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"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy
)
125 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
130 static inline int task_has_rt_policy(struct task_struct
*p
)
132 return rt_policy(p
->policy
);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array
{
139 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
140 struct list_head queue
[MAX_RT_PRIO
];
143 struct rt_bandwidth
{
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock
;
148 struct hrtimer rt_period_timer
;
151 static struct rt_bandwidth def_rt_bandwidth
;
153 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
155 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
157 struct rt_bandwidth
*rt_b
=
158 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
164 now
= hrtimer_cb_get_time(timer
);
165 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
170 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
173 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
177 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
179 rt_b
->rt_period
= ns_to_ktime(period
);
180 rt_b
->rt_runtime
= runtime
;
182 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
184 hrtimer_init(&rt_b
->rt_period_timer
,
185 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
186 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime
>= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
198 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
201 if (hrtimer_active(&rt_b
->rt_period_timer
))
204 raw_spin_lock(&rt_b
->rt_runtime_lock
);
209 if (hrtimer_active(&rt_b
->rt_period_timer
))
212 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
213 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
215 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
216 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
217 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
218 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
219 HRTIMER_MODE_ABS_PINNED
, 0);
221 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
227 hrtimer_cancel(&rt_b
->rt_period_timer
);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex
);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups
);
245 /* task group related information */
247 struct cgroup_subsys_state css
;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity
**se
;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq
**cfs_rq
;
254 unsigned long shares
;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity
**rt_se
;
259 struct rt_rq
**rt_rq
;
261 struct rt_bandwidth rt_bandwidth
;
265 struct list_head list
;
267 struct task_group
*parent
;
268 struct list_head siblings
;
269 struct list_head children
;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock
);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group
.children
);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group
;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load
;
314 unsigned long nr_running
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
322 struct list_head tasks
;
323 struct list_head
*balance_iterator
;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity
*curr
, *next
, *last
;
331 unsigned int nr_spread_over
;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list
;
345 struct task_group
*tg
; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight
;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load
;
362 * this cpu's part of tg->shares
364 unsigned long shares
;
367 * load.weight at the time we set shares
369 unsigned long rq_weight
;
374 /* Real-Time classes' related field in a runqueue: */
376 struct rt_prio_array active
;
377 unsigned long rt_nr_running
;
378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
380 int curr
; /* highest queued rt task prio */
382 int next
; /* next highest */
387 unsigned long rt_nr_migratory
;
388 unsigned long rt_nr_total
;
390 struct plist_head pushable_tasks
;
395 /* Nests inside the rq lock: */
396 raw_spinlock_t rt_runtime_lock
;
398 #ifdef CONFIG_RT_GROUP_SCHED
399 unsigned long rt_nr_boosted
;
402 struct list_head leaf_rt_rq_list
;
403 struct task_group
*tg
;
410 * We add the notion of a root-domain which will be used to define per-domain
411 * variables. Each exclusive cpuset essentially defines an island domain by
412 * fully partitioning the member cpus from any other cpuset. Whenever a new
413 * exclusive cpuset is created, we also create and attach a new root-domain
420 cpumask_var_t online
;
423 * The "RT overload" flag: it gets set if a CPU has more than
424 * one runnable RT task.
426 cpumask_var_t rto_mask
;
429 struct cpupri cpupri
;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain
;
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running
;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
461 unsigned char in_nohz_recently
;
463 unsigned int skip_clock_update
;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load
;
467 unsigned long nr_load_updates
;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list
;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list
;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible
;
489 struct task_struct
*curr
, *idle
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
499 struct root_domain
*rd
;
500 struct sched_domain
*sd
;
502 unsigned long cpu_power
;
504 unsigned char idle_at_tick
;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work
;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task
;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update
;
528 long calc_load_active
;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending
;
533 struct call_single_data hrtick_csd
;
535 struct hrtimer hrtick_timer
;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info
;
541 unsigned long long rq_cpu_time
;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count
;
547 /* schedule() stats */
548 unsigned int sched_switch
;
549 unsigned int sched_count
;
550 unsigned int sched_goidle
;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count
;
554 unsigned int ttwu_local
;
557 unsigned int bkl_count
;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
564 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
566 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
569 * A queue event has occurred, and we're going to schedule. In
570 * this case, we can save a useless back to back clock update.
572 if (rq
->curr
->se
.on_rq
&& test_tsk_need_resched(rq
->curr
))
573 rq
->skip_clock_update
= 1;
576 static inline int cpu_of(struct rq
*rq
)
585 #define rcu_dereference_check_sched_domain(p) \
586 rcu_dereference_check((p), \
587 rcu_read_lock_sched_held() || \
588 lockdep_is_held(&sched_domains_mutex))
591 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
592 * See detach_destroy_domains: synchronize_sched for details.
594 * The domain tree of any CPU may only be accessed from within
595 * preempt-disabled sections.
597 #define for_each_domain(cpu, __sd) \
598 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
600 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
601 #define this_rq() (&__get_cpu_var(runqueues))
602 #define task_rq(p) cpu_rq(task_cpu(p))
603 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
604 #define raw_rq() (&__raw_get_cpu_var(runqueues))
606 #ifdef CONFIG_CGROUP_SCHED
609 * Return the group to which this tasks belongs.
611 * We use task_subsys_state_check() and extend the RCU verification
612 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
613 * holds that lock for each task it moves into the cgroup. Therefore
614 * by holding that lock, we pin the task to the current cgroup.
616 static inline struct task_group
*task_group(struct task_struct
*p
)
618 struct cgroup_subsys_state
*css
;
620 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
621 lockdep_is_held(&task_rq(p
)->lock
));
622 return container_of(css
, struct task_group
, css
);
625 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
626 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
628 #ifdef CONFIG_FAIR_GROUP_SCHED
629 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
630 p
->se
.parent
= task_group(p
)->se
[cpu
];
633 #ifdef CONFIG_RT_GROUP_SCHED
634 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
635 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
639 #else /* CONFIG_CGROUP_SCHED */
641 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
642 static inline struct task_group
*task_group(struct task_struct
*p
)
647 #endif /* CONFIG_CGROUP_SCHED */
649 static u64
irq_time_cpu(int cpu
);
650 static void sched_irq_time_avg_update(struct rq
*rq
, u64 irq_time
);
652 inline void update_rq_clock(struct rq
*rq
)
654 int cpu
= cpu_of(rq
);
657 if (!rq
->skip_clock_update
)
658 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
659 irq_time
= irq_time_cpu(cpu
);
660 if (rq
->clock
- irq_time
> rq
->clock_task
)
661 rq
->clock_task
= rq
->clock
- irq_time
;
663 sched_irq_time_avg_update(rq
, irq_time
);
667 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
669 #ifdef CONFIG_SCHED_DEBUG
670 # define const_debug __read_mostly
672 # define const_debug static const
677 * @cpu: the processor in question.
679 * Returns true if the current cpu runqueue is locked.
680 * This interface allows printk to be called with the runqueue lock
681 * held and know whether or not it is OK to wake up the klogd.
683 int runqueue_is_locked(int cpu
)
685 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
689 * Debugging: various feature bits
692 #define SCHED_FEAT(name, enabled) \
693 __SCHED_FEAT_##name ,
696 #include "sched_features.h"
701 #define SCHED_FEAT(name, enabled) \
702 (1UL << __SCHED_FEAT_##name) * enabled |
704 const_debug
unsigned int sysctl_sched_features
=
705 #include "sched_features.h"
710 #ifdef CONFIG_SCHED_DEBUG
711 #define SCHED_FEAT(name, enabled) \
714 static __read_mostly
char *sched_feat_names
[] = {
715 #include "sched_features.h"
721 static int sched_feat_show(struct seq_file
*m
, void *v
)
725 for (i
= 0; sched_feat_names
[i
]; i
++) {
726 if (!(sysctl_sched_features
& (1UL << i
)))
728 seq_printf(m
, "%s ", sched_feat_names
[i
]);
736 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
737 size_t cnt
, loff_t
*ppos
)
747 if (copy_from_user(&buf
, ubuf
, cnt
))
753 if (strncmp(buf
, "NO_", 3) == 0) {
758 for (i
= 0; sched_feat_names
[i
]; i
++) {
759 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
761 sysctl_sched_features
&= ~(1UL << i
);
763 sysctl_sched_features
|= (1UL << i
);
768 if (!sched_feat_names
[i
])
776 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
778 return single_open(filp
, sched_feat_show
, NULL
);
781 static const struct file_operations sched_feat_fops
= {
782 .open
= sched_feat_open
,
783 .write
= sched_feat_write
,
786 .release
= single_release
,
789 static __init
int sched_init_debug(void)
791 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
796 late_initcall(sched_init_debug
);
800 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
803 * Number of tasks to iterate in a single balance run.
804 * Limited because this is done with IRQs disabled.
806 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
809 * ratelimit for updating the group shares.
812 unsigned int sysctl_sched_shares_ratelimit
= 250000;
813 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
816 * Inject some fuzzyness into changing the per-cpu group shares
817 * this avoids remote rq-locks at the expense of fairness.
820 unsigned int sysctl_sched_shares_thresh
= 4;
823 * period over which we average the RT time consumption, measured
828 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
831 * period over which we measure -rt task cpu usage in us.
834 unsigned int sysctl_sched_rt_period
= 1000000;
836 static __read_mostly
int scheduler_running
;
839 * part of the period that we allow rt tasks to run in us.
842 int sysctl_sched_rt_runtime
= 950000;
844 static inline u64
global_rt_period(void)
846 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
849 static inline u64
global_rt_runtime(void)
851 if (sysctl_sched_rt_runtime
< 0)
854 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
857 #ifndef prepare_arch_switch
858 # define prepare_arch_switch(next) do { } while (0)
860 #ifndef finish_arch_switch
861 # define finish_arch_switch(prev) do { } while (0)
864 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
866 return rq
->curr
== p
;
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
872 return task_current(rq
, p
);
875 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
879 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
881 #ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq
->lock
.owner
= current
;
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
890 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
892 raw_spin_unlock_irq(&rq
->lock
);
895 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
896 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
901 return task_current(rq
, p
);
905 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
909 * We can optimise this out completely for !SMP, because the
910 * SMP rebalancing from interrupt is the only thing that cares
915 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
916 raw_spin_unlock_irq(&rq
->lock
);
918 raw_spin_unlock(&rq
->lock
);
922 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
926 * After ->oncpu is cleared, the task can be moved to a different CPU.
927 * We must ensure this doesn't happen until the switch is completely
933 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
940 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
943 static inline int task_is_waking(struct task_struct
*p
)
945 return unlikely(p
->state
== TASK_WAKING
);
949 * __task_rq_lock - lock the runqueue a given task resides on.
950 * Must be called interrupts disabled.
952 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
959 raw_spin_lock(&rq
->lock
);
960 if (likely(rq
== task_rq(p
)))
962 raw_spin_unlock(&rq
->lock
);
967 * task_rq_lock - lock the runqueue a given task resides on and disable
968 * interrupts. Note the ordering: we can safely lookup the task_rq without
969 * explicitly disabling preemption.
971 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
977 local_irq_save(*flags
);
979 raw_spin_lock(&rq
->lock
);
980 if (likely(rq
== task_rq(p
)))
982 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
986 static void __task_rq_unlock(struct rq
*rq
)
989 raw_spin_unlock(&rq
->lock
);
992 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
995 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq
*this_rq_lock(void)
1002 __acquires(rq
->lock
)
1006 local_irq_disable();
1008 raw_spin_lock(&rq
->lock
);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq
*rq
)
1032 if (!sched_feat(HRTICK
))
1034 if (!cpu_active(cpu_of(rq
)))
1036 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1039 static void hrtick_clear(struct rq
*rq
)
1041 if (hrtimer_active(&rq
->hrtick_timer
))
1042 hrtimer_cancel(&rq
->hrtick_timer
);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1051 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1053 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1055 raw_spin_lock(&rq
->lock
);
1056 update_rq_clock(rq
);
1057 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1058 raw_spin_unlock(&rq
->lock
);
1060 return HRTIMER_NORESTART
;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg
)
1069 struct rq
*rq
= arg
;
1071 raw_spin_lock(&rq
->lock
);
1072 hrtimer_restart(&rq
->hrtick_timer
);
1073 rq
->hrtick_csd_pending
= 0;
1074 raw_spin_unlock(&rq
->lock
);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq
*rq
, u64 delay
)
1084 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1085 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1087 hrtimer_set_expires(timer
, time
);
1089 if (rq
== this_rq()) {
1090 hrtimer_restart(timer
);
1091 } else if (!rq
->hrtick_csd_pending
) {
1092 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1093 rq
->hrtick_csd_pending
= 1;
1098 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1100 int cpu
= (int)(long)hcpu
;
1103 case CPU_UP_CANCELED
:
1104 case CPU_UP_CANCELED_FROZEN
:
1105 case CPU_DOWN_PREPARE
:
1106 case CPU_DOWN_PREPARE_FROZEN
:
1108 case CPU_DEAD_FROZEN
:
1109 hrtick_clear(cpu_rq(cpu
));
1116 static __init
void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick
, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq
*rq
, u64 delay
)
1128 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1129 HRTIMER_MODE_REL_PINNED
, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq
*rq
)
1140 rq
->hrtick_csd_pending
= 0;
1142 rq
->hrtick_csd
.flags
= 0;
1143 rq
->hrtick_csd
.func
= __hrtick_start
;
1144 rq
->hrtick_csd
.info
= rq
;
1147 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1148 rq
->hrtick_timer
.function
= hrtick
;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq
*rq
)
1155 static inline void init_rq_hrtick(struct rq
*rq
)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct
*p
)
1181 assert_raw_spin_locked(&task_rq(p
)->lock
);
1183 if (test_tsk_need_resched(p
))
1186 set_tsk_need_resched(p
);
1189 if (cpu
== smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p
))
1195 smp_send_reschedule(cpu
);
1198 static void resched_cpu(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1201 unsigned long flags
;
1203 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1205 resched_task(cpu_curr(cpu
));
1206 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu
)
1222 struct rq
*rq
= cpu_rq(cpu
);
1224 if (cpu
== smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq
->curr
!= rq
->idle
)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq
->idle
);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq
->idle
))
1247 smp_send_reschedule(cpu
);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64
sched_avg_period(void)
1254 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1257 static void sched_avg_update(struct rq
*rq
)
1259 s64 period
= sched_avg_period();
1261 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1263 * Inline assembly required to prevent the compiler
1264 * optimising this loop into a divmod call.
1265 * See __iter_div_u64_rem() for another example of this.
1267 asm("" : "+rm" (rq
->age_stamp
));
1268 rq
->age_stamp
+= period
;
1273 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1275 rq
->rt_avg
+= rt_delta
;
1276 sched_avg_update(rq
);
1279 #else /* !CONFIG_SMP */
1280 static void resched_task(struct task_struct
*p
)
1282 assert_raw_spin_locked(&task_rq(p
)->lock
);
1283 set_tsk_need_resched(p
);
1286 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1290 static void sched_avg_update(struct rq
*rq
)
1293 #endif /* CONFIG_SMP */
1295 #if BITS_PER_LONG == 32
1296 # define WMULT_CONST (~0UL)
1298 # define WMULT_CONST (1UL << 32)
1301 #define WMULT_SHIFT 32
1304 * Shift right and round:
1306 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1309 * delta *= weight / lw
1311 static unsigned long
1312 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1313 struct load_weight
*lw
)
1317 if (!lw
->inv_weight
) {
1318 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1321 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1325 tmp
= (u64
)delta_exec
* weight
;
1327 * Check whether we'd overflow the 64-bit multiplication:
1329 if (unlikely(tmp
> WMULT_CONST
))
1330 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1333 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1335 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1338 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1344 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1351 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1352 * of tasks with abnormal "nice" values across CPUs the contribution that
1353 * each task makes to its run queue's load is weighted according to its
1354 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1355 * scaled version of the new time slice allocation that they receive on time
1359 #define WEIGHT_IDLEPRIO 3
1360 #define WMULT_IDLEPRIO 1431655765
1363 * Nice levels are multiplicative, with a gentle 10% change for every
1364 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1365 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1366 * that remained on nice 0.
1368 * The "10% effect" is relative and cumulative: from _any_ nice level,
1369 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1370 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1371 * If a task goes up by ~10% and another task goes down by ~10% then
1372 * the relative distance between them is ~25%.)
1374 static const int prio_to_weight
[40] = {
1375 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1376 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1377 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1378 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1379 /* 0 */ 1024, 820, 655, 526, 423,
1380 /* 5 */ 335, 272, 215, 172, 137,
1381 /* 10 */ 110, 87, 70, 56, 45,
1382 /* 15 */ 36, 29, 23, 18, 15,
1386 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1388 * In cases where the weight does not change often, we can use the
1389 * precalculated inverse to speed up arithmetics by turning divisions
1390 * into multiplications:
1392 static const u32 prio_to_wmult
[40] = {
1393 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1394 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1395 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1396 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1397 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1398 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1399 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1400 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1403 /* Time spent by the tasks of the cpu accounting group executing in ... */
1404 enum cpuacct_stat_index
{
1405 CPUACCT_STAT_USER
, /* ... user mode */
1406 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1408 CPUACCT_STAT_NSTATS
,
1411 #ifdef CONFIG_CGROUP_CPUACCT
1412 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1413 static void cpuacct_update_stats(struct task_struct
*tsk
,
1414 enum cpuacct_stat_index idx
, cputime_t val
);
1416 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1417 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1418 enum cpuacct_stat_index idx
, cputime_t val
) {}
1421 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1423 update_load_add(&rq
->load
, load
);
1426 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1428 update_load_sub(&rq
->load
, load
);
1431 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1432 typedef int (*tg_visitor
)(struct task_group
*, void *);
1435 * Iterate the full tree, calling @down when first entering a node and @up when
1436 * leaving it for the final time.
1438 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1440 struct task_group
*parent
, *child
;
1444 parent
= &root_task_group
;
1446 ret
= (*down
)(parent
, data
);
1449 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1456 ret
= (*up
)(parent
, data
);
1461 parent
= parent
->parent
;
1470 static int tg_nop(struct task_group
*tg
, void *data
)
1477 /* Used instead of source_load when we know the type == 0 */
1478 static unsigned long weighted_cpuload(const int cpu
)
1480 return cpu_rq(cpu
)->load
.weight
;
1484 * Return a low guess at the load of a migration-source cpu weighted
1485 * according to the scheduling class and "nice" value.
1487 * We want to under-estimate the load of migration sources, to
1488 * balance conservatively.
1490 static unsigned long source_load(int cpu
, int type
)
1492 struct rq
*rq
= cpu_rq(cpu
);
1493 unsigned long total
= weighted_cpuload(cpu
);
1495 if (type
== 0 || !sched_feat(LB_BIAS
))
1498 return min(rq
->cpu_load
[type
-1], total
);
1502 * Return a high guess at the load of a migration-target cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 static unsigned long target_load(int cpu
, int type
)
1507 struct rq
*rq
= cpu_rq(cpu
);
1508 unsigned long total
= weighted_cpuload(cpu
);
1510 if (type
== 0 || !sched_feat(LB_BIAS
))
1513 return max(rq
->cpu_load
[type
-1], total
);
1516 static unsigned long power_of(int cpu
)
1518 return cpu_rq(cpu
)->cpu_power
;
1521 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1523 static unsigned long cpu_avg_load_per_task(int cpu
)
1525 struct rq
*rq
= cpu_rq(cpu
);
1526 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1529 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1531 rq
->avg_load_per_task
= 0;
1533 return rq
->avg_load_per_task
;
1536 #ifdef CONFIG_FAIR_GROUP_SCHED
1538 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1540 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1543 * Calculate and set the cpu's group shares.
1545 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1546 unsigned long sd_shares
,
1547 unsigned long sd_rq_weight
,
1548 unsigned long *usd_rq_weight
)
1550 unsigned long shares
, rq_weight
;
1553 rq_weight
= usd_rq_weight
[cpu
];
1556 rq_weight
= NICE_0_LOAD
;
1560 * \Sum_j shares_j * rq_weight_i
1561 * shares_i = -----------------------------
1562 * \Sum_j rq_weight_j
1564 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1565 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1567 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1568 sysctl_sched_shares_thresh
) {
1569 struct rq
*rq
= cpu_rq(cpu
);
1570 unsigned long flags
;
1572 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1573 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1574 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1575 __set_se_shares(tg
->se
[cpu
], shares
);
1576 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1581 * Re-compute the task group their per cpu shares over the given domain.
1582 * This needs to be done in a bottom-up fashion because the rq weight of a
1583 * parent group depends on the shares of its child groups.
1585 static int tg_shares_up(struct task_group
*tg
, void *data
)
1587 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1588 unsigned long *usd_rq_weight
;
1589 struct sched_domain
*sd
= data
;
1590 unsigned long flags
;
1596 local_irq_save(flags
);
1597 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1599 for_each_cpu(i
, sched_domain_span(sd
)) {
1600 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1601 usd_rq_weight
[i
] = weight
;
1603 rq_weight
+= weight
;
1605 * If there are currently no tasks on the cpu pretend there
1606 * is one of average load so that when a new task gets to
1607 * run here it will not get delayed by group starvation.
1610 weight
= NICE_0_LOAD
;
1612 sum_weight
+= weight
;
1613 shares
+= tg
->cfs_rq
[i
]->shares
;
1617 rq_weight
= sum_weight
;
1619 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1620 shares
= tg
->shares
;
1622 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1623 shares
= tg
->shares
;
1625 for_each_cpu(i
, sched_domain_span(sd
))
1626 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1628 local_irq_restore(flags
);
1634 * Compute the cpu's hierarchical load factor for each task group.
1635 * This needs to be done in a top-down fashion because the load of a child
1636 * group is a fraction of its parents load.
1638 static int tg_load_down(struct task_group
*tg
, void *data
)
1641 long cpu
= (long)data
;
1644 load
= cpu_rq(cpu
)->load
.weight
;
1646 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1647 load
*= tg
->cfs_rq
[cpu
]->shares
;
1648 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1651 tg
->cfs_rq
[cpu
]->h_load
= load
;
1656 static void update_shares(struct sched_domain
*sd
)
1661 if (root_task_group_empty())
1664 now
= cpu_clock(raw_smp_processor_id());
1665 elapsed
= now
- sd
->last_update
;
1667 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1668 sd
->last_update
= now
;
1669 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1673 static void update_h_load(long cpu
)
1675 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1680 static inline void update_shares(struct sched_domain
*sd
)
1686 #ifdef CONFIG_PREEMPT
1688 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1691 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1692 * way at the expense of forcing extra atomic operations in all
1693 * invocations. This assures that the double_lock is acquired using the
1694 * same underlying policy as the spinlock_t on this architecture, which
1695 * reduces latency compared to the unfair variant below. However, it
1696 * also adds more overhead and therefore may reduce throughput.
1698 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1699 __releases(this_rq
->lock
)
1700 __acquires(busiest
->lock
)
1701 __acquires(this_rq
->lock
)
1703 raw_spin_unlock(&this_rq
->lock
);
1704 double_rq_lock(this_rq
, busiest
);
1711 * Unfair double_lock_balance: Optimizes throughput at the expense of
1712 * latency by eliminating extra atomic operations when the locks are
1713 * already in proper order on entry. This favors lower cpu-ids and will
1714 * grant the double lock to lower cpus over higher ids under contention,
1715 * regardless of entry order into the function.
1717 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1718 __releases(this_rq
->lock
)
1719 __acquires(busiest
->lock
)
1720 __acquires(this_rq
->lock
)
1724 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1725 if (busiest
< this_rq
) {
1726 raw_spin_unlock(&this_rq
->lock
);
1727 raw_spin_lock(&busiest
->lock
);
1728 raw_spin_lock_nested(&this_rq
->lock
,
1729 SINGLE_DEPTH_NESTING
);
1732 raw_spin_lock_nested(&busiest
->lock
,
1733 SINGLE_DEPTH_NESTING
);
1738 #endif /* CONFIG_PREEMPT */
1741 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1743 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1745 if (unlikely(!irqs_disabled())) {
1746 /* printk() doesn't work good under rq->lock */
1747 raw_spin_unlock(&this_rq
->lock
);
1751 return _double_lock_balance(this_rq
, busiest
);
1754 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1755 __releases(busiest
->lock
)
1757 raw_spin_unlock(&busiest
->lock
);
1758 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1762 * double_rq_lock - safely lock two runqueues
1764 * Note this does not disable interrupts like task_rq_lock,
1765 * you need to do so manually before calling.
1767 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1768 __acquires(rq1
->lock
)
1769 __acquires(rq2
->lock
)
1771 BUG_ON(!irqs_disabled());
1773 raw_spin_lock(&rq1
->lock
);
1774 __acquire(rq2
->lock
); /* Fake it out ;) */
1777 raw_spin_lock(&rq1
->lock
);
1778 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1780 raw_spin_lock(&rq2
->lock
);
1781 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1787 * double_rq_unlock - safely unlock two runqueues
1789 * Note this does not restore interrupts like task_rq_unlock,
1790 * you need to do so manually after calling.
1792 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1793 __releases(rq1
->lock
)
1794 __releases(rq2
->lock
)
1796 raw_spin_unlock(&rq1
->lock
);
1798 raw_spin_unlock(&rq2
->lock
);
1800 __release(rq2
->lock
);
1805 #ifdef CONFIG_FAIR_GROUP_SCHED
1806 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1809 cfs_rq
->shares
= shares
;
1814 static void calc_load_account_idle(struct rq
*this_rq
);
1815 static void update_sysctl(void);
1816 static int get_update_sysctl_factor(void);
1818 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1820 set_task_rq(p
, cpu
);
1823 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1824 * successfuly executed on another CPU. We must ensure that updates of
1825 * per-task data have been completed by this moment.
1828 task_thread_info(p
)->cpu
= cpu
;
1832 static const struct sched_class rt_sched_class
;
1834 #define sched_class_highest (&rt_sched_class)
1835 #define for_each_class(class) \
1836 for (class = sched_class_highest; class; class = class->next)
1838 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1841 * There are no locks covering percpu hardirq/softirq time.
1842 * They are only modified in account_system_vtime, on corresponding CPU
1843 * with interrupts disabled. So, writes are safe.
1844 * They are read and saved off onto struct rq in update_rq_clock().
1845 * This may result in other CPU reading this CPU's irq time and can
1846 * race with irq/account_system_vtime on this CPU. We would either get old
1847 * or new value (or semi updated value on 32 bit) with a side effect of
1848 * accounting a slice of irq time to wrong task when irq is in progress
1849 * while we read rq->clock. That is a worthy compromise in place of having
1850 * locks on each irq in account_system_time.
1852 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1853 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1855 static DEFINE_PER_CPU(u64
, irq_start_time
);
1856 static int sched_clock_irqtime
;
1858 void enable_sched_clock_irqtime(void)
1860 sched_clock_irqtime
= 1;
1863 void disable_sched_clock_irqtime(void)
1865 sched_clock_irqtime
= 0;
1868 static u64
irq_time_cpu(int cpu
)
1870 if (!sched_clock_irqtime
)
1873 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1876 void account_system_vtime(struct task_struct
*curr
)
1878 unsigned long flags
;
1882 if (!sched_clock_irqtime
)
1885 local_irq_save(flags
);
1887 cpu
= smp_processor_id();
1888 now
= sched_clock_cpu(cpu
);
1889 delta
= now
- per_cpu(irq_start_time
, cpu
);
1890 per_cpu(irq_start_time
, cpu
) = now
;
1892 * We do not account for softirq time from ksoftirqd here.
1893 * We want to continue accounting softirq time to ksoftirqd thread
1894 * in that case, so as not to confuse scheduler with a special task
1895 * that do not consume any time, but still wants to run.
1897 if (hardirq_count())
1898 per_cpu(cpu_hardirq_time
, cpu
) += delta
;
1899 else if (in_serving_softirq() && !(curr
->flags
& PF_KSOFTIRQD
))
1900 per_cpu(cpu_softirq_time
, cpu
) += delta
;
1902 local_irq_restore(flags
);
1904 EXPORT_SYMBOL_GPL(account_system_vtime
);
1906 static void sched_irq_time_avg_update(struct rq
*rq
, u64 curr_irq_time
)
1908 if (sched_clock_irqtime
&& sched_feat(NONIRQ_POWER
)) {
1909 u64 delta_irq
= curr_irq_time
- rq
->prev_irq_time
;
1910 rq
->prev_irq_time
= curr_irq_time
;
1911 sched_rt_avg_update(rq
, delta_irq
);
1917 static u64
irq_time_cpu(int cpu
)
1922 static void sched_irq_time_avg_update(struct rq
*rq
, u64 curr_irq_time
) { }
1926 #include "sched_stats.h"
1928 static void inc_nr_running(struct rq
*rq
)
1933 static void dec_nr_running(struct rq
*rq
)
1938 static void set_load_weight(struct task_struct
*p
)
1941 * SCHED_IDLE tasks get minimal weight:
1943 if (p
->policy
== SCHED_IDLE
) {
1944 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1945 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1949 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1950 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1953 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1955 update_rq_clock(rq
);
1956 sched_info_queued(p
);
1957 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1961 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1963 update_rq_clock(rq
);
1964 sched_info_dequeued(p
);
1965 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1970 * activate_task - move a task to the runqueue.
1972 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1974 if (task_contributes_to_load(p
))
1975 rq
->nr_uninterruptible
--;
1977 enqueue_task(rq
, p
, flags
);
1982 * deactivate_task - remove a task from the runqueue.
1984 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1986 if (task_contributes_to_load(p
))
1987 rq
->nr_uninterruptible
++;
1989 dequeue_task(rq
, p
, flags
);
1993 #include "sched_idletask.c"
1994 #include "sched_fair.c"
1995 #include "sched_rt.c"
1996 #ifdef CONFIG_SCHED_DEBUG
1997 # include "sched_debug.c"
2001 * __normal_prio - return the priority that is based on the static prio
2003 static inline int __normal_prio(struct task_struct
*p
)
2005 return p
->static_prio
;
2009 * Calculate the expected normal priority: i.e. priority
2010 * without taking RT-inheritance into account. Might be
2011 * boosted by interactivity modifiers. Changes upon fork,
2012 * setprio syscalls, and whenever the interactivity
2013 * estimator recalculates.
2015 static inline int normal_prio(struct task_struct
*p
)
2019 if (task_has_rt_policy(p
))
2020 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2022 prio
= __normal_prio(p
);
2027 * Calculate the current priority, i.e. the priority
2028 * taken into account by the scheduler. This value might
2029 * be boosted by RT tasks, or might be boosted by
2030 * interactivity modifiers. Will be RT if the task got
2031 * RT-boosted. If not then it returns p->normal_prio.
2033 static int effective_prio(struct task_struct
*p
)
2035 p
->normal_prio
= normal_prio(p
);
2037 * If we are RT tasks or we were boosted to RT priority,
2038 * keep the priority unchanged. Otherwise, update priority
2039 * to the normal priority:
2041 if (!rt_prio(p
->prio
))
2042 return p
->normal_prio
;
2047 * task_curr - is this task currently executing on a CPU?
2048 * @p: the task in question.
2050 inline int task_curr(const struct task_struct
*p
)
2052 return cpu_curr(task_cpu(p
)) == p
;
2055 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2056 const struct sched_class
*prev_class
,
2057 int oldprio
, int running
)
2059 if (prev_class
!= p
->sched_class
) {
2060 if (prev_class
->switched_from
)
2061 prev_class
->switched_from(rq
, p
, running
);
2062 p
->sched_class
->switched_to(rq
, p
, running
);
2064 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2069 * Is this task likely cache-hot:
2072 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2076 if (p
->sched_class
!= &fair_sched_class
)
2079 if (unlikely(p
->policy
== SCHED_IDLE
))
2083 * Buddy candidates are cache hot:
2085 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2086 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2087 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2090 if (sysctl_sched_migration_cost
== -1)
2092 if (sysctl_sched_migration_cost
== 0)
2095 delta
= now
- p
->se
.exec_start
;
2097 return delta
< (s64
)sysctl_sched_migration_cost
;
2100 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2102 #ifdef CONFIG_SCHED_DEBUG
2104 * We should never call set_task_cpu() on a blocked task,
2105 * ttwu() will sort out the placement.
2107 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2108 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2111 trace_sched_migrate_task(p
, new_cpu
);
2113 if (task_cpu(p
) != new_cpu
) {
2114 p
->se
.nr_migrations
++;
2115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2118 __set_task_cpu(p
, new_cpu
);
2121 struct migration_arg
{
2122 struct task_struct
*task
;
2126 static int migration_cpu_stop(void *data
);
2129 * The task's runqueue lock must be held.
2130 * Returns true if you have to wait for migration thread.
2132 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2134 struct rq
*rq
= task_rq(p
);
2137 * If the task is not on a runqueue (and not running), then
2138 * the next wake-up will properly place the task.
2140 return p
->se
.on_rq
|| task_running(rq
, p
);
2144 * wait_task_inactive - wait for a thread to unschedule.
2146 * If @match_state is nonzero, it's the @p->state value just checked and
2147 * not expected to change. If it changes, i.e. @p might have woken up,
2148 * then return zero. When we succeed in waiting for @p to be off its CPU,
2149 * we return a positive number (its total switch count). If a second call
2150 * a short while later returns the same number, the caller can be sure that
2151 * @p has remained unscheduled the whole time.
2153 * The caller must ensure that the task *will* unschedule sometime soon,
2154 * else this function might spin for a *long* time. This function can't
2155 * be called with interrupts off, or it may introduce deadlock with
2156 * smp_call_function() if an IPI is sent by the same process we are
2157 * waiting to become inactive.
2159 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2161 unsigned long flags
;
2168 * We do the initial early heuristics without holding
2169 * any task-queue locks at all. We'll only try to get
2170 * the runqueue lock when things look like they will
2176 * If the task is actively running on another CPU
2177 * still, just relax and busy-wait without holding
2180 * NOTE! Since we don't hold any locks, it's not
2181 * even sure that "rq" stays as the right runqueue!
2182 * But we don't care, since "task_running()" will
2183 * return false if the runqueue has changed and p
2184 * is actually now running somewhere else!
2186 while (task_running(rq
, p
)) {
2187 if (match_state
&& unlikely(p
->state
!= match_state
))
2193 * Ok, time to look more closely! We need the rq
2194 * lock now, to be *sure*. If we're wrong, we'll
2195 * just go back and repeat.
2197 rq
= task_rq_lock(p
, &flags
);
2198 trace_sched_wait_task(p
);
2199 running
= task_running(rq
, p
);
2200 on_rq
= p
->se
.on_rq
;
2202 if (!match_state
|| p
->state
== match_state
)
2203 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2204 task_rq_unlock(rq
, &flags
);
2207 * If it changed from the expected state, bail out now.
2209 if (unlikely(!ncsw
))
2213 * Was it really running after all now that we
2214 * checked with the proper locks actually held?
2216 * Oops. Go back and try again..
2218 if (unlikely(running
)) {
2224 * It's not enough that it's not actively running,
2225 * it must be off the runqueue _entirely_, and not
2228 * So if it was still runnable (but just not actively
2229 * running right now), it's preempted, and we should
2230 * yield - it could be a while.
2232 if (unlikely(on_rq
)) {
2233 schedule_timeout_uninterruptible(1);
2238 * Ahh, all good. It wasn't running, and it wasn't
2239 * runnable, which means that it will never become
2240 * running in the future either. We're all done!
2249 * kick_process - kick a running thread to enter/exit the kernel
2250 * @p: the to-be-kicked thread
2252 * Cause a process which is running on another CPU to enter
2253 * kernel-mode, without any delay. (to get signals handled.)
2255 * NOTE: this function doesnt have to take the runqueue lock,
2256 * because all it wants to ensure is that the remote task enters
2257 * the kernel. If the IPI races and the task has been migrated
2258 * to another CPU then no harm is done and the purpose has been
2261 void kick_process(struct task_struct
*p
)
2267 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2268 smp_send_reschedule(cpu
);
2271 EXPORT_SYMBOL_GPL(kick_process
);
2272 #endif /* CONFIG_SMP */
2275 * task_oncpu_function_call - call a function on the cpu on which a task runs
2276 * @p: the task to evaluate
2277 * @func: the function to be called
2278 * @info: the function call argument
2280 * Calls the function @func when the task is currently running. This might
2281 * be on the current CPU, which just calls the function directly
2283 void task_oncpu_function_call(struct task_struct
*p
,
2284 void (*func
) (void *info
), void *info
)
2291 smp_call_function_single(cpu
, func
, info
, 1);
2297 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2299 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2302 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2304 /* Look for allowed, online CPU in same node. */
2305 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2306 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2309 /* Any allowed, online CPU? */
2310 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2311 if (dest_cpu
< nr_cpu_ids
)
2314 /* No more Mr. Nice Guy. */
2315 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2316 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2318 * Don't tell them about moving exiting tasks or
2319 * kernel threads (both mm NULL), since they never
2322 if (p
->mm
&& printk_ratelimit()) {
2323 printk(KERN_INFO
"process %d (%s) no "
2324 "longer affine to cpu%d\n",
2325 task_pid_nr(p
), p
->comm
, cpu
);
2333 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2336 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2338 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2341 * In order not to call set_task_cpu() on a blocking task we need
2342 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2345 * Since this is common to all placement strategies, this lives here.
2347 * [ this allows ->select_task() to simply return task_cpu(p) and
2348 * not worry about this generic constraint ]
2350 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2352 cpu
= select_fallback_rq(task_cpu(p
), p
);
2357 static void update_avg(u64
*avg
, u64 sample
)
2359 s64 diff
= sample
- *avg
;
2365 * try_to_wake_up - wake up a thread
2366 * @p: the to-be-woken-up thread
2367 * @state: the mask of task states that can be woken
2368 * @sync: do a synchronous wakeup?
2370 * Put it on the run-queue if it's not already there. The "current"
2371 * thread is always on the run-queue (except when the actual
2372 * re-schedule is in progress), and as such you're allowed to do
2373 * the simpler "current->state = TASK_RUNNING" to mark yourself
2374 * runnable without the overhead of this.
2376 * returns failure only if the task is already active.
2378 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2381 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2382 unsigned long flags
;
2383 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2386 this_cpu
= get_cpu();
2389 rq
= task_rq_lock(p
, &flags
);
2390 if (!(p
->state
& state
))
2400 if (unlikely(task_running(rq
, p
)))
2404 * In order to handle concurrent wakeups and release the rq->lock
2405 * we put the task in TASK_WAKING state.
2407 * First fix up the nr_uninterruptible count:
2409 if (task_contributes_to_load(p
)) {
2410 if (likely(cpu_online(orig_cpu
)))
2411 rq
->nr_uninterruptible
--;
2413 this_rq()->nr_uninterruptible
--;
2415 p
->state
= TASK_WAKING
;
2417 if (p
->sched_class
->task_waking
) {
2418 p
->sched_class
->task_waking(rq
, p
);
2419 en_flags
|= ENQUEUE_WAKING
;
2422 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2423 if (cpu
!= orig_cpu
)
2424 set_task_cpu(p
, cpu
);
2425 __task_rq_unlock(rq
);
2428 raw_spin_lock(&rq
->lock
);
2431 * We migrated the task without holding either rq->lock, however
2432 * since the task is not on the task list itself, nobody else
2433 * will try and migrate the task, hence the rq should match the
2434 * cpu we just moved it to.
2436 WARN_ON(task_cpu(p
) != cpu
);
2437 WARN_ON(p
->state
!= TASK_WAKING
);
2439 #ifdef CONFIG_SCHEDSTATS
2440 schedstat_inc(rq
, ttwu_count
);
2441 if (cpu
== this_cpu
)
2442 schedstat_inc(rq
, ttwu_local
);
2444 struct sched_domain
*sd
;
2445 for_each_domain(this_cpu
, sd
) {
2446 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2447 schedstat_inc(sd
, ttwu_wake_remote
);
2452 #endif /* CONFIG_SCHEDSTATS */
2455 #endif /* CONFIG_SMP */
2456 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2457 if (wake_flags
& WF_SYNC
)
2458 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2459 if (orig_cpu
!= cpu
)
2460 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2461 if (cpu
== this_cpu
)
2462 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2464 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2465 activate_task(rq
, p
, en_flags
);
2469 trace_sched_wakeup(p
, success
);
2470 check_preempt_curr(rq
, p
, wake_flags
);
2472 p
->state
= TASK_RUNNING
;
2474 if (p
->sched_class
->task_woken
)
2475 p
->sched_class
->task_woken(rq
, p
);
2477 if (unlikely(rq
->idle_stamp
)) {
2478 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2479 u64 max
= 2*sysctl_sched_migration_cost
;
2484 update_avg(&rq
->avg_idle
, delta
);
2489 task_rq_unlock(rq
, &flags
);
2496 * wake_up_process - Wake up a specific process
2497 * @p: The process to be woken up.
2499 * Attempt to wake up the nominated process and move it to the set of runnable
2500 * processes. Returns 1 if the process was woken up, 0 if it was already
2503 * It may be assumed that this function implies a write memory barrier before
2504 * changing the task state if and only if any tasks are woken up.
2506 int wake_up_process(struct task_struct
*p
)
2508 return try_to_wake_up(p
, TASK_ALL
, 0);
2510 EXPORT_SYMBOL(wake_up_process
);
2512 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2514 return try_to_wake_up(p
, state
, 0);
2518 * Perform scheduler related setup for a newly forked process p.
2519 * p is forked by current.
2521 * __sched_fork() is basic setup used by init_idle() too:
2523 static void __sched_fork(struct task_struct
*p
)
2525 p
->se
.exec_start
= 0;
2526 p
->se
.sum_exec_runtime
= 0;
2527 p
->se
.prev_sum_exec_runtime
= 0;
2528 p
->se
.nr_migrations
= 0;
2530 #ifdef CONFIG_SCHEDSTATS
2531 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2534 INIT_LIST_HEAD(&p
->rt
.run_list
);
2536 INIT_LIST_HEAD(&p
->se
.group_node
);
2538 #ifdef CONFIG_PREEMPT_NOTIFIERS
2539 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2544 * fork()/clone()-time setup:
2546 void sched_fork(struct task_struct
*p
, int clone_flags
)
2548 int cpu
= get_cpu();
2552 * We mark the process as running here. This guarantees that
2553 * nobody will actually run it, and a signal or other external
2554 * event cannot wake it up and insert it on the runqueue either.
2556 p
->state
= TASK_RUNNING
;
2559 * Revert to default priority/policy on fork if requested.
2561 if (unlikely(p
->sched_reset_on_fork
)) {
2562 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2563 p
->policy
= SCHED_NORMAL
;
2564 p
->normal_prio
= p
->static_prio
;
2567 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2568 p
->static_prio
= NICE_TO_PRIO(0);
2569 p
->normal_prio
= p
->static_prio
;
2574 * We don't need the reset flag anymore after the fork. It has
2575 * fulfilled its duty:
2577 p
->sched_reset_on_fork
= 0;
2581 * Make sure we do not leak PI boosting priority to the child.
2583 p
->prio
= current
->normal_prio
;
2585 if (!rt_prio(p
->prio
))
2586 p
->sched_class
= &fair_sched_class
;
2588 if (p
->sched_class
->task_fork
)
2589 p
->sched_class
->task_fork(p
);
2592 * The child is not yet in the pid-hash so no cgroup attach races,
2593 * and the cgroup is pinned to this child due to cgroup_fork()
2594 * is ran before sched_fork().
2596 * Silence PROVE_RCU.
2599 set_task_cpu(p
, cpu
);
2602 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2603 if (likely(sched_info_on()))
2604 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2606 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2609 #ifdef CONFIG_PREEMPT
2610 /* Want to start with kernel preemption disabled. */
2611 task_thread_info(p
)->preempt_count
= 1;
2613 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2619 * wake_up_new_task - wake up a newly created task for the first time.
2621 * This function will do some initial scheduler statistics housekeeping
2622 * that must be done for every newly created context, then puts the task
2623 * on the runqueue and wakes it.
2625 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2627 unsigned long flags
;
2629 int cpu __maybe_unused
= get_cpu();
2632 rq
= task_rq_lock(p
, &flags
);
2633 p
->state
= TASK_WAKING
;
2636 * Fork balancing, do it here and not earlier because:
2637 * - cpus_allowed can change in the fork path
2638 * - any previously selected cpu might disappear through hotplug
2640 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2641 * without people poking at ->cpus_allowed.
2643 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2644 set_task_cpu(p
, cpu
);
2646 p
->state
= TASK_RUNNING
;
2647 task_rq_unlock(rq
, &flags
);
2650 rq
= task_rq_lock(p
, &flags
);
2651 activate_task(rq
, p
, 0);
2652 trace_sched_wakeup_new(p
, 1);
2653 check_preempt_curr(rq
, p
, WF_FORK
);
2655 if (p
->sched_class
->task_woken
)
2656 p
->sched_class
->task_woken(rq
, p
);
2658 task_rq_unlock(rq
, &flags
);
2662 #ifdef CONFIG_PREEMPT_NOTIFIERS
2665 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2666 * @notifier: notifier struct to register
2668 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2670 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2672 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2675 * preempt_notifier_unregister - no longer interested in preemption notifications
2676 * @notifier: notifier struct to unregister
2678 * This is safe to call from within a preemption notifier.
2680 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2682 hlist_del(¬ifier
->link
);
2684 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2686 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2688 struct preempt_notifier
*notifier
;
2689 struct hlist_node
*node
;
2691 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2692 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2696 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2697 struct task_struct
*next
)
2699 struct preempt_notifier
*notifier
;
2700 struct hlist_node
*node
;
2702 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2703 notifier
->ops
->sched_out(notifier
, next
);
2706 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2708 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2713 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2714 struct task_struct
*next
)
2718 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2721 * prepare_task_switch - prepare to switch tasks
2722 * @rq: the runqueue preparing to switch
2723 * @prev: the current task that is being switched out
2724 * @next: the task we are going to switch to.
2726 * This is called with the rq lock held and interrupts off. It must
2727 * be paired with a subsequent finish_task_switch after the context
2730 * prepare_task_switch sets up locking and calls architecture specific
2734 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2735 struct task_struct
*next
)
2737 fire_sched_out_preempt_notifiers(prev
, next
);
2738 prepare_lock_switch(rq
, next
);
2739 prepare_arch_switch(next
);
2743 * finish_task_switch - clean up after a task-switch
2744 * @rq: runqueue associated with task-switch
2745 * @prev: the thread we just switched away from.
2747 * finish_task_switch must be called after the context switch, paired
2748 * with a prepare_task_switch call before the context switch.
2749 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2750 * and do any other architecture-specific cleanup actions.
2752 * Note that we may have delayed dropping an mm in context_switch(). If
2753 * so, we finish that here outside of the runqueue lock. (Doing it
2754 * with the lock held can cause deadlocks; see schedule() for
2757 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2758 __releases(rq
->lock
)
2760 struct mm_struct
*mm
= rq
->prev_mm
;
2766 * A task struct has one reference for the use as "current".
2767 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2768 * schedule one last time. The schedule call will never return, and
2769 * the scheduled task must drop that reference.
2770 * The test for TASK_DEAD must occur while the runqueue locks are
2771 * still held, otherwise prev could be scheduled on another cpu, die
2772 * there before we look at prev->state, and then the reference would
2774 * Manfred Spraul <manfred@colorfullife.com>
2776 prev_state
= prev
->state
;
2777 finish_arch_switch(prev
);
2778 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2779 local_irq_disable();
2780 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2781 perf_event_task_sched_in(current
);
2782 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2784 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2785 finish_lock_switch(rq
, prev
);
2787 fire_sched_in_preempt_notifiers(current
);
2790 if (unlikely(prev_state
== TASK_DEAD
)) {
2792 * Remove function-return probe instances associated with this
2793 * task and put them back on the free list.
2795 kprobe_flush_task(prev
);
2796 put_task_struct(prev
);
2802 /* assumes rq->lock is held */
2803 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2805 if (prev
->sched_class
->pre_schedule
)
2806 prev
->sched_class
->pre_schedule(rq
, prev
);
2809 /* rq->lock is NOT held, but preemption is disabled */
2810 static inline void post_schedule(struct rq
*rq
)
2812 if (rq
->post_schedule
) {
2813 unsigned long flags
;
2815 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2816 if (rq
->curr
->sched_class
->post_schedule
)
2817 rq
->curr
->sched_class
->post_schedule(rq
);
2818 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2820 rq
->post_schedule
= 0;
2826 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2830 static inline void post_schedule(struct rq
*rq
)
2837 * schedule_tail - first thing a freshly forked thread must call.
2838 * @prev: the thread we just switched away from.
2840 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2841 __releases(rq
->lock
)
2843 struct rq
*rq
= this_rq();
2845 finish_task_switch(rq
, prev
);
2848 * FIXME: do we need to worry about rq being invalidated by the
2853 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2854 /* In this case, finish_task_switch does not reenable preemption */
2857 if (current
->set_child_tid
)
2858 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2862 * context_switch - switch to the new MM and the new
2863 * thread's register state.
2866 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2867 struct task_struct
*next
)
2869 struct mm_struct
*mm
, *oldmm
;
2871 prepare_task_switch(rq
, prev
, next
);
2872 trace_sched_switch(prev
, next
);
2874 oldmm
= prev
->active_mm
;
2876 * For paravirt, this is coupled with an exit in switch_to to
2877 * combine the page table reload and the switch backend into
2880 arch_start_context_switch(prev
);
2883 next
->active_mm
= oldmm
;
2884 atomic_inc(&oldmm
->mm_count
);
2885 enter_lazy_tlb(oldmm
, next
);
2887 switch_mm(oldmm
, mm
, next
);
2889 if (likely(!prev
->mm
)) {
2890 prev
->active_mm
= NULL
;
2891 rq
->prev_mm
= oldmm
;
2894 * Since the runqueue lock will be released by the next
2895 * task (which is an invalid locking op but in the case
2896 * of the scheduler it's an obvious special-case), so we
2897 * do an early lockdep release here:
2899 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2900 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2903 /* Here we just switch the register state and the stack. */
2904 switch_to(prev
, next
, prev
);
2908 * this_rq must be evaluated again because prev may have moved
2909 * CPUs since it called schedule(), thus the 'rq' on its stack
2910 * frame will be invalid.
2912 finish_task_switch(this_rq(), prev
);
2916 * nr_running, nr_uninterruptible and nr_context_switches:
2918 * externally visible scheduler statistics: current number of runnable
2919 * threads, current number of uninterruptible-sleeping threads, total
2920 * number of context switches performed since bootup.
2922 unsigned long nr_running(void)
2924 unsigned long i
, sum
= 0;
2926 for_each_online_cpu(i
)
2927 sum
+= cpu_rq(i
)->nr_running
;
2932 unsigned long nr_uninterruptible(void)
2934 unsigned long i
, sum
= 0;
2936 for_each_possible_cpu(i
)
2937 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2940 * Since we read the counters lockless, it might be slightly
2941 * inaccurate. Do not allow it to go below zero though:
2943 if (unlikely((long)sum
< 0))
2949 unsigned long long nr_context_switches(void)
2952 unsigned long long sum
= 0;
2954 for_each_possible_cpu(i
)
2955 sum
+= cpu_rq(i
)->nr_switches
;
2960 unsigned long nr_iowait(void)
2962 unsigned long i
, sum
= 0;
2964 for_each_possible_cpu(i
)
2965 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2970 unsigned long nr_iowait_cpu(int cpu
)
2972 struct rq
*this = cpu_rq(cpu
);
2973 return atomic_read(&this->nr_iowait
);
2976 unsigned long this_cpu_load(void)
2978 struct rq
*this = this_rq();
2979 return this->cpu_load
[0];
2983 /* Variables and functions for calc_load */
2984 static atomic_long_t calc_load_tasks
;
2985 static unsigned long calc_load_update
;
2986 unsigned long avenrun
[3];
2987 EXPORT_SYMBOL(avenrun
);
2989 static long calc_load_fold_active(struct rq
*this_rq
)
2991 long nr_active
, delta
= 0;
2993 nr_active
= this_rq
->nr_running
;
2994 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2996 if (nr_active
!= this_rq
->calc_load_active
) {
2997 delta
= nr_active
- this_rq
->calc_load_active
;
2998 this_rq
->calc_load_active
= nr_active
;
3004 static unsigned long
3005 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3008 load
+= active
* (FIXED_1
- exp
);
3009 load
+= 1UL << (FSHIFT
- 1);
3010 return load
>> FSHIFT
;
3015 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3017 * When making the ILB scale, we should try to pull this in as well.
3019 static atomic_long_t calc_load_tasks_idle
;
3021 static void calc_load_account_idle(struct rq
*this_rq
)
3025 delta
= calc_load_fold_active(this_rq
);
3027 atomic_long_add(delta
, &calc_load_tasks_idle
);
3030 static long calc_load_fold_idle(void)
3035 * Its got a race, we don't care...
3037 if (atomic_long_read(&calc_load_tasks_idle
))
3038 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3044 * fixed_power_int - compute: x^n, in O(log n) time
3046 * @x: base of the power
3047 * @frac_bits: fractional bits of @x
3048 * @n: power to raise @x to.
3050 * By exploiting the relation between the definition of the natural power
3051 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3052 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3053 * (where: n_i \elem {0, 1}, the binary vector representing n),
3054 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3055 * of course trivially computable in O(log_2 n), the length of our binary
3058 static unsigned long
3059 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3061 unsigned long result
= 1UL << frac_bits
;
3066 result
+= 1UL << (frac_bits
- 1);
3067 result
>>= frac_bits
;
3073 x
+= 1UL << (frac_bits
- 1);
3081 * a1 = a0 * e + a * (1 - e)
3083 * a2 = a1 * e + a * (1 - e)
3084 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3085 * = a0 * e^2 + a * (1 - e) * (1 + e)
3087 * a3 = a2 * e + a * (1 - e)
3088 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3089 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3093 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3094 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3095 * = a0 * e^n + a * (1 - e^n)
3097 * [1] application of the geometric series:
3100 * S_n := \Sum x^i = -------------
3103 static unsigned long
3104 calc_load_n(unsigned long load
, unsigned long exp
,
3105 unsigned long active
, unsigned int n
)
3108 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3112 * NO_HZ can leave us missing all per-cpu ticks calling
3113 * calc_load_account_active(), but since an idle CPU folds its delta into
3114 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3115 * in the pending idle delta if our idle period crossed a load cycle boundary.
3117 * Once we've updated the global active value, we need to apply the exponential
3118 * weights adjusted to the number of cycles missed.
3120 static void calc_global_nohz(unsigned long ticks
)
3122 long delta
, active
, n
;
3124 if (time_before(jiffies
, calc_load_update
))
3128 * If we crossed a calc_load_update boundary, make sure to fold
3129 * any pending idle changes, the respective CPUs might have
3130 * missed the tick driven calc_load_account_active() update
3133 delta
= calc_load_fold_idle();
3135 atomic_long_add(delta
, &calc_load_tasks
);
3138 * If we were idle for multiple load cycles, apply them.
3140 if (ticks
>= LOAD_FREQ
) {
3141 n
= ticks
/ LOAD_FREQ
;
3143 active
= atomic_long_read(&calc_load_tasks
);
3144 active
= active
> 0 ? active
* FIXED_1
: 0;
3146 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3147 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3148 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3150 calc_load_update
+= n
* LOAD_FREQ
;
3154 * Its possible the remainder of the above division also crosses
3155 * a LOAD_FREQ period, the regular check in calc_global_load()
3156 * which comes after this will take care of that.
3158 * Consider us being 11 ticks before a cycle completion, and us
3159 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3160 * age us 4 cycles, and the test in calc_global_load() will
3161 * pick up the final one.
3165 static void calc_load_account_idle(struct rq
*this_rq
)
3169 static inline long calc_load_fold_idle(void)
3174 static void calc_global_nohz(unsigned long ticks
)
3180 * get_avenrun - get the load average array
3181 * @loads: pointer to dest load array
3182 * @offset: offset to add
3183 * @shift: shift count to shift the result left
3185 * These values are estimates at best, so no need for locking.
3187 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3189 loads
[0] = (avenrun
[0] + offset
) << shift
;
3190 loads
[1] = (avenrun
[1] + offset
) << shift
;
3191 loads
[2] = (avenrun
[2] + offset
) << shift
;
3195 * calc_load - update the avenrun load estimates 10 ticks after the
3196 * CPUs have updated calc_load_tasks.
3198 void calc_global_load(unsigned long ticks
)
3202 calc_global_nohz(ticks
);
3204 if (time_before(jiffies
, calc_load_update
+ 10))
3207 active
= atomic_long_read(&calc_load_tasks
);
3208 active
= active
> 0 ? active
* FIXED_1
: 0;
3210 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3211 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3212 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3214 calc_load_update
+= LOAD_FREQ
;
3218 * Called from update_cpu_load() to periodically update this CPU's
3221 static void calc_load_account_active(struct rq
*this_rq
)
3225 if (time_before(jiffies
, this_rq
->calc_load_update
))
3228 delta
= calc_load_fold_active(this_rq
);
3229 delta
+= calc_load_fold_idle();
3231 atomic_long_add(delta
, &calc_load_tasks
);
3233 this_rq
->calc_load_update
+= LOAD_FREQ
;
3237 * Update rq->cpu_load[] statistics. This function is usually called every
3238 * scheduler tick (TICK_NSEC).
3240 static void update_cpu_load(struct rq
*this_rq
)
3242 unsigned long this_load
= this_rq
->load
.weight
;
3245 this_rq
->nr_load_updates
++;
3247 /* Update our load: */
3248 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3249 unsigned long old_load
, new_load
;
3251 /* scale is effectively 1 << i now, and >> i divides by scale */
3253 old_load
= this_rq
->cpu_load
[i
];
3254 new_load
= this_load
;
3256 * Round up the averaging division if load is increasing. This
3257 * prevents us from getting stuck on 9 if the load is 10, for
3260 if (new_load
> old_load
)
3261 new_load
+= scale
-1;
3262 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3265 calc_load_account_active(this_rq
);
3267 sched_avg_update(this_rq
);
3273 * sched_exec - execve() is a valuable balancing opportunity, because at
3274 * this point the task has the smallest effective memory and cache footprint.
3276 void sched_exec(void)
3278 struct task_struct
*p
= current
;
3279 unsigned long flags
;
3283 rq
= task_rq_lock(p
, &flags
);
3284 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3285 if (dest_cpu
== smp_processor_id())
3289 * select_task_rq() can race against ->cpus_allowed
3291 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3292 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3293 struct migration_arg arg
= { p
, dest_cpu
};
3295 task_rq_unlock(rq
, &flags
);
3296 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3300 task_rq_unlock(rq
, &flags
);
3305 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3307 EXPORT_PER_CPU_SYMBOL(kstat
);
3310 * Return any ns on the sched_clock that have not yet been accounted in
3311 * @p in case that task is currently running.
3313 * Called with task_rq_lock() held on @rq.
3315 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3319 if (task_current(rq
, p
)) {
3320 update_rq_clock(rq
);
3321 ns
= rq
->clock_task
- p
->se
.exec_start
;
3329 unsigned long long task_delta_exec(struct task_struct
*p
)
3331 unsigned long flags
;
3335 rq
= task_rq_lock(p
, &flags
);
3336 ns
= do_task_delta_exec(p
, rq
);
3337 task_rq_unlock(rq
, &flags
);
3343 * Return accounted runtime for the task.
3344 * In case the task is currently running, return the runtime plus current's
3345 * pending runtime that have not been accounted yet.
3347 unsigned long long task_sched_runtime(struct task_struct
*p
)
3349 unsigned long flags
;
3353 rq
= task_rq_lock(p
, &flags
);
3354 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3355 task_rq_unlock(rq
, &flags
);
3361 * Return sum_exec_runtime for the thread group.
3362 * In case the task is currently running, return the sum plus current's
3363 * pending runtime that have not been accounted yet.
3365 * Note that the thread group might have other running tasks as well,
3366 * so the return value not includes other pending runtime that other
3367 * running tasks might have.
3369 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3371 struct task_cputime totals
;
3372 unsigned long flags
;
3376 rq
= task_rq_lock(p
, &flags
);
3377 thread_group_cputime(p
, &totals
);
3378 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3379 task_rq_unlock(rq
, &flags
);
3385 * Account user cpu time to a process.
3386 * @p: the process that the cpu time gets accounted to
3387 * @cputime: the cpu time spent in user space since the last update
3388 * @cputime_scaled: cputime scaled by cpu frequency
3390 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3391 cputime_t cputime_scaled
)
3393 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3396 /* Add user time to process. */
3397 p
->utime
= cputime_add(p
->utime
, cputime
);
3398 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3399 account_group_user_time(p
, cputime
);
3401 /* Add user time to cpustat. */
3402 tmp
= cputime_to_cputime64(cputime
);
3403 if (TASK_NICE(p
) > 0)
3404 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3406 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3408 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3409 /* Account for user time used */
3410 acct_update_integrals(p
);
3414 * Account guest cpu time to a process.
3415 * @p: the process that the cpu time gets accounted to
3416 * @cputime: the cpu time spent in virtual machine since the last update
3417 * @cputime_scaled: cputime scaled by cpu frequency
3419 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3420 cputime_t cputime_scaled
)
3423 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3425 tmp
= cputime_to_cputime64(cputime
);
3427 /* Add guest time to process. */
3428 p
->utime
= cputime_add(p
->utime
, cputime
);
3429 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3430 account_group_user_time(p
, cputime
);
3431 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3433 /* Add guest time to cpustat. */
3434 if (TASK_NICE(p
) > 0) {
3435 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3436 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3438 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3439 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3444 * Account system cpu time to a process.
3445 * @p: the process that the cpu time gets accounted to
3446 * @hardirq_offset: the offset to subtract from hardirq_count()
3447 * @cputime: the cpu time spent in kernel space since the last update
3448 * @cputime_scaled: cputime scaled by cpu frequency
3450 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3451 cputime_t cputime
, cputime_t cputime_scaled
)
3453 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3456 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3457 account_guest_time(p
, cputime
, cputime_scaled
);
3461 /* Add system time to process. */
3462 p
->stime
= cputime_add(p
->stime
, cputime
);
3463 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3464 account_group_system_time(p
, cputime
);
3466 /* Add system time to cpustat. */
3467 tmp
= cputime_to_cputime64(cputime
);
3468 if (hardirq_count() - hardirq_offset
)
3469 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3470 else if (in_serving_softirq())
3471 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3473 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3475 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3477 /* Account for system time used */
3478 acct_update_integrals(p
);
3482 * Account for involuntary wait time.
3483 * @steal: the cpu time spent in involuntary wait
3485 void account_steal_time(cputime_t cputime
)
3487 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3488 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3490 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3494 * Account for idle time.
3495 * @cputime: the cpu time spent in idle wait
3497 void account_idle_time(cputime_t cputime
)
3499 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3500 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3501 struct rq
*rq
= this_rq();
3503 if (atomic_read(&rq
->nr_iowait
) > 0)
3504 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3506 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3509 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3512 * Account a single tick of cpu time.
3513 * @p: the process that the cpu time gets accounted to
3514 * @user_tick: indicates if the tick is a user or a system tick
3516 void account_process_tick(struct task_struct
*p
, int user_tick
)
3518 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3519 struct rq
*rq
= this_rq();
3522 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3523 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3524 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3527 account_idle_time(cputime_one_jiffy
);
3531 * Account multiple ticks of steal time.
3532 * @p: the process from which the cpu time has been stolen
3533 * @ticks: number of stolen ticks
3535 void account_steal_ticks(unsigned long ticks
)
3537 account_steal_time(jiffies_to_cputime(ticks
));
3541 * Account multiple ticks of idle time.
3542 * @ticks: number of stolen ticks
3544 void account_idle_ticks(unsigned long ticks
)
3546 account_idle_time(jiffies_to_cputime(ticks
));
3552 * Use precise platform statistics if available:
3554 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3555 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3561 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3563 struct task_cputime cputime
;
3565 thread_group_cputime(p
, &cputime
);
3567 *ut
= cputime
.utime
;
3568 *st
= cputime
.stime
;
3572 #ifndef nsecs_to_cputime
3573 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3576 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3578 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3581 * Use CFS's precise accounting:
3583 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3589 do_div(temp
, total
);
3590 utime
= (cputime_t
)temp
;
3595 * Compare with previous values, to keep monotonicity:
3597 p
->prev_utime
= max(p
->prev_utime
, utime
);
3598 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3600 *ut
= p
->prev_utime
;
3601 *st
= p
->prev_stime
;
3605 * Must be called with siglock held.
3607 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3609 struct signal_struct
*sig
= p
->signal
;
3610 struct task_cputime cputime
;
3611 cputime_t rtime
, utime
, total
;
3613 thread_group_cputime(p
, &cputime
);
3615 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3616 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3621 temp
*= cputime
.utime
;
3622 do_div(temp
, total
);
3623 utime
= (cputime_t
)temp
;
3627 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3628 sig
->prev_stime
= max(sig
->prev_stime
,
3629 cputime_sub(rtime
, sig
->prev_utime
));
3631 *ut
= sig
->prev_utime
;
3632 *st
= sig
->prev_stime
;
3637 * This function gets called by the timer code, with HZ frequency.
3638 * We call it with interrupts disabled.
3640 * It also gets called by the fork code, when changing the parent's
3643 void scheduler_tick(void)
3645 int cpu
= smp_processor_id();
3646 struct rq
*rq
= cpu_rq(cpu
);
3647 struct task_struct
*curr
= rq
->curr
;
3651 raw_spin_lock(&rq
->lock
);
3652 update_rq_clock(rq
);
3653 update_cpu_load(rq
);
3654 curr
->sched_class
->task_tick(rq
, curr
, 0);
3655 raw_spin_unlock(&rq
->lock
);
3657 perf_event_task_tick(curr
);
3660 rq
->idle_at_tick
= idle_cpu(cpu
);
3661 trigger_load_balance(rq
, cpu
);
3665 notrace
unsigned long get_parent_ip(unsigned long addr
)
3667 if (in_lock_functions(addr
)) {
3668 addr
= CALLER_ADDR2
;
3669 if (in_lock_functions(addr
))
3670 addr
= CALLER_ADDR3
;
3675 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3676 defined(CONFIG_PREEMPT_TRACER))
3678 void __kprobes
add_preempt_count(int val
)
3680 #ifdef CONFIG_DEBUG_PREEMPT
3684 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3687 preempt_count() += val
;
3688 #ifdef CONFIG_DEBUG_PREEMPT
3690 * Spinlock count overflowing soon?
3692 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3695 if (preempt_count() == val
)
3696 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3698 EXPORT_SYMBOL(add_preempt_count
);
3700 void __kprobes
sub_preempt_count(int val
)
3702 #ifdef CONFIG_DEBUG_PREEMPT
3706 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3709 * Is the spinlock portion underflowing?
3711 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3712 !(preempt_count() & PREEMPT_MASK
)))
3716 if (preempt_count() == val
)
3717 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3718 preempt_count() -= val
;
3720 EXPORT_SYMBOL(sub_preempt_count
);
3725 * Print scheduling while atomic bug:
3727 static noinline
void __schedule_bug(struct task_struct
*prev
)
3729 struct pt_regs
*regs
= get_irq_regs();
3731 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3732 prev
->comm
, prev
->pid
, preempt_count());
3734 debug_show_held_locks(prev
);
3736 if (irqs_disabled())
3737 print_irqtrace_events(prev
);
3746 * Various schedule()-time debugging checks and statistics:
3748 static inline void schedule_debug(struct task_struct
*prev
)
3751 * Test if we are atomic. Since do_exit() needs to call into
3752 * schedule() atomically, we ignore that path for now.
3753 * Otherwise, whine if we are scheduling when we should not be.
3755 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3756 __schedule_bug(prev
);
3758 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3760 schedstat_inc(this_rq(), sched_count
);
3761 #ifdef CONFIG_SCHEDSTATS
3762 if (unlikely(prev
->lock_depth
>= 0)) {
3763 schedstat_inc(this_rq(), bkl_count
);
3764 schedstat_inc(prev
, sched_info
.bkl_count
);
3769 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3772 update_rq_clock(rq
);
3773 prev
->sched_class
->put_prev_task(rq
, prev
);
3777 * Pick up the highest-prio task:
3779 static inline struct task_struct
*
3780 pick_next_task(struct rq
*rq
)
3782 const struct sched_class
*class;
3783 struct task_struct
*p
;
3786 * Optimization: we know that if all tasks are in
3787 * the fair class we can call that function directly:
3789 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3790 p
= fair_sched_class
.pick_next_task(rq
);
3795 class = sched_class_highest
;
3797 p
= class->pick_next_task(rq
);
3801 * Will never be NULL as the idle class always
3802 * returns a non-NULL p:
3804 class = class->next
;
3809 * schedule() is the main scheduler function.
3811 asmlinkage
void __sched
schedule(void)
3813 struct task_struct
*prev
, *next
;
3814 unsigned long *switch_count
;
3820 cpu
= smp_processor_id();
3822 rcu_note_context_switch(cpu
);
3824 switch_count
= &prev
->nivcsw
;
3826 release_kernel_lock(prev
);
3827 need_resched_nonpreemptible
:
3829 schedule_debug(prev
);
3831 if (sched_feat(HRTICK
))
3834 raw_spin_lock_irq(&rq
->lock
);
3836 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3837 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3838 prev
->state
= TASK_RUNNING
;
3840 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3841 switch_count
= &prev
->nvcsw
;
3844 pre_schedule(rq
, prev
);
3846 if (unlikely(!rq
->nr_running
))
3847 idle_balance(cpu
, rq
);
3849 put_prev_task(rq
, prev
);
3850 next
= pick_next_task(rq
);
3851 clear_tsk_need_resched(prev
);
3852 rq
->skip_clock_update
= 0;
3854 if (likely(prev
!= next
)) {
3855 sched_info_switch(prev
, next
);
3856 perf_event_task_sched_out(prev
, next
);
3862 context_switch(rq
, prev
, next
); /* unlocks the rq */
3864 * the context switch might have flipped the stack from under
3865 * us, hence refresh the local variables.
3867 cpu
= smp_processor_id();
3870 raw_spin_unlock_irq(&rq
->lock
);
3874 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3876 switch_count
= &prev
->nivcsw
;
3877 goto need_resched_nonpreemptible
;
3880 preempt_enable_no_resched();
3884 EXPORT_SYMBOL(schedule
);
3886 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3888 * Look out! "owner" is an entirely speculative pointer
3889 * access and not reliable.
3891 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3896 if (!sched_feat(OWNER_SPIN
))
3899 #ifdef CONFIG_DEBUG_PAGEALLOC
3901 * Need to access the cpu field knowing that
3902 * DEBUG_PAGEALLOC could have unmapped it if
3903 * the mutex owner just released it and exited.
3905 if (probe_kernel_address(&owner
->cpu
, cpu
))
3912 * Even if the access succeeded (likely case),
3913 * the cpu field may no longer be valid.
3915 if (cpu
>= nr_cpumask_bits
)
3919 * We need to validate that we can do a
3920 * get_cpu() and that we have the percpu area.
3922 if (!cpu_online(cpu
))
3929 * Owner changed, break to re-assess state.
3931 if (lock
->owner
!= owner
) {
3933 * If the lock has switched to a different owner,
3934 * we likely have heavy contention. Return 0 to quit
3935 * optimistic spinning and not contend further:
3943 * Is that owner really running on that cpu?
3945 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3955 #ifdef CONFIG_PREEMPT
3957 * this is the entry point to schedule() from in-kernel preemption
3958 * off of preempt_enable. Kernel preemptions off return from interrupt
3959 * occur there and call schedule directly.
3961 asmlinkage
void __sched
preempt_schedule(void)
3963 struct thread_info
*ti
= current_thread_info();
3966 * If there is a non-zero preempt_count or interrupts are disabled,
3967 * we do not want to preempt the current task. Just return..
3969 if (likely(ti
->preempt_count
|| irqs_disabled()))
3973 add_preempt_count(PREEMPT_ACTIVE
);
3975 sub_preempt_count(PREEMPT_ACTIVE
);
3978 * Check again in case we missed a preemption opportunity
3979 * between schedule and now.
3982 } while (need_resched());
3984 EXPORT_SYMBOL(preempt_schedule
);
3987 * this is the entry point to schedule() from kernel preemption
3988 * off of irq context.
3989 * Note, that this is called and return with irqs disabled. This will
3990 * protect us against recursive calling from irq.
3992 asmlinkage
void __sched
preempt_schedule_irq(void)
3994 struct thread_info
*ti
= current_thread_info();
3996 /* Catch callers which need to be fixed */
3997 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4000 add_preempt_count(PREEMPT_ACTIVE
);
4003 local_irq_disable();
4004 sub_preempt_count(PREEMPT_ACTIVE
);
4007 * Check again in case we missed a preemption opportunity
4008 * between schedule and now.
4011 } while (need_resched());
4014 #endif /* CONFIG_PREEMPT */
4016 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4019 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4021 EXPORT_SYMBOL(default_wake_function
);
4024 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4025 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4026 * number) then we wake all the non-exclusive tasks and one exclusive task.
4028 * There are circumstances in which we can try to wake a task which has already
4029 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4030 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4032 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4033 int nr_exclusive
, int wake_flags
, void *key
)
4035 wait_queue_t
*curr
, *next
;
4037 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4038 unsigned flags
= curr
->flags
;
4040 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4041 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4047 * __wake_up - wake up threads blocked on a waitqueue.
4049 * @mode: which threads
4050 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4051 * @key: is directly passed to the wakeup function
4053 * It may be assumed that this function implies a write memory barrier before
4054 * changing the task state if and only if any tasks are woken up.
4056 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4057 int nr_exclusive
, void *key
)
4059 unsigned long flags
;
4061 spin_lock_irqsave(&q
->lock
, flags
);
4062 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4063 spin_unlock_irqrestore(&q
->lock
, flags
);
4065 EXPORT_SYMBOL(__wake_up
);
4068 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4070 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4072 __wake_up_common(q
, mode
, 1, 0, NULL
);
4074 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4076 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4078 __wake_up_common(q
, mode
, 1, 0, key
);
4082 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4084 * @mode: which threads
4085 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4086 * @key: opaque value to be passed to wakeup targets
4088 * The sync wakeup differs that the waker knows that it will schedule
4089 * away soon, so while the target thread will be woken up, it will not
4090 * be migrated to another CPU - ie. the two threads are 'synchronized'
4091 * with each other. This can prevent needless bouncing between CPUs.
4093 * On UP it can prevent extra preemption.
4095 * It may be assumed that this function implies a write memory barrier before
4096 * changing the task state if and only if any tasks are woken up.
4098 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4099 int nr_exclusive
, void *key
)
4101 unsigned long flags
;
4102 int wake_flags
= WF_SYNC
;
4107 if (unlikely(!nr_exclusive
))
4110 spin_lock_irqsave(&q
->lock
, flags
);
4111 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4112 spin_unlock_irqrestore(&q
->lock
, flags
);
4114 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4117 * __wake_up_sync - see __wake_up_sync_key()
4119 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4121 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4123 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4126 * complete: - signals a single thread waiting on this completion
4127 * @x: holds the state of this particular completion
4129 * This will wake up a single thread waiting on this completion. Threads will be
4130 * awakened in the same order in which they were queued.
4132 * See also complete_all(), wait_for_completion() and related routines.
4134 * It may be assumed that this function implies a write memory barrier before
4135 * changing the task state if and only if any tasks are woken up.
4137 void complete(struct completion
*x
)
4139 unsigned long flags
;
4141 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4143 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4144 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4146 EXPORT_SYMBOL(complete
);
4149 * complete_all: - signals all threads waiting on this completion
4150 * @x: holds the state of this particular completion
4152 * This will wake up all threads waiting on this particular completion event.
4154 * It may be assumed that this function implies a write memory barrier before
4155 * changing the task state if and only if any tasks are woken up.
4157 void complete_all(struct completion
*x
)
4159 unsigned long flags
;
4161 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4162 x
->done
+= UINT_MAX
/2;
4163 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4164 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4166 EXPORT_SYMBOL(complete_all
);
4168 static inline long __sched
4169 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4172 DECLARE_WAITQUEUE(wait
, current
);
4174 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4176 if (signal_pending_state(state
, current
)) {
4177 timeout
= -ERESTARTSYS
;
4180 __set_current_state(state
);
4181 spin_unlock_irq(&x
->wait
.lock
);
4182 timeout
= schedule_timeout(timeout
);
4183 spin_lock_irq(&x
->wait
.lock
);
4184 } while (!x
->done
&& timeout
);
4185 __remove_wait_queue(&x
->wait
, &wait
);
4190 return timeout
?: 1;
4194 wait_for_common(struct completion
*x
, long timeout
, int state
)
4198 spin_lock_irq(&x
->wait
.lock
);
4199 timeout
= do_wait_for_common(x
, timeout
, state
);
4200 spin_unlock_irq(&x
->wait
.lock
);
4205 * wait_for_completion: - waits for completion of a task
4206 * @x: holds the state of this particular completion
4208 * This waits to be signaled for completion of a specific task. It is NOT
4209 * interruptible and there is no timeout.
4211 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4212 * and interrupt capability. Also see complete().
4214 void __sched
wait_for_completion(struct completion
*x
)
4216 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4218 EXPORT_SYMBOL(wait_for_completion
);
4221 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
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 signaled or for a
4226 * specified timeout to expire. The timeout is in jiffies. It is not
4229 unsigned long __sched
4230 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4232 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4234 EXPORT_SYMBOL(wait_for_completion_timeout
);
4237 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4238 * @x: holds the state of this particular completion
4240 * This waits for completion of a specific task to be signaled. It is
4243 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4245 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4246 if (t
== -ERESTARTSYS
)
4250 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4253 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4254 * @x: holds the state of this particular completion
4255 * @timeout: timeout value in jiffies
4257 * This waits for either a completion of a specific task to be signaled or for a
4258 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4260 unsigned long __sched
4261 wait_for_completion_interruptible_timeout(struct completion
*x
,
4262 unsigned long timeout
)
4264 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4266 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4269 * wait_for_completion_killable: - waits for completion of a task (killable)
4270 * @x: holds the state of this particular completion
4272 * This waits to be signaled for completion of a specific task. It can be
4273 * interrupted by a kill signal.
4275 int __sched
wait_for_completion_killable(struct completion
*x
)
4277 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4278 if (t
== -ERESTARTSYS
)
4282 EXPORT_SYMBOL(wait_for_completion_killable
);
4285 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4286 * @x: holds the state of this particular completion
4287 * @timeout: timeout value in jiffies
4289 * This waits for either a completion of a specific task to be
4290 * signaled or for a specified timeout to expire. It can be
4291 * interrupted by a kill signal. The timeout is in jiffies.
4293 unsigned long __sched
4294 wait_for_completion_killable_timeout(struct completion
*x
,
4295 unsigned long timeout
)
4297 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4299 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4302 * try_wait_for_completion - try to decrement a completion without blocking
4303 * @x: completion structure
4305 * Returns: 0 if a decrement cannot be done without blocking
4306 * 1 if a decrement succeeded.
4308 * If a completion is being used as a counting completion,
4309 * attempt to decrement the counter without blocking. This
4310 * enables us to avoid waiting if the resource the completion
4311 * is protecting is not available.
4313 bool try_wait_for_completion(struct completion
*x
)
4315 unsigned long flags
;
4318 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4323 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4326 EXPORT_SYMBOL(try_wait_for_completion
);
4329 * completion_done - Test to see if a completion has any waiters
4330 * @x: completion structure
4332 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4333 * 1 if there are no waiters.
4336 bool completion_done(struct completion
*x
)
4338 unsigned long flags
;
4341 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4344 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4347 EXPORT_SYMBOL(completion_done
);
4350 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4352 unsigned long flags
;
4355 init_waitqueue_entry(&wait
, current
);
4357 __set_current_state(state
);
4359 spin_lock_irqsave(&q
->lock
, flags
);
4360 __add_wait_queue(q
, &wait
);
4361 spin_unlock(&q
->lock
);
4362 timeout
= schedule_timeout(timeout
);
4363 spin_lock_irq(&q
->lock
);
4364 __remove_wait_queue(q
, &wait
);
4365 spin_unlock_irqrestore(&q
->lock
, flags
);
4370 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4372 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4374 EXPORT_SYMBOL(interruptible_sleep_on
);
4377 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4379 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4381 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4383 void __sched
sleep_on(wait_queue_head_t
*q
)
4385 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4387 EXPORT_SYMBOL(sleep_on
);
4389 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4391 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4393 EXPORT_SYMBOL(sleep_on_timeout
);
4395 #ifdef CONFIG_RT_MUTEXES
4398 * rt_mutex_setprio - set the current priority of a task
4400 * @prio: prio value (kernel-internal form)
4402 * This function changes the 'effective' priority of a task. It does
4403 * not touch ->normal_prio like __setscheduler().
4405 * Used by the rt_mutex code to implement priority inheritance logic.
4407 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4409 unsigned long flags
;
4410 int oldprio
, on_rq
, running
;
4412 const struct sched_class
*prev_class
;
4414 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4416 rq
= task_rq_lock(p
, &flags
);
4419 prev_class
= p
->sched_class
;
4420 on_rq
= p
->se
.on_rq
;
4421 running
= task_current(rq
, p
);
4423 dequeue_task(rq
, p
, 0);
4425 p
->sched_class
->put_prev_task(rq
, p
);
4428 p
->sched_class
= &rt_sched_class
;
4430 p
->sched_class
= &fair_sched_class
;
4435 p
->sched_class
->set_curr_task(rq
);
4437 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4439 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4441 task_rq_unlock(rq
, &flags
);
4446 void set_user_nice(struct task_struct
*p
, long nice
)
4448 int old_prio
, delta
, on_rq
;
4449 unsigned long flags
;
4452 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4455 * We have to be careful, if called from sys_setpriority(),
4456 * the task might be in the middle of scheduling on another CPU.
4458 rq
= task_rq_lock(p
, &flags
);
4460 * The RT priorities are set via sched_setscheduler(), but we still
4461 * allow the 'normal' nice value to be set - but as expected
4462 * it wont have any effect on scheduling until the task is
4463 * SCHED_FIFO/SCHED_RR:
4465 if (task_has_rt_policy(p
)) {
4466 p
->static_prio
= NICE_TO_PRIO(nice
);
4469 on_rq
= p
->se
.on_rq
;
4471 dequeue_task(rq
, p
, 0);
4473 p
->static_prio
= NICE_TO_PRIO(nice
);
4476 p
->prio
= effective_prio(p
);
4477 delta
= p
->prio
- old_prio
;
4480 enqueue_task(rq
, p
, 0);
4482 * If the task increased its priority or is running and
4483 * lowered its priority, then reschedule its CPU:
4485 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4486 resched_task(rq
->curr
);
4489 task_rq_unlock(rq
, &flags
);
4491 EXPORT_SYMBOL(set_user_nice
);
4494 * can_nice - check if a task can reduce its nice value
4498 int can_nice(const struct task_struct
*p
, const int nice
)
4500 /* convert nice value [19,-20] to rlimit style value [1,40] */
4501 int nice_rlim
= 20 - nice
;
4503 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4504 capable(CAP_SYS_NICE
));
4507 #ifdef __ARCH_WANT_SYS_NICE
4510 * sys_nice - change the priority of the current process.
4511 * @increment: priority increment
4513 * sys_setpriority is a more generic, but much slower function that
4514 * does similar things.
4516 SYSCALL_DEFINE1(nice
, int, increment
)
4521 * Setpriority might change our priority at the same moment.
4522 * We don't have to worry. Conceptually one call occurs first
4523 * and we have a single winner.
4525 if (increment
< -40)
4530 nice
= TASK_NICE(current
) + increment
;
4536 if (increment
< 0 && !can_nice(current
, nice
))
4539 retval
= security_task_setnice(current
, nice
);
4543 set_user_nice(current
, nice
);
4550 * task_prio - return the priority value of a given task.
4551 * @p: the task in question.
4553 * This is the priority value as seen by users in /proc.
4554 * RT tasks are offset by -200. Normal tasks are centered
4555 * around 0, value goes from -16 to +15.
4557 int task_prio(const struct task_struct
*p
)
4559 return p
->prio
- MAX_RT_PRIO
;
4563 * task_nice - return the nice value of a given task.
4564 * @p: the task in question.
4566 int task_nice(const struct task_struct
*p
)
4568 return TASK_NICE(p
);
4570 EXPORT_SYMBOL(task_nice
);
4573 * idle_cpu - is a given cpu idle currently?
4574 * @cpu: the processor in question.
4576 int idle_cpu(int cpu
)
4578 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4582 * idle_task - return the idle task for a given cpu.
4583 * @cpu: the processor in question.
4585 struct task_struct
*idle_task(int cpu
)
4587 return cpu_rq(cpu
)->idle
;
4591 * find_process_by_pid - find a process with a matching PID value.
4592 * @pid: the pid in question.
4594 static struct task_struct
*find_process_by_pid(pid_t pid
)
4596 return pid
? find_task_by_vpid(pid
) : current
;
4599 /* Actually do priority change: must hold rq lock. */
4601 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4603 BUG_ON(p
->se
.on_rq
);
4606 p
->rt_priority
= prio
;
4607 p
->normal_prio
= normal_prio(p
);
4608 /* we are holding p->pi_lock already */
4609 p
->prio
= rt_mutex_getprio(p
);
4610 if (rt_prio(p
->prio
))
4611 p
->sched_class
= &rt_sched_class
;
4613 p
->sched_class
= &fair_sched_class
;
4618 * check the target process has a UID that matches the current process's
4620 static bool check_same_owner(struct task_struct
*p
)
4622 const struct cred
*cred
= current_cred(), *pcred
;
4626 pcred
= __task_cred(p
);
4627 match
= (cred
->euid
== pcred
->euid
||
4628 cred
->euid
== pcred
->uid
);
4633 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4634 struct sched_param
*param
, bool user
)
4636 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4637 unsigned long flags
;
4638 const struct sched_class
*prev_class
;
4642 /* may grab non-irq protected spin_locks */
4643 BUG_ON(in_interrupt());
4645 /* double check policy once rq lock held */
4647 reset_on_fork
= p
->sched_reset_on_fork
;
4648 policy
= oldpolicy
= p
->policy
;
4650 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4651 policy
&= ~SCHED_RESET_ON_FORK
;
4653 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4654 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4655 policy
!= SCHED_IDLE
)
4660 * Valid priorities for SCHED_FIFO and SCHED_RR are
4661 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4662 * SCHED_BATCH and SCHED_IDLE is 0.
4664 if (param
->sched_priority
< 0 ||
4665 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4666 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4668 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4672 * Allow unprivileged RT tasks to decrease priority:
4674 if (user
&& !capable(CAP_SYS_NICE
)) {
4675 if (rt_policy(policy
)) {
4676 unsigned long rlim_rtprio
;
4678 if (!lock_task_sighand(p
, &flags
))
4680 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4681 unlock_task_sighand(p
, &flags
);
4683 /* can't set/change the rt policy */
4684 if (policy
!= p
->policy
&& !rlim_rtprio
)
4687 /* can't increase priority */
4688 if (param
->sched_priority
> p
->rt_priority
&&
4689 param
->sched_priority
> rlim_rtprio
)
4693 * Like positive nice levels, dont allow tasks to
4694 * move out of SCHED_IDLE either:
4696 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4699 /* can't change other user's priorities */
4700 if (!check_same_owner(p
))
4703 /* Normal users shall not reset the sched_reset_on_fork flag */
4704 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4709 retval
= security_task_setscheduler(p
, policy
, param
);
4715 * make sure no PI-waiters arrive (or leave) while we are
4716 * changing the priority of the task:
4718 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4720 * To be able to change p->policy safely, the apropriate
4721 * runqueue lock must be held.
4723 rq
= __task_rq_lock(p
);
4725 #ifdef CONFIG_RT_GROUP_SCHED
4728 * Do not allow realtime tasks into groups that have no runtime
4731 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4732 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4733 __task_rq_unlock(rq
);
4734 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4740 /* recheck policy now with rq lock held */
4741 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4742 policy
= oldpolicy
= -1;
4743 __task_rq_unlock(rq
);
4744 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4747 on_rq
= p
->se
.on_rq
;
4748 running
= task_current(rq
, p
);
4750 deactivate_task(rq
, p
, 0);
4752 p
->sched_class
->put_prev_task(rq
, p
);
4754 p
->sched_reset_on_fork
= reset_on_fork
;
4757 prev_class
= p
->sched_class
;
4758 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4761 p
->sched_class
->set_curr_task(rq
);
4763 activate_task(rq
, p
, 0);
4765 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4767 __task_rq_unlock(rq
);
4768 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4770 rt_mutex_adjust_pi(p
);
4776 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4777 * @p: the task in question.
4778 * @policy: new policy.
4779 * @param: structure containing the new RT priority.
4781 * NOTE that the task may be already dead.
4783 int sched_setscheduler(struct task_struct
*p
, int policy
,
4784 struct sched_param
*param
)
4786 return __sched_setscheduler(p
, policy
, param
, true);
4788 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4791 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4792 * @p: the task in question.
4793 * @policy: new policy.
4794 * @param: structure containing the new RT priority.
4796 * Just like sched_setscheduler, only don't bother checking if the
4797 * current context has permission. For example, this is needed in
4798 * stop_machine(): we create temporary high priority worker threads,
4799 * but our caller might not have that capability.
4801 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4802 struct sched_param
*param
)
4804 return __sched_setscheduler(p
, policy
, param
, false);
4808 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4810 struct sched_param lparam
;
4811 struct task_struct
*p
;
4814 if (!param
|| pid
< 0)
4816 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4821 p
= find_process_by_pid(pid
);
4823 retval
= sched_setscheduler(p
, policy
, &lparam
);
4830 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4831 * @pid: the pid in question.
4832 * @policy: new policy.
4833 * @param: structure containing the new RT priority.
4835 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4836 struct sched_param __user
*, param
)
4838 /* negative values for policy are not valid */
4842 return do_sched_setscheduler(pid
, policy
, param
);
4846 * sys_sched_setparam - set/change the RT priority of a thread
4847 * @pid: the pid in question.
4848 * @param: structure containing the new RT priority.
4850 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4852 return do_sched_setscheduler(pid
, -1, param
);
4856 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4857 * @pid: the pid in question.
4859 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4861 struct task_struct
*p
;
4869 p
= find_process_by_pid(pid
);
4871 retval
= security_task_getscheduler(p
);
4874 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4881 * sys_sched_getparam - get the RT priority of a thread
4882 * @pid: the pid in question.
4883 * @param: structure containing the RT priority.
4885 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4887 struct sched_param lp
;
4888 struct task_struct
*p
;
4891 if (!param
|| pid
< 0)
4895 p
= find_process_by_pid(pid
);
4900 retval
= security_task_getscheduler(p
);
4904 lp
.sched_priority
= p
->rt_priority
;
4908 * This one might sleep, we cannot do it with a spinlock held ...
4910 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4919 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4921 cpumask_var_t cpus_allowed
, new_mask
;
4922 struct task_struct
*p
;
4928 p
= find_process_by_pid(pid
);
4935 /* Prevent p going away */
4939 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4943 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4945 goto out_free_cpus_allowed
;
4948 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4951 retval
= security_task_setscheduler(p
, 0, NULL
);
4955 cpuset_cpus_allowed(p
, cpus_allowed
);
4956 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4958 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4961 cpuset_cpus_allowed(p
, cpus_allowed
);
4962 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4964 * We must have raced with a concurrent cpuset
4965 * update. Just reset the cpus_allowed to the
4966 * cpuset's cpus_allowed
4968 cpumask_copy(new_mask
, cpus_allowed
);
4973 free_cpumask_var(new_mask
);
4974 out_free_cpus_allowed
:
4975 free_cpumask_var(cpus_allowed
);
4982 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4983 struct cpumask
*new_mask
)
4985 if (len
< cpumask_size())
4986 cpumask_clear(new_mask
);
4987 else if (len
> cpumask_size())
4988 len
= cpumask_size();
4990 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4994 * sys_sched_setaffinity - set the cpu affinity of a process
4995 * @pid: pid of the process
4996 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4997 * @user_mask_ptr: user-space pointer to the new cpu mask
4999 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5000 unsigned long __user
*, user_mask_ptr
)
5002 cpumask_var_t new_mask
;
5005 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5008 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5010 retval
= sched_setaffinity(pid
, new_mask
);
5011 free_cpumask_var(new_mask
);
5015 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5017 struct task_struct
*p
;
5018 unsigned long flags
;
5026 p
= find_process_by_pid(pid
);
5030 retval
= security_task_getscheduler(p
);
5034 rq
= task_rq_lock(p
, &flags
);
5035 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5036 task_rq_unlock(rq
, &flags
);
5046 * sys_sched_getaffinity - get the cpu affinity of a process
5047 * @pid: pid of the process
5048 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5049 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5051 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5052 unsigned long __user
*, user_mask_ptr
)
5057 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5059 if (len
& (sizeof(unsigned long)-1))
5062 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5065 ret
= sched_getaffinity(pid
, mask
);
5067 size_t retlen
= min_t(size_t, len
, cpumask_size());
5069 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5074 free_cpumask_var(mask
);
5080 * sys_sched_yield - yield the current processor to other threads.
5082 * This function yields the current CPU to other tasks. If there are no
5083 * other threads running on this CPU then this function will return.
5085 SYSCALL_DEFINE0(sched_yield
)
5087 struct rq
*rq
= this_rq_lock();
5089 schedstat_inc(rq
, yld_count
);
5090 current
->sched_class
->yield_task(rq
);
5093 * Since we are going to call schedule() anyway, there's
5094 * no need to preempt or enable interrupts:
5096 __release(rq
->lock
);
5097 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5098 do_raw_spin_unlock(&rq
->lock
);
5099 preempt_enable_no_resched();
5106 static inline int should_resched(void)
5108 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5111 static void __cond_resched(void)
5113 add_preempt_count(PREEMPT_ACTIVE
);
5115 sub_preempt_count(PREEMPT_ACTIVE
);
5118 int __sched
_cond_resched(void)
5120 if (should_resched()) {
5126 EXPORT_SYMBOL(_cond_resched
);
5129 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5130 * call schedule, and on return reacquire the lock.
5132 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5133 * operations here to prevent schedule() from being called twice (once via
5134 * spin_unlock(), once by hand).
5136 int __cond_resched_lock(spinlock_t
*lock
)
5138 int resched
= should_resched();
5141 lockdep_assert_held(lock
);
5143 if (spin_needbreak(lock
) || resched
) {
5154 EXPORT_SYMBOL(__cond_resched_lock
);
5156 int __sched
__cond_resched_softirq(void)
5158 BUG_ON(!in_softirq());
5160 if (should_resched()) {
5168 EXPORT_SYMBOL(__cond_resched_softirq
);
5171 * yield - yield the current processor to other threads.
5173 * This is a shortcut for kernel-space yielding - it marks the
5174 * thread runnable and calls sys_sched_yield().
5176 void __sched
yield(void)
5178 set_current_state(TASK_RUNNING
);
5181 EXPORT_SYMBOL(yield
);
5184 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5185 * that process accounting knows that this is a task in IO wait state.
5187 void __sched
io_schedule(void)
5189 struct rq
*rq
= raw_rq();
5191 delayacct_blkio_start();
5192 atomic_inc(&rq
->nr_iowait
);
5193 current
->in_iowait
= 1;
5195 current
->in_iowait
= 0;
5196 atomic_dec(&rq
->nr_iowait
);
5197 delayacct_blkio_end();
5199 EXPORT_SYMBOL(io_schedule
);
5201 long __sched
io_schedule_timeout(long timeout
)
5203 struct rq
*rq
= raw_rq();
5206 delayacct_blkio_start();
5207 atomic_inc(&rq
->nr_iowait
);
5208 current
->in_iowait
= 1;
5209 ret
= schedule_timeout(timeout
);
5210 current
->in_iowait
= 0;
5211 atomic_dec(&rq
->nr_iowait
);
5212 delayacct_blkio_end();
5217 * sys_sched_get_priority_max - return maximum RT priority.
5218 * @policy: scheduling class.
5220 * this syscall returns the maximum rt_priority that can be used
5221 * by a given scheduling class.
5223 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5230 ret
= MAX_USER_RT_PRIO
-1;
5242 * sys_sched_get_priority_min - return minimum RT priority.
5243 * @policy: scheduling class.
5245 * this syscall returns the minimum rt_priority that can be used
5246 * by a given scheduling class.
5248 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5266 * sys_sched_rr_get_interval - return the default timeslice of a process.
5267 * @pid: pid of the process.
5268 * @interval: userspace pointer to the timeslice value.
5270 * this syscall writes the default timeslice value of a given process
5271 * into the user-space timespec buffer. A value of '0' means infinity.
5273 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5274 struct timespec __user
*, interval
)
5276 struct task_struct
*p
;
5277 unsigned int time_slice
;
5278 unsigned long flags
;
5288 p
= find_process_by_pid(pid
);
5292 retval
= security_task_getscheduler(p
);
5296 rq
= task_rq_lock(p
, &flags
);
5297 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5298 task_rq_unlock(rq
, &flags
);
5301 jiffies_to_timespec(time_slice
, &t
);
5302 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5310 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5312 void sched_show_task(struct task_struct
*p
)
5314 unsigned long free
= 0;
5317 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5318 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5319 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5320 #if BITS_PER_LONG == 32
5321 if (state
== TASK_RUNNING
)
5322 printk(KERN_CONT
" running ");
5324 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5326 if (state
== TASK_RUNNING
)
5327 printk(KERN_CONT
" running task ");
5329 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5331 #ifdef CONFIG_DEBUG_STACK_USAGE
5332 free
= stack_not_used(p
);
5334 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5335 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5336 (unsigned long)task_thread_info(p
)->flags
);
5338 show_stack(p
, NULL
);
5341 void show_state_filter(unsigned long state_filter
)
5343 struct task_struct
*g
, *p
;
5345 #if BITS_PER_LONG == 32
5347 " task PC stack pid father\n");
5350 " task PC stack pid father\n");
5352 read_lock(&tasklist_lock
);
5353 do_each_thread(g
, p
) {
5355 * reset the NMI-timeout, listing all files on a slow
5356 * console might take alot of time:
5358 touch_nmi_watchdog();
5359 if (!state_filter
|| (p
->state
& state_filter
))
5361 } while_each_thread(g
, p
);
5363 touch_all_softlockup_watchdogs();
5365 #ifdef CONFIG_SCHED_DEBUG
5366 sysrq_sched_debug_show();
5368 read_unlock(&tasklist_lock
);
5370 * Only show locks if all tasks are dumped:
5373 debug_show_all_locks();
5376 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5378 idle
->sched_class
= &idle_sched_class
;
5382 * init_idle - set up an idle thread for a given CPU
5383 * @idle: task in question
5384 * @cpu: cpu the idle task belongs to
5386 * NOTE: this function does not set the idle thread's NEED_RESCHED
5387 * flag, to make booting more robust.
5389 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5391 struct rq
*rq
= cpu_rq(cpu
);
5392 unsigned long flags
;
5394 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5397 idle
->state
= TASK_RUNNING
;
5398 idle
->se
.exec_start
= sched_clock();
5400 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5402 * We're having a chicken and egg problem, even though we are
5403 * holding rq->lock, the cpu isn't yet set to this cpu so the
5404 * lockdep check in task_group() will fail.
5406 * Similar case to sched_fork(). / Alternatively we could
5407 * use task_rq_lock() here and obtain the other rq->lock.
5412 __set_task_cpu(idle
, cpu
);
5415 rq
->curr
= rq
->idle
= idle
;
5416 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5419 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5421 /* Set the preempt count _outside_ the spinlocks! */
5422 #if defined(CONFIG_PREEMPT)
5423 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5425 task_thread_info(idle
)->preempt_count
= 0;
5428 * The idle tasks have their own, simple scheduling class:
5430 idle
->sched_class
= &idle_sched_class
;
5431 ftrace_graph_init_idle_task(idle
, cpu
);
5435 * In a system that switches off the HZ timer nohz_cpu_mask
5436 * indicates which cpus entered this state. This is used
5437 * in the rcu update to wait only for active cpus. For system
5438 * which do not switch off the HZ timer nohz_cpu_mask should
5439 * always be CPU_BITS_NONE.
5441 cpumask_var_t nohz_cpu_mask
;
5444 * Increase the granularity value when there are more CPUs,
5445 * because with more CPUs the 'effective latency' as visible
5446 * to users decreases. But the relationship is not linear,
5447 * so pick a second-best guess by going with the log2 of the
5450 * This idea comes from the SD scheduler of Con Kolivas:
5452 static int get_update_sysctl_factor(void)
5454 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5455 unsigned int factor
;
5457 switch (sysctl_sched_tunable_scaling
) {
5458 case SCHED_TUNABLESCALING_NONE
:
5461 case SCHED_TUNABLESCALING_LINEAR
:
5464 case SCHED_TUNABLESCALING_LOG
:
5466 factor
= 1 + ilog2(cpus
);
5473 static void update_sysctl(void)
5475 unsigned int factor
= get_update_sysctl_factor();
5477 #define SET_SYSCTL(name) \
5478 (sysctl_##name = (factor) * normalized_sysctl_##name)
5479 SET_SYSCTL(sched_min_granularity
);
5480 SET_SYSCTL(sched_latency
);
5481 SET_SYSCTL(sched_wakeup_granularity
);
5482 SET_SYSCTL(sched_shares_ratelimit
);
5486 static inline void sched_init_granularity(void)
5493 * This is how migration works:
5495 * 1) we invoke migration_cpu_stop() on the target CPU using
5497 * 2) stopper starts to run (implicitly forcing the migrated thread
5499 * 3) it checks whether the migrated task is still in the wrong runqueue.
5500 * 4) if it's in the wrong runqueue then the migration thread removes
5501 * it and puts it into the right queue.
5502 * 5) stopper completes and stop_one_cpu() returns and the migration
5507 * Change a given task's CPU affinity. Migrate the thread to a
5508 * proper CPU and schedule it away if the CPU it's executing on
5509 * is removed from the allowed bitmask.
5511 * NOTE: the caller must have a valid reference to the task, the
5512 * task must not exit() & deallocate itself prematurely. The
5513 * call is not atomic; no spinlocks may be held.
5515 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5517 unsigned long flags
;
5519 unsigned int dest_cpu
;
5523 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5524 * drop the rq->lock and still rely on ->cpus_allowed.
5527 while (task_is_waking(p
))
5529 rq
= task_rq_lock(p
, &flags
);
5530 if (task_is_waking(p
)) {
5531 task_rq_unlock(rq
, &flags
);
5535 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5540 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5541 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5546 if (p
->sched_class
->set_cpus_allowed
)
5547 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5549 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5550 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5553 /* Can the task run on the task's current CPU? If so, we're done */
5554 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5557 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5558 if (migrate_task(p
, dest_cpu
)) {
5559 struct migration_arg arg
= { p
, dest_cpu
};
5560 /* Need help from migration thread: drop lock and wait. */
5561 task_rq_unlock(rq
, &flags
);
5562 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5563 tlb_migrate_finish(p
->mm
);
5567 task_rq_unlock(rq
, &flags
);
5571 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5574 * Move (not current) task off this cpu, onto dest cpu. We're doing
5575 * this because either it can't run here any more (set_cpus_allowed()
5576 * away from this CPU, or CPU going down), or because we're
5577 * attempting to rebalance this task on exec (sched_exec).
5579 * So we race with normal scheduler movements, but that's OK, as long
5580 * as the task is no longer on this CPU.
5582 * Returns non-zero if task was successfully migrated.
5584 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5586 struct rq
*rq_dest
, *rq_src
;
5589 if (unlikely(!cpu_active(dest_cpu
)))
5592 rq_src
= cpu_rq(src_cpu
);
5593 rq_dest
= cpu_rq(dest_cpu
);
5595 double_rq_lock(rq_src
, rq_dest
);
5596 /* Already moved. */
5597 if (task_cpu(p
) != src_cpu
)
5599 /* Affinity changed (again). */
5600 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5604 * If we're not on a rq, the next wake-up will ensure we're
5608 deactivate_task(rq_src
, p
, 0);
5609 set_task_cpu(p
, dest_cpu
);
5610 activate_task(rq_dest
, p
, 0);
5611 check_preempt_curr(rq_dest
, p
, 0);
5616 double_rq_unlock(rq_src
, rq_dest
);
5621 * migration_cpu_stop - this will be executed by a highprio stopper thread
5622 * and performs thread migration by bumping thread off CPU then
5623 * 'pushing' onto another runqueue.
5625 static int migration_cpu_stop(void *data
)
5627 struct migration_arg
*arg
= data
;
5630 * The original target cpu might have gone down and we might
5631 * be on another cpu but it doesn't matter.
5633 local_irq_disable();
5634 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5639 #ifdef CONFIG_HOTPLUG_CPU
5641 * Figure out where task on dead CPU should go, use force if necessary.
5643 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5645 struct rq
*rq
= cpu_rq(dead_cpu
);
5646 int needs_cpu
, uninitialized_var(dest_cpu
);
5647 unsigned long flags
;
5649 local_irq_save(flags
);
5651 raw_spin_lock(&rq
->lock
);
5652 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5654 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5655 raw_spin_unlock(&rq
->lock
);
5657 * It can only fail if we race with set_cpus_allowed(),
5658 * in the racer should migrate the task anyway.
5661 __migrate_task(p
, dead_cpu
, dest_cpu
);
5662 local_irq_restore(flags
);
5666 * While a dead CPU has no uninterruptible tasks queued at this point,
5667 * it might still have a nonzero ->nr_uninterruptible counter, because
5668 * for performance reasons the counter is not stricly tracking tasks to
5669 * their home CPUs. So we just add the counter to another CPU's counter,
5670 * to keep the global sum constant after CPU-down:
5672 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5674 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5675 unsigned long flags
;
5677 local_irq_save(flags
);
5678 double_rq_lock(rq_src
, rq_dest
);
5679 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5680 rq_src
->nr_uninterruptible
= 0;
5681 double_rq_unlock(rq_src
, rq_dest
);
5682 local_irq_restore(flags
);
5685 /* Run through task list and migrate tasks from the dead cpu. */
5686 static void migrate_live_tasks(int src_cpu
)
5688 struct task_struct
*p
, *t
;
5690 read_lock(&tasklist_lock
);
5692 do_each_thread(t
, p
) {
5696 if (task_cpu(p
) == src_cpu
)
5697 move_task_off_dead_cpu(src_cpu
, p
);
5698 } while_each_thread(t
, p
);
5700 read_unlock(&tasklist_lock
);
5704 * Schedules idle task to be the next runnable task on current CPU.
5705 * It does so by boosting its priority to highest possible.
5706 * Used by CPU offline code.
5708 void sched_idle_next(void)
5710 int this_cpu
= smp_processor_id();
5711 struct rq
*rq
= cpu_rq(this_cpu
);
5712 struct task_struct
*p
= rq
->idle
;
5713 unsigned long flags
;
5715 /* cpu has to be offline */
5716 BUG_ON(cpu_online(this_cpu
));
5719 * Strictly not necessary since rest of the CPUs are stopped by now
5720 * and interrupts disabled on the current cpu.
5722 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5724 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5726 activate_task(rq
, p
, 0);
5728 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5732 * Ensures that the idle task is using init_mm right before its cpu goes
5735 void idle_task_exit(void)
5737 struct mm_struct
*mm
= current
->active_mm
;
5739 BUG_ON(cpu_online(smp_processor_id()));
5742 switch_mm(mm
, &init_mm
, current
);
5746 /* called under rq->lock with disabled interrupts */
5747 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5749 struct rq
*rq
= cpu_rq(dead_cpu
);
5751 /* Must be exiting, otherwise would be on tasklist. */
5752 BUG_ON(!p
->exit_state
);
5754 /* Cannot have done final schedule yet: would have vanished. */
5755 BUG_ON(p
->state
== TASK_DEAD
);
5760 * Drop lock around migration; if someone else moves it,
5761 * that's OK. No task can be added to this CPU, so iteration is
5764 raw_spin_unlock_irq(&rq
->lock
);
5765 move_task_off_dead_cpu(dead_cpu
, p
);
5766 raw_spin_lock_irq(&rq
->lock
);
5771 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5772 static void migrate_dead_tasks(unsigned int dead_cpu
)
5774 struct rq
*rq
= cpu_rq(dead_cpu
);
5775 struct task_struct
*next
;
5778 if (!rq
->nr_running
)
5780 next
= pick_next_task(rq
);
5783 next
->sched_class
->put_prev_task(rq
, next
);
5784 migrate_dead(dead_cpu
, next
);
5790 * remove the tasks which were accounted by rq from calc_load_tasks.
5792 static void calc_global_load_remove(struct rq
*rq
)
5794 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5795 rq
->calc_load_active
= 0;
5797 #endif /* CONFIG_HOTPLUG_CPU */
5799 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5801 static struct ctl_table sd_ctl_dir
[] = {
5803 .procname
= "sched_domain",
5809 static struct ctl_table sd_ctl_root
[] = {
5811 .procname
= "kernel",
5813 .child
= sd_ctl_dir
,
5818 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5820 struct ctl_table
*entry
=
5821 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5826 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5828 struct ctl_table
*entry
;
5831 * In the intermediate directories, both the child directory and
5832 * procname are dynamically allocated and could fail but the mode
5833 * will always be set. In the lowest directory the names are
5834 * static strings and all have proc handlers.
5836 for (entry
= *tablep
; entry
->mode
; entry
++) {
5838 sd_free_ctl_entry(&entry
->child
);
5839 if (entry
->proc_handler
== NULL
)
5840 kfree(entry
->procname
);
5848 set_table_entry(struct ctl_table
*entry
,
5849 const char *procname
, void *data
, int maxlen
,
5850 mode_t mode
, proc_handler
*proc_handler
)
5852 entry
->procname
= procname
;
5854 entry
->maxlen
= maxlen
;
5856 entry
->proc_handler
= proc_handler
;
5859 static struct ctl_table
*
5860 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5862 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5867 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5868 sizeof(long), 0644, proc_doulongvec_minmax
);
5869 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5870 sizeof(long), 0644, proc_doulongvec_minmax
);
5871 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5872 sizeof(int), 0644, proc_dointvec_minmax
);
5873 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5874 sizeof(int), 0644, proc_dointvec_minmax
);
5875 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5876 sizeof(int), 0644, proc_dointvec_minmax
);
5877 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5878 sizeof(int), 0644, proc_dointvec_minmax
);
5879 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5880 sizeof(int), 0644, proc_dointvec_minmax
);
5881 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5882 sizeof(int), 0644, proc_dointvec_minmax
);
5883 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5884 sizeof(int), 0644, proc_dointvec_minmax
);
5885 set_table_entry(&table
[9], "cache_nice_tries",
5886 &sd
->cache_nice_tries
,
5887 sizeof(int), 0644, proc_dointvec_minmax
);
5888 set_table_entry(&table
[10], "flags", &sd
->flags
,
5889 sizeof(int), 0644, proc_dointvec_minmax
);
5890 set_table_entry(&table
[11], "name", sd
->name
,
5891 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5892 /* &table[12] is terminator */
5897 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5899 struct ctl_table
*entry
, *table
;
5900 struct sched_domain
*sd
;
5901 int domain_num
= 0, i
;
5904 for_each_domain(cpu
, sd
)
5906 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5911 for_each_domain(cpu
, sd
) {
5912 snprintf(buf
, 32, "domain%d", i
);
5913 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5915 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5922 static struct ctl_table_header
*sd_sysctl_header
;
5923 static void register_sched_domain_sysctl(void)
5925 int i
, cpu_num
= num_possible_cpus();
5926 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5929 WARN_ON(sd_ctl_dir
[0].child
);
5930 sd_ctl_dir
[0].child
= entry
;
5935 for_each_possible_cpu(i
) {
5936 snprintf(buf
, 32, "cpu%d", i
);
5937 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5939 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5943 WARN_ON(sd_sysctl_header
);
5944 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5947 /* may be called multiple times per register */
5948 static void unregister_sched_domain_sysctl(void)
5950 if (sd_sysctl_header
)
5951 unregister_sysctl_table(sd_sysctl_header
);
5952 sd_sysctl_header
= NULL
;
5953 if (sd_ctl_dir
[0].child
)
5954 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5957 static void register_sched_domain_sysctl(void)
5960 static void unregister_sched_domain_sysctl(void)
5965 static void set_rq_online(struct rq
*rq
)
5968 const struct sched_class
*class;
5970 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5973 for_each_class(class) {
5974 if (class->rq_online
)
5975 class->rq_online(rq
);
5980 static void set_rq_offline(struct rq
*rq
)
5983 const struct sched_class
*class;
5985 for_each_class(class) {
5986 if (class->rq_offline
)
5987 class->rq_offline(rq
);
5990 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5996 * migration_call - callback that gets triggered when a CPU is added.
5997 * Here we can start up the necessary migration thread for the new CPU.
5999 static int __cpuinit
6000 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6002 int cpu
= (long)hcpu
;
6003 unsigned long flags
;
6004 struct rq
*rq
= cpu_rq(cpu
);
6008 case CPU_UP_PREPARE
:
6009 case CPU_UP_PREPARE_FROZEN
:
6010 rq
->calc_load_update
= calc_load_update
;
6014 case CPU_ONLINE_FROZEN
:
6015 /* Update our root-domain */
6016 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6018 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6022 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6025 #ifdef CONFIG_HOTPLUG_CPU
6027 case CPU_DEAD_FROZEN
:
6028 migrate_live_tasks(cpu
);
6029 /* Idle task back to normal (off runqueue, low prio) */
6030 raw_spin_lock_irq(&rq
->lock
);
6031 deactivate_task(rq
, rq
->idle
, 0);
6032 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6033 rq
->idle
->sched_class
= &idle_sched_class
;
6034 migrate_dead_tasks(cpu
);
6035 raw_spin_unlock_irq(&rq
->lock
);
6036 migrate_nr_uninterruptible(rq
);
6037 BUG_ON(rq
->nr_running
!= 0);
6038 calc_global_load_remove(rq
);
6042 case CPU_DYING_FROZEN
:
6043 /* Update our root-domain */
6044 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6046 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6049 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6057 * Register at high priority so that task migration (migrate_all_tasks)
6058 * happens before everything else. This has to be lower priority than
6059 * the notifier in the perf_event subsystem, though.
6061 static struct notifier_block __cpuinitdata migration_notifier
= {
6062 .notifier_call
= migration_call
,
6066 static int __init
migration_init(void)
6068 void *cpu
= (void *)(long)smp_processor_id();
6071 /* Start one for the boot CPU: */
6072 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6073 BUG_ON(err
== NOTIFY_BAD
);
6074 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6075 register_cpu_notifier(&migration_notifier
);
6079 early_initcall(migration_init
);
6084 #ifdef CONFIG_SCHED_DEBUG
6086 static __read_mostly
int sched_domain_debug_enabled
;
6088 static int __init
sched_domain_debug_setup(char *str
)
6090 sched_domain_debug_enabled
= 1;
6094 early_param("sched_debug", sched_domain_debug_setup
);
6096 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6097 struct cpumask
*groupmask
)
6099 struct sched_group
*group
= sd
->groups
;
6102 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6103 cpumask_clear(groupmask
);
6105 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6107 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6108 printk("does not load-balance\n");
6110 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6115 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6117 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6118 printk(KERN_ERR
"ERROR: domain->span does not contain "
6121 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6122 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6126 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6130 printk(KERN_ERR
"ERROR: group is NULL\n");
6134 if (!group
->cpu_power
) {
6135 printk(KERN_CONT
"\n");
6136 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6141 if (!cpumask_weight(sched_group_cpus(group
))) {
6142 printk(KERN_CONT
"\n");
6143 printk(KERN_ERR
"ERROR: empty group\n");
6147 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6148 printk(KERN_CONT
"\n");
6149 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6153 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6155 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6157 printk(KERN_CONT
" %s", str
);
6158 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6159 printk(KERN_CONT
" (cpu_power = %d)",
6163 group
= group
->next
;
6164 } while (group
!= sd
->groups
);
6165 printk(KERN_CONT
"\n");
6167 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6168 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6171 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6172 printk(KERN_ERR
"ERROR: parent span is not a superset "
6173 "of domain->span\n");
6177 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6179 cpumask_var_t groupmask
;
6182 if (!sched_domain_debug_enabled
)
6186 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6190 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6192 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6193 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6198 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6205 free_cpumask_var(groupmask
);
6207 #else /* !CONFIG_SCHED_DEBUG */
6208 # define sched_domain_debug(sd, cpu) do { } while (0)
6209 #endif /* CONFIG_SCHED_DEBUG */
6211 static int sd_degenerate(struct sched_domain
*sd
)
6213 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6216 /* Following flags need at least 2 groups */
6217 if (sd
->flags
& (SD_LOAD_BALANCE
|
6218 SD_BALANCE_NEWIDLE
|
6222 SD_SHARE_PKG_RESOURCES
)) {
6223 if (sd
->groups
!= sd
->groups
->next
)
6227 /* Following flags don't use groups */
6228 if (sd
->flags
& (SD_WAKE_AFFINE
))
6235 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6237 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6239 if (sd_degenerate(parent
))
6242 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6245 /* Flags needing groups don't count if only 1 group in parent */
6246 if (parent
->groups
== parent
->groups
->next
) {
6247 pflags
&= ~(SD_LOAD_BALANCE
|
6248 SD_BALANCE_NEWIDLE
|
6252 SD_SHARE_PKG_RESOURCES
);
6253 if (nr_node_ids
== 1)
6254 pflags
&= ~SD_SERIALIZE
;
6256 if (~cflags
& pflags
)
6262 static void free_rootdomain(struct root_domain
*rd
)
6264 synchronize_sched();
6266 cpupri_cleanup(&rd
->cpupri
);
6268 free_cpumask_var(rd
->rto_mask
);
6269 free_cpumask_var(rd
->online
);
6270 free_cpumask_var(rd
->span
);
6274 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6276 struct root_domain
*old_rd
= NULL
;
6277 unsigned long flags
;
6279 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6284 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6287 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6290 * If we dont want to free the old_rt yet then
6291 * set old_rd to NULL to skip the freeing later
6294 if (!atomic_dec_and_test(&old_rd
->refcount
))
6298 atomic_inc(&rd
->refcount
);
6301 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6302 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6305 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6308 free_rootdomain(old_rd
);
6311 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6313 gfp_t gfp
= GFP_KERNEL
;
6315 memset(rd
, 0, sizeof(*rd
));
6320 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6322 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6324 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6327 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6332 free_cpumask_var(rd
->rto_mask
);
6334 free_cpumask_var(rd
->online
);
6336 free_cpumask_var(rd
->span
);
6341 static void init_defrootdomain(void)
6343 init_rootdomain(&def_root_domain
, true);
6345 atomic_set(&def_root_domain
.refcount
, 1);
6348 static struct root_domain
*alloc_rootdomain(void)
6350 struct root_domain
*rd
;
6352 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6356 if (init_rootdomain(rd
, false) != 0) {
6365 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6366 * hold the hotplug lock.
6369 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6371 struct rq
*rq
= cpu_rq(cpu
);
6372 struct sched_domain
*tmp
;
6374 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6375 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6377 /* Remove the sched domains which do not contribute to scheduling. */
6378 for (tmp
= sd
; tmp
; ) {
6379 struct sched_domain
*parent
= tmp
->parent
;
6383 if (sd_parent_degenerate(tmp
, parent
)) {
6384 tmp
->parent
= parent
->parent
;
6386 parent
->parent
->child
= tmp
;
6391 if (sd
&& sd_degenerate(sd
)) {
6397 sched_domain_debug(sd
, cpu
);
6399 rq_attach_root(rq
, rd
);
6400 rcu_assign_pointer(rq
->sd
, sd
);
6403 /* cpus with isolated domains */
6404 static cpumask_var_t cpu_isolated_map
;
6406 /* Setup the mask of cpus configured for isolated domains */
6407 static int __init
isolated_cpu_setup(char *str
)
6409 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6410 cpulist_parse(str
, cpu_isolated_map
);
6414 __setup("isolcpus=", isolated_cpu_setup
);
6417 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6418 * to a function which identifies what group(along with sched group) a CPU
6419 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6420 * (due to the fact that we keep track of groups covered with a struct cpumask).
6422 * init_sched_build_groups will build a circular linked list of the groups
6423 * covered by the given span, and will set each group's ->cpumask correctly,
6424 * and ->cpu_power to 0.
6427 init_sched_build_groups(const struct cpumask
*span
,
6428 const struct cpumask
*cpu_map
,
6429 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6430 struct sched_group
**sg
,
6431 struct cpumask
*tmpmask
),
6432 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6434 struct sched_group
*first
= NULL
, *last
= NULL
;
6437 cpumask_clear(covered
);
6439 for_each_cpu(i
, span
) {
6440 struct sched_group
*sg
;
6441 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6444 if (cpumask_test_cpu(i
, covered
))
6447 cpumask_clear(sched_group_cpus(sg
));
6450 for_each_cpu(j
, span
) {
6451 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6454 cpumask_set_cpu(j
, covered
);
6455 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6466 #define SD_NODES_PER_DOMAIN 16
6471 * find_next_best_node - find the next node to include in a sched_domain
6472 * @node: node whose sched_domain we're building
6473 * @used_nodes: nodes already in the sched_domain
6475 * Find the next node to include in a given scheduling domain. Simply
6476 * finds the closest node not already in the @used_nodes map.
6478 * Should use nodemask_t.
6480 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6482 int i
, n
, val
, min_val
, best_node
= 0;
6486 for (i
= 0; i
< nr_node_ids
; i
++) {
6487 /* Start at @node */
6488 n
= (node
+ i
) % nr_node_ids
;
6490 if (!nr_cpus_node(n
))
6493 /* Skip already used nodes */
6494 if (node_isset(n
, *used_nodes
))
6497 /* Simple min distance search */
6498 val
= node_distance(node
, n
);
6500 if (val
< min_val
) {
6506 node_set(best_node
, *used_nodes
);
6511 * sched_domain_node_span - get a cpumask for a node's sched_domain
6512 * @node: node whose cpumask we're constructing
6513 * @span: resulting cpumask
6515 * Given a node, construct a good cpumask for its sched_domain to span. It
6516 * should be one that prevents unnecessary balancing, but also spreads tasks
6519 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6521 nodemask_t used_nodes
;
6524 cpumask_clear(span
);
6525 nodes_clear(used_nodes
);
6527 cpumask_or(span
, span
, cpumask_of_node(node
));
6528 node_set(node
, used_nodes
);
6530 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6531 int next_node
= find_next_best_node(node
, &used_nodes
);
6533 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6536 #endif /* CONFIG_NUMA */
6538 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6541 * The cpus mask in sched_group and sched_domain hangs off the end.
6543 * ( See the the comments in include/linux/sched.h:struct sched_group
6544 * and struct sched_domain. )
6546 struct static_sched_group
{
6547 struct sched_group sg
;
6548 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6551 struct static_sched_domain
{
6552 struct sched_domain sd
;
6553 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6559 cpumask_var_t domainspan
;
6560 cpumask_var_t covered
;
6561 cpumask_var_t notcovered
;
6563 cpumask_var_t nodemask
;
6564 cpumask_var_t this_sibling_map
;
6565 cpumask_var_t this_core_map
;
6566 cpumask_var_t send_covered
;
6567 cpumask_var_t tmpmask
;
6568 struct sched_group
**sched_group_nodes
;
6569 struct root_domain
*rd
;
6573 sa_sched_groups
= 0,
6578 sa_this_sibling_map
,
6580 sa_sched_group_nodes
,
6590 * SMT sched-domains:
6592 #ifdef CONFIG_SCHED_SMT
6593 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6594 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6597 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6598 struct sched_group
**sg
, struct cpumask
*unused
)
6601 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6604 #endif /* CONFIG_SCHED_SMT */
6607 * multi-core sched-domains:
6609 #ifdef CONFIG_SCHED_MC
6610 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6611 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6612 #endif /* CONFIG_SCHED_MC */
6614 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6616 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6617 struct sched_group
**sg
, struct cpumask
*mask
)
6621 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6622 group
= cpumask_first(mask
);
6624 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6627 #elif defined(CONFIG_SCHED_MC)
6629 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6630 struct sched_group
**sg
, struct cpumask
*unused
)
6633 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6638 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6639 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6642 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6643 struct sched_group
**sg
, struct cpumask
*mask
)
6646 #ifdef CONFIG_SCHED_MC
6647 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6648 group
= cpumask_first(mask
);
6649 #elif defined(CONFIG_SCHED_SMT)
6650 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6651 group
= cpumask_first(mask
);
6656 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6662 * The init_sched_build_groups can't handle what we want to do with node
6663 * groups, so roll our own. Now each node has its own list of groups which
6664 * gets dynamically allocated.
6666 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6667 static struct sched_group
***sched_group_nodes_bycpu
;
6669 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6670 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6672 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6673 struct sched_group
**sg
,
6674 struct cpumask
*nodemask
)
6678 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6679 group
= cpumask_first(nodemask
);
6682 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6686 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6688 struct sched_group
*sg
= group_head
;
6694 for_each_cpu(j
, sched_group_cpus(sg
)) {
6695 struct sched_domain
*sd
;
6697 sd
= &per_cpu(phys_domains
, j
).sd
;
6698 if (j
!= group_first_cpu(sd
->groups
)) {
6700 * Only add "power" once for each
6706 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6709 } while (sg
!= group_head
);
6712 static int build_numa_sched_groups(struct s_data
*d
,
6713 const struct cpumask
*cpu_map
, int num
)
6715 struct sched_domain
*sd
;
6716 struct sched_group
*sg
, *prev
;
6719 cpumask_clear(d
->covered
);
6720 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6721 if (cpumask_empty(d
->nodemask
)) {
6722 d
->sched_group_nodes
[num
] = NULL
;
6726 sched_domain_node_span(num
, d
->domainspan
);
6727 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6729 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6732 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6736 d
->sched_group_nodes
[num
] = sg
;
6738 for_each_cpu(j
, d
->nodemask
) {
6739 sd
= &per_cpu(node_domains
, j
).sd
;
6744 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6746 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6749 for (j
= 0; j
< nr_node_ids
; j
++) {
6750 n
= (num
+ j
) % nr_node_ids
;
6751 cpumask_complement(d
->notcovered
, d
->covered
);
6752 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6753 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6754 if (cpumask_empty(d
->tmpmask
))
6756 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6757 if (cpumask_empty(d
->tmpmask
))
6759 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6763 "Can not alloc domain group for node %d\n", j
);
6767 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6768 sg
->next
= prev
->next
;
6769 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6776 #endif /* CONFIG_NUMA */
6779 /* Free memory allocated for various sched_group structures */
6780 static void free_sched_groups(const struct cpumask
*cpu_map
,
6781 struct cpumask
*nodemask
)
6785 for_each_cpu(cpu
, cpu_map
) {
6786 struct sched_group
**sched_group_nodes
6787 = sched_group_nodes_bycpu
[cpu
];
6789 if (!sched_group_nodes
)
6792 for (i
= 0; i
< nr_node_ids
; i
++) {
6793 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6795 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6796 if (cpumask_empty(nodemask
))
6806 if (oldsg
!= sched_group_nodes
[i
])
6809 kfree(sched_group_nodes
);
6810 sched_group_nodes_bycpu
[cpu
] = NULL
;
6813 #else /* !CONFIG_NUMA */
6814 static void free_sched_groups(const struct cpumask
*cpu_map
,
6815 struct cpumask
*nodemask
)
6818 #endif /* CONFIG_NUMA */
6821 * Initialize sched groups cpu_power.
6823 * cpu_power indicates the capacity of sched group, which is used while
6824 * distributing the load between different sched groups in a sched domain.
6825 * Typically cpu_power for all the groups in a sched domain will be same unless
6826 * there are asymmetries in the topology. If there are asymmetries, group
6827 * having more cpu_power will pickup more load compared to the group having
6830 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6832 struct sched_domain
*child
;
6833 struct sched_group
*group
;
6837 WARN_ON(!sd
|| !sd
->groups
);
6839 if (cpu
!= group_first_cpu(sd
->groups
))
6842 sd
->groups
->group_weight
= cpumask_weight(sched_group_cpus(sd
->groups
));
6846 sd
->groups
->cpu_power
= 0;
6849 power
= SCHED_LOAD_SCALE
;
6850 weight
= cpumask_weight(sched_domain_span(sd
));
6852 * SMT siblings share the power of a single core.
6853 * Usually multiple threads get a better yield out of
6854 * that one core than a single thread would have,
6855 * reflect that in sd->smt_gain.
6857 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6858 power
*= sd
->smt_gain
;
6860 power
>>= SCHED_LOAD_SHIFT
;
6862 sd
->groups
->cpu_power
+= power
;
6867 * Add cpu_power of each child group to this groups cpu_power.
6869 group
= child
->groups
;
6871 sd
->groups
->cpu_power
+= group
->cpu_power
;
6872 group
= group
->next
;
6873 } while (group
!= child
->groups
);
6877 * Initializers for schedule domains
6878 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6881 #ifdef CONFIG_SCHED_DEBUG
6882 # define SD_INIT_NAME(sd, type) sd->name = #type
6884 # define SD_INIT_NAME(sd, type) do { } while (0)
6887 #define SD_INIT(sd, type) sd_init_##type(sd)
6889 #define SD_INIT_FUNC(type) \
6890 static noinline void sd_init_##type(struct sched_domain *sd) \
6892 memset(sd, 0, sizeof(*sd)); \
6893 *sd = SD_##type##_INIT; \
6894 sd->level = SD_LV_##type; \
6895 SD_INIT_NAME(sd, type); \
6900 SD_INIT_FUNC(ALLNODES
)
6903 #ifdef CONFIG_SCHED_SMT
6904 SD_INIT_FUNC(SIBLING
)
6906 #ifdef CONFIG_SCHED_MC
6910 static int default_relax_domain_level
= -1;
6912 static int __init
setup_relax_domain_level(char *str
)
6916 val
= simple_strtoul(str
, NULL
, 0);
6917 if (val
< SD_LV_MAX
)
6918 default_relax_domain_level
= val
;
6922 __setup("relax_domain_level=", setup_relax_domain_level
);
6924 static void set_domain_attribute(struct sched_domain
*sd
,
6925 struct sched_domain_attr
*attr
)
6929 if (!attr
|| attr
->relax_domain_level
< 0) {
6930 if (default_relax_domain_level
< 0)
6933 request
= default_relax_domain_level
;
6935 request
= attr
->relax_domain_level
;
6936 if (request
< sd
->level
) {
6937 /* turn off idle balance on this domain */
6938 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6940 /* turn on idle balance on this domain */
6941 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6945 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6946 const struct cpumask
*cpu_map
)
6949 case sa_sched_groups
:
6950 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6951 d
->sched_group_nodes
= NULL
;
6953 free_rootdomain(d
->rd
); /* fall through */
6955 free_cpumask_var(d
->tmpmask
); /* fall through */
6956 case sa_send_covered
:
6957 free_cpumask_var(d
->send_covered
); /* fall through */
6958 case sa_this_core_map
:
6959 free_cpumask_var(d
->this_core_map
); /* fall through */
6960 case sa_this_sibling_map
:
6961 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6963 free_cpumask_var(d
->nodemask
); /* fall through */
6964 case sa_sched_group_nodes
:
6966 kfree(d
->sched_group_nodes
); /* fall through */
6968 free_cpumask_var(d
->notcovered
); /* fall through */
6970 free_cpumask_var(d
->covered
); /* fall through */
6972 free_cpumask_var(d
->domainspan
); /* fall through */
6979 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6980 const struct cpumask
*cpu_map
)
6983 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6985 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6986 return sa_domainspan
;
6987 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6989 /* Allocate the per-node list of sched groups */
6990 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6991 sizeof(struct sched_group
*), GFP_KERNEL
);
6992 if (!d
->sched_group_nodes
) {
6993 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6994 return sa_notcovered
;
6996 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6998 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6999 return sa_sched_group_nodes
;
7000 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
7002 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
7003 return sa_this_sibling_map
;
7004 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
7005 return sa_this_core_map
;
7006 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
7007 return sa_send_covered
;
7008 d
->rd
= alloc_rootdomain();
7010 printk(KERN_WARNING
"Cannot alloc root domain\n");
7013 return sa_rootdomain
;
7016 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
7017 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
7019 struct sched_domain
*sd
= NULL
;
7021 struct sched_domain
*parent
;
7024 if (cpumask_weight(cpu_map
) >
7025 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7026 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7027 SD_INIT(sd
, ALLNODES
);
7028 set_domain_attribute(sd
, attr
);
7029 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7030 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7035 sd
= &per_cpu(node_domains
, i
).sd
;
7037 set_domain_attribute(sd
, attr
);
7038 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7039 sd
->parent
= parent
;
7042 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7047 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7048 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7049 struct sched_domain
*parent
, int i
)
7051 struct sched_domain
*sd
;
7052 sd
= &per_cpu(phys_domains
, i
).sd
;
7054 set_domain_attribute(sd
, attr
);
7055 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7056 sd
->parent
= parent
;
7059 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7063 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7064 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7065 struct sched_domain
*parent
, int i
)
7067 struct sched_domain
*sd
= parent
;
7068 #ifdef CONFIG_SCHED_MC
7069 sd
= &per_cpu(core_domains
, i
).sd
;
7071 set_domain_attribute(sd
, attr
);
7072 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7073 sd
->parent
= parent
;
7075 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7080 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7081 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7082 struct sched_domain
*parent
, int i
)
7084 struct sched_domain
*sd
= parent
;
7085 #ifdef CONFIG_SCHED_SMT
7086 sd
= &per_cpu(cpu_domains
, i
).sd
;
7087 SD_INIT(sd
, SIBLING
);
7088 set_domain_attribute(sd
, attr
);
7089 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7090 sd
->parent
= parent
;
7092 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7097 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7098 const struct cpumask
*cpu_map
, int cpu
)
7101 #ifdef CONFIG_SCHED_SMT
7102 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7103 cpumask_and(d
->this_sibling_map
, cpu_map
,
7104 topology_thread_cpumask(cpu
));
7105 if (cpu
== cpumask_first(d
->this_sibling_map
))
7106 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7108 d
->send_covered
, d
->tmpmask
);
7111 #ifdef CONFIG_SCHED_MC
7112 case SD_LV_MC
: /* set up multi-core groups */
7113 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7114 if (cpu
== cpumask_first(d
->this_core_map
))
7115 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7117 d
->send_covered
, d
->tmpmask
);
7120 case SD_LV_CPU
: /* set up physical groups */
7121 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7122 if (!cpumask_empty(d
->nodemask
))
7123 init_sched_build_groups(d
->nodemask
, cpu_map
,
7125 d
->send_covered
, d
->tmpmask
);
7128 case SD_LV_ALLNODES
:
7129 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7130 d
->send_covered
, d
->tmpmask
);
7139 * Build sched domains for a given set of cpus and attach the sched domains
7140 * to the individual cpus
7142 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7143 struct sched_domain_attr
*attr
)
7145 enum s_alloc alloc_state
= sa_none
;
7147 struct sched_domain
*sd
;
7153 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7154 if (alloc_state
!= sa_rootdomain
)
7156 alloc_state
= sa_sched_groups
;
7159 * Set up domains for cpus specified by the cpu_map.
7161 for_each_cpu(i
, cpu_map
) {
7162 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7165 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7166 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7167 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7168 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7171 for_each_cpu(i
, cpu_map
) {
7172 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7173 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7176 /* Set up physical groups */
7177 for (i
= 0; i
< nr_node_ids
; i
++)
7178 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7181 /* Set up node groups */
7183 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7185 for (i
= 0; i
< nr_node_ids
; i
++)
7186 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7190 /* Calculate CPU power for physical packages and nodes */
7191 #ifdef CONFIG_SCHED_SMT
7192 for_each_cpu(i
, cpu_map
) {
7193 sd
= &per_cpu(cpu_domains
, i
).sd
;
7194 init_sched_groups_power(i
, sd
);
7197 #ifdef CONFIG_SCHED_MC
7198 for_each_cpu(i
, cpu_map
) {
7199 sd
= &per_cpu(core_domains
, i
).sd
;
7200 init_sched_groups_power(i
, sd
);
7204 for_each_cpu(i
, cpu_map
) {
7205 sd
= &per_cpu(phys_domains
, i
).sd
;
7206 init_sched_groups_power(i
, sd
);
7210 for (i
= 0; i
< nr_node_ids
; i
++)
7211 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7213 if (d
.sd_allnodes
) {
7214 struct sched_group
*sg
;
7216 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7218 init_numa_sched_groups_power(sg
);
7222 /* Attach the domains */
7223 for_each_cpu(i
, cpu_map
) {
7224 #ifdef CONFIG_SCHED_SMT
7225 sd
= &per_cpu(cpu_domains
, i
).sd
;
7226 #elif defined(CONFIG_SCHED_MC)
7227 sd
= &per_cpu(core_domains
, i
).sd
;
7229 sd
= &per_cpu(phys_domains
, i
).sd
;
7231 cpu_attach_domain(sd
, d
.rd
, i
);
7234 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7235 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7239 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7243 static int build_sched_domains(const struct cpumask
*cpu_map
)
7245 return __build_sched_domains(cpu_map
, NULL
);
7248 static cpumask_var_t
*doms_cur
; /* current sched domains */
7249 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7250 static struct sched_domain_attr
*dattr_cur
;
7251 /* attribues of custom domains in 'doms_cur' */
7254 * Special case: If a kmalloc of a doms_cur partition (array of
7255 * cpumask) fails, then fallback to a single sched domain,
7256 * as determined by the single cpumask fallback_doms.
7258 static cpumask_var_t fallback_doms
;
7261 * arch_update_cpu_topology lets virtualized architectures update the
7262 * cpu core maps. It is supposed to return 1 if the topology changed
7263 * or 0 if it stayed the same.
7265 int __attribute__((weak
)) arch_update_cpu_topology(void)
7270 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7273 cpumask_var_t
*doms
;
7275 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7278 for (i
= 0; i
< ndoms
; i
++) {
7279 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7280 free_sched_domains(doms
, i
);
7287 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7290 for (i
= 0; i
< ndoms
; i
++)
7291 free_cpumask_var(doms
[i
]);
7296 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7297 * For now this just excludes isolated cpus, but could be used to
7298 * exclude other special cases in the future.
7300 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7304 arch_update_cpu_topology();
7306 doms_cur
= alloc_sched_domains(ndoms_cur
);
7308 doms_cur
= &fallback_doms
;
7309 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7311 err
= build_sched_domains(doms_cur
[0]);
7312 register_sched_domain_sysctl();
7317 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7318 struct cpumask
*tmpmask
)
7320 free_sched_groups(cpu_map
, tmpmask
);
7324 * Detach sched domains from a group of cpus specified in cpu_map
7325 * These cpus will now be attached to the NULL domain
7327 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7329 /* Save because hotplug lock held. */
7330 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7333 for_each_cpu(i
, cpu_map
)
7334 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7335 synchronize_sched();
7336 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7339 /* handle null as "default" */
7340 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7341 struct sched_domain_attr
*new, int idx_new
)
7343 struct sched_domain_attr tmp
;
7350 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7351 new ? (new + idx_new
) : &tmp
,
7352 sizeof(struct sched_domain_attr
));
7356 * Partition sched domains as specified by the 'ndoms_new'
7357 * cpumasks in the array doms_new[] of cpumasks. This compares
7358 * doms_new[] to the current sched domain partitioning, doms_cur[].
7359 * It destroys each deleted domain and builds each new domain.
7361 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7362 * The masks don't intersect (don't overlap.) We should setup one
7363 * sched domain for each mask. CPUs not in any of the cpumasks will
7364 * not be load balanced. If the same cpumask appears both in the
7365 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7368 * The passed in 'doms_new' should be allocated using
7369 * alloc_sched_domains. This routine takes ownership of it and will
7370 * free_sched_domains it when done with it. If the caller failed the
7371 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7372 * and partition_sched_domains() will fallback to the single partition
7373 * 'fallback_doms', it also forces the domains to be rebuilt.
7375 * If doms_new == NULL it will be replaced with cpu_online_mask.
7376 * ndoms_new == 0 is a special case for destroying existing domains,
7377 * and it will not create the default domain.
7379 * Call with hotplug lock held
7381 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7382 struct sched_domain_attr
*dattr_new
)
7387 mutex_lock(&sched_domains_mutex
);
7389 /* always unregister in case we don't destroy any domains */
7390 unregister_sched_domain_sysctl();
7392 /* Let architecture update cpu core mappings. */
7393 new_topology
= arch_update_cpu_topology();
7395 n
= doms_new
? ndoms_new
: 0;
7397 /* Destroy deleted domains */
7398 for (i
= 0; i
< ndoms_cur
; i
++) {
7399 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7400 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7401 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7404 /* no match - a current sched domain not in new doms_new[] */
7405 detach_destroy_domains(doms_cur
[i
]);
7410 if (doms_new
== NULL
) {
7412 doms_new
= &fallback_doms
;
7413 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7414 WARN_ON_ONCE(dattr_new
);
7417 /* Build new domains */
7418 for (i
= 0; i
< ndoms_new
; i
++) {
7419 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7420 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7421 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7424 /* no match - add a new doms_new */
7425 __build_sched_domains(doms_new
[i
],
7426 dattr_new
? dattr_new
+ i
: NULL
);
7431 /* Remember the new sched domains */
7432 if (doms_cur
!= &fallback_doms
)
7433 free_sched_domains(doms_cur
, ndoms_cur
);
7434 kfree(dattr_cur
); /* kfree(NULL) is safe */
7435 doms_cur
= doms_new
;
7436 dattr_cur
= dattr_new
;
7437 ndoms_cur
= ndoms_new
;
7439 register_sched_domain_sysctl();
7441 mutex_unlock(&sched_domains_mutex
);
7444 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7445 static void arch_reinit_sched_domains(void)
7449 /* Destroy domains first to force the rebuild */
7450 partition_sched_domains(0, NULL
, NULL
);
7452 rebuild_sched_domains();
7456 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7458 unsigned int level
= 0;
7460 if (sscanf(buf
, "%u", &level
) != 1)
7464 * level is always be positive so don't check for
7465 * level < POWERSAVINGS_BALANCE_NONE which is 0
7466 * What happens on 0 or 1 byte write,
7467 * need to check for count as well?
7470 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7474 sched_smt_power_savings
= level
;
7476 sched_mc_power_savings
= level
;
7478 arch_reinit_sched_domains();
7483 #ifdef CONFIG_SCHED_MC
7484 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7485 struct sysdev_class_attribute
*attr
,
7488 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7490 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7491 struct sysdev_class_attribute
*attr
,
7492 const char *buf
, size_t count
)
7494 return sched_power_savings_store(buf
, count
, 0);
7496 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7497 sched_mc_power_savings_show
,
7498 sched_mc_power_savings_store
);
7501 #ifdef CONFIG_SCHED_SMT
7502 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7503 struct sysdev_class_attribute
*attr
,
7506 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7508 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7509 struct sysdev_class_attribute
*attr
,
7510 const char *buf
, size_t count
)
7512 return sched_power_savings_store(buf
, count
, 1);
7514 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7515 sched_smt_power_savings_show
,
7516 sched_smt_power_savings_store
);
7519 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7523 #ifdef CONFIG_SCHED_SMT
7525 err
= sysfs_create_file(&cls
->kset
.kobj
,
7526 &attr_sched_smt_power_savings
.attr
);
7528 #ifdef CONFIG_SCHED_MC
7529 if (!err
&& mc_capable())
7530 err
= sysfs_create_file(&cls
->kset
.kobj
,
7531 &attr_sched_mc_power_savings
.attr
);
7535 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7537 #ifndef CONFIG_CPUSETS
7539 * Add online and remove offline CPUs from the scheduler domains.
7540 * When cpusets are enabled they take over this function.
7542 static int update_sched_domains(struct notifier_block
*nfb
,
7543 unsigned long action
, void *hcpu
)
7547 case CPU_ONLINE_FROZEN
:
7548 case CPU_DOWN_PREPARE
:
7549 case CPU_DOWN_PREPARE_FROZEN
:
7550 case CPU_DOWN_FAILED
:
7551 case CPU_DOWN_FAILED_FROZEN
:
7552 partition_sched_domains(1, NULL
, NULL
);
7561 static int update_runtime(struct notifier_block
*nfb
,
7562 unsigned long action
, void *hcpu
)
7564 int cpu
= (int)(long)hcpu
;
7567 case CPU_DOWN_PREPARE
:
7568 case CPU_DOWN_PREPARE_FROZEN
:
7569 disable_runtime(cpu_rq(cpu
));
7572 case CPU_DOWN_FAILED
:
7573 case CPU_DOWN_FAILED_FROZEN
:
7575 case CPU_ONLINE_FROZEN
:
7576 enable_runtime(cpu_rq(cpu
));
7584 void __init
sched_init_smp(void)
7586 cpumask_var_t non_isolated_cpus
;
7588 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7589 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7591 #if defined(CONFIG_NUMA)
7592 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7594 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7597 mutex_lock(&sched_domains_mutex
);
7598 arch_init_sched_domains(cpu_active_mask
);
7599 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7600 if (cpumask_empty(non_isolated_cpus
))
7601 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7602 mutex_unlock(&sched_domains_mutex
);
7605 #ifndef CONFIG_CPUSETS
7606 /* XXX: Theoretical race here - CPU may be hotplugged now */
7607 hotcpu_notifier(update_sched_domains
, 0);
7610 /* RT runtime code needs to handle some hotplug events */
7611 hotcpu_notifier(update_runtime
, 0);
7615 /* Move init over to a non-isolated CPU */
7616 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7618 sched_init_granularity();
7619 free_cpumask_var(non_isolated_cpus
);
7621 init_sched_rt_class();
7624 void __init
sched_init_smp(void)
7626 sched_init_granularity();
7628 #endif /* CONFIG_SMP */
7630 const_debug
unsigned int sysctl_timer_migration
= 1;
7632 int in_sched_functions(unsigned long addr
)
7634 return in_lock_functions(addr
) ||
7635 (addr
>= (unsigned long)__sched_text_start
7636 && addr
< (unsigned long)__sched_text_end
);
7639 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7641 cfs_rq
->tasks_timeline
= RB_ROOT
;
7642 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7646 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7649 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7651 struct rt_prio_array
*array
;
7654 array
= &rt_rq
->active
;
7655 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7656 INIT_LIST_HEAD(array
->queue
+ i
);
7657 __clear_bit(i
, array
->bitmap
);
7659 /* delimiter for bitsearch: */
7660 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7662 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7663 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7665 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7669 rt_rq
->rt_nr_migratory
= 0;
7670 rt_rq
->overloaded
= 0;
7671 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7675 rt_rq
->rt_throttled
= 0;
7676 rt_rq
->rt_runtime
= 0;
7677 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7679 #ifdef CONFIG_RT_GROUP_SCHED
7680 rt_rq
->rt_nr_boosted
= 0;
7685 #ifdef CONFIG_FAIR_GROUP_SCHED
7686 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7687 struct sched_entity
*se
, int cpu
, int add
,
7688 struct sched_entity
*parent
)
7690 struct rq
*rq
= cpu_rq(cpu
);
7691 tg
->cfs_rq
[cpu
] = cfs_rq
;
7692 init_cfs_rq(cfs_rq
, rq
);
7695 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7698 /* se could be NULL for init_task_group */
7703 se
->cfs_rq
= &rq
->cfs
;
7705 se
->cfs_rq
= parent
->my_q
;
7708 se
->load
.weight
= tg
->shares
;
7709 se
->load
.inv_weight
= 0;
7710 se
->parent
= parent
;
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7716 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7717 struct sched_rt_entity
*parent
)
7719 struct rq
*rq
= cpu_rq(cpu
);
7721 tg
->rt_rq
[cpu
] = rt_rq
;
7722 init_rt_rq(rt_rq
, rq
);
7724 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7726 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7728 tg
->rt_se
[cpu
] = rt_se
;
7733 rt_se
->rt_rq
= &rq
->rt
;
7735 rt_se
->rt_rq
= parent
->my_q
;
7737 rt_se
->my_q
= rt_rq
;
7738 rt_se
->parent
= parent
;
7739 INIT_LIST_HEAD(&rt_se
->run_list
);
7743 void __init
sched_init(void)
7746 unsigned long alloc_size
= 0, ptr
;
7748 #ifdef CONFIG_FAIR_GROUP_SCHED
7749 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7751 #ifdef CONFIG_RT_GROUP_SCHED
7752 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7754 #ifdef CONFIG_CPUMASK_OFFSTACK
7755 alloc_size
+= num_possible_cpus() * cpumask_size();
7758 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7760 #ifdef CONFIG_FAIR_GROUP_SCHED
7761 init_task_group
.se
= (struct sched_entity
**)ptr
;
7762 ptr
+= nr_cpu_ids
* sizeof(void **);
7764 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7765 ptr
+= nr_cpu_ids
* sizeof(void **);
7767 #endif /* CONFIG_FAIR_GROUP_SCHED */
7768 #ifdef CONFIG_RT_GROUP_SCHED
7769 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7770 ptr
+= nr_cpu_ids
* sizeof(void **);
7772 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7773 ptr
+= nr_cpu_ids
* sizeof(void **);
7775 #endif /* CONFIG_RT_GROUP_SCHED */
7776 #ifdef CONFIG_CPUMASK_OFFSTACK
7777 for_each_possible_cpu(i
) {
7778 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7779 ptr
+= cpumask_size();
7781 #endif /* CONFIG_CPUMASK_OFFSTACK */
7785 init_defrootdomain();
7788 init_rt_bandwidth(&def_rt_bandwidth
,
7789 global_rt_period(), global_rt_runtime());
7791 #ifdef CONFIG_RT_GROUP_SCHED
7792 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7793 global_rt_period(), global_rt_runtime());
7794 #endif /* CONFIG_RT_GROUP_SCHED */
7796 #ifdef CONFIG_CGROUP_SCHED
7797 list_add(&init_task_group
.list
, &task_groups
);
7798 INIT_LIST_HEAD(&init_task_group
.children
);
7800 #endif /* CONFIG_CGROUP_SCHED */
7802 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7803 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7804 __alignof__(unsigned long));
7806 for_each_possible_cpu(i
) {
7810 raw_spin_lock_init(&rq
->lock
);
7812 rq
->calc_load_active
= 0;
7813 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7814 init_cfs_rq(&rq
->cfs
, rq
);
7815 init_rt_rq(&rq
->rt
, rq
);
7816 #ifdef CONFIG_FAIR_GROUP_SCHED
7817 init_task_group
.shares
= init_task_group_load
;
7818 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7819 #ifdef CONFIG_CGROUP_SCHED
7821 * How much cpu bandwidth does init_task_group get?
7823 * In case of task-groups formed thr' the cgroup filesystem, it
7824 * gets 100% of the cpu resources in the system. This overall
7825 * system cpu resource is divided among the tasks of
7826 * init_task_group and its child task-groups in a fair manner,
7827 * based on each entity's (task or task-group's) weight
7828 * (se->load.weight).
7830 * In other words, if init_task_group has 10 tasks of weight
7831 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7832 * then A0's share of the cpu resource is:
7834 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7836 * We achieve this by letting init_task_group's tasks sit
7837 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7839 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7841 #endif /* CONFIG_FAIR_GROUP_SCHED */
7843 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7844 #ifdef CONFIG_RT_GROUP_SCHED
7845 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7846 #ifdef CONFIG_CGROUP_SCHED
7847 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7851 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7852 rq
->cpu_load
[j
] = 0;
7856 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7857 rq
->post_schedule
= 0;
7858 rq
->active_balance
= 0;
7859 rq
->next_balance
= jiffies
;
7864 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7865 rq_attach_root(rq
, &def_root_domain
);
7868 atomic_set(&rq
->nr_iowait
, 0);
7871 set_load_weight(&init_task
);
7873 #ifdef CONFIG_PREEMPT_NOTIFIERS
7874 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7878 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7881 #ifdef CONFIG_RT_MUTEXES
7882 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7886 * The boot idle thread does lazy MMU switching as well:
7888 atomic_inc(&init_mm
.mm_count
);
7889 enter_lazy_tlb(&init_mm
, current
);
7892 * Make us the idle thread. Technically, schedule() should not be
7893 * called from this thread, however somewhere below it might be,
7894 * but because we are the idle thread, we just pick up running again
7895 * when this runqueue becomes "idle".
7897 init_idle(current
, smp_processor_id());
7899 calc_load_update
= jiffies
+ LOAD_FREQ
;
7902 * During early bootup we pretend to be a normal task:
7904 current
->sched_class
= &fair_sched_class
;
7906 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7907 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7910 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7911 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7913 /* May be allocated at isolcpus cmdline parse time */
7914 if (cpu_isolated_map
== NULL
)
7915 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7920 scheduler_running
= 1;
7923 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7924 static inline int preempt_count_equals(int preempt_offset
)
7926 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7928 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7931 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7934 static unsigned long prev_jiffy
; /* ratelimiting */
7936 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7937 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7939 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7941 prev_jiffy
= jiffies
;
7944 "BUG: sleeping function called from invalid context at %s:%d\n",
7947 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7948 in_atomic(), irqs_disabled(),
7949 current
->pid
, current
->comm
);
7951 debug_show_held_locks(current
);
7952 if (irqs_disabled())
7953 print_irqtrace_events(current
);
7957 EXPORT_SYMBOL(__might_sleep
);
7960 #ifdef CONFIG_MAGIC_SYSRQ
7961 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7965 on_rq
= p
->se
.on_rq
;
7967 deactivate_task(rq
, p
, 0);
7968 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7970 activate_task(rq
, p
, 0);
7971 resched_task(rq
->curr
);
7975 void normalize_rt_tasks(void)
7977 struct task_struct
*g
, *p
;
7978 unsigned long flags
;
7981 read_lock_irqsave(&tasklist_lock
, flags
);
7982 do_each_thread(g
, p
) {
7984 * Only normalize user tasks:
7989 p
->se
.exec_start
= 0;
7990 #ifdef CONFIG_SCHEDSTATS
7991 p
->se
.statistics
.wait_start
= 0;
7992 p
->se
.statistics
.sleep_start
= 0;
7993 p
->se
.statistics
.block_start
= 0;
7998 * Renice negative nice level userspace
8001 if (TASK_NICE(p
) < 0 && p
->mm
)
8002 set_user_nice(p
, 0);
8006 raw_spin_lock(&p
->pi_lock
);
8007 rq
= __task_rq_lock(p
);
8009 normalize_task(rq
, p
);
8011 __task_rq_unlock(rq
);
8012 raw_spin_unlock(&p
->pi_lock
);
8013 } while_each_thread(g
, p
);
8015 read_unlock_irqrestore(&tasklist_lock
, flags
);
8018 #endif /* CONFIG_MAGIC_SYSRQ */
8020 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8022 * These functions are only useful for the IA64 MCA handling, or kdb.
8024 * They can only be called when the whole system has been
8025 * stopped - every CPU needs to be quiescent, and no scheduling
8026 * activity can take place. Using them for anything else would
8027 * be a serious bug, and as a result, they aren't even visible
8028 * under any other configuration.
8032 * curr_task - return the current task for a given cpu.
8033 * @cpu: the processor in question.
8035 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8037 struct task_struct
*curr_task(int cpu
)
8039 return cpu_curr(cpu
);
8042 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8046 * set_curr_task - set the current task for a given cpu.
8047 * @cpu: the processor in question.
8048 * @p: the task pointer to set.
8050 * Description: This function must only be used when non-maskable interrupts
8051 * are serviced on a separate stack. It allows the architecture to switch the
8052 * notion of the current task on a cpu in a non-blocking manner. This function
8053 * must be called with all CPU's synchronized, and interrupts disabled, the
8054 * and caller must save the original value of the current task (see
8055 * curr_task() above) and restore that value before reenabling interrupts and
8056 * re-starting the system.
8058 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8060 void set_curr_task(int cpu
, struct task_struct
*p
)
8067 #ifdef CONFIG_FAIR_GROUP_SCHED
8068 static void free_fair_sched_group(struct task_group
*tg
)
8072 for_each_possible_cpu(i
) {
8074 kfree(tg
->cfs_rq
[i
]);
8084 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8086 struct cfs_rq
*cfs_rq
;
8087 struct sched_entity
*se
;
8091 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8094 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8098 tg
->shares
= NICE_0_LOAD
;
8100 for_each_possible_cpu(i
) {
8103 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8104 GFP_KERNEL
, cpu_to_node(i
));
8108 se
= kzalloc_node(sizeof(struct sched_entity
),
8109 GFP_KERNEL
, cpu_to_node(i
));
8113 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8124 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8126 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8127 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8130 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8132 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8134 #else /* !CONFG_FAIR_GROUP_SCHED */
8135 static inline void free_fair_sched_group(struct task_group
*tg
)
8140 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8145 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8149 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8152 #endif /* CONFIG_FAIR_GROUP_SCHED */
8154 #ifdef CONFIG_RT_GROUP_SCHED
8155 static void free_rt_sched_group(struct task_group
*tg
)
8159 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8161 for_each_possible_cpu(i
) {
8163 kfree(tg
->rt_rq
[i
]);
8165 kfree(tg
->rt_se
[i
]);
8173 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8175 struct rt_rq
*rt_rq
;
8176 struct sched_rt_entity
*rt_se
;
8180 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8183 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8187 init_rt_bandwidth(&tg
->rt_bandwidth
,
8188 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8190 for_each_possible_cpu(i
) {
8193 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8194 GFP_KERNEL
, cpu_to_node(i
));
8198 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8199 GFP_KERNEL
, cpu_to_node(i
));
8203 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8214 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8216 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8217 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8220 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8222 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8224 #else /* !CONFIG_RT_GROUP_SCHED */
8225 static inline void free_rt_sched_group(struct task_group
*tg
)
8230 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8235 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8239 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8242 #endif /* CONFIG_RT_GROUP_SCHED */
8244 #ifdef CONFIG_CGROUP_SCHED
8245 static void free_sched_group(struct task_group
*tg
)
8247 free_fair_sched_group(tg
);
8248 free_rt_sched_group(tg
);
8252 /* allocate runqueue etc for a new task group */
8253 struct task_group
*sched_create_group(struct task_group
*parent
)
8255 struct task_group
*tg
;
8256 unsigned long flags
;
8259 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8261 return ERR_PTR(-ENOMEM
);
8263 if (!alloc_fair_sched_group(tg
, parent
))
8266 if (!alloc_rt_sched_group(tg
, parent
))
8269 spin_lock_irqsave(&task_group_lock
, flags
);
8270 for_each_possible_cpu(i
) {
8271 register_fair_sched_group(tg
, i
);
8272 register_rt_sched_group(tg
, i
);
8274 list_add_rcu(&tg
->list
, &task_groups
);
8276 WARN_ON(!parent
); /* root should already exist */
8278 tg
->parent
= parent
;
8279 INIT_LIST_HEAD(&tg
->children
);
8280 list_add_rcu(&tg
->siblings
, &parent
->children
);
8281 spin_unlock_irqrestore(&task_group_lock
, flags
);
8286 free_sched_group(tg
);
8287 return ERR_PTR(-ENOMEM
);
8290 /* rcu callback to free various structures associated with a task group */
8291 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8293 /* now it should be safe to free those cfs_rqs */
8294 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8297 /* Destroy runqueue etc associated with a task group */
8298 void sched_destroy_group(struct task_group
*tg
)
8300 unsigned long flags
;
8303 spin_lock_irqsave(&task_group_lock
, flags
);
8304 for_each_possible_cpu(i
) {
8305 unregister_fair_sched_group(tg
, i
);
8306 unregister_rt_sched_group(tg
, i
);
8308 list_del_rcu(&tg
->list
);
8309 list_del_rcu(&tg
->siblings
);
8310 spin_unlock_irqrestore(&task_group_lock
, flags
);
8312 /* wait for possible concurrent references to cfs_rqs complete */
8313 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8316 /* change task's runqueue when it moves between groups.
8317 * The caller of this function should have put the task in its new group
8318 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8319 * reflect its new group.
8321 void sched_move_task(struct task_struct
*tsk
)
8324 unsigned long flags
;
8327 rq
= task_rq_lock(tsk
, &flags
);
8329 running
= task_current(rq
, tsk
);
8330 on_rq
= tsk
->se
.on_rq
;
8333 dequeue_task(rq
, tsk
, 0);
8334 if (unlikely(running
))
8335 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8337 #ifdef CONFIG_FAIR_GROUP_SCHED
8338 if (tsk
->sched_class
->task_move_group
)
8339 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8342 set_task_rq(tsk
, task_cpu(tsk
));
8344 if (unlikely(running
))
8345 tsk
->sched_class
->set_curr_task(rq
);
8347 enqueue_task(rq
, tsk
, 0);
8349 task_rq_unlock(rq
, &flags
);
8351 #endif /* CONFIG_CGROUP_SCHED */
8353 #ifdef CONFIG_FAIR_GROUP_SCHED
8354 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8356 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8361 dequeue_entity(cfs_rq
, se
, 0);
8363 se
->load
.weight
= shares
;
8364 se
->load
.inv_weight
= 0;
8367 enqueue_entity(cfs_rq
, se
, 0);
8370 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8372 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8373 struct rq
*rq
= cfs_rq
->rq
;
8374 unsigned long flags
;
8376 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8377 __set_se_shares(se
, shares
);
8378 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8381 static DEFINE_MUTEX(shares_mutex
);
8383 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8386 unsigned long flags
;
8389 * We can't change the weight of the root cgroup.
8394 if (shares
< MIN_SHARES
)
8395 shares
= MIN_SHARES
;
8396 else if (shares
> MAX_SHARES
)
8397 shares
= MAX_SHARES
;
8399 mutex_lock(&shares_mutex
);
8400 if (tg
->shares
== shares
)
8403 spin_lock_irqsave(&task_group_lock
, flags
);
8404 for_each_possible_cpu(i
)
8405 unregister_fair_sched_group(tg
, i
);
8406 list_del_rcu(&tg
->siblings
);
8407 spin_unlock_irqrestore(&task_group_lock
, flags
);
8409 /* wait for any ongoing reference to this group to finish */
8410 synchronize_sched();
8413 * Now we are free to modify the group's share on each cpu
8414 * w/o tripping rebalance_share or load_balance_fair.
8416 tg
->shares
= shares
;
8417 for_each_possible_cpu(i
) {
8421 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8422 set_se_shares(tg
->se
[i
], shares
);
8426 * Enable load balance activity on this group, by inserting it back on
8427 * each cpu's rq->leaf_cfs_rq_list.
8429 spin_lock_irqsave(&task_group_lock
, flags
);
8430 for_each_possible_cpu(i
)
8431 register_fair_sched_group(tg
, i
);
8432 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8433 spin_unlock_irqrestore(&task_group_lock
, flags
);
8435 mutex_unlock(&shares_mutex
);
8439 unsigned long sched_group_shares(struct task_group
*tg
)
8445 #ifdef CONFIG_RT_GROUP_SCHED
8447 * Ensure that the real time constraints are schedulable.
8449 static DEFINE_MUTEX(rt_constraints_mutex
);
8451 static unsigned long to_ratio(u64 period
, u64 runtime
)
8453 if (runtime
== RUNTIME_INF
)
8456 return div64_u64(runtime
<< 20, period
);
8459 /* Must be called with tasklist_lock held */
8460 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8462 struct task_struct
*g
, *p
;
8464 do_each_thread(g
, p
) {
8465 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8467 } while_each_thread(g
, p
);
8472 struct rt_schedulable_data
{
8473 struct task_group
*tg
;
8478 static int tg_schedulable(struct task_group
*tg
, void *data
)
8480 struct rt_schedulable_data
*d
= data
;
8481 struct task_group
*child
;
8482 unsigned long total
, sum
= 0;
8483 u64 period
, runtime
;
8485 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8486 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8489 period
= d
->rt_period
;
8490 runtime
= d
->rt_runtime
;
8494 * Cannot have more runtime than the period.
8496 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8500 * Ensure we don't starve existing RT tasks.
8502 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8505 total
= to_ratio(period
, runtime
);
8508 * Nobody can have more than the global setting allows.
8510 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8514 * The sum of our children's runtime should not exceed our own.
8516 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8517 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8518 runtime
= child
->rt_bandwidth
.rt_runtime
;
8520 if (child
== d
->tg
) {
8521 period
= d
->rt_period
;
8522 runtime
= d
->rt_runtime
;
8525 sum
+= to_ratio(period
, runtime
);
8534 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8536 struct rt_schedulable_data data
= {
8538 .rt_period
= period
,
8539 .rt_runtime
= runtime
,
8542 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8545 static int tg_set_bandwidth(struct task_group
*tg
,
8546 u64 rt_period
, u64 rt_runtime
)
8550 mutex_lock(&rt_constraints_mutex
);
8551 read_lock(&tasklist_lock
);
8552 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8556 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8557 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8558 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8560 for_each_possible_cpu(i
) {
8561 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8563 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8564 rt_rq
->rt_runtime
= rt_runtime
;
8565 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8567 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8569 read_unlock(&tasklist_lock
);
8570 mutex_unlock(&rt_constraints_mutex
);
8575 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8577 u64 rt_runtime
, rt_period
;
8579 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8580 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8581 if (rt_runtime_us
< 0)
8582 rt_runtime
= RUNTIME_INF
;
8584 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8587 long sched_group_rt_runtime(struct task_group
*tg
)
8591 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8594 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8595 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8596 return rt_runtime_us
;
8599 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8601 u64 rt_runtime
, rt_period
;
8603 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8604 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8609 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8612 long sched_group_rt_period(struct task_group
*tg
)
8616 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8617 do_div(rt_period_us
, NSEC_PER_USEC
);
8618 return rt_period_us
;
8621 static int sched_rt_global_constraints(void)
8623 u64 runtime
, period
;
8626 if (sysctl_sched_rt_period
<= 0)
8629 runtime
= global_rt_runtime();
8630 period
= global_rt_period();
8633 * Sanity check on the sysctl variables.
8635 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8638 mutex_lock(&rt_constraints_mutex
);
8639 read_lock(&tasklist_lock
);
8640 ret
= __rt_schedulable(NULL
, 0, 0);
8641 read_unlock(&tasklist_lock
);
8642 mutex_unlock(&rt_constraints_mutex
);
8647 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8649 /* Don't accept realtime tasks when there is no way for them to run */
8650 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8656 #else /* !CONFIG_RT_GROUP_SCHED */
8657 static int sched_rt_global_constraints(void)
8659 unsigned long flags
;
8662 if (sysctl_sched_rt_period
<= 0)
8666 * There's always some RT tasks in the root group
8667 * -- migration, kstopmachine etc..
8669 if (sysctl_sched_rt_runtime
== 0)
8672 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8673 for_each_possible_cpu(i
) {
8674 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8676 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8677 rt_rq
->rt_runtime
= global_rt_runtime();
8678 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8680 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8684 #endif /* CONFIG_RT_GROUP_SCHED */
8686 int sched_rt_handler(struct ctl_table
*table
, int write
,
8687 void __user
*buffer
, size_t *lenp
,
8691 int old_period
, old_runtime
;
8692 static DEFINE_MUTEX(mutex
);
8695 old_period
= sysctl_sched_rt_period
;
8696 old_runtime
= sysctl_sched_rt_runtime
;
8698 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8700 if (!ret
&& write
) {
8701 ret
= sched_rt_global_constraints();
8703 sysctl_sched_rt_period
= old_period
;
8704 sysctl_sched_rt_runtime
= old_runtime
;
8706 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8707 def_rt_bandwidth
.rt_period
=
8708 ns_to_ktime(global_rt_period());
8711 mutex_unlock(&mutex
);
8716 #ifdef CONFIG_CGROUP_SCHED
8718 /* return corresponding task_group object of a cgroup */
8719 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8721 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8722 struct task_group
, css
);
8725 static struct cgroup_subsys_state
*
8726 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8728 struct task_group
*tg
, *parent
;
8730 if (!cgrp
->parent
) {
8731 /* This is early initialization for the top cgroup */
8732 return &init_task_group
.css
;
8735 parent
= cgroup_tg(cgrp
->parent
);
8736 tg
= sched_create_group(parent
);
8738 return ERR_PTR(-ENOMEM
);
8744 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8746 struct task_group
*tg
= cgroup_tg(cgrp
);
8748 sched_destroy_group(tg
);
8752 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8754 #ifdef CONFIG_RT_GROUP_SCHED
8755 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8758 /* We don't support RT-tasks being in separate groups */
8759 if (tsk
->sched_class
!= &fair_sched_class
)
8766 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8767 struct task_struct
*tsk
, bool threadgroup
)
8769 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8773 struct task_struct
*c
;
8775 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8776 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8788 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8789 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8792 sched_move_task(tsk
);
8794 struct task_struct
*c
;
8796 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8803 #ifdef CONFIG_FAIR_GROUP_SCHED
8804 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8807 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8810 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8812 struct task_group
*tg
= cgroup_tg(cgrp
);
8814 return (u64
) tg
->shares
;
8816 #endif /* CONFIG_FAIR_GROUP_SCHED */
8818 #ifdef CONFIG_RT_GROUP_SCHED
8819 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8822 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8825 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8827 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8830 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8833 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8836 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8838 return sched_group_rt_period(cgroup_tg(cgrp
));
8840 #endif /* CONFIG_RT_GROUP_SCHED */
8842 static struct cftype cpu_files
[] = {
8843 #ifdef CONFIG_FAIR_GROUP_SCHED
8846 .read_u64
= cpu_shares_read_u64
,
8847 .write_u64
= cpu_shares_write_u64
,
8850 #ifdef CONFIG_RT_GROUP_SCHED
8852 .name
= "rt_runtime_us",
8853 .read_s64
= cpu_rt_runtime_read
,
8854 .write_s64
= cpu_rt_runtime_write
,
8857 .name
= "rt_period_us",
8858 .read_u64
= cpu_rt_period_read_uint
,
8859 .write_u64
= cpu_rt_period_write_uint
,
8864 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8866 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8869 struct cgroup_subsys cpu_cgroup_subsys
= {
8871 .create
= cpu_cgroup_create
,
8872 .destroy
= cpu_cgroup_destroy
,
8873 .can_attach
= cpu_cgroup_can_attach
,
8874 .attach
= cpu_cgroup_attach
,
8875 .populate
= cpu_cgroup_populate
,
8876 .subsys_id
= cpu_cgroup_subsys_id
,
8880 #endif /* CONFIG_CGROUP_SCHED */
8882 #ifdef CONFIG_CGROUP_CPUACCT
8885 * CPU accounting code for task groups.
8887 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8888 * (balbir@in.ibm.com).
8891 /* track cpu usage of a group of tasks and its child groups */
8893 struct cgroup_subsys_state css
;
8894 /* cpuusage holds pointer to a u64-type object on every cpu */
8895 u64 __percpu
*cpuusage
;
8896 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8897 struct cpuacct
*parent
;
8900 struct cgroup_subsys cpuacct_subsys
;
8902 /* return cpu accounting group corresponding to this container */
8903 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8905 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8906 struct cpuacct
, css
);
8909 /* return cpu accounting group to which this task belongs */
8910 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8912 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8913 struct cpuacct
, css
);
8916 /* create a new cpu accounting group */
8917 static struct cgroup_subsys_state
*cpuacct_create(
8918 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8920 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8926 ca
->cpuusage
= alloc_percpu(u64
);
8930 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8931 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8932 goto out_free_counters
;
8935 ca
->parent
= cgroup_ca(cgrp
->parent
);
8941 percpu_counter_destroy(&ca
->cpustat
[i
]);
8942 free_percpu(ca
->cpuusage
);
8946 return ERR_PTR(-ENOMEM
);
8949 /* destroy an existing cpu accounting group */
8951 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8953 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8956 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8957 percpu_counter_destroy(&ca
->cpustat
[i
]);
8958 free_percpu(ca
->cpuusage
);
8962 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8964 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8967 #ifndef CONFIG_64BIT
8969 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8971 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8973 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8981 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8983 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8985 #ifndef CONFIG_64BIT
8987 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8989 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8991 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8997 /* return total cpu usage (in nanoseconds) of a group */
8998 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9000 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9001 u64 totalcpuusage
= 0;
9004 for_each_present_cpu(i
)
9005 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9007 return totalcpuusage
;
9010 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9013 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9022 for_each_present_cpu(i
)
9023 cpuacct_cpuusage_write(ca
, i
, 0);
9029 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9032 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9036 for_each_present_cpu(i
) {
9037 percpu
= cpuacct_cpuusage_read(ca
, i
);
9038 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9040 seq_printf(m
, "\n");
9044 static const char *cpuacct_stat_desc
[] = {
9045 [CPUACCT_STAT_USER
] = "user",
9046 [CPUACCT_STAT_SYSTEM
] = "system",
9049 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9050 struct cgroup_map_cb
*cb
)
9052 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9055 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9056 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9057 val
= cputime64_to_clock_t(val
);
9058 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9063 static struct cftype files
[] = {
9066 .read_u64
= cpuusage_read
,
9067 .write_u64
= cpuusage_write
,
9070 .name
= "usage_percpu",
9071 .read_seq_string
= cpuacct_percpu_seq_read
,
9075 .read_map
= cpuacct_stats_show
,
9079 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9081 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9085 * charge this task's execution time to its accounting group.
9087 * called with rq->lock held.
9089 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9094 if (unlikely(!cpuacct_subsys
.active
))
9097 cpu
= task_cpu(tsk
);
9103 for (; ca
; ca
= ca
->parent
) {
9104 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9105 *cpuusage
+= cputime
;
9112 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9113 * in cputime_t units. As a result, cpuacct_update_stats calls
9114 * percpu_counter_add with values large enough to always overflow the
9115 * per cpu batch limit causing bad SMP scalability.
9117 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9118 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9119 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9122 #define CPUACCT_BATCH \
9123 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9125 #define CPUACCT_BATCH 0
9129 * Charge the system/user time to the task's accounting group.
9131 static void cpuacct_update_stats(struct task_struct
*tsk
,
9132 enum cpuacct_stat_index idx
, cputime_t val
)
9135 int batch
= CPUACCT_BATCH
;
9137 if (unlikely(!cpuacct_subsys
.active
))
9144 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9150 struct cgroup_subsys cpuacct_subsys
= {
9152 .create
= cpuacct_create
,
9153 .destroy
= cpuacct_destroy
,
9154 .populate
= cpuacct_populate
,
9155 .subsys_id
= cpuacct_subsys_id
,
9157 #endif /* CONFIG_CGROUP_CPUACCT */
9161 void synchronize_sched_expedited(void)
9165 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9167 #else /* #ifndef CONFIG_SMP */
9169 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9171 static int synchronize_sched_expedited_cpu_stop(void *data
)
9174 * There must be a full memory barrier on each affected CPU
9175 * between the time that try_stop_cpus() is called and the
9176 * time that it returns.
9178 * In the current initial implementation of cpu_stop, the
9179 * above condition is already met when the control reaches
9180 * this point and the following smp_mb() is not strictly
9181 * necessary. Do smp_mb() anyway for documentation and
9182 * robustness against future implementation changes.
9184 smp_mb(); /* See above comment block. */
9189 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9190 * approach to force grace period to end quickly. This consumes
9191 * significant time on all CPUs, and is thus not recommended for
9192 * any sort of common-case code.
9194 * Note that it is illegal to call this function while holding any
9195 * lock that is acquired by a CPU-hotplug notifier. Failing to
9196 * observe this restriction will result in deadlock.
9198 void synchronize_sched_expedited(void)
9200 int snap
, trycount
= 0;
9202 smp_mb(); /* ensure prior mod happens before capturing snap. */
9203 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9205 while (try_stop_cpus(cpu_online_mask
,
9206 synchronize_sched_expedited_cpu_stop
,
9209 if (trycount
++ < 10)
9210 udelay(trycount
* num_online_cpus());
9212 synchronize_sched();
9215 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9216 smp_mb(); /* ensure test happens before caller kfree */
9221 atomic_inc(&synchronize_sched_expedited_count
);
9222 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9225 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9227 #endif /* #else #ifndef CONFIG_SMP */