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 /* return group to which a task belongs */
310 static inline struct task_group
*task_group(struct task_struct
*p
)
312 struct task_group
*tg
;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
316 struct task_group
, css
);
318 tg
= &init_task_group
;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
327 * Strictly speaking this rcu_read_lock() is not needed since the
328 * task_group is tied to the cgroup, which in turn can never go away
329 * as long as there are tasks attached to it.
331 * However since task_group() uses task_subsys_state() which is an
332 * rcu_dereference() user, this quiets CONFIG_PROVE_RCU.
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
337 p
->se
.parent
= task_group(p
)->se
[cpu
];
340 #ifdef CONFIG_RT_GROUP_SCHED
341 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
342 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
349 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
350 static inline struct task_group
*task_group(struct task_struct
*p
)
355 #endif /* CONFIG_CGROUP_SCHED */
357 /* CFS-related fields in a runqueue */
359 struct load_weight load
;
360 unsigned long nr_running
;
365 struct rb_root tasks_timeline
;
366 struct rb_node
*rb_leftmost
;
368 struct list_head tasks
;
369 struct list_head
*balance_iterator
;
372 * 'curr' points to currently running entity on this cfs_rq.
373 * It is set to NULL otherwise (i.e when none are currently running).
375 struct sched_entity
*curr
, *next
, *last
;
377 unsigned int nr_spread_over
;
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
385 * (like users, containers etc.)
387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
388 * list is used during load balance.
390 struct list_head leaf_cfs_rq_list
;
391 struct task_group
*tg
; /* group that "owns" this runqueue */
395 * the part of load.weight contributed by tasks
397 unsigned long task_weight
;
400 * h_load = weight * f(tg)
402 * Where f(tg) is the recursive weight fraction assigned to
405 unsigned long h_load
;
408 * this cpu's part of tg->shares
410 unsigned long shares
;
413 * load.weight at the time we set shares
415 unsigned long rq_weight
;
420 /* Real-Time classes' related field in a runqueue: */
422 struct rt_prio_array active
;
423 unsigned long rt_nr_running
;
424 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
426 int curr
; /* highest queued rt task prio */
428 int next
; /* next highest */
433 unsigned long rt_nr_migratory
;
434 unsigned long rt_nr_total
;
436 struct plist_head pushable_tasks
;
441 /* Nests inside the rq lock: */
442 raw_spinlock_t rt_runtime_lock
;
444 #ifdef CONFIG_RT_GROUP_SCHED
445 unsigned long rt_nr_boosted
;
448 struct list_head leaf_rt_rq_list
;
449 struct task_group
*tg
;
456 * We add the notion of a root-domain which will be used to define per-domain
457 * variables. Each exclusive cpuset essentially defines an island domain by
458 * fully partitioning the member cpus from any other cpuset. Whenever a new
459 * exclusive cpuset is created, we also create and attach a new root-domain
466 cpumask_var_t online
;
469 * The "RT overload" flag: it gets set if a CPU has more than
470 * one runnable RT task.
472 cpumask_var_t rto_mask
;
475 struct cpupri cpupri
;
480 * By default the system creates a single root-domain with all cpus as
481 * members (mimicking the global state we have today).
483 static struct root_domain def_root_domain
;
488 * This is the main, per-CPU runqueue data structure.
490 * Locking rule: those places that want to lock multiple runqueues
491 * (such as the load balancing or the thread migration code), lock
492 * acquire operations must be ordered by ascending &runqueue.
499 * nr_running and cpu_load should be in the same cacheline because
500 * remote CPUs use both these fields when doing load calculation.
502 unsigned long nr_running
;
503 #define CPU_LOAD_IDX_MAX 5
504 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
507 unsigned char in_nohz_recently
;
509 unsigned int skip_clock_update
;
511 /* capture load from *all* tasks on this cpu: */
512 struct load_weight load
;
513 unsigned long nr_load_updates
;
519 #ifdef CONFIG_FAIR_GROUP_SCHED
520 /* list of leaf cfs_rq on this cpu: */
521 struct list_head leaf_cfs_rq_list
;
523 #ifdef CONFIG_RT_GROUP_SCHED
524 struct list_head leaf_rt_rq_list
;
528 * This is part of a global counter where only the total sum
529 * over all CPUs matters. A task can increase this counter on
530 * one CPU and if it got migrated afterwards it may decrease
531 * it on another CPU. Always updated under the runqueue lock:
533 unsigned long nr_uninterruptible
;
535 struct task_struct
*curr
, *idle
;
536 unsigned long next_balance
;
537 struct mm_struct
*prev_mm
;
544 struct root_domain
*rd
;
545 struct sched_domain
*sd
;
547 unsigned long cpu_power
;
549 unsigned char idle_at_tick
;
550 /* For active balancing */
554 struct cpu_stop_work active_balance_work
;
555 /* cpu of this runqueue: */
559 unsigned long avg_load_per_task
;
567 /* calc_load related fields */
568 unsigned long calc_load_update
;
569 long calc_load_active
;
571 #ifdef CONFIG_SCHED_HRTICK
573 int hrtick_csd_pending
;
574 struct call_single_data hrtick_csd
;
576 struct hrtimer hrtick_timer
;
579 #ifdef CONFIG_SCHEDSTATS
581 struct sched_info rq_sched_info
;
582 unsigned long long rq_cpu_time
;
583 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
585 /* sys_sched_yield() stats */
586 unsigned int yld_count
;
588 /* schedule() stats */
589 unsigned int sched_switch
;
590 unsigned int sched_count
;
591 unsigned int sched_goidle
;
593 /* try_to_wake_up() stats */
594 unsigned int ttwu_count
;
595 unsigned int ttwu_local
;
598 unsigned int bkl_count
;
602 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
605 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
607 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
610 * A queue event has occurred, and we're going to schedule. In
611 * this case, we can save a useless back to back clock update.
613 if (test_tsk_need_resched(p
))
614 rq
->skip_clock_update
= 1;
617 static inline int cpu_of(struct rq
*rq
)
626 #define rcu_dereference_check_sched_domain(p) \
627 rcu_dereference_check((p), \
628 rcu_read_lock_sched_held() || \
629 lockdep_is_held(&sched_domains_mutex))
632 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
633 * See detach_destroy_domains: synchronize_sched for details.
635 * The domain tree of any CPU may only be accessed from within
636 * preempt-disabled sections.
638 #define for_each_domain(cpu, __sd) \
639 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
641 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
642 #define this_rq() (&__get_cpu_var(runqueues))
643 #define task_rq(p) cpu_rq(task_cpu(p))
644 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
645 #define raw_rq() (&__raw_get_cpu_var(runqueues))
647 inline void update_rq_clock(struct rq
*rq
)
649 if (!rq
->skip_clock_update
)
650 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
654 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
656 #ifdef CONFIG_SCHED_DEBUG
657 # define const_debug __read_mostly
659 # define const_debug static const
664 * @cpu: the processor in question.
666 * Returns true if the current cpu runqueue is locked.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(int cpu
)
672 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug
unsigned int sysctl_sched_features
=
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly
char *sched_feat_names
[] = {
702 #include "sched_features.h"
708 static int sched_feat_show(struct seq_file
*m
, void *v
)
712 for (i
= 0; sched_feat_names
[i
]; i
++) {
713 if (!(sysctl_sched_features
& (1UL << i
)))
715 seq_printf(m
, "%s ", sched_feat_names
[i
]);
723 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
724 size_t cnt
, loff_t
*ppos
)
734 if (copy_from_user(&buf
, ubuf
, cnt
))
739 if (strncmp(buf
, "NO_", 3) == 0) {
744 for (i
= 0; sched_feat_names
[i
]; i
++) {
745 int len
= strlen(sched_feat_names
[i
]);
747 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
749 sysctl_sched_features
&= ~(1UL << i
);
751 sysctl_sched_features
|= (1UL << i
);
756 if (!sched_feat_names
[i
])
764 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
766 return single_open(filp
, sched_feat_show
, NULL
);
769 static const struct file_operations sched_feat_fops
= {
770 .open
= sched_feat_open
,
771 .write
= sched_feat_write
,
774 .release
= single_release
,
777 static __init
int sched_init_debug(void)
779 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
784 late_initcall(sched_init_debug
);
788 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
791 * Number of tasks to iterate in a single balance run.
792 * Limited because this is done with IRQs disabled.
794 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
797 * ratelimit for updating the group shares.
800 unsigned int sysctl_sched_shares_ratelimit
= 250000;
801 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
804 * Inject some fuzzyness into changing the per-cpu group shares
805 * this avoids remote rq-locks at the expense of fairness.
808 unsigned int sysctl_sched_shares_thresh
= 4;
811 * period over which we average the RT time consumption, measured
816 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
819 * period over which we measure -rt task cpu usage in us.
822 unsigned int sysctl_sched_rt_period
= 1000000;
824 static __read_mostly
int scheduler_running
;
827 * part of the period that we allow rt tasks to run in us.
830 int sysctl_sched_rt_runtime
= 950000;
832 static inline u64
global_rt_period(void)
834 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
837 static inline u64
global_rt_runtime(void)
839 if (sysctl_sched_rt_runtime
< 0)
842 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
845 #ifndef prepare_arch_switch
846 # define prepare_arch_switch(next) do { } while (0)
848 #ifndef finish_arch_switch
849 # define finish_arch_switch(prev) do { } while (0)
852 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
854 return rq
->curr
== p
;
857 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
858 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
860 return task_current(rq
, p
);
863 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
867 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
869 #ifdef CONFIG_DEBUG_SPINLOCK
870 /* this is a valid case when another task releases the spinlock */
871 rq
->lock
.owner
= current
;
874 * If we are tracking spinlock dependencies then we have to
875 * fix up the runqueue lock - which gets 'carried over' from
878 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
880 raw_spin_unlock_irq(&rq
->lock
);
883 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
884 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
889 return task_current(rq
, p
);
893 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
897 * We can optimise this out completely for !SMP, because the
898 * SMP rebalancing from interrupt is the only thing that cares
903 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
904 raw_spin_unlock_irq(&rq
->lock
);
906 raw_spin_unlock(&rq
->lock
);
910 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
914 * After ->oncpu is cleared, the task can be moved to a different CPU.
915 * We must ensure this doesn't happen until the switch is completely
921 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
925 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
928 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
931 static inline int task_is_waking(struct task_struct
*p
)
933 return unlikely(p
->state
== TASK_WAKING
);
937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
940 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
947 raw_spin_lock(&rq
->lock
);
948 if (likely(rq
== task_rq(p
)))
950 raw_spin_unlock(&rq
->lock
);
955 * task_rq_lock - lock the runqueue a given task resides on and disable
956 * interrupts. Note the ordering: we can safely lookup the task_rq without
957 * explicitly disabling preemption.
959 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
965 local_irq_save(*flags
);
967 raw_spin_lock(&rq
->lock
);
968 if (likely(rq
== task_rq(p
)))
970 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
974 static void __task_rq_unlock(struct rq
*rq
)
977 raw_spin_unlock(&rq
->lock
);
980 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
983 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
987 * this_rq_lock - lock this runqueue and disable interrupts.
989 static struct rq
*this_rq_lock(void)
996 raw_spin_lock(&rq
->lock
);
1001 #ifdef CONFIG_SCHED_HRTICK
1003 * Use HR-timers to deliver accurate preemption points.
1005 * Its all a bit involved since we cannot program an hrt while holding the
1006 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1009 * When we get rescheduled we reprogram the hrtick_timer outside of the
1015 * - enabled by features
1016 * - hrtimer is actually high res
1018 static inline int hrtick_enabled(struct rq
*rq
)
1020 if (!sched_feat(HRTICK
))
1022 if (!cpu_active(cpu_of(rq
)))
1024 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1027 static void hrtick_clear(struct rq
*rq
)
1029 if (hrtimer_active(&rq
->hrtick_timer
))
1030 hrtimer_cancel(&rq
->hrtick_timer
);
1034 * High-resolution timer tick.
1035 * Runs from hardirq context with interrupts disabled.
1037 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1039 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1041 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1043 raw_spin_lock(&rq
->lock
);
1044 update_rq_clock(rq
);
1045 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1046 raw_spin_unlock(&rq
->lock
);
1048 return HRTIMER_NORESTART
;
1053 * called from hardirq (IPI) context
1055 static void __hrtick_start(void *arg
)
1057 struct rq
*rq
= arg
;
1059 raw_spin_lock(&rq
->lock
);
1060 hrtimer_restart(&rq
->hrtick_timer
);
1061 rq
->hrtick_csd_pending
= 0;
1062 raw_spin_unlock(&rq
->lock
);
1066 * Called to set the hrtick timer state.
1068 * called with rq->lock held and irqs disabled
1070 static void hrtick_start(struct rq
*rq
, u64 delay
)
1072 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1073 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1075 hrtimer_set_expires(timer
, time
);
1077 if (rq
== this_rq()) {
1078 hrtimer_restart(timer
);
1079 } else if (!rq
->hrtick_csd_pending
) {
1080 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1081 rq
->hrtick_csd_pending
= 1;
1086 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1088 int cpu
= (int)(long)hcpu
;
1091 case CPU_UP_CANCELED
:
1092 case CPU_UP_CANCELED_FROZEN
:
1093 case CPU_DOWN_PREPARE
:
1094 case CPU_DOWN_PREPARE_FROZEN
:
1096 case CPU_DEAD_FROZEN
:
1097 hrtick_clear(cpu_rq(cpu
));
1104 static __init
void init_hrtick(void)
1106 hotcpu_notifier(hotplug_hrtick
, 0);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq
*rq
, u64 delay
)
1116 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1117 HRTIMER_MODE_REL_PINNED
, 0);
1120 static inline void init_hrtick(void)
1123 #endif /* CONFIG_SMP */
1125 static void init_rq_hrtick(struct rq
*rq
)
1128 rq
->hrtick_csd_pending
= 0;
1130 rq
->hrtick_csd
.flags
= 0;
1131 rq
->hrtick_csd
.func
= __hrtick_start
;
1132 rq
->hrtick_csd
.info
= rq
;
1135 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1136 rq
->hrtick_timer
.function
= hrtick
;
1138 #else /* CONFIG_SCHED_HRTICK */
1139 static inline void hrtick_clear(struct rq
*rq
)
1143 static inline void init_rq_hrtick(struct rq
*rq
)
1147 static inline void init_hrtick(void)
1150 #endif /* CONFIG_SCHED_HRTICK */
1153 * resched_task - mark a task 'to be rescheduled now'.
1155 * On UP this means the setting of the need_resched flag, on SMP it
1156 * might also involve a cross-CPU call to trigger the scheduler on
1161 #ifndef tsk_is_polling
1162 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1165 static void resched_task(struct task_struct
*p
)
1169 assert_raw_spin_locked(&task_rq(p
)->lock
);
1171 if (test_tsk_need_resched(p
))
1174 set_tsk_need_resched(p
);
1177 if (cpu
== smp_processor_id())
1180 /* NEED_RESCHED must be visible before we test polling */
1182 if (!tsk_is_polling(p
))
1183 smp_send_reschedule(cpu
);
1186 static void resched_cpu(int cpu
)
1188 struct rq
*rq
= cpu_rq(cpu
);
1189 unsigned long flags
;
1191 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1193 resched_task(cpu_curr(cpu
));
1194 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1199 * When add_timer_on() enqueues a timer into the timer wheel of an
1200 * idle CPU then this timer might expire before the next timer event
1201 * which is scheduled to wake up that CPU. In case of a completely
1202 * idle system the next event might even be infinite time into the
1203 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1204 * leaves the inner idle loop so the newly added timer is taken into
1205 * account when the CPU goes back to idle and evaluates the timer
1206 * wheel for the next timer event.
1208 void wake_up_idle_cpu(int cpu
)
1210 struct rq
*rq
= cpu_rq(cpu
);
1212 if (cpu
== smp_processor_id())
1216 * This is safe, as this function is called with the timer
1217 * wheel base lock of (cpu) held. When the CPU is on the way
1218 * to idle and has not yet set rq->curr to idle then it will
1219 * be serialized on the timer wheel base lock and take the new
1220 * timer into account automatically.
1222 if (rq
->curr
!= rq
->idle
)
1226 * We can set TIF_RESCHED on the idle task of the other CPU
1227 * lockless. The worst case is that the other CPU runs the
1228 * idle task through an additional NOOP schedule()
1230 set_tsk_need_resched(rq
->idle
);
1232 /* NEED_RESCHED must be visible before we test polling */
1234 if (!tsk_is_polling(rq
->idle
))
1235 smp_send_reschedule(cpu
);
1238 int nohz_ratelimit(int cpu
)
1240 struct rq
*rq
= cpu_rq(cpu
);
1241 u64 diff
= rq
->clock
- rq
->nohz_stamp
;
1243 rq
->nohz_stamp
= rq
->clock
;
1245 return diff
< (NSEC_PER_SEC
/ HZ
) >> 1;
1248 #endif /* CONFIG_NO_HZ */
1250 static u64
sched_avg_period(void)
1252 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1255 static void sched_avg_update(struct rq
*rq
)
1257 s64 period
= sched_avg_period();
1259 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1260 rq
->age_stamp
+= period
;
1265 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1267 rq
->rt_avg
+= rt_delta
;
1268 sched_avg_update(rq
);
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct
*p
)
1274 assert_raw_spin_locked(&task_rq(p
)->lock
);
1275 set_tsk_need_resched(p
);
1278 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1301 struct load_weight
*lw
)
1305 if (!lw
->inv_weight
) {
1306 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1309 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1313 tmp
= (u64
)delta_exec
* weight
;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp
> WMULT_CONST
))
1318 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1321 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1323 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1326 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1332 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight
[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult
[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 /* Time spent by the tasks of the cpu accounting group executing in ... */
1392 enum cpuacct_stat_index
{
1393 CPUACCT_STAT_USER
, /* ... user mode */
1394 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1396 CPUACCT_STAT_NSTATS
,
1399 #ifdef CONFIG_CGROUP_CPUACCT
1400 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1401 static void cpuacct_update_stats(struct task_struct
*tsk
,
1402 enum cpuacct_stat_index idx
, cputime_t val
);
1404 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1405 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1406 enum cpuacct_stat_index idx
, cputime_t val
) {}
1409 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1411 update_load_add(&rq
->load
, load
);
1414 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1416 update_load_sub(&rq
->load
, load
);
1419 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1420 typedef int (*tg_visitor
)(struct task_group
*, void *);
1423 * Iterate the full tree, calling @down when first entering a node and @up when
1424 * leaving it for the final time.
1426 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1428 struct task_group
*parent
, *child
;
1432 parent
= &root_task_group
;
1434 ret
= (*down
)(parent
, data
);
1437 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1444 ret
= (*up
)(parent
, data
);
1449 parent
= parent
->parent
;
1458 static int tg_nop(struct task_group
*tg
, void *data
)
1465 /* Used instead of source_load when we know the type == 0 */
1466 static unsigned long weighted_cpuload(const int cpu
)
1468 return cpu_rq(cpu
)->load
.weight
;
1472 * Return a low guess at the load of a migration-source cpu weighted
1473 * according to the scheduling class and "nice" value.
1475 * We want to under-estimate the load of migration sources, to
1476 * balance conservatively.
1478 static unsigned long source_load(int cpu
, int type
)
1480 struct rq
*rq
= cpu_rq(cpu
);
1481 unsigned long total
= weighted_cpuload(cpu
);
1483 if (type
== 0 || !sched_feat(LB_BIAS
))
1486 return min(rq
->cpu_load
[type
-1], total
);
1490 * Return a high guess at the load of a migration-target cpu weighted
1491 * according to the scheduling class and "nice" value.
1493 static unsigned long target_load(int cpu
, int type
)
1495 struct rq
*rq
= cpu_rq(cpu
);
1496 unsigned long total
= weighted_cpuload(cpu
);
1498 if (type
== 0 || !sched_feat(LB_BIAS
))
1501 return max(rq
->cpu_load
[type
-1], total
);
1504 static unsigned long power_of(int cpu
)
1506 return cpu_rq(cpu
)->cpu_power
;
1509 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1511 static unsigned long cpu_avg_load_per_task(int cpu
)
1513 struct rq
*rq
= cpu_rq(cpu
);
1514 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1517 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1519 rq
->avg_load_per_task
= 0;
1521 return rq
->avg_load_per_task
;
1524 #ifdef CONFIG_FAIR_GROUP_SCHED
1526 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1528 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1531 * Calculate and set the cpu's group shares.
1533 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1534 unsigned long sd_shares
,
1535 unsigned long sd_rq_weight
,
1536 unsigned long *usd_rq_weight
)
1538 unsigned long shares
, rq_weight
;
1541 rq_weight
= usd_rq_weight
[cpu
];
1544 rq_weight
= NICE_0_LOAD
;
1548 * \Sum_j shares_j * rq_weight_i
1549 * shares_i = -----------------------------
1550 * \Sum_j rq_weight_j
1552 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1553 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1555 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1556 sysctl_sched_shares_thresh
) {
1557 struct rq
*rq
= cpu_rq(cpu
);
1558 unsigned long flags
;
1560 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1561 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1562 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1563 __set_se_shares(tg
->se
[cpu
], shares
);
1564 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1569 * Re-compute the task group their per cpu shares over the given domain.
1570 * This needs to be done in a bottom-up fashion because the rq weight of a
1571 * parent group depends on the shares of its child groups.
1573 static int tg_shares_up(struct task_group
*tg
, void *data
)
1575 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1576 unsigned long *usd_rq_weight
;
1577 struct sched_domain
*sd
= data
;
1578 unsigned long flags
;
1584 local_irq_save(flags
);
1585 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1587 for_each_cpu(i
, sched_domain_span(sd
)) {
1588 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1589 usd_rq_weight
[i
] = weight
;
1591 rq_weight
+= weight
;
1593 * If there are currently no tasks on the cpu pretend there
1594 * is one of average load so that when a new task gets to
1595 * run here it will not get delayed by group starvation.
1598 weight
= NICE_0_LOAD
;
1600 sum_weight
+= weight
;
1601 shares
+= tg
->cfs_rq
[i
]->shares
;
1605 rq_weight
= sum_weight
;
1607 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1608 shares
= tg
->shares
;
1610 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1611 shares
= tg
->shares
;
1613 for_each_cpu(i
, sched_domain_span(sd
))
1614 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1616 local_irq_restore(flags
);
1622 * Compute the cpu's hierarchical load factor for each task group.
1623 * This needs to be done in a top-down fashion because the load of a child
1624 * group is a fraction of its parents load.
1626 static int tg_load_down(struct task_group
*tg
, void *data
)
1629 long cpu
= (long)data
;
1632 load
= cpu_rq(cpu
)->load
.weight
;
1634 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1635 load
*= tg
->cfs_rq
[cpu
]->shares
;
1636 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1639 tg
->cfs_rq
[cpu
]->h_load
= load
;
1644 static void update_shares(struct sched_domain
*sd
)
1649 if (root_task_group_empty())
1652 now
= cpu_clock(raw_smp_processor_id());
1653 elapsed
= now
- sd
->last_update
;
1655 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1656 sd
->last_update
= now
;
1657 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1661 static void update_h_load(long cpu
)
1663 if (root_task_group_empty())
1666 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1671 static inline void update_shares(struct sched_domain
*sd
)
1677 #ifdef CONFIG_PREEMPT
1679 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1682 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1683 * way at the expense of forcing extra atomic operations in all
1684 * invocations. This assures that the double_lock is acquired using the
1685 * same underlying policy as the spinlock_t on this architecture, which
1686 * reduces latency compared to the unfair variant below. However, it
1687 * also adds more overhead and therefore may reduce throughput.
1689 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1690 __releases(this_rq
->lock
)
1691 __acquires(busiest
->lock
)
1692 __acquires(this_rq
->lock
)
1694 raw_spin_unlock(&this_rq
->lock
);
1695 double_rq_lock(this_rq
, busiest
);
1702 * Unfair double_lock_balance: Optimizes throughput at the expense of
1703 * latency by eliminating extra atomic operations when the locks are
1704 * already in proper order on entry. This favors lower cpu-ids and will
1705 * grant the double lock to lower cpus over higher ids under contention,
1706 * regardless of entry order into the function.
1708 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1709 __releases(this_rq
->lock
)
1710 __acquires(busiest
->lock
)
1711 __acquires(this_rq
->lock
)
1715 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1716 if (busiest
< this_rq
) {
1717 raw_spin_unlock(&this_rq
->lock
);
1718 raw_spin_lock(&busiest
->lock
);
1719 raw_spin_lock_nested(&this_rq
->lock
,
1720 SINGLE_DEPTH_NESTING
);
1723 raw_spin_lock_nested(&busiest
->lock
,
1724 SINGLE_DEPTH_NESTING
);
1729 #endif /* CONFIG_PREEMPT */
1732 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1734 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1736 if (unlikely(!irqs_disabled())) {
1737 /* printk() doesn't work good under rq->lock */
1738 raw_spin_unlock(&this_rq
->lock
);
1742 return _double_lock_balance(this_rq
, busiest
);
1745 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1746 __releases(busiest
->lock
)
1748 raw_spin_unlock(&busiest
->lock
);
1749 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1753 * double_rq_lock - safely lock two runqueues
1755 * Note this does not disable interrupts like task_rq_lock,
1756 * you need to do so manually before calling.
1758 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1759 __acquires(rq1
->lock
)
1760 __acquires(rq2
->lock
)
1762 BUG_ON(!irqs_disabled());
1764 raw_spin_lock(&rq1
->lock
);
1765 __acquire(rq2
->lock
); /* Fake it out ;) */
1768 raw_spin_lock(&rq1
->lock
);
1769 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1771 raw_spin_lock(&rq2
->lock
);
1772 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1778 * double_rq_unlock - safely unlock two runqueues
1780 * Note this does not restore interrupts like task_rq_unlock,
1781 * you need to do so manually after calling.
1783 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1784 __releases(rq1
->lock
)
1785 __releases(rq2
->lock
)
1787 raw_spin_unlock(&rq1
->lock
);
1789 raw_spin_unlock(&rq2
->lock
);
1791 __release(rq2
->lock
);
1796 #ifdef CONFIG_FAIR_GROUP_SCHED
1797 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1800 cfs_rq
->shares
= shares
;
1805 static void calc_load_account_idle(struct rq
*this_rq
);
1806 static void update_sysctl(void);
1807 static int get_update_sysctl_factor(void);
1809 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1811 set_task_rq(p
, cpu
);
1814 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1815 * successfuly executed on another CPU. We must ensure that updates of
1816 * per-task data have been completed by this moment.
1819 task_thread_info(p
)->cpu
= cpu
;
1823 static const struct sched_class rt_sched_class
;
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 #include "sched_stats.h"
1831 static void inc_nr_running(struct rq
*rq
)
1836 static void dec_nr_running(struct rq
*rq
)
1841 static void set_load_weight(struct task_struct
*p
)
1843 if (task_has_rt_policy(p
)) {
1844 p
->se
.load
.weight
= 0;
1845 p
->se
.load
.inv_weight
= WMULT_CONST
;
1850 * SCHED_IDLE tasks get minimal weight:
1852 if (p
->policy
== SCHED_IDLE
) {
1853 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1854 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1858 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1859 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1862 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1864 update_rq_clock(rq
);
1865 sched_info_queued(p
);
1866 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1870 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1872 update_rq_clock(rq
);
1873 sched_info_dequeued(p
);
1874 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1879 * activate_task - move a task to the runqueue.
1881 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1883 if (task_contributes_to_load(p
))
1884 rq
->nr_uninterruptible
--;
1886 enqueue_task(rq
, p
, flags
);
1891 * deactivate_task - remove a task from the runqueue.
1893 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1895 if (task_contributes_to_load(p
))
1896 rq
->nr_uninterruptible
++;
1898 dequeue_task(rq
, p
, flags
);
1902 #include "sched_idletask.c"
1903 #include "sched_fair.c"
1904 #include "sched_rt.c"
1905 #ifdef CONFIG_SCHED_DEBUG
1906 # include "sched_debug.c"
1910 * __normal_prio - return the priority that is based on the static prio
1912 static inline int __normal_prio(struct task_struct
*p
)
1914 return p
->static_prio
;
1918 * Calculate the expected normal priority: i.e. priority
1919 * without taking RT-inheritance into account. Might be
1920 * boosted by interactivity modifiers. Changes upon fork,
1921 * setprio syscalls, and whenever the interactivity
1922 * estimator recalculates.
1924 static inline int normal_prio(struct task_struct
*p
)
1928 if (task_has_rt_policy(p
))
1929 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1931 prio
= __normal_prio(p
);
1936 * Calculate the current priority, i.e. the priority
1937 * taken into account by the scheduler. This value might
1938 * be boosted by RT tasks, or might be boosted by
1939 * interactivity modifiers. Will be RT if the task got
1940 * RT-boosted. If not then it returns p->normal_prio.
1942 static int effective_prio(struct task_struct
*p
)
1944 p
->normal_prio
= normal_prio(p
);
1946 * If we are RT tasks or we were boosted to RT priority,
1947 * keep the priority unchanged. Otherwise, update priority
1948 * to the normal priority:
1950 if (!rt_prio(p
->prio
))
1951 return p
->normal_prio
;
1956 * task_curr - is this task currently executing on a CPU?
1957 * @p: the task in question.
1959 inline int task_curr(const struct task_struct
*p
)
1961 return cpu_curr(task_cpu(p
)) == p
;
1964 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1965 const struct sched_class
*prev_class
,
1966 int oldprio
, int running
)
1968 if (prev_class
!= p
->sched_class
) {
1969 if (prev_class
->switched_from
)
1970 prev_class
->switched_from(rq
, p
, running
);
1971 p
->sched_class
->switched_to(rq
, p
, running
);
1973 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1978 * Is this task likely cache-hot:
1981 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1985 if (p
->sched_class
!= &fair_sched_class
)
1989 * Buddy candidates are cache hot:
1991 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
1992 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1993 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1996 if (sysctl_sched_migration_cost
== -1)
1998 if (sysctl_sched_migration_cost
== 0)
2001 delta
= now
- p
->se
.exec_start
;
2003 return delta
< (s64
)sysctl_sched_migration_cost
;
2006 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2008 #ifdef CONFIG_SCHED_DEBUG
2010 * We should never call set_task_cpu() on a blocked task,
2011 * ttwu() will sort out the placement.
2013 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2014 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2017 trace_sched_migrate_task(p
, new_cpu
);
2019 if (task_cpu(p
) != new_cpu
) {
2020 p
->se
.nr_migrations
++;
2021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2024 __set_task_cpu(p
, new_cpu
);
2027 struct migration_arg
{
2028 struct task_struct
*task
;
2032 static int migration_cpu_stop(void *data
);
2035 * The task's runqueue lock must be held.
2036 * Returns true if you have to wait for migration thread.
2038 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2040 struct rq
*rq
= task_rq(p
);
2043 * If the task is not on a runqueue (and not running), then
2044 * the next wake-up will properly place the task.
2046 return p
->se
.on_rq
|| task_running(rq
, p
);
2050 * wait_task_inactive - wait for a thread to unschedule.
2052 * If @match_state is nonzero, it's the @p->state value just checked and
2053 * not expected to change. If it changes, i.e. @p might have woken up,
2054 * then return zero. When we succeed in waiting for @p to be off its CPU,
2055 * we return a positive number (its total switch count). If a second call
2056 * a short while later returns the same number, the caller can be sure that
2057 * @p has remained unscheduled the whole time.
2059 * The caller must ensure that the task *will* unschedule sometime soon,
2060 * else this function might spin for a *long* time. This function can't
2061 * be called with interrupts off, or it may introduce deadlock with
2062 * smp_call_function() if an IPI is sent by the same process we are
2063 * waiting to become inactive.
2065 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2067 unsigned long flags
;
2074 * We do the initial early heuristics without holding
2075 * any task-queue locks at all. We'll only try to get
2076 * the runqueue lock when things look like they will
2082 * If the task is actively running on another CPU
2083 * still, just relax and busy-wait without holding
2086 * NOTE! Since we don't hold any locks, it's not
2087 * even sure that "rq" stays as the right runqueue!
2088 * But we don't care, since "task_running()" will
2089 * return false if the runqueue has changed and p
2090 * is actually now running somewhere else!
2092 while (task_running(rq
, p
)) {
2093 if (match_state
&& unlikely(p
->state
!= match_state
))
2099 * Ok, time to look more closely! We need the rq
2100 * lock now, to be *sure*. If we're wrong, we'll
2101 * just go back and repeat.
2103 rq
= task_rq_lock(p
, &flags
);
2104 trace_sched_wait_task(p
);
2105 running
= task_running(rq
, p
);
2106 on_rq
= p
->se
.on_rq
;
2108 if (!match_state
|| p
->state
== match_state
)
2109 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2110 task_rq_unlock(rq
, &flags
);
2113 * If it changed from the expected state, bail out now.
2115 if (unlikely(!ncsw
))
2119 * Was it really running after all now that we
2120 * checked with the proper locks actually held?
2122 * Oops. Go back and try again..
2124 if (unlikely(running
)) {
2130 * It's not enough that it's not actively running,
2131 * it must be off the runqueue _entirely_, and not
2134 * So if it was still runnable (but just not actively
2135 * running right now), it's preempted, and we should
2136 * yield - it could be a while.
2138 if (unlikely(on_rq
)) {
2139 schedule_timeout_uninterruptible(1);
2144 * Ahh, all good. It wasn't running, and it wasn't
2145 * runnable, which means that it will never become
2146 * running in the future either. We're all done!
2155 * kick_process - kick a running thread to enter/exit the kernel
2156 * @p: the to-be-kicked thread
2158 * Cause a process which is running on another CPU to enter
2159 * kernel-mode, without any delay. (to get signals handled.)
2161 * NOTE: this function doesnt have to take the runqueue lock,
2162 * because all it wants to ensure is that the remote task enters
2163 * the kernel. If the IPI races and the task has been migrated
2164 * to another CPU then no harm is done and the purpose has been
2167 void kick_process(struct task_struct
*p
)
2173 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2174 smp_send_reschedule(cpu
);
2177 EXPORT_SYMBOL_GPL(kick_process
);
2178 #endif /* CONFIG_SMP */
2181 * task_oncpu_function_call - call a function on the cpu on which a task runs
2182 * @p: the task to evaluate
2183 * @func: the function to be called
2184 * @info: the function call argument
2186 * Calls the function @func when the task is currently running. This might
2187 * be on the current CPU, which just calls the function directly
2189 void task_oncpu_function_call(struct task_struct
*p
,
2190 void (*func
) (void *info
), void *info
)
2197 smp_call_function_single(cpu
, func
, info
, 1);
2203 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2205 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2208 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2210 /* Look for allowed, online CPU in same node. */
2211 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2212 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2215 /* Any allowed, online CPU? */
2216 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2217 if (dest_cpu
< nr_cpu_ids
)
2220 /* No more Mr. Nice Guy. */
2221 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2222 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2224 * Don't tell them about moving exiting tasks or
2225 * kernel threads (both mm NULL), since they never
2228 if (p
->mm
&& printk_ratelimit()) {
2229 printk(KERN_INFO
"process %d (%s) no "
2230 "longer affine to cpu%d\n",
2231 task_pid_nr(p
), p
->comm
, cpu
);
2239 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2242 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2244 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2247 * In order not to call set_task_cpu() on a blocking task we need
2248 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2251 * Since this is common to all placement strategies, this lives here.
2253 * [ this allows ->select_task() to simply return task_cpu(p) and
2254 * not worry about this generic constraint ]
2256 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2258 cpu
= select_fallback_rq(task_cpu(p
), p
);
2263 static void update_avg(u64
*avg
, u64 sample
)
2265 s64 diff
= sample
- *avg
;
2271 * try_to_wake_up - wake up a thread
2272 * @p: the to-be-woken-up thread
2273 * @state: the mask of task states that can be woken
2274 * @sync: do a synchronous wakeup?
2276 * Put it on the run-queue if it's not already there. The "current"
2277 * thread is always on the run-queue (except when the actual
2278 * re-schedule is in progress), and as such you're allowed to do
2279 * the simpler "current->state = TASK_RUNNING" to mark yourself
2280 * runnable without the overhead of this.
2282 * returns failure only if the task is already active.
2284 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2287 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2288 unsigned long flags
;
2289 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2292 this_cpu
= get_cpu();
2295 rq
= task_rq_lock(p
, &flags
);
2296 if (!(p
->state
& state
))
2306 if (unlikely(task_running(rq
, p
)))
2310 * In order to handle concurrent wakeups and release the rq->lock
2311 * we put the task in TASK_WAKING state.
2313 * First fix up the nr_uninterruptible count:
2315 if (task_contributes_to_load(p
)) {
2316 if (likely(cpu_online(orig_cpu
)))
2317 rq
->nr_uninterruptible
--;
2319 this_rq()->nr_uninterruptible
--;
2321 p
->state
= TASK_WAKING
;
2323 if (p
->sched_class
->task_waking
) {
2324 p
->sched_class
->task_waking(rq
, p
);
2325 en_flags
|= ENQUEUE_WAKING
;
2328 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2329 if (cpu
!= orig_cpu
)
2330 set_task_cpu(p
, cpu
);
2331 __task_rq_unlock(rq
);
2334 raw_spin_lock(&rq
->lock
);
2337 * We migrated the task without holding either rq->lock, however
2338 * since the task is not on the task list itself, nobody else
2339 * will try and migrate the task, hence the rq should match the
2340 * cpu we just moved it to.
2342 WARN_ON(task_cpu(p
) != cpu
);
2343 WARN_ON(p
->state
!= TASK_WAKING
);
2345 #ifdef CONFIG_SCHEDSTATS
2346 schedstat_inc(rq
, ttwu_count
);
2347 if (cpu
== this_cpu
)
2348 schedstat_inc(rq
, ttwu_local
);
2350 struct sched_domain
*sd
;
2351 for_each_domain(this_cpu
, sd
) {
2352 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2353 schedstat_inc(sd
, ttwu_wake_remote
);
2358 #endif /* CONFIG_SCHEDSTATS */
2361 #endif /* CONFIG_SMP */
2362 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2363 if (wake_flags
& WF_SYNC
)
2364 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2365 if (orig_cpu
!= cpu
)
2366 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2367 if (cpu
== this_cpu
)
2368 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2370 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2371 activate_task(rq
, p
, en_flags
);
2375 trace_sched_wakeup(p
, success
);
2376 check_preempt_curr(rq
, p
, wake_flags
);
2378 p
->state
= TASK_RUNNING
;
2380 if (p
->sched_class
->task_woken
)
2381 p
->sched_class
->task_woken(rq
, p
);
2383 if (unlikely(rq
->idle_stamp
)) {
2384 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2385 u64 max
= 2*sysctl_sched_migration_cost
;
2390 update_avg(&rq
->avg_idle
, delta
);
2395 task_rq_unlock(rq
, &flags
);
2402 * wake_up_process - Wake up a specific process
2403 * @p: The process to be woken up.
2405 * Attempt to wake up the nominated process and move it to the set of runnable
2406 * processes. Returns 1 if the process was woken up, 0 if it was already
2409 * It may be assumed that this function implies a write memory barrier before
2410 * changing the task state if and only if any tasks are woken up.
2412 int wake_up_process(struct task_struct
*p
)
2414 return try_to_wake_up(p
, TASK_ALL
, 0);
2416 EXPORT_SYMBOL(wake_up_process
);
2418 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2420 return try_to_wake_up(p
, state
, 0);
2424 * Perform scheduler related setup for a newly forked process p.
2425 * p is forked by current.
2427 * __sched_fork() is basic setup used by init_idle() too:
2429 static void __sched_fork(struct task_struct
*p
)
2431 p
->se
.exec_start
= 0;
2432 p
->se
.sum_exec_runtime
= 0;
2433 p
->se
.prev_sum_exec_runtime
= 0;
2434 p
->se
.nr_migrations
= 0;
2436 #ifdef CONFIG_SCHEDSTATS
2437 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2440 INIT_LIST_HEAD(&p
->rt
.run_list
);
2442 INIT_LIST_HEAD(&p
->se
.group_node
);
2444 #ifdef CONFIG_PREEMPT_NOTIFIERS
2445 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2450 * fork()/clone()-time setup:
2452 void sched_fork(struct task_struct
*p
, int clone_flags
)
2454 int cpu
= get_cpu();
2458 * We mark the process as running here. This guarantees that
2459 * nobody will actually run it, and a signal or other external
2460 * event cannot wake it up and insert it on the runqueue either.
2462 p
->state
= TASK_RUNNING
;
2465 * Revert to default priority/policy on fork if requested.
2467 if (unlikely(p
->sched_reset_on_fork
)) {
2468 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2469 p
->policy
= SCHED_NORMAL
;
2470 p
->normal_prio
= p
->static_prio
;
2473 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2474 p
->static_prio
= NICE_TO_PRIO(0);
2475 p
->normal_prio
= p
->static_prio
;
2480 * We don't need the reset flag anymore after the fork. It has
2481 * fulfilled its duty:
2483 p
->sched_reset_on_fork
= 0;
2487 * Make sure we do not leak PI boosting priority to the child.
2489 p
->prio
= current
->normal_prio
;
2491 if (!rt_prio(p
->prio
))
2492 p
->sched_class
= &fair_sched_class
;
2494 if (p
->sched_class
->task_fork
)
2495 p
->sched_class
->task_fork(p
);
2498 * The child is not yet in the pid-hash so no cgroup attach races,
2499 * and the cgroup is pinned to this child due to cgroup_fork()
2500 * is ran before sched_fork().
2502 * Silence PROVE_RCU.
2505 set_task_cpu(p
, cpu
);
2508 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2509 if (likely(sched_info_on()))
2510 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2512 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2515 #ifdef CONFIG_PREEMPT
2516 /* Want to start with kernel preemption disabled. */
2517 task_thread_info(p
)->preempt_count
= 1;
2519 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2525 * wake_up_new_task - wake up a newly created task for the first time.
2527 * This function will do some initial scheduler statistics housekeeping
2528 * that must be done for every newly created context, then puts the task
2529 * on the runqueue and wakes it.
2531 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2533 unsigned long flags
;
2535 int cpu __maybe_unused
= get_cpu();
2538 rq
= task_rq_lock(p
, &flags
);
2539 p
->state
= TASK_WAKING
;
2542 * Fork balancing, do it here and not earlier because:
2543 * - cpus_allowed can change in the fork path
2544 * - any previously selected cpu might disappear through hotplug
2546 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2547 * without people poking at ->cpus_allowed.
2549 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2550 set_task_cpu(p
, cpu
);
2552 p
->state
= TASK_RUNNING
;
2553 task_rq_unlock(rq
, &flags
);
2556 rq
= task_rq_lock(p
, &flags
);
2557 activate_task(rq
, p
, 0);
2558 trace_sched_wakeup_new(p
, 1);
2559 check_preempt_curr(rq
, p
, WF_FORK
);
2561 if (p
->sched_class
->task_woken
)
2562 p
->sched_class
->task_woken(rq
, p
);
2564 task_rq_unlock(rq
, &flags
);
2568 #ifdef CONFIG_PREEMPT_NOTIFIERS
2571 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2572 * @notifier: notifier struct to register
2574 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2576 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2578 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2581 * preempt_notifier_unregister - no longer interested in preemption notifications
2582 * @notifier: notifier struct to unregister
2584 * This is safe to call from within a preemption notifier.
2586 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2588 hlist_del(¬ifier
->link
);
2590 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2592 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2594 struct preempt_notifier
*notifier
;
2595 struct hlist_node
*node
;
2597 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2598 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2602 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2603 struct task_struct
*next
)
2605 struct preempt_notifier
*notifier
;
2606 struct hlist_node
*node
;
2608 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2609 notifier
->ops
->sched_out(notifier
, next
);
2612 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2614 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2619 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2620 struct task_struct
*next
)
2624 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2627 * prepare_task_switch - prepare to switch tasks
2628 * @rq: the runqueue preparing to switch
2629 * @prev: the current task that is being switched out
2630 * @next: the task we are going to switch to.
2632 * This is called with the rq lock held and interrupts off. It must
2633 * be paired with a subsequent finish_task_switch after the context
2636 * prepare_task_switch sets up locking and calls architecture specific
2640 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2641 struct task_struct
*next
)
2643 fire_sched_out_preempt_notifiers(prev
, next
);
2644 prepare_lock_switch(rq
, next
);
2645 prepare_arch_switch(next
);
2649 * finish_task_switch - clean up after a task-switch
2650 * @rq: runqueue associated with task-switch
2651 * @prev: the thread we just switched away from.
2653 * finish_task_switch must be called after the context switch, paired
2654 * with a prepare_task_switch call before the context switch.
2655 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2656 * and do any other architecture-specific cleanup actions.
2658 * Note that we may have delayed dropping an mm in context_switch(). If
2659 * so, we finish that here outside of the runqueue lock. (Doing it
2660 * with the lock held can cause deadlocks; see schedule() for
2663 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2664 __releases(rq
->lock
)
2666 struct mm_struct
*mm
= rq
->prev_mm
;
2672 * A task struct has one reference for the use as "current".
2673 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2674 * schedule one last time. The schedule call will never return, and
2675 * the scheduled task must drop that reference.
2676 * The test for TASK_DEAD must occur while the runqueue locks are
2677 * still held, otherwise prev could be scheduled on another cpu, die
2678 * there before we look at prev->state, and then the reference would
2680 * Manfred Spraul <manfred@colorfullife.com>
2682 prev_state
= prev
->state
;
2683 finish_arch_switch(prev
);
2684 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2685 local_irq_disable();
2686 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2687 perf_event_task_sched_in(current
);
2688 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2690 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2691 finish_lock_switch(rq
, prev
);
2693 fire_sched_in_preempt_notifiers(current
);
2696 if (unlikely(prev_state
== TASK_DEAD
)) {
2698 * Remove function-return probe instances associated with this
2699 * task and put them back on the free list.
2701 kprobe_flush_task(prev
);
2702 put_task_struct(prev
);
2708 /* assumes rq->lock is held */
2709 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2711 if (prev
->sched_class
->pre_schedule
)
2712 prev
->sched_class
->pre_schedule(rq
, prev
);
2715 /* rq->lock is NOT held, but preemption is disabled */
2716 static inline void post_schedule(struct rq
*rq
)
2718 if (rq
->post_schedule
) {
2719 unsigned long flags
;
2721 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2722 if (rq
->curr
->sched_class
->post_schedule
)
2723 rq
->curr
->sched_class
->post_schedule(rq
);
2724 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2726 rq
->post_schedule
= 0;
2732 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2736 static inline void post_schedule(struct rq
*rq
)
2743 * schedule_tail - first thing a freshly forked thread must call.
2744 * @prev: the thread we just switched away from.
2746 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2747 __releases(rq
->lock
)
2749 struct rq
*rq
= this_rq();
2751 finish_task_switch(rq
, prev
);
2754 * FIXME: do we need to worry about rq being invalidated by the
2759 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2760 /* In this case, finish_task_switch does not reenable preemption */
2763 if (current
->set_child_tid
)
2764 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2768 * context_switch - switch to the new MM and the new
2769 * thread's register state.
2772 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2773 struct task_struct
*next
)
2775 struct mm_struct
*mm
, *oldmm
;
2777 prepare_task_switch(rq
, prev
, next
);
2778 trace_sched_switch(prev
, next
);
2780 oldmm
= prev
->active_mm
;
2782 * For paravirt, this is coupled with an exit in switch_to to
2783 * combine the page table reload and the switch backend into
2786 arch_start_context_switch(prev
);
2789 next
->active_mm
= oldmm
;
2790 atomic_inc(&oldmm
->mm_count
);
2791 enter_lazy_tlb(oldmm
, next
);
2793 switch_mm(oldmm
, mm
, next
);
2795 if (likely(!prev
->mm
)) {
2796 prev
->active_mm
= NULL
;
2797 rq
->prev_mm
= oldmm
;
2800 * Since the runqueue lock will be released by the next
2801 * task (which is an invalid locking op but in the case
2802 * of the scheduler it's an obvious special-case), so we
2803 * do an early lockdep release here:
2805 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2806 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2809 /* Here we just switch the register state and the stack. */
2810 switch_to(prev
, next
, prev
);
2814 * this_rq must be evaluated again because prev may have moved
2815 * CPUs since it called schedule(), thus the 'rq' on its stack
2816 * frame will be invalid.
2818 finish_task_switch(this_rq(), prev
);
2822 * nr_running, nr_uninterruptible and nr_context_switches:
2824 * externally visible scheduler statistics: current number of runnable
2825 * threads, current number of uninterruptible-sleeping threads, total
2826 * number of context switches performed since bootup.
2828 unsigned long nr_running(void)
2830 unsigned long i
, sum
= 0;
2832 for_each_online_cpu(i
)
2833 sum
+= cpu_rq(i
)->nr_running
;
2838 unsigned long nr_uninterruptible(void)
2840 unsigned long i
, sum
= 0;
2842 for_each_possible_cpu(i
)
2843 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2846 * Since we read the counters lockless, it might be slightly
2847 * inaccurate. Do not allow it to go below zero though:
2849 if (unlikely((long)sum
< 0))
2855 unsigned long long nr_context_switches(void)
2858 unsigned long long sum
= 0;
2860 for_each_possible_cpu(i
)
2861 sum
+= cpu_rq(i
)->nr_switches
;
2866 unsigned long nr_iowait(void)
2868 unsigned long i
, sum
= 0;
2870 for_each_possible_cpu(i
)
2871 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2876 unsigned long nr_iowait_cpu(void)
2878 struct rq
*this = this_rq();
2879 return atomic_read(&this->nr_iowait
);
2882 unsigned long this_cpu_load(void)
2884 struct rq
*this = this_rq();
2885 return this->cpu_load
[0];
2889 /* Variables and functions for calc_load */
2890 static atomic_long_t calc_load_tasks
;
2891 static unsigned long calc_load_update
;
2892 unsigned long avenrun
[3];
2893 EXPORT_SYMBOL(avenrun
);
2895 static long calc_load_fold_active(struct rq
*this_rq
)
2897 long nr_active
, delta
= 0;
2899 nr_active
= this_rq
->nr_running
;
2900 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2902 if (nr_active
!= this_rq
->calc_load_active
) {
2903 delta
= nr_active
- this_rq
->calc_load_active
;
2904 this_rq
->calc_load_active
= nr_active
;
2912 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2914 * When making the ILB scale, we should try to pull this in as well.
2916 static atomic_long_t calc_load_tasks_idle
;
2918 static void calc_load_account_idle(struct rq
*this_rq
)
2922 delta
= calc_load_fold_active(this_rq
);
2924 atomic_long_add(delta
, &calc_load_tasks_idle
);
2927 static long calc_load_fold_idle(void)
2932 * Its got a race, we don't care...
2934 if (atomic_long_read(&calc_load_tasks_idle
))
2935 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2940 static void calc_load_account_idle(struct rq
*this_rq
)
2944 static inline long calc_load_fold_idle(void)
2951 * get_avenrun - get the load average array
2952 * @loads: pointer to dest load array
2953 * @offset: offset to add
2954 * @shift: shift count to shift the result left
2956 * These values are estimates at best, so no need for locking.
2958 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2960 loads
[0] = (avenrun
[0] + offset
) << shift
;
2961 loads
[1] = (avenrun
[1] + offset
) << shift
;
2962 loads
[2] = (avenrun
[2] + offset
) << shift
;
2965 static unsigned long
2966 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2969 load
+= active
* (FIXED_1
- exp
);
2970 return load
>> FSHIFT
;
2974 * calc_load - update the avenrun load estimates 10 ticks after the
2975 * CPUs have updated calc_load_tasks.
2977 void calc_global_load(void)
2979 unsigned long upd
= calc_load_update
+ 10;
2982 if (time_before(jiffies
, upd
))
2985 active
= atomic_long_read(&calc_load_tasks
);
2986 active
= active
> 0 ? active
* FIXED_1
: 0;
2988 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2989 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2990 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2992 calc_load_update
+= LOAD_FREQ
;
2996 * Called from update_cpu_load() to periodically update this CPU's
2999 static void calc_load_account_active(struct rq
*this_rq
)
3003 if (time_before(jiffies
, this_rq
->calc_load_update
))
3006 delta
= calc_load_fold_active(this_rq
);
3007 delta
+= calc_load_fold_idle();
3009 atomic_long_add(delta
, &calc_load_tasks
);
3011 this_rq
->calc_load_update
+= LOAD_FREQ
;
3015 * Update rq->cpu_load[] statistics. This function is usually called every
3016 * scheduler tick (TICK_NSEC).
3018 static void update_cpu_load(struct rq
*this_rq
)
3020 unsigned long this_load
= this_rq
->load
.weight
;
3023 this_rq
->nr_load_updates
++;
3025 /* Update our load: */
3026 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3027 unsigned long old_load
, new_load
;
3029 /* scale is effectively 1 << i now, and >> i divides by scale */
3031 old_load
= this_rq
->cpu_load
[i
];
3032 new_load
= this_load
;
3034 * Round up the averaging division if load is increasing. This
3035 * prevents us from getting stuck on 9 if the load is 10, for
3038 if (new_load
> old_load
)
3039 new_load
+= scale
-1;
3040 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3043 calc_load_account_active(this_rq
);
3049 * sched_exec - execve() is a valuable balancing opportunity, because at
3050 * this point the task has the smallest effective memory and cache footprint.
3052 void sched_exec(void)
3054 struct task_struct
*p
= current
;
3055 unsigned long flags
;
3059 rq
= task_rq_lock(p
, &flags
);
3060 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3061 if (dest_cpu
== smp_processor_id())
3065 * select_task_rq() can race against ->cpus_allowed
3067 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3068 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3069 struct migration_arg arg
= { p
, dest_cpu
};
3071 task_rq_unlock(rq
, &flags
);
3072 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3076 task_rq_unlock(rq
, &flags
);
3081 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3083 EXPORT_PER_CPU_SYMBOL(kstat
);
3086 * Return any ns on the sched_clock that have not yet been accounted in
3087 * @p in case that task is currently running.
3089 * Called with task_rq_lock() held on @rq.
3091 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3095 if (task_current(rq
, p
)) {
3096 update_rq_clock(rq
);
3097 ns
= rq
->clock
- p
->se
.exec_start
;
3105 unsigned long long task_delta_exec(struct task_struct
*p
)
3107 unsigned long flags
;
3111 rq
= task_rq_lock(p
, &flags
);
3112 ns
= do_task_delta_exec(p
, rq
);
3113 task_rq_unlock(rq
, &flags
);
3119 * Return accounted runtime for the task.
3120 * In case the task is currently running, return the runtime plus current's
3121 * pending runtime that have not been accounted yet.
3123 unsigned long long task_sched_runtime(struct task_struct
*p
)
3125 unsigned long flags
;
3129 rq
= task_rq_lock(p
, &flags
);
3130 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3131 task_rq_unlock(rq
, &flags
);
3137 * Return sum_exec_runtime for the thread group.
3138 * In case the task is currently running, return the sum plus current's
3139 * pending runtime that have not been accounted yet.
3141 * Note that the thread group might have other running tasks as well,
3142 * so the return value not includes other pending runtime that other
3143 * running tasks might have.
3145 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3147 struct task_cputime totals
;
3148 unsigned long flags
;
3152 rq
= task_rq_lock(p
, &flags
);
3153 thread_group_cputime(p
, &totals
);
3154 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3155 task_rq_unlock(rq
, &flags
);
3161 * Account user cpu time to a process.
3162 * @p: the process that the cpu time gets accounted to
3163 * @cputime: the cpu time spent in user space since the last update
3164 * @cputime_scaled: cputime scaled by cpu frequency
3166 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3167 cputime_t cputime_scaled
)
3169 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3172 /* Add user time to process. */
3173 p
->utime
= cputime_add(p
->utime
, cputime
);
3174 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3175 account_group_user_time(p
, cputime
);
3177 /* Add user time to cpustat. */
3178 tmp
= cputime_to_cputime64(cputime
);
3179 if (TASK_NICE(p
) > 0)
3180 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3182 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3184 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3185 /* Account for user time used */
3186 acct_update_integrals(p
);
3190 * Account guest cpu time to a process.
3191 * @p: the process that the cpu time gets accounted to
3192 * @cputime: the cpu time spent in virtual machine since the last update
3193 * @cputime_scaled: cputime scaled by cpu frequency
3195 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3196 cputime_t cputime_scaled
)
3199 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3201 tmp
= cputime_to_cputime64(cputime
);
3203 /* Add guest time to process. */
3204 p
->utime
= cputime_add(p
->utime
, cputime
);
3205 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3206 account_group_user_time(p
, cputime
);
3207 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3209 /* Add guest time to cpustat. */
3210 if (TASK_NICE(p
) > 0) {
3211 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3212 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3214 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3215 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3220 * Account system cpu time to a process.
3221 * @p: the process that the cpu time gets accounted to
3222 * @hardirq_offset: the offset to subtract from hardirq_count()
3223 * @cputime: the cpu time spent in kernel space since the last update
3224 * @cputime_scaled: cputime scaled by cpu frequency
3226 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3227 cputime_t cputime
, cputime_t cputime_scaled
)
3229 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3232 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3233 account_guest_time(p
, cputime
, cputime_scaled
);
3237 /* Add system time to process. */
3238 p
->stime
= cputime_add(p
->stime
, cputime
);
3239 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3240 account_group_system_time(p
, cputime
);
3242 /* Add system time to cpustat. */
3243 tmp
= cputime_to_cputime64(cputime
);
3244 if (hardirq_count() - hardirq_offset
)
3245 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3246 else if (softirq_count())
3247 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3249 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3251 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3253 /* Account for system time used */
3254 acct_update_integrals(p
);
3258 * Account for involuntary wait time.
3259 * @steal: the cpu time spent in involuntary wait
3261 void account_steal_time(cputime_t cputime
)
3263 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3264 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3266 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3270 * Account for idle time.
3271 * @cputime: the cpu time spent in idle wait
3273 void account_idle_time(cputime_t cputime
)
3275 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3276 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3277 struct rq
*rq
= this_rq();
3279 if (atomic_read(&rq
->nr_iowait
) > 0)
3280 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3282 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3285 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3288 * Account a single tick of cpu time.
3289 * @p: the process that the cpu time gets accounted to
3290 * @user_tick: indicates if the tick is a user or a system tick
3292 void account_process_tick(struct task_struct
*p
, int user_tick
)
3294 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3295 struct rq
*rq
= this_rq();
3298 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3299 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3300 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3303 account_idle_time(cputime_one_jiffy
);
3307 * Account multiple ticks of steal time.
3308 * @p: the process from which the cpu time has been stolen
3309 * @ticks: number of stolen ticks
3311 void account_steal_ticks(unsigned long ticks
)
3313 account_steal_time(jiffies_to_cputime(ticks
));
3317 * Account multiple ticks of idle time.
3318 * @ticks: number of stolen ticks
3320 void account_idle_ticks(unsigned long ticks
)
3322 account_idle_time(jiffies_to_cputime(ticks
));
3328 * Use precise platform statistics if available:
3330 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3331 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3337 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3339 struct task_cputime cputime
;
3341 thread_group_cputime(p
, &cputime
);
3343 *ut
= cputime
.utime
;
3344 *st
= cputime
.stime
;
3348 #ifndef nsecs_to_cputime
3349 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3352 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3354 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3357 * Use CFS's precise accounting:
3359 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3364 temp
= (u64
)(rtime
* utime
);
3365 do_div(temp
, total
);
3366 utime
= (cputime_t
)temp
;
3371 * Compare with previous values, to keep monotonicity:
3373 p
->prev_utime
= max(p
->prev_utime
, utime
);
3374 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3376 *ut
= p
->prev_utime
;
3377 *st
= p
->prev_stime
;
3381 * Must be called with siglock held.
3383 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3385 struct signal_struct
*sig
= p
->signal
;
3386 struct task_cputime cputime
;
3387 cputime_t rtime
, utime
, total
;
3389 thread_group_cputime(p
, &cputime
);
3391 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3392 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3397 temp
= (u64
)(rtime
* cputime
.utime
);
3398 do_div(temp
, total
);
3399 utime
= (cputime_t
)temp
;
3403 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3404 sig
->prev_stime
= max(sig
->prev_stime
,
3405 cputime_sub(rtime
, sig
->prev_utime
));
3407 *ut
= sig
->prev_utime
;
3408 *st
= sig
->prev_stime
;
3413 * This function gets called by the timer code, with HZ frequency.
3414 * We call it with interrupts disabled.
3416 * It also gets called by the fork code, when changing the parent's
3419 void scheduler_tick(void)
3421 int cpu
= smp_processor_id();
3422 struct rq
*rq
= cpu_rq(cpu
);
3423 struct task_struct
*curr
= rq
->curr
;
3427 raw_spin_lock(&rq
->lock
);
3428 update_rq_clock(rq
);
3429 update_cpu_load(rq
);
3430 curr
->sched_class
->task_tick(rq
, curr
, 0);
3431 raw_spin_unlock(&rq
->lock
);
3433 perf_event_task_tick(curr
);
3436 rq
->idle_at_tick
= idle_cpu(cpu
);
3437 trigger_load_balance(rq
, cpu
);
3441 notrace
unsigned long get_parent_ip(unsigned long addr
)
3443 if (in_lock_functions(addr
)) {
3444 addr
= CALLER_ADDR2
;
3445 if (in_lock_functions(addr
))
3446 addr
= CALLER_ADDR3
;
3451 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3452 defined(CONFIG_PREEMPT_TRACER))
3454 void __kprobes
add_preempt_count(int val
)
3456 #ifdef CONFIG_DEBUG_PREEMPT
3460 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3463 preempt_count() += val
;
3464 #ifdef CONFIG_DEBUG_PREEMPT
3466 * Spinlock count overflowing soon?
3468 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3471 if (preempt_count() == val
)
3472 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3474 EXPORT_SYMBOL(add_preempt_count
);
3476 void __kprobes
sub_preempt_count(int val
)
3478 #ifdef CONFIG_DEBUG_PREEMPT
3482 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3485 * Is the spinlock portion underflowing?
3487 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3488 !(preempt_count() & PREEMPT_MASK
)))
3492 if (preempt_count() == val
)
3493 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3494 preempt_count() -= val
;
3496 EXPORT_SYMBOL(sub_preempt_count
);
3501 * Print scheduling while atomic bug:
3503 static noinline
void __schedule_bug(struct task_struct
*prev
)
3505 struct pt_regs
*regs
= get_irq_regs();
3507 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3508 prev
->comm
, prev
->pid
, preempt_count());
3510 debug_show_held_locks(prev
);
3512 if (irqs_disabled())
3513 print_irqtrace_events(prev
);
3522 * Various schedule()-time debugging checks and statistics:
3524 static inline void schedule_debug(struct task_struct
*prev
)
3527 * Test if we are atomic. Since do_exit() needs to call into
3528 * schedule() atomically, we ignore that path for now.
3529 * Otherwise, whine if we are scheduling when we should not be.
3531 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3532 __schedule_bug(prev
);
3534 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3536 schedstat_inc(this_rq(), sched_count
);
3537 #ifdef CONFIG_SCHEDSTATS
3538 if (unlikely(prev
->lock_depth
>= 0)) {
3539 schedstat_inc(this_rq(), bkl_count
);
3540 schedstat_inc(prev
, sched_info
.bkl_count
);
3545 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3548 update_rq_clock(rq
);
3549 rq
->skip_clock_update
= 0;
3550 prev
->sched_class
->put_prev_task(rq
, prev
);
3554 * Pick up the highest-prio task:
3556 static inline struct task_struct
*
3557 pick_next_task(struct rq
*rq
)
3559 const struct sched_class
*class;
3560 struct task_struct
*p
;
3563 * Optimization: we know that if all tasks are in
3564 * the fair class we can call that function directly:
3566 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3567 p
= fair_sched_class
.pick_next_task(rq
);
3572 class = sched_class_highest
;
3574 p
= class->pick_next_task(rq
);
3578 * Will never be NULL as the idle class always
3579 * returns a non-NULL p:
3581 class = class->next
;
3586 * schedule() is the main scheduler function.
3588 asmlinkage
void __sched
schedule(void)
3590 struct task_struct
*prev
, *next
;
3591 unsigned long *switch_count
;
3597 cpu
= smp_processor_id();
3599 rcu_note_context_switch(cpu
);
3601 switch_count
= &prev
->nivcsw
;
3603 release_kernel_lock(prev
);
3604 need_resched_nonpreemptible
:
3606 schedule_debug(prev
);
3608 if (sched_feat(HRTICK
))
3611 raw_spin_lock_irq(&rq
->lock
);
3612 clear_tsk_need_resched(prev
);
3614 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3615 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3616 prev
->state
= TASK_RUNNING
;
3618 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3619 switch_count
= &prev
->nvcsw
;
3622 pre_schedule(rq
, prev
);
3624 if (unlikely(!rq
->nr_running
))
3625 idle_balance(cpu
, rq
);
3627 put_prev_task(rq
, prev
);
3628 next
= pick_next_task(rq
);
3630 if (likely(prev
!= next
)) {
3631 sched_info_switch(prev
, next
);
3632 perf_event_task_sched_out(prev
, next
);
3638 context_switch(rq
, prev
, next
); /* unlocks the rq */
3640 * the context switch might have flipped the stack from under
3641 * us, hence refresh the local variables.
3643 cpu
= smp_processor_id();
3646 raw_spin_unlock_irq(&rq
->lock
);
3650 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3652 switch_count
= &prev
->nivcsw
;
3653 goto need_resched_nonpreemptible
;
3656 preempt_enable_no_resched();
3660 EXPORT_SYMBOL(schedule
);
3662 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3664 * Look out! "owner" is an entirely speculative pointer
3665 * access and not reliable.
3667 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3672 if (!sched_feat(OWNER_SPIN
))
3675 #ifdef CONFIG_DEBUG_PAGEALLOC
3677 * Need to access the cpu field knowing that
3678 * DEBUG_PAGEALLOC could have unmapped it if
3679 * the mutex owner just released it and exited.
3681 if (probe_kernel_address(&owner
->cpu
, cpu
))
3688 * Even if the access succeeded (likely case),
3689 * the cpu field may no longer be valid.
3691 if (cpu
>= nr_cpumask_bits
)
3695 * We need to validate that we can do a
3696 * get_cpu() and that we have the percpu area.
3698 if (!cpu_online(cpu
))
3705 * Owner changed, break to re-assess state.
3707 if (lock
->owner
!= owner
)
3711 * Is that owner really running on that cpu?
3713 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3723 #ifdef CONFIG_PREEMPT
3725 * this is the entry point to schedule() from in-kernel preemption
3726 * off of preempt_enable. Kernel preemptions off return from interrupt
3727 * occur there and call schedule directly.
3729 asmlinkage
void __sched
preempt_schedule(void)
3731 struct thread_info
*ti
= current_thread_info();
3734 * If there is a non-zero preempt_count or interrupts are disabled,
3735 * we do not want to preempt the current task. Just return..
3737 if (likely(ti
->preempt_count
|| irqs_disabled()))
3741 add_preempt_count(PREEMPT_ACTIVE
);
3743 sub_preempt_count(PREEMPT_ACTIVE
);
3746 * Check again in case we missed a preemption opportunity
3747 * between schedule and now.
3750 } while (need_resched());
3752 EXPORT_SYMBOL(preempt_schedule
);
3755 * this is the entry point to schedule() from kernel preemption
3756 * off of irq context.
3757 * Note, that this is called and return with irqs disabled. This will
3758 * protect us against recursive calling from irq.
3760 asmlinkage
void __sched
preempt_schedule_irq(void)
3762 struct thread_info
*ti
= current_thread_info();
3764 /* Catch callers which need to be fixed */
3765 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3768 add_preempt_count(PREEMPT_ACTIVE
);
3771 local_irq_disable();
3772 sub_preempt_count(PREEMPT_ACTIVE
);
3775 * Check again in case we missed a preemption opportunity
3776 * between schedule and now.
3779 } while (need_resched());
3782 #endif /* CONFIG_PREEMPT */
3784 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3787 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3789 EXPORT_SYMBOL(default_wake_function
);
3792 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3793 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3794 * number) then we wake all the non-exclusive tasks and one exclusive task.
3796 * There are circumstances in which we can try to wake a task which has already
3797 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3798 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3800 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3801 int nr_exclusive
, int wake_flags
, void *key
)
3803 wait_queue_t
*curr
, *next
;
3805 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3806 unsigned flags
= curr
->flags
;
3808 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3809 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3815 * __wake_up - wake up threads blocked on a waitqueue.
3817 * @mode: which threads
3818 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3819 * @key: is directly passed to the wakeup function
3821 * It may be assumed that this function implies a write memory barrier before
3822 * changing the task state if and only if any tasks are woken up.
3824 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3825 int nr_exclusive
, void *key
)
3827 unsigned long flags
;
3829 spin_lock_irqsave(&q
->lock
, flags
);
3830 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3831 spin_unlock_irqrestore(&q
->lock
, flags
);
3833 EXPORT_SYMBOL(__wake_up
);
3836 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3838 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3840 __wake_up_common(q
, mode
, 1, 0, NULL
);
3842 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3844 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3846 __wake_up_common(q
, mode
, 1, 0, key
);
3850 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3852 * @mode: which threads
3853 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3854 * @key: opaque value to be passed to wakeup targets
3856 * The sync wakeup differs that the waker knows that it will schedule
3857 * away soon, so while the target thread will be woken up, it will not
3858 * be migrated to another CPU - ie. the two threads are 'synchronized'
3859 * with each other. This can prevent needless bouncing between CPUs.
3861 * On UP it can prevent extra preemption.
3863 * It may be assumed that this function implies a write memory barrier before
3864 * changing the task state if and only if any tasks are woken up.
3866 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3867 int nr_exclusive
, void *key
)
3869 unsigned long flags
;
3870 int wake_flags
= WF_SYNC
;
3875 if (unlikely(!nr_exclusive
))
3878 spin_lock_irqsave(&q
->lock
, flags
);
3879 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3880 spin_unlock_irqrestore(&q
->lock
, flags
);
3882 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3885 * __wake_up_sync - see __wake_up_sync_key()
3887 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3889 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3891 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3894 * complete: - signals a single thread waiting on this completion
3895 * @x: holds the state of this particular completion
3897 * This will wake up a single thread waiting on this completion. Threads will be
3898 * awakened in the same order in which they were queued.
3900 * See also complete_all(), wait_for_completion() and related routines.
3902 * It may be assumed that this function implies a write memory barrier before
3903 * changing the task state if and only if any tasks are woken up.
3905 void complete(struct completion
*x
)
3907 unsigned long flags
;
3909 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3911 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3912 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3914 EXPORT_SYMBOL(complete
);
3917 * complete_all: - signals all threads waiting on this completion
3918 * @x: holds the state of this particular completion
3920 * This will wake up all threads waiting on this particular completion event.
3922 * It may be assumed that this function implies a write memory barrier before
3923 * changing the task state if and only if any tasks are woken up.
3925 void complete_all(struct completion
*x
)
3927 unsigned long flags
;
3929 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3930 x
->done
+= UINT_MAX
/2;
3931 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3932 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3934 EXPORT_SYMBOL(complete_all
);
3936 static inline long __sched
3937 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3940 DECLARE_WAITQUEUE(wait
, current
);
3942 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3944 if (signal_pending_state(state
, current
)) {
3945 timeout
= -ERESTARTSYS
;
3948 __set_current_state(state
);
3949 spin_unlock_irq(&x
->wait
.lock
);
3950 timeout
= schedule_timeout(timeout
);
3951 spin_lock_irq(&x
->wait
.lock
);
3952 } while (!x
->done
&& timeout
);
3953 __remove_wait_queue(&x
->wait
, &wait
);
3958 return timeout
?: 1;
3962 wait_for_common(struct completion
*x
, long timeout
, int state
)
3966 spin_lock_irq(&x
->wait
.lock
);
3967 timeout
= do_wait_for_common(x
, timeout
, state
);
3968 spin_unlock_irq(&x
->wait
.lock
);
3973 * wait_for_completion: - waits for completion of a task
3974 * @x: holds the state of this particular completion
3976 * This waits to be signaled for completion of a specific task. It is NOT
3977 * interruptible and there is no timeout.
3979 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3980 * and interrupt capability. Also see complete().
3982 void __sched
wait_for_completion(struct completion
*x
)
3984 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3986 EXPORT_SYMBOL(wait_for_completion
);
3989 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3990 * @x: holds the state of this particular completion
3991 * @timeout: timeout value in jiffies
3993 * This waits for either a completion of a specific task to be signaled or for a
3994 * specified timeout to expire. The timeout is in jiffies. It is not
3997 unsigned long __sched
3998 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4000 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4002 EXPORT_SYMBOL(wait_for_completion_timeout
);
4005 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4006 * @x: holds the state of this particular completion
4008 * This waits for completion of a specific task to be signaled. It is
4011 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4013 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4014 if (t
== -ERESTARTSYS
)
4018 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4021 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4022 * @x: holds the state of this particular completion
4023 * @timeout: timeout value in jiffies
4025 * This waits for either a completion of a specific task to be signaled or for a
4026 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4028 unsigned long __sched
4029 wait_for_completion_interruptible_timeout(struct completion
*x
,
4030 unsigned long timeout
)
4032 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4034 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4037 * wait_for_completion_killable: - waits for completion of a task (killable)
4038 * @x: holds the state of this particular completion
4040 * This waits to be signaled for completion of a specific task. It can be
4041 * interrupted by a kill signal.
4043 int __sched
wait_for_completion_killable(struct completion
*x
)
4045 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4046 if (t
== -ERESTARTSYS
)
4050 EXPORT_SYMBOL(wait_for_completion_killable
);
4053 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4054 * @x: holds the state of this particular completion
4055 * @timeout: timeout value in jiffies
4057 * This waits for either a completion of a specific task to be
4058 * signaled or for a specified timeout to expire. It can be
4059 * interrupted by a kill signal. The timeout is in jiffies.
4061 unsigned long __sched
4062 wait_for_completion_killable_timeout(struct completion
*x
,
4063 unsigned long timeout
)
4065 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4067 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4070 * try_wait_for_completion - try to decrement a completion without blocking
4071 * @x: completion structure
4073 * Returns: 0 if a decrement cannot be done without blocking
4074 * 1 if a decrement succeeded.
4076 * If a completion is being used as a counting completion,
4077 * attempt to decrement the counter without blocking. This
4078 * enables us to avoid waiting if the resource the completion
4079 * is protecting is not available.
4081 bool try_wait_for_completion(struct completion
*x
)
4083 unsigned long flags
;
4086 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4091 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4094 EXPORT_SYMBOL(try_wait_for_completion
);
4097 * completion_done - Test to see if a completion has any waiters
4098 * @x: completion structure
4100 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4101 * 1 if there are no waiters.
4104 bool completion_done(struct completion
*x
)
4106 unsigned long flags
;
4109 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4112 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4115 EXPORT_SYMBOL(completion_done
);
4118 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4120 unsigned long flags
;
4123 init_waitqueue_entry(&wait
, current
);
4125 __set_current_state(state
);
4127 spin_lock_irqsave(&q
->lock
, flags
);
4128 __add_wait_queue(q
, &wait
);
4129 spin_unlock(&q
->lock
);
4130 timeout
= schedule_timeout(timeout
);
4131 spin_lock_irq(&q
->lock
);
4132 __remove_wait_queue(q
, &wait
);
4133 spin_unlock_irqrestore(&q
->lock
, flags
);
4138 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4140 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4142 EXPORT_SYMBOL(interruptible_sleep_on
);
4145 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4147 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4149 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4151 void __sched
sleep_on(wait_queue_head_t
*q
)
4153 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4155 EXPORT_SYMBOL(sleep_on
);
4157 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4159 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4161 EXPORT_SYMBOL(sleep_on_timeout
);
4163 #ifdef CONFIG_RT_MUTEXES
4166 * rt_mutex_setprio - set the current priority of a task
4168 * @prio: prio value (kernel-internal form)
4170 * This function changes the 'effective' priority of a task. It does
4171 * not touch ->normal_prio like __setscheduler().
4173 * Used by the rt_mutex code to implement priority inheritance logic.
4175 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4177 unsigned long flags
;
4178 int oldprio
, on_rq
, running
;
4180 const struct sched_class
*prev_class
;
4182 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4184 rq
= task_rq_lock(p
, &flags
);
4187 prev_class
= p
->sched_class
;
4188 on_rq
= p
->se
.on_rq
;
4189 running
= task_current(rq
, p
);
4191 dequeue_task(rq
, p
, 0);
4193 p
->sched_class
->put_prev_task(rq
, p
);
4196 p
->sched_class
= &rt_sched_class
;
4198 p
->sched_class
= &fair_sched_class
;
4203 p
->sched_class
->set_curr_task(rq
);
4205 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4207 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4209 task_rq_unlock(rq
, &flags
);
4214 void set_user_nice(struct task_struct
*p
, long nice
)
4216 int old_prio
, delta
, on_rq
;
4217 unsigned long flags
;
4220 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4223 * We have to be careful, if called from sys_setpriority(),
4224 * the task might be in the middle of scheduling on another CPU.
4226 rq
= task_rq_lock(p
, &flags
);
4228 * The RT priorities are set via sched_setscheduler(), but we still
4229 * allow the 'normal' nice value to be set - but as expected
4230 * it wont have any effect on scheduling until the task is
4231 * SCHED_FIFO/SCHED_RR:
4233 if (task_has_rt_policy(p
)) {
4234 p
->static_prio
= NICE_TO_PRIO(nice
);
4237 on_rq
= p
->se
.on_rq
;
4239 dequeue_task(rq
, p
, 0);
4241 p
->static_prio
= NICE_TO_PRIO(nice
);
4244 p
->prio
= effective_prio(p
);
4245 delta
= p
->prio
- old_prio
;
4248 enqueue_task(rq
, p
, 0);
4250 * If the task increased its priority or is running and
4251 * lowered its priority, then reschedule its CPU:
4253 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4254 resched_task(rq
->curr
);
4257 task_rq_unlock(rq
, &flags
);
4259 EXPORT_SYMBOL(set_user_nice
);
4262 * can_nice - check if a task can reduce its nice value
4266 int can_nice(const struct task_struct
*p
, const int nice
)
4268 /* convert nice value [19,-20] to rlimit style value [1,40] */
4269 int nice_rlim
= 20 - nice
;
4271 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4272 capable(CAP_SYS_NICE
));
4275 #ifdef __ARCH_WANT_SYS_NICE
4278 * sys_nice - change the priority of the current process.
4279 * @increment: priority increment
4281 * sys_setpriority is a more generic, but much slower function that
4282 * does similar things.
4284 SYSCALL_DEFINE1(nice
, int, increment
)
4289 * Setpriority might change our priority at the same moment.
4290 * We don't have to worry. Conceptually one call occurs first
4291 * and we have a single winner.
4293 if (increment
< -40)
4298 nice
= TASK_NICE(current
) + increment
;
4304 if (increment
< 0 && !can_nice(current
, nice
))
4307 retval
= security_task_setnice(current
, nice
);
4311 set_user_nice(current
, nice
);
4318 * task_prio - return the priority value of a given task.
4319 * @p: the task in question.
4321 * This is the priority value as seen by users in /proc.
4322 * RT tasks are offset by -200. Normal tasks are centered
4323 * around 0, value goes from -16 to +15.
4325 int task_prio(const struct task_struct
*p
)
4327 return p
->prio
- MAX_RT_PRIO
;
4331 * task_nice - return the nice value of a given task.
4332 * @p: the task in question.
4334 int task_nice(const struct task_struct
*p
)
4336 return TASK_NICE(p
);
4338 EXPORT_SYMBOL(task_nice
);
4341 * idle_cpu - is a given cpu idle currently?
4342 * @cpu: the processor in question.
4344 int idle_cpu(int cpu
)
4346 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4350 * idle_task - return the idle task for a given cpu.
4351 * @cpu: the processor in question.
4353 struct task_struct
*idle_task(int cpu
)
4355 return cpu_rq(cpu
)->idle
;
4359 * find_process_by_pid - find a process with a matching PID value.
4360 * @pid: the pid in question.
4362 static struct task_struct
*find_process_by_pid(pid_t pid
)
4364 return pid
? find_task_by_vpid(pid
) : current
;
4367 /* Actually do priority change: must hold rq lock. */
4369 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4371 BUG_ON(p
->se
.on_rq
);
4374 p
->rt_priority
= prio
;
4375 p
->normal_prio
= normal_prio(p
);
4376 /* we are holding p->pi_lock already */
4377 p
->prio
= rt_mutex_getprio(p
);
4378 if (rt_prio(p
->prio
))
4379 p
->sched_class
= &rt_sched_class
;
4381 p
->sched_class
= &fair_sched_class
;
4386 * check the target process has a UID that matches the current process's
4388 static bool check_same_owner(struct task_struct
*p
)
4390 const struct cred
*cred
= current_cred(), *pcred
;
4394 pcred
= __task_cred(p
);
4395 match
= (cred
->euid
== pcred
->euid
||
4396 cred
->euid
== pcred
->uid
);
4401 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4402 struct sched_param
*param
, bool user
)
4404 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4405 unsigned long flags
;
4406 const struct sched_class
*prev_class
;
4410 /* may grab non-irq protected spin_locks */
4411 BUG_ON(in_interrupt());
4413 /* double check policy once rq lock held */
4415 reset_on_fork
= p
->sched_reset_on_fork
;
4416 policy
= oldpolicy
= p
->policy
;
4418 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4419 policy
&= ~SCHED_RESET_ON_FORK
;
4421 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4422 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4423 policy
!= SCHED_IDLE
)
4428 * Valid priorities for SCHED_FIFO and SCHED_RR are
4429 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4430 * SCHED_BATCH and SCHED_IDLE is 0.
4432 if (param
->sched_priority
< 0 ||
4433 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4434 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4436 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4440 * Allow unprivileged RT tasks to decrease priority:
4442 if (user
&& !capable(CAP_SYS_NICE
)) {
4443 if (rt_policy(policy
)) {
4444 unsigned long rlim_rtprio
;
4446 if (!lock_task_sighand(p
, &flags
))
4448 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4449 unlock_task_sighand(p
, &flags
);
4451 /* can't set/change the rt policy */
4452 if (policy
!= p
->policy
&& !rlim_rtprio
)
4455 /* can't increase priority */
4456 if (param
->sched_priority
> p
->rt_priority
&&
4457 param
->sched_priority
> rlim_rtprio
)
4461 * Like positive nice levels, dont allow tasks to
4462 * move out of SCHED_IDLE either:
4464 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4467 /* can't change other user's priorities */
4468 if (!check_same_owner(p
))
4471 /* Normal users shall not reset the sched_reset_on_fork flag */
4472 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4477 #ifdef CONFIG_RT_GROUP_SCHED
4479 * Do not allow realtime tasks into groups that have no runtime
4482 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4483 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4487 retval
= security_task_setscheduler(p
, policy
, param
);
4493 * make sure no PI-waiters arrive (or leave) while we are
4494 * changing the priority of the task:
4496 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4498 * To be able to change p->policy safely, the apropriate
4499 * runqueue lock must be held.
4501 rq
= __task_rq_lock(p
);
4502 /* recheck policy now with rq lock held */
4503 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4504 policy
= oldpolicy
= -1;
4505 __task_rq_unlock(rq
);
4506 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4509 on_rq
= p
->se
.on_rq
;
4510 running
= task_current(rq
, p
);
4512 deactivate_task(rq
, p
, 0);
4514 p
->sched_class
->put_prev_task(rq
, p
);
4516 p
->sched_reset_on_fork
= reset_on_fork
;
4519 prev_class
= p
->sched_class
;
4520 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4523 p
->sched_class
->set_curr_task(rq
);
4525 activate_task(rq
, p
, 0);
4527 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4529 __task_rq_unlock(rq
);
4530 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4532 rt_mutex_adjust_pi(p
);
4538 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4539 * @p: the task in question.
4540 * @policy: new policy.
4541 * @param: structure containing the new RT priority.
4543 * NOTE that the task may be already dead.
4545 int sched_setscheduler(struct task_struct
*p
, int policy
,
4546 struct sched_param
*param
)
4548 return __sched_setscheduler(p
, policy
, param
, true);
4550 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4553 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4554 * @p: the task in question.
4555 * @policy: new policy.
4556 * @param: structure containing the new RT priority.
4558 * Just like sched_setscheduler, only don't bother checking if the
4559 * current context has permission. For example, this is needed in
4560 * stop_machine(): we create temporary high priority worker threads,
4561 * but our caller might not have that capability.
4563 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4564 struct sched_param
*param
)
4566 return __sched_setscheduler(p
, policy
, param
, false);
4570 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4572 struct sched_param lparam
;
4573 struct task_struct
*p
;
4576 if (!param
|| pid
< 0)
4578 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4583 p
= find_process_by_pid(pid
);
4585 retval
= sched_setscheduler(p
, policy
, &lparam
);
4592 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4593 * @pid: the pid in question.
4594 * @policy: new policy.
4595 * @param: structure containing the new RT priority.
4597 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4598 struct sched_param __user
*, param
)
4600 /* negative values for policy are not valid */
4604 return do_sched_setscheduler(pid
, policy
, param
);
4608 * sys_sched_setparam - set/change the RT priority of a thread
4609 * @pid: the pid in question.
4610 * @param: structure containing the new RT priority.
4612 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4614 return do_sched_setscheduler(pid
, -1, param
);
4618 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4619 * @pid: the pid in question.
4621 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4623 struct task_struct
*p
;
4631 p
= find_process_by_pid(pid
);
4633 retval
= security_task_getscheduler(p
);
4636 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4643 * sys_sched_getparam - get the RT priority of a thread
4644 * @pid: the pid in question.
4645 * @param: structure containing the RT priority.
4647 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4649 struct sched_param lp
;
4650 struct task_struct
*p
;
4653 if (!param
|| pid
< 0)
4657 p
= find_process_by_pid(pid
);
4662 retval
= security_task_getscheduler(p
);
4666 lp
.sched_priority
= p
->rt_priority
;
4670 * This one might sleep, we cannot do it with a spinlock held ...
4672 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4681 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4683 cpumask_var_t cpus_allowed
, new_mask
;
4684 struct task_struct
*p
;
4690 p
= find_process_by_pid(pid
);
4697 /* Prevent p going away */
4701 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4705 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4707 goto out_free_cpus_allowed
;
4710 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4713 retval
= security_task_setscheduler(p
, 0, NULL
);
4717 cpuset_cpus_allowed(p
, cpus_allowed
);
4718 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4720 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4723 cpuset_cpus_allowed(p
, cpus_allowed
);
4724 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4726 * We must have raced with a concurrent cpuset
4727 * update. Just reset the cpus_allowed to the
4728 * cpuset's cpus_allowed
4730 cpumask_copy(new_mask
, cpus_allowed
);
4735 free_cpumask_var(new_mask
);
4736 out_free_cpus_allowed
:
4737 free_cpumask_var(cpus_allowed
);
4744 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4745 struct cpumask
*new_mask
)
4747 if (len
< cpumask_size())
4748 cpumask_clear(new_mask
);
4749 else if (len
> cpumask_size())
4750 len
= cpumask_size();
4752 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4756 * sys_sched_setaffinity - set the cpu affinity of a process
4757 * @pid: pid of the process
4758 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4759 * @user_mask_ptr: user-space pointer to the new cpu mask
4761 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4762 unsigned long __user
*, user_mask_ptr
)
4764 cpumask_var_t new_mask
;
4767 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4770 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4772 retval
= sched_setaffinity(pid
, new_mask
);
4773 free_cpumask_var(new_mask
);
4777 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4779 struct task_struct
*p
;
4780 unsigned long flags
;
4788 p
= find_process_by_pid(pid
);
4792 retval
= security_task_getscheduler(p
);
4796 rq
= task_rq_lock(p
, &flags
);
4797 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4798 task_rq_unlock(rq
, &flags
);
4808 * sys_sched_getaffinity - get the cpu affinity of a process
4809 * @pid: pid of the process
4810 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4811 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4813 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4814 unsigned long __user
*, user_mask_ptr
)
4819 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4821 if (len
& (sizeof(unsigned long)-1))
4824 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4827 ret
= sched_getaffinity(pid
, mask
);
4829 size_t retlen
= min_t(size_t, len
, cpumask_size());
4831 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4836 free_cpumask_var(mask
);
4842 * sys_sched_yield - yield the current processor to other threads.
4844 * This function yields the current CPU to other tasks. If there are no
4845 * other threads running on this CPU then this function will return.
4847 SYSCALL_DEFINE0(sched_yield
)
4849 struct rq
*rq
= this_rq_lock();
4851 schedstat_inc(rq
, yld_count
);
4852 current
->sched_class
->yield_task(rq
);
4855 * Since we are going to call schedule() anyway, there's
4856 * no need to preempt or enable interrupts:
4858 __release(rq
->lock
);
4859 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4860 do_raw_spin_unlock(&rq
->lock
);
4861 preempt_enable_no_resched();
4868 static inline int should_resched(void)
4870 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4873 static void __cond_resched(void)
4875 add_preempt_count(PREEMPT_ACTIVE
);
4877 sub_preempt_count(PREEMPT_ACTIVE
);
4880 int __sched
_cond_resched(void)
4882 if (should_resched()) {
4888 EXPORT_SYMBOL(_cond_resched
);
4891 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4892 * call schedule, and on return reacquire the lock.
4894 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4895 * operations here to prevent schedule() from being called twice (once via
4896 * spin_unlock(), once by hand).
4898 int __cond_resched_lock(spinlock_t
*lock
)
4900 int resched
= should_resched();
4903 lockdep_assert_held(lock
);
4905 if (spin_needbreak(lock
) || resched
) {
4916 EXPORT_SYMBOL(__cond_resched_lock
);
4918 int __sched
__cond_resched_softirq(void)
4920 BUG_ON(!in_softirq());
4922 if (should_resched()) {
4930 EXPORT_SYMBOL(__cond_resched_softirq
);
4933 * yield - yield the current processor to other threads.
4935 * This is a shortcut for kernel-space yielding - it marks the
4936 * thread runnable and calls sys_sched_yield().
4938 void __sched
yield(void)
4940 set_current_state(TASK_RUNNING
);
4943 EXPORT_SYMBOL(yield
);
4946 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4947 * that process accounting knows that this is a task in IO wait state.
4949 void __sched
io_schedule(void)
4951 struct rq
*rq
= raw_rq();
4953 delayacct_blkio_start();
4954 atomic_inc(&rq
->nr_iowait
);
4955 current
->in_iowait
= 1;
4957 current
->in_iowait
= 0;
4958 atomic_dec(&rq
->nr_iowait
);
4959 delayacct_blkio_end();
4961 EXPORT_SYMBOL(io_schedule
);
4963 long __sched
io_schedule_timeout(long timeout
)
4965 struct rq
*rq
= raw_rq();
4968 delayacct_blkio_start();
4969 atomic_inc(&rq
->nr_iowait
);
4970 current
->in_iowait
= 1;
4971 ret
= schedule_timeout(timeout
);
4972 current
->in_iowait
= 0;
4973 atomic_dec(&rq
->nr_iowait
);
4974 delayacct_blkio_end();
4979 * sys_sched_get_priority_max - return maximum RT priority.
4980 * @policy: scheduling class.
4982 * this syscall returns the maximum rt_priority that can be used
4983 * by a given scheduling class.
4985 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4992 ret
= MAX_USER_RT_PRIO
-1;
5004 * sys_sched_get_priority_min - return minimum RT priority.
5005 * @policy: scheduling class.
5007 * this syscall returns the minimum rt_priority that can be used
5008 * by a given scheduling class.
5010 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5028 * sys_sched_rr_get_interval - return the default timeslice of a process.
5029 * @pid: pid of the process.
5030 * @interval: userspace pointer to the timeslice value.
5032 * this syscall writes the default timeslice value of a given process
5033 * into the user-space timespec buffer. A value of '0' means infinity.
5035 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5036 struct timespec __user
*, interval
)
5038 struct task_struct
*p
;
5039 unsigned int time_slice
;
5040 unsigned long flags
;
5050 p
= find_process_by_pid(pid
);
5054 retval
= security_task_getscheduler(p
);
5058 rq
= task_rq_lock(p
, &flags
);
5059 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5060 task_rq_unlock(rq
, &flags
);
5063 jiffies_to_timespec(time_slice
, &t
);
5064 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5072 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5074 void sched_show_task(struct task_struct
*p
)
5076 unsigned long free
= 0;
5079 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5080 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5081 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5082 #if BITS_PER_LONG == 32
5083 if (state
== TASK_RUNNING
)
5084 printk(KERN_CONT
" running ");
5086 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5088 if (state
== TASK_RUNNING
)
5089 printk(KERN_CONT
" running task ");
5091 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5093 #ifdef CONFIG_DEBUG_STACK_USAGE
5094 free
= stack_not_used(p
);
5096 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5097 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5098 (unsigned long)task_thread_info(p
)->flags
);
5100 show_stack(p
, NULL
);
5103 void show_state_filter(unsigned long state_filter
)
5105 struct task_struct
*g
, *p
;
5107 #if BITS_PER_LONG == 32
5109 " task PC stack pid father\n");
5112 " task PC stack pid father\n");
5114 read_lock(&tasklist_lock
);
5115 do_each_thread(g
, p
) {
5117 * reset the NMI-timeout, listing all files on a slow
5118 * console might take alot of time:
5120 touch_nmi_watchdog();
5121 if (!state_filter
|| (p
->state
& state_filter
))
5123 } while_each_thread(g
, p
);
5125 touch_all_softlockup_watchdogs();
5127 #ifdef CONFIG_SCHED_DEBUG
5128 sysrq_sched_debug_show();
5130 read_unlock(&tasklist_lock
);
5132 * Only show locks if all tasks are dumped:
5135 debug_show_all_locks();
5138 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5140 idle
->sched_class
= &idle_sched_class
;
5144 * init_idle - set up an idle thread for a given CPU
5145 * @idle: task in question
5146 * @cpu: cpu the idle task belongs to
5148 * NOTE: this function does not set the idle thread's NEED_RESCHED
5149 * flag, to make booting more robust.
5151 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5153 struct rq
*rq
= cpu_rq(cpu
);
5154 unsigned long flags
;
5156 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5159 idle
->state
= TASK_RUNNING
;
5160 idle
->se
.exec_start
= sched_clock();
5162 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5163 __set_task_cpu(idle
, cpu
);
5165 rq
->curr
= rq
->idle
= idle
;
5166 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5169 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5171 /* Set the preempt count _outside_ the spinlocks! */
5172 #if defined(CONFIG_PREEMPT)
5173 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5175 task_thread_info(idle
)->preempt_count
= 0;
5178 * The idle tasks have their own, simple scheduling class:
5180 idle
->sched_class
= &idle_sched_class
;
5181 ftrace_graph_init_task(idle
);
5185 * In a system that switches off the HZ timer nohz_cpu_mask
5186 * indicates which cpus entered this state. This is used
5187 * in the rcu update to wait only for active cpus. For system
5188 * which do not switch off the HZ timer nohz_cpu_mask should
5189 * always be CPU_BITS_NONE.
5191 cpumask_var_t nohz_cpu_mask
;
5194 * Increase the granularity value when there are more CPUs,
5195 * because with more CPUs the 'effective latency' as visible
5196 * to users decreases. But the relationship is not linear,
5197 * so pick a second-best guess by going with the log2 of the
5200 * This idea comes from the SD scheduler of Con Kolivas:
5202 static int get_update_sysctl_factor(void)
5204 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5205 unsigned int factor
;
5207 switch (sysctl_sched_tunable_scaling
) {
5208 case SCHED_TUNABLESCALING_NONE
:
5211 case SCHED_TUNABLESCALING_LINEAR
:
5214 case SCHED_TUNABLESCALING_LOG
:
5216 factor
= 1 + ilog2(cpus
);
5223 static void update_sysctl(void)
5225 unsigned int factor
= get_update_sysctl_factor();
5227 #define SET_SYSCTL(name) \
5228 (sysctl_##name = (factor) * normalized_sysctl_##name)
5229 SET_SYSCTL(sched_min_granularity
);
5230 SET_SYSCTL(sched_latency
);
5231 SET_SYSCTL(sched_wakeup_granularity
);
5232 SET_SYSCTL(sched_shares_ratelimit
);
5236 static inline void sched_init_granularity(void)
5243 * This is how migration works:
5245 * 1) we invoke migration_cpu_stop() on the target CPU using
5247 * 2) stopper starts to run (implicitly forcing the migrated thread
5249 * 3) it checks whether the migrated task is still in the wrong runqueue.
5250 * 4) if it's in the wrong runqueue then the migration thread removes
5251 * it and puts it into the right queue.
5252 * 5) stopper completes and stop_one_cpu() returns and the migration
5257 * Change a given task's CPU affinity. Migrate the thread to a
5258 * proper CPU and schedule it away if the CPU it's executing on
5259 * is removed from the allowed bitmask.
5261 * NOTE: the caller must have a valid reference to the task, the
5262 * task must not exit() & deallocate itself prematurely. The
5263 * call is not atomic; no spinlocks may be held.
5265 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5267 unsigned long flags
;
5269 unsigned int dest_cpu
;
5273 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5274 * drop the rq->lock and still rely on ->cpus_allowed.
5277 while (task_is_waking(p
))
5279 rq
= task_rq_lock(p
, &flags
);
5280 if (task_is_waking(p
)) {
5281 task_rq_unlock(rq
, &flags
);
5285 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5290 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5291 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5296 if (p
->sched_class
->set_cpus_allowed
)
5297 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5299 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5300 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5303 /* Can the task run on the task's current CPU? If so, we're done */
5304 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5307 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5308 if (migrate_task(p
, dest_cpu
)) {
5309 struct migration_arg arg
= { p
, dest_cpu
};
5310 /* Need help from migration thread: drop lock and wait. */
5311 task_rq_unlock(rq
, &flags
);
5312 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5313 tlb_migrate_finish(p
->mm
);
5317 task_rq_unlock(rq
, &flags
);
5321 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5324 * Move (not current) task off this cpu, onto dest cpu. We're doing
5325 * this because either it can't run here any more (set_cpus_allowed()
5326 * away from this CPU, or CPU going down), or because we're
5327 * attempting to rebalance this task on exec (sched_exec).
5329 * So we race with normal scheduler movements, but that's OK, as long
5330 * as the task is no longer on this CPU.
5332 * Returns non-zero if task was successfully migrated.
5334 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5336 struct rq
*rq_dest
, *rq_src
;
5339 if (unlikely(!cpu_active(dest_cpu
)))
5342 rq_src
= cpu_rq(src_cpu
);
5343 rq_dest
= cpu_rq(dest_cpu
);
5345 double_rq_lock(rq_src
, rq_dest
);
5346 /* Already moved. */
5347 if (task_cpu(p
) != src_cpu
)
5349 /* Affinity changed (again). */
5350 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5354 * If we're not on a rq, the next wake-up will ensure we're
5358 deactivate_task(rq_src
, p
, 0);
5359 set_task_cpu(p
, dest_cpu
);
5360 activate_task(rq_dest
, p
, 0);
5361 check_preempt_curr(rq_dest
, p
, 0);
5366 double_rq_unlock(rq_src
, rq_dest
);
5371 * migration_cpu_stop - this will be executed by a highprio stopper thread
5372 * and performs thread migration by bumping thread off CPU then
5373 * 'pushing' onto another runqueue.
5375 static int migration_cpu_stop(void *data
)
5377 struct migration_arg
*arg
= data
;
5380 * The original target cpu might have gone down and we might
5381 * be on another cpu but it doesn't matter.
5383 local_irq_disable();
5384 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5389 #ifdef CONFIG_HOTPLUG_CPU
5391 * Figure out where task on dead CPU should go, use force if necessary.
5393 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5395 struct rq
*rq
= cpu_rq(dead_cpu
);
5396 int needs_cpu
, uninitialized_var(dest_cpu
);
5397 unsigned long flags
;
5399 local_irq_save(flags
);
5401 raw_spin_lock(&rq
->lock
);
5402 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5404 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5405 raw_spin_unlock(&rq
->lock
);
5407 * It can only fail if we race with set_cpus_allowed(),
5408 * in the racer should migrate the task anyway.
5411 __migrate_task(p
, dead_cpu
, dest_cpu
);
5412 local_irq_restore(flags
);
5416 * While a dead CPU has no uninterruptible tasks queued at this point,
5417 * it might still have a nonzero ->nr_uninterruptible counter, because
5418 * for performance reasons the counter is not stricly tracking tasks to
5419 * their home CPUs. So we just add the counter to another CPU's counter,
5420 * to keep the global sum constant after CPU-down:
5422 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5424 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5425 unsigned long flags
;
5427 local_irq_save(flags
);
5428 double_rq_lock(rq_src
, rq_dest
);
5429 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5430 rq_src
->nr_uninterruptible
= 0;
5431 double_rq_unlock(rq_src
, rq_dest
);
5432 local_irq_restore(flags
);
5435 /* Run through task list and migrate tasks from the dead cpu. */
5436 static void migrate_live_tasks(int src_cpu
)
5438 struct task_struct
*p
, *t
;
5440 read_lock(&tasklist_lock
);
5442 do_each_thread(t
, p
) {
5446 if (task_cpu(p
) == src_cpu
)
5447 move_task_off_dead_cpu(src_cpu
, p
);
5448 } while_each_thread(t
, p
);
5450 read_unlock(&tasklist_lock
);
5454 * Schedules idle task to be the next runnable task on current CPU.
5455 * It does so by boosting its priority to highest possible.
5456 * Used by CPU offline code.
5458 void sched_idle_next(void)
5460 int this_cpu
= smp_processor_id();
5461 struct rq
*rq
= cpu_rq(this_cpu
);
5462 struct task_struct
*p
= rq
->idle
;
5463 unsigned long flags
;
5465 /* cpu has to be offline */
5466 BUG_ON(cpu_online(this_cpu
));
5469 * Strictly not necessary since rest of the CPUs are stopped by now
5470 * and interrupts disabled on the current cpu.
5472 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5474 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5476 activate_task(rq
, p
, 0);
5478 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5482 * Ensures that the idle task is using init_mm right before its cpu goes
5485 void idle_task_exit(void)
5487 struct mm_struct
*mm
= current
->active_mm
;
5489 BUG_ON(cpu_online(smp_processor_id()));
5492 switch_mm(mm
, &init_mm
, current
);
5496 /* called under rq->lock with disabled interrupts */
5497 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5499 struct rq
*rq
= cpu_rq(dead_cpu
);
5501 /* Must be exiting, otherwise would be on tasklist. */
5502 BUG_ON(!p
->exit_state
);
5504 /* Cannot have done final schedule yet: would have vanished. */
5505 BUG_ON(p
->state
== TASK_DEAD
);
5510 * Drop lock around migration; if someone else moves it,
5511 * that's OK. No task can be added to this CPU, so iteration is
5514 raw_spin_unlock_irq(&rq
->lock
);
5515 move_task_off_dead_cpu(dead_cpu
, p
);
5516 raw_spin_lock_irq(&rq
->lock
);
5521 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5522 static void migrate_dead_tasks(unsigned int dead_cpu
)
5524 struct rq
*rq
= cpu_rq(dead_cpu
);
5525 struct task_struct
*next
;
5528 if (!rq
->nr_running
)
5530 next
= pick_next_task(rq
);
5533 next
->sched_class
->put_prev_task(rq
, next
);
5534 migrate_dead(dead_cpu
, next
);
5540 * remove the tasks which were accounted by rq from calc_load_tasks.
5542 static void calc_global_load_remove(struct rq
*rq
)
5544 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5545 rq
->calc_load_active
= 0;
5547 #endif /* CONFIG_HOTPLUG_CPU */
5549 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5551 static struct ctl_table sd_ctl_dir
[] = {
5553 .procname
= "sched_domain",
5559 static struct ctl_table sd_ctl_root
[] = {
5561 .procname
= "kernel",
5563 .child
= sd_ctl_dir
,
5568 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5570 struct ctl_table
*entry
=
5571 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5576 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5578 struct ctl_table
*entry
;
5581 * In the intermediate directories, both the child directory and
5582 * procname are dynamically allocated and could fail but the mode
5583 * will always be set. In the lowest directory the names are
5584 * static strings and all have proc handlers.
5586 for (entry
= *tablep
; entry
->mode
; entry
++) {
5588 sd_free_ctl_entry(&entry
->child
);
5589 if (entry
->proc_handler
== NULL
)
5590 kfree(entry
->procname
);
5598 set_table_entry(struct ctl_table
*entry
,
5599 const char *procname
, void *data
, int maxlen
,
5600 mode_t mode
, proc_handler
*proc_handler
)
5602 entry
->procname
= procname
;
5604 entry
->maxlen
= maxlen
;
5606 entry
->proc_handler
= proc_handler
;
5609 static struct ctl_table
*
5610 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5612 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5617 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5618 sizeof(long), 0644, proc_doulongvec_minmax
);
5619 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5620 sizeof(long), 0644, proc_doulongvec_minmax
);
5621 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5622 sizeof(int), 0644, proc_dointvec_minmax
);
5623 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5624 sizeof(int), 0644, proc_dointvec_minmax
);
5625 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5626 sizeof(int), 0644, proc_dointvec_minmax
);
5627 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5628 sizeof(int), 0644, proc_dointvec_minmax
);
5629 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5630 sizeof(int), 0644, proc_dointvec_minmax
);
5631 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5632 sizeof(int), 0644, proc_dointvec_minmax
);
5633 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5634 sizeof(int), 0644, proc_dointvec_minmax
);
5635 set_table_entry(&table
[9], "cache_nice_tries",
5636 &sd
->cache_nice_tries
,
5637 sizeof(int), 0644, proc_dointvec_minmax
);
5638 set_table_entry(&table
[10], "flags", &sd
->flags
,
5639 sizeof(int), 0644, proc_dointvec_minmax
);
5640 set_table_entry(&table
[11], "name", sd
->name
,
5641 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5642 /* &table[12] is terminator */
5647 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5649 struct ctl_table
*entry
, *table
;
5650 struct sched_domain
*sd
;
5651 int domain_num
= 0, i
;
5654 for_each_domain(cpu
, sd
)
5656 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5661 for_each_domain(cpu
, sd
) {
5662 snprintf(buf
, 32, "domain%d", i
);
5663 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5665 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5672 static struct ctl_table_header
*sd_sysctl_header
;
5673 static void register_sched_domain_sysctl(void)
5675 int i
, cpu_num
= num_possible_cpus();
5676 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5679 WARN_ON(sd_ctl_dir
[0].child
);
5680 sd_ctl_dir
[0].child
= entry
;
5685 for_each_possible_cpu(i
) {
5686 snprintf(buf
, 32, "cpu%d", i
);
5687 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5689 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5693 WARN_ON(sd_sysctl_header
);
5694 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5697 /* may be called multiple times per register */
5698 static void unregister_sched_domain_sysctl(void)
5700 if (sd_sysctl_header
)
5701 unregister_sysctl_table(sd_sysctl_header
);
5702 sd_sysctl_header
= NULL
;
5703 if (sd_ctl_dir
[0].child
)
5704 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5707 static void register_sched_domain_sysctl(void)
5710 static void unregister_sched_domain_sysctl(void)
5715 static void set_rq_online(struct rq
*rq
)
5718 const struct sched_class
*class;
5720 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5723 for_each_class(class) {
5724 if (class->rq_online
)
5725 class->rq_online(rq
);
5730 static void set_rq_offline(struct rq
*rq
)
5733 const struct sched_class
*class;
5735 for_each_class(class) {
5736 if (class->rq_offline
)
5737 class->rq_offline(rq
);
5740 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5746 * migration_call - callback that gets triggered when a CPU is added.
5747 * Here we can start up the necessary migration thread for the new CPU.
5749 static int __cpuinit
5750 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5752 int cpu
= (long)hcpu
;
5753 unsigned long flags
;
5754 struct rq
*rq
= cpu_rq(cpu
);
5758 case CPU_UP_PREPARE
:
5759 case CPU_UP_PREPARE_FROZEN
:
5760 rq
->calc_load_update
= calc_load_update
;
5764 case CPU_ONLINE_FROZEN
:
5765 /* Update our root-domain */
5766 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5768 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5772 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5775 #ifdef CONFIG_HOTPLUG_CPU
5777 case CPU_DEAD_FROZEN
:
5778 migrate_live_tasks(cpu
);
5779 /* Idle task back to normal (off runqueue, low prio) */
5780 raw_spin_lock_irq(&rq
->lock
);
5781 deactivate_task(rq
, rq
->idle
, 0);
5782 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5783 rq
->idle
->sched_class
= &idle_sched_class
;
5784 migrate_dead_tasks(cpu
);
5785 raw_spin_unlock_irq(&rq
->lock
);
5786 migrate_nr_uninterruptible(rq
);
5787 BUG_ON(rq
->nr_running
!= 0);
5788 calc_global_load_remove(rq
);
5792 case CPU_DYING_FROZEN
:
5793 /* Update our root-domain */
5794 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5796 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5799 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5807 * Register at high priority so that task migration (migrate_all_tasks)
5808 * happens before everything else. This has to be lower priority than
5809 * the notifier in the perf_event subsystem, though.
5811 static struct notifier_block __cpuinitdata migration_notifier
= {
5812 .notifier_call
= migration_call
,
5816 static int __init
migration_init(void)
5818 void *cpu
= (void *)(long)smp_processor_id();
5821 /* Start one for the boot CPU: */
5822 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5823 BUG_ON(err
== NOTIFY_BAD
);
5824 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5825 register_cpu_notifier(&migration_notifier
);
5829 early_initcall(migration_init
);
5834 #ifdef CONFIG_SCHED_DEBUG
5836 static __read_mostly
int sched_domain_debug_enabled
;
5838 static int __init
sched_domain_debug_setup(char *str
)
5840 sched_domain_debug_enabled
= 1;
5844 early_param("sched_debug", sched_domain_debug_setup
);
5846 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5847 struct cpumask
*groupmask
)
5849 struct sched_group
*group
= sd
->groups
;
5852 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5853 cpumask_clear(groupmask
);
5855 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5857 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5858 printk("does not load-balance\n");
5860 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5865 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5867 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5868 printk(KERN_ERR
"ERROR: domain->span does not contain "
5871 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5872 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5876 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5880 printk(KERN_ERR
"ERROR: group is NULL\n");
5884 if (!group
->cpu_power
) {
5885 printk(KERN_CONT
"\n");
5886 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5891 if (!cpumask_weight(sched_group_cpus(group
))) {
5892 printk(KERN_CONT
"\n");
5893 printk(KERN_ERR
"ERROR: empty group\n");
5897 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5898 printk(KERN_CONT
"\n");
5899 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5903 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5905 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5907 printk(KERN_CONT
" %s", str
);
5908 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
5909 printk(KERN_CONT
" (cpu_power = %d)",
5913 group
= group
->next
;
5914 } while (group
!= sd
->groups
);
5915 printk(KERN_CONT
"\n");
5917 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5918 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5921 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5922 printk(KERN_ERR
"ERROR: parent span is not a superset "
5923 "of domain->span\n");
5927 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5929 cpumask_var_t groupmask
;
5932 if (!sched_domain_debug_enabled
)
5936 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5940 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5942 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
5943 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
5948 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
5955 free_cpumask_var(groupmask
);
5957 #else /* !CONFIG_SCHED_DEBUG */
5958 # define sched_domain_debug(sd, cpu) do { } while (0)
5959 #endif /* CONFIG_SCHED_DEBUG */
5961 static int sd_degenerate(struct sched_domain
*sd
)
5963 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5966 /* Following flags need at least 2 groups */
5967 if (sd
->flags
& (SD_LOAD_BALANCE
|
5968 SD_BALANCE_NEWIDLE
|
5972 SD_SHARE_PKG_RESOURCES
)) {
5973 if (sd
->groups
!= sd
->groups
->next
)
5977 /* Following flags don't use groups */
5978 if (sd
->flags
& (SD_WAKE_AFFINE
))
5985 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5987 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5989 if (sd_degenerate(parent
))
5992 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5995 /* Flags needing groups don't count if only 1 group in parent */
5996 if (parent
->groups
== parent
->groups
->next
) {
5997 pflags
&= ~(SD_LOAD_BALANCE
|
5998 SD_BALANCE_NEWIDLE
|
6002 SD_SHARE_PKG_RESOURCES
);
6003 if (nr_node_ids
== 1)
6004 pflags
&= ~SD_SERIALIZE
;
6006 if (~cflags
& pflags
)
6012 static void free_rootdomain(struct root_domain
*rd
)
6014 synchronize_sched();
6016 cpupri_cleanup(&rd
->cpupri
);
6018 free_cpumask_var(rd
->rto_mask
);
6019 free_cpumask_var(rd
->online
);
6020 free_cpumask_var(rd
->span
);
6024 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6026 struct root_domain
*old_rd
= NULL
;
6027 unsigned long flags
;
6029 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6034 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6037 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6040 * If we dont want to free the old_rt yet then
6041 * set old_rd to NULL to skip the freeing later
6044 if (!atomic_dec_and_test(&old_rd
->refcount
))
6048 atomic_inc(&rd
->refcount
);
6051 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6052 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6055 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6058 free_rootdomain(old_rd
);
6061 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6063 gfp_t gfp
= GFP_KERNEL
;
6065 memset(rd
, 0, sizeof(*rd
));
6070 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6072 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6074 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6077 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6082 free_cpumask_var(rd
->rto_mask
);
6084 free_cpumask_var(rd
->online
);
6086 free_cpumask_var(rd
->span
);
6091 static void init_defrootdomain(void)
6093 init_rootdomain(&def_root_domain
, true);
6095 atomic_set(&def_root_domain
.refcount
, 1);
6098 static struct root_domain
*alloc_rootdomain(void)
6100 struct root_domain
*rd
;
6102 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6106 if (init_rootdomain(rd
, false) != 0) {
6115 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6116 * hold the hotplug lock.
6119 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6121 struct rq
*rq
= cpu_rq(cpu
);
6122 struct sched_domain
*tmp
;
6124 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6125 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6127 /* Remove the sched domains which do not contribute to scheduling. */
6128 for (tmp
= sd
; tmp
; ) {
6129 struct sched_domain
*parent
= tmp
->parent
;
6133 if (sd_parent_degenerate(tmp
, parent
)) {
6134 tmp
->parent
= parent
->parent
;
6136 parent
->parent
->child
= tmp
;
6141 if (sd
&& sd_degenerate(sd
)) {
6147 sched_domain_debug(sd
, cpu
);
6149 rq_attach_root(rq
, rd
);
6150 rcu_assign_pointer(rq
->sd
, sd
);
6153 /* cpus with isolated domains */
6154 static cpumask_var_t cpu_isolated_map
;
6156 /* Setup the mask of cpus configured for isolated domains */
6157 static int __init
isolated_cpu_setup(char *str
)
6159 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6160 cpulist_parse(str
, cpu_isolated_map
);
6164 __setup("isolcpus=", isolated_cpu_setup
);
6167 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6168 * to a function which identifies what group(along with sched group) a CPU
6169 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6170 * (due to the fact that we keep track of groups covered with a struct cpumask).
6172 * init_sched_build_groups will build a circular linked list of the groups
6173 * covered by the given span, and will set each group's ->cpumask correctly,
6174 * and ->cpu_power to 0.
6177 init_sched_build_groups(const struct cpumask
*span
,
6178 const struct cpumask
*cpu_map
,
6179 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6180 struct sched_group
**sg
,
6181 struct cpumask
*tmpmask
),
6182 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6184 struct sched_group
*first
= NULL
, *last
= NULL
;
6187 cpumask_clear(covered
);
6189 for_each_cpu(i
, span
) {
6190 struct sched_group
*sg
;
6191 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6194 if (cpumask_test_cpu(i
, covered
))
6197 cpumask_clear(sched_group_cpus(sg
));
6200 for_each_cpu(j
, span
) {
6201 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6204 cpumask_set_cpu(j
, covered
);
6205 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6216 #define SD_NODES_PER_DOMAIN 16
6221 * find_next_best_node - find the next node to include in a sched_domain
6222 * @node: node whose sched_domain we're building
6223 * @used_nodes: nodes already in the sched_domain
6225 * Find the next node to include in a given scheduling domain. Simply
6226 * finds the closest node not already in the @used_nodes map.
6228 * Should use nodemask_t.
6230 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6232 int i
, n
, val
, min_val
, best_node
= 0;
6236 for (i
= 0; i
< nr_node_ids
; i
++) {
6237 /* Start at @node */
6238 n
= (node
+ i
) % nr_node_ids
;
6240 if (!nr_cpus_node(n
))
6243 /* Skip already used nodes */
6244 if (node_isset(n
, *used_nodes
))
6247 /* Simple min distance search */
6248 val
= node_distance(node
, n
);
6250 if (val
< min_val
) {
6256 node_set(best_node
, *used_nodes
);
6261 * sched_domain_node_span - get a cpumask for a node's sched_domain
6262 * @node: node whose cpumask we're constructing
6263 * @span: resulting cpumask
6265 * Given a node, construct a good cpumask for its sched_domain to span. It
6266 * should be one that prevents unnecessary balancing, but also spreads tasks
6269 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6271 nodemask_t used_nodes
;
6274 cpumask_clear(span
);
6275 nodes_clear(used_nodes
);
6277 cpumask_or(span
, span
, cpumask_of_node(node
));
6278 node_set(node
, used_nodes
);
6280 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6281 int next_node
= find_next_best_node(node
, &used_nodes
);
6283 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6286 #endif /* CONFIG_NUMA */
6288 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6291 * The cpus mask in sched_group and sched_domain hangs off the end.
6293 * ( See the the comments in include/linux/sched.h:struct sched_group
6294 * and struct sched_domain. )
6296 struct static_sched_group
{
6297 struct sched_group sg
;
6298 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6301 struct static_sched_domain
{
6302 struct sched_domain sd
;
6303 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6309 cpumask_var_t domainspan
;
6310 cpumask_var_t covered
;
6311 cpumask_var_t notcovered
;
6313 cpumask_var_t nodemask
;
6314 cpumask_var_t this_sibling_map
;
6315 cpumask_var_t this_core_map
;
6316 cpumask_var_t send_covered
;
6317 cpumask_var_t tmpmask
;
6318 struct sched_group
**sched_group_nodes
;
6319 struct root_domain
*rd
;
6323 sa_sched_groups
= 0,
6328 sa_this_sibling_map
,
6330 sa_sched_group_nodes
,
6340 * SMT sched-domains:
6342 #ifdef CONFIG_SCHED_SMT
6343 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6344 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6347 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6348 struct sched_group
**sg
, struct cpumask
*unused
)
6351 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6354 #endif /* CONFIG_SCHED_SMT */
6357 * multi-core sched-domains:
6359 #ifdef CONFIG_SCHED_MC
6360 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6361 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6362 #endif /* CONFIG_SCHED_MC */
6364 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6366 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6367 struct sched_group
**sg
, struct cpumask
*mask
)
6371 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6372 group
= cpumask_first(mask
);
6374 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6377 #elif defined(CONFIG_SCHED_MC)
6379 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6380 struct sched_group
**sg
, struct cpumask
*unused
)
6383 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6388 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6389 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6392 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6393 struct sched_group
**sg
, struct cpumask
*mask
)
6396 #ifdef CONFIG_SCHED_MC
6397 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6398 group
= cpumask_first(mask
);
6399 #elif defined(CONFIG_SCHED_SMT)
6400 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6401 group
= cpumask_first(mask
);
6406 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6412 * The init_sched_build_groups can't handle what we want to do with node
6413 * groups, so roll our own. Now each node has its own list of groups which
6414 * gets dynamically allocated.
6416 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6417 static struct sched_group
***sched_group_nodes_bycpu
;
6419 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6420 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6422 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6423 struct sched_group
**sg
,
6424 struct cpumask
*nodemask
)
6428 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6429 group
= cpumask_first(nodemask
);
6432 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6436 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6438 struct sched_group
*sg
= group_head
;
6444 for_each_cpu(j
, sched_group_cpus(sg
)) {
6445 struct sched_domain
*sd
;
6447 sd
= &per_cpu(phys_domains
, j
).sd
;
6448 if (j
!= group_first_cpu(sd
->groups
)) {
6450 * Only add "power" once for each
6456 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6459 } while (sg
!= group_head
);
6462 static int build_numa_sched_groups(struct s_data
*d
,
6463 const struct cpumask
*cpu_map
, int num
)
6465 struct sched_domain
*sd
;
6466 struct sched_group
*sg
, *prev
;
6469 cpumask_clear(d
->covered
);
6470 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6471 if (cpumask_empty(d
->nodemask
)) {
6472 d
->sched_group_nodes
[num
] = NULL
;
6476 sched_domain_node_span(num
, d
->domainspan
);
6477 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6479 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6482 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6486 d
->sched_group_nodes
[num
] = sg
;
6488 for_each_cpu(j
, d
->nodemask
) {
6489 sd
= &per_cpu(node_domains
, j
).sd
;
6494 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6496 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6499 for (j
= 0; j
< nr_node_ids
; j
++) {
6500 n
= (num
+ j
) % nr_node_ids
;
6501 cpumask_complement(d
->notcovered
, d
->covered
);
6502 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6503 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6504 if (cpumask_empty(d
->tmpmask
))
6506 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6507 if (cpumask_empty(d
->tmpmask
))
6509 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6513 "Can not alloc domain group for node %d\n", j
);
6517 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6518 sg
->next
= prev
->next
;
6519 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6526 #endif /* CONFIG_NUMA */
6529 /* Free memory allocated for various sched_group structures */
6530 static void free_sched_groups(const struct cpumask
*cpu_map
,
6531 struct cpumask
*nodemask
)
6535 for_each_cpu(cpu
, cpu_map
) {
6536 struct sched_group
**sched_group_nodes
6537 = sched_group_nodes_bycpu
[cpu
];
6539 if (!sched_group_nodes
)
6542 for (i
= 0; i
< nr_node_ids
; i
++) {
6543 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6545 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6546 if (cpumask_empty(nodemask
))
6556 if (oldsg
!= sched_group_nodes
[i
])
6559 kfree(sched_group_nodes
);
6560 sched_group_nodes_bycpu
[cpu
] = NULL
;
6563 #else /* !CONFIG_NUMA */
6564 static void free_sched_groups(const struct cpumask
*cpu_map
,
6565 struct cpumask
*nodemask
)
6568 #endif /* CONFIG_NUMA */
6571 * Initialize sched groups cpu_power.
6573 * cpu_power indicates the capacity of sched group, which is used while
6574 * distributing the load between different sched groups in a sched domain.
6575 * Typically cpu_power for all the groups in a sched domain will be same unless
6576 * there are asymmetries in the topology. If there are asymmetries, group
6577 * having more cpu_power will pickup more load compared to the group having
6580 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6582 struct sched_domain
*child
;
6583 struct sched_group
*group
;
6587 WARN_ON(!sd
|| !sd
->groups
);
6589 if (cpu
!= group_first_cpu(sd
->groups
))
6594 sd
->groups
->cpu_power
= 0;
6597 power
= SCHED_LOAD_SCALE
;
6598 weight
= cpumask_weight(sched_domain_span(sd
));
6600 * SMT siblings share the power of a single core.
6601 * Usually multiple threads get a better yield out of
6602 * that one core than a single thread would have,
6603 * reflect that in sd->smt_gain.
6605 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6606 power
*= sd
->smt_gain
;
6608 power
>>= SCHED_LOAD_SHIFT
;
6610 sd
->groups
->cpu_power
+= power
;
6615 * Add cpu_power of each child group to this groups cpu_power.
6617 group
= child
->groups
;
6619 sd
->groups
->cpu_power
+= group
->cpu_power
;
6620 group
= group
->next
;
6621 } while (group
!= child
->groups
);
6625 * Initializers for schedule domains
6626 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6629 #ifdef CONFIG_SCHED_DEBUG
6630 # define SD_INIT_NAME(sd, type) sd->name = #type
6632 # define SD_INIT_NAME(sd, type) do { } while (0)
6635 #define SD_INIT(sd, type) sd_init_##type(sd)
6637 #define SD_INIT_FUNC(type) \
6638 static noinline void sd_init_##type(struct sched_domain *sd) \
6640 memset(sd, 0, sizeof(*sd)); \
6641 *sd = SD_##type##_INIT; \
6642 sd->level = SD_LV_##type; \
6643 SD_INIT_NAME(sd, type); \
6648 SD_INIT_FUNC(ALLNODES
)
6651 #ifdef CONFIG_SCHED_SMT
6652 SD_INIT_FUNC(SIBLING
)
6654 #ifdef CONFIG_SCHED_MC
6658 static int default_relax_domain_level
= -1;
6660 static int __init
setup_relax_domain_level(char *str
)
6664 val
= simple_strtoul(str
, NULL
, 0);
6665 if (val
< SD_LV_MAX
)
6666 default_relax_domain_level
= val
;
6670 __setup("relax_domain_level=", setup_relax_domain_level
);
6672 static void set_domain_attribute(struct sched_domain
*sd
,
6673 struct sched_domain_attr
*attr
)
6677 if (!attr
|| attr
->relax_domain_level
< 0) {
6678 if (default_relax_domain_level
< 0)
6681 request
= default_relax_domain_level
;
6683 request
= attr
->relax_domain_level
;
6684 if (request
< sd
->level
) {
6685 /* turn off idle balance on this domain */
6686 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6688 /* turn on idle balance on this domain */
6689 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6693 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6694 const struct cpumask
*cpu_map
)
6697 case sa_sched_groups
:
6698 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6699 d
->sched_group_nodes
= NULL
;
6701 free_rootdomain(d
->rd
); /* fall through */
6703 free_cpumask_var(d
->tmpmask
); /* fall through */
6704 case sa_send_covered
:
6705 free_cpumask_var(d
->send_covered
); /* fall through */
6706 case sa_this_core_map
:
6707 free_cpumask_var(d
->this_core_map
); /* fall through */
6708 case sa_this_sibling_map
:
6709 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6711 free_cpumask_var(d
->nodemask
); /* fall through */
6712 case sa_sched_group_nodes
:
6714 kfree(d
->sched_group_nodes
); /* fall through */
6716 free_cpumask_var(d
->notcovered
); /* fall through */
6718 free_cpumask_var(d
->covered
); /* fall through */
6720 free_cpumask_var(d
->domainspan
); /* fall through */
6727 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6728 const struct cpumask
*cpu_map
)
6731 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6733 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6734 return sa_domainspan
;
6735 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6737 /* Allocate the per-node list of sched groups */
6738 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6739 sizeof(struct sched_group
*), GFP_KERNEL
);
6740 if (!d
->sched_group_nodes
) {
6741 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6742 return sa_notcovered
;
6744 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6746 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6747 return sa_sched_group_nodes
;
6748 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6750 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6751 return sa_this_sibling_map
;
6752 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6753 return sa_this_core_map
;
6754 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6755 return sa_send_covered
;
6756 d
->rd
= alloc_rootdomain();
6758 printk(KERN_WARNING
"Cannot alloc root domain\n");
6761 return sa_rootdomain
;
6764 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6765 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6767 struct sched_domain
*sd
= NULL
;
6769 struct sched_domain
*parent
;
6772 if (cpumask_weight(cpu_map
) >
6773 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6774 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6775 SD_INIT(sd
, ALLNODES
);
6776 set_domain_attribute(sd
, attr
);
6777 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6778 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6783 sd
= &per_cpu(node_domains
, i
).sd
;
6785 set_domain_attribute(sd
, attr
);
6786 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6787 sd
->parent
= parent
;
6790 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6795 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6796 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6797 struct sched_domain
*parent
, int i
)
6799 struct sched_domain
*sd
;
6800 sd
= &per_cpu(phys_domains
, i
).sd
;
6802 set_domain_attribute(sd
, attr
);
6803 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6804 sd
->parent
= parent
;
6807 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6811 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6812 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6813 struct sched_domain
*parent
, int i
)
6815 struct sched_domain
*sd
= parent
;
6816 #ifdef CONFIG_SCHED_MC
6817 sd
= &per_cpu(core_domains
, i
).sd
;
6819 set_domain_attribute(sd
, attr
);
6820 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6821 sd
->parent
= parent
;
6823 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6828 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6829 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6830 struct sched_domain
*parent
, int i
)
6832 struct sched_domain
*sd
= parent
;
6833 #ifdef CONFIG_SCHED_SMT
6834 sd
= &per_cpu(cpu_domains
, i
).sd
;
6835 SD_INIT(sd
, SIBLING
);
6836 set_domain_attribute(sd
, attr
);
6837 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6838 sd
->parent
= parent
;
6840 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6845 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6846 const struct cpumask
*cpu_map
, int cpu
)
6849 #ifdef CONFIG_SCHED_SMT
6850 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6851 cpumask_and(d
->this_sibling_map
, cpu_map
,
6852 topology_thread_cpumask(cpu
));
6853 if (cpu
== cpumask_first(d
->this_sibling_map
))
6854 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
6856 d
->send_covered
, d
->tmpmask
);
6859 #ifdef CONFIG_SCHED_MC
6860 case SD_LV_MC
: /* set up multi-core groups */
6861 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
6862 if (cpu
== cpumask_first(d
->this_core_map
))
6863 init_sched_build_groups(d
->this_core_map
, cpu_map
,
6865 d
->send_covered
, d
->tmpmask
);
6868 case SD_LV_CPU
: /* set up physical groups */
6869 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
6870 if (!cpumask_empty(d
->nodemask
))
6871 init_sched_build_groups(d
->nodemask
, cpu_map
,
6873 d
->send_covered
, d
->tmpmask
);
6876 case SD_LV_ALLNODES
:
6877 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
6878 d
->send_covered
, d
->tmpmask
);
6887 * Build sched domains for a given set of cpus and attach the sched domains
6888 * to the individual cpus
6890 static int __build_sched_domains(const struct cpumask
*cpu_map
,
6891 struct sched_domain_attr
*attr
)
6893 enum s_alloc alloc_state
= sa_none
;
6895 struct sched_domain
*sd
;
6901 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6902 if (alloc_state
!= sa_rootdomain
)
6904 alloc_state
= sa_sched_groups
;
6907 * Set up domains for cpus specified by the cpu_map.
6909 for_each_cpu(i
, cpu_map
) {
6910 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
6913 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
6914 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6915 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6916 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6919 for_each_cpu(i
, cpu_map
) {
6920 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
6921 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
6924 /* Set up physical groups */
6925 for (i
= 0; i
< nr_node_ids
; i
++)
6926 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
6929 /* Set up node groups */
6931 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
6933 for (i
= 0; i
< nr_node_ids
; i
++)
6934 if (build_numa_sched_groups(&d
, cpu_map
, i
))
6938 /* Calculate CPU power for physical packages and nodes */
6939 #ifdef CONFIG_SCHED_SMT
6940 for_each_cpu(i
, cpu_map
) {
6941 sd
= &per_cpu(cpu_domains
, i
).sd
;
6942 init_sched_groups_power(i
, sd
);
6945 #ifdef CONFIG_SCHED_MC
6946 for_each_cpu(i
, cpu_map
) {
6947 sd
= &per_cpu(core_domains
, i
).sd
;
6948 init_sched_groups_power(i
, sd
);
6952 for_each_cpu(i
, cpu_map
) {
6953 sd
= &per_cpu(phys_domains
, i
).sd
;
6954 init_sched_groups_power(i
, sd
);
6958 for (i
= 0; i
< nr_node_ids
; i
++)
6959 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
6961 if (d
.sd_allnodes
) {
6962 struct sched_group
*sg
;
6964 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
6966 init_numa_sched_groups_power(sg
);
6970 /* Attach the domains */
6971 for_each_cpu(i
, cpu_map
) {
6972 #ifdef CONFIG_SCHED_SMT
6973 sd
= &per_cpu(cpu_domains
, i
).sd
;
6974 #elif defined(CONFIG_SCHED_MC)
6975 sd
= &per_cpu(core_domains
, i
).sd
;
6977 sd
= &per_cpu(phys_domains
, i
).sd
;
6979 cpu_attach_domain(sd
, d
.rd
, i
);
6982 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
6983 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
6987 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6991 static int build_sched_domains(const struct cpumask
*cpu_map
)
6993 return __build_sched_domains(cpu_map
, NULL
);
6996 static cpumask_var_t
*doms_cur
; /* current sched domains */
6997 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6998 static struct sched_domain_attr
*dattr_cur
;
6999 /* attribues of custom domains in 'doms_cur' */
7002 * Special case: If a kmalloc of a doms_cur partition (array of
7003 * cpumask) fails, then fallback to a single sched domain,
7004 * as determined by the single cpumask fallback_doms.
7006 static cpumask_var_t fallback_doms
;
7009 * arch_update_cpu_topology lets virtualized architectures update the
7010 * cpu core maps. It is supposed to return 1 if the topology changed
7011 * or 0 if it stayed the same.
7013 int __attribute__((weak
)) arch_update_cpu_topology(void)
7018 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7021 cpumask_var_t
*doms
;
7023 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7026 for (i
= 0; i
< ndoms
; i
++) {
7027 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7028 free_sched_domains(doms
, i
);
7035 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7038 for (i
= 0; i
< ndoms
; i
++)
7039 free_cpumask_var(doms
[i
]);
7044 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7045 * For now this just excludes isolated cpus, but could be used to
7046 * exclude other special cases in the future.
7048 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7052 arch_update_cpu_topology();
7054 doms_cur
= alloc_sched_domains(ndoms_cur
);
7056 doms_cur
= &fallback_doms
;
7057 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7059 err
= build_sched_domains(doms_cur
[0]);
7060 register_sched_domain_sysctl();
7065 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7066 struct cpumask
*tmpmask
)
7068 free_sched_groups(cpu_map
, tmpmask
);
7072 * Detach sched domains from a group of cpus specified in cpu_map
7073 * These cpus will now be attached to the NULL domain
7075 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7077 /* Save because hotplug lock held. */
7078 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7081 for_each_cpu(i
, cpu_map
)
7082 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7083 synchronize_sched();
7084 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7087 /* handle null as "default" */
7088 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7089 struct sched_domain_attr
*new, int idx_new
)
7091 struct sched_domain_attr tmp
;
7098 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7099 new ? (new + idx_new
) : &tmp
,
7100 sizeof(struct sched_domain_attr
));
7104 * Partition sched domains as specified by the 'ndoms_new'
7105 * cpumasks in the array doms_new[] of cpumasks. This compares
7106 * doms_new[] to the current sched domain partitioning, doms_cur[].
7107 * It destroys each deleted domain and builds each new domain.
7109 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7110 * The masks don't intersect (don't overlap.) We should setup one
7111 * sched domain for each mask. CPUs not in any of the cpumasks will
7112 * not be load balanced. If the same cpumask appears both in the
7113 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7116 * The passed in 'doms_new' should be allocated using
7117 * alloc_sched_domains. This routine takes ownership of it and will
7118 * free_sched_domains it when done with it. If the caller failed the
7119 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7120 * and partition_sched_domains() will fallback to the single partition
7121 * 'fallback_doms', it also forces the domains to be rebuilt.
7123 * If doms_new == NULL it will be replaced with cpu_online_mask.
7124 * ndoms_new == 0 is a special case for destroying existing domains,
7125 * and it will not create the default domain.
7127 * Call with hotplug lock held
7129 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7130 struct sched_domain_attr
*dattr_new
)
7135 mutex_lock(&sched_domains_mutex
);
7137 /* always unregister in case we don't destroy any domains */
7138 unregister_sched_domain_sysctl();
7140 /* Let architecture update cpu core mappings. */
7141 new_topology
= arch_update_cpu_topology();
7143 n
= doms_new
? ndoms_new
: 0;
7145 /* Destroy deleted domains */
7146 for (i
= 0; i
< ndoms_cur
; i
++) {
7147 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7148 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7149 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7152 /* no match - a current sched domain not in new doms_new[] */
7153 detach_destroy_domains(doms_cur
[i
]);
7158 if (doms_new
== NULL
) {
7160 doms_new
= &fallback_doms
;
7161 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7162 WARN_ON_ONCE(dattr_new
);
7165 /* Build new domains */
7166 for (i
= 0; i
< ndoms_new
; i
++) {
7167 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7168 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7169 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7172 /* no match - add a new doms_new */
7173 __build_sched_domains(doms_new
[i
],
7174 dattr_new
? dattr_new
+ i
: NULL
);
7179 /* Remember the new sched domains */
7180 if (doms_cur
!= &fallback_doms
)
7181 free_sched_domains(doms_cur
, ndoms_cur
);
7182 kfree(dattr_cur
); /* kfree(NULL) is safe */
7183 doms_cur
= doms_new
;
7184 dattr_cur
= dattr_new
;
7185 ndoms_cur
= ndoms_new
;
7187 register_sched_domain_sysctl();
7189 mutex_unlock(&sched_domains_mutex
);
7192 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7193 static void arch_reinit_sched_domains(void)
7197 /* Destroy domains first to force the rebuild */
7198 partition_sched_domains(0, NULL
, NULL
);
7200 rebuild_sched_domains();
7204 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7206 unsigned int level
= 0;
7208 if (sscanf(buf
, "%u", &level
) != 1)
7212 * level is always be positive so don't check for
7213 * level < POWERSAVINGS_BALANCE_NONE which is 0
7214 * What happens on 0 or 1 byte write,
7215 * need to check for count as well?
7218 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7222 sched_smt_power_savings
= level
;
7224 sched_mc_power_savings
= level
;
7226 arch_reinit_sched_domains();
7231 #ifdef CONFIG_SCHED_MC
7232 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7233 struct sysdev_class_attribute
*attr
,
7236 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7238 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7239 struct sysdev_class_attribute
*attr
,
7240 const char *buf
, size_t count
)
7242 return sched_power_savings_store(buf
, count
, 0);
7244 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7245 sched_mc_power_savings_show
,
7246 sched_mc_power_savings_store
);
7249 #ifdef CONFIG_SCHED_SMT
7250 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7251 struct sysdev_class_attribute
*attr
,
7254 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7256 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7257 struct sysdev_class_attribute
*attr
,
7258 const char *buf
, size_t count
)
7260 return sched_power_savings_store(buf
, count
, 1);
7262 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7263 sched_smt_power_savings_show
,
7264 sched_smt_power_savings_store
);
7267 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7271 #ifdef CONFIG_SCHED_SMT
7273 err
= sysfs_create_file(&cls
->kset
.kobj
,
7274 &attr_sched_smt_power_savings
.attr
);
7276 #ifdef CONFIG_SCHED_MC
7277 if (!err
&& mc_capable())
7278 err
= sysfs_create_file(&cls
->kset
.kobj
,
7279 &attr_sched_mc_power_savings
.attr
);
7283 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7285 #ifndef CONFIG_CPUSETS
7287 * Add online and remove offline CPUs from the scheduler domains.
7288 * When cpusets are enabled they take over this function.
7290 static int update_sched_domains(struct notifier_block
*nfb
,
7291 unsigned long action
, void *hcpu
)
7295 case CPU_ONLINE_FROZEN
:
7296 case CPU_DOWN_PREPARE
:
7297 case CPU_DOWN_PREPARE_FROZEN
:
7298 case CPU_DOWN_FAILED
:
7299 case CPU_DOWN_FAILED_FROZEN
:
7300 partition_sched_domains(1, NULL
, NULL
);
7309 static int update_runtime(struct notifier_block
*nfb
,
7310 unsigned long action
, void *hcpu
)
7312 int cpu
= (int)(long)hcpu
;
7315 case CPU_DOWN_PREPARE
:
7316 case CPU_DOWN_PREPARE_FROZEN
:
7317 disable_runtime(cpu_rq(cpu
));
7320 case CPU_DOWN_FAILED
:
7321 case CPU_DOWN_FAILED_FROZEN
:
7323 case CPU_ONLINE_FROZEN
:
7324 enable_runtime(cpu_rq(cpu
));
7332 void __init
sched_init_smp(void)
7334 cpumask_var_t non_isolated_cpus
;
7336 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7337 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7339 #if defined(CONFIG_NUMA)
7340 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7342 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7345 mutex_lock(&sched_domains_mutex
);
7346 arch_init_sched_domains(cpu_active_mask
);
7347 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7348 if (cpumask_empty(non_isolated_cpus
))
7349 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7350 mutex_unlock(&sched_domains_mutex
);
7353 #ifndef CONFIG_CPUSETS
7354 /* XXX: Theoretical race here - CPU may be hotplugged now */
7355 hotcpu_notifier(update_sched_domains
, 0);
7358 /* RT runtime code needs to handle some hotplug events */
7359 hotcpu_notifier(update_runtime
, 0);
7363 /* Move init over to a non-isolated CPU */
7364 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7366 sched_init_granularity();
7367 free_cpumask_var(non_isolated_cpus
);
7369 init_sched_rt_class();
7372 void __init
sched_init_smp(void)
7374 sched_init_granularity();
7376 #endif /* CONFIG_SMP */
7378 const_debug
unsigned int sysctl_timer_migration
= 1;
7380 int in_sched_functions(unsigned long addr
)
7382 return in_lock_functions(addr
) ||
7383 (addr
>= (unsigned long)__sched_text_start
7384 && addr
< (unsigned long)__sched_text_end
);
7387 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7389 cfs_rq
->tasks_timeline
= RB_ROOT
;
7390 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7391 #ifdef CONFIG_FAIR_GROUP_SCHED
7394 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7397 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7399 struct rt_prio_array
*array
;
7402 array
= &rt_rq
->active
;
7403 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7404 INIT_LIST_HEAD(array
->queue
+ i
);
7405 __clear_bit(i
, array
->bitmap
);
7407 /* delimiter for bitsearch: */
7408 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7410 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7411 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7413 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7417 rt_rq
->rt_nr_migratory
= 0;
7418 rt_rq
->overloaded
= 0;
7419 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7423 rt_rq
->rt_throttled
= 0;
7424 rt_rq
->rt_runtime
= 0;
7425 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7427 #ifdef CONFIG_RT_GROUP_SCHED
7428 rt_rq
->rt_nr_boosted
= 0;
7433 #ifdef CONFIG_FAIR_GROUP_SCHED
7434 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7435 struct sched_entity
*se
, int cpu
, int add
,
7436 struct sched_entity
*parent
)
7438 struct rq
*rq
= cpu_rq(cpu
);
7439 tg
->cfs_rq
[cpu
] = cfs_rq
;
7440 init_cfs_rq(cfs_rq
, rq
);
7443 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7446 /* se could be NULL for init_task_group */
7451 se
->cfs_rq
= &rq
->cfs
;
7453 se
->cfs_rq
= parent
->my_q
;
7456 se
->load
.weight
= tg
->shares
;
7457 se
->load
.inv_weight
= 0;
7458 se
->parent
= parent
;
7462 #ifdef CONFIG_RT_GROUP_SCHED
7463 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7464 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7465 struct sched_rt_entity
*parent
)
7467 struct rq
*rq
= cpu_rq(cpu
);
7469 tg
->rt_rq
[cpu
] = rt_rq
;
7470 init_rt_rq(rt_rq
, rq
);
7472 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7474 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7476 tg
->rt_se
[cpu
] = rt_se
;
7481 rt_se
->rt_rq
= &rq
->rt
;
7483 rt_se
->rt_rq
= parent
->my_q
;
7485 rt_se
->my_q
= rt_rq
;
7486 rt_se
->parent
= parent
;
7487 INIT_LIST_HEAD(&rt_se
->run_list
);
7491 void __init
sched_init(void)
7494 unsigned long alloc_size
= 0, ptr
;
7496 #ifdef CONFIG_FAIR_GROUP_SCHED
7497 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7499 #ifdef CONFIG_RT_GROUP_SCHED
7500 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7502 #ifdef CONFIG_CPUMASK_OFFSTACK
7503 alloc_size
+= num_possible_cpus() * cpumask_size();
7506 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7508 #ifdef CONFIG_FAIR_GROUP_SCHED
7509 init_task_group
.se
= (struct sched_entity
**)ptr
;
7510 ptr
+= nr_cpu_ids
* sizeof(void **);
7512 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7513 ptr
+= nr_cpu_ids
* sizeof(void **);
7515 #endif /* CONFIG_FAIR_GROUP_SCHED */
7516 #ifdef CONFIG_RT_GROUP_SCHED
7517 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7518 ptr
+= nr_cpu_ids
* sizeof(void **);
7520 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7521 ptr
+= nr_cpu_ids
* sizeof(void **);
7523 #endif /* CONFIG_RT_GROUP_SCHED */
7524 #ifdef CONFIG_CPUMASK_OFFSTACK
7525 for_each_possible_cpu(i
) {
7526 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7527 ptr
+= cpumask_size();
7529 #endif /* CONFIG_CPUMASK_OFFSTACK */
7533 init_defrootdomain();
7536 init_rt_bandwidth(&def_rt_bandwidth
,
7537 global_rt_period(), global_rt_runtime());
7539 #ifdef CONFIG_RT_GROUP_SCHED
7540 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7541 global_rt_period(), global_rt_runtime());
7542 #endif /* CONFIG_RT_GROUP_SCHED */
7544 #ifdef CONFIG_CGROUP_SCHED
7545 list_add(&init_task_group
.list
, &task_groups
);
7546 INIT_LIST_HEAD(&init_task_group
.children
);
7548 #endif /* CONFIG_CGROUP_SCHED */
7550 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7551 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7552 __alignof__(unsigned long));
7554 for_each_possible_cpu(i
) {
7558 raw_spin_lock_init(&rq
->lock
);
7560 rq
->calc_load_active
= 0;
7561 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7562 init_cfs_rq(&rq
->cfs
, rq
);
7563 init_rt_rq(&rq
->rt
, rq
);
7564 #ifdef CONFIG_FAIR_GROUP_SCHED
7565 init_task_group
.shares
= init_task_group_load
;
7566 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7567 #ifdef CONFIG_CGROUP_SCHED
7569 * How much cpu bandwidth does init_task_group get?
7571 * In case of task-groups formed thr' the cgroup filesystem, it
7572 * gets 100% of the cpu resources in the system. This overall
7573 * system cpu resource is divided among the tasks of
7574 * init_task_group and its child task-groups in a fair manner,
7575 * based on each entity's (task or task-group's) weight
7576 * (se->load.weight).
7578 * In other words, if init_task_group has 10 tasks of weight
7579 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7580 * then A0's share of the cpu resource is:
7582 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7584 * We achieve this by letting init_task_group's tasks sit
7585 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7587 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7589 #endif /* CONFIG_FAIR_GROUP_SCHED */
7591 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7592 #ifdef CONFIG_RT_GROUP_SCHED
7593 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7594 #ifdef CONFIG_CGROUP_SCHED
7595 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7599 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7600 rq
->cpu_load
[j
] = 0;
7604 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7605 rq
->post_schedule
= 0;
7606 rq
->active_balance
= 0;
7607 rq
->next_balance
= jiffies
;
7612 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7613 rq_attach_root(rq
, &def_root_domain
);
7616 atomic_set(&rq
->nr_iowait
, 0);
7619 set_load_weight(&init_task
);
7621 #ifdef CONFIG_PREEMPT_NOTIFIERS
7622 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7626 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7629 #ifdef CONFIG_RT_MUTEXES
7630 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7634 * The boot idle thread does lazy MMU switching as well:
7636 atomic_inc(&init_mm
.mm_count
);
7637 enter_lazy_tlb(&init_mm
, current
);
7640 * Make us the idle thread. Technically, schedule() should not be
7641 * called from this thread, however somewhere below it might be,
7642 * but because we are the idle thread, we just pick up running again
7643 * when this runqueue becomes "idle".
7645 init_idle(current
, smp_processor_id());
7647 calc_load_update
= jiffies
+ LOAD_FREQ
;
7650 * During early bootup we pretend to be a normal task:
7652 current
->sched_class
= &fair_sched_class
;
7654 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7655 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7658 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7659 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7661 /* May be allocated at isolcpus cmdline parse time */
7662 if (cpu_isolated_map
== NULL
)
7663 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7668 scheduler_running
= 1;
7671 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7672 static inline int preempt_count_equals(int preempt_offset
)
7674 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7676 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7679 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7682 static unsigned long prev_jiffy
; /* ratelimiting */
7684 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7685 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7687 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7689 prev_jiffy
= jiffies
;
7692 "BUG: sleeping function called from invalid context at %s:%d\n",
7695 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7696 in_atomic(), irqs_disabled(),
7697 current
->pid
, current
->comm
);
7699 debug_show_held_locks(current
);
7700 if (irqs_disabled())
7701 print_irqtrace_events(current
);
7705 EXPORT_SYMBOL(__might_sleep
);
7708 #ifdef CONFIG_MAGIC_SYSRQ
7709 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7713 on_rq
= p
->se
.on_rq
;
7715 deactivate_task(rq
, p
, 0);
7716 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7718 activate_task(rq
, p
, 0);
7719 resched_task(rq
->curr
);
7723 void normalize_rt_tasks(void)
7725 struct task_struct
*g
, *p
;
7726 unsigned long flags
;
7729 read_lock_irqsave(&tasklist_lock
, flags
);
7730 do_each_thread(g
, p
) {
7732 * Only normalize user tasks:
7737 p
->se
.exec_start
= 0;
7738 #ifdef CONFIG_SCHEDSTATS
7739 p
->se
.statistics
.wait_start
= 0;
7740 p
->se
.statistics
.sleep_start
= 0;
7741 p
->se
.statistics
.block_start
= 0;
7746 * Renice negative nice level userspace
7749 if (TASK_NICE(p
) < 0 && p
->mm
)
7750 set_user_nice(p
, 0);
7754 raw_spin_lock(&p
->pi_lock
);
7755 rq
= __task_rq_lock(p
);
7757 normalize_task(rq
, p
);
7759 __task_rq_unlock(rq
);
7760 raw_spin_unlock(&p
->pi_lock
);
7761 } while_each_thread(g
, p
);
7763 read_unlock_irqrestore(&tasklist_lock
, flags
);
7766 #endif /* CONFIG_MAGIC_SYSRQ */
7768 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7770 * These functions are only useful for the IA64 MCA handling, or kdb.
7772 * They can only be called when the whole system has been
7773 * stopped - every CPU needs to be quiescent, and no scheduling
7774 * activity can take place. Using them for anything else would
7775 * be a serious bug, and as a result, they aren't even visible
7776 * under any other configuration.
7780 * curr_task - return the current task for a given cpu.
7781 * @cpu: the processor in question.
7783 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7785 struct task_struct
*curr_task(int cpu
)
7787 return cpu_curr(cpu
);
7790 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7794 * set_curr_task - set the current task for a given cpu.
7795 * @cpu: the processor in question.
7796 * @p: the task pointer to set.
7798 * Description: This function must only be used when non-maskable interrupts
7799 * are serviced on a separate stack. It allows the architecture to switch the
7800 * notion of the current task on a cpu in a non-blocking manner. This function
7801 * must be called with all CPU's synchronized, and interrupts disabled, the
7802 * and caller must save the original value of the current task (see
7803 * curr_task() above) and restore that value before reenabling interrupts and
7804 * re-starting the system.
7806 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7808 void set_curr_task(int cpu
, struct task_struct
*p
)
7815 #ifdef CONFIG_FAIR_GROUP_SCHED
7816 static void free_fair_sched_group(struct task_group
*tg
)
7820 for_each_possible_cpu(i
) {
7822 kfree(tg
->cfs_rq
[i
]);
7832 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7834 struct cfs_rq
*cfs_rq
;
7835 struct sched_entity
*se
;
7839 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7842 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7846 tg
->shares
= NICE_0_LOAD
;
7848 for_each_possible_cpu(i
) {
7851 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7852 GFP_KERNEL
, cpu_to_node(i
));
7856 se
= kzalloc_node(sizeof(struct sched_entity
),
7857 GFP_KERNEL
, cpu_to_node(i
));
7861 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
7872 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7874 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7875 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7878 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7880 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7882 #else /* !CONFG_FAIR_GROUP_SCHED */
7883 static inline void free_fair_sched_group(struct task_group
*tg
)
7888 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7893 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7897 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7900 #endif /* CONFIG_FAIR_GROUP_SCHED */
7902 #ifdef CONFIG_RT_GROUP_SCHED
7903 static void free_rt_sched_group(struct task_group
*tg
)
7907 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
7909 for_each_possible_cpu(i
) {
7911 kfree(tg
->rt_rq
[i
]);
7913 kfree(tg
->rt_se
[i
]);
7921 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7923 struct rt_rq
*rt_rq
;
7924 struct sched_rt_entity
*rt_se
;
7928 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7931 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
7935 init_rt_bandwidth(&tg
->rt_bandwidth
,
7936 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
7938 for_each_possible_cpu(i
) {
7941 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
7942 GFP_KERNEL
, cpu_to_node(i
));
7946 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
7947 GFP_KERNEL
, cpu_to_node(i
));
7951 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
7962 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7964 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
7965 &cpu_rq(cpu
)->leaf_rt_rq_list
);
7968 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7970 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
7972 #else /* !CONFIG_RT_GROUP_SCHED */
7973 static inline void free_rt_sched_group(struct task_group
*tg
)
7978 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7983 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7987 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7990 #endif /* CONFIG_RT_GROUP_SCHED */
7992 #ifdef CONFIG_CGROUP_SCHED
7993 static void free_sched_group(struct task_group
*tg
)
7995 free_fair_sched_group(tg
);
7996 free_rt_sched_group(tg
);
8000 /* allocate runqueue etc for a new task group */
8001 struct task_group
*sched_create_group(struct task_group
*parent
)
8003 struct task_group
*tg
;
8004 unsigned long flags
;
8007 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8009 return ERR_PTR(-ENOMEM
);
8011 if (!alloc_fair_sched_group(tg
, parent
))
8014 if (!alloc_rt_sched_group(tg
, parent
))
8017 spin_lock_irqsave(&task_group_lock
, flags
);
8018 for_each_possible_cpu(i
) {
8019 register_fair_sched_group(tg
, i
);
8020 register_rt_sched_group(tg
, i
);
8022 list_add_rcu(&tg
->list
, &task_groups
);
8024 WARN_ON(!parent
); /* root should already exist */
8026 tg
->parent
= parent
;
8027 INIT_LIST_HEAD(&tg
->children
);
8028 list_add_rcu(&tg
->siblings
, &parent
->children
);
8029 spin_unlock_irqrestore(&task_group_lock
, flags
);
8034 free_sched_group(tg
);
8035 return ERR_PTR(-ENOMEM
);
8038 /* rcu callback to free various structures associated with a task group */
8039 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8041 /* now it should be safe to free those cfs_rqs */
8042 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8045 /* Destroy runqueue etc associated with a task group */
8046 void sched_destroy_group(struct task_group
*tg
)
8048 unsigned long flags
;
8051 spin_lock_irqsave(&task_group_lock
, flags
);
8052 for_each_possible_cpu(i
) {
8053 unregister_fair_sched_group(tg
, i
);
8054 unregister_rt_sched_group(tg
, i
);
8056 list_del_rcu(&tg
->list
);
8057 list_del_rcu(&tg
->siblings
);
8058 spin_unlock_irqrestore(&task_group_lock
, flags
);
8060 /* wait for possible concurrent references to cfs_rqs complete */
8061 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8064 /* change task's runqueue when it moves between groups.
8065 * The caller of this function should have put the task in its new group
8066 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8067 * reflect its new group.
8069 void sched_move_task(struct task_struct
*tsk
)
8072 unsigned long flags
;
8075 rq
= task_rq_lock(tsk
, &flags
);
8077 running
= task_current(rq
, tsk
);
8078 on_rq
= tsk
->se
.on_rq
;
8081 dequeue_task(rq
, tsk
, 0);
8082 if (unlikely(running
))
8083 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8085 set_task_rq(tsk
, task_cpu(tsk
));
8087 #ifdef CONFIG_FAIR_GROUP_SCHED
8088 if (tsk
->sched_class
->moved_group
)
8089 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8092 if (unlikely(running
))
8093 tsk
->sched_class
->set_curr_task(rq
);
8095 enqueue_task(rq
, tsk
, 0);
8097 task_rq_unlock(rq
, &flags
);
8099 #endif /* CONFIG_CGROUP_SCHED */
8101 #ifdef CONFIG_FAIR_GROUP_SCHED
8102 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8104 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8109 dequeue_entity(cfs_rq
, se
, 0);
8111 se
->load
.weight
= shares
;
8112 se
->load
.inv_weight
= 0;
8115 enqueue_entity(cfs_rq
, se
, 0);
8118 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8120 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8121 struct rq
*rq
= cfs_rq
->rq
;
8122 unsigned long flags
;
8124 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8125 __set_se_shares(se
, shares
);
8126 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8129 static DEFINE_MUTEX(shares_mutex
);
8131 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8134 unsigned long flags
;
8137 * We can't change the weight of the root cgroup.
8142 if (shares
< MIN_SHARES
)
8143 shares
= MIN_SHARES
;
8144 else if (shares
> MAX_SHARES
)
8145 shares
= MAX_SHARES
;
8147 mutex_lock(&shares_mutex
);
8148 if (tg
->shares
== shares
)
8151 spin_lock_irqsave(&task_group_lock
, flags
);
8152 for_each_possible_cpu(i
)
8153 unregister_fair_sched_group(tg
, i
);
8154 list_del_rcu(&tg
->siblings
);
8155 spin_unlock_irqrestore(&task_group_lock
, flags
);
8157 /* wait for any ongoing reference to this group to finish */
8158 synchronize_sched();
8161 * Now we are free to modify the group's share on each cpu
8162 * w/o tripping rebalance_share or load_balance_fair.
8164 tg
->shares
= shares
;
8165 for_each_possible_cpu(i
) {
8169 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8170 set_se_shares(tg
->se
[i
], shares
);
8174 * Enable load balance activity on this group, by inserting it back on
8175 * each cpu's rq->leaf_cfs_rq_list.
8177 spin_lock_irqsave(&task_group_lock
, flags
);
8178 for_each_possible_cpu(i
)
8179 register_fair_sched_group(tg
, i
);
8180 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8181 spin_unlock_irqrestore(&task_group_lock
, flags
);
8183 mutex_unlock(&shares_mutex
);
8187 unsigned long sched_group_shares(struct task_group
*tg
)
8193 #ifdef CONFIG_RT_GROUP_SCHED
8195 * Ensure that the real time constraints are schedulable.
8197 static DEFINE_MUTEX(rt_constraints_mutex
);
8199 static unsigned long to_ratio(u64 period
, u64 runtime
)
8201 if (runtime
== RUNTIME_INF
)
8204 return div64_u64(runtime
<< 20, period
);
8207 /* Must be called with tasklist_lock held */
8208 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8210 struct task_struct
*g
, *p
;
8212 do_each_thread(g
, p
) {
8213 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8215 } while_each_thread(g
, p
);
8220 struct rt_schedulable_data
{
8221 struct task_group
*tg
;
8226 static int tg_schedulable(struct task_group
*tg
, void *data
)
8228 struct rt_schedulable_data
*d
= data
;
8229 struct task_group
*child
;
8230 unsigned long total
, sum
= 0;
8231 u64 period
, runtime
;
8233 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8234 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8237 period
= d
->rt_period
;
8238 runtime
= d
->rt_runtime
;
8242 * Cannot have more runtime than the period.
8244 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8248 * Ensure we don't starve existing RT tasks.
8250 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8253 total
= to_ratio(period
, runtime
);
8256 * Nobody can have more than the global setting allows.
8258 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8262 * The sum of our children's runtime should not exceed our own.
8264 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8265 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8266 runtime
= child
->rt_bandwidth
.rt_runtime
;
8268 if (child
== d
->tg
) {
8269 period
= d
->rt_period
;
8270 runtime
= d
->rt_runtime
;
8273 sum
+= to_ratio(period
, runtime
);
8282 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8284 struct rt_schedulable_data data
= {
8286 .rt_period
= period
,
8287 .rt_runtime
= runtime
,
8290 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8293 static int tg_set_bandwidth(struct task_group
*tg
,
8294 u64 rt_period
, u64 rt_runtime
)
8298 mutex_lock(&rt_constraints_mutex
);
8299 read_lock(&tasklist_lock
);
8300 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8304 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8305 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8306 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8308 for_each_possible_cpu(i
) {
8309 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8311 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8312 rt_rq
->rt_runtime
= rt_runtime
;
8313 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8315 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8317 read_unlock(&tasklist_lock
);
8318 mutex_unlock(&rt_constraints_mutex
);
8323 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8325 u64 rt_runtime
, rt_period
;
8327 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8328 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8329 if (rt_runtime_us
< 0)
8330 rt_runtime
= RUNTIME_INF
;
8332 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8335 long sched_group_rt_runtime(struct task_group
*tg
)
8339 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8342 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8343 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8344 return rt_runtime_us
;
8347 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8349 u64 rt_runtime
, rt_period
;
8351 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8352 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8357 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8360 long sched_group_rt_period(struct task_group
*tg
)
8364 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8365 do_div(rt_period_us
, NSEC_PER_USEC
);
8366 return rt_period_us
;
8369 static int sched_rt_global_constraints(void)
8371 u64 runtime
, period
;
8374 if (sysctl_sched_rt_period
<= 0)
8377 runtime
= global_rt_runtime();
8378 period
= global_rt_period();
8381 * Sanity check on the sysctl variables.
8383 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8386 mutex_lock(&rt_constraints_mutex
);
8387 read_lock(&tasklist_lock
);
8388 ret
= __rt_schedulable(NULL
, 0, 0);
8389 read_unlock(&tasklist_lock
);
8390 mutex_unlock(&rt_constraints_mutex
);
8395 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8397 /* Don't accept realtime tasks when there is no way for them to run */
8398 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8404 #else /* !CONFIG_RT_GROUP_SCHED */
8405 static int sched_rt_global_constraints(void)
8407 unsigned long flags
;
8410 if (sysctl_sched_rt_period
<= 0)
8414 * There's always some RT tasks in the root group
8415 * -- migration, kstopmachine etc..
8417 if (sysctl_sched_rt_runtime
== 0)
8420 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8421 for_each_possible_cpu(i
) {
8422 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8424 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8425 rt_rq
->rt_runtime
= global_rt_runtime();
8426 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8428 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8432 #endif /* CONFIG_RT_GROUP_SCHED */
8434 int sched_rt_handler(struct ctl_table
*table
, int write
,
8435 void __user
*buffer
, size_t *lenp
,
8439 int old_period
, old_runtime
;
8440 static DEFINE_MUTEX(mutex
);
8443 old_period
= sysctl_sched_rt_period
;
8444 old_runtime
= sysctl_sched_rt_runtime
;
8446 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8448 if (!ret
&& write
) {
8449 ret
= sched_rt_global_constraints();
8451 sysctl_sched_rt_period
= old_period
;
8452 sysctl_sched_rt_runtime
= old_runtime
;
8454 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8455 def_rt_bandwidth
.rt_period
=
8456 ns_to_ktime(global_rt_period());
8459 mutex_unlock(&mutex
);
8464 #ifdef CONFIG_CGROUP_SCHED
8466 /* return corresponding task_group object of a cgroup */
8467 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8469 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8470 struct task_group
, css
);
8473 static struct cgroup_subsys_state
*
8474 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8476 struct task_group
*tg
, *parent
;
8478 if (!cgrp
->parent
) {
8479 /* This is early initialization for the top cgroup */
8480 return &init_task_group
.css
;
8483 parent
= cgroup_tg(cgrp
->parent
);
8484 tg
= sched_create_group(parent
);
8486 return ERR_PTR(-ENOMEM
);
8492 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8494 struct task_group
*tg
= cgroup_tg(cgrp
);
8496 sched_destroy_group(tg
);
8500 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8502 #ifdef CONFIG_RT_GROUP_SCHED
8503 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8506 /* We don't support RT-tasks being in separate groups */
8507 if (tsk
->sched_class
!= &fair_sched_class
)
8514 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8515 struct task_struct
*tsk
, bool threadgroup
)
8517 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8521 struct task_struct
*c
;
8523 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8524 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8536 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8537 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8540 sched_move_task(tsk
);
8542 struct task_struct
*c
;
8544 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8551 #ifdef CONFIG_FAIR_GROUP_SCHED
8552 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8555 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8558 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8560 struct task_group
*tg
= cgroup_tg(cgrp
);
8562 return (u64
) tg
->shares
;
8564 #endif /* CONFIG_FAIR_GROUP_SCHED */
8566 #ifdef CONFIG_RT_GROUP_SCHED
8567 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8570 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8573 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8575 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8578 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8581 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8584 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8586 return sched_group_rt_period(cgroup_tg(cgrp
));
8588 #endif /* CONFIG_RT_GROUP_SCHED */
8590 static struct cftype cpu_files
[] = {
8591 #ifdef CONFIG_FAIR_GROUP_SCHED
8594 .read_u64
= cpu_shares_read_u64
,
8595 .write_u64
= cpu_shares_write_u64
,
8598 #ifdef CONFIG_RT_GROUP_SCHED
8600 .name
= "rt_runtime_us",
8601 .read_s64
= cpu_rt_runtime_read
,
8602 .write_s64
= cpu_rt_runtime_write
,
8605 .name
= "rt_period_us",
8606 .read_u64
= cpu_rt_period_read_uint
,
8607 .write_u64
= cpu_rt_period_write_uint
,
8612 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8614 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8617 struct cgroup_subsys cpu_cgroup_subsys
= {
8619 .create
= cpu_cgroup_create
,
8620 .destroy
= cpu_cgroup_destroy
,
8621 .can_attach
= cpu_cgroup_can_attach
,
8622 .attach
= cpu_cgroup_attach
,
8623 .populate
= cpu_cgroup_populate
,
8624 .subsys_id
= cpu_cgroup_subsys_id
,
8628 #endif /* CONFIG_CGROUP_SCHED */
8630 #ifdef CONFIG_CGROUP_CPUACCT
8633 * CPU accounting code for task groups.
8635 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8636 * (balbir@in.ibm.com).
8639 /* track cpu usage of a group of tasks and its child groups */
8641 struct cgroup_subsys_state css
;
8642 /* cpuusage holds pointer to a u64-type object on every cpu */
8643 u64 __percpu
*cpuusage
;
8644 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8645 struct cpuacct
*parent
;
8648 struct cgroup_subsys cpuacct_subsys
;
8650 /* return cpu accounting group corresponding to this container */
8651 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8653 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8654 struct cpuacct
, css
);
8657 /* return cpu accounting group to which this task belongs */
8658 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8660 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8661 struct cpuacct
, css
);
8664 /* create a new cpu accounting group */
8665 static struct cgroup_subsys_state
*cpuacct_create(
8666 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8668 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8674 ca
->cpuusage
= alloc_percpu(u64
);
8678 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8679 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8680 goto out_free_counters
;
8683 ca
->parent
= cgroup_ca(cgrp
->parent
);
8689 percpu_counter_destroy(&ca
->cpustat
[i
]);
8690 free_percpu(ca
->cpuusage
);
8694 return ERR_PTR(-ENOMEM
);
8697 /* destroy an existing cpu accounting group */
8699 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8701 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8704 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8705 percpu_counter_destroy(&ca
->cpustat
[i
]);
8706 free_percpu(ca
->cpuusage
);
8710 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8712 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8715 #ifndef CONFIG_64BIT
8717 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8719 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8721 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8729 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8731 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8733 #ifndef CONFIG_64BIT
8735 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8737 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8739 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8745 /* return total cpu usage (in nanoseconds) of a group */
8746 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8748 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8749 u64 totalcpuusage
= 0;
8752 for_each_present_cpu(i
)
8753 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8755 return totalcpuusage
;
8758 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8761 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8770 for_each_present_cpu(i
)
8771 cpuacct_cpuusage_write(ca
, i
, 0);
8777 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8780 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8784 for_each_present_cpu(i
) {
8785 percpu
= cpuacct_cpuusage_read(ca
, i
);
8786 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8788 seq_printf(m
, "\n");
8792 static const char *cpuacct_stat_desc
[] = {
8793 [CPUACCT_STAT_USER
] = "user",
8794 [CPUACCT_STAT_SYSTEM
] = "system",
8797 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8798 struct cgroup_map_cb
*cb
)
8800 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8803 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8804 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8805 val
= cputime64_to_clock_t(val
);
8806 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8811 static struct cftype files
[] = {
8814 .read_u64
= cpuusage_read
,
8815 .write_u64
= cpuusage_write
,
8818 .name
= "usage_percpu",
8819 .read_seq_string
= cpuacct_percpu_seq_read
,
8823 .read_map
= cpuacct_stats_show
,
8827 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8829 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8833 * charge this task's execution time to its accounting group.
8835 * called with rq->lock held.
8837 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8842 if (unlikely(!cpuacct_subsys
.active
))
8845 cpu
= task_cpu(tsk
);
8851 for (; ca
; ca
= ca
->parent
) {
8852 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8853 *cpuusage
+= cputime
;
8860 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8861 * in cputime_t units. As a result, cpuacct_update_stats calls
8862 * percpu_counter_add with values large enough to always overflow the
8863 * per cpu batch limit causing bad SMP scalability.
8865 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8866 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8867 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8870 #define CPUACCT_BATCH \
8871 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8873 #define CPUACCT_BATCH 0
8877 * Charge the system/user time to the task's accounting group.
8879 static void cpuacct_update_stats(struct task_struct
*tsk
,
8880 enum cpuacct_stat_index idx
, cputime_t val
)
8883 int batch
= CPUACCT_BATCH
;
8885 if (unlikely(!cpuacct_subsys
.active
))
8892 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
8898 struct cgroup_subsys cpuacct_subsys
= {
8900 .create
= cpuacct_create
,
8901 .destroy
= cpuacct_destroy
,
8902 .populate
= cpuacct_populate
,
8903 .subsys_id
= cpuacct_subsys_id
,
8905 #endif /* CONFIG_CGROUP_CPUACCT */
8909 void synchronize_sched_expedited(void)
8913 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
8915 #else /* #ifndef CONFIG_SMP */
8917 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
8919 static int synchronize_sched_expedited_cpu_stop(void *data
)
8922 * There must be a full memory barrier on each affected CPU
8923 * between the time that try_stop_cpus() is called and the
8924 * time that it returns.
8926 * In the current initial implementation of cpu_stop, the
8927 * above condition is already met when the control reaches
8928 * this point and the following smp_mb() is not strictly
8929 * necessary. Do smp_mb() anyway for documentation and
8930 * robustness against future implementation changes.
8932 smp_mb(); /* See above comment block. */
8937 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
8938 * approach to force grace period to end quickly. This consumes
8939 * significant time on all CPUs, and is thus not recommended for
8940 * any sort of common-case code.
8942 * Note that it is illegal to call this function while holding any
8943 * lock that is acquired by a CPU-hotplug notifier. Failing to
8944 * observe this restriction will result in deadlock.
8946 void synchronize_sched_expedited(void)
8948 int snap
, trycount
= 0;
8950 smp_mb(); /* ensure prior mod happens before capturing snap. */
8951 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
8953 while (try_stop_cpus(cpu_online_mask
,
8954 synchronize_sched_expedited_cpu_stop
,
8957 if (trycount
++ < 10)
8958 udelay(trycount
* num_online_cpus());
8960 synchronize_sched();
8963 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
8964 smp_mb(); /* ensure test happens before caller kfree */
8969 atomic_inc(&synchronize_sched_expedited_count
);
8970 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
8973 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
8975 #endif /* #else #ifndef CONFIG_SMP */