the ILB scale, we should try to pull this in as well.
2976 static atomic_long_t calc_load_tasks_idle
;
2978 static void calc_load_account_idle(struct rq
*this_rq
)
2982 delta
= calc_load_fold_active(this_rq
);
2984 atomic_long_add(delta
, &calc_load_tasks_idle
);
2987 static long calc_load_fold_idle(void)
2992 * Its got a race, we don't care...
2994 if (atomic_long_read(&calc_load_tasks_idle
))
2995 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3000 static void calc_load_account_idle(struct rq
*this_rq
)
3004 static inline long calc_load_fold_idle(void)
3011 * get_avenrun - get the load average array
3012 * @loads: pointer to dest load array
3013 * @offset: offset to add
3014 * @shift: shift count to shift the result left
3016 * These values are estimates at best, so no need for locking.
3018 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3020 loads
[0] = (avenrun
[0] + offset
) << shift
;
3021 loads
[1] = (avenrun
[1] + offset
) << shift
;
3022 loads
[2] = (avenrun
[2] + offset
) << shift
;
3025 static unsigned long
3026 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3029 load
+= active
* (FIXED_1
- exp
);
3030 return load
>> FSHIFT
;
3034 * calc_load - update the avenrun load estimates 10 ticks after the
3035 * CPUs have updated calc_load_tasks.
3037 void calc_global_load(void)
3039 unsigned long upd
= calc_load_update
+ 10;
3042 if (time_before(jiffies
, upd
))
3045 active
= atomic_long_read(&calc_load_tasks
);
3046 active
= active
> 0 ? active
* FIXED_1
: 0;
3048 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3049 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3050 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3052 calc_load_update
+= LOAD_FREQ
;
3056 * Called from update_cpu_load() to periodically update this CPU's
3059 static void calc_load_account_active(struct rq
*this_rq
)
3063 if (time_before(jiffies
, this_rq
->calc_load_update
))
3066 delta
= calc_load_fold_active(this_rq
);
3067 delta
+= calc_load_fold_idle();
3069 atomic_long_add(delta
, &calc_load_tasks
);
3071 this_rq
->calc_load_update
+= LOAD_FREQ
;
3075 * The exact cpuload at various idx values, calculated at every tick would be
3076 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3078 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3079 * on nth tick when cpu may be busy, then we have:
3080 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3081 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3083 * decay_load_missed() below does efficient calculation of
3084 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3085 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3087 * The calculation is approximated on a 128 point scale.
3088 * degrade_zero_ticks is the number of ticks after which load at any
3089 * particular idx is approximated to be zero.
3090 * degrade_factor is a precomputed table, a row for each load idx.
3091 * Each column corresponds to degradation factor for a power of two ticks,
3092 * based on 128 point scale.
3094 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3095 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3097 * With this power of 2 load factors, we can degrade the load n times
3098 * by looking at 1 bits in n and doing as many mult/shift instead of
3099 * n mult/shifts needed by the exact degradation.
3101 #define DEGRADE_SHIFT 7
3102 static const unsigned char
3103 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3104 static const unsigned char
3105 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3106 {0, 0, 0, 0, 0, 0, 0, 0},
3107 {64, 32, 8, 0, 0, 0, 0, 0},
3108 {96, 72, 40, 12, 1, 0, 0},
3109 {112, 98, 75, 43, 15, 1, 0},
3110 {120, 112, 98, 76, 45, 16, 2} };
3113 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3114 * would be when CPU is idle and so we just decay the old load without
3115 * adding any new load.
3117 static unsigned long
3118 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3122 if (!missed_updates
)
3125 if (missed_updates
>= degrade_zero_ticks
[idx
])
3129 return load
>> missed_updates
;
3131 while (missed_updates
) {
3132 if (missed_updates
% 2)
3133 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3135 missed_updates
>>= 1;
3142 * Update rq->cpu_load[] statistics. This function is usually called every
3143 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3144 * every tick. We fix it up based on jiffies.
3146 static void update_cpu_load(struct rq
*this_rq
)
3148 unsigned long this_load
= this_rq
->load
.weight
;
3149 unsigned long curr_jiffies
= jiffies
;
3150 unsigned long pending_updates
;
3153 this_rq
->nr_load_updates
++;
3155 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3156 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3159 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3160 this_rq
->last_load_update_tick
= curr_jiffies
;
3162 /* Update our load: */
3163 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3164 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3165 unsigned long old_load
, new_load
;
3167 /* scale is effectively 1 << i now, and >> i divides by scale */
3169 old_load
= this_rq
->cpu_load
[i
];
3170 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3171 new_load
= this_load
;
3173 * Round up the averaging division if load is increasing. This
3174 * prevents us from getting stuck on 9 if the load is 10, for
3177 if (new_load
> old_load
)
3178 new_load
+= scale
- 1;
3180 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3183 sched_avg_update(this_rq
);
3186 static void update_cpu_load_active(struct rq
*this_rq
)
3188 update_cpu_load(this_rq
);
3190 calc_load_account_active(this_rq
);
3196 * sched_exec - execve() is a valuable balancing opportunity, because at
3197 * this point the task has the smallest effective memory and cache footprint.
3199 void sched_exec(void)
3201 struct task_struct
*p
= current
;
3202 unsigned long flags
;
3206 rq
= task_rq_lock(p
, &flags
);
3207 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3208 if (dest_cpu
== smp_processor_id())
3212 * select_task_rq() can race against ->cpus_allowed
3214 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3215 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3216 struct migration_arg arg
= { p
, dest_cpu
};
3218 task_rq_unlock(rq
, &flags
);
3219 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3223 task_rq_unlock(rq
, &flags
);
3228 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3230 EXPORT_PER_CPU_SYMBOL(kstat
);
3233 * Return any ns on the sched_clock that have not yet been accounted in
3234 * @p in case that task is currently running.
3236 * Called with task_rq_lock() held on @rq.
3238 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3242 if (task_current(rq
, p
)) {
3243 update_rq_clock(rq
);
3244 ns
= rq
->clock
- p
->se
.exec_start
;
3252 unsigned long long task_delta_exec(struct task_struct
*p
)
3254 unsigned long flags
;
3258 rq
= task_rq_lock(p
, &flags
);
3259 ns
= do_task_delta_exec(p
, rq
);
3260 task_rq_unlock(rq
, &flags
);
3266 * Return accounted runtime for the task.
3267 * In case the task is currently running, return the runtime plus current's
3268 * pending runtime that have not been accounted yet.
3270 unsigned long long task_sched_runtime(struct task_struct
*p
)
3272 unsigned long flags
;
3276 rq
= task_rq_lock(p
, &flags
);
3277 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3278 task_rq_unlock(rq
, &flags
);
3284 * Return sum_exec_runtime for the thread group.
3285 * In case the task is currently running, return the sum plus current's
3286 * pending runtime that have not been accounted yet.
3288 * Note that the thread group might have other running tasks as well,
3289 * so the return value not includes other pending runtime that other
3290 * running tasks might have.
3292 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3294 struct task_cputime totals
;
3295 unsigned long flags
;
3299 rq
= task_rq_lock(p
, &flags
);
3300 thread_group_cputime(p
, &totals
);
3301 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3302 task_rq_unlock(rq
, &flags
);
3308 * Account user cpu time to a process.
3309 * @p: the process that the cpu time gets accounted to
3310 * @cputime: the cpu time spent in user space since the last update
3311 * @cputime_scaled: cputime scaled by cpu frequency
3313 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3314 cputime_t cputime_scaled
)
3316 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3319 /* Add user time to process. */
3320 p
->utime
= cputime_add(p
->utime
, cputime
);
3321 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3322 account_group_user_time(p
, cputime
);
3324 /* Add user time to cpustat. */
3325 tmp
= cputime_to_cputime64(cputime
);
3326 if (TASK_NICE(p
) > 0)
3327 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3329 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3331 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3332 /* Account for user time used */
3333 acct_update_integrals(p
);
3337 * Account guest cpu time to a process.
3338 * @p: the process that the cpu time gets accounted to
3339 * @cputime: the cpu time spent in virtual machine since the last update
3340 * @cputime_scaled: cputime scaled by cpu frequency
3342 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3343 cputime_t cputime_scaled
)
3346 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3348 tmp
= cputime_to_cputime64(cputime
);
3350 /* Add guest time to process. */
3351 p
->utime
= cputime_add(p
->utime
, cputime
);
3352 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3353 account_group_user_time(p
, cputime
);
3354 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3356 /* Add guest time to cpustat. */
3357 if (TASK_NICE(p
) > 0) {
3358 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3359 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3361 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3362 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3367 * Account system cpu time to a process.
3368 * @p: the process that the cpu time gets accounted to
3369 * @hardirq_offset: the offset to subtract from hardirq_count()
3370 * @cputime: the cpu time spent in kernel space since the last update
3371 * @cputime_scaled: cputime scaled by cpu frequency
3373 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3374 cputime_t cputime
, cputime_t cputime_scaled
)
3376 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3379 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3380 account_guest_time(p
, cputime
, cputime_scaled
);
3384 /* Add system time to process. */
3385 p
->stime
= cputime_add(p
->stime
, cputime
);
3386 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3387 account_group_system_time(p
, cputime
);
3389 /* Add system time to cpustat. */
3390 tmp
= cputime_to_cputime64(cputime
);
3391 if (hardirq_count() - hardirq_offset
)
3392 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3393 else if (softirq_count())
3394 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3396 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3398 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3400 /* Account for system time used */
3401 acct_update_integrals(p
);
3405 * Account for involuntary wait time.
3406 * @steal: the cpu time spent in involuntary wait
3408 void account_steal_time(cputime_t cputime
)
3410 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3411 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3413 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3417 * Account for idle time.
3418 * @cputime: the cpu time spent in idle wait
3420 void account_idle_time(cputime_t cputime
)
3422 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3423 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3424 struct rq
*rq
= this_rq();
3426 if (atomic_read(&rq
->nr_iowait
) > 0)
3427 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3429 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3432 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3435 * Account a single tick of cpu time.
3436 * @p: the process that the cpu time gets accounted to
3437 * @user_tick: indicates if the tick is a user or a system tick
3439 void account_process_tick(struct task_struct
*p
, int user_tick
)
3441 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3442 struct rq
*rq
= this_rq();
3445 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3446 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3447 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3450 account_idle_time(cputime_one_jiffy
);
3454 * Account multiple ticks of steal time.
3455 * @p: the process from which the cpu time has been stolen
3456 * @ticks: number of stolen ticks
3458 void account_steal_ticks(unsigned long ticks
)
3460 account_steal_time(jiffies_to_cputime(ticks
));
3464 * Account multiple ticks of idle time.
3465 * @ticks: number of stolen ticks
3467 void account_idle_ticks(unsigned long ticks
)
3469 account_idle_time(jiffies_to_cputime(ticks
));
3475 * Use precise platform statistics if available:
3477 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3478 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3484 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3486 struct task_cputime cputime
;
3488 thread_group_cputime(p
, &cputime
);
3490 *ut
= cputime
.utime
;
3491 *st
= cputime
.stime
;
3495 #ifndef nsecs_to_cputime
3496 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3499 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3501 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3504 * Use CFS's precise accounting:
3506 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3512 do_div(temp
, total
);
3513 utime
= (cputime_t
)temp
;
3518 * Compare with previous values, to keep monotonicity:
3520 p
->prev_utime
= max(p
->prev_utime
, utime
);
3521 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3523 *ut
= p
->prev_utime
;
3524 *st
= p
->prev_stime
;
3528 * Must be called with siglock held.
3530 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3532 struct signal_struct
*sig
= p
->signal
;
3533 struct task_cputime cputime
;
3534 cputime_t rtime
, utime
, total
;
3536 thread_group_cputime(p
, &cputime
);
3538 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3539 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3544 temp
*= cputime
.utime
;
3545 do_div(temp
, total
);
3546 utime
= (cputime_t
)temp
;
3550 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3551 sig
->prev_stime
= max(sig
->prev_stime
,
3552 cputime_sub(rtime
, sig
->prev_utime
));
3554 *ut
= sig
->prev_utime
;
3555 *st
= sig
->prev_stime
;
3560 * This function gets called by the timer code, with HZ frequency.
3561 * We call it with interrupts disabled.
3563 * It also gets called by the fork code, when changing the parent's
3566 void scheduler_tick(void)
3568 int cpu
= smp_processor_id();
3569 struct rq
*rq
= cpu_rq(cpu
);
3570 struct task_struct
*curr
= rq
->curr
;
3574 raw_spin_lock(&rq
->lock
);
3575 update_rq_clock(rq
);
3576 update_cpu_load_active(rq
);
3577 curr
->sched_class
->task_tick(rq
, curr
, 0);
3578 raw_spin_unlock(&rq
->lock
);
3580 perf_event_task_tick(curr
);
3583 rq
->idle_at_tick
= idle_cpu(cpu
);
3584 trigger_load_balance(rq
, cpu
);
3588 notrace
unsigned long get_parent_ip(unsigned long addr
)
3590 if (in_lock_functions(addr
)) {
3591 addr
= CALLER_ADDR2
;
3592 if (in_lock_functions(addr
))
3593 addr
= CALLER_ADDR3
;
3598 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3599 defined(CONFIG_PREEMPT_TRACER))
3601 void __kprobes
add_preempt_count(int val
)
3603 #ifdef CONFIG_DEBUG_PREEMPT
3607 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3610 preempt_count() += val
;
3611 #ifdef CONFIG_DEBUG_PREEMPT
3613 * Spinlock count overflowing soon?
3615 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3618 if (preempt_count() == val
)
3619 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3621 EXPORT_SYMBOL(add_preempt_count
);
3623 void __kprobes
sub_preempt_count(int val
)
3625 #ifdef CONFIG_DEBUG_PREEMPT
3629 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3632 * Is the spinlock portion underflowing?
3634 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3635 !(preempt_count() & PREEMPT_MASK
)))
3639 if (preempt_count() == val
)
3640 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3641 preempt_count() -= val
;
3643 EXPORT_SYMBOL(sub_preempt_count
);
3648 * Print scheduling while atomic bug:
3650 static noinline
void __schedule_bug(struct task_struct
*prev
)
3652 struct pt_regs
*regs
= get_irq_regs();
3654 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3655 prev
->comm
, prev
->pid
, preempt_count());
3657 debug_show_held_locks(prev
);
3659 if (irqs_disabled())
3660 print_irqtrace_events(prev
);
3669 * Various schedule()-time debugging checks and statistics:
3671 static inline void schedule_debug(struct task_struct
*prev
)
3674 * Test if we are atomic. Since do_exit() needs to call into
3675 * schedule() atomically, we ignore that path for now.
3676 * Otherwise, whine if we are scheduling when we should not be.
3678 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3679 __schedule_bug(prev
);
3681 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3683 schedstat_inc(this_rq(), sched_count
);
3684 #ifdef CONFIG_SCHEDSTATS
3685 if (unlikely(prev
->lock_depth
>= 0)) {
3686 schedstat_inc(this_rq(), bkl_count
);
3687 schedstat_inc(prev
, sched_info
.bkl_count
);
3692 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3695 update_rq_clock(rq
);
3696 rq
->skip_clock_update
= 0;
3697 prev
->sched_class
->put_prev_task(rq
, prev
);
3701 * Pick up the highest-prio task:
3703 static inline struct task_struct
*
3704 pick_next_task(struct rq
*rq
)
3706 const struct sched_class
*class;
3707 struct task_struct
*p
;
3710 * Optimization: we know that if all tasks are in
3711 * the fair class we can call that function directly:
3713 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3714 p
= fair_sched_class
.pick_next_task(rq
);
3719 class = sched_class_highest
;
3721 p
= class->pick_next_task(rq
);
3725 * Will never be NULL as the idle class always
3726 * returns a non-NULL p:
3728 class = class->next
;
3733 * schedule() is the main scheduler function.
3735 asmlinkage
void __sched
schedule(void)
3737 struct task_struct
*prev
, *next
;
3738 unsigned long *switch_count
;
3744 cpu
= smp_processor_id();
3746 rcu_note_context_switch(cpu
);
3749 release_kernel_lock(prev
);
3750 need_resched_nonpreemptible
:
3752 schedule_debug(prev
);
3754 if (sched_feat(HRTICK
))
3757 raw_spin_lock_irq(&rq
->lock
);
3758 clear_tsk_need_resched(prev
);
3760 switch_count
= &prev
->nivcsw
;
3761 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3762 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3763 prev
->state
= TASK_RUNNING
;
3766 * If a worker is going to sleep, notify and
3767 * ask workqueue whether it wants to wake up a
3768 * task to maintain concurrency. If so, wake
3771 if (prev
->flags
& PF_WQ_WORKER
) {
3772 struct task_struct
*to_wakeup
;
3774 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3776 try_to_wake_up_local(to_wakeup
);
3778 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3780 switch_count
= &prev
->nvcsw
;
3783 pre_schedule(rq
, prev
);
3785 if (unlikely(!rq
->nr_running
))
3786 idle_balance(cpu
, rq
);
3788 put_prev_task(rq
, prev
);
3789 next
= pick_next_task(rq
);
3791 if (likely(prev
!= next
)) {
3792 sched_info_switch(prev
, next
);
3793 perf_event_task_sched_out(prev
, next
);
3799 context_switch(rq
, prev
, next
); /* unlocks the rq */
3801 * The context switch have flipped the stack from under us
3802 * and restored the local variables which were saved when
3803 * this task called schedule() in the past. prev == current
3804 * is still correct, but it can be moved to another cpu/rq.
3806 cpu
= smp_processor_id();
3809 raw_spin_unlock_irq(&rq
->lock
);
3813 if (unlikely(reacquire_kernel_lock(prev
)))
3814 goto need_resched_nonpreemptible
;
3816 preempt_enable_no_resched();
3820 EXPORT_SYMBOL(schedule
);
3822 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3824 * Look out! "owner" is an entirely speculative pointer
3825 * access and not reliable.
3827 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3832 if (!sched_feat(OWNER_SPIN
))
3835 #ifdef CONFIG_DEBUG_PAGEALLOC
3837 * Need to access the cpu field knowing that
3838 * DEBUG_PAGEALLOC could have unmapped it if
3839 * the mutex owner just released it and exited.
3841 if (probe_kernel_address(&owner
->cpu
, cpu
))
3848 * Even if the access succeeded (likely case),
3849 * the cpu field may no longer be valid.
3851 if (cpu
>= nr_cpumask_bits
)
3855 * We need to validate that we can do a
3856 * get_cpu() and that we have the percpu area.
3858 if (!cpu_online(cpu
))
3865 * Owner changed, break to re-assess state.
3867 if (lock
->owner
!= owner
) {
3869 * If the lock has switched to a different owner,
3870 * we likely have heavy contention. Return 0 to quit
3871 * optimistic spinning and not contend further:
3879 * Is that owner really running on that cpu?
3881 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3891 #ifdef CONFIG_PREEMPT
3893 * this is the entry point to schedule() from in-kernel preemption
3894 * off of preempt_enable. Kernel preemptions off return from interrupt
3895 * occur there and call schedule directly.
3897 asmlinkage
void __sched notrace
preempt_schedule(void)
3899 struct thread_info
*ti
= current_thread_info();
3902 * If there is a non-zero preempt_count or interrupts are disabled,
3903 * we do not want to preempt the current task. Just return..
3905 if (likely(ti
->preempt_count
|| irqs_disabled()))
3909 add_preempt_count_notrace(PREEMPT_ACTIVE
);
3911 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
3914 * Check again in case we missed a preemption opportunity
3915 * between schedule and now.
3918 } while (need_resched());
3920 EXPORT_SYMBOL(preempt_schedule
);
3923 * this is the entry point to schedule() from kernel preemption
3924 * off of irq context.
3925 * Note, that this is called and return with irqs disabled. This will
3926 * protect us against recursive calling from irq.
3928 asmlinkage
void __sched
preempt_schedule_irq(void)
3930 struct thread_info
*ti
= current_thread_info();
3932 /* Catch callers which need to be fixed */
3933 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3936 add_preempt_count(PREEMPT_ACTIVE
);
3939 local_irq_disable();
3940 sub_preempt_count(PREEMPT_ACTIVE
);
3943 * Check again in case we missed a preemption opportunity
3944 * between schedule and now.
3947 } while (need_resched());
3950 #endif /* CONFIG_PREEMPT */
3952 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3955 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3957 EXPORT_SYMBOL(default_wake_function
);
3960 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3961 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3962 * number) then we wake all the non-exclusive tasks and one exclusive task.
3964 * There are circumstances in which we can try to wake a task which has already
3965 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3966 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3968 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3969 int nr_exclusive
, int wake_flags
, void *key
)
3971 wait_queue_t
*curr
, *next
;
3973 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3974 unsigned flags
= curr
->flags
;
3976 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3977 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3983 * __wake_up - wake up threads blocked on a waitqueue.
3985 * @mode: which threads
3986 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3987 * @key: is directly passed to the wakeup function
3989 * It may be assumed that this function implies a write memory barrier before
3990 * changing the task state if and only if any tasks are woken up.
3992 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3993 int nr_exclusive
, void *key
)
3995 unsigned long flags
;
3997 spin_lock_irqsave(&q
->lock
, flags
);
3998 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3999 spin_unlock_irqrestore(&q
->lock
, flags
);
4001 EXPORT_SYMBOL(__wake_up
);
4004 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4006 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4008 __wake_up_common(q
, mode
, 1, 0, NULL
);
4010 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4012 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4014 __wake_up_common(q
, mode
, 1, 0, key
);
4018 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4020 * @mode: which threads
4021 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4022 * @key: opaque value to be passed to wakeup targets
4024 * The sync wakeup differs that the waker knows that it will schedule
4025 * away soon, so while the target thread will be woken up, it will not
4026 * be migrated to another CPU - ie. the two threads are 'synchronized'
4027 * with each other. This can prevent needless bouncing between CPUs.
4029 * On UP it can prevent extra preemption.
4031 * It may be assumed that this function implies a write memory barrier before
4032 * changing the task state if and only if any tasks are woken up.
4034 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4035 int nr_exclusive
, void *key
)
4037 unsigned long flags
;
4038 int wake_flags
= WF_SYNC
;
4043 if (unlikely(!nr_exclusive
))
4046 spin_lock_irqsave(&q
->lock
, flags
);
4047 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4048 spin_unlock_irqrestore(&q
->lock
, flags
);
4050 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4053 * __wake_up_sync - see __wake_up_sync_key()
4055 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4057 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4059 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4062 * complete: - signals a single thread waiting on this completion
4063 * @x: holds the state of this particular completion
4065 * This will wake up a single thread waiting on this completion. Threads will be
4066 * awakened in the same order in which they were queued.
4068 * See also complete_all(), wait_for_completion() and related routines.
4070 * It may be assumed that this function implies a write memory barrier before
4071 * changing the task state if and only if any tasks are woken up.
4073 void complete(struct completion
*x
)
4075 unsigned long flags
;
4077 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4079 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4080 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4082 EXPORT_SYMBOL(complete
);
4085 * complete_all: - signals all threads waiting on this completion
4086 * @x: holds the state of this particular completion
4088 * This will wake up all threads waiting on this particular completion event.
4090 * It may be assumed that this function implies a write memory barrier before
4091 * changing the task state if and only if any tasks are woken up.
4093 void complete_all(struct completion
*x
)
4095 unsigned long flags
;
4097 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4098 x
->done
+= UINT_MAX
/2;
4099 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4100 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4102 EXPORT_SYMBOL(complete_all
);
4104 static inline long __sched
4105 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4108 DECLARE_WAITQUEUE(wait
, current
);
4110 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4112 if (signal_pending_state(state
, current
)) {
4113 timeout
= -ERESTARTSYS
;
4116 __set_current_state(state
);
4117 spin_unlock_irq(&x
->wait
.lock
);
4118 timeout
= schedule_timeout(timeout
);
4119 spin_lock_irq(&x
->wait
.lock
);
4120 } while (!x
->done
&& timeout
);
4121 __remove_wait_queue(&x
->wait
, &wait
);
4126 return timeout
?: 1;
4130 wait_for_common(struct completion
*x
, long timeout
, int state
)
4134 spin_lock_irq(&x
->wait
.lock
);
4135 timeout
= do_wait_for_common(x
, timeout
, state
);
4136 spin_unlock_irq(&x
->wait
.lock
);
4141 * wait_for_completion: - waits for completion of a task
4142 * @x: holds the state of this particular completion
4144 * This waits to be signaled for completion of a specific task. It is NOT
4145 * interruptible and there is no timeout.
4147 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4148 * and interrupt capability. Also see complete().
4150 void __sched
wait_for_completion(struct completion
*x
)
4152 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4154 EXPORT_SYMBOL(wait_for_completion
);
4157 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4158 * @x: holds the state of this particular completion
4159 * @timeout: timeout value in jiffies
4161 * This waits for either a completion of a specific task to be signaled or for a
4162 * specified timeout to expire. The timeout is in jiffies. It is not
4165 unsigned long __sched
4166 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4168 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4170 EXPORT_SYMBOL(wait_for_completion_timeout
);
4173 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4174 * @x: holds the state of this particular completion
4176 * This waits for completion of a specific task to be signaled. It is
4179 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4181 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4182 if (t
== -ERESTARTSYS
)
4186 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4189 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4190 * @x: holds the state of this particular completion
4191 * @timeout: timeout value in jiffies
4193 * This waits for either a completion of a specific task to be signaled or for a
4194 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4196 unsigned long __sched
4197 wait_for_completion_interruptible_timeout(struct completion
*x
,
4198 unsigned long timeout
)
4200 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4202 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4205 * wait_for_completion_killable: - waits for completion of a task (killable)
4206 * @x: holds the state of this particular completion
4208 * This waits to be signaled for completion of a specific task. It can be
4209 * interrupted by a kill signal.
4211 int __sched
wait_for_completion_killable(struct completion
*x
)
4213 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4214 if (t
== -ERESTARTSYS
)
4218 EXPORT_SYMBOL(wait_for_completion_killable
);
4221 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4222 * @x: holds the state of this particular completion
4223 * @timeout: timeout value in jiffies
4225 * This waits for either a completion of a specific task to be
4226 * signaled or for a specified timeout to expire. It can be
4227 * interrupted by a kill signal. The timeout is in jiffies.
4229 unsigned long __sched
4230 wait_for_completion_killable_timeout(struct completion
*x
,
4231 unsigned long timeout
)
4233 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4235 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4238 * try_wait_for_completion - try to decrement a completion without blocking
4239 * @x: completion structure
4241 * Returns: 0 if a decrement cannot be done without blocking
4242 * 1 if a decrement succeeded.
4244 * If a completion is being used as a counting completion,
4245 * attempt to decrement the counter without blocking. This
4246 * enables us to avoid waiting if the resource the completion
4247 * is protecting is not available.
4249 bool try_wait_for_completion(struct completion
*x
)
4251 unsigned long flags
;
4254 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4259 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4262 EXPORT_SYMBOL(try_wait_for_completion
);
4265 * completion_done - Test to see if a completion has any waiters
4266 * @x: completion structure
4268 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4269 * 1 if there are no waiters.
4272 bool completion_done(struct completion
*x
)
4274 unsigned long flags
;
4277 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4280 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4283 EXPORT_SYMBOL(completion_done
);
4286 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4288 unsigned long flags
;
4291 init_waitqueue_entry(&wait
, current
);
4293 __set_current_state(state
);
4295 spin_lock_irqsave(&q
->lock
, flags
);
4296 __add_wait_queue(q
, &wait
);
4297 spin_unlock(&q
->lock
);
4298 timeout
= schedule_timeout(timeout
);
4299 spin_lock_irq(&q
->lock
);
4300 __remove_wait_queue(q
, &wait
);
4301 spin_unlock_irqrestore(&q
->lock
, flags
);
4306 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4308 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4310 EXPORT_SYMBOL(interruptible_sleep_on
);
4313 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4315 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4317 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4319 void __sched
sleep_on(wait_queue_head_t
*q
)
4321 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4323 EXPORT_SYMBOL(sleep_on
);
4325 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4327 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4329 EXPORT_SYMBOL(sleep_on_timeout
);
4331 #ifdef CONFIG_RT_MUTEXES
4334 * rt_mutex_setprio - set the current priority of a task
4336 * @prio: prio value (kernel-internal form)
4338 * This function changes the 'effective' priority of a task. It does
4339 * not touch ->normal_prio like __setscheduler().
4341 * Used by the rt_mutex code to implement priority inheritance logic.
4343 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4345 unsigned long flags
;
4346 int oldprio
, on_rq
, running
;
4348 const struct sched_class
*prev_class
;
4350 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4352 rq
= task_rq_lock(p
, &flags
);
4355 prev_class
= p
->sched_class
;
4356 on_rq
= p
->se
.on_rq
;
4357 running
= task_current(rq
, p
);
4359 dequeue_task(rq
, p
, 0);
4361 p
->sched_class
->put_prev_task(rq
, p
);
4364 p
->sched_class
= &rt_sched_class
;
4366 p
->sched_class
= &fair_sched_class
;
4371 p
->sched_class
->set_curr_task(rq
);
4373 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4375 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4377 task_rq_unlock(rq
, &flags
);
4382 void set_user_nice(struct task_struct
*p
, long nice
)
4384 int old_prio
, delta
, on_rq
;
4385 unsigned long flags
;
4388 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4391 * We have to be careful, if called from sys_setpriority(),
4392 * the task might be in the middle of scheduling on another CPU.
4394 rq
= task_rq_lock(p
, &flags
);
4396 * The RT priorities are set via sched_setscheduler(), but we still
4397 * allow the 'normal' nice value to be set - but as expected
4398 * it wont have any effect on scheduling until the task is
4399 * SCHED_FIFO/SCHED_RR:
4401 if (task_has_rt_policy(p
)) {
4402 p
->static_prio
= NICE_TO_PRIO(nice
);
4405 on_rq
= p
->se
.on_rq
;
4407 dequeue_task(rq
, p
, 0);
4409 p
->static_prio
= NICE_TO_PRIO(nice
);
4412 p
->prio
= effective_prio(p
);
4413 delta
= p
->prio
- old_prio
;
4416 enqueue_task(rq
, p
, 0);
4418 * If the task increased its priority or is running and
4419 * lowered its priority, then reschedule its CPU:
4421 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4422 resched_task(rq
->curr
);
4425 task_rq_unlock(rq
, &flags
);
4427 EXPORT_SYMBOL(set_user_nice
);
4430 * can_nice - check if a task can reduce its nice value
4434 int can_nice(const struct task_struct
*p
, const int nice
)
4436 /* convert nice value [19,-20] to rlimit style value [1,40] */
4437 int nice_rlim
= 20 - nice
;
4439 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4440 capable(CAP_SYS_NICE
));
4443 #ifdef __ARCH_WANT_SYS_NICE
4446 * sys_nice - change the priority of the current process.
4447 * @increment: priority increment
4449 * sys_setpriority is a more generic, but much slower function that
4450 * does similar things.
4452 SYSCALL_DEFINE1(nice
, int, increment
)
4457 * Setpriority might change our priority at the same moment.
4458 * We don't have to worry. Conceptually one call occurs first
4459 * and we have a single winner.
4461 if (increment
< -40)
4466 nice
= TASK_NICE(current
) + increment
;
4472 if (increment
< 0 && !can_nice(current
, nice
))
4475 retval
= security_task_setnice(current
, nice
);
4479 set_user_nice(current
, nice
);
4486 * task_prio - return the priority value of a given task.
4487 * @p: the task in question.
4489 * This is the priority value as seen by users in /proc.
4490 * RT tasks are offset by -200. Normal tasks are centered
4491 * around 0, value goes from -16 to +15.
4493 int task_prio(const struct task_struct
*p
)
4495 return p
->prio
- MAX_RT_PRIO
;
4499 * task_nice - return the nice value of a given task.
4500 * @p: the task in question.
4502 int task_nice(const struct task_struct
*p
)
4504 return TASK_NICE(p
);
4506 EXPORT_SYMBOL(task_nice
);
4509 * idle_cpu - is a given cpu idle currently?
4510 * @cpu: the processor in question.
4512 int idle_cpu(int cpu
)
4514 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4518 * idle_task - return the idle task for a given cpu.
4519 * @cpu: the processor in question.
4521 struct task_struct
*idle_task(int cpu
)
4523 return cpu_rq(cpu
)->idle
;
4527 * find_process_by_pid - find a process with a matching PID value.
4528 * @pid: the pid in question.
4530 static struct task_struct
*find_process_by_pid(pid_t pid
)
4532 return pid
? find_task_by_vpid(pid
) : current
;
4535 /* Actually do priority change: must hold rq lock. */
4537 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4539 BUG_ON(p
->se
.on_rq
);
4542 p
->rt_priority
= prio
;
4543 p
->normal_prio
= normal_prio(p
);
4544 /* we are holding p->pi_lock already */
4545 p
->prio
= rt_mutex_getprio(p
);
4546 if (rt_prio(p
->prio
))
4547 p
->sched_class
= &rt_sched_class
;
4549 p
->sched_class
= &fair_sched_class
;
4554 * check the target process has a UID that matches the current process's
4556 static bool check_same_owner(struct task_struct
*p
)
4558 const struct cred
*cred
= current_cred(), *pcred
;
4562 pcred
= __task_cred(p
);
4563 match
= (cred
->euid
== pcred
->euid
||
4564 cred
->euid
== pcred
->uid
);
4569 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4570 struct sched_param
*param
, bool user
)
4572 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4573 unsigned long flags
;
4574 const struct sched_class
*prev_class
;
4578 /* may grab non-irq protected spin_locks */
4579 BUG_ON(in_interrupt());
4581 /* double check policy once rq lock held */
4583 reset_on_fork
= p
->sched_reset_on_fork
;
4584 policy
= oldpolicy
= p
->policy
;
4586 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4587 policy
&= ~SCHED_RESET_ON_FORK
;
4589 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4590 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4591 policy
!= SCHED_IDLE
)
4596 * Valid priorities for SCHED_FIFO and SCHED_RR are
4597 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4598 * SCHED_BATCH and SCHED_IDLE is 0.
4600 if (param
->sched_priority
< 0 ||
4601 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4602 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4604 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4608 * Allow unprivileged RT tasks to decrease priority:
4610 if (user
&& !capable(CAP_SYS_NICE
)) {
4611 if (rt_policy(policy
)) {
4612 unsigned long rlim_rtprio
=
4613 task_rlimit(p
, RLIMIT_RTPRIO
);
4615 /* can't set/change the rt policy */
4616 if (policy
!= p
->policy
&& !rlim_rtprio
)
4619 /* can't increase priority */
4620 if (param
->sched_priority
> p
->rt_priority
&&
4621 param
->sched_priority
> rlim_rtprio
)
4625 * Like positive nice levels, dont allow tasks to
4626 * move out of SCHED_IDLE either:
4628 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4631 /* can't change other user's priorities */
4632 if (!check_same_owner(p
))
4635 /* Normal users shall not reset the sched_reset_on_fork flag */
4636 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4641 retval
= security_task_setscheduler(p
, policy
, param
);
4647 * make sure no PI-waiters arrive (or leave) while we are
4648 * changing the priority of the task:
4650 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4652 * To be able to change p->policy safely, the apropriate
4653 * runqueue lock must be held.
4655 rq
= __task_rq_lock(p
);
4657 #ifdef CONFIG_RT_GROUP_SCHED
4660 * Do not allow realtime tasks into groups that have no runtime
4663 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4664 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4665 __task_rq_unlock(rq
);
4666 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4672 /* recheck policy now with rq lock held */
4673 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4674 policy
= oldpolicy
= -1;
4675 __task_rq_unlock(rq
);
4676 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4679 on_rq
= p
->se
.on_rq
;
4680 running
= task_current(rq
, p
);
4682 deactivate_task(rq
, p
, 0);
4684 p
->sched_class
->put_prev_task(rq
, p
);
4686 p
->sched_reset_on_fork
= reset_on_fork
;
4689 prev_class
= p
->sched_class
;
4690 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4693 p
->sched_class
->set_curr_task(rq
);
4695 activate_task(rq
, p
, 0);
4697 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4699 __task_rq_unlock(rq
);
4700 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4702 rt_mutex_adjust_pi(p
);
4708 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4709 * @p: the task in question.
4710 * @policy: new policy.
4711 * @param: structure containing the new RT priority.
4713 * NOTE that the task may be already dead.
4715 int sched_setscheduler(struct task_struct
*p
, int policy
,
4716 struct sched_param
*param
)
4718 return __sched_setscheduler(p
, policy
, param
, true);
4720 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4723 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4724 * @p: the task in question.
4725 * @policy: new policy.
4726 * @param: structure containing the new RT priority.
4728 * Just like sched_setscheduler, only don't bother checking if the
4729 * current context has permission. For example, this is needed in
4730 * stop_machine(): we create temporary high priority worker threads,
4731 * but our caller might not have that capability.
4733 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4734 struct sched_param
*param
)
4736 return __sched_setscheduler(p
, policy
, param
, false);
4740 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4742 struct sched_param lparam
;
4743 struct task_struct
*p
;
4746 if (!param
|| pid
< 0)
4748 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4753 p
= find_process_by_pid(pid
);
4755 retval
= sched_setscheduler(p
, policy
, &lparam
);
4762 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4763 * @pid: the pid in question.
4764 * @policy: new policy.
4765 * @param: structure containing the new RT priority.
4767 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4768 struct sched_param __user
*, param
)
4770 /* negative values for policy are not valid */
4774 return do_sched_setscheduler(pid
, policy
, param
);
4778 * sys_sched_setparam - set/change the RT priority of a thread
4779 * @pid: the pid in question.
4780 * @param: structure containing the new RT priority.
4782 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4784 return do_sched_setscheduler(pid
, -1, param
);
4788 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4789 * @pid: the pid in question.
4791 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4793 struct task_struct
*p
;
4801 p
= find_process_by_pid(pid
);
4803 retval
= security_task_getscheduler(p
);
4806 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4813 * sys_sched_getparam - get the RT priority of a thread
4814 * @pid: the pid in question.
4815 * @param: structure containing the RT priority.
4817 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4819 struct sched_param lp
;
4820 struct task_struct
*p
;
4823 if (!param
|| pid
< 0)
4827 p
= find_process_by_pid(pid
);
4832 retval
= security_task_getscheduler(p
);
4836 lp
.sched_priority
= p
->rt_priority
;
4840 * This one might sleep, we cannot do it with a spinlock held ...
4842 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4851 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4853 cpumask_var_t cpus_allowed
, new_mask
;
4854 struct task_struct
*p
;
4860 p
= find_process_by_pid(pid
);
4867 /* Prevent p going away */
4871 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4875 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4877 goto out_free_cpus_allowed
;
4880 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4883 retval
= security_task_setscheduler(p
, 0, NULL
);
4887 cpuset_cpus_allowed(p
, cpus_allowed
);
4888 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4890 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4893 cpuset_cpus_allowed(p
, cpus_allowed
);
4894 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4896 * We must have raced with a concurrent cpuset
4897 * update. Just reset the cpus_allowed to the
4898 * cpuset's cpus_allowed
4900 cpumask_copy(new_mask
, cpus_allowed
);
4905 free_cpumask_var(new_mask
);
4906 out_free_cpus_allowed
:
4907 free_cpumask_var(cpus_allowed
);
4914 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4915 struct cpumask
*new_mask
)
4917 if (len
< cpumask_size())
4918 cpumask_clear(new_mask
);
4919 else if (len
> cpumask_size())
4920 len
= cpumask_size();
4922 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4926 * sys_sched_setaffinity - set the cpu affinity of a process
4927 * @pid: pid of the process
4928 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4929 * @user_mask_ptr: user-space pointer to the new cpu mask
4931 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4932 unsigned long __user
*, user_mask_ptr
)
4934 cpumask_var_t new_mask
;
4937 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4940 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4942 retval
= sched_setaffinity(pid
, new_mask
);
4943 free_cpumask_var(new_mask
);
4947 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4949 struct task_struct
*p
;
4950 unsigned long flags
;
4958 p
= find_process_by_pid(pid
);
4962 retval
= security_task_getscheduler(p
);
4966 rq
= task_rq_lock(p
, &flags
);
4967 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4968 task_rq_unlock(rq
, &flags
);
4978 * sys_sched_getaffinity - get the cpu affinity of a process
4979 * @pid: pid of the process
4980 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4981 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4983 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4984 unsigned long __user
*, user_mask_ptr
)
4989 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4991 if (len
& (sizeof(unsigned long)-1))
4994 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4997 ret
= sched_getaffinity(pid
, mask
);
4999 size_t retlen
= min_t(size_t, len
, cpumask_size());
5001 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5006 free_cpumask_var(mask
);
5012 * sys_sched_yield - yield the current processor to other threads.
5014 * This function yields the current CPU to other tasks. If there are no
5015 * other threads running on this CPU then this function will return.
5017 SYSCALL_DEFINE0(sched_yield
)
5019 struct rq
*rq
= this_rq_lock();
5021 schedstat_inc(rq
, yld_count
);
5022 current
->sched_class
->yield_task(rq
);
5025 * Since we are going to call schedule() anyway, there's
5026 * no need to preempt or enable interrupts:
5028 __release(rq
->lock
);
5029 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5030 do_raw_spin_unlock(&rq
->lock
);
5031 preempt_enable_no_resched();
5038 static inline int should_resched(void)
5040 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5043 static void __cond_resched(void)
5045 add_preempt_count(PREEMPT_ACTIVE
);
5047 sub_preempt_count(PREEMPT_ACTIVE
);
5050 int __sched
_cond_resched(void)
5052 if (should_resched()) {
5058 EXPORT_SYMBOL(_cond_resched
);
5061 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5062 * call schedule, and on return reacquire the lock.
5064 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5065 * operations here to prevent schedule() from being called twice (once via
5066 * spin_unlock(), once by hand).
5068 int __cond_resched_lock(spinlock_t
*lock
)
5070 int resched
= should_resched();
5073 lockdep_assert_held(lock
);
5075 if (spin_needbreak(lock
) || resched
) {
5086 EXPORT_SYMBOL(__cond_resched_lock
);
5088 int __sched
__cond_resched_softirq(void)
5090 BUG_ON(!in_softirq());
5092 if (should_resched()) {
5100 EXPORT_SYMBOL(__cond_resched_softirq
);
5103 * yield - yield the current processor to other threads.
5105 * This is a shortcut for kernel-space yielding - it marks the
5106 * thread runnable and calls sys_sched_yield().
5108 void __sched
yield(void)
5110 set_current_state(TASK_RUNNING
);
5113 EXPORT_SYMBOL(yield
);
5116 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5117 * that process accounting knows that this is a task in IO wait state.
5119 void __sched
io_schedule(void)
5121 struct rq
*rq
= raw_rq();
5123 delayacct_blkio_start();
5124 atomic_inc(&rq
->nr_iowait
);
5125 current
->in_iowait
= 1;
5127 current
->in_iowait
= 0;
5128 atomic_dec(&rq
->nr_iowait
);
5129 delayacct_blkio_end();
5131 EXPORT_SYMBOL(io_schedule
);
5133 long __sched
io_schedule_timeout(long timeout
)
5135 struct rq
*rq
= raw_rq();
5138 delayacct_blkio_start();
5139 atomic_inc(&rq
->nr_iowait
);
5140 current
->in_iowait
= 1;
5141 ret
= schedule_timeout(timeout
);
5142 current
->in_iowait
= 0;
5143 atomic_dec(&rq
->nr_iowait
);
5144 delayacct_blkio_end();
5149 * sys_sched_get_priority_max - return maximum RT priority.
5150 * @policy: scheduling class.
5152 * this syscall returns the maximum rt_priority that can be used
5153 * by a given scheduling class.
5155 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5162 ret
= MAX_USER_RT_PRIO
-1;
5174 * sys_sched_get_priority_min - return minimum RT priority.
5175 * @policy: scheduling class.
5177 * this syscall returns the minimum rt_priority that can be used
5178 * by a given scheduling class.
5180 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5198 * sys_sched_rr_get_interval - return the default timeslice of a process.
5199 * @pid: pid of the process.
5200 * @interval: userspace pointer to the timeslice value.
5202 * this syscall writes the default timeslice value of a given process
5203 * into the user-space timespec buffer. A value of '0' means infinity.
5205 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5206 struct timespec __user
*, interval
)
5208 struct task_struct
*p
;
5209 unsigned int time_slice
;
5210 unsigned long flags
;
5220 p
= find_process_by_pid(pid
);
5224 retval
= security_task_getscheduler(p
);
5228 rq
= task_rq_lock(p
, &flags
);
5229 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5230 task_rq_unlock(rq
, &flags
);
5233 jiffies_to_timespec(time_slice
, &t
);
5234 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5242 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5244 void sched_show_task(struct task_struct
*p
)
5246 unsigned long free
= 0;
5249 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5250 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5251 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5252 #if BITS_PER_LONG == 32
5253 if (state
== TASK_RUNNING
)
5254 printk(KERN_CONT
" running ");
5256 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5258 if (state
== TASK_RUNNING
)
5259 printk(KERN_CONT
" running task ");
5261 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5263 #ifdef CONFIG_DEBUG_STACK_USAGE
5264 free
= stack_not_used(p
);
5266 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5267 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5268 (unsigned long)task_thread_info(p
)->flags
);
5270 show_stack(p
, NULL
);
5273 void show_state_filter(unsigned long state_filter
)
5275 struct task_struct
*g
, *p
;
5277 #if BITS_PER_LONG == 32
5279 " task PC stack pid father\n");
5282 " task PC stack pid father\n");
5284 read_lock(&tasklist_lock
);
5285 do_each_thread(g
, p
) {
5287 * reset the NMI-timeout, listing all files on a slow
5288 * console might take alot of time:
5290 touch_nmi_watchdog();
5291 if (!state_filter
|| (p
->state
& state_filter
))
5293 } while_each_thread(g
, p
);
5295 touch_all_softlockup_watchdogs();
5297 #ifdef CONFIG_SCHED_DEBUG
5298 sysrq_sched_debug_show();
5300 read_unlock(&tasklist_lock
);
5302 * Only show locks if all tasks are dumped:
5305 debug_show_all_locks();
5308 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5310 idle
->sched_class
= &idle_sched_class
;
5314 * init_idle - set up an idle thread for a given CPU
5315 * @idle: task in question
5316 * @cpu: cpu the idle task belongs to
5318 * NOTE: this function does not set the idle thread's NEED_RESCHED
5319 * flag, to make booting more robust.
5321 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5323 struct rq
*rq
= cpu_rq(cpu
);
5324 unsigned long flags
;
5326 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5329 idle
->state
= TASK_RUNNING
;
5330 idle
->se
.exec_start
= sched_clock();
5332 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5334 * We're having a chicken and egg problem, even though we are
5335 * holding rq->lock, the cpu isn't yet set to this cpu so the
5336 * lockdep check in task_group() will fail.
5338 * Similar case to sched_fork(). / Alternatively we could
5339 * use task_rq_lock() here and obtain the other rq->lock.
5344 __set_task_cpu(idle
, cpu
);
5347 rq
->curr
= rq
->idle
= idle
;
5348 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5351 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5353 /* Set the preempt count _outside_ the spinlocks! */
5354 #if defined(CONFIG_PREEMPT)
5355 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5357 task_thread_info(idle
)->preempt_count
= 0;
5360 * The idle tasks have their own, simple scheduling class:
5362 idle
->sched_class
= &idle_sched_class
;
5363 ftrace_graph_init_task(idle
);
5367 * In a system that switches off the HZ timer nohz_cpu_mask
5368 * indicates which cpus entered this state. This is used
5369 * in the rcu update to wait only for active cpus. For system
5370 * which do not switch off the HZ timer nohz_cpu_mask should
5371 * always be CPU_BITS_NONE.
5373 cpumask_var_t nohz_cpu_mask
;
5376 * Increase the granularity value when there are more CPUs,
5377 * because with more CPUs the 'effective latency' as visible
5378 * to users decreases. But the relationship is not linear,
5379 * so pick a second-best guess by going with the log2 of the
5382 * This idea comes from the SD scheduler of Con Kolivas:
5384 static int get_update_sysctl_factor(void)
5386 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5387 unsigned int factor
;
5389 switch (sysctl_sched_tunable_scaling
) {
5390 case SCHED_TUNABLESCALING_NONE
:
5393 case SCHED_TUNABLESCALING_LINEAR
:
5396 case SCHED_TUNABLESCALING_LOG
:
5398 factor
= 1 + ilog2(cpus
);
5405 static void update_sysctl(void)
5407 unsigned int factor
= get_update_sysctl_factor();
5409 #define SET_SYSCTL(name) \
5410 (sysctl_##name = (factor) * normalized_sysctl_##name)
5411 SET_SYSCTL(sched_min_granularity
);
5412 SET_SYSCTL(sched_latency
);
5413 SET_SYSCTL(sched_wakeup_granularity
);
5414 SET_SYSCTL(sched_shares_ratelimit
);
5418 static inline void sched_init_granularity(void)
5425 * This is how migration works:
5427 * 1) we invoke migration_cpu_stop() on the target CPU using
5429 * 2) stopper starts to run (implicitly forcing the migrated thread
5431 * 3) it checks whether the migrated task is still in the wrong runqueue.
5432 * 4) if it's in the wrong runqueue then the migration thread removes
5433 * it and puts it into the right queue.
5434 * 5) stopper completes and stop_one_cpu() returns and the migration
5439 * Change a given task's CPU affinity. Migrate the thread to a
5440 * proper CPU and schedule it away if the CPU it's executing on
5441 * is removed from the allowed bitmask.
5443 * NOTE: the caller must have a valid reference to the task, the
5444 * task must not exit() & deallocate itself prematurely. The
5445 * call is not atomic; no spinlocks may be held.
5447 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5449 unsigned long flags
;
5451 unsigned int dest_cpu
;
5455 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5456 * drop the rq->lock and still rely on ->cpus_allowed.
5459 while (task_is_waking(p
))
5461 rq
= task_rq_lock(p
, &flags
);
5462 if (task_is_waking(p
)) {
5463 task_rq_unlock(rq
, &flags
);
5467 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5472 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5473 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5478 if (p
->sched_class
->set_cpus_allowed
)
5479 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5481 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5482 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5485 /* Can the task run on the task's current CPU? If so, we're done */
5486 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5489 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5490 if (migrate_task(p
, dest_cpu
)) {
5491 struct migration_arg arg
= { p
, dest_cpu
};
5492 /* Need help from migration thread: drop lock and wait. */
5493 task_rq_unlock(rq
, &flags
);
5494 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5495 tlb_migrate_finish(p
->mm
);
5499 task_rq_unlock(rq
, &flags
);
5503 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5506 * Move (not current) task off this cpu, onto dest cpu. We're doing
5507 * this because either it can't run here any more (set_cpus_allowed()
5508 * away from this CPU, or CPU going down), or because we're
5509 * attempting to rebalance this task on exec (sched_exec).
5511 * So we race with normal scheduler movements, but that's OK, as long
5512 * as the task is no longer on this CPU.
5514 * Returns non-zero if task was successfully migrated.
5516 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5518 struct rq
*rq_dest
, *rq_src
;
5521 if (unlikely(!cpu_active(dest_cpu
)))
5524 rq_src
= cpu_rq(src_cpu
);
5525 rq_dest
= cpu_rq(dest_cpu
);
5527 double_rq_lock(rq_src
, rq_dest
);
5528 /* Already moved. */
5529 if (task_cpu(p
) != src_cpu
)
5531 /* Affinity changed (again). */
5532 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5536 * If we're not on a rq, the next wake-up will ensure we're
5540 deactivate_task(rq_src
, p
, 0);
5541 set_task_cpu(p
, dest_cpu
);
5542 activate_task(rq_dest
, p
, 0);
5543 check_preempt_curr(rq_dest
, p
, 0);
5548 double_rq_unlock(rq_src
, rq_dest
);
5553 * migration_cpu_stop - this will be executed by a highprio stopper thread
5554 * and performs thread migration by bumping thread off CPU then
5555 * 'pushing' onto another runqueue.
5557 static int migration_cpu_stop(void *data
)
5559 struct migration_arg
*arg
= data
;
5562 * The original target cpu might have gone down and we might
5563 * be on another cpu but it doesn't matter.
5565 local_irq_disable();
5566 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5571 #ifdef CONFIG_HOTPLUG_CPU
5573 * Figure out where task on dead CPU should go, use force if necessary.
5575 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5577 struct rq
*rq
= cpu_rq(dead_cpu
);
5578 int needs_cpu
, uninitialized_var(dest_cpu
);
5579 unsigned long flags
;
5581 local_irq_save(flags
);
5583 raw_spin_lock(&rq
->lock
);
5584 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5586 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5587 raw_spin_unlock(&rq
->lock
);
5589 * It can only fail if we race with set_cpus_allowed(),
5590 * in the racer should migrate the task anyway.
5593 __migrate_task(p
, dead_cpu
, dest_cpu
);
5594 local_irq_restore(flags
);
5598 * While a dead CPU has no uninterruptible tasks queued at this point,
5599 * it might still have a nonzero ->nr_uninterruptible counter, because
5600 * for performance reasons the counter is not stricly tracking tasks to
5601 * their home CPUs. So we just add the counter to another CPU's counter,
5602 * to keep the global sum constant after CPU-down:
5604 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5606 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5607 unsigned long flags
;
5609 local_irq_save(flags
);
5610 double_rq_lock(rq_src
, rq_dest
);
5611 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5612 rq_src
->nr_uninterruptible
= 0;
5613 double_rq_unlock(rq_src
, rq_dest
);
5614 local_irq_restore(flags
);
5617 /* Run through task list and migrate tasks from the dead cpu. */
5618 static void migrate_live_tasks(int src_cpu
)
5620 struct task_struct
*p
, *t
;
5622 read_lock(&tasklist_lock
);
5624 do_each_thread(t
, p
) {
5628 if (task_cpu(p
) == src_cpu
)
5629 move_task_off_dead_cpu(src_cpu
, p
);
5630 } while_each_thread(t
, p
);
5632 read_unlock(&tasklist_lock
);
5636 * Schedules idle task to be the next runnable task on current CPU.
5637 * It does so by boosting its priority to highest possible.
5638 * Used by CPU offline code.
5640 void sched_idle_next(void)
5642 int this_cpu
= smp_processor_id();
5643 struct rq
*rq
= cpu_rq(this_cpu
);
5644 struct task_struct
*p
= rq
->idle
;
5645 unsigned long flags
;
5647 /* cpu has to be offline */
5648 BUG_ON(cpu_online(this_cpu
));
5651 * Strictly not necessary since rest of the CPUs are stopped by now
5652 * and interrupts disabled on the current cpu.
5654 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5656 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5658 activate_task(rq
, p
, 0);
5660 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5664 * Ensures that the idle task is using init_mm right before its cpu goes
5667 void idle_task_exit(void)
5669 struct mm_struct
*mm
= current
->active_mm
;
5671 BUG_ON(cpu_online(smp_processor_id()));
5674 switch_mm(mm
, &init_mm
, current
);
5678 /* called under rq->lock with disabled interrupts */
5679 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5681 struct rq
*rq
= cpu_rq(dead_cpu
);
5683 /* Must be exiting, otherwise would be on tasklist. */
5684 BUG_ON(!p
->exit_state
);
5686 /* Cannot have done final schedule yet: would have vanished. */
5687 BUG_ON(p
->state
== TASK_DEAD
);
5692 * Drop lock around migration; if someone else moves it,
5693 * that's OK. No task can be added to this CPU, so iteration is
5696 raw_spin_unlock_irq(&rq
->lock
);
5697 move_task_off_dead_cpu(dead_cpu
, p
);
5698 raw_spin_lock_irq(&rq
->lock
);
5703 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5704 static void migrate_dead_tasks(unsigned int dead_cpu
)
5706 struct rq
*rq
= cpu_rq(dead_cpu
);
5707 struct task_struct
*next
;
5710 if (!rq
->nr_running
)
5712 next
= pick_next_task(rq
);
5715 next
->sched_class
->put_prev_task(rq
, next
);
5716 migrate_dead(dead_cpu
, next
);
5722 * remove the tasks which were accounted by rq from calc_load_tasks.
5724 static void calc_global_load_remove(struct rq
*rq
)
5726 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5727 rq
->calc_load_active
= 0;
5729 #endif /* CONFIG_HOTPLUG_CPU */
5731 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5733 static struct ctl_table sd_ctl_dir
[] = {
5735 .procname
= "sched_domain",
5741 static struct ctl_table sd_ctl_root
[] = {
5743 .procname
= "kernel",
5745 .child
= sd_ctl_dir
,
5750 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5752 struct ctl_table
*entry
=
5753 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5758 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5760 struct ctl_table
*entry
;
5763 * In the intermediate directories, both the child directory and
5764 * procname are dynamically allocated and could fail but the mode
5765 * will always be set. In the lowest directory the names are
5766 * static strings and all have proc handlers.
5768 for (entry
= *tablep
; entry
->mode
; entry
++) {
5770 sd_free_ctl_entry(&entry
->child
);
5771 if (entry
->proc_handler
== NULL
)
5772 kfree(entry
->procname
);
5780 set_table_entry(struct ctl_table
*entry
,
5781 const char *procname
, void *data
, int maxlen
,
5782 mode_t mode
, proc_handler
*proc_handler
)
5784 entry
->procname
= procname
;
5786 entry
->maxlen
= maxlen
;
5788 entry
->proc_handler
= proc_handler
;
5791 static struct ctl_table
*
5792 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5794 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5799 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5800 sizeof(long), 0644, proc_doulongvec_minmax
);
5801 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5802 sizeof(long), 0644, proc_doulongvec_minmax
);
5803 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5804 sizeof(int), 0644, proc_dointvec_minmax
);
5805 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5806 sizeof(int), 0644, proc_dointvec_minmax
);
5807 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5808 sizeof(int), 0644, proc_dointvec_minmax
);
5809 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5810 sizeof(int), 0644, proc_dointvec_minmax
);
5811 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5812 sizeof(int), 0644, proc_dointvec_minmax
);
5813 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5814 sizeof(int), 0644, proc_dointvec_minmax
);
5815 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5816 sizeof(int), 0644, proc_dointvec_minmax
);
5817 set_table_entry(&table
[9], "cache_nice_tries",
5818 &sd
->cache_nice_tries
,
5819 sizeof(int), 0644, proc_dointvec_minmax
);
5820 set_table_entry(&table
[10], "flags", &sd
->flags
,
5821 sizeof(int), 0644, proc_dointvec_minmax
);
5822 set_table_entry(&table
[11], "name", sd
->name
,
5823 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5824 /* &table[12] is terminator */
5829 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5831 struct ctl_table
*entry
, *table
;
5832 struct sched_domain
*sd
;
5833 int domain_num
= 0, i
;
5836 for_each_domain(cpu
, sd
)
5838 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5843 for_each_domain(cpu
, sd
) {
5844 snprintf(buf
, 32, "domain%d", i
);
5845 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5847 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5854 static struct ctl_table_header
*sd_sysctl_header
;
5855 static void register_sched_domain_sysctl(void)
5857 int i
, cpu_num
= num_possible_cpus();
5858 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5861 WARN_ON(sd_ctl_dir
[0].child
);
5862 sd_ctl_dir
[0].child
= entry
;
5867 for_each_possible_cpu(i
) {
5868 snprintf(buf
, 32, "cpu%d", i
);
5869 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5871 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5875 WARN_ON(sd_sysctl_header
);
5876 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5879 /* may be called multiple times per register */
5880 static void unregister_sched_domain_sysctl(void)
5882 if (sd_sysctl_header
)
5883 unregister_sysctl_table(sd_sysctl_header
);
5884 sd_sysctl_header
= NULL
;
5885 if (sd_ctl_dir
[0].child
)
5886 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5889 static void register_sched_domain_sysctl(void)
5892 static void unregister_sched_domain_sysctl(void)
5897 static void set_rq_online(struct rq
*rq
)
5900 const struct sched_class
*class;
5902 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5905 for_each_class(class) {
5906 if (class->rq_online
)
5907 class->rq_online(rq
);
5912 static void set_rq_offline(struct rq
*rq
)
5915 const struct sched_class
*class;
5917 for_each_class(class) {
5918 if (class->rq_offline
)
5919 class->rq_offline(rq
);
5922 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5928 * migration_call - callback that gets triggered when a CPU is added.
5929 * Here we can start up the necessary migration thread for the new CPU.
5931 static int __cpuinit
5932 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5934 int cpu
= (long)hcpu
;
5935 unsigned long flags
;
5936 struct rq
*rq
= cpu_rq(cpu
);
5940 case CPU_UP_PREPARE
:
5941 case CPU_UP_PREPARE_FROZEN
:
5942 rq
->calc_load_update
= calc_load_update
;
5946 case CPU_ONLINE_FROZEN
:
5947 /* Update our root-domain */
5948 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5950 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5954 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5957 #ifdef CONFIG_HOTPLUG_CPU
5959 case CPU_DEAD_FROZEN
:
5960 migrate_live_tasks(cpu
);
5961 /* Idle task back to normal (off runqueue, low prio) */
5962 raw_spin_lock_irq(&rq
->lock
);
5963 deactivate_task(rq
, rq
->idle
, 0);
5964 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5965 rq
->idle
->sched_class
= &idle_sched_class
;
5966 migrate_dead_tasks(cpu
);
5967 raw_spin_unlock_irq(&rq
->lock
);
5968 migrate_nr_uninterruptible(rq
);
5969 BUG_ON(rq
->nr_running
!= 0);
5970 calc_global_load_remove(rq
);
5974 case CPU_DYING_FROZEN
:
5975 /* Update our root-domain */
5976 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5978 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5981 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5989 * Register at high priority so that task migration (migrate_all_tasks)
5990 * happens before everything else. This has to be lower priority than
5991 * the notifier in the perf_event subsystem, though.
5993 static struct notifier_block __cpuinitdata migration_notifier
= {
5994 .notifier_call
= migration_call
,
5995 .priority
= CPU_PRI_MIGRATION
,
5998 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
5999 unsigned long action
, void *hcpu
)
6001 switch (action
& ~CPU_TASKS_FROZEN
) {
6003 case CPU_DOWN_FAILED
:
6004 set_cpu_active((long)hcpu
, true);
6011 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6012 unsigned long action
, void *hcpu
)
6014 switch (action
& ~CPU_TASKS_FROZEN
) {
6015 case CPU_DOWN_PREPARE
:
6016 set_cpu_active((long)hcpu
, false);
6023 static int __init
migration_init(void)
6025 void *cpu
= (void *)(long)smp_processor_id();
6028 /* Initialize migration for the boot CPU */
6029 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6030 BUG_ON(err
== NOTIFY_BAD
);
6031 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6032 register_cpu_notifier(&migration_notifier
);
6034 /* Register cpu active notifiers */
6035 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6036 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6040 early_initcall(migration_init
);
6045 #ifdef CONFIG_SCHED_DEBUG
6047 static __read_mostly
int sched_domain_debug_enabled
;
6049 static int __init
sched_domain_debug_setup(char *str
)
6051 sched_domain_debug_enabled
= 1;
6055 early_param("sched_debug", sched_domain_debug_setup
);
6057 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6058 struct cpumask
*groupmask
)
6060 struct sched_group
*group
= sd
->groups
;
6063 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6064 cpumask_clear(groupmask
);
6066 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6068 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6069 printk("does not load-balance\n");
6071 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6076 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6078 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6079 printk(KERN_ERR
"ERROR: domain->span does not contain "
6082 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6083 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6087 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6091 printk(KERN_ERR
"ERROR: group is NULL\n");
6095 if (!group
->cpu_power
) {
6096 printk(KERN_CONT
"\n");
6097 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6102 if (!cpumask_weight(sched_group_cpus(group
))) {
6103 printk(KERN_CONT
"\n");
6104 printk(KERN_ERR
"ERROR: empty group\n");
6108 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6109 printk(KERN_CONT
"\n");
6110 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6114 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6116 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6118 printk(KERN_CONT
" %s", str
);
6119 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6120 printk(KERN_CONT
" (cpu_power = %d)",
6124 group
= group
->next
;
6125 } while (group
!= sd
->groups
);
6126 printk(KERN_CONT
"\n");
6128 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6129 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6132 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6133 printk(KERN_ERR
"ERROR: parent span is not a superset "
6134 "of domain->span\n");
6138 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6140 cpumask_var_t groupmask
;
6143 if (!sched_domain_debug_enabled
)
6147 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6151 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6153 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6154 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6159 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6166 free_cpumask_var(groupmask
);
6168 #else /* !CONFIG_SCHED_DEBUG */
6169 # define sched_domain_debug(sd, cpu) do { } while (0)
6170 #endif /* CONFIG_SCHED_DEBUG */
6172 static int sd_degenerate(struct sched_domain
*sd
)
6174 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6177 /* Following flags need at least 2 groups */
6178 if (sd
->flags
& (SD_LOAD_BALANCE
|
6179 SD_BALANCE_NEWIDLE
|
6183 SD_SHARE_PKG_RESOURCES
)) {
6184 if (sd
->groups
!= sd
->groups
->next
)
6188 /* Following flags don't use groups */
6189 if (sd
->flags
& (SD_WAKE_AFFINE
))
6196 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6198 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6200 if (sd_degenerate(parent
))
6203 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6206 /* Flags needing groups don't count if only 1 group in parent */
6207 if (parent
->groups
== parent
->groups
->next
) {
6208 pflags
&= ~(SD_LOAD_BALANCE
|
6209 SD_BALANCE_NEWIDLE
|
6213 SD_SHARE_PKG_RESOURCES
);
6214 if (nr_node_ids
== 1)
6215 pflags
&= ~SD_SERIALIZE
;
6217 if (~cflags
& pflags
)
6223 static void free_rootdomain(struct root_domain
*rd
)
6225 synchronize_sched();
6227 cpupri_cleanup(&rd
->cpupri
);
6229 free_cpumask_var(rd
->rto_mask
);
6230 free_cpumask_var(rd
->online
);
6231 free_cpumask_var(rd
->span
);
6235 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6237 struct root_domain
*old_rd
= NULL
;
6238 unsigned long flags
;
6240 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6245 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6248 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6251 * If we dont want to free the old_rt yet then
6252 * set old_rd to NULL to skip the freeing later
6255 if (!atomic_dec_and_test(&old_rd
->refcount
))
6259 atomic_inc(&rd
->refcount
);
6262 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6263 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6266 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6269 free_rootdomain(old_rd
);
6272 static int init_rootdomain(struct root_domain
*rd
)
6274 memset(rd
, 0, sizeof(*rd
));
6276 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6278 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6280 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6283 if (cpupri_init(&rd
->cpupri
) != 0)
6288 free_cpumask_var(rd
->rto_mask
);
6290 free_cpumask_var(rd
->online
);
6292 free_cpumask_var(rd
->span
);
6297 static void init_defrootdomain(void)
6299 init_rootdomain(&def_root_domain
);
6301 atomic_set(&def_root_domain
.refcount
, 1);
6304 static struct root_domain
*alloc_rootdomain(void)
6306 struct root_domain
*rd
;
6308 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6312 if (init_rootdomain(rd
) != 0) {
6321 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6322 * hold the hotplug lock.
6325 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6327 struct rq
*rq
= cpu_rq(cpu
);
6328 struct sched_domain
*tmp
;
6330 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6331 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6333 /* Remove the sched domains which do not contribute to scheduling. */
6334 for (tmp
= sd
; tmp
; ) {
6335 struct sched_domain
*parent
= tmp
->parent
;
6339 if (sd_parent_degenerate(tmp
, parent
)) {
6340 tmp
->parent
= parent
->parent
;
6342 parent
->parent
->child
= tmp
;
6347 if (sd
&& sd_degenerate(sd
)) {
6353 sched_domain_debug(sd
, cpu
);
6355 rq_attach_root(rq
, rd
);
6356 rcu_assign_pointer(rq
->sd
, sd
);
6359 /* cpus with isolated domains */
6360 static cpumask_var_t cpu_isolated_map
;
6362 /* Setup the mask of cpus configured for isolated domains */
6363 static int __init
isolated_cpu_setup(char *str
)
6365 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6366 cpulist_parse(str
, cpu_isolated_map
);
6370 __setup("isolcpus=", isolated_cpu_setup
);
6373 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6374 * to a function which identifies what group(along with sched group) a CPU
6375 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6376 * (due to the fact that we keep track of groups covered with a struct cpumask).
6378 * init_sched_build_groups will build a circular linked list of the groups
6379 * covered by the given span, and will set each group's ->cpumask correctly,
6380 * and ->cpu_power to 0.
6383 init_sched_build_groups(const struct cpumask
*span
,
6384 const struct cpumask
*cpu_map
,
6385 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6386 struct sched_group
**sg
,
6387 struct cpumask
*tmpmask
),
6388 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6390 struct sched_group
*first
= NULL
, *last
= NULL
;
6393 cpumask_clear(covered
);
6395 for_each_cpu(i
, span
) {
6396 struct sched_group
*sg
;
6397 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6400 if (cpumask_test_cpu(i
, covered
))
6403 cpumask_clear(sched_group_cpus(sg
));
6406 for_each_cpu(j
, span
) {
6407 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6410 cpumask_set_cpu(j
, covered
);
6411 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6422 #define SD_NODES_PER_DOMAIN 16
6427 * find_next_best_node - find the next node to include in a sched_domain
6428 * @node: node whose sched_domain we're building
6429 * @used_nodes: nodes already in the sched_domain
6431 * Find the next node to include in a given scheduling domain. Simply
6432 * finds the closest node not already in the @used_nodes map.
6434 * Should use nodemask_t.
6436 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6438 int i
, n
, val
, min_val
, best_node
= 0;
6442 for (i
= 0; i
< nr_node_ids
; i
++) {
6443 /* Start at @node */
6444 n
= (node
+ i
) % nr_node_ids
;
6446 if (!nr_cpus_node(n
))
6449 /* Skip already used nodes */
6450 if (node_isset(n
, *used_nodes
))
6453 /* Simple min distance search */
6454 val
= node_distance(node
, n
);
6456 if (val
< min_val
) {
6462 node_set(best_node
, *used_nodes
);
6467 * sched_domain_node_span - get a cpumask for a node's sched_domain
6468 * @node: node whose cpumask we're constructing
6469 * @span: resulting cpumask
6471 * Given a node, construct a good cpumask for its sched_domain to span. It
6472 * should be one that prevents unnecessary balancing, but also spreads tasks
6475 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6477 nodemask_t used_nodes
;
6480 cpumask_clear(span
);
6481 nodes_clear(used_nodes
);
6483 cpumask_or(span
, span
, cpumask_of_node(node
));
6484 node_set(node
, used_nodes
);
6486 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6487 int next_node
= find_next_best_node(node
, &used_nodes
);
6489 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6492 #endif /* CONFIG_NUMA */
6494 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6497 * The cpus mask in sched_group and sched_domain hangs off the end.
6499 * ( See the the comments in include/linux/sched.h:struct sched_group
6500 * and struct sched_domain. )
6502 struct static_sched_group
{
6503 struct sched_group sg
;
6504 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6507 struct static_sched_domain
{
6508 struct sched_domain sd
;
6509 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6515 cpumask_var_t domainspan
;
6516 cpumask_var_t covered
;
6517 cpumask_var_t notcovered
;
6519 cpumask_var_t nodemask
;
6520 cpumask_var_t this_sibling_map
;
6521 cpumask_var_t this_core_map
;
6522 cpumask_var_t send_covered
;
6523 cpumask_var_t tmpmask
;
6524 struct sched_group
**sched_group_nodes
;
6525 struct root_domain
*rd
;
6529 sa_sched_groups
= 0,
6534 sa_this_sibling_map
,
6536 sa_sched_group_nodes
,
6546 * SMT sched-domains:
6548 #ifdef CONFIG_SCHED_SMT
6549 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6550 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6553 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6554 struct sched_group
**sg
, struct cpumask
*unused
)
6557 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6560 #endif /* CONFIG_SCHED_SMT */
6563 * multi-core sched-domains:
6565 #ifdef CONFIG_SCHED_MC
6566 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6567 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6568 #endif /* CONFIG_SCHED_MC */
6570 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6572 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6573 struct sched_group
**sg
, struct cpumask
*mask
)
6577 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6578 group
= cpumask_first(mask
);
6580 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6583 #elif defined(CONFIG_SCHED_MC)
6585 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6586 struct sched_group
**sg
, struct cpumask
*unused
)
6589 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6594 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6595 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6598 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6599 struct sched_group
**sg
, struct cpumask
*mask
)
6602 #ifdef CONFIG_SCHED_MC
6603 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6604 group
= cpumask_first(mask
);
6605 #elif defined(CONFIG_SCHED_SMT)
6606 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6607 group
= cpumask_first(mask
);
6612 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6618 * The init_sched_build_groups can't handle what we want to do with node
6619 * groups, so roll our own. Now each node has its own list of groups which
6620 * gets dynamically allocated.
6622 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6623 static struct sched_group
***sched_group_nodes_bycpu
;
6625 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6626 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6628 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6629 struct sched_group
**sg
,
6630 struct cpumask
*nodemask
)
6634 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6635 group
= cpumask_first(nodemask
);
6638 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6642 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6644 struct sched_group
*sg
= group_head
;
6650 for_each_cpu(j
, sched_group_cpus(sg
)) {
6651 struct sched_domain
*sd
;
6653 sd
= &per_cpu(phys_domains
, j
).sd
;
6654 if (j
!= group_first_cpu(sd
->groups
)) {
6656 * Only add "power" once for each
6662 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6665 } while (sg
!= group_head
);
6668 static int build_numa_sched_groups(struct s_data
*d
,
6669 const struct cpumask
*cpu_map
, int num
)
6671 struct sched_domain
*sd
;
6672 struct sched_group
*sg
, *prev
;
6675 cpumask_clear(d
->covered
);
6676 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6677 if (cpumask_empty(d
->nodemask
)) {
6678 d
->sched_group_nodes
[num
] = NULL
;
6682 sched_domain_node_span(num
, d
->domainspan
);
6683 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6685 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6688 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6692 d
->sched_group_nodes
[num
] = sg
;
6694 for_each_cpu(j
, d
->nodemask
) {
6695 sd
= &per_cpu(node_domains
, j
).sd
;
6700 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6702 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6705 for (j
= 0; j
< nr_node_ids
; j
++) {
6706 n
= (num
+ j
) % nr_node_ids
;
6707 cpumask_complement(d
->notcovered
, d
->covered
);
6708 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6709 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6710 if (cpumask_empty(d
->tmpmask
))
6712 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6713 if (cpumask_empty(d
->tmpmask
))
6715 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6719 "Can not alloc domain group for node %d\n", j
);
6723 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6724 sg
->next
= prev
->next
;
6725 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6732 #endif /* CONFIG_NUMA */
6735 /* Free memory allocated for various sched_group structures */
6736 static void free_sched_groups(const struct cpumask
*cpu_map
,
6737 struct cpumask
*nodemask
)
6741 for_each_cpu(cpu
, cpu_map
) {
6742 struct sched_group
**sched_group_nodes
6743 = sched_group_nodes_bycpu
[cpu
];
6745 if (!sched_group_nodes
)
6748 for (i
= 0; i
< nr_node_ids
; i
++) {
6749 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6751 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6752 if (cpumask_empty(nodemask
))
6762 if (oldsg
!= sched_group_nodes
[i
])
6765 kfree(sched_group_nodes
);
6766 sched_group_nodes_bycpu
[cpu
] = NULL
;
6769 #else /* !CONFIG_NUMA */
6770 static void free_sched_groups(const struct cpumask
*cpu_map
,
6771 struct cpumask
*nodemask
)
6774 #endif /* CONFIG_NUMA */
6777 * Initialize sched groups cpu_power.
6779 * cpu_power indicates the capacity of sched group, which is used while
6780 * distributing the load between different sched groups in a sched domain.
6781 * Typically cpu_power for all the groups in a sched domain will be same unless
6782 * there are asymmetries in the topology. If there are asymmetries, group
6783 * having more cpu_power will pickup more load compared to the group having
6786 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6788 struct sched_domain
*child
;
6789 struct sched_group
*group
;
6793 WARN_ON(!sd
|| !sd
->groups
);
6795 if (cpu
!= group_first_cpu(sd
->groups
))
6800 sd
->groups
->cpu_power
= 0;
6803 power
= SCHED_LOAD_SCALE
;
6804 weight
= cpumask_weight(sched_domain_span(sd
));
6806 * SMT siblings share the power of a single core.
6807 * Usually multiple threads get a better yield out of
6808 * that one core than a single thread would have,
6809 * reflect that in sd->smt_gain.
6811 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6812 power
*= sd
->smt_gain
;
6814 power
>>= SCHED_LOAD_SHIFT
;
6816 sd
->groups
->cpu_power
+= power
;
6821 * Add cpu_power of each child group to this groups cpu_power.
6823 group
= child
->groups
;
6825 sd
->groups
->cpu_power
+= group
->cpu_power
;
6826 group
= group
->next
;
6827 } while (group
!= child
->groups
);
6831 * Initializers for schedule domains
6832 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6835 #ifdef CONFIG_SCHED_DEBUG
6836 # define SD_INIT_NAME(sd, type) sd->name = #type
6838 # define SD_INIT_NAME(sd, type) do { } while (0)
6841 #define SD_INIT(sd, type) sd_init_##type(sd)
6843 #define SD_INIT_FUNC(type) \
6844 static noinline void sd_init_##type(struct sched_domain *sd) \
6846 memset(sd, 0, sizeof(*sd)); \
6847 *sd = SD_##type##_INIT; \
6848 sd->level = SD_LV_##type; \
6849 SD_INIT_NAME(sd, type); \
6854 SD_INIT_FUNC(ALLNODES
)
6857 #ifdef CONFIG_SCHED_SMT
6858 SD_INIT_FUNC(SIBLING
)
6860 #ifdef CONFIG_SCHED_MC
6864 static int default_relax_domain_level
= -1;
6866 static int __init
setup_relax_domain_level(char *str
)
6870 val
= simple_strtoul(str
, NULL
, 0);
6871 if (val
< SD_LV_MAX
)
6872 default_relax_domain_level
= val
;
6876 __setup("relax_domain_level=", setup_relax_domain_level
);
6878 static void set_domain_attribute(struct sched_domain
*sd
,
6879 struct sched_domain_attr
*attr
)
6883 if (!attr
|| attr
->relax_domain_level
< 0) {
6884 if (default_relax_domain_level
< 0)
6887 request
= default_relax_domain_level
;
6889 request
= attr
->relax_domain_level
;
6890 if (request
< sd
->level
) {
6891 /* turn off idle balance on this domain */
6892 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6894 /* turn on idle balance on this domain */
6895 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6899 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6900 const struct cpumask
*cpu_map
)
6903 case sa_sched_groups
:
6904 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6905 d
->sched_group_nodes
= NULL
;
6907 free_rootdomain(d
->rd
); /* fall through */
6909 free_cpumask_var(d
->tmpmask
); /* fall through */
6910 case sa_send_covered
:
6911 free_cpumask_var(d
->send_covered
); /* fall through */
6912 case sa_this_core_map
:
6913 free_cpumask_var(d
->this_core_map
); /* fall through */
6914 case sa_this_sibling_map
:
6915 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6917 free_cpumask_var(d
->nodemask
); /* fall through */
6918 case sa_sched_group_nodes
:
6920 kfree(d
->sched_group_nodes
); /* fall through */
6922 free_cpumask_var(d
->notcovered
); /* fall through */
6924 free_cpumask_var(d
->covered
); /* fall through */
6926 free_cpumask_var(d
->domainspan
); /* fall through */
6933 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6934 const struct cpumask
*cpu_map
)
6937 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6939 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6940 return sa_domainspan
;
6941 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6943 /* Allocate the per-node list of sched groups */
6944 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6945 sizeof(struct sched_group
*), GFP_KERNEL
);
6946 if (!d
->sched_group_nodes
) {
6947 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6948 return sa_notcovered
;
6950 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6952 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6953 return sa_sched_group_nodes
;
6954 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6956 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6957 return sa_this_sibling_map
;
6958 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6959 return sa_this_core_map
;
6960 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6961 return sa_send_covered
;
6962 d
->rd
= alloc_rootdomain();
6964 printk(KERN_WARNING
"Cannot alloc root domain\n");
6967 return sa_rootdomain
;
6970 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6971 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6973 struct sched_domain
*sd
= NULL
;
6975 struct sched_domain
*parent
;
6978 if (cpumask_weight(cpu_map
) >
6979 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6980 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6981 SD_INIT(sd
, ALLNODES
);
6982 set_domain_attribute(sd
, attr
);
6983 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6984 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6989 sd
= &per_cpu(node_domains
, i
).sd
;
6991 set_domain_attribute(sd
, attr
);
6992 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6993 sd
->parent
= parent
;
6996 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7001 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7002 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7003 struct sched_domain
*parent
, int i
)
7005 struct sched_domain
*sd
;
7006 sd
= &per_cpu(phys_domains
, i
).sd
;
7008 set_domain_attribute(sd
, attr
);
7009 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7010 sd
->parent
= parent
;
7013 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7017 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7018 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7019 struct sched_domain
*parent
, int i
)
7021 struct sched_domain
*sd
= parent
;
7022 #ifdef CONFIG_SCHED_MC
7023 sd
= &per_cpu(core_domains
, i
).sd
;
7025 set_domain_attribute(sd
, attr
);
7026 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7027 sd
->parent
= parent
;
7029 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7034 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7035 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7036 struct sched_domain
*parent
, int i
)
7038 struct sched_domain
*sd
= parent
;
7039 #ifdef CONFIG_SCHED_SMT
7040 sd
= &per_cpu(cpu_domains
, i
).sd
;
7041 SD_INIT(sd
, SIBLING
);
7042 set_domain_attribute(sd
, attr
);
7043 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7044 sd
->parent
= parent
;
7046 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7051 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7052 const struct cpumask
*cpu_map
, int cpu
)
7055 #ifdef CONFIG_SCHED_SMT
7056 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7057 cpumask_and(d
->this_sibling_map
, cpu_map
,
7058 topology_thread_cpumask(cpu
));
7059 if (cpu
== cpumask_first(d
->this_sibling_map
))
7060 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7062 d
->send_covered
, d
->tmpmask
);
7065 #ifdef CONFIG_SCHED_MC
7066 case SD_LV_MC
: /* set up multi-core groups */
7067 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7068 if (cpu
== cpumask_first(d
->this_core_map
))
7069 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7071 d
->send_covered
, d
->tmpmask
);
7074 case SD_LV_CPU
: /* set up physical groups */
7075 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7076 if (!cpumask_empty(d
->nodemask
))
7077 init_sched_build_groups(d
->nodemask
, cpu_map
,
7079 d
->send_covered
, d
->tmpmask
);
7082 case SD_LV_ALLNODES
:
7083 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7084 d
->send_covered
, d
->tmpmask
);
7093 * Build sched domains for a given set of cpus and attach the sched domains
7094 * to the individual cpus
7096 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7097 struct sched_domain_attr
*attr
)
7099 enum s_alloc alloc_state
= sa_none
;
7101 struct sched_domain
*sd
;
7107 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7108 if (alloc_state
!= sa_rootdomain
)
7110 alloc_state
= sa_sched_groups
;
7113 * Set up domains for cpus specified by the cpu_map.
7115 for_each_cpu(i
, cpu_map
) {
7116 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7119 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7120 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7121 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7122 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7125 for_each_cpu(i
, cpu_map
) {
7126 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7127 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7130 /* Set up physical groups */
7131 for (i
= 0; i
< nr_node_ids
; i
++)
7132 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7135 /* Set up node groups */
7137 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7139 for (i
= 0; i
< nr_node_ids
; i
++)
7140 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7144 /* Calculate CPU power for physical packages and nodes */
7145 #ifdef CONFIG_SCHED_SMT
7146 for_each_cpu(i
, cpu_map
) {
7147 sd
= &per_cpu(cpu_domains
, i
).sd
;
7148 init_sched_groups_power(i
, sd
);
7151 #ifdef CONFIG_SCHED_MC
7152 for_each_cpu(i
, cpu_map
) {
7153 sd
= &per_cpu(core_domains
, i
).sd
;
7154 init_sched_groups_power(i
, sd
);
7158 for_each_cpu(i
, cpu_map
) {
7159 sd
= &per_cpu(phys_domains
, i
).sd
;
7160 init_sched_groups_power(i
, sd
);
7164 for (i
= 0; i
< nr_node_ids
; i
++)
7165 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7167 if (d
.sd_allnodes
) {
7168 struct sched_group
*sg
;
7170 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7172 init_numa_sched_groups_power(sg
);
7176 /* Attach the domains */
7177 for_each_cpu(i
, cpu_map
) {
7178 #ifdef CONFIG_SCHED_SMT
7179 sd
= &per_cpu(cpu_domains
, i
).sd
;
7180 #elif defined(CONFIG_SCHED_MC)
7181 sd
= &per_cpu(core_domains
, i
).sd
;
7183 sd
= &per_cpu(phys_domains
, i
).sd
;
7185 cpu_attach_domain(sd
, d
.rd
, i
);
7188 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7189 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7193 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7197 static int build_sched_domains(const struct cpumask
*cpu_map
)
7199 return __build_sched_domains(cpu_map
, NULL
);
7202 static cpumask_var_t
*doms_cur
; /* current sched domains */
7203 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7204 static struct sched_domain_attr
*dattr_cur
;
7205 /* attribues of custom domains in 'doms_cur' */
7208 * Special case: If a kmalloc of a doms_cur partition (array of
7209 * cpumask) fails, then fallback to a single sched domain,
7210 * as determined by the single cpumask fallback_doms.
7212 static cpumask_var_t fallback_doms
;
7215 * arch_update_cpu_topology lets virtualized architectures update the
7216 * cpu core maps. It is supposed to return 1 if the topology changed
7217 * or 0 if it stayed the same.
7219 int __attribute__((weak
)) arch_update_cpu_topology(void)
7224 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7227 cpumask_var_t
*doms
;
7229 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7232 for (i
= 0; i
< ndoms
; i
++) {
7233 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7234 free_sched_domains(doms
, i
);
7241 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7244 for (i
= 0; i
< ndoms
; i
++)
7245 free_cpumask_var(doms
[i
]);
7250 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7251 * For now this just excludes isolated cpus, but could be used to
7252 * exclude other special cases in the future.
7254 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7258 arch_update_cpu_topology();
7260 doms_cur
= alloc_sched_domains(ndoms_cur
);
7262 doms_cur
= &fallback_doms
;
7263 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7265 err
= build_sched_domains(doms_cur
[0]);
7266 register_sched_domain_sysctl();
7271 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7272 struct cpumask
*tmpmask
)
7274 free_sched_groups(cpu_map
, tmpmask
);
7278 * Detach sched domains from a group of cpus specified in cpu_map
7279 * These cpus will now be attached to the NULL domain
7281 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7283 /* Save because hotplug lock held. */
7284 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7287 for_each_cpu(i
, cpu_map
)
7288 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7289 synchronize_sched();
7290 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7293 /* handle null as "default" */
7294 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7295 struct sched_domain_attr
*new, int idx_new
)
7297 struct sched_domain_attr tmp
;
7304 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7305 new ? (new + idx_new
) : &tmp
,
7306 sizeof(struct sched_domain_attr
));
7310 * Partition sched domains as specified by the 'ndoms_new'
7311 * cpumasks in the array doms_new[] of cpumasks. This compares
7312 * doms_new[] to the current sched domain partitioning, doms_cur[].
7313 * It destroys each deleted domain and builds each new domain.
7315 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7316 * The masks don't intersect (don't overlap.) We should setup one
7317 * sched domain for each mask. CPUs not in any of the cpumasks will
7318 * not be load balanced. If the same cpumask appears both in the
7319 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7322 * The passed in 'doms_new' should be allocated using
7323 * alloc_sched_domains. This routine takes ownership of it and will
7324 * free_sched_domains it when done with it. If the caller failed the
7325 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7326 * and partition_sched_domains() will fallback to the single partition
7327 * 'fallback_doms', it also forces the domains to be rebuilt.
7329 * If doms_new == NULL it will be replaced with cpu_online_mask.
7330 * ndoms_new == 0 is a special case for destroying existing domains,
7331 * and it will not create the default domain.
7333 * Call with hotplug lock held
7335 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7336 struct sched_domain_attr
*dattr_new
)
7341 mutex_lock(&sched_domains_mutex
);
7343 /* always unregister in case we don't destroy any domains */
7344 unregister_sched_domain_sysctl();
7346 /* Let architecture update cpu core mappings. */
7347 new_topology
= arch_update_cpu_topology();
7349 n
= doms_new
? ndoms_new
: 0;
7351 /* Destroy deleted domains */
7352 for (i
= 0; i
< ndoms_cur
; i
++) {
7353 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7354 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7355 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7358 /* no match - a current sched domain not in new doms_new[] */
7359 detach_destroy_domains(doms_cur
[i
]);
7364 if (doms_new
== NULL
) {
7366 doms_new
= &fallback_doms
;
7367 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7368 WARN_ON_ONCE(dattr_new
);
7371 /* Build new domains */
7372 for (i
= 0; i
< ndoms_new
; i
++) {
7373 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7374 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7375 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7378 /* no match - add a new doms_new */
7379 __build_sched_domains(doms_new
[i
],
7380 dattr_new
? dattr_new
+ i
: NULL
);
7385 /* Remember the new sched domains */
7386 if (doms_cur
!= &fallback_doms
)
7387 free_sched_domains(doms_cur
, ndoms_cur
);
7388 kfree(dattr_cur
); /* kfree(NULL) is safe */
7389 doms_cur
= doms_new
;
7390 dattr_cur
= dattr_new
;
7391 ndoms_cur
= ndoms_new
;
7393 register_sched_domain_sysctl();
7395 mutex_unlock(&sched_domains_mutex
);
7398 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7399 static void arch_reinit_sched_domains(void)
7403 /* Destroy domains first to force the rebuild */
7404 partition_sched_domains(0, NULL
, NULL
);
7406 rebuild_sched_domains();
7410 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7412 unsigned int level
= 0;
7414 if (sscanf(buf
, "%u", &level
) != 1)
7418 * level is always be positive so don't check for
7419 * level < POWERSAVINGS_BALANCE_NONE which is 0
7420 * What happens on 0 or 1 byte write,
7421 * need to check for count as well?
7424 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7428 sched_smt_power_savings
= level
;
7430 sched_mc_power_savings
= level
;
7432 arch_reinit_sched_domains();
7437 #ifdef CONFIG_SCHED_MC
7438 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7439 struct sysdev_class_attribute
*attr
,
7442 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7444 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7445 struct sysdev_class_attribute
*attr
,
7446 const char *buf
, size_t count
)
7448 return sched_power_savings_store(buf
, count
, 0);
7450 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7451 sched_mc_power_savings_show
,
7452 sched_mc_power_savings_store
);
7455 #ifdef CONFIG_SCHED_SMT
7456 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7457 struct sysdev_class_attribute
*attr
,
7460 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7462 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7463 struct sysdev_class_attribute
*attr
,
7464 const char *buf
, size_t count
)
7466 return sched_power_savings_store(buf
, count
, 1);
7468 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7469 sched_smt_power_savings_show
,
7470 sched_smt_power_savings_store
);
7473 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7477 #ifdef CONFIG_SCHED_SMT
7479 err
= sysfs_create_file(&cls
->kset
.kobj
,
7480 &attr_sched_smt_power_savings
.attr
);
7482 #ifdef CONFIG_SCHED_MC
7483 if (!err
&& mc_capable())
7484 err
= sysfs_create_file(&cls
->kset
.kobj
,
7485 &attr_sched_mc_power_savings
.attr
);
7489 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7492 * Update cpusets according to cpu_active mask. If cpusets are
7493 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7494 * around partition_sched_domains().
7496 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7499 switch (action
& ~CPU_TASKS_FROZEN
) {
7501 case CPU_DOWN_FAILED
:
7502 cpuset_update_active_cpus();
7509 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7512 switch (action
& ~CPU_TASKS_FROZEN
) {
7513 case CPU_DOWN_PREPARE
:
7514 cpuset_update_active_cpus();
7521 static int update_runtime(struct notifier_block
*nfb
,
7522 unsigned long action
, void *hcpu
)
7524 int cpu
= (int)(long)hcpu
;
7527 case CPU_DOWN_PREPARE
:
7528 case CPU_DOWN_PREPARE_FROZEN
:
7529 disable_runtime(cpu_rq(cpu
));
7532 case CPU_DOWN_FAILED
:
7533 case CPU_DOWN_FAILED_FROZEN
:
7535 case CPU_ONLINE_FROZEN
:
7536 enable_runtime(cpu_rq(cpu
));
7544 void __init
sched_init_smp(void)
7546 cpumask_var_t non_isolated_cpus
;
7548 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7549 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7551 #if defined(CONFIG_NUMA)
7552 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7554 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7557 mutex_lock(&sched_domains_mutex
);
7558 arch_init_sched_domains(cpu_active_mask
);
7559 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7560 if (cpumask_empty(non_isolated_cpus
))
7561 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7562 mutex_unlock(&sched_domains_mutex
);
7565 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7566 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7568 /* RT runtime code needs to handle some hotplug events */
7569 hotcpu_notifier(update_runtime
, 0);
7573 /* Move init over to a non-isolated CPU */
7574 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7576 sched_init_granularity();
7577 free_cpumask_var(non_isolated_cpus
);
7579 init_sched_rt_class();
7582 void __init
sched_init_smp(void)
7584 sched_init_granularity();
7586 #endif /* CONFIG_SMP */
7588 const_debug
unsigned int sysctl_timer_migration
= 1;
7590 int in_sched_functions(unsigned long addr
)
7592 return in_lock_functions(addr
) ||
7593 (addr
>= (unsigned long)__sched_text_start
7594 && addr
< (unsigned long)__sched_text_end
);
7597 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7599 cfs_rq
->tasks_timeline
= RB_ROOT
;
7600 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7601 #ifdef CONFIG_FAIR_GROUP_SCHED
7604 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7607 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7609 struct rt_prio_array
*array
;
7612 array
= &rt_rq
->active
;
7613 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7614 INIT_LIST_HEAD(array
->queue
+ i
);
7615 __clear_bit(i
, array
->bitmap
);
7617 /* delimiter for bitsearch: */
7618 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7620 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7621 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7623 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7627 rt_rq
->rt_nr_migratory
= 0;
7628 rt_rq
->overloaded
= 0;
7629 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7633 rt_rq
->rt_throttled
= 0;
7634 rt_rq
->rt_runtime
= 0;
7635 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7637 #ifdef CONFIG_RT_GROUP_SCHED
7638 rt_rq
->rt_nr_boosted
= 0;
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7645 struct sched_entity
*se
, int cpu
, int add
,
7646 struct sched_entity
*parent
)
7648 struct rq
*rq
= cpu_rq(cpu
);
7649 tg
->cfs_rq
[cpu
] = cfs_rq
;
7650 init_cfs_rq(cfs_rq
, rq
);
7653 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7656 /* se could be NULL for init_task_group */
7661 se
->cfs_rq
= &rq
->cfs
;
7663 se
->cfs_rq
= parent
->my_q
;
7666 se
->load
.weight
= tg
->shares
;
7667 se
->load
.inv_weight
= 0;
7668 se
->parent
= parent
;
7672 #ifdef CONFIG_RT_GROUP_SCHED
7673 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7674 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7675 struct sched_rt_entity
*parent
)
7677 struct rq
*rq
= cpu_rq(cpu
);
7679 tg
->rt_rq
[cpu
] = rt_rq
;
7680 init_rt_rq(rt_rq
, rq
);
7682 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7684 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7686 tg
->rt_se
[cpu
] = rt_se
;
7691 rt_se
->rt_rq
= &rq
->rt
;
7693 rt_se
->rt_rq
= parent
->my_q
;
7695 rt_se
->my_q
= rt_rq
;
7696 rt_se
->parent
= parent
;
7697 INIT_LIST_HEAD(&rt_se
->run_list
);
7701 void __init
sched_init(void)
7704 unsigned long alloc_size
= 0, ptr
;
7706 #ifdef CONFIG_FAIR_GROUP_SCHED
7707 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7709 #ifdef CONFIG_RT_GROUP_SCHED
7710 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7712 #ifdef CONFIG_CPUMASK_OFFSTACK
7713 alloc_size
+= num_possible_cpus() * cpumask_size();
7716 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7718 #ifdef CONFIG_FAIR_GROUP_SCHED
7719 init_task_group
.se
= (struct sched_entity
**)ptr
;
7720 ptr
+= nr_cpu_ids
* sizeof(void **);
7722 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7723 ptr
+= nr_cpu_ids
* sizeof(void **);
7725 #endif /* CONFIG_FAIR_GROUP_SCHED */
7726 #ifdef CONFIG_RT_GROUP_SCHED
7727 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7728 ptr
+= nr_cpu_ids
* sizeof(void **);
7730 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7731 ptr
+= nr_cpu_ids
* sizeof(void **);
7733 #endif /* CONFIG_RT_GROUP_SCHED */
7734 #ifdef CONFIG_CPUMASK_OFFSTACK
7735 for_each_possible_cpu(i
) {
7736 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7737 ptr
+= cpumask_size();
7739 #endif /* CONFIG_CPUMASK_OFFSTACK */
7743 init_defrootdomain();
7746 init_rt_bandwidth(&def_rt_bandwidth
,
7747 global_rt_period(), global_rt_runtime());
7749 #ifdef CONFIG_RT_GROUP_SCHED
7750 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7751 global_rt_period(), global_rt_runtime());
7752 #endif /* CONFIG_RT_GROUP_SCHED */
7754 #ifdef CONFIG_CGROUP_SCHED
7755 list_add(&init_task_group
.list
, &task_groups
);
7756 INIT_LIST_HEAD(&init_task_group
.children
);
7758 #endif /* CONFIG_CGROUP_SCHED */
7760 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7761 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7762 __alignof__(unsigned long));
7764 for_each_possible_cpu(i
) {
7768 raw_spin_lock_init(&rq
->lock
);
7770 rq
->calc_load_active
= 0;
7771 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7772 init_cfs_rq(&rq
->cfs
, rq
);
7773 init_rt_rq(&rq
->rt
, rq
);
7774 #ifdef CONFIG_FAIR_GROUP_SCHED
7775 init_task_group
.shares
= init_task_group_load
;
7776 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7777 #ifdef CONFIG_CGROUP_SCHED
7779 * How much cpu bandwidth does init_task_group get?
7781 * In case of task-groups formed thr' the cgroup filesystem, it
7782 * gets 100% of the cpu resources in the system. This overall
7783 * system cpu resource is divided among the tasks of
7784 * init_task_group and its child task-groups in a fair manner,
7785 * based on each entity's (task or task-group's) weight
7786 * (se->load.weight).
7788 * In other words, if init_task_group has 10 tasks of weight
7789 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7790 * then A0's share of the cpu resource is:
7792 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7794 * We achieve this by letting init_task_group's tasks sit
7795 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7797 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7799 #endif /* CONFIG_FAIR_GROUP_SCHED */
7801 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7802 #ifdef CONFIG_RT_GROUP_SCHED
7803 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7804 #ifdef CONFIG_CGROUP_SCHED
7805 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7809 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7810 rq
->cpu_load
[j
] = 0;
7812 rq
->last_load_update_tick
= jiffies
;
7817 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7818 rq
->post_schedule
= 0;
7819 rq
->active_balance
= 0;
7820 rq
->next_balance
= jiffies
;
7825 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7826 rq_attach_root(rq
, &def_root_domain
);
7828 rq
->nohz_balance_kick
= 0;
7829 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
7833 atomic_set(&rq
->nr_iowait
, 0);
7836 set_load_weight(&init_task
);
7838 #ifdef CONFIG_PREEMPT_NOTIFIERS
7839 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7843 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7846 #ifdef CONFIG_RT_MUTEXES
7847 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7851 * The boot idle thread does lazy MMU switching as well:
7853 atomic_inc(&init_mm
.mm_count
);
7854 enter_lazy_tlb(&init_mm
, current
);
7857 * Make us the idle thread. Technically, schedule() should not be
7858 * called from this thread, however somewhere below it might be,
7859 * but because we are the idle thread, we just pick up running again
7860 * when this runqueue becomes "idle".
7862 init_idle(current
, smp_processor_id());
7864 calc_load_update
= jiffies
+ LOAD_FREQ
;
7867 * During early bootup we pretend to be a normal task:
7869 current
->sched_class
= &fair_sched_class
;
7871 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7872 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7875 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
7876 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
7877 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
7878 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
7879 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
7881 /* May be allocated at isolcpus cmdline parse time */
7882 if (cpu_isolated_map
== NULL
)
7883 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7888 scheduler_running
= 1;
7891 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7892 static inline int preempt_count_equals(int preempt_offset
)
7894 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7896 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7899 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7902 static unsigned long prev_jiffy
; /* ratelimiting */
7904 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7905 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7907 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7909 prev_jiffy
= jiffies
;
7912 "BUG: sleeping function called from invalid context at %s:%d\n",
7915 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7916 in_atomic(), irqs_disabled(),
7917 current
->pid
, current
->comm
);
7919 debug_show_held_locks(current
);
7920 if (irqs_disabled())
7921 print_irqtrace_events(current
);
7925 EXPORT_SYMBOL(__might_sleep
);
7928 #ifdef CONFIG_MAGIC_SYSRQ
7929 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7933 on_rq
= p
->se
.on_rq
;
7935 deactivate_task(rq
, p
, 0);
7936 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7938 activate_task(rq
, p
, 0);
7939 resched_task(rq
->curr
);
7943 void normalize_rt_tasks(void)
7945 struct task_struct
*g
, *p
;
7946 unsigned long flags
;
7949 read_lock_irqsave(&tasklist_lock
, flags
);
7950 do_each_thread(g
, p
) {
7952 * Only normalize user tasks:
7957 p
->se
.exec_start
= 0;
7958 #ifdef CONFIG_SCHEDSTATS
7959 p
->se
.statistics
.wait_start
= 0;
7960 p
->se
.statistics
.sleep_start
= 0;
7961 p
->se
.statistics
.block_start
= 0;
7966 * Renice negative nice level userspace
7969 if (TASK_NICE(p
) < 0 && p
->mm
)
7970 set_user_nice(p
, 0);
7974 raw_spin_lock(&p
->pi_lock
);
7975 rq
= __task_rq_lock(p
);
7977 normalize_task(rq
, p
);
7979 __task_rq_unlock(rq
);
7980 raw_spin_unlock(&p
->pi_lock
);
7981 } while_each_thread(g
, p
);
7983 read_unlock_irqrestore(&tasklist_lock
, flags
);
7986 #endif /* CONFIG_MAGIC_SYSRQ */
7988 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7990 * These functions are only useful for the IA64 MCA handling, or kdb.
7992 * They can only be called when the whole system has been
7993 * stopped - every CPU needs to be quiescent, and no scheduling
7994 * activity can take place. Using them for anything else would
7995 * be a serious bug, and as a result, they aren't even visible
7996 * under any other configuration.
8000 * curr_task - return the current task for a given cpu.
8001 * @cpu: the processor in question.
8003 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8005 struct task_struct
*curr_task(int cpu
)
8007 return cpu_curr(cpu
);
8010 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8014 * set_curr_task - set the current task for a given cpu.
8015 * @cpu: the processor in question.
8016 * @p: the task pointer to set.
8018 * Description: This function must only be used when non-maskable interrupts
8019 * are serviced on a separate stack. It allows the architecture to switch the
8020 * notion of the current task on a cpu in a non-blocking manner. This function
8021 * must be called with all CPU's synchronized, and interrupts disabled, the
8022 * and caller must save the original value of the current task (see
8023 * curr_task() above) and restore that value before reenabling interrupts and
8024 * re-starting the system.
8026 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8028 void set_curr_task(int cpu
, struct task_struct
*p
)
8035 #ifdef CONFIG_FAIR_GROUP_SCHED
8036 static void free_fair_sched_group(struct task_group
*tg
)
8040 for_each_possible_cpu(i
) {
8042 kfree(tg
->cfs_rq
[i
]);
8052 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8054 struct cfs_rq
*cfs_rq
;
8055 struct sched_entity
*se
;
8059 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8062 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8066 tg
->shares
= NICE_0_LOAD
;
8068 for_each_possible_cpu(i
) {
8071 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8072 GFP_KERNEL
, cpu_to_node(i
));
8076 se
= kzalloc_node(sizeof(struct sched_entity
),
8077 GFP_KERNEL
, cpu_to_node(i
));
8081 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8092 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8094 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8095 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8098 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8100 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8102 #else /* !CONFG_FAIR_GROUP_SCHED */
8103 static inline void free_fair_sched_group(struct task_group
*tg
)
8108 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8113 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8117 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8120 #endif /* CONFIG_FAIR_GROUP_SCHED */
8122 #ifdef CONFIG_RT_GROUP_SCHED
8123 static void free_rt_sched_group(struct task_group
*tg
)
8127 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8129 for_each_possible_cpu(i
) {
8131 kfree(tg
->rt_rq
[i
]);
8133 kfree(tg
->rt_se
[i
]);
8141 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8143 struct rt_rq
*rt_rq
;
8144 struct sched_rt_entity
*rt_se
;
8148 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8151 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8155 init_rt_bandwidth(&tg
->rt_bandwidth
,
8156 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8158 for_each_possible_cpu(i
) {
8161 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8162 GFP_KERNEL
, cpu_to_node(i
));
8166 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8167 GFP_KERNEL
, cpu_to_node(i
));
8171 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8182 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8184 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8185 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8188 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8190 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8192 #else /* !CONFIG_RT_GROUP_SCHED */
8193 static inline void free_rt_sched_group(struct task_group
*tg
)
8198 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8203 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8207 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8210 #endif /* CONFIG_RT_GROUP_SCHED */
8212 #ifdef CONFIG_CGROUP_SCHED
8213 static void free_sched_group(struct task_group
*tg
)
8215 free_fair_sched_group(tg
);
8216 free_rt_sched_group(tg
);
8220 /* allocate runqueue etc for a new task group */
8221 struct task_group
*sched_create_group(struct task_group
*parent
)
8223 struct task_group
*tg
;
8224 unsigned long flags
;
8227 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8229 return ERR_PTR(-ENOMEM
);
8231 if (!alloc_fair_sched_group(tg
, parent
))
8234 if (!alloc_rt_sched_group(tg
, parent
))
8237 spin_lock_irqsave(&task_group_lock
, flags
);
8238 for_each_possible_cpu(i
) {
8239 register_fair_sched_group(tg
, i
);
8240 register_rt_sched_group(tg
, i
);
8242 list_add_rcu(&tg
->list
, &task_groups
);
8244 WARN_ON(!parent
); /* root should already exist */
8246 tg
->parent
= parent
;
8247 INIT_LIST_HEAD(&tg
->children
);
8248 list_add_rcu(&tg
->siblings
, &parent
->children
);
8249 spin_unlock_irqrestore(&task_group_lock
, flags
);
8254 free_sched_group(tg
);
8255 return ERR_PTR(-ENOMEM
);
8258 /* rcu callback to free various structures associated with a task group */
8259 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8261 /* now it should be safe to free those cfs_rqs */
8262 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8265 /* Destroy runqueue etc associated with a task group */
8266 void sched_destroy_group(struct task_group
*tg
)
8268 unsigned long flags
;
8271 spin_lock_irqsave(&task_group_lock
, flags
);
8272 for_each_possible_cpu(i
) {
8273 unregister_fair_sched_group(tg
, i
);
8274 unregister_rt_sched_group(tg
, i
);
8276 list_del_rcu(&tg
->list
);
8277 list_del_rcu(&tg
->siblings
);
8278 spin_unlock_irqrestore(&task_group_lock
, flags
);
8280 /* wait for possible concurrent references to cfs_rqs complete */
8281 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8284 /* change task's runqueue when it moves between groups.
8285 * The caller of this function should have put the task in its new group
8286 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8287 * reflect its new group.
8289 void sched_move_task(struct task_struct
*tsk
)
8292 unsigned long flags
;
8295 rq
= task_rq_lock(tsk
, &flags
);
8297 running
= task_current(rq
, tsk
);
8298 on_rq
= tsk
->se
.on_rq
;
8301 dequeue_task(rq
, tsk
, 0);
8302 if (unlikely(running
))
8303 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8305 set_task_rq(tsk
, task_cpu(tsk
));
8307 #ifdef CONFIG_FAIR_GROUP_SCHED
8308 if (tsk
->sched_class
->moved_group
)
8309 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8312 if (unlikely(running
))
8313 tsk
->sched_class
->set_curr_task(rq
);
8315 enqueue_task(rq
, tsk
, 0);
8317 task_rq_unlock(rq
, &flags
);
8319 #endif /* CONFIG_CGROUP_SCHED */
8321 #ifdef CONFIG_FAIR_GROUP_SCHED
8322 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8324 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8329 dequeue_entity(cfs_rq
, se
, 0);
8331 se
->load
.weight
= shares
;
8332 se
->load
.inv_weight
= 0;
8335 enqueue_entity(cfs_rq
, se
, 0);
8338 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8340 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8341 struct rq
*rq
= cfs_rq
->rq
;
8342 unsigned long flags
;
8344 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8345 __set_se_shares(se
, shares
);
8346 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8349 static DEFINE_MUTEX(shares_mutex
);
8351 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8354 unsigned long flags
;
8357 * We can't change the weight of the root cgroup.
8362 if (shares
< MIN_SHARES
)
8363 shares
= MIN_SHARES
;
8364 else if (shares
> MAX_SHARES
)
8365 shares
= MAX_SHARES
;
8367 mutex_lock(&shares_mutex
);
8368 if (tg
->shares
== shares
)
8371 spin_lock_irqsave(&task_group_lock
, flags
);
8372 for_each_possible_cpu(i
)
8373 unregister_fair_sched_group(tg
, i
);
8374 list_del_rcu(&tg
->siblings
);
8375 spin_unlock_irqrestore(&task_group_lock
, flags
);
8377 /* wait for any ongoing reference to this group to finish */
8378 synchronize_sched();
8381 * Now we are free to modify the group's share on each cpu
8382 * w/o tripping rebalance_share or load_balance_fair.
8384 tg
->shares
= shares
;
8385 for_each_possible_cpu(i
) {
8389 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8390 set_se_shares(tg
->se
[i
], shares
);
8394 * Enable load balance activity on this group, by inserting it back on
8395 * each cpu's rq->leaf_cfs_rq_list.
8397 spin_lock_irqsave(&task_group_lock
, flags
);
8398 for_each_possible_cpu(i
)
8399 register_fair_sched_group(tg
, i
);
8400 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8401 spin_unlock_irqrestore(&task_group_lock
, flags
);
8403 mutex_unlock(&shares_mutex
);
8407 unsigned long sched_group_shares(struct task_group
*tg
)
8413 #ifdef CONFIG_RT_GROUP_SCHED
8415 * Ensure that the real time constraints are schedulable.
8417 static DEFINE_MUTEX(rt_constraints_mutex
);
8419 static unsigned long to_ratio(u64 period
, u64 runtime
)
8421 if (runtime
== RUNTIME_INF
)
8424 return div64_u64(runtime
<< 20, period
);
8427 /* Must be called with tasklist_lock held */
8428 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8430 struct task_struct
*g
, *p
;
8432 do_each_thread(g
, p
) {
8433 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8435 } while_each_thread(g
, p
);
8440 struct rt_schedulable_data
{
8441 struct task_group
*tg
;
8446 static int tg_schedulable(struct task_group
*tg
, void *data
)
8448 struct rt_schedulable_data
*d
= data
;
8449 struct task_group
*child
;
8450 unsigned long total
, sum
= 0;
8451 u64 period
, runtime
;
8453 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8454 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8457 period
= d
->rt_period
;
8458 runtime
= d
->rt_runtime
;
8462 * Cannot have more runtime than the period.
8464 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8468 * Ensure we don't starve existing RT tasks.
8470 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8473 total
= to_ratio(period
, runtime
);
8476 * Nobody can have more than the global setting allows.
8478 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8482 * The sum of our children's runtime should not exceed our own.
8484 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8485 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8486 runtime
= child
->rt_bandwidth
.rt_runtime
;
8488 if (child
== d
->tg
) {
8489 period
= d
->rt_period
;
8490 runtime
= d
->rt_runtime
;
8493 sum
+= to_ratio(period
, runtime
);
8502 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8504 struct rt_schedulable_data data
= {
8506 .rt_period
= period
,
8507 .rt_runtime
= runtime
,
8510 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8513 static int tg_set_bandwidth(struct task_group
*tg
,
8514 u64 rt_period
, u64 rt_runtime
)
8518 mutex_lock(&rt_constraints_mutex
);
8519 read_lock(&tasklist_lock
);
8520 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8524 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8525 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8526 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8528 for_each_possible_cpu(i
) {
8529 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8531 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8532 rt_rq
->rt_runtime
= rt_runtime
;
8533 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8535 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8537 read_unlock(&tasklist_lock
);
8538 mutex_unlock(&rt_constraints_mutex
);
8543 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8545 u64 rt_runtime
, rt_period
;
8547 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8548 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8549 if (rt_runtime_us
< 0)
8550 rt_runtime
= RUNTIME_INF
;
8552 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8555 long sched_group_rt_runtime(struct task_group
*tg
)
8559 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8562 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8563 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8564 return rt_runtime_us
;
8567 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8569 u64 rt_runtime
, rt_period
;
8571 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8572 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8577 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8580 long sched_group_rt_period(struct task_group
*tg
)
8584 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8585 do_div(rt_period_us
, NSEC_PER_USEC
);
8586 return rt_period_us
;
8589 static int sched_rt_global_constraints(void)
8591 u64 runtime
, period
;
8594 if (sysctl_sched_rt_period
<= 0)
8597 runtime
= global_rt_runtime();
8598 period
= global_rt_period();
8601 * Sanity check on the sysctl variables.
8603 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8606 mutex_lock(&rt_constraints_mutex
);
8607 read_lock(&tasklist_lock
);
8608 ret
= __rt_schedulable(NULL
, 0, 0);
8609 read_unlock(&tasklist_lock
);
8610 mutex_unlock(&rt_constraints_mutex
);
8615 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8617 /* Don't accept realtime tasks when there is no way for them to run */
8618 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8624 #else /* !CONFIG_RT_GROUP_SCHED */
8625 static int sched_rt_global_constraints(void)
8627 unsigned long flags
;
8630 if (sysctl_sched_rt_period
<= 0)
8634 * There's always some RT tasks in the root group
8635 * -- migration, kstopmachine etc..
8637 if (sysctl_sched_rt_runtime
== 0)
8640 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8641 for_each_possible_cpu(i
) {
8642 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8644 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8645 rt_rq
->rt_runtime
= global_rt_runtime();
8646 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8648 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8652 #endif /* CONFIG_RT_GROUP_SCHED */
8654 int sched_rt_handler(struct ctl_table
*table
, int write
,
8655 void __user
*buffer
, size_t *lenp
,
8659 int old_period
, old_runtime
;
8660 static DEFINE_MUTEX(mutex
);
8663 old_period
= sysctl_sched_rt_period
;
8664 old_runtime
= sysctl_sched_rt_runtime
;
8666 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8668 if (!ret
&& write
) {
8669 ret
= sched_rt_global_constraints();
8671 sysctl_sched_rt_period
= old_period
;
8672 sysctl_sched_rt_runtime
= old_runtime
;
8674 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8675 def_rt_bandwidth
.rt_period
=
8676 ns_to_ktime(global_rt_period());
8679 mutex_unlock(&mutex
);
8684 #ifdef CONFIG_CGROUP_SCHED
8686 /* return corresponding task_group object of a cgroup */
8687 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8689 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8690 struct task_group
, css
);
8693 static struct cgroup_subsys_state
*
8694 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8696 struct task_group
*tg
, *parent
;
8698 if (!cgrp
->parent
) {
8699 /* This is early initialization for the top cgroup */
8700 return &init_task_group
.css
;
8703 parent
= cgroup_tg(cgrp
->parent
);
8704 tg
= sched_create_group(parent
);
8706 return ERR_PTR(-ENOMEM
);
8712 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8714 struct task_group
*tg
= cgroup_tg(cgrp
);
8716 sched_destroy_group(tg
);
8720 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8722 #ifdef CONFIG_RT_GROUP_SCHED
8723 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8726 /* We don't support RT-tasks being in separate groups */
8727 if (tsk
->sched_class
!= &fair_sched_class
)
8734 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8735 struct task_struct
*tsk
, bool threadgroup
)
8737 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8741 struct task_struct
*c
;
8743 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8744 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8756 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8757 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8760 sched_move_task(tsk
);
8762 struct task_struct
*c
;
8764 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8771 #ifdef CONFIG_FAIR_GROUP_SCHED
8772 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8775 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8778 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8780 struct task_group
*tg
= cgroup_tg(cgrp
);
8782 return (u64
) tg
->shares
;
8784 #endif /* CONFIG_FAIR_GROUP_SCHED */
8786 #ifdef CONFIG_RT_GROUP_SCHED
8787 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8790 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8793 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8795 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8798 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8801 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8804 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8806 return sched_group_rt_period(cgroup_tg(cgrp
));
8808 #endif /* CONFIG_RT_GROUP_SCHED */
8810 static struct cftype cpu_files
[] = {
8811 #ifdef CONFIG_FAIR_GROUP_SCHED
8814 .read_u64
= cpu_shares_read_u64
,
8815 .write_u64
= cpu_shares_write_u64
,
8818 #ifdef CONFIG_RT_GROUP_SCHED
8820 .name
= "rt_runtime_us",
8821 .read_s64
= cpu_rt_runtime_read
,
8822 .write_s64
= cpu_rt_runtime_write
,
8825 .name
= "rt_period_us",
8826 .read_u64
= cpu_rt_period_read_uint
,
8827 .write_u64
= cpu_rt_period_write_uint
,
8832 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8834 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8837 struct cgroup_subsys cpu_cgroup_subsys
= {
8839 .create
= cpu_cgroup_create
,
8840 .destroy
= cpu_cgroup_destroy
,
8841 .can_attach
= cpu_cgroup_can_attach
,
8842 .attach
= cpu_cgroup_attach
,
8843 .populate
= cpu_cgroup_populate
,
8844 .subsys_id
= cpu_cgroup_subsys_id
,
8848 #endif /* CONFIG_CGROUP_SCHED */
8850 #ifdef CONFIG_CGROUP_CPUACCT
8853 * CPU accounting code for task groups.
8855 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8856 * (balbir@in.ibm.com).
8859 /* track cpu usage of a group of tasks and its child groups */
8861 struct cgroup_subsys_state css
;
8862 /* cpuusage holds pointer to a u64-type object on every cpu */
8863 u64 __percpu
*cpuusage
;
8864 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8865 struct cpuacct
*parent
;
8868 struct cgroup_subsys cpuacct_subsys
;
8870 /* return cpu accounting group corresponding to this container */
8871 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8873 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8874 struct cpuacct
, css
);
8877 /* return cpu accounting group to which this task belongs */
8878 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8880 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8881 struct cpuacct
, css
);
8884 /* create a new cpu accounting group */
8885 static struct cgroup_subsys_state
*cpuacct_create(
8886 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8888 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8894 ca
->cpuusage
= alloc_percpu(u64
);
8898 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8899 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8900 goto out_free_counters
;
8903 ca
->parent
= cgroup_ca(cgrp
->parent
);
8909 percpu_counter_destroy(&ca
->cpustat
[i
]);
8910 free_percpu(ca
->cpuusage
);
8914 return ERR_PTR(-ENOMEM
);
8917 /* destroy an existing cpu accounting group */
8919 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8921 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8924 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8925 percpu_counter_destroy(&ca
->cpustat
[i
]);
8926 free_percpu(ca
->cpuusage
);
8930 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8932 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8935 #ifndef CONFIG_64BIT
8937 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8939 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8941 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8949 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8951 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8953 #ifndef CONFIG_64BIT
8955 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8957 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8959 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8965 /* return total cpu usage (in nanoseconds) of a group */
8966 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8968 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8969 u64 totalcpuusage
= 0;
8972 for_each_present_cpu(i
)
8973 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8975 return totalcpuusage
;
8978 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8981 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8990 for_each_present_cpu(i
)
8991 cpuacct_cpuusage_write(ca
, i
, 0);
8997 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9000 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9004 for_each_present_cpu(i
) {
9005 percpu
= cpuacct_cpuusage_read(ca
, i
);
9006 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9008 seq_printf(m
, "\n");
9012 static const char *cpuacct_stat_desc
[] = {
9013 [CPUACCT_STAT_USER
] = "user",
9014 [CPUACCT_STAT_SYSTEM
] = "system",
9017 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9018 struct cgroup_map_cb
*cb
)
9020 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9023 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9024 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9025 val
= cputime64_to_clock_t(val
);
9026 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9031 static struct cftype files
[] = {
9034 .read_u64
= cpuusage_read
,
9035 .write_u64
= cpuusage_write
,
9038 .name
= "usage_percpu",
9039 .read_seq_string
= cpuacct_percpu_seq_read
,
9043 .read_map
= cpuacct_stats_show
,
9047 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9049 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9053 * charge this task's execution time to its accounting group.
9055 * called with rq->lock held.
9057 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9062 if (unlikely(!cpuacct_subsys
.active
))
9065 cpu
= task_cpu(tsk
);
9071 for (; ca
; ca
= ca
->parent
) {
9072 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9073 *cpuusage
+= cputime
;
9080 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9081 * in cputime_t units. As a result, cpuacct_update_stats calls
9082 * percpu_counter_add with values large enough to always overflow the
9083 * per cpu batch limit causing bad SMP scalability.
9085 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9086 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9087 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9090 #define CPUACCT_BATCH \
9091 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9093 #define CPUACCT_BATCH 0
9097 * Charge the system/user time to the task's accounting group.
9099 static void cpuacct_update_stats(struct task_struct
*tsk
,
9100 enum cpuacct_stat_index idx
, cputime_t val
)
9103 int batch
= CPUACCT_BATCH
;
9105 if (unlikely(!cpuacct_subsys
.active
))
9112 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9118 struct cgroup_subsys cpuacct_subsys
= {
9120 .create
= cpuacct_create
,
9121 .destroy
= cpuacct_destroy
,
9122 .populate
= cpuacct_populate
,
9123 .subsys_id
= cpuacct_subsys_id
,
9125 #endif /* CONFIG_CGROUP_CPUACCT */
9129 void synchronize_sched_expedited(void)
9133 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9135 #else /* #ifndef CONFIG_SMP */
9137 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9139 static int synchronize_sched_expedited_cpu_stop(void *data
)
9142 * There must be a full memory barrier on each affected CPU
9143 * between the time that try_stop_cpus() is called and the
9144 * time that it returns.
9146 * In the current initial implementation of cpu_stop, the
9147 * above condition is already met when the control reaches
9148 * this point and the following smp_mb() is not strictly
9149 * necessary. Do smp_mb() anyway for documentation and
9150 * robustness against future implementation changes.
9152 smp_mb(); /* See above comment block. */
9157 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9158 * approach to force grace period to end quickly. This consumes
9159 * significant time on all CPUs, and is thus not recommended for
9160 * any sort of common-case code.
9162 * Note that it is illegal to call this function while holding any
9163 * lock that is acquired by a CPU-hotplug notifier. Failing to
9164 * observe this restriction will result in deadlock.
9166 void synchronize_sched_expedited(void)
9168 int snap
, trycount
= 0;
9170 smp_mb(); /* ensure prior mod happens before capturing snap. */
9171 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9173 while (try_stop_cpus(cpu_online_mask
,
9174 synchronize_sched_expedited_cpu_stop
,
9177 if (trycount
++ < 10)
9178 udelay(trycount
* num_online_cpus());
9180 synchronize_sched();
9183 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9184 smp_mb(); /* ensure test happens before caller kfree */
9189 atomic_inc(&synchronize_sched_expedited_count
);
9190 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9193 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9195 #endif /* #else #ifndef CONFIG_SMP */