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/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.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
];
506 unsigned char in_nohz_recently
;
508 /* capture load from *all* tasks on this cpu: */
509 struct load_weight load
;
510 unsigned long nr_load_updates
;
516 #ifdef CONFIG_FAIR_GROUP_SCHED
517 /* list of leaf cfs_rq on this cpu: */
518 struct list_head leaf_cfs_rq_list
;
520 #ifdef CONFIG_RT_GROUP_SCHED
521 struct list_head leaf_rt_rq_list
;
525 * This is part of a global counter where only the total sum
526 * over all CPUs matters. A task can increase this counter on
527 * one CPU and if it got migrated afterwards it may decrease
528 * it on another CPU. Always updated under the runqueue lock:
530 unsigned long nr_uninterruptible
;
532 struct task_struct
*curr
, *idle
;
533 unsigned long next_balance
;
534 struct mm_struct
*prev_mm
;
541 struct root_domain
*rd
;
542 struct sched_domain
*sd
;
544 unsigned long cpu_power
;
546 unsigned char idle_at_tick
;
547 /* For active balancing */
551 /* cpu of this runqueue: */
555 unsigned long avg_load_per_task
;
557 struct task_struct
*migration_thread
;
558 struct list_head migration_queue
;
566 /* calc_load related fields */
567 unsigned long calc_load_update
;
568 long calc_load_active
;
570 #ifdef CONFIG_SCHED_HRTICK
572 int hrtick_csd_pending
;
573 struct call_single_data hrtick_csd
;
575 struct hrtimer hrtick_timer
;
578 #ifdef CONFIG_SCHEDSTATS
580 struct sched_info rq_sched_info
;
581 unsigned long long rq_cpu_time
;
582 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
584 /* sys_sched_yield() stats */
585 unsigned int yld_count
;
587 /* schedule() stats */
588 unsigned int sched_switch
;
589 unsigned int sched_count
;
590 unsigned int sched_goidle
;
592 /* try_to_wake_up() stats */
593 unsigned int ttwu_count
;
594 unsigned int ttwu_local
;
597 unsigned int bkl_count
;
601 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
604 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
606 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
609 static inline int cpu_of(struct rq
*rq
)
618 #define rcu_dereference_check_sched_domain(p) \
619 rcu_dereference_check((p), \
620 rcu_read_lock_sched_held() || \
621 lockdep_is_held(&sched_domains_mutex))
624 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
625 * See detach_destroy_domains: synchronize_sched for details.
627 * The domain tree of any CPU may only be accessed from within
628 * preempt-disabled sections.
630 #define for_each_domain(cpu, __sd) \
631 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
633 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
634 #define this_rq() (&__get_cpu_var(runqueues))
635 #define task_rq(p) cpu_rq(task_cpu(p))
636 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
637 #define raw_rq() (&__raw_get_cpu_var(runqueues))
639 inline void update_rq_clock(struct rq
*rq
)
641 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
645 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
647 #ifdef CONFIG_SCHED_DEBUG
648 # define const_debug __read_mostly
650 # define const_debug static const
655 * @cpu: the processor in question.
657 * Returns true if the current cpu runqueue is locked.
658 * This interface allows printk to be called with the runqueue lock
659 * held and know whether or not it is OK to wake up the klogd.
661 int runqueue_is_locked(int cpu
)
663 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
667 * Debugging: various feature bits
670 #define SCHED_FEAT(name, enabled) \
671 __SCHED_FEAT_##name ,
674 #include "sched_features.h"
679 #define SCHED_FEAT(name, enabled) \
680 (1UL << __SCHED_FEAT_##name) * enabled |
682 const_debug
unsigned int sysctl_sched_features
=
683 #include "sched_features.h"
688 #ifdef CONFIG_SCHED_DEBUG
689 #define SCHED_FEAT(name, enabled) \
692 static __read_mostly
char *sched_feat_names
[] = {
693 #include "sched_features.h"
699 static int sched_feat_show(struct seq_file
*m
, void *v
)
703 for (i
= 0; sched_feat_names
[i
]; i
++) {
704 if (!(sysctl_sched_features
& (1UL << i
)))
706 seq_printf(m
, "%s ", sched_feat_names
[i
]);
714 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
715 size_t cnt
, loff_t
*ppos
)
725 if (copy_from_user(&buf
, ubuf
, cnt
))
731 if (strncmp(buf
, "NO_", 3) == 0) {
736 for (i
= 0; sched_feat_names
[i
]; i
++) {
737 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
739 sysctl_sched_features
&= ~(1UL << i
);
741 sysctl_sched_features
|= (1UL << i
);
746 if (!sched_feat_names
[i
])
754 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
756 return single_open(filp
, sched_feat_show
, NULL
);
759 static const struct file_operations sched_feat_fops
= {
760 .open
= sched_feat_open
,
761 .write
= sched_feat_write
,
764 .release
= single_release
,
767 static __init
int sched_init_debug(void)
769 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
774 late_initcall(sched_init_debug
);
778 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
781 * Number of tasks to iterate in a single balance run.
782 * Limited because this is done with IRQs disabled.
784 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
787 * ratelimit for updating the group shares.
790 unsigned int sysctl_sched_shares_ratelimit
= 250000;
791 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
794 * Inject some fuzzyness into changing the per-cpu group shares
795 * this avoids remote rq-locks at the expense of fairness.
798 unsigned int sysctl_sched_shares_thresh
= 4;
801 * period over which we average the RT time consumption, measured
806 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
809 * period over which we measure -rt task cpu usage in us.
812 unsigned int sysctl_sched_rt_period
= 1000000;
814 static __read_mostly
int scheduler_running
;
817 * part of the period that we allow rt tasks to run in us.
820 int sysctl_sched_rt_runtime
= 950000;
822 static inline u64
global_rt_period(void)
824 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
827 static inline u64
global_rt_runtime(void)
829 if (sysctl_sched_rt_runtime
< 0)
832 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
835 #ifndef prepare_arch_switch
836 # define prepare_arch_switch(next) do { } while (0)
838 #ifndef finish_arch_switch
839 # define finish_arch_switch(prev) do { } while (0)
842 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
844 return rq
->curr
== p
;
847 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
848 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
850 return task_current(rq
, p
);
853 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
857 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
859 #ifdef CONFIG_DEBUG_SPINLOCK
860 /* this is a valid case when another task releases the spinlock */
861 rq
->lock
.owner
= current
;
864 * If we are tracking spinlock dependencies then we have to
865 * fix up the runqueue lock - which gets 'carried over' from
868 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
870 raw_spin_unlock_irq(&rq
->lock
);
873 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
874 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
879 return task_current(rq
, p
);
883 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
887 * We can optimise this out completely for !SMP, because the
888 * SMP rebalancing from interrupt is the only thing that cares
893 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
894 raw_spin_unlock_irq(&rq
->lock
);
896 raw_spin_unlock(&rq
->lock
);
900 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
904 * After ->oncpu is cleared, the task can be moved to a different CPU.
905 * We must ensure this doesn't happen until the switch is completely
911 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
918 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
921 static inline int task_is_waking(struct task_struct
*p
)
923 return unlikely(p
->state
== TASK_WAKING
);
927 * __task_rq_lock - lock the runqueue a given task resides on.
928 * Must be called interrupts disabled.
930 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
937 raw_spin_lock(&rq
->lock
);
938 if (likely(rq
== task_rq(p
)))
940 raw_spin_unlock(&rq
->lock
);
945 * task_rq_lock - lock the runqueue a given task resides on and disable
946 * interrupts. Note the ordering: we can safely lookup the task_rq without
947 * explicitly disabling preemption.
949 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
955 local_irq_save(*flags
);
957 raw_spin_lock(&rq
->lock
);
958 if (likely(rq
== task_rq(p
)))
960 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
964 void task_rq_unlock_wait(struct task_struct
*p
)
966 struct rq
*rq
= task_rq(p
);
968 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
969 raw_spin_unlock_wait(&rq
->lock
);
972 static void __task_rq_unlock(struct rq
*rq
)
975 raw_spin_unlock(&rq
->lock
);
978 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
981 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq
*this_rq_lock(void)
994 raw_spin_lock(&rq
->lock
);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq
*rq
)
1018 if (!sched_feat(HRTICK
))
1020 if (!cpu_active(cpu_of(rq
)))
1022 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1025 static void hrtick_clear(struct rq
*rq
)
1027 if (hrtimer_active(&rq
->hrtick_timer
))
1028 hrtimer_cancel(&rq
->hrtick_timer
);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1037 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1039 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1041 raw_spin_lock(&rq
->lock
);
1042 update_rq_clock(rq
);
1043 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1044 raw_spin_unlock(&rq
->lock
);
1046 return HRTIMER_NORESTART
;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg
)
1055 struct rq
*rq
= arg
;
1057 raw_spin_lock(&rq
->lock
);
1058 hrtimer_restart(&rq
->hrtick_timer
);
1059 rq
->hrtick_csd_pending
= 0;
1060 raw_spin_unlock(&rq
->lock
);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq
*rq
, u64 delay
)
1070 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1071 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1073 hrtimer_set_expires(timer
, time
);
1075 if (rq
== this_rq()) {
1076 hrtimer_restart(timer
);
1077 } else if (!rq
->hrtick_csd_pending
) {
1078 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1079 rq
->hrtick_csd_pending
= 1;
1084 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1086 int cpu
= (int)(long)hcpu
;
1089 case CPU_UP_CANCELED
:
1090 case CPU_UP_CANCELED_FROZEN
:
1091 case CPU_DOWN_PREPARE
:
1092 case CPU_DOWN_PREPARE_FROZEN
:
1094 case CPU_DEAD_FROZEN
:
1095 hrtick_clear(cpu_rq(cpu
));
1102 static __init
void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick
, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq
*rq
, u64 delay
)
1114 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1115 HRTIMER_MODE_REL_PINNED
, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq
*rq
)
1126 rq
->hrtick_csd_pending
= 0;
1128 rq
->hrtick_csd
.flags
= 0;
1129 rq
->hrtick_csd
.func
= __hrtick_start
;
1130 rq
->hrtick_csd
.info
= rq
;
1133 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1134 rq
->hrtick_timer
.function
= hrtick
;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq
*rq
)
1141 static inline void init_rq_hrtick(struct rq
*rq
)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct
*p
)
1167 assert_raw_spin_locked(&task_rq(p
)->lock
);
1169 if (test_tsk_need_resched(p
))
1172 set_tsk_need_resched(p
);
1175 if (cpu
== smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p
))
1181 smp_send_reschedule(cpu
);
1184 static void resched_cpu(int cpu
)
1186 struct rq
*rq
= cpu_rq(cpu
);
1187 unsigned long flags
;
1189 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1191 resched_task(cpu_curr(cpu
));
1192 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1197 * When add_timer_on() enqueues a timer into the timer wheel of an
1198 * idle CPU then this timer might expire before the next timer event
1199 * which is scheduled to wake up that CPU. In case of a completely
1200 * idle system the next event might even be infinite time into the
1201 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1202 * leaves the inner idle loop so the newly added timer is taken into
1203 * account when the CPU goes back to idle and evaluates the timer
1204 * wheel for the next timer event.
1206 void wake_up_idle_cpu(int cpu
)
1208 struct rq
*rq
= cpu_rq(cpu
);
1210 if (cpu
== smp_processor_id())
1214 * This is safe, as this function is called with the timer
1215 * wheel base lock of (cpu) held. When the CPU is on the way
1216 * to idle and has not yet set rq->curr to idle then it will
1217 * be serialized on the timer wheel base lock and take the new
1218 * timer into account automatically.
1220 if (rq
->curr
!= rq
->idle
)
1224 * We can set TIF_RESCHED on the idle task of the other CPU
1225 * lockless. The worst case is that the other CPU runs the
1226 * idle task through an additional NOOP schedule()
1228 set_tsk_need_resched(rq
->idle
);
1230 /* NEED_RESCHED must be visible before we test polling */
1232 if (!tsk_is_polling(rq
->idle
))
1233 smp_send_reschedule(cpu
);
1235 #endif /* CONFIG_NO_HZ */
1237 static u64
sched_avg_period(void)
1239 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1242 static void sched_avg_update(struct rq
*rq
)
1244 s64 period
= sched_avg_period();
1246 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1248 * Inline assembly required to prevent the compiler
1249 * optimising this loop into a divmod call.
1250 * See __iter_div_u64_rem() for another example of this.
1252 asm("" : "+rm" (rq
->age_stamp
));
1253 rq
->age_stamp
+= period
;
1258 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1260 rq
->rt_avg
+= rt_delta
;
1261 sched_avg_update(rq
);
1264 #else /* !CONFIG_SMP */
1265 static void resched_task(struct task_struct
*p
)
1267 assert_raw_spin_locked(&task_rq(p
)->lock
);
1268 set_tsk_need_resched(p
);
1271 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1275 static void sched_avg_update(struct rq
*rq
)
1278 #endif /* CONFIG_SMP */
1280 #if BITS_PER_LONG == 32
1281 # define WMULT_CONST (~0UL)
1283 # define WMULT_CONST (1UL << 32)
1286 #define WMULT_SHIFT 32
1289 * Shift right and round:
1291 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1294 * delta *= weight / lw
1296 static unsigned long
1297 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1298 struct load_weight
*lw
)
1302 if (!lw
->inv_weight
) {
1303 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1306 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1310 tmp
= (u64
)delta_exec
* weight
;
1312 * Check whether we'd overflow the 64-bit multiplication:
1314 if (unlikely(tmp
> WMULT_CONST
))
1315 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1318 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1320 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1323 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1329 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1336 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1337 * of tasks with abnormal "nice" values across CPUs the contribution that
1338 * each task makes to its run queue's load is weighted according to its
1339 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1340 * scaled version of the new time slice allocation that they receive on time
1344 #define WEIGHT_IDLEPRIO 3
1345 #define WMULT_IDLEPRIO 1431655765
1348 * Nice levels are multiplicative, with a gentle 10% change for every
1349 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1350 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1351 * that remained on nice 0.
1353 * The "10% effect" is relative and cumulative: from _any_ nice level,
1354 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1355 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1356 * If a task goes up by ~10% and another task goes down by ~10% then
1357 * the relative distance between them is ~25%.)
1359 static const int prio_to_weight
[40] = {
1360 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1361 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1362 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1363 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1364 /* 0 */ 1024, 820, 655, 526, 423,
1365 /* 5 */ 335, 272, 215, 172, 137,
1366 /* 10 */ 110, 87, 70, 56, 45,
1367 /* 15 */ 36, 29, 23, 18, 15,
1371 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1373 * In cases where the weight does not change often, we can use the
1374 * precalculated inverse to speed up arithmetics by turning divisions
1375 * into multiplications:
1377 static const u32 prio_to_wmult
[40] = {
1378 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1379 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1380 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1381 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1382 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1383 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1384 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1385 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1388 /* Time spent by the tasks of the cpu accounting group executing in ... */
1389 enum cpuacct_stat_index
{
1390 CPUACCT_STAT_USER
, /* ... user mode */
1391 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1393 CPUACCT_STAT_NSTATS
,
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1398 static void cpuacct_update_stats(struct task_struct
*tsk
,
1399 enum cpuacct_stat_index idx
, cputime_t val
);
1401 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1402 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1403 enum cpuacct_stat_index idx
, cputime_t val
) {}
1406 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1408 update_load_add(&rq
->load
, load
);
1411 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1413 update_load_sub(&rq
->load
, load
);
1416 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1417 typedef int (*tg_visitor
)(struct task_group
*, void *);
1420 * Iterate the full tree, calling @down when first entering a node and @up when
1421 * leaving it for the final time.
1423 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1425 struct task_group
*parent
, *child
;
1429 parent
= &root_task_group
;
1431 ret
= (*down
)(parent
, data
);
1434 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1441 ret
= (*up
)(parent
, data
);
1446 parent
= parent
->parent
;
1455 static int tg_nop(struct task_group
*tg
, void *data
)
1462 /* Used instead of source_load when we know the type == 0 */
1463 static unsigned long weighted_cpuload(const int cpu
)
1465 return cpu_rq(cpu
)->load
.weight
;
1469 * Return a low guess at the load of a migration-source cpu weighted
1470 * according to the scheduling class and "nice" value.
1472 * We want to under-estimate the load of migration sources, to
1473 * balance conservatively.
1475 static unsigned long source_load(int cpu
, int type
)
1477 struct rq
*rq
= cpu_rq(cpu
);
1478 unsigned long total
= weighted_cpuload(cpu
);
1480 if (type
== 0 || !sched_feat(LB_BIAS
))
1483 return min(rq
->cpu_load
[type
-1], total
);
1487 * Return a high guess at the load of a migration-target cpu weighted
1488 * according to the scheduling class and "nice" value.
1490 static unsigned long target_load(int cpu
, int type
)
1492 struct rq
*rq
= cpu_rq(cpu
);
1493 unsigned long total
= weighted_cpuload(cpu
);
1495 if (type
== 0 || !sched_feat(LB_BIAS
))
1498 return max(rq
->cpu_load
[type
-1], total
);
1501 static unsigned long power_of(int cpu
)
1503 return cpu_rq(cpu
)->cpu_power
;
1506 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1508 static unsigned long cpu_avg_load_per_task(int cpu
)
1510 struct rq
*rq
= cpu_rq(cpu
);
1511 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1514 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1516 rq
->avg_load_per_task
= 0;
1518 return rq
->avg_load_per_task
;
1521 #ifdef CONFIG_FAIR_GROUP_SCHED
1523 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1525 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1528 * Calculate and set the cpu's group shares.
1530 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1531 unsigned long sd_shares
,
1532 unsigned long sd_rq_weight
,
1533 unsigned long *usd_rq_weight
)
1535 unsigned long shares
, rq_weight
;
1538 rq_weight
= usd_rq_weight
[cpu
];
1541 rq_weight
= NICE_0_LOAD
;
1545 * \Sum_j shares_j * rq_weight_i
1546 * shares_i = -----------------------------
1547 * \Sum_j rq_weight_j
1549 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1550 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1552 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1553 sysctl_sched_shares_thresh
) {
1554 struct rq
*rq
= cpu_rq(cpu
);
1555 unsigned long flags
;
1557 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1558 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1559 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1560 __set_se_shares(tg
->se
[cpu
], shares
);
1561 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1566 * Re-compute the task group their per cpu shares over the given domain.
1567 * This needs to be done in a bottom-up fashion because the rq weight of a
1568 * parent group depends on the shares of its child groups.
1570 static int tg_shares_up(struct task_group
*tg
, void *data
)
1572 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1573 unsigned long *usd_rq_weight
;
1574 struct sched_domain
*sd
= data
;
1575 unsigned long flags
;
1581 local_irq_save(flags
);
1582 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1584 for_each_cpu(i
, sched_domain_span(sd
)) {
1585 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1586 usd_rq_weight
[i
] = weight
;
1588 rq_weight
+= weight
;
1590 * If there are currently no tasks on the cpu pretend there
1591 * is one of average load so that when a new task gets to
1592 * run here it will not get delayed by group starvation.
1595 weight
= NICE_0_LOAD
;
1597 sum_weight
+= weight
;
1598 shares
+= tg
->cfs_rq
[i
]->shares
;
1602 rq_weight
= sum_weight
;
1604 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1605 shares
= tg
->shares
;
1607 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1608 shares
= tg
->shares
;
1610 for_each_cpu(i
, sched_domain_span(sd
))
1611 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1613 local_irq_restore(flags
);
1619 * Compute the cpu's hierarchical load factor for each task group.
1620 * This needs to be done in a top-down fashion because the load of a child
1621 * group is a fraction of its parents load.
1623 static int tg_load_down(struct task_group
*tg
, void *data
)
1626 long cpu
= (long)data
;
1629 load
= cpu_rq(cpu
)->load
.weight
;
1631 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1632 load
*= tg
->cfs_rq
[cpu
]->shares
;
1633 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1636 tg
->cfs_rq
[cpu
]->h_load
= load
;
1641 static void update_shares(struct sched_domain
*sd
)
1646 if (root_task_group_empty())
1649 now
= cpu_clock(raw_smp_processor_id());
1650 elapsed
= now
- sd
->last_update
;
1652 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1653 sd
->last_update
= now
;
1654 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1658 static void update_h_load(long cpu
)
1660 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1665 static inline void update_shares(struct sched_domain
*sd
)
1671 #ifdef CONFIG_PREEMPT
1673 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1676 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1677 * way at the expense of forcing extra atomic operations in all
1678 * invocations. This assures that the double_lock is acquired using the
1679 * same underlying policy as the spinlock_t on this architecture, which
1680 * reduces latency compared to the unfair variant below. However, it
1681 * also adds more overhead and therefore may reduce throughput.
1683 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1684 __releases(this_rq
->lock
)
1685 __acquires(busiest
->lock
)
1686 __acquires(this_rq
->lock
)
1688 raw_spin_unlock(&this_rq
->lock
);
1689 double_rq_lock(this_rq
, busiest
);
1696 * Unfair double_lock_balance: Optimizes throughput at the expense of
1697 * latency by eliminating extra atomic operations when the locks are
1698 * already in proper order on entry. This favors lower cpu-ids and will
1699 * grant the double lock to lower cpus over higher ids under contention,
1700 * regardless of entry order into the function.
1702 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1703 __releases(this_rq
->lock
)
1704 __acquires(busiest
->lock
)
1705 __acquires(this_rq
->lock
)
1709 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1710 if (busiest
< this_rq
) {
1711 raw_spin_unlock(&this_rq
->lock
);
1712 raw_spin_lock(&busiest
->lock
);
1713 raw_spin_lock_nested(&this_rq
->lock
,
1714 SINGLE_DEPTH_NESTING
);
1717 raw_spin_lock_nested(&busiest
->lock
,
1718 SINGLE_DEPTH_NESTING
);
1723 #endif /* CONFIG_PREEMPT */
1726 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1728 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1730 if (unlikely(!irqs_disabled())) {
1731 /* printk() doesn't work good under rq->lock */
1732 raw_spin_unlock(&this_rq
->lock
);
1736 return _double_lock_balance(this_rq
, busiest
);
1739 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1740 __releases(busiest
->lock
)
1742 raw_spin_unlock(&busiest
->lock
);
1743 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1747 * double_rq_lock - safely lock two runqueues
1749 * Note this does not disable interrupts like task_rq_lock,
1750 * you need to do so manually before calling.
1752 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1753 __acquires(rq1
->lock
)
1754 __acquires(rq2
->lock
)
1756 BUG_ON(!irqs_disabled());
1758 raw_spin_lock(&rq1
->lock
);
1759 __acquire(rq2
->lock
); /* Fake it out ;) */
1762 raw_spin_lock(&rq1
->lock
);
1763 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1765 raw_spin_lock(&rq2
->lock
);
1766 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1769 update_rq_clock(rq1
);
1770 update_rq_clock(rq2
);
1774 * double_rq_unlock - safely unlock two runqueues
1776 * Note this does not restore interrupts like task_rq_unlock,
1777 * you need to do so manually after calling.
1779 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1780 __releases(rq1
->lock
)
1781 __releases(rq2
->lock
)
1783 raw_spin_unlock(&rq1
->lock
);
1785 raw_spin_unlock(&rq2
->lock
);
1787 __release(rq2
->lock
);
1792 #ifdef CONFIG_FAIR_GROUP_SCHED
1793 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1796 cfs_rq
->shares
= shares
;
1801 static void calc_load_account_active(struct rq
*this_rq
);
1802 static void update_sysctl(void);
1803 static int get_update_sysctl_factor(void);
1805 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1807 set_task_rq(p
, cpu
);
1810 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1811 * successfuly executed on another CPU. We must ensure that updates of
1812 * per-task data have been completed by this moment.
1815 task_thread_info(p
)->cpu
= cpu
;
1819 static const struct sched_class rt_sched_class
;
1821 #define sched_class_highest (&rt_sched_class)
1822 #define for_each_class(class) \
1823 for (class = sched_class_highest; class; class = class->next)
1825 #include "sched_stats.h"
1827 static void inc_nr_running(struct rq
*rq
)
1832 static void dec_nr_running(struct rq
*rq
)
1837 static void set_load_weight(struct task_struct
*p
)
1839 if (task_has_rt_policy(p
)) {
1840 p
->se
.load
.weight
= 0;
1841 p
->se
.load
.inv_weight
= WMULT_CONST
;
1846 * SCHED_IDLE tasks get minimal weight:
1848 if (p
->policy
== SCHED_IDLE
) {
1849 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1850 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1854 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1855 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1858 static void update_avg(u64
*avg
, u64 sample
)
1860 s64 diff
= sample
- *avg
;
1865 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1868 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1870 sched_info_queued(p
);
1871 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1875 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1878 if (p
->se
.last_wakeup
) {
1879 update_avg(&p
->se
.avg_overlap
,
1880 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1881 p
->se
.last_wakeup
= 0;
1883 update_avg(&p
->se
.avg_wakeup
,
1884 sysctl_sched_wakeup_granularity
);
1888 sched_info_dequeued(p
);
1889 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1894 * activate_task - move a task to the runqueue.
1896 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1898 if (task_contributes_to_load(p
))
1899 rq
->nr_uninterruptible
--;
1901 enqueue_task(rq
, p
, wakeup
, false);
1906 * deactivate_task - remove a task from the runqueue.
1908 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1910 if (task_contributes_to_load(p
))
1911 rq
->nr_uninterruptible
++;
1913 dequeue_task(rq
, p
, sleep
);
1917 #include "sched_idletask.c"
1918 #include "sched_fair.c"
1919 #include "sched_rt.c"
1920 #ifdef CONFIG_SCHED_DEBUG
1921 # include "sched_debug.c"
1925 * __normal_prio - return the priority that is based on the static prio
1927 static inline int __normal_prio(struct task_struct
*p
)
1929 return p
->static_prio
;
1933 * Calculate the expected normal priority: i.e. priority
1934 * without taking RT-inheritance into account. Might be
1935 * boosted by interactivity modifiers. Changes upon fork,
1936 * setprio syscalls, and whenever the interactivity
1937 * estimator recalculates.
1939 static inline int normal_prio(struct task_struct
*p
)
1943 if (task_has_rt_policy(p
))
1944 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1946 prio
= __normal_prio(p
);
1951 * Calculate the current priority, i.e. the priority
1952 * taken into account by the scheduler. This value might
1953 * be boosted by RT tasks, or might be boosted by
1954 * interactivity modifiers. Will be RT if the task got
1955 * RT-boosted. If not then it returns p->normal_prio.
1957 static int effective_prio(struct task_struct
*p
)
1959 p
->normal_prio
= normal_prio(p
);
1961 * If we are RT tasks or we were boosted to RT priority,
1962 * keep the priority unchanged. Otherwise, update priority
1963 * to the normal priority:
1965 if (!rt_prio(p
->prio
))
1966 return p
->normal_prio
;
1971 * task_curr - is this task currently executing on a CPU?
1972 * @p: the task in question.
1974 inline int task_curr(const struct task_struct
*p
)
1976 return cpu_curr(task_cpu(p
)) == p
;
1979 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1980 const struct sched_class
*prev_class
,
1981 int oldprio
, int running
)
1983 if (prev_class
!= p
->sched_class
) {
1984 if (prev_class
->switched_from
)
1985 prev_class
->switched_from(rq
, p
, running
);
1986 p
->sched_class
->switched_to(rq
, p
, running
);
1988 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1993 * Is this task likely cache-hot:
1996 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2000 if (p
->sched_class
!= &fair_sched_class
)
2004 * Buddy candidates are cache hot:
2006 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2007 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2008 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2011 if (sysctl_sched_migration_cost
== -1)
2013 if (sysctl_sched_migration_cost
== 0)
2016 delta
= now
- p
->se
.exec_start
;
2018 return delta
< (s64
)sysctl_sched_migration_cost
;
2021 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2023 #ifdef CONFIG_SCHED_DEBUG
2025 * We should never call set_task_cpu() on a blocked task,
2026 * ttwu() will sort out the placement.
2028 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2029 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2032 trace_sched_migrate_task(p
, new_cpu
);
2034 if (task_cpu(p
) != new_cpu
) {
2035 p
->se
.nr_migrations
++;
2036 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2039 __set_task_cpu(p
, new_cpu
);
2042 struct migration_req
{
2043 struct list_head list
;
2045 struct task_struct
*task
;
2048 struct completion done
;
2052 * The task's runqueue lock must be held.
2053 * Returns true if you have to wait for migration thread.
2056 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2058 struct rq
*rq
= task_rq(p
);
2061 * If the task is not on a runqueue (and not running), then
2062 * the next wake-up will properly place the task.
2064 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2067 init_completion(&req
->done
);
2069 req
->dest_cpu
= dest_cpu
;
2070 list_add(&req
->list
, &rq
->migration_queue
);
2076 * wait_task_context_switch - wait for a thread to complete at least one
2079 * @p must not be current.
2081 void wait_task_context_switch(struct task_struct
*p
)
2083 unsigned long nvcsw
, nivcsw
, flags
;
2091 * The runqueue is assigned before the actual context
2092 * switch. We need to take the runqueue lock.
2094 * We could check initially without the lock but it is
2095 * very likely that we need to take the lock in every
2098 rq
= task_rq_lock(p
, &flags
);
2099 running
= task_running(rq
, p
);
2100 task_rq_unlock(rq
, &flags
);
2102 if (likely(!running
))
2105 * The switch count is incremented before the actual
2106 * context switch. We thus wait for two switches to be
2107 * sure at least one completed.
2109 if ((p
->nvcsw
- nvcsw
) > 1)
2111 if ((p
->nivcsw
- nivcsw
) > 1)
2119 * wait_task_inactive - wait for a thread to unschedule.
2121 * If @match_state is nonzero, it's the @p->state value just checked and
2122 * not expected to change. If it changes, i.e. @p might have woken up,
2123 * then return zero. When we succeed in waiting for @p to be off its CPU,
2124 * we return a positive number (its total switch count). If a second call
2125 * a short while later returns the same number, the caller can be sure that
2126 * @p has remained unscheduled the whole time.
2128 * The caller must ensure that the task *will* unschedule sometime soon,
2129 * else this function might spin for a *long* time. This function can't
2130 * be called with interrupts off, or it may introduce deadlock with
2131 * smp_call_function() if an IPI is sent by the same process we are
2132 * waiting to become inactive.
2134 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2136 unsigned long flags
;
2143 * We do the initial early heuristics without holding
2144 * any task-queue locks at all. We'll only try to get
2145 * the runqueue lock when things look like they will
2151 * If the task is actively running on another CPU
2152 * still, just relax and busy-wait without holding
2155 * NOTE! Since we don't hold any locks, it's not
2156 * even sure that "rq" stays as the right runqueue!
2157 * But we don't care, since "task_running()" will
2158 * return false if the runqueue has changed and p
2159 * is actually now running somewhere else!
2161 while (task_running(rq
, p
)) {
2162 if (match_state
&& unlikely(p
->state
!= match_state
))
2168 * Ok, time to look more closely! We need the rq
2169 * lock now, to be *sure*. If we're wrong, we'll
2170 * just go back and repeat.
2172 rq
= task_rq_lock(p
, &flags
);
2173 trace_sched_wait_task(rq
, p
);
2174 running
= task_running(rq
, p
);
2175 on_rq
= p
->se
.on_rq
;
2177 if (!match_state
|| p
->state
== match_state
)
2178 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2179 task_rq_unlock(rq
, &flags
);
2182 * If it changed from the expected state, bail out now.
2184 if (unlikely(!ncsw
))
2188 * Was it really running after all now that we
2189 * checked with the proper locks actually held?
2191 * Oops. Go back and try again..
2193 if (unlikely(running
)) {
2199 * It's not enough that it's not actively running,
2200 * it must be off the runqueue _entirely_, and not
2203 * So if it was still runnable (but just not actively
2204 * running right now), it's preempted, and we should
2205 * yield - it could be a while.
2207 if (unlikely(on_rq
)) {
2208 schedule_timeout_uninterruptible(1);
2213 * Ahh, all good. It wasn't running, and it wasn't
2214 * runnable, which means that it will never become
2215 * running in the future either. We're all done!
2224 * kick_process - kick a running thread to enter/exit the kernel
2225 * @p: the to-be-kicked thread
2227 * Cause a process which is running on another CPU to enter
2228 * kernel-mode, without any delay. (to get signals handled.)
2230 * NOTE: this function doesnt have to take the runqueue lock,
2231 * because all it wants to ensure is that the remote task enters
2232 * the kernel. If the IPI races and the task has been migrated
2233 * to another CPU then no harm is done and the purpose has been
2236 void kick_process(struct task_struct
*p
)
2242 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2243 smp_send_reschedule(cpu
);
2246 EXPORT_SYMBOL_GPL(kick_process
);
2247 #endif /* CONFIG_SMP */
2250 * task_oncpu_function_call - call a function on the cpu on which a task runs
2251 * @p: the task to evaluate
2252 * @func: the function to be called
2253 * @info: the function call argument
2255 * Calls the function @func when the task is currently running. This might
2256 * be on the current CPU, which just calls the function directly
2258 void task_oncpu_function_call(struct task_struct
*p
,
2259 void (*func
) (void *info
), void *info
)
2266 smp_call_function_single(cpu
, func
, info
, 1);
2272 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2274 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2277 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2279 /* Look for allowed, online CPU in same node. */
2280 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2281 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2284 /* Any allowed, online CPU? */
2285 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2286 if (dest_cpu
< nr_cpu_ids
)
2289 /* No more Mr. Nice Guy. */
2290 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2291 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2293 * Don't tell them about moving exiting tasks or
2294 * kernel threads (both mm NULL), since they never
2297 if (p
->mm
&& printk_ratelimit()) {
2298 printk(KERN_INFO
"process %d (%s) no "
2299 "longer affine to cpu%d\n",
2300 task_pid_nr(p
), p
->comm
, cpu
);
2308 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2311 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2313 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2316 * In order not to call set_task_cpu() on a blocking task we need
2317 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2320 * Since this is common to all placement strategies, this lives here.
2322 * [ this allows ->select_task() to simply return task_cpu(p) and
2323 * not worry about this generic constraint ]
2325 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2327 cpu
= select_fallback_rq(task_cpu(p
), p
);
2334 * try_to_wake_up - wake up a thread
2335 * @p: the to-be-woken-up thread
2336 * @state: the mask of task states that can be woken
2337 * @sync: do a synchronous wakeup?
2339 * Put it on the run-queue if it's not already there. The "current"
2340 * thread is always on the run-queue (except when the actual
2341 * re-schedule is in progress), and as such you're allowed to do
2342 * the simpler "current->state = TASK_RUNNING" to mark yourself
2343 * runnable without the overhead of this.
2345 * returns failure only if the task is already active.
2347 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2350 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2351 unsigned long flags
;
2354 if (!sched_feat(SYNC_WAKEUPS
))
2355 wake_flags
&= ~WF_SYNC
;
2357 this_cpu
= get_cpu();
2360 rq
= task_rq_lock(p
, &flags
);
2361 update_rq_clock(rq
);
2362 if (!(p
->state
& state
))
2372 if (unlikely(task_running(rq
, p
)))
2376 * In order to handle concurrent wakeups and release the rq->lock
2377 * we put the task in TASK_WAKING state.
2379 * First fix up the nr_uninterruptible count:
2381 if (task_contributes_to_load(p
)) {
2382 if (likely(cpu_online(orig_cpu
)))
2383 rq
->nr_uninterruptible
--;
2385 this_rq()->nr_uninterruptible
--;
2387 p
->state
= TASK_WAKING
;
2389 if (p
->sched_class
->task_waking
)
2390 p
->sched_class
->task_waking(rq
, p
);
2392 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2393 if (cpu
!= orig_cpu
)
2394 set_task_cpu(p
, cpu
);
2395 __task_rq_unlock(rq
);
2398 raw_spin_lock(&rq
->lock
);
2399 update_rq_clock(rq
);
2402 * We migrated the task without holding either rq->lock, however
2403 * since the task is not on the task list itself, nobody else
2404 * will try and migrate the task, hence the rq should match the
2405 * cpu we just moved it to.
2407 WARN_ON(task_cpu(p
) != cpu
);
2408 WARN_ON(p
->state
!= TASK_WAKING
);
2410 #ifdef CONFIG_SCHEDSTATS
2411 schedstat_inc(rq
, ttwu_count
);
2412 if (cpu
== this_cpu
)
2413 schedstat_inc(rq
, ttwu_local
);
2415 struct sched_domain
*sd
;
2416 for_each_domain(this_cpu
, sd
) {
2417 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2418 schedstat_inc(sd
, ttwu_wake_remote
);
2423 #endif /* CONFIG_SCHEDSTATS */
2426 #endif /* CONFIG_SMP */
2427 schedstat_inc(p
, se
.nr_wakeups
);
2428 if (wake_flags
& WF_SYNC
)
2429 schedstat_inc(p
, se
.nr_wakeups_sync
);
2430 if (orig_cpu
!= cpu
)
2431 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2432 if (cpu
== this_cpu
)
2433 schedstat_inc(p
, se
.nr_wakeups_local
);
2435 schedstat_inc(p
, se
.nr_wakeups_remote
);
2436 activate_task(rq
, p
, 1);
2440 * Only attribute actual wakeups done by this task.
2442 if (!in_interrupt()) {
2443 struct sched_entity
*se
= ¤t
->se
;
2444 u64 sample
= se
->sum_exec_runtime
;
2446 if (se
->last_wakeup
)
2447 sample
-= se
->last_wakeup
;
2449 sample
-= se
->start_runtime
;
2450 update_avg(&se
->avg_wakeup
, sample
);
2452 se
->last_wakeup
= se
->sum_exec_runtime
;
2456 trace_sched_wakeup(rq
, p
, success
);
2457 check_preempt_curr(rq
, p
, wake_flags
);
2459 p
->state
= TASK_RUNNING
;
2461 if (p
->sched_class
->task_woken
)
2462 p
->sched_class
->task_woken(rq
, p
);
2464 if (unlikely(rq
->idle_stamp
)) {
2465 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2466 u64 max
= 2*sysctl_sched_migration_cost
;
2471 update_avg(&rq
->avg_idle
, delta
);
2476 task_rq_unlock(rq
, &flags
);
2483 * wake_up_process - Wake up a specific process
2484 * @p: The process to be woken up.
2486 * Attempt to wake up the nominated process and move it to the set of runnable
2487 * processes. Returns 1 if the process was woken up, 0 if it was already
2490 * It may be assumed that this function implies a write memory barrier before
2491 * changing the task state if and only if any tasks are woken up.
2493 int wake_up_process(struct task_struct
*p
)
2495 return try_to_wake_up(p
, TASK_ALL
, 0);
2497 EXPORT_SYMBOL(wake_up_process
);
2499 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2501 return try_to_wake_up(p
, state
, 0);
2505 * Perform scheduler related setup for a newly forked process p.
2506 * p is forked by current.
2508 * __sched_fork() is basic setup used by init_idle() too:
2510 static void __sched_fork(struct task_struct
*p
)
2512 p
->se
.exec_start
= 0;
2513 p
->se
.sum_exec_runtime
= 0;
2514 p
->se
.prev_sum_exec_runtime
= 0;
2515 p
->se
.nr_migrations
= 0;
2516 p
->se
.last_wakeup
= 0;
2517 p
->se
.avg_overlap
= 0;
2518 p
->se
.start_runtime
= 0;
2519 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2521 #ifdef CONFIG_SCHEDSTATS
2522 p
->se
.wait_start
= 0;
2524 p
->se
.wait_count
= 0;
2527 p
->se
.sleep_start
= 0;
2528 p
->se
.sleep_max
= 0;
2529 p
->se
.sum_sleep_runtime
= 0;
2531 p
->se
.block_start
= 0;
2532 p
->se
.block_max
= 0;
2534 p
->se
.slice_max
= 0;
2536 p
->se
.nr_migrations_cold
= 0;
2537 p
->se
.nr_failed_migrations_affine
= 0;
2538 p
->se
.nr_failed_migrations_running
= 0;
2539 p
->se
.nr_failed_migrations_hot
= 0;
2540 p
->se
.nr_forced_migrations
= 0;
2542 p
->se
.nr_wakeups
= 0;
2543 p
->se
.nr_wakeups_sync
= 0;
2544 p
->se
.nr_wakeups_migrate
= 0;
2545 p
->se
.nr_wakeups_local
= 0;
2546 p
->se
.nr_wakeups_remote
= 0;
2547 p
->se
.nr_wakeups_affine
= 0;
2548 p
->se
.nr_wakeups_affine_attempts
= 0;
2549 p
->se
.nr_wakeups_passive
= 0;
2550 p
->se
.nr_wakeups_idle
= 0;
2554 INIT_LIST_HEAD(&p
->rt
.run_list
);
2556 INIT_LIST_HEAD(&p
->se
.group_node
);
2558 #ifdef CONFIG_PREEMPT_NOTIFIERS
2559 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2564 * fork()/clone()-time setup:
2566 void sched_fork(struct task_struct
*p
, int clone_flags
)
2568 int cpu
= get_cpu();
2572 * We mark the process as running here. This guarantees that
2573 * nobody will actually run it, and a signal or other external
2574 * event cannot wake it up and insert it on the runqueue either.
2576 p
->state
= TASK_RUNNING
;
2579 * Revert to default priority/policy on fork if requested.
2581 if (unlikely(p
->sched_reset_on_fork
)) {
2582 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2583 p
->policy
= SCHED_NORMAL
;
2584 p
->normal_prio
= p
->static_prio
;
2587 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2588 p
->static_prio
= NICE_TO_PRIO(0);
2589 p
->normal_prio
= p
->static_prio
;
2594 * We don't need the reset flag anymore after the fork. It has
2595 * fulfilled its duty:
2597 p
->sched_reset_on_fork
= 0;
2601 * Make sure we do not leak PI boosting priority to the child.
2603 p
->prio
= current
->normal_prio
;
2605 if (!rt_prio(p
->prio
))
2606 p
->sched_class
= &fair_sched_class
;
2608 if (p
->sched_class
->task_fork
)
2609 p
->sched_class
->task_fork(p
);
2611 set_task_cpu(p
, cpu
);
2613 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2614 if (likely(sched_info_on()))
2615 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2617 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2620 #ifdef CONFIG_PREEMPT
2621 /* Want to start with kernel preemption disabled. */
2622 task_thread_info(p
)->preempt_count
= 1;
2624 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2630 * wake_up_new_task - wake up a newly created task for the first time.
2632 * This function will do some initial scheduler statistics housekeeping
2633 * that must be done for every newly created context, then puts the task
2634 * on the runqueue and wakes it.
2636 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2638 unsigned long flags
;
2640 int cpu __maybe_unused
= get_cpu();
2643 rq
= task_rq_lock(p
, &flags
);
2644 p
->state
= TASK_WAKING
;
2647 * Fork balancing, do it here and not earlier because:
2648 * - cpus_allowed can change in the fork path
2649 * - any previously selected cpu might disappear through hotplug
2651 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2652 * without people poking at ->cpus_allowed.
2654 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2655 set_task_cpu(p
, cpu
);
2657 p
->state
= TASK_RUNNING
;
2658 task_rq_unlock(rq
, &flags
);
2661 rq
= task_rq_lock(p
, &flags
);
2662 update_rq_clock(rq
);
2663 activate_task(rq
, p
, 0);
2664 trace_sched_wakeup_new(rq
, p
, 1);
2665 check_preempt_curr(rq
, p
, WF_FORK
);
2667 if (p
->sched_class
->task_woken
)
2668 p
->sched_class
->task_woken(rq
, p
);
2670 task_rq_unlock(rq
, &flags
);
2674 #ifdef CONFIG_PREEMPT_NOTIFIERS
2677 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2678 * @notifier: notifier struct to register
2680 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2682 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2684 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2687 * preempt_notifier_unregister - no longer interested in preemption notifications
2688 * @notifier: notifier struct to unregister
2690 * This is safe to call from within a preemption notifier.
2692 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2694 hlist_del(¬ifier
->link
);
2696 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2698 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2700 struct preempt_notifier
*notifier
;
2701 struct hlist_node
*node
;
2703 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2704 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2708 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2709 struct task_struct
*next
)
2711 struct preempt_notifier
*notifier
;
2712 struct hlist_node
*node
;
2714 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2715 notifier
->ops
->sched_out(notifier
, next
);
2718 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2720 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2725 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2726 struct task_struct
*next
)
2730 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2733 * prepare_task_switch - prepare to switch tasks
2734 * @rq: the runqueue preparing to switch
2735 * @prev: the current task that is being switched out
2736 * @next: the task we are going to switch to.
2738 * This is called with the rq lock held and interrupts off. It must
2739 * be paired with a subsequent finish_task_switch after the context
2742 * prepare_task_switch sets up locking and calls architecture specific
2746 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2747 struct task_struct
*next
)
2749 fire_sched_out_preempt_notifiers(prev
, next
);
2750 prepare_lock_switch(rq
, next
);
2751 prepare_arch_switch(next
);
2755 * finish_task_switch - clean up after a task-switch
2756 * @rq: runqueue associated with task-switch
2757 * @prev: the thread we just switched away from.
2759 * finish_task_switch must be called after the context switch, paired
2760 * with a prepare_task_switch call before the context switch.
2761 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2762 * and do any other architecture-specific cleanup actions.
2764 * Note that we may have delayed dropping an mm in context_switch(). If
2765 * so, we finish that here outside of the runqueue lock. (Doing it
2766 * with the lock held can cause deadlocks; see schedule() for
2769 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2770 __releases(rq
->lock
)
2772 struct mm_struct
*mm
= rq
->prev_mm
;
2778 * A task struct has one reference for the use as "current".
2779 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2780 * schedule one last time. The schedule call will never return, and
2781 * the scheduled task must drop that reference.
2782 * The test for TASK_DEAD must occur while the runqueue locks are
2783 * still held, otherwise prev could be scheduled on another cpu, die
2784 * there before we look at prev->state, and then the reference would
2786 * Manfred Spraul <manfred@colorfullife.com>
2788 prev_state
= prev
->state
;
2789 finish_arch_switch(prev
);
2790 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2791 local_irq_disable();
2792 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2793 perf_event_task_sched_in(current
);
2794 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2796 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2797 finish_lock_switch(rq
, prev
);
2799 fire_sched_in_preempt_notifiers(current
);
2802 if (unlikely(prev_state
== TASK_DEAD
)) {
2804 * Remove function-return probe instances associated with this
2805 * task and put them back on the free list.
2807 kprobe_flush_task(prev
);
2808 put_task_struct(prev
);
2814 /* assumes rq->lock is held */
2815 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2817 if (prev
->sched_class
->pre_schedule
)
2818 prev
->sched_class
->pre_schedule(rq
, prev
);
2821 /* rq->lock is NOT held, but preemption is disabled */
2822 static inline void post_schedule(struct rq
*rq
)
2824 if (rq
->post_schedule
) {
2825 unsigned long flags
;
2827 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2828 if (rq
->curr
->sched_class
->post_schedule
)
2829 rq
->curr
->sched_class
->post_schedule(rq
);
2830 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2832 rq
->post_schedule
= 0;
2838 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2842 static inline void post_schedule(struct rq
*rq
)
2849 * schedule_tail - first thing a freshly forked thread must call.
2850 * @prev: the thread we just switched away from.
2852 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2853 __releases(rq
->lock
)
2855 struct rq
*rq
= this_rq();
2857 finish_task_switch(rq
, prev
);
2860 * FIXME: do we need to worry about rq being invalidated by the
2865 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2866 /* In this case, finish_task_switch does not reenable preemption */
2869 if (current
->set_child_tid
)
2870 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2874 * context_switch - switch to the new MM and the new
2875 * thread's register state.
2878 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2879 struct task_struct
*next
)
2881 struct mm_struct
*mm
, *oldmm
;
2883 prepare_task_switch(rq
, prev
, next
);
2884 trace_sched_switch(rq
, prev
, next
);
2886 oldmm
= prev
->active_mm
;
2888 * For paravirt, this is coupled with an exit in switch_to to
2889 * combine the page table reload and the switch backend into
2892 arch_start_context_switch(prev
);
2895 next
->active_mm
= oldmm
;
2896 atomic_inc(&oldmm
->mm_count
);
2897 enter_lazy_tlb(oldmm
, next
);
2899 switch_mm(oldmm
, mm
, next
);
2901 if (likely(!prev
->mm
)) {
2902 prev
->active_mm
= NULL
;
2903 rq
->prev_mm
= oldmm
;
2906 * Since the runqueue lock will be released by the next
2907 * task (which is an invalid locking op but in the case
2908 * of the scheduler it's an obvious special-case), so we
2909 * do an early lockdep release here:
2911 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2912 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2915 /* Here we just switch the register state and the stack. */
2916 switch_to(prev
, next
, prev
);
2920 * this_rq must be evaluated again because prev may have moved
2921 * CPUs since it called schedule(), thus the 'rq' on its stack
2922 * frame will be invalid.
2924 finish_task_switch(this_rq(), prev
);
2928 * nr_running, nr_uninterruptible and nr_context_switches:
2930 * externally visible scheduler statistics: current number of runnable
2931 * threads, current number of uninterruptible-sleeping threads, total
2932 * number of context switches performed since bootup.
2934 unsigned long nr_running(void)
2936 unsigned long i
, sum
= 0;
2938 for_each_online_cpu(i
)
2939 sum
+= cpu_rq(i
)->nr_running
;
2944 unsigned long nr_uninterruptible(void)
2946 unsigned long i
, sum
= 0;
2948 for_each_possible_cpu(i
)
2949 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2952 * Since we read the counters lockless, it might be slightly
2953 * inaccurate. Do not allow it to go below zero though:
2955 if (unlikely((long)sum
< 0))
2961 unsigned long long nr_context_switches(void)
2964 unsigned long long sum
= 0;
2966 for_each_possible_cpu(i
)
2967 sum
+= cpu_rq(i
)->nr_switches
;
2972 unsigned long nr_iowait(void)
2974 unsigned long i
, sum
= 0;
2976 for_each_possible_cpu(i
)
2977 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2982 unsigned long nr_iowait_cpu(void)
2984 struct rq
*this = this_rq();
2985 return atomic_read(&this->nr_iowait
);
2988 unsigned long this_cpu_load(void)
2990 struct rq
*this = this_rq();
2991 return this->cpu_load
[0];
2995 /* Variables and functions for calc_load */
2996 static atomic_long_t calc_load_tasks
;
2997 static unsigned long calc_load_update
;
2998 unsigned long avenrun
[3];
2999 EXPORT_SYMBOL(avenrun
);
3002 * get_avenrun - get the load average array
3003 * @loads: pointer to dest load array
3004 * @offset: offset to add
3005 * @shift: shift count to shift the result left
3007 * These values are estimates at best, so no need for locking.
3009 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3011 loads
[0] = (avenrun
[0] + offset
) << shift
;
3012 loads
[1] = (avenrun
[1] + offset
) << shift
;
3013 loads
[2] = (avenrun
[2] + offset
) << shift
;
3016 static unsigned long
3017 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3020 load
+= active
* (FIXED_1
- exp
);
3021 return load
>> FSHIFT
;
3025 * calc_load - update the avenrun load estimates 10 ticks after the
3026 * CPUs have updated calc_load_tasks.
3028 void calc_global_load(void)
3030 unsigned long upd
= calc_load_update
+ 10;
3033 if (time_before(jiffies
, upd
))
3036 active
= atomic_long_read(&calc_load_tasks
);
3037 active
= active
> 0 ? active
* FIXED_1
: 0;
3039 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3040 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3041 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3043 calc_load_update
+= LOAD_FREQ
;
3047 * Either called from update_cpu_load() or from a cpu going idle
3049 static void calc_load_account_active(struct rq
*this_rq
)
3051 long nr_active
, delta
;
3053 nr_active
= this_rq
->nr_running
;
3054 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3056 if (nr_active
!= this_rq
->calc_load_active
) {
3057 delta
= nr_active
- this_rq
->calc_load_active
;
3058 this_rq
->calc_load_active
= nr_active
;
3059 atomic_long_add(delta
, &calc_load_tasks
);
3064 * Update rq->cpu_load[] statistics. This function is usually called every
3065 * scheduler tick (TICK_NSEC).
3067 static void update_cpu_load(struct rq
*this_rq
)
3069 unsigned long this_load
= this_rq
->load
.weight
;
3072 this_rq
->nr_load_updates
++;
3074 /* Update our load: */
3075 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3076 unsigned long old_load
, new_load
;
3078 /* scale is effectively 1 << i now, and >> i divides by scale */
3080 old_load
= this_rq
->cpu_load
[i
];
3081 new_load
= this_load
;
3083 * Round up the averaging division if load is increasing. This
3084 * prevents us from getting stuck on 9 if the load is 10, for
3087 if (new_load
> old_load
)
3088 new_load
+= scale
-1;
3089 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3092 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3093 this_rq
->calc_load_update
+= LOAD_FREQ
;
3094 calc_load_account_active(this_rq
);
3097 sched_avg_update(this_rq
);
3103 * sched_exec - execve() is a valuable balancing opportunity, because at
3104 * this point the task has the smallest effective memory and cache footprint.
3106 void sched_exec(void)
3108 struct task_struct
*p
= current
;
3109 struct migration_req req
;
3110 unsigned long flags
;
3114 rq
= task_rq_lock(p
, &flags
);
3115 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3116 if (dest_cpu
== smp_processor_id())
3120 * select_task_rq() can race against ->cpus_allowed
3122 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3123 likely(cpu_active(dest_cpu
)) &&
3124 migrate_task(p
, dest_cpu
, &req
)) {
3125 /* Need to wait for migration thread (might exit: take ref). */
3126 struct task_struct
*mt
= rq
->migration_thread
;
3128 get_task_struct(mt
);
3129 task_rq_unlock(rq
, &flags
);
3130 wake_up_process(mt
);
3131 put_task_struct(mt
);
3132 wait_for_completion(&req
.done
);
3137 task_rq_unlock(rq
, &flags
);
3142 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3144 EXPORT_PER_CPU_SYMBOL(kstat
);
3147 * Return any ns on the sched_clock that have not yet been accounted in
3148 * @p in case that task is currently running.
3150 * Called with task_rq_lock() held on @rq.
3152 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3156 if (task_current(rq
, p
)) {
3157 update_rq_clock(rq
);
3158 ns
= rq
->clock
- p
->se
.exec_start
;
3166 unsigned long long task_delta_exec(struct task_struct
*p
)
3168 unsigned long flags
;
3172 rq
= task_rq_lock(p
, &flags
);
3173 ns
= do_task_delta_exec(p
, rq
);
3174 task_rq_unlock(rq
, &flags
);
3180 * Return accounted runtime for the task.
3181 * In case the task is currently running, return the runtime plus current's
3182 * pending runtime that have not been accounted yet.
3184 unsigned long long task_sched_runtime(struct task_struct
*p
)
3186 unsigned long flags
;
3190 rq
= task_rq_lock(p
, &flags
);
3191 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3192 task_rq_unlock(rq
, &flags
);
3198 * Return sum_exec_runtime for the thread group.
3199 * In case the task is currently running, return the sum plus current's
3200 * pending runtime that have not been accounted yet.
3202 * Note that the thread group might have other running tasks as well,
3203 * so the return value not includes other pending runtime that other
3204 * running tasks might have.
3206 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3208 struct task_cputime totals
;
3209 unsigned long flags
;
3213 rq
= task_rq_lock(p
, &flags
);
3214 thread_group_cputime(p
, &totals
);
3215 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3216 task_rq_unlock(rq
, &flags
);
3222 * Account user cpu time to a process.
3223 * @p: the process that the cpu time gets accounted to
3224 * @cputime: the cpu time spent in user space since the last update
3225 * @cputime_scaled: cputime scaled by cpu frequency
3227 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3228 cputime_t cputime_scaled
)
3230 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3233 /* Add user time to process. */
3234 p
->utime
= cputime_add(p
->utime
, cputime
);
3235 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3236 account_group_user_time(p
, cputime
);
3238 /* Add user time to cpustat. */
3239 tmp
= cputime_to_cputime64(cputime
);
3240 if (TASK_NICE(p
) > 0)
3241 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3243 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3245 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3246 /* Account for user time used */
3247 acct_update_integrals(p
);
3251 * Account guest cpu time to a process.
3252 * @p: the process that the cpu time gets accounted to
3253 * @cputime: the cpu time spent in virtual machine since the last update
3254 * @cputime_scaled: cputime scaled by cpu frequency
3256 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3257 cputime_t cputime_scaled
)
3260 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3262 tmp
= cputime_to_cputime64(cputime
);
3264 /* Add guest time to process. */
3265 p
->utime
= cputime_add(p
->utime
, cputime
);
3266 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3267 account_group_user_time(p
, cputime
);
3268 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3270 /* Add guest time to cpustat. */
3271 if (TASK_NICE(p
) > 0) {
3272 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3273 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3275 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3276 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3281 * Account system cpu time to a process.
3282 * @p: the process that the cpu time gets accounted to
3283 * @hardirq_offset: the offset to subtract from hardirq_count()
3284 * @cputime: the cpu time spent in kernel space since the last update
3285 * @cputime_scaled: cputime scaled by cpu frequency
3287 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3288 cputime_t cputime
, cputime_t cputime_scaled
)
3290 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3293 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3294 account_guest_time(p
, cputime
, cputime_scaled
);
3298 /* Add system time to process. */
3299 p
->stime
= cputime_add(p
->stime
, cputime
);
3300 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3301 account_group_system_time(p
, cputime
);
3303 /* Add system time to cpustat. */
3304 tmp
= cputime_to_cputime64(cputime
);
3305 if (hardirq_count() - hardirq_offset
)
3306 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3307 else if (softirq_count())
3308 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3310 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3312 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3314 /* Account for system time used */
3315 acct_update_integrals(p
);
3319 * Account for involuntary wait time.
3320 * @steal: the cpu time spent in involuntary wait
3322 void account_steal_time(cputime_t cputime
)
3324 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3325 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3327 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3331 * Account for idle time.
3332 * @cputime: the cpu time spent in idle wait
3334 void account_idle_time(cputime_t cputime
)
3336 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3337 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3338 struct rq
*rq
= this_rq();
3340 if (atomic_read(&rq
->nr_iowait
) > 0)
3341 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3343 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3346 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3349 * Account a single tick of cpu time.
3350 * @p: the process that the cpu time gets accounted to
3351 * @user_tick: indicates if the tick is a user or a system tick
3353 void account_process_tick(struct task_struct
*p
, int user_tick
)
3355 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3356 struct rq
*rq
= this_rq();
3359 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3360 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3361 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3364 account_idle_time(cputime_one_jiffy
);
3368 * Account multiple ticks of steal time.
3369 * @p: the process from which the cpu time has been stolen
3370 * @ticks: number of stolen ticks
3372 void account_steal_ticks(unsigned long ticks
)
3374 account_steal_time(jiffies_to_cputime(ticks
));
3378 * Account multiple ticks of idle time.
3379 * @ticks: number of stolen ticks
3381 void account_idle_ticks(unsigned long ticks
)
3383 account_idle_time(jiffies_to_cputime(ticks
));
3389 * Use precise platform statistics if available:
3391 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3392 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3398 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3400 struct task_cputime cputime
;
3402 thread_group_cputime(p
, &cputime
);
3404 *ut
= cputime
.utime
;
3405 *st
= cputime
.stime
;
3409 #ifndef nsecs_to_cputime
3410 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3413 static cputime_t
scale_utime(cputime_t utime
, cputime_t rtime
, cputime_t total
)
3415 u64 temp
= (__force u64
) rtime
;
3417 temp
*= (__force u64
) utime
;
3419 if (sizeof(cputime_t
) == 4)
3420 temp
= div_u64(temp
, (__force u32
) total
);
3422 temp
= div64_u64(temp
, (__force u64
) total
);
3424 return (__force cputime_t
) temp
;
3427 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3429 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3432 * Use CFS's precise accounting:
3434 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3437 utime
= scale_utime(utime
, rtime
, total
);
3442 * Compare with previous values, to keep monotonicity:
3444 p
->prev_utime
= max(p
->prev_utime
, utime
);
3445 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3447 *ut
= p
->prev_utime
;
3448 *st
= p
->prev_stime
;
3452 * Must be called with siglock held.
3454 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3456 struct signal_struct
*sig
= p
->signal
;
3457 struct task_cputime cputime
;
3458 cputime_t rtime
, utime
, total
;
3460 thread_group_cputime(p
, &cputime
);
3462 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3463 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3466 utime
= scale_utime(cputime
.utime
, rtime
, total
);
3470 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3471 sig
->prev_stime
= max(sig
->prev_stime
,
3472 cputime_sub(rtime
, sig
->prev_utime
));
3474 *ut
= sig
->prev_utime
;
3475 *st
= sig
->prev_stime
;
3480 * This function gets called by the timer code, with HZ frequency.
3481 * We call it with interrupts disabled.
3483 * It also gets called by the fork code, when changing the parent's
3486 void scheduler_tick(void)
3488 int cpu
= smp_processor_id();
3489 struct rq
*rq
= cpu_rq(cpu
);
3490 struct task_struct
*curr
= rq
->curr
;
3494 raw_spin_lock(&rq
->lock
);
3495 update_rq_clock(rq
);
3496 update_cpu_load(rq
);
3497 curr
->sched_class
->task_tick(rq
, curr
, 0);
3498 raw_spin_unlock(&rq
->lock
);
3500 perf_event_task_tick(curr
);
3503 rq
->idle_at_tick
= idle_cpu(cpu
);
3504 trigger_load_balance(rq
, cpu
);
3508 notrace
unsigned long get_parent_ip(unsigned long addr
)
3510 if (in_lock_functions(addr
)) {
3511 addr
= CALLER_ADDR2
;
3512 if (in_lock_functions(addr
))
3513 addr
= CALLER_ADDR3
;
3518 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3519 defined(CONFIG_PREEMPT_TRACER))
3521 void __kprobes
add_preempt_count(int val
)
3523 #ifdef CONFIG_DEBUG_PREEMPT
3527 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3530 preempt_count() += val
;
3531 #ifdef CONFIG_DEBUG_PREEMPT
3533 * Spinlock count overflowing soon?
3535 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3538 if (preempt_count() == val
)
3539 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3541 EXPORT_SYMBOL(add_preempt_count
);
3543 void __kprobes
sub_preempt_count(int val
)
3545 #ifdef CONFIG_DEBUG_PREEMPT
3549 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3552 * Is the spinlock portion underflowing?
3554 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3555 !(preempt_count() & PREEMPT_MASK
)))
3559 if (preempt_count() == val
)
3560 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3561 preempt_count() -= val
;
3563 EXPORT_SYMBOL(sub_preempt_count
);
3568 * Print scheduling while atomic bug:
3570 static noinline
void __schedule_bug(struct task_struct
*prev
)
3572 struct pt_regs
*regs
= get_irq_regs();
3574 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3575 prev
->comm
, prev
->pid
, preempt_count());
3577 debug_show_held_locks(prev
);
3579 if (irqs_disabled())
3580 print_irqtrace_events(prev
);
3589 * Various schedule()-time debugging checks and statistics:
3591 static inline void schedule_debug(struct task_struct
*prev
)
3594 * Test if we are atomic. Since do_exit() needs to call into
3595 * schedule() atomically, we ignore that path for now.
3596 * Otherwise, whine if we are scheduling when we should not be.
3598 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3599 __schedule_bug(prev
);
3601 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3603 schedstat_inc(this_rq(), sched_count
);
3604 #ifdef CONFIG_SCHEDSTATS
3605 if (unlikely(prev
->lock_depth
>= 0)) {
3606 schedstat_inc(this_rq(), bkl_count
);
3607 schedstat_inc(prev
, sched_info
.bkl_count
);
3612 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3614 if (prev
->state
== TASK_RUNNING
) {
3615 u64 runtime
= prev
->se
.sum_exec_runtime
;
3617 runtime
-= prev
->se
.prev_sum_exec_runtime
;
3618 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
3621 * In order to avoid avg_overlap growing stale when we are
3622 * indeed overlapping and hence not getting put to sleep, grow
3623 * the avg_overlap on preemption.
3625 * We use the average preemption runtime because that
3626 * correlates to the amount of cache footprint a task can
3629 update_avg(&prev
->se
.avg_overlap
, runtime
);
3631 prev
->sched_class
->put_prev_task(rq
, prev
);
3635 * Pick up the highest-prio task:
3637 static inline struct task_struct
*
3638 pick_next_task(struct rq
*rq
)
3640 const struct sched_class
*class;
3641 struct task_struct
*p
;
3644 * Optimization: we know that if all tasks are in
3645 * the fair class we can call that function directly:
3647 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3648 p
= fair_sched_class
.pick_next_task(rq
);
3653 class = sched_class_highest
;
3655 p
= class->pick_next_task(rq
);
3659 * Will never be NULL as the idle class always
3660 * returns a non-NULL p:
3662 class = class->next
;
3667 * schedule() is the main scheduler function.
3669 asmlinkage
void __sched
schedule(void)
3671 struct task_struct
*prev
, *next
;
3672 unsigned long *switch_count
;
3678 cpu
= smp_processor_id();
3682 switch_count
= &prev
->nivcsw
;
3684 release_kernel_lock(prev
);
3685 need_resched_nonpreemptible
:
3687 schedule_debug(prev
);
3689 if (sched_feat(HRTICK
))
3692 raw_spin_lock_irq(&rq
->lock
);
3693 update_rq_clock(rq
);
3694 clear_tsk_need_resched(prev
);
3696 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3697 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3698 prev
->state
= TASK_RUNNING
;
3700 deactivate_task(rq
, prev
, 1);
3701 switch_count
= &prev
->nvcsw
;
3704 pre_schedule(rq
, prev
);
3706 if (unlikely(!rq
->nr_running
))
3707 idle_balance(cpu
, rq
);
3709 put_prev_task(rq
, prev
);
3710 next
= pick_next_task(rq
);
3712 if (likely(prev
!= next
)) {
3713 sched_info_switch(prev
, next
);
3714 perf_event_task_sched_out(prev
, next
);
3720 context_switch(rq
, prev
, next
); /* unlocks the rq */
3722 * the context switch might have flipped the stack from under
3723 * us, hence refresh the local variables.
3725 cpu
= smp_processor_id();
3728 raw_spin_unlock_irq(&rq
->lock
);
3732 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3734 switch_count
= &prev
->nivcsw
;
3735 goto need_resched_nonpreemptible
;
3738 preempt_enable_no_resched();
3742 EXPORT_SYMBOL(schedule
);
3744 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3746 * Look out! "owner" is an entirely speculative pointer
3747 * access and not reliable.
3749 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3754 if (!sched_feat(OWNER_SPIN
))
3757 #ifdef CONFIG_DEBUG_PAGEALLOC
3759 * Need to access the cpu field knowing that
3760 * DEBUG_PAGEALLOC could have unmapped it if
3761 * the mutex owner just released it and exited.
3763 if (probe_kernel_address(&owner
->cpu
, cpu
))
3770 * Even if the access succeeded (likely case),
3771 * the cpu field may no longer be valid.
3773 if (cpu
>= nr_cpumask_bits
)
3777 * We need to validate that we can do a
3778 * get_cpu() and that we have the percpu area.
3780 if (!cpu_online(cpu
))
3787 * Owner changed, break to re-assess state.
3789 if (lock
->owner
!= owner
)
3793 * Is that owner really running on that cpu?
3795 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3805 #ifdef CONFIG_PREEMPT
3807 * this is the entry point to schedule() from in-kernel preemption
3808 * off of preempt_enable. Kernel preemptions off return from interrupt
3809 * occur there and call schedule directly.
3811 asmlinkage
void __sched
preempt_schedule(void)
3813 struct thread_info
*ti
= current_thread_info();
3816 * If there is a non-zero preempt_count or interrupts are disabled,
3817 * we do not want to preempt the current task. Just return..
3819 if (likely(ti
->preempt_count
|| irqs_disabled()))
3823 add_preempt_count(PREEMPT_ACTIVE
);
3825 sub_preempt_count(PREEMPT_ACTIVE
);
3828 * Check again in case we missed a preemption opportunity
3829 * between schedule and now.
3832 } while (need_resched());
3834 EXPORT_SYMBOL(preempt_schedule
);
3837 * this is the entry point to schedule() from kernel preemption
3838 * off of irq context.
3839 * Note, that this is called and return with irqs disabled. This will
3840 * protect us against recursive calling from irq.
3842 asmlinkage
void __sched
preempt_schedule_irq(void)
3844 struct thread_info
*ti
= current_thread_info();
3846 /* Catch callers which need to be fixed */
3847 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3850 add_preempt_count(PREEMPT_ACTIVE
);
3853 local_irq_disable();
3854 sub_preempt_count(PREEMPT_ACTIVE
);
3857 * Check again in case we missed a preemption opportunity
3858 * between schedule and now.
3861 } while (need_resched());
3864 #endif /* CONFIG_PREEMPT */
3866 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3869 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3871 EXPORT_SYMBOL(default_wake_function
);
3874 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3875 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3876 * number) then we wake all the non-exclusive tasks and one exclusive task.
3878 * There are circumstances in which we can try to wake a task which has already
3879 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3880 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3882 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3883 int nr_exclusive
, int wake_flags
, void *key
)
3885 wait_queue_t
*curr
, *next
;
3887 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3888 unsigned flags
= curr
->flags
;
3890 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3891 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3897 * __wake_up - wake up threads blocked on a waitqueue.
3899 * @mode: which threads
3900 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3901 * @key: is directly passed to the wakeup function
3903 * It may be assumed that this function implies a write memory barrier before
3904 * changing the task state if and only if any tasks are woken up.
3906 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3907 int nr_exclusive
, void *key
)
3909 unsigned long flags
;
3911 spin_lock_irqsave(&q
->lock
, flags
);
3912 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3913 spin_unlock_irqrestore(&q
->lock
, flags
);
3915 EXPORT_SYMBOL(__wake_up
);
3918 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3920 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3922 __wake_up_common(q
, mode
, 1, 0, NULL
);
3925 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3927 __wake_up_common(q
, mode
, 1, 0, key
);
3931 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3933 * @mode: which threads
3934 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3935 * @key: opaque value to be passed to wakeup targets
3937 * The sync wakeup differs that the waker knows that it will schedule
3938 * away soon, so while the target thread will be woken up, it will not
3939 * be migrated to another CPU - ie. the two threads are 'synchronized'
3940 * with each other. This can prevent needless bouncing between CPUs.
3942 * On UP it can prevent extra preemption.
3944 * It may be assumed that this function implies a write memory barrier before
3945 * changing the task state if and only if any tasks are woken up.
3947 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3948 int nr_exclusive
, void *key
)
3950 unsigned long flags
;
3951 int wake_flags
= WF_SYNC
;
3956 if (unlikely(!nr_exclusive
))
3959 spin_lock_irqsave(&q
->lock
, flags
);
3960 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3961 spin_unlock_irqrestore(&q
->lock
, flags
);
3963 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3966 * __wake_up_sync - see __wake_up_sync_key()
3968 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3970 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3972 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3975 * complete: - signals a single thread waiting on this completion
3976 * @x: holds the state of this particular completion
3978 * This will wake up a single thread waiting on this completion. Threads will be
3979 * awakened in the same order in which they were queued.
3981 * See also complete_all(), wait_for_completion() and related routines.
3983 * It may be assumed that this function implies a write memory barrier before
3984 * changing the task state if and only if any tasks are woken up.
3986 void complete(struct completion
*x
)
3988 unsigned long flags
;
3990 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3992 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3993 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3995 EXPORT_SYMBOL(complete
);
3998 * complete_all: - signals all threads waiting on this completion
3999 * @x: holds the state of this particular completion
4001 * This will wake up all threads waiting on this particular completion event.
4003 * It may be assumed that this function implies a write memory barrier before
4004 * changing the task state if and only if any tasks are woken up.
4006 void complete_all(struct completion
*x
)
4008 unsigned long flags
;
4010 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4011 x
->done
+= UINT_MAX
/2;
4012 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4013 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4015 EXPORT_SYMBOL(complete_all
);
4017 static inline long __sched
4018 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4021 DECLARE_WAITQUEUE(wait
, current
);
4023 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4024 __add_wait_queue_tail(&x
->wait
, &wait
);
4026 if (signal_pending_state(state
, current
)) {
4027 timeout
= -ERESTARTSYS
;
4030 __set_current_state(state
);
4031 spin_unlock_irq(&x
->wait
.lock
);
4032 timeout
= schedule_timeout(timeout
);
4033 spin_lock_irq(&x
->wait
.lock
);
4034 } while (!x
->done
&& timeout
);
4035 __remove_wait_queue(&x
->wait
, &wait
);
4040 return timeout
?: 1;
4044 wait_for_common(struct completion
*x
, long timeout
, int state
)
4048 spin_lock_irq(&x
->wait
.lock
);
4049 timeout
= do_wait_for_common(x
, timeout
, state
);
4050 spin_unlock_irq(&x
->wait
.lock
);
4055 * wait_for_completion: - waits for completion of a task
4056 * @x: holds the state of this particular completion
4058 * This waits to be signaled for completion of a specific task. It is NOT
4059 * interruptible and there is no timeout.
4061 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4062 * and interrupt capability. Also see complete().
4064 void __sched
wait_for_completion(struct completion
*x
)
4066 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4068 EXPORT_SYMBOL(wait_for_completion
);
4071 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4072 * @x: holds the state of this particular completion
4073 * @timeout: timeout value in jiffies
4075 * This waits for either a completion of a specific task to be signaled or for a
4076 * specified timeout to expire. The timeout is in jiffies. It is not
4079 unsigned long __sched
4080 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4082 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4084 EXPORT_SYMBOL(wait_for_completion_timeout
);
4087 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4088 * @x: holds the state of this particular completion
4090 * This waits for completion of a specific task to be signaled. It is
4093 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4095 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4096 if (t
== -ERESTARTSYS
)
4100 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4103 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4104 * @x: holds the state of this particular completion
4105 * @timeout: timeout value in jiffies
4107 * This waits for either a completion of a specific task to be signaled or for a
4108 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4110 unsigned long __sched
4111 wait_for_completion_interruptible_timeout(struct completion
*x
,
4112 unsigned long timeout
)
4114 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4116 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4119 * wait_for_completion_killable: - waits for completion of a task (killable)
4120 * @x: holds the state of this particular completion
4122 * This waits to be signaled for completion of a specific task. It can be
4123 * interrupted by a kill signal.
4125 int __sched
wait_for_completion_killable(struct completion
*x
)
4127 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4128 if (t
== -ERESTARTSYS
)
4132 EXPORT_SYMBOL(wait_for_completion_killable
);
4135 * try_wait_for_completion - try to decrement a completion without blocking
4136 * @x: completion structure
4138 * Returns: 0 if a decrement cannot be done without blocking
4139 * 1 if a decrement succeeded.
4141 * If a completion is being used as a counting completion,
4142 * attempt to decrement the counter without blocking. This
4143 * enables us to avoid waiting if the resource the completion
4144 * is protecting is not available.
4146 bool try_wait_for_completion(struct completion
*x
)
4148 unsigned long flags
;
4151 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4156 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4159 EXPORT_SYMBOL(try_wait_for_completion
);
4162 * completion_done - Test to see if a completion has any waiters
4163 * @x: completion structure
4165 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4166 * 1 if there are no waiters.
4169 bool completion_done(struct completion
*x
)
4171 unsigned long flags
;
4174 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4177 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4180 EXPORT_SYMBOL(completion_done
);
4183 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4185 unsigned long flags
;
4188 init_waitqueue_entry(&wait
, current
);
4190 __set_current_state(state
);
4192 spin_lock_irqsave(&q
->lock
, flags
);
4193 __add_wait_queue(q
, &wait
);
4194 spin_unlock(&q
->lock
);
4195 timeout
= schedule_timeout(timeout
);
4196 spin_lock_irq(&q
->lock
);
4197 __remove_wait_queue(q
, &wait
);
4198 spin_unlock_irqrestore(&q
->lock
, flags
);
4203 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4205 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4207 EXPORT_SYMBOL(interruptible_sleep_on
);
4210 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4212 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4214 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4216 void __sched
sleep_on(wait_queue_head_t
*q
)
4218 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4220 EXPORT_SYMBOL(sleep_on
);
4222 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4224 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4226 EXPORT_SYMBOL(sleep_on_timeout
);
4228 #ifdef CONFIG_RT_MUTEXES
4231 * rt_mutex_setprio - set the current priority of a task
4233 * @prio: prio value (kernel-internal form)
4235 * This function changes the 'effective' priority of a task. It does
4236 * not touch ->normal_prio like __setscheduler().
4238 * Used by the rt_mutex code to implement priority inheritance logic.
4240 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4242 unsigned long flags
;
4243 int oldprio
, on_rq
, running
;
4245 const struct sched_class
*prev_class
;
4247 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4249 rq
= task_rq_lock(p
, &flags
);
4250 update_rq_clock(rq
);
4253 prev_class
= p
->sched_class
;
4254 on_rq
= p
->se
.on_rq
;
4255 running
= task_current(rq
, p
);
4257 dequeue_task(rq
, p
, 0);
4259 p
->sched_class
->put_prev_task(rq
, p
);
4262 p
->sched_class
= &rt_sched_class
;
4264 p
->sched_class
= &fair_sched_class
;
4269 p
->sched_class
->set_curr_task(rq
);
4271 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4273 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4275 task_rq_unlock(rq
, &flags
);
4280 void set_user_nice(struct task_struct
*p
, long nice
)
4282 int old_prio
, delta
, on_rq
;
4283 unsigned long flags
;
4286 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4289 * We have to be careful, if called from sys_setpriority(),
4290 * the task might be in the middle of scheduling on another CPU.
4292 rq
= task_rq_lock(p
, &flags
);
4293 update_rq_clock(rq
);
4295 * The RT priorities are set via sched_setscheduler(), but we still
4296 * allow the 'normal' nice value to be set - but as expected
4297 * it wont have any effect on scheduling until the task is
4298 * SCHED_FIFO/SCHED_RR:
4300 if (task_has_rt_policy(p
)) {
4301 p
->static_prio
= NICE_TO_PRIO(nice
);
4304 on_rq
= p
->se
.on_rq
;
4306 dequeue_task(rq
, p
, 0);
4308 p
->static_prio
= NICE_TO_PRIO(nice
);
4311 p
->prio
= effective_prio(p
);
4312 delta
= p
->prio
- old_prio
;
4315 enqueue_task(rq
, p
, 0, false);
4317 * If the task increased its priority or is running and
4318 * lowered its priority, then reschedule its CPU:
4320 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4321 resched_task(rq
->curr
);
4324 task_rq_unlock(rq
, &flags
);
4326 EXPORT_SYMBOL(set_user_nice
);
4329 * can_nice - check if a task can reduce its nice value
4333 int can_nice(const struct task_struct
*p
, const int nice
)
4335 /* convert nice value [19,-20] to rlimit style value [1,40] */
4336 int nice_rlim
= 20 - nice
;
4338 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4339 capable(CAP_SYS_NICE
));
4342 #ifdef __ARCH_WANT_SYS_NICE
4345 * sys_nice - change the priority of the current process.
4346 * @increment: priority increment
4348 * sys_setpriority is a more generic, but much slower function that
4349 * does similar things.
4351 SYSCALL_DEFINE1(nice
, int, increment
)
4356 * Setpriority might change our priority at the same moment.
4357 * We don't have to worry. Conceptually one call occurs first
4358 * and we have a single winner.
4360 if (increment
< -40)
4365 nice
= TASK_NICE(current
) + increment
;
4371 if (increment
< 0 && !can_nice(current
, nice
))
4374 retval
= security_task_setnice(current
, nice
);
4378 set_user_nice(current
, nice
);
4385 * task_prio - return the priority value of a given task.
4386 * @p: the task in question.
4388 * This is the priority value as seen by users in /proc.
4389 * RT tasks are offset by -200. Normal tasks are centered
4390 * around 0, value goes from -16 to +15.
4392 int task_prio(const struct task_struct
*p
)
4394 return p
->prio
- MAX_RT_PRIO
;
4398 * task_nice - return the nice value of a given task.
4399 * @p: the task in question.
4401 int task_nice(const struct task_struct
*p
)
4403 return TASK_NICE(p
);
4405 EXPORT_SYMBOL(task_nice
);
4408 * idle_cpu - is a given cpu idle currently?
4409 * @cpu: the processor in question.
4411 int idle_cpu(int cpu
)
4413 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4417 * idle_task - return the idle task for a given cpu.
4418 * @cpu: the processor in question.
4420 struct task_struct
*idle_task(int cpu
)
4422 return cpu_rq(cpu
)->idle
;
4426 * find_process_by_pid - find a process with a matching PID value.
4427 * @pid: the pid in question.
4429 static struct task_struct
*find_process_by_pid(pid_t pid
)
4431 return pid
? find_task_by_vpid(pid
) : current
;
4434 /* Actually do priority change: must hold rq lock. */
4436 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4438 BUG_ON(p
->se
.on_rq
);
4441 p
->rt_priority
= prio
;
4442 p
->normal_prio
= normal_prio(p
);
4443 /* we are holding p->pi_lock already */
4444 p
->prio
= rt_mutex_getprio(p
);
4445 if (rt_prio(p
->prio
))
4446 p
->sched_class
= &rt_sched_class
;
4448 p
->sched_class
= &fair_sched_class
;
4453 * check the target process has a UID that matches the current process's
4455 static bool check_same_owner(struct task_struct
*p
)
4457 const struct cred
*cred
= current_cred(), *pcred
;
4461 pcred
= __task_cred(p
);
4462 match
= (cred
->euid
== pcred
->euid
||
4463 cred
->euid
== pcred
->uid
);
4468 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4469 struct sched_param
*param
, bool user
)
4471 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4472 unsigned long flags
;
4473 const struct sched_class
*prev_class
;
4477 /* may grab non-irq protected spin_locks */
4478 BUG_ON(in_interrupt());
4480 /* double check policy once rq lock held */
4482 reset_on_fork
= p
->sched_reset_on_fork
;
4483 policy
= oldpolicy
= p
->policy
;
4485 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4486 policy
&= ~SCHED_RESET_ON_FORK
;
4488 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4489 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4490 policy
!= SCHED_IDLE
)
4495 * Valid priorities for SCHED_FIFO and SCHED_RR are
4496 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4497 * SCHED_BATCH and SCHED_IDLE is 0.
4499 if (param
->sched_priority
< 0 ||
4500 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4501 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4503 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4507 * Allow unprivileged RT tasks to decrease priority:
4509 if (user
&& !capable(CAP_SYS_NICE
)) {
4510 if (rt_policy(policy
)) {
4511 unsigned long rlim_rtprio
;
4513 if (!lock_task_sighand(p
, &flags
))
4515 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4516 unlock_task_sighand(p
, &flags
);
4518 /* can't set/change the rt policy */
4519 if (policy
!= p
->policy
&& !rlim_rtprio
)
4522 /* can't increase priority */
4523 if (param
->sched_priority
> p
->rt_priority
&&
4524 param
->sched_priority
> rlim_rtprio
)
4528 * Like positive nice levels, dont allow tasks to
4529 * move out of SCHED_IDLE either:
4531 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4534 /* can't change other user's priorities */
4535 if (!check_same_owner(p
))
4538 /* Normal users shall not reset the sched_reset_on_fork flag */
4539 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4544 #ifdef CONFIG_RT_GROUP_SCHED
4546 * Do not allow realtime tasks into groups that have no runtime
4549 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4550 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4554 retval
= security_task_setscheduler(p
, policy
, param
);
4560 * make sure no PI-waiters arrive (or leave) while we are
4561 * changing the priority of the task:
4563 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4565 * To be able to change p->policy safely, the apropriate
4566 * runqueue lock must be held.
4568 rq
= __task_rq_lock(p
);
4569 /* recheck policy now with rq lock held */
4570 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4571 policy
= oldpolicy
= -1;
4572 __task_rq_unlock(rq
);
4573 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4576 update_rq_clock(rq
);
4577 on_rq
= p
->se
.on_rq
;
4578 running
= task_current(rq
, p
);
4580 deactivate_task(rq
, p
, 0);
4582 p
->sched_class
->put_prev_task(rq
, p
);
4584 p
->sched_reset_on_fork
= reset_on_fork
;
4587 prev_class
= p
->sched_class
;
4588 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4591 p
->sched_class
->set_curr_task(rq
);
4593 activate_task(rq
, p
, 0);
4595 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4597 __task_rq_unlock(rq
);
4598 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4600 rt_mutex_adjust_pi(p
);
4606 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4607 * @p: the task in question.
4608 * @policy: new policy.
4609 * @param: structure containing the new RT priority.
4611 * NOTE that the task may be already dead.
4613 int sched_setscheduler(struct task_struct
*p
, int policy
,
4614 struct sched_param
*param
)
4616 return __sched_setscheduler(p
, policy
, param
, true);
4618 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4621 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4622 * @p: the task in question.
4623 * @policy: new policy.
4624 * @param: structure containing the new RT priority.
4626 * Just like sched_setscheduler, only don't bother checking if the
4627 * current context has permission. For example, this is needed in
4628 * stop_machine(): we create temporary high priority worker threads,
4629 * but our caller might not have that capability.
4631 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4632 struct sched_param
*param
)
4634 return __sched_setscheduler(p
, policy
, param
, false);
4638 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4640 struct sched_param lparam
;
4641 struct task_struct
*p
;
4644 if (!param
|| pid
< 0)
4646 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4651 p
= find_process_by_pid(pid
);
4653 retval
= sched_setscheduler(p
, policy
, &lparam
);
4660 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4661 * @pid: the pid in question.
4662 * @policy: new policy.
4663 * @param: structure containing the new RT priority.
4665 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4666 struct sched_param __user
*, param
)
4668 /* negative values for policy are not valid */
4672 return do_sched_setscheduler(pid
, policy
, param
);
4676 * sys_sched_setparam - set/change the RT priority of a thread
4677 * @pid: the pid in question.
4678 * @param: structure containing the new RT priority.
4680 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4682 return do_sched_setscheduler(pid
, -1, param
);
4686 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4687 * @pid: the pid in question.
4689 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4691 struct task_struct
*p
;
4699 p
= find_process_by_pid(pid
);
4701 retval
= security_task_getscheduler(p
);
4704 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4711 * sys_sched_getparam - get the RT priority of a thread
4712 * @pid: the pid in question.
4713 * @param: structure containing the RT priority.
4715 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4717 struct sched_param lp
;
4718 struct task_struct
*p
;
4721 if (!param
|| pid
< 0)
4725 p
= find_process_by_pid(pid
);
4730 retval
= security_task_getscheduler(p
);
4734 lp
.sched_priority
= p
->rt_priority
;
4738 * This one might sleep, we cannot do it with a spinlock held ...
4740 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4749 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4751 cpumask_var_t cpus_allowed
, new_mask
;
4752 struct task_struct
*p
;
4758 p
= find_process_by_pid(pid
);
4765 /* Prevent p going away */
4769 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4773 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4775 goto out_free_cpus_allowed
;
4778 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4781 retval
= security_task_setscheduler(p
, 0, NULL
);
4785 cpuset_cpus_allowed(p
, cpus_allowed
);
4786 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4788 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4791 cpuset_cpus_allowed(p
, cpus_allowed
);
4792 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4794 * We must have raced with a concurrent cpuset
4795 * update. Just reset the cpus_allowed to the
4796 * cpuset's cpus_allowed
4798 cpumask_copy(new_mask
, cpus_allowed
);
4803 free_cpumask_var(new_mask
);
4804 out_free_cpus_allowed
:
4805 free_cpumask_var(cpus_allowed
);
4812 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4813 struct cpumask
*new_mask
)
4815 if (len
< cpumask_size())
4816 cpumask_clear(new_mask
);
4817 else if (len
> cpumask_size())
4818 len
= cpumask_size();
4820 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4824 * sys_sched_setaffinity - set the cpu affinity of a process
4825 * @pid: pid of the process
4826 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4827 * @user_mask_ptr: user-space pointer to the new cpu mask
4829 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4830 unsigned long __user
*, user_mask_ptr
)
4832 cpumask_var_t new_mask
;
4835 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4838 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4840 retval
= sched_setaffinity(pid
, new_mask
);
4841 free_cpumask_var(new_mask
);
4845 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4847 struct task_struct
*p
;
4848 unsigned long flags
;
4856 p
= find_process_by_pid(pid
);
4860 retval
= security_task_getscheduler(p
);
4864 rq
= task_rq_lock(p
, &flags
);
4865 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4866 task_rq_unlock(rq
, &flags
);
4876 * sys_sched_getaffinity - get the cpu affinity of a process
4877 * @pid: pid of the process
4878 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4879 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4881 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4882 unsigned long __user
*, user_mask_ptr
)
4887 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4889 if (len
& (sizeof(unsigned long)-1))
4892 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4895 ret
= sched_getaffinity(pid
, mask
);
4897 size_t retlen
= min_t(size_t, len
, cpumask_size());
4899 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4904 free_cpumask_var(mask
);
4910 * sys_sched_yield - yield the current processor to other threads.
4912 * This function yields the current CPU to other tasks. If there are no
4913 * other threads running on this CPU then this function will return.
4915 SYSCALL_DEFINE0(sched_yield
)
4917 struct rq
*rq
= this_rq_lock();
4919 schedstat_inc(rq
, yld_count
);
4920 current
->sched_class
->yield_task(rq
);
4923 * Since we are going to call schedule() anyway, there's
4924 * no need to preempt or enable interrupts:
4926 __release(rq
->lock
);
4927 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4928 do_raw_spin_unlock(&rq
->lock
);
4929 preempt_enable_no_resched();
4936 static inline int should_resched(void)
4938 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4941 static void __cond_resched(void)
4943 add_preempt_count(PREEMPT_ACTIVE
);
4945 sub_preempt_count(PREEMPT_ACTIVE
);
4948 int __sched
_cond_resched(void)
4950 if (should_resched()) {
4956 EXPORT_SYMBOL(_cond_resched
);
4959 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4960 * call schedule, and on return reacquire the lock.
4962 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4963 * operations here to prevent schedule() from being called twice (once via
4964 * spin_unlock(), once by hand).
4966 int __cond_resched_lock(spinlock_t
*lock
)
4968 int resched
= should_resched();
4971 lockdep_assert_held(lock
);
4973 if (spin_needbreak(lock
) || resched
) {
4984 EXPORT_SYMBOL(__cond_resched_lock
);
4986 int __sched
__cond_resched_softirq(void)
4988 BUG_ON(!in_softirq());
4990 if (should_resched()) {
4998 EXPORT_SYMBOL(__cond_resched_softirq
);
5001 * yield - yield the current processor to other threads.
5003 * This is a shortcut for kernel-space yielding - it marks the
5004 * thread runnable and calls sys_sched_yield().
5006 void __sched
yield(void)
5008 set_current_state(TASK_RUNNING
);
5011 EXPORT_SYMBOL(yield
);
5014 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5015 * that process accounting knows that this is a task in IO wait state.
5017 void __sched
io_schedule(void)
5019 struct rq
*rq
= raw_rq();
5021 delayacct_blkio_start();
5022 atomic_inc(&rq
->nr_iowait
);
5023 current
->in_iowait
= 1;
5025 current
->in_iowait
= 0;
5026 atomic_dec(&rq
->nr_iowait
);
5027 delayacct_blkio_end();
5029 EXPORT_SYMBOL(io_schedule
);
5031 long __sched
io_schedule_timeout(long timeout
)
5033 struct rq
*rq
= raw_rq();
5036 delayacct_blkio_start();
5037 atomic_inc(&rq
->nr_iowait
);
5038 current
->in_iowait
= 1;
5039 ret
= schedule_timeout(timeout
);
5040 current
->in_iowait
= 0;
5041 atomic_dec(&rq
->nr_iowait
);
5042 delayacct_blkio_end();
5047 * sys_sched_get_priority_max - return maximum RT priority.
5048 * @policy: scheduling class.
5050 * this syscall returns the maximum rt_priority that can be used
5051 * by a given scheduling class.
5053 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5060 ret
= MAX_USER_RT_PRIO
-1;
5072 * sys_sched_get_priority_min - return minimum RT priority.
5073 * @policy: scheduling class.
5075 * this syscall returns the minimum rt_priority that can be used
5076 * by a given scheduling class.
5078 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5096 * sys_sched_rr_get_interval - return the default timeslice of a process.
5097 * @pid: pid of the process.
5098 * @interval: userspace pointer to the timeslice value.
5100 * this syscall writes the default timeslice value of a given process
5101 * into the user-space timespec buffer. A value of '0' means infinity.
5103 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5104 struct timespec __user
*, interval
)
5106 struct task_struct
*p
;
5107 unsigned int time_slice
;
5108 unsigned long flags
;
5118 p
= find_process_by_pid(pid
);
5122 retval
= security_task_getscheduler(p
);
5126 rq
= task_rq_lock(p
, &flags
);
5127 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5128 task_rq_unlock(rq
, &flags
);
5131 jiffies_to_timespec(time_slice
, &t
);
5132 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5140 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5142 void sched_show_task(struct task_struct
*p
)
5144 unsigned long free
= 0;
5147 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5148 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5149 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5150 #if BITS_PER_LONG == 32
5151 if (state
== TASK_RUNNING
)
5152 printk(KERN_CONT
" running ");
5154 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5156 if (state
== TASK_RUNNING
)
5157 printk(KERN_CONT
" running task ");
5159 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5161 #ifdef CONFIG_DEBUG_STACK_USAGE
5162 free
= stack_not_used(p
);
5164 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5165 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5166 (unsigned long)task_thread_info(p
)->flags
);
5168 show_stack(p
, NULL
);
5171 void show_state_filter(unsigned long state_filter
)
5173 struct task_struct
*g
, *p
;
5175 #if BITS_PER_LONG == 32
5177 " task PC stack pid father\n");
5180 " task PC stack pid father\n");
5182 read_lock(&tasklist_lock
);
5183 do_each_thread(g
, p
) {
5185 * reset the NMI-timeout, listing all files on a slow
5186 * console might take alot of time:
5188 touch_nmi_watchdog();
5189 if (!state_filter
|| (p
->state
& state_filter
))
5191 } while_each_thread(g
, p
);
5193 touch_all_softlockup_watchdogs();
5195 #ifdef CONFIG_SCHED_DEBUG
5196 sysrq_sched_debug_show();
5198 read_unlock(&tasklist_lock
);
5200 * Only show locks if all tasks are dumped:
5203 debug_show_all_locks();
5206 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5208 idle
->sched_class
= &idle_sched_class
;
5212 * init_idle - set up an idle thread for a given CPU
5213 * @idle: task in question
5214 * @cpu: cpu the idle task belongs to
5216 * NOTE: this function does not set the idle thread's NEED_RESCHED
5217 * flag, to make booting more robust.
5219 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5221 struct rq
*rq
= cpu_rq(cpu
);
5222 unsigned long flags
;
5224 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5227 idle
->state
= TASK_RUNNING
;
5228 idle
->se
.exec_start
= sched_clock();
5230 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5232 * We're having a chicken and egg problem, even though we are
5233 * holding rq->lock, the cpu isn't yet set to this cpu so the
5234 * lockdep check in task_group() will fail.
5236 * Similar case to sched_fork(). / Alternatively we could
5237 * use task_rq_lock() here and obtain the other rq->lock.
5242 __set_task_cpu(idle
, cpu
);
5245 rq
->curr
= rq
->idle
= idle
;
5246 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5249 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5251 /* Set the preempt count _outside_ the spinlocks! */
5252 #if defined(CONFIG_PREEMPT)
5253 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5255 task_thread_info(idle
)->preempt_count
= 0;
5258 * The idle tasks have their own, simple scheduling class:
5260 idle
->sched_class
= &idle_sched_class
;
5261 ftrace_graph_init_idle_task(idle
, cpu
);
5265 * In a system that switches off the HZ timer nohz_cpu_mask
5266 * indicates which cpus entered this state. This is used
5267 * in the rcu update to wait only for active cpus. For system
5268 * which do not switch off the HZ timer nohz_cpu_mask should
5269 * always be CPU_BITS_NONE.
5271 cpumask_var_t nohz_cpu_mask
;
5274 * Increase the granularity value when there are more CPUs,
5275 * because with more CPUs the 'effective latency' as visible
5276 * to users decreases. But the relationship is not linear,
5277 * so pick a second-best guess by going with the log2 of the
5280 * This idea comes from the SD scheduler of Con Kolivas:
5282 static int get_update_sysctl_factor(void)
5284 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5285 unsigned int factor
;
5287 switch (sysctl_sched_tunable_scaling
) {
5288 case SCHED_TUNABLESCALING_NONE
:
5291 case SCHED_TUNABLESCALING_LINEAR
:
5294 case SCHED_TUNABLESCALING_LOG
:
5296 factor
= 1 + ilog2(cpus
);
5303 static void update_sysctl(void)
5305 unsigned int factor
= get_update_sysctl_factor();
5307 #define SET_SYSCTL(name) \
5308 (sysctl_##name = (factor) * normalized_sysctl_##name)
5309 SET_SYSCTL(sched_min_granularity
);
5310 SET_SYSCTL(sched_latency
);
5311 SET_SYSCTL(sched_wakeup_granularity
);
5312 SET_SYSCTL(sched_shares_ratelimit
);
5316 static inline void sched_init_granularity(void)
5323 * This is how migration works:
5325 * 1) we queue a struct migration_req structure in the source CPU's
5326 * runqueue and wake up that CPU's migration thread.
5327 * 2) we down() the locked semaphore => thread blocks.
5328 * 3) migration thread wakes up (implicitly it forces the migrated
5329 * thread off the CPU)
5330 * 4) it gets the migration request and checks whether the migrated
5331 * task is still in the wrong runqueue.
5332 * 5) if it's in the wrong runqueue then the migration thread removes
5333 * it and puts it into the right queue.
5334 * 6) migration thread up()s the semaphore.
5335 * 7) we wake up and the migration is done.
5339 * Change a given task's CPU affinity. Migrate the thread to a
5340 * proper CPU and schedule it away if the CPU it's executing on
5341 * is removed from the allowed bitmask.
5343 * NOTE: the caller must have a valid reference to the task, the
5344 * task must not exit() & deallocate itself prematurely. The
5345 * call is not atomic; no spinlocks may be held.
5347 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5349 struct migration_req req
;
5350 unsigned long flags
;
5355 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5356 * drop the rq->lock and still rely on ->cpus_allowed.
5359 while (task_is_waking(p
))
5361 rq
= task_rq_lock(p
, &flags
);
5362 if (task_is_waking(p
)) {
5363 task_rq_unlock(rq
, &flags
);
5367 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5372 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5373 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5378 if (p
->sched_class
->set_cpus_allowed
)
5379 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5381 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5382 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5385 /* Can the task run on the task's current CPU? If so, we're done */
5386 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5389 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5390 /* Need help from migration thread: drop lock and wait. */
5391 struct task_struct
*mt
= rq
->migration_thread
;
5393 get_task_struct(mt
);
5394 task_rq_unlock(rq
, &flags
);
5395 wake_up_process(mt
);
5396 put_task_struct(mt
);
5397 wait_for_completion(&req
.done
);
5398 tlb_migrate_finish(p
->mm
);
5402 task_rq_unlock(rq
, &flags
);
5406 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5409 * Move (not current) task off this cpu, onto dest cpu. We're doing
5410 * this because either it can't run here any more (set_cpus_allowed()
5411 * away from this CPU, or CPU going down), or because we're
5412 * attempting to rebalance this task on exec (sched_exec).
5414 * So we race with normal scheduler movements, but that's OK, as long
5415 * as the task is no longer on this CPU.
5417 * Returns non-zero if task was successfully migrated.
5419 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5421 struct rq
*rq_dest
, *rq_src
;
5424 if (unlikely(!cpu_active(dest_cpu
)))
5427 rq_src
= cpu_rq(src_cpu
);
5428 rq_dest
= cpu_rq(dest_cpu
);
5430 double_rq_lock(rq_src
, rq_dest
);
5431 /* Already moved. */
5432 if (task_cpu(p
) != src_cpu
)
5434 /* Affinity changed (again). */
5435 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5439 * If we're not on a rq, the next wake-up will ensure we're
5443 deactivate_task(rq_src
, p
, 0);
5444 set_task_cpu(p
, dest_cpu
);
5445 activate_task(rq_dest
, p
, 0);
5446 check_preempt_curr(rq_dest
, p
, 0);
5451 double_rq_unlock(rq_src
, rq_dest
);
5455 #define RCU_MIGRATION_IDLE 0
5456 #define RCU_MIGRATION_NEED_QS 1
5457 #define RCU_MIGRATION_GOT_QS 2
5458 #define RCU_MIGRATION_MUST_SYNC 3
5461 * migration_thread - this is a highprio system thread that performs
5462 * thread migration by bumping thread off CPU then 'pushing' onto
5465 static int migration_thread(void *data
)
5468 int cpu
= (long)data
;
5472 BUG_ON(rq
->migration_thread
!= current
);
5474 set_current_state(TASK_INTERRUPTIBLE
);
5475 while (!kthread_should_stop()) {
5476 struct migration_req
*req
;
5477 struct list_head
*head
;
5479 raw_spin_lock_irq(&rq
->lock
);
5481 if (cpu_is_offline(cpu
)) {
5482 raw_spin_unlock_irq(&rq
->lock
);
5486 if (rq
->active_balance
) {
5487 active_load_balance(rq
, cpu
);
5488 rq
->active_balance
= 0;
5491 head
= &rq
->migration_queue
;
5493 if (list_empty(head
)) {
5494 raw_spin_unlock_irq(&rq
->lock
);
5496 set_current_state(TASK_INTERRUPTIBLE
);
5499 req
= list_entry(head
->next
, struct migration_req
, list
);
5500 list_del_init(head
->next
);
5502 if (req
->task
!= NULL
) {
5503 raw_spin_unlock(&rq
->lock
);
5504 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5505 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5506 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5507 raw_spin_unlock(&rq
->lock
);
5509 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5510 raw_spin_unlock(&rq
->lock
);
5511 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5515 complete(&req
->done
);
5517 __set_current_state(TASK_RUNNING
);
5522 #ifdef CONFIG_HOTPLUG_CPU
5524 * Figure out where task on dead CPU should go, use force if necessary.
5526 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5528 struct rq
*rq
= cpu_rq(dead_cpu
);
5529 int needs_cpu
, uninitialized_var(dest_cpu
);
5530 unsigned long flags
;
5532 local_irq_save(flags
);
5534 raw_spin_lock(&rq
->lock
);
5535 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5537 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5538 raw_spin_unlock(&rq
->lock
);
5540 * It can only fail if we race with set_cpus_allowed(),
5541 * in the racer should migrate the task anyway.
5544 __migrate_task(p
, dead_cpu
, dest_cpu
);
5545 local_irq_restore(flags
);
5549 * While a dead CPU has no uninterruptible tasks queued at this point,
5550 * it might still have a nonzero ->nr_uninterruptible counter, because
5551 * for performance reasons the counter is not stricly tracking tasks to
5552 * their home CPUs. So we just add the counter to another CPU's counter,
5553 * to keep the global sum constant after CPU-down:
5555 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5557 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5558 unsigned long flags
;
5560 local_irq_save(flags
);
5561 double_rq_lock(rq_src
, rq_dest
);
5562 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5563 rq_src
->nr_uninterruptible
= 0;
5564 double_rq_unlock(rq_src
, rq_dest
);
5565 local_irq_restore(flags
);
5568 /* Run through task list and migrate tasks from the dead cpu. */
5569 static void migrate_live_tasks(int src_cpu
)
5571 struct task_struct
*p
, *t
;
5573 read_lock(&tasklist_lock
);
5575 do_each_thread(t
, p
) {
5579 if (task_cpu(p
) == src_cpu
)
5580 move_task_off_dead_cpu(src_cpu
, p
);
5581 } while_each_thread(t
, p
);
5583 read_unlock(&tasklist_lock
);
5587 * Schedules idle task to be the next runnable task on current CPU.
5588 * It does so by boosting its priority to highest possible.
5589 * Used by CPU offline code.
5591 void sched_idle_next(void)
5593 int this_cpu
= smp_processor_id();
5594 struct rq
*rq
= cpu_rq(this_cpu
);
5595 struct task_struct
*p
= rq
->idle
;
5596 unsigned long flags
;
5598 /* cpu has to be offline */
5599 BUG_ON(cpu_online(this_cpu
));
5602 * Strictly not necessary since rest of the CPUs are stopped by now
5603 * and interrupts disabled on the current cpu.
5605 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5607 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5609 update_rq_clock(rq
);
5610 activate_task(rq
, p
, 0);
5612 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5616 * Ensures that the idle task is using init_mm right before its cpu goes
5619 void idle_task_exit(void)
5621 struct mm_struct
*mm
= current
->active_mm
;
5623 BUG_ON(cpu_online(smp_processor_id()));
5626 switch_mm(mm
, &init_mm
, current
);
5630 /* called under rq->lock with disabled interrupts */
5631 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5633 struct rq
*rq
= cpu_rq(dead_cpu
);
5635 /* Must be exiting, otherwise would be on tasklist. */
5636 BUG_ON(!p
->exit_state
);
5638 /* Cannot have done final schedule yet: would have vanished. */
5639 BUG_ON(p
->state
== TASK_DEAD
);
5644 * Drop lock around migration; if someone else moves it,
5645 * that's OK. No task can be added to this CPU, so iteration is
5648 raw_spin_unlock_irq(&rq
->lock
);
5649 move_task_off_dead_cpu(dead_cpu
, p
);
5650 raw_spin_lock_irq(&rq
->lock
);
5655 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5656 static void migrate_dead_tasks(unsigned int dead_cpu
)
5658 struct rq
*rq
= cpu_rq(dead_cpu
);
5659 struct task_struct
*next
;
5662 if (!rq
->nr_running
)
5664 update_rq_clock(rq
);
5665 next
= pick_next_task(rq
);
5668 next
->sched_class
->put_prev_task(rq
, next
);
5669 migrate_dead(dead_cpu
, next
);
5675 * remove the tasks which were accounted by rq from calc_load_tasks.
5677 static void calc_global_load_remove(struct rq
*rq
)
5679 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5680 rq
->calc_load_active
= 0;
5682 #endif /* CONFIG_HOTPLUG_CPU */
5684 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5686 static struct ctl_table sd_ctl_dir
[] = {
5688 .procname
= "sched_domain",
5694 static struct ctl_table sd_ctl_root
[] = {
5696 .procname
= "kernel",
5698 .child
= sd_ctl_dir
,
5703 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5705 struct ctl_table
*entry
=
5706 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5711 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5713 struct ctl_table
*entry
;
5716 * In the intermediate directories, both the child directory and
5717 * procname are dynamically allocated and could fail but the mode
5718 * will always be set. In the lowest directory the names are
5719 * static strings and all have proc handlers.
5721 for (entry
= *tablep
; entry
->mode
; entry
++) {
5723 sd_free_ctl_entry(&entry
->child
);
5724 if (entry
->proc_handler
== NULL
)
5725 kfree(entry
->procname
);
5733 set_table_entry(struct ctl_table
*entry
,
5734 const char *procname
, void *data
, int maxlen
,
5735 mode_t mode
, proc_handler
*proc_handler
)
5737 entry
->procname
= procname
;
5739 entry
->maxlen
= maxlen
;
5741 entry
->proc_handler
= proc_handler
;
5744 static struct ctl_table
*
5745 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5747 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5752 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5753 sizeof(long), 0644, proc_doulongvec_minmax
);
5754 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5755 sizeof(long), 0644, proc_doulongvec_minmax
);
5756 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5757 sizeof(int), 0644, proc_dointvec_minmax
);
5758 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5759 sizeof(int), 0644, proc_dointvec_minmax
);
5760 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5761 sizeof(int), 0644, proc_dointvec_minmax
);
5762 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5763 sizeof(int), 0644, proc_dointvec_minmax
);
5764 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5765 sizeof(int), 0644, proc_dointvec_minmax
);
5766 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5767 sizeof(int), 0644, proc_dointvec_minmax
);
5768 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5769 sizeof(int), 0644, proc_dointvec_minmax
);
5770 set_table_entry(&table
[9], "cache_nice_tries",
5771 &sd
->cache_nice_tries
,
5772 sizeof(int), 0644, proc_dointvec_minmax
);
5773 set_table_entry(&table
[10], "flags", &sd
->flags
,
5774 sizeof(int), 0644, proc_dointvec_minmax
);
5775 set_table_entry(&table
[11], "name", sd
->name
,
5776 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5777 /* &table[12] is terminator */
5782 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5784 struct ctl_table
*entry
, *table
;
5785 struct sched_domain
*sd
;
5786 int domain_num
= 0, i
;
5789 for_each_domain(cpu
, sd
)
5791 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5796 for_each_domain(cpu
, sd
) {
5797 snprintf(buf
, 32, "domain%d", i
);
5798 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5800 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5807 static struct ctl_table_header
*sd_sysctl_header
;
5808 static void register_sched_domain_sysctl(void)
5810 int i
, cpu_num
= num_possible_cpus();
5811 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5814 WARN_ON(sd_ctl_dir
[0].child
);
5815 sd_ctl_dir
[0].child
= entry
;
5820 for_each_possible_cpu(i
) {
5821 snprintf(buf
, 32, "cpu%d", i
);
5822 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5824 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5828 WARN_ON(sd_sysctl_header
);
5829 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5832 /* may be called multiple times per register */
5833 static void unregister_sched_domain_sysctl(void)
5835 if (sd_sysctl_header
)
5836 unregister_sysctl_table(sd_sysctl_header
);
5837 sd_sysctl_header
= NULL
;
5838 if (sd_ctl_dir
[0].child
)
5839 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5842 static void register_sched_domain_sysctl(void)
5845 static void unregister_sched_domain_sysctl(void)
5850 static void set_rq_online(struct rq
*rq
)
5853 const struct sched_class
*class;
5855 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5858 for_each_class(class) {
5859 if (class->rq_online
)
5860 class->rq_online(rq
);
5865 static void set_rq_offline(struct rq
*rq
)
5868 const struct sched_class
*class;
5870 for_each_class(class) {
5871 if (class->rq_offline
)
5872 class->rq_offline(rq
);
5875 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5881 * migration_call - callback that gets triggered when a CPU is added.
5882 * Here we can start up the necessary migration thread for the new CPU.
5884 static int __cpuinit
5885 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5887 struct task_struct
*p
;
5888 int cpu
= (long)hcpu
;
5889 unsigned long flags
;
5894 case CPU_UP_PREPARE
:
5895 case CPU_UP_PREPARE_FROZEN
:
5896 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5899 kthread_bind(p
, cpu
);
5900 /* Must be high prio: stop_machine expects to yield to it. */
5901 rq
= task_rq_lock(p
, &flags
);
5902 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5903 task_rq_unlock(rq
, &flags
);
5905 cpu_rq(cpu
)->migration_thread
= p
;
5906 rq
->calc_load_update
= calc_load_update
;
5910 case CPU_ONLINE_FROZEN
:
5911 /* Strictly unnecessary, as first user will wake it. */
5912 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5914 /* Update our root-domain */
5916 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5918 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5922 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5925 #ifdef CONFIG_HOTPLUG_CPU
5926 case CPU_UP_CANCELED
:
5927 case CPU_UP_CANCELED_FROZEN
:
5928 if (!cpu_rq(cpu
)->migration_thread
)
5930 /* Unbind it from offline cpu so it can run. Fall thru. */
5931 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5932 cpumask_any(cpu_online_mask
));
5933 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5934 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5935 cpu_rq(cpu
)->migration_thread
= NULL
;
5939 case CPU_DEAD_FROZEN
:
5940 migrate_live_tasks(cpu
);
5942 kthread_stop(rq
->migration_thread
);
5943 put_task_struct(rq
->migration_thread
);
5944 rq
->migration_thread
= NULL
;
5945 /* Idle task back to normal (off runqueue, low prio) */
5946 raw_spin_lock_irq(&rq
->lock
);
5947 update_rq_clock(rq
);
5948 deactivate_task(rq
, rq
->idle
, 0);
5949 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5950 rq
->idle
->sched_class
= &idle_sched_class
;
5951 migrate_dead_tasks(cpu
);
5952 raw_spin_unlock_irq(&rq
->lock
);
5953 migrate_nr_uninterruptible(rq
);
5954 BUG_ON(rq
->nr_running
!= 0);
5955 calc_global_load_remove(rq
);
5957 * No need to migrate the tasks: it was best-effort if
5958 * they didn't take sched_hotcpu_mutex. Just wake up
5961 raw_spin_lock_irq(&rq
->lock
);
5962 while (!list_empty(&rq
->migration_queue
)) {
5963 struct migration_req
*req
;
5965 req
= list_entry(rq
->migration_queue
.next
,
5966 struct migration_req
, list
);
5967 list_del_init(&req
->list
);
5968 raw_spin_unlock_irq(&rq
->lock
);
5969 complete(&req
->done
);
5970 raw_spin_lock_irq(&rq
->lock
);
5972 raw_spin_unlock_irq(&rq
->lock
);
5976 case CPU_DYING_FROZEN
:
5977 /* Update our root-domain */
5979 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5981 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5984 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5992 * Register at high priority so that task migration (migrate_all_tasks)
5993 * happens before everything else. This has to be lower priority than
5994 * the notifier in the perf_event subsystem, though.
5996 static struct notifier_block __cpuinitdata migration_notifier
= {
5997 .notifier_call
= migration_call
,
6001 static int __init
migration_init(void)
6003 void *cpu
= (void *)(long)smp_processor_id();
6006 /* Start one for the boot CPU: */
6007 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6008 BUG_ON(err
== NOTIFY_BAD
);
6009 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6010 register_cpu_notifier(&migration_notifier
);
6014 early_initcall(migration_init
);
6019 #ifdef CONFIG_SCHED_DEBUG
6021 static __read_mostly
int sched_domain_debug_enabled
;
6023 static int __init
sched_domain_debug_setup(char *str
)
6025 sched_domain_debug_enabled
= 1;
6029 early_param("sched_debug", sched_domain_debug_setup
);
6031 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6032 struct cpumask
*groupmask
)
6034 struct sched_group
*group
= sd
->groups
;
6037 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6038 cpumask_clear(groupmask
);
6040 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6042 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6043 printk("does not load-balance\n");
6045 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6050 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6052 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6053 printk(KERN_ERR
"ERROR: domain->span does not contain "
6056 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6057 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6061 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6065 printk(KERN_ERR
"ERROR: group is NULL\n");
6069 if (!group
->cpu_power
) {
6070 printk(KERN_CONT
"\n");
6071 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6076 if (!cpumask_weight(sched_group_cpus(group
))) {
6077 printk(KERN_CONT
"\n");
6078 printk(KERN_ERR
"ERROR: empty group\n");
6082 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6083 printk(KERN_CONT
"\n");
6084 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6088 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6090 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6092 printk(KERN_CONT
" %s", str
);
6093 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6094 printk(KERN_CONT
" (cpu_power = %d)",
6098 group
= group
->next
;
6099 } while (group
!= sd
->groups
);
6100 printk(KERN_CONT
"\n");
6102 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6103 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6106 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6107 printk(KERN_ERR
"ERROR: parent span is not a superset "
6108 "of domain->span\n");
6112 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6114 cpumask_var_t groupmask
;
6117 if (!sched_domain_debug_enabled
)
6121 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6125 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6127 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6128 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6133 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6140 free_cpumask_var(groupmask
);
6142 #else /* !CONFIG_SCHED_DEBUG */
6143 # define sched_domain_debug(sd, cpu) do { } while (0)
6144 #endif /* CONFIG_SCHED_DEBUG */
6146 static int sd_degenerate(struct sched_domain
*sd
)
6148 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6151 /* Following flags need at least 2 groups */
6152 if (sd
->flags
& (SD_LOAD_BALANCE
|
6153 SD_BALANCE_NEWIDLE
|
6157 SD_SHARE_PKG_RESOURCES
)) {
6158 if (sd
->groups
!= sd
->groups
->next
)
6162 /* Following flags don't use groups */
6163 if (sd
->flags
& (SD_WAKE_AFFINE
))
6170 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6172 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6174 if (sd_degenerate(parent
))
6177 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6180 /* Flags needing groups don't count if only 1 group in parent */
6181 if (parent
->groups
== parent
->groups
->next
) {
6182 pflags
&= ~(SD_LOAD_BALANCE
|
6183 SD_BALANCE_NEWIDLE
|
6187 SD_SHARE_PKG_RESOURCES
);
6188 if (nr_node_ids
== 1)
6189 pflags
&= ~SD_SERIALIZE
;
6191 if (~cflags
& pflags
)
6197 static void free_rootdomain(struct root_domain
*rd
)
6199 synchronize_sched();
6201 cpupri_cleanup(&rd
->cpupri
);
6203 free_cpumask_var(rd
->rto_mask
);
6204 free_cpumask_var(rd
->online
);
6205 free_cpumask_var(rd
->span
);
6209 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6211 struct root_domain
*old_rd
= NULL
;
6212 unsigned long flags
;
6214 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6219 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6222 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6225 * If we dont want to free the old_rt yet then
6226 * set old_rd to NULL to skip the freeing later
6229 if (!atomic_dec_and_test(&old_rd
->refcount
))
6233 atomic_inc(&rd
->refcount
);
6236 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6237 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6240 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6243 free_rootdomain(old_rd
);
6246 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6248 gfp_t gfp
= GFP_KERNEL
;
6250 memset(rd
, 0, sizeof(*rd
));
6255 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6257 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6259 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6262 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6267 free_cpumask_var(rd
->rto_mask
);
6269 free_cpumask_var(rd
->online
);
6271 free_cpumask_var(rd
->span
);
6276 static void init_defrootdomain(void)
6278 init_rootdomain(&def_root_domain
, true);
6280 atomic_set(&def_root_domain
.refcount
, 1);
6283 static struct root_domain
*alloc_rootdomain(void)
6285 struct root_domain
*rd
;
6287 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6291 if (init_rootdomain(rd
, false) != 0) {
6300 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6301 * hold the hotplug lock.
6304 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6306 struct rq
*rq
= cpu_rq(cpu
);
6307 struct sched_domain
*tmp
;
6309 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6310 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6312 /* Remove the sched domains which do not contribute to scheduling. */
6313 for (tmp
= sd
; tmp
; ) {
6314 struct sched_domain
*parent
= tmp
->parent
;
6318 if (sd_parent_degenerate(tmp
, parent
)) {
6319 tmp
->parent
= parent
->parent
;
6321 parent
->parent
->child
= tmp
;
6326 if (sd
&& sd_degenerate(sd
)) {
6332 sched_domain_debug(sd
, cpu
);
6334 rq_attach_root(rq
, rd
);
6335 rcu_assign_pointer(rq
->sd
, sd
);
6338 /* cpus with isolated domains */
6339 static cpumask_var_t cpu_isolated_map
;
6341 /* Setup the mask of cpus configured for isolated domains */
6342 static int __init
isolated_cpu_setup(char *str
)
6344 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6345 cpulist_parse(str
, cpu_isolated_map
);
6349 __setup("isolcpus=", isolated_cpu_setup
);
6352 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6353 * to a function which identifies what group(along with sched group) a CPU
6354 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6355 * (due to the fact that we keep track of groups covered with a struct cpumask).
6357 * init_sched_build_groups will build a circular linked list of the groups
6358 * covered by the given span, and will set each group's ->cpumask correctly,
6359 * and ->cpu_power to 0.
6362 init_sched_build_groups(const struct cpumask
*span
,
6363 const struct cpumask
*cpu_map
,
6364 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6365 struct sched_group
**sg
,
6366 struct cpumask
*tmpmask
),
6367 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6369 struct sched_group
*first
= NULL
, *last
= NULL
;
6372 cpumask_clear(covered
);
6374 for_each_cpu(i
, span
) {
6375 struct sched_group
*sg
;
6376 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6379 if (cpumask_test_cpu(i
, covered
))
6382 cpumask_clear(sched_group_cpus(sg
));
6385 for_each_cpu(j
, span
) {
6386 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6389 cpumask_set_cpu(j
, covered
);
6390 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6401 #define SD_NODES_PER_DOMAIN 16
6406 * find_next_best_node - find the next node to include in a sched_domain
6407 * @node: node whose sched_domain we're building
6408 * @used_nodes: nodes already in the sched_domain
6410 * Find the next node to include in a given scheduling domain. Simply
6411 * finds the closest node not already in the @used_nodes map.
6413 * Should use nodemask_t.
6415 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6417 int i
, n
, val
, min_val
, best_node
= 0;
6421 for (i
= 0; i
< nr_node_ids
; i
++) {
6422 /* Start at @node */
6423 n
= (node
+ i
) % nr_node_ids
;
6425 if (!nr_cpus_node(n
))
6428 /* Skip already used nodes */
6429 if (node_isset(n
, *used_nodes
))
6432 /* Simple min distance search */
6433 val
= node_distance(node
, n
);
6435 if (val
< min_val
) {
6441 node_set(best_node
, *used_nodes
);
6446 * sched_domain_node_span - get a cpumask for a node's sched_domain
6447 * @node: node whose cpumask we're constructing
6448 * @span: resulting cpumask
6450 * Given a node, construct a good cpumask for its sched_domain to span. It
6451 * should be one that prevents unnecessary balancing, but also spreads tasks
6454 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6456 nodemask_t used_nodes
;
6459 cpumask_clear(span
);
6460 nodes_clear(used_nodes
);
6462 cpumask_or(span
, span
, cpumask_of_node(node
));
6463 node_set(node
, used_nodes
);
6465 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6466 int next_node
= find_next_best_node(node
, &used_nodes
);
6468 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6471 #endif /* CONFIG_NUMA */
6473 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6476 * The cpus mask in sched_group and sched_domain hangs off the end.
6478 * ( See the the comments in include/linux/sched.h:struct sched_group
6479 * and struct sched_domain. )
6481 struct static_sched_group
{
6482 struct sched_group sg
;
6483 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6486 struct static_sched_domain
{
6487 struct sched_domain sd
;
6488 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6494 cpumask_var_t domainspan
;
6495 cpumask_var_t covered
;
6496 cpumask_var_t notcovered
;
6498 cpumask_var_t nodemask
;
6499 cpumask_var_t this_sibling_map
;
6500 cpumask_var_t this_core_map
;
6501 cpumask_var_t send_covered
;
6502 cpumask_var_t tmpmask
;
6503 struct sched_group
**sched_group_nodes
;
6504 struct root_domain
*rd
;
6508 sa_sched_groups
= 0,
6513 sa_this_sibling_map
,
6515 sa_sched_group_nodes
,
6525 * SMT sched-domains:
6527 #ifdef CONFIG_SCHED_SMT
6528 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6529 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6532 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6533 struct sched_group
**sg
, struct cpumask
*unused
)
6536 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6539 #endif /* CONFIG_SCHED_SMT */
6542 * multi-core sched-domains:
6544 #ifdef CONFIG_SCHED_MC
6545 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6546 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6547 #endif /* CONFIG_SCHED_MC */
6549 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6551 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6552 struct sched_group
**sg
, struct cpumask
*mask
)
6556 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6557 group
= cpumask_first(mask
);
6559 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6562 #elif defined(CONFIG_SCHED_MC)
6564 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6565 struct sched_group
**sg
, struct cpumask
*unused
)
6568 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6573 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6574 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6577 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6578 struct sched_group
**sg
, struct cpumask
*mask
)
6581 #ifdef CONFIG_SCHED_MC
6582 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6583 group
= cpumask_first(mask
);
6584 #elif defined(CONFIG_SCHED_SMT)
6585 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6586 group
= cpumask_first(mask
);
6591 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6597 * The init_sched_build_groups can't handle what we want to do with node
6598 * groups, so roll our own. Now each node has its own list of groups which
6599 * gets dynamically allocated.
6601 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6602 static struct sched_group
***sched_group_nodes_bycpu
;
6604 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6605 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6607 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6608 struct sched_group
**sg
,
6609 struct cpumask
*nodemask
)
6613 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6614 group
= cpumask_first(nodemask
);
6617 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6621 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6623 struct sched_group
*sg
= group_head
;
6629 for_each_cpu(j
, sched_group_cpus(sg
)) {
6630 struct sched_domain
*sd
;
6632 sd
= &per_cpu(phys_domains
, j
).sd
;
6633 if (j
!= group_first_cpu(sd
->groups
)) {
6635 * Only add "power" once for each
6641 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6644 } while (sg
!= group_head
);
6647 static int build_numa_sched_groups(struct s_data
*d
,
6648 const struct cpumask
*cpu_map
, int num
)
6650 struct sched_domain
*sd
;
6651 struct sched_group
*sg
, *prev
;
6654 cpumask_clear(d
->covered
);
6655 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6656 if (cpumask_empty(d
->nodemask
)) {
6657 d
->sched_group_nodes
[num
] = NULL
;
6661 sched_domain_node_span(num
, d
->domainspan
);
6662 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6664 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6667 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6671 d
->sched_group_nodes
[num
] = sg
;
6673 for_each_cpu(j
, d
->nodemask
) {
6674 sd
= &per_cpu(node_domains
, j
).sd
;
6679 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6681 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6684 for (j
= 0; j
< nr_node_ids
; j
++) {
6685 n
= (num
+ j
) % nr_node_ids
;
6686 cpumask_complement(d
->notcovered
, d
->covered
);
6687 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6688 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6689 if (cpumask_empty(d
->tmpmask
))
6691 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6692 if (cpumask_empty(d
->tmpmask
))
6694 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6698 "Can not alloc domain group for node %d\n", j
);
6702 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6703 sg
->next
= prev
->next
;
6704 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6711 #endif /* CONFIG_NUMA */
6714 /* Free memory allocated for various sched_group structures */
6715 static void free_sched_groups(const struct cpumask
*cpu_map
,
6716 struct cpumask
*nodemask
)
6720 for_each_cpu(cpu
, cpu_map
) {
6721 struct sched_group
**sched_group_nodes
6722 = sched_group_nodes_bycpu
[cpu
];
6724 if (!sched_group_nodes
)
6727 for (i
= 0; i
< nr_node_ids
; i
++) {
6728 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6730 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6731 if (cpumask_empty(nodemask
))
6741 if (oldsg
!= sched_group_nodes
[i
])
6744 kfree(sched_group_nodes
);
6745 sched_group_nodes_bycpu
[cpu
] = NULL
;
6748 #else /* !CONFIG_NUMA */
6749 static void free_sched_groups(const struct cpumask
*cpu_map
,
6750 struct cpumask
*nodemask
)
6753 #endif /* CONFIG_NUMA */
6756 * Initialize sched groups cpu_power.
6758 * cpu_power indicates the capacity of sched group, which is used while
6759 * distributing the load between different sched groups in a sched domain.
6760 * Typically cpu_power for all the groups in a sched domain will be same unless
6761 * there are asymmetries in the topology. If there are asymmetries, group
6762 * having more cpu_power will pickup more load compared to the group having
6765 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6767 struct sched_domain
*child
;
6768 struct sched_group
*group
;
6772 WARN_ON(!sd
|| !sd
->groups
);
6774 if (cpu
!= group_first_cpu(sd
->groups
))
6779 sd
->groups
->cpu_power
= 0;
6782 power
= SCHED_LOAD_SCALE
;
6783 weight
= cpumask_weight(sched_domain_span(sd
));
6785 * SMT siblings share the power of a single core.
6786 * Usually multiple threads get a better yield out of
6787 * that one core than a single thread would have,
6788 * reflect that in sd->smt_gain.
6790 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6791 power
*= sd
->smt_gain
;
6793 power
>>= SCHED_LOAD_SHIFT
;
6795 sd
->groups
->cpu_power
+= power
;
6800 * Add cpu_power of each child group to this groups cpu_power.
6802 group
= child
->groups
;
6804 sd
->groups
->cpu_power
+= group
->cpu_power
;
6805 group
= group
->next
;
6806 } while (group
!= child
->groups
);
6810 * Initializers for schedule domains
6811 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6814 #ifdef CONFIG_SCHED_DEBUG
6815 # define SD_INIT_NAME(sd, type) sd->name = #type
6817 # define SD_INIT_NAME(sd, type) do { } while (0)
6820 #define SD_INIT(sd, type) sd_init_##type(sd)
6822 #define SD_INIT_FUNC(type) \
6823 static noinline void sd_init_##type(struct sched_domain *sd) \
6825 memset(sd, 0, sizeof(*sd)); \
6826 *sd = SD_##type##_INIT; \
6827 sd->level = SD_LV_##type; \
6828 SD_INIT_NAME(sd, type); \
6833 SD_INIT_FUNC(ALLNODES
)
6836 #ifdef CONFIG_SCHED_SMT
6837 SD_INIT_FUNC(SIBLING
)
6839 #ifdef CONFIG_SCHED_MC
6843 static int default_relax_domain_level
= -1;
6845 static int __init
setup_relax_domain_level(char *str
)
6849 val
= simple_strtoul(str
, NULL
, 0);
6850 if (val
< SD_LV_MAX
)
6851 default_relax_domain_level
= val
;
6855 __setup("relax_domain_level=", setup_relax_domain_level
);
6857 static void set_domain_attribute(struct sched_domain
*sd
,
6858 struct sched_domain_attr
*attr
)
6862 if (!attr
|| attr
->relax_domain_level
< 0) {
6863 if (default_relax_domain_level
< 0)
6866 request
= default_relax_domain_level
;
6868 request
= attr
->relax_domain_level
;
6869 if (request
< sd
->level
) {
6870 /* turn off idle balance on this domain */
6871 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6873 /* turn on idle balance on this domain */
6874 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6878 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6879 const struct cpumask
*cpu_map
)
6882 case sa_sched_groups
:
6883 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6884 d
->sched_group_nodes
= NULL
;
6886 free_rootdomain(d
->rd
); /* fall through */
6888 free_cpumask_var(d
->tmpmask
); /* fall through */
6889 case sa_send_covered
:
6890 free_cpumask_var(d
->send_covered
); /* fall through */
6891 case sa_this_core_map
:
6892 free_cpumask_var(d
->this_core_map
); /* fall through */
6893 case sa_this_sibling_map
:
6894 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6896 free_cpumask_var(d
->nodemask
); /* fall through */
6897 case sa_sched_group_nodes
:
6899 kfree(d
->sched_group_nodes
); /* fall through */
6901 free_cpumask_var(d
->notcovered
); /* fall through */
6903 free_cpumask_var(d
->covered
); /* fall through */
6905 free_cpumask_var(d
->domainspan
); /* fall through */
6912 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6913 const struct cpumask
*cpu_map
)
6916 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6918 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6919 return sa_domainspan
;
6920 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6922 /* Allocate the per-node list of sched groups */
6923 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6924 sizeof(struct sched_group
*), GFP_KERNEL
);
6925 if (!d
->sched_group_nodes
) {
6926 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6927 return sa_notcovered
;
6929 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6931 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6932 return sa_sched_group_nodes
;
6933 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6935 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6936 return sa_this_sibling_map
;
6937 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6938 return sa_this_core_map
;
6939 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6940 return sa_send_covered
;
6941 d
->rd
= alloc_rootdomain();
6943 printk(KERN_WARNING
"Cannot alloc root domain\n");
6946 return sa_rootdomain
;
6949 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6950 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6952 struct sched_domain
*sd
= NULL
;
6954 struct sched_domain
*parent
;
6957 if (cpumask_weight(cpu_map
) >
6958 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6959 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6960 SD_INIT(sd
, ALLNODES
);
6961 set_domain_attribute(sd
, attr
);
6962 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6963 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6968 sd
= &per_cpu(node_domains
, i
).sd
;
6970 set_domain_attribute(sd
, attr
);
6971 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6972 sd
->parent
= parent
;
6975 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6980 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6981 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6982 struct sched_domain
*parent
, int i
)
6984 struct sched_domain
*sd
;
6985 sd
= &per_cpu(phys_domains
, i
).sd
;
6987 set_domain_attribute(sd
, attr
);
6988 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6989 sd
->parent
= parent
;
6992 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6996 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6997 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6998 struct sched_domain
*parent
, int i
)
7000 struct sched_domain
*sd
= parent
;
7001 #ifdef CONFIG_SCHED_MC
7002 sd
= &per_cpu(core_domains
, i
).sd
;
7004 set_domain_attribute(sd
, attr
);
7005 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7006 sd
->parent
= parent
;
7008 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7013 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7014 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7015 struct sched_domain
*parent
, int i
)
7017 struct sched_domain
*sd
= parent
;
7018 #ifdef CONFIG_SCHED_SMT
7019 sd
= &per_cpu(cpu_domains
, i
).sd
;
7020 SD_INIT(sd
, SIBLING
);
7021 set_domain_attribute(sd
, attr
);
7022 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7023 sd
->parent
= parent
;
7025 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7030 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7031 const struct cpumask
*cpu_map
, int cpu
)
7034 #ifdef CONFIG_SCHED_SMT
7035 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7036 cpumask_and(d
->this_sibling_map
, cpu_map
,
7037 topology_thread_cpumask(cpu
));
7038 if (cpu
== cpumask_first(d
->this_sibling_map
))
7039 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7041 d
->send_covered
, d
->tmpmask
);
7044 #ifdef CONFIG_SCHED_MC
7045 case SD_LV_MC
: /* set up multi-core groups */
7046 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7047 if (cpu
== cpumask_first(d
->this_core_map
))
7048 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7050 d
->send_covered
, d
->tmpmask
);
7053 case SD_LV_CPU
: /* set up physical groups */
7054 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7055 if (!cpumask_empty(d
->nodemask
))
7056 init_sched_build_groups(d
->nodemask
, cpu_map
,
7058 d
->send_covered
, d
->tmpmask
);
7061 case SD_LV_ALLNODES
:
7062 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7063 d
->send_covered
, d
->tmpmask
);
7072 * Build sched domains for a given set of cpus and attach the sched domains
7073 * to the individual cpus
7075 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7076 struct sched_domain_attr
*attr
)
7078 enum s_alloc alloc_state
= sa_none
;
7080 struct sched_domain
*sd
;
7086 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7087 if (alloc_state
!= sa_rootdomain
)
7089 alloc_state
= sa_sched_groups
;
7092 * Set up domains for cpus specified by the cpu_map.
7094 for_each_cpu(i
, cpu_map
) {
7095 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7098 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7099 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7100 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7101 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7104 for_each_cpu(i
, cpu_map
) {
7105 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7106 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7109 /* Set up physical groups */
7110 for (i
= 0; i
< nr_node_ids
; i
++)
7111 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7114 /* Set up node groups */
7116 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7118 for (i
= 0; i
< nr_node_ids
; i
++)
7119 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7123 /* Calculate CPU power for physical packages and nodes */
7124 #ifdef CONFIG_SCHED_SMT
7125 for_each_cpu(i
, cpu_map
) {
7126 sd
= &per_cpu(cpu_domains
, i
).sd
;
7127 init_sched_groups_power(i
, sd
);
7130 #ifdef CONFIG_SCHED_MC
7131 for_each_cpu(i
, cpu_map
) {
7132 sd
= &per_cpu(core_domains
, i
).sd
;
7133 init_sched_groups_power(i
, sd
);
7137 for_each_cpu(i
, cpu_map
) {
7138 sd
= &per_cpu(phys_domains
, i
).sd
;
7139 init_sched_groups_power(i
, sd
);
7143 for (i
= 0; i
< nr_node_ids
; i
++)
7144 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7146 if (d
.sd_allnodes
) {
7147 struct sched_group
*sg
;
7149 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7151 init_numa_sched_groups_power(sg
);
7155 /* Attach the domains */
7156 for_each_cpu(i
, cpu_map
) {
7157 #ifdef CONFIG_SCHED_SMT
7158 sd
= &per_cpu(cpu_domains
, i
).sd
;
7159 #elif defined(CONFIG_SCHED_MC)
7160 sd
= &per_cpu(core_domains
, i
).sd
;
7162 sd
= &per_cpu(phys_domains
, i
).sd
;
7164 cpu_attach_domain(sd
, d
.rd
, i
);
7167 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7168 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7172 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7176 static int build_sched_domains(const struct cpumask
*cpu_map
)
7178 return __build_sched_domains(cpu_map
, NULL
);
7181 static cpumask_var_t
*doms_cur
; /* current sched domains */
7182 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7183 static struct sched_domain_attr
*dattr_cur
;
7184 /* attribues of custom domains in 'doms_cur' */
7187 * Special case: If a kmalloc of a doms_cur partition (array of
7188 * cpumask) fails, then fallback to a single sched domain,
7189 * as determined by the single cpumask fallback_doms.
7191 static cpumask_var_t fallback_doms
;
7194 * arch_update_cpu_topology lets virtualized architectures update the
7195 * cpu core maps. It is supposed to return 1 if the topology changed
7196 * or 0 if it stayed the same.
7198 int __attribute__((weak
)) arch_update_cpu_topology(void)
7203 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7206 cpumask_var_t
*doms
;
7208 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7211 for (i
= 0; i
< ndoms
; i
++) {
7212 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7213 free_sched_domains(doms
, i
);
7220 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7223 for (i
= 0; i
< ndoms
; i
++)
7224 free_cpumask_var(doms
[i
]);
7229 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7230 * For now this just excludes isolated cpus, but could be used to
7231 * exclude other special cases in the future.
7233 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7237 arch_update_cpu_topology();
7239 doms_cur
= alloc_sched_domains(ndoms_cur
);
7241 doms_cur
= &fallback_doms
;
7242 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7244 err
= build_sched_domains(doms_cur
[0]);
7245 register_sched_domain_sysctl();
7250 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7251 struct cpumask
*tmpmask
)
7253 free_sched_groups(cpu_map
, tmpmask
);
7257 * Detach sched domains from a group of cpus specified in cpu_map
7258 * These cpus will now be attached to the NULL domain
7260 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7262 /* Save because hotplug lock held. */
7263 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7266 for_each_cpu(i
, cpu_map
)
7267 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7268 synchronize_sched();
7269 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7272 /* handle null as "default" */
7273 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7274 struct sched_domain_attr
*new, int idx_new
)
7276 struct sched_domain_attr tmp
;
7283 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7284 new ? (new + idx_new
) : &tmp
,
7285 sizeof(struct sched_domain_attr
));
7289 * Partition sched domains as specified by the 'ndoms_new'
7290 * cpumasks in the array doms_new[] of cpumasks. This compares
7291 * doms_new[] to the current sched domain partitioning, doms_cur[].
7292 * It destroys each deleted domain and builds each new domain.
7294 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7295 * The masks don't intersect (don't overlap.) We should setup one
7296 * sched domain for each mask. CPUs not in any of the cpumasks will
7297 * not be load balanced. If the same cpumask appears both in the
7298 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7301 * The passed in 'doms_new' should be allocated using
7302 * alloc_sched_domains. This routine takes ownership of it and will
7303 * free_sched_domains it when done with it. If the caller failed the
7304 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7305 * and partition_sched_domains() will fallback to the single partition
7306 * 'fallback_doms', it also forces the domains to be rebuilt.
7308 * If doms_new == NULL it will be replaced with cpu_online_mask.
7309 * ndoms_new == 0 is a special case for destroying existing domains,
7310 * and it will not create the default domain.
7312 * Call with hotplug lock held
7314 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7315 struct sched_domain_attr
*dattr_new
)
7320 mutex_lock(&sched_domains_mutex
);
7322 /* always unregister in case we don't destroy any domains */
7323 unregister_sched_domain_sysctl();
7325 /* Let architecture update cpu core mappings. */
7326 new_topology
= arch_update_cpu_topology();
7328 n
= doms_new
? ndoms_new
: 0;
7330 /* Destroy deleted domains */
7331 for (i
= 0; i
< ndoms_cur
; i
++) {
7332 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7333 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7334 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7337 /* no match - a current sched domain not in new doms_new[] */
7338 detach_destroy_domains(doms_cur
[i
]);
7343 if (doms_new
== NULL
) {
7345 doms_new
= &fallback_doms
;
7346 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7347 WARN_ON_ONCE(dattr_new
);
7350 /* Build new domains */
7351 for (i
= 0; i
< ndoms_new
; i
++) {
7352 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7353 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7354 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7357 /* no match - add a new doms_new */
7358 __build_sched_domains(doms_new
[i
],
7359 dattr_new
? dattr_new
+ i
: NULL
);
7364 /* Remember the new sched domains */
7365 if (doms_cur
!= &fallback_doms
)
7366 free_sched_domains(doms_cur
, ndoms_cur
);
7367 kfree(dattr_cur
); /* kfree(NULL) is safe */
7368 doms_cur
= doms_new
;
7369 dattr_cur
= dattr_new
;
7370 ndoms_cur
= ndoms_new
;
7372 register_sched_domain_sysctl();
7374 mutex_unlock(&sched_domains_mutex
);
7377 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7378 static void arch_reinit_sched_domains(void)
7382 /* Destroy domains first to force the rebuild */
7383 partition_sched_domains(0, NULL
, NULL
);
7385 rebuild_sched_domains();
7389 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7391 unsigned int level
= 0;
7393 if (sscanf(buf
, "%u", &level
) != 1)
7397 * level is always be positive so don't check for
7398 * level < POWERSAVINGS_BALANCE_NONE which is 0
7399 * What happens on 0 or 1 byte write,
7400 * need to check for count as well?
7403 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7407 sched_smt_power_savings
= level
;
7409 sched_mc_power_savings
= level
;
7411 arch_reinit_sched_domains();
7416 #ifdef CONFIG_SCHED_MC
7417 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7418 struct sysdev_class_attribute
*attr
,
7421 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7423 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7424 struct sysdev_class_attribute
*attr
,
7425 const char *buf
, size_t count
)
7427 return sched_power_savings_store(buf
, count
, 0);
7429 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7430 sched_mc_power_savings_show
,
7431 sched_mc_power_savings_store
);
7434 #ifdef CONFIG_SCHED_SMT
7435 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7436 struct sysdev_class_attribute
*attr
,
7439 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7441 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7442 struct sysdev_class_attribute
*attr
,
7443 const char *buf
, size_t count
)
7445 return sched_power_savings_store(buf
, count
, 1);
7447 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7448 sched_smt_power_savings_show
,
7449 sched_smt_power_savings_store
);
7452 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7456 #ifdef CONFIG_SCHED_SMT
7458 err
= sysfs_create_file(&cls
->kset
.kobj
,
7459 &attr_sched_smt_power_savings
.attr
);
7461 #ifdef CONFIG_SCHED_MC
7462 if (!err
&& mc_capable())
7463 err
= sysfs_create_file(&cls
->kset
.kobj
,
7464 &attr_sched_mc_power_savings
.attr
);
7468 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7470 #ifndef CONFIG_CPUSETS
7472 * Add online and remove offline CPUs from the scheduler domains.
7473 * When cpusets are enabled they take over this function.
7475 static int update_sched_domains(struct notifier_block
*nfb
,
7476 unsigned long action
, void *hcpu
)
7480 case CPU_ONLINE_FROZEN
:
7481 case CPU_DOWN_PREPARE
:
7482 case CPU_DOWN_PREPARE_FROZEN
:
7483 case CPU_DOWN_FAILED
:
7484 case CPU_DOWN_FAILED_FROZEN
:
7485 partition_sched_domains(1, NULL
, NULL
);
7494 static int update_runtime(struct notifier_block
*nfb
,
7495 unsigned long action
, void *hcpu
)
7497 int cpu
= (int)(long)hcpu
;
7500 case CPU_DOWN_PREPARE
:
7501 case CPU_DOWN_PREPARE_FROZEN
:
7502 disable_runtime(cpu_rq(cpu
));
7505 case CPU_DOWN_FAILED
:
7506 case CPU_DOWN_FAILED_FROZEN
:
7508 case CPU_ONLINE_FROZEN
:
7509 enable_runtime(cpu_rq(cpu
));
7517 void __init
sched_init_smp(void)
7519 cpumask_var_t non_isolated_cpus
;
7521 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7522 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7524 #if defined(CONFIG_NUMA)
7525 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7527 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7530 mutex_lock(&sched_domains_mutex
);
7531 arch_init_sched_domains(cpu_active_mask
);
7532 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7533 if (cpumask_empty(non_isolated_cpus
))
7534 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7535 mutex_unlock(&sched_domains_mutex
);
7538 #ifndef CONFIG_CPUSETS
7539 /* XXX: Theoretical race here - CPU may be hotplugged now */
7540 hotcpu_notifier(update_sched_domains
, 0);
7543 /* RT runtime code needs to handle some hotplug events */
7544 hotcpu_notifier(update_runtime
, 0);
7548 /* Move init over to a non-isolated CPU */
7549 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7551 sched_init_granularity();
7552 free_cpumask_var(non_isolated_cpus
);
7554 init_sched_rt_class();
7557 void __init
sched_init_smp(void)
7559 sched_init_granularity();
7561 #endif /* CONFIG_SMP */
7563 const_debug
unsigned int sysctl_timer_migration
= 1;
7565 int in_sched_functions(unsigned long addr
)
7567 return in_lock_functions(addr
) ||
7568 (addr
>= (unsigned long)__sched_text_start
7569 && addr
< (unsigned long)__sched_text_end
);
7572 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7574 cfs_rq
->tasks_timeline
= RB_ROOT
;
7575 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7576 #ifdef CONFIG_FAIR_GROUP_SCHED
7579 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7582 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7584 struct rt_prio_array
*array
;
7587 array
= &rt_rq
->active
;
7588 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7589 INIT_LIST_HEAD(array
->queue
+ i
);
7590 __clear_bit(i
, array
->bitmap
);
7592 /* delimiter for bitsearch: */
7593 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7595 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7596 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7598 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7602 rt_rq
->rt_nr_migratory
= 0;
7603 rt_rq
->overloaded
= 0;
7604 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7608 rt_rq
->rt_throttled
= 0;
7609 rt_rq
->rt_runtime
= 0;
7610 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7612 #ifdef CONFIG_RT_GROUP_SCHED
7613 rt_rq
->rt_nr_boosted
= 0;
7618 #ifdef CONFIG_FAIR_GROUP_SCHED
7619 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7620 struct sched_entity
*se
, int cpu
, int add
,
7621 struct sched_entity
*parent
)
7623 struct rq
*rq
= cpu_rq(cpu
);
7624 tg
->cfs_rq
[cpu
] = cfs_rq
;
7625 init_cfs_rq(cfs_rq
, rq
);
7628 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7631 /* se could be NULL for init_task_group */
7636 se
->cfs_rq
= &rq
->cfs
;
7638 se
->cfs_rq
= parent
->my_q
;
7641 se
->load
.weight
= tg
->shares
;
7642 se
->load
.inv_weight
= 0;
7643 se
->parent
= parent
;
7647 #ifdef CONFIG_RT_GROUP_SCHED
7648 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7649 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7650 struct sched_rt_entity
*parent
)
7652 struct rq
*rq
= cpu_rq(cpu
);
7654 tg
->rt_rq
[cpu
] = rt_rq
;
7655 init_rt_rq(rt_rq
, rq
);
7657 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7659 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7661 tg
->rt_se
[cpu
] = rt_se
;
7666 rt_se
->rt_rq
= &rq
->rt
;
7668 rt_se
->rt_rq
= parent
->my_q
;
7670 rt_se
->my_q
= rt_rq
;
7671 rt_se
->parent
= parent
;
7672 INIT_LIST_HEAD(&rt_se
->run_list
);
7676 void __init
sched_init(void)
7679 unsigned long alloc_size
= 0, ptr
;
7681 #ifdef CONFIG_FAIR_GROUP_SCHED
7682 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7684 #ifdef CONFIG_RT_GROUP_SCHED
7685 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7687 #ifdef CONFIG_CPUMASK_OFFSTACK
7688 alloc_size
+= num_possible_cpus() * cpumask_size();
7691 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7693 #ifdef CONFIG_FAIR_GROUP_SCHED
7694 init_task_group
.se
= (struct sched_entity
**)ptr
;
7695 ptr
+= nr_cpu_ids
* sizeof(void **);
7697 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7698 ptr
+= nr_cpu_ids
* sizeof(void **);
7700 #endif /* CONFIG_FAIR_GROUP_SCHED */
7701 #ifdef CONFIG_RT_GROUP_SCHED
7702 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7703 ptr
+= nr_cpu_ids
* sizeof(void **);
7705 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7706 ptr
+= nr_cpu_ids
* sizeof(void **);
7708 #endif /* CONFIG_RT_GROUP_SCHED */
7709 #ifdef CONFIG_CPUMASK_OFFSTACK
7710 for_each_possible_cpu(i
) {
7711 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7712 ptr
+= cpumask_size();
7714 #endif /* CONFIG_CPUMASK_OFFSTACK */
7718 init_defrootdomain();
7721 init_rt_bandwidth(&def_rt_bandwidth
,
7722 global_rt_period(), global_rt_runtime());
7724 #ifdef CONFIG_RT_GROUP_SCHED
7725 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7726 global_rt_period(), global_rt_runtime());
7727 #endif /* CONFIG_RT_GROUP_SCHED */
7729 #ifdef CONFIG_CGROUP_SCHED
7730 list_add(&init_task_group
.list
, &task_groups
);
7731 INIT_LIST_HEAD(&init_task_group
.children
);
7733 #endif /* CONFIG_CGROUP_SCHED */
7735 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7736 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7737 __alignof__(unsigned long));
7739 for_each_possible_cpu(i
) {
7743 raw_spin_lock_init(&rq
->lock
);
7745 rq
->calc_load_active
= 0;
7746 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7747 init_cfs_rq(&rq
->cfs
, rq
);
7748 init_rt_rq(&rq
->rt
, rq
);
7749 #ifdef CONFIG_FAIR_GROUP_SCHED
7750 init_task_group
.shares
= init_task_group_load
;
7751 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7752 #ifdef CONFIG_CGROUP_SCHED
7754 * How much cpu bandwidth does init_task_group get?
7756 * In case of task-groups formed thr' the cgroup filesystem, it
7757 * gets 100% of the cpu resources in the system. This overall
7758 * system cpu resource is divided among the tasks of
7759 * init_task_group and its child task-groups in a fair manner,
7760 * based on each entity's (task or task-group's) weight
7761 * (se->load.weight).
7763 * In other words, if init_task_group has 10 tasks of weight
7764 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7765 * then A0's share of the cpu resource is:
7767 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7769 * We achieve this by letting init_task_group's tasks sit
7770 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7772 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7774 #endif /* CONFIG_FAIR_GROUP_SCHED */
7776 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7777 #ifdef CONFIG_RT_GROUP_SCHED
7778 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7779 #ifdef CONFIG_CGROUP_SCHED
7780 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7784 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7785 rq
->cpu_load
[j
] = 0;
7789 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7790 rq
->post_schedule
= 0;
7791 rq
->active_balance
= 0;
7792 rq
->next_balance
= jiffies
;
7796 rq
->migration_thread
= NULL
;
7798 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7799 INIT_LIST_HEAD(&rq
->migration_queue
);
7800 rq_attach_root(rq
, &def_root_domain
);
7803 atomic_set(&rq
->nr_iowait
, 0);
7806 set_load_weight(&init_task
);
7808 #ifdef CONFIG_PREEMPT_NOTIFIERS
7809 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7813 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7816 #ifdef CONFIG_RT_MUTEXES
7817 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7821 * The boot idle thread does lazy MMU switching as well:
7823 atomic_inc(&init_mm
.mm_count
);
7824 enter_lazy_tlb(&init_mm
, current
);
7827 * Make us the idle thread. Technically, schedule() should not be
7828 * called from this thread, however somewhere below it might be,
7829 * but because we are the idle thread, we just pick up running again
7830 * when this runqueue becomes "idle".
7832 init_idle(current
, smp_processor_id());
7834 calc_load_update
= jiffies
+ LOAD_FREQ
;
7837 * During early bootup we pretend to be a normal task:
7839 current
->sched_class
= &fair_sched_class
;
7841 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7842 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7845 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7846 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7848 /* May be allocated at isolcpus cmdline parse time */
7849 if (cpu_isolated_map
== NULL
)
7850 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7855 scheduler_running
= 1;
7858 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7859 static inline int preempt_count_equals(int preempt_offset
)
7861 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7863 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7866 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7869 static unsigned long prev_jiffy
; /* ratelimiting */
7871 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7872 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7874 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7876 prev_jiffy
= jiffies
;
7879 "BUG: sleeping function called from invalid context at %s:%d\n",
7882 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7883 in_atomic(), irqs_disabled(),
7884 current
->pid
, current
->comm
);
7886 debug_show_held_locks(current
);
7887 if (irqs_disabled())
7888 print_irqtrace_events(current
);
7892 EXPORT_SYMBOL(__might_sleep
);
7895 #ifdef CONFIG_MAGIC_SYSRQ
7896 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7900 update_rq_clock(rq
);
7901 on_rq
= p
->se
.on_rq
;
7903 deactivate_task(rq
, p
, 0);
7904 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7906 activate_task(rq
, p
, 0);
7907 resched_task(rq
->curr
);
7911 void normalize_rt_tasks(void)
7913 struct task_struct
*g
, *p
;
7914 unsigned long flags
;
7917 read_lock_irqsave(&tasklist_lock
, flags
);
7918 do_each_thread(g
, p
) {
7920 * Only normalize user tasks:
7925 p
->se
.exec_start
= 0;
7926 #ifdef CONFIG_SCHEDSTATS
7927 p
->se
.wait_start
= 0;
7928 p
->se
.sleep_start
= 0;
7929 p
->se
.block_start
= 0;
7934 * Renice negative nice level userspace
7937 if (TASK_NICE(p
) < 0 && p
->mm
)
7938 set_user_nice(p
, 0);
7942 raw_spin_lock(&p
->pi_lock
);
7943 rq
= __task_rq_lock(p
);
7945 normalize_task(rq
, p
);
7947 __task_rq_unlock(rq
);
7948 raw_spin_unlock(&p
->pi_lock
);
7949 } while_each_thread(g
, p
);
7951 read_unlock_irqrestore(&tasklist_lock
, flags
);
7954 #endif /* CONFIG_MAGIC_SYSRQ */
7958 * These functions are only useful for the IA64 MCA handling.
7960 * They can only be called when the whole system has been
7961 * stopped - every CPU needs to be quiescent, and no scheduling
7962 * activity can take place. Using them for anything else would
7963 * be a serious bug, and as a result, they aren't even visible
7964 * under any other configuration.
7968 * curr_task - return the current task for a given cpu.
7969 * @cpu: the processor in question.
7971 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7973 struct task_struct
*curr_task(int cpu
)
7975 return cpu_curr(cpu
);
7979 * set_curr_task - set the current task for a given cpu.
7980 * @cpu: the processor in question.
7981 * @p: the task pointer to set.
7983 * Description: This function must only be used when non-maskable interrupts
7984 * are serviced on a separate stack. It allows the architecture to switch the
7985 * notion of the current task on a cpu in a non-blocking manner. This function
7986 * must be called with all CPU's synchronized, and interrupts disabled, the
7987 * and caller must save the original value of the current task (see
7988 * curr_task() above) and restore that value before reenabling interrupts and
7989 * re-starting the system.
7991 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7993 void set_curr_task(int cpu
, struct task_struct
*p
)
8000 #ifdef CONFIG_FAIR_GROUP_SCHED
8001 static void free_fair_sched_group(struct task_group
*tg
)
8005 for_each_possible_cpu(i
) {
8007 kfree(tg
->cfs_rq
[i
]);
8017 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8019 struct cfs_rq
*cfs_rq
;
8020 struct sched_entity
*se
;
8024 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8027 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8031 tg
->shares
= NICE_0_LOAD
;
8033 for_each_possible_cpu(i
) {
8036 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8037 GFP_KERNEL
, cpu_to_node(i
));
8041 se
= kzalloc_node(sizeof(struct sched_entity
),
8042 GFP_KERNEL
, cpu_to_node(i
));
8046 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8057 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8059 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8060 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8063 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8065 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8067 #else /* !CONFG_FAIR_GROUP_SCHED */
8068 static inline void free_fair_sched_group(struct task_group
*tg
)
8073 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8078 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8082 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8085 #endif /* CONFIG_FAIR_GROUP_SCHED */
8087 #ifdef CONFIG_RT_GROUP_SCHED
8088 static void free_rt_sched_group(struct task_group
*tg
)
8092 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8094 for_each_possible_cpu(i
) {
8096 kfree(tg
->rt_rq
[i
]);
8098 kfree(tg
->rt_se
[i
]);
8106 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8108 struct rt_rq
*rt_rq
;
8109 struct sched_rt_entity
*rt_se
;
8113 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8116 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8120 init_rt_bandwidth(&tg
->rt_bandwidth
,
8121 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8123 for_each_possible_cpu(i
) {
8126 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8127 GFP_KERNEL
, cpu_to_node(i
));
8131 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8132 GFP_KERNEL
, cpu_to_node(i
));
8136 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8147 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8149 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8150 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8153 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8155 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8157 #else /* !CONFIG_RT_GROUP_SCHED */
8158 static inline void free_rt_sched_group(struct task_group
*tg
)
8163 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8168 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8172 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8175 #endif /* CONFIG_RT_GROUP_SCHED */
8177 #ifdef CONFIG_CGROUP_SCHED
8178 static void free_sched_group(struct task_group
*tg
)
8180 free_fair_sched_group(tg
);
8181 free_rt_sched_group(tg
);
8185 /* allocate runqueue etc for a new task group */
8186 struct task_group
*sched_create_group(struct task_group
*parent
)
8188 struct task_group
*tg
;
8189 unsigned long flags
;
8192 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8194 return ERR_PTR(-ENOMEM
);
8196 if (!alloc_fair_sched_group(tg
, parent
))
8199 if (!alloc_rt_sched_group(tg
, parent
))
8202 spin_lock_irqsave(&task_group_lock
, flags
);
8203 for_each_possible_cpu(i
) {
8204 register_fair_sched_group(tg
, i
);
8205 register_rt_sched_group(tg
, i
);
8207 list_add_rcu(&tg
->list
, &task_groups
);
8209 WARN_ON(!parent
); /* root should already exist */
8211 tg
->parent
= parent
;
8212 INIT_LIST_HEAD(&tg
->children
);
8213 list_add_rcu(&tg
->siblings
, &parent
->children
);
8214 spin_unlock_irqrestore(&task_group_lock
, flags
);
8219 free_sched_group(tg
);
8220 return ERR_PTR(-ENOMEM
);
8223 /* rcu callback to free various structures associated with a task group */
8224 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8226 /* now it should be safe to free those cfs_rqs */
8227 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8230 /* Destroy runqueue etc associated with a task group */
8231 void sched_destroy_group(struct task_group
*tg
)
8233 unsigned long flags
;
8236 spin_lock_irqsave(&task_group_lock
, flags
);
8237 for_each_possible_cpu(i
) {
8238 unregister_fair_sched_group(tg
, i
);
8239 unregister_rt_sched_group(tg
, i
);
8241 list_del_rcu(&tg
->list
);
8242 list_del_rcu(&tg
->siblings
);
8243 spin_unlock_irqrestore(&task_group_lock
, flags
);
8245 /* wait for possible concurrent references to cfs_rqs complete */
8246 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8249 /* change task's runqueue when it moves between groups.
8250 * The caller of this function should have put the task in its new group
8251 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8252 * reflect its new group.
8254 void sched_move_task(struct task_struct
*tsk
)
8257 unsigned long flags
;
8260 rq
= task_rq_lock(tsk
, &flags
);
8262 update_rq_clock(rq
);
8264 running
= task_current(rq
, tsk
);
8265 on_rq
= tsk
->se
.on_rq
;
8268 dequeue_task(rq
, tsk
, 0);
8269 if (unlikely(running
))
8270 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8272 #ifdef CONFIG_FAIR_GROUP_SCHED
8273 if (tsk
->sched_class
->task_move_group
)
8274 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8277 set_task_rq(tsk
, task_cpu(tsk
));
8279 if (unlikely(running
))
8280 tsk
->sched_class
->set_curr_task(rq
);
8282 enqueue_task(rq
, tsk
, 0, false);
8284 task_rq_unlock(rq
, &flags
);
8286 #endif /* CONFIG_CGROUP_SCHED */
8288 #ifdef CONFIG_FAIR_GROUP_SCHED
8289 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8291 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8296 dequeue_entity(cfs_rq
, se
, 0);
8298 se
->load
.weight
= shares
;
8299 se
->load
.inv_weight
= 0;
8302 enqueue_entity(cfs_rq
, se
, 0);
8305 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8307 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8308 struct rq
*rq
= cfs_rq
->rq
;
8309 unsigned long flags
;
8311 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8312 __set_se_shares(se
, shares
);
8313 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8316 static DEFINE_MUTEX(shares_mutex
);
8318 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8321 unsigned long flags
;
8324 * We can't change the weight of the root cgroup.
8329 if (shares
< MIN_SHARES
)
8330 shares
= MIN_SHARES
;
8331 else if (shares
> MAX_SHARES
)
8332 shares
= MAX_SHARES
;
8334 mutex_lock(&shares_mutex
);
8335 if (tg
->shares
== shares
)
8338 spin_lock_irqsave(&task_group_lock
, flags
);
8339 for_each_possible_cpu(i
)
8340 unregister_fair_sched_group(tg
, i
);
8341 list_del_rcu(&tg
->siblings
);
8342 spin_unlock_irqrestore(&task_group_lock
, flags
);
8344 /* wait for any ongoing reference to this group to finish */
8345 synchronize_sched();
8348 * Now we are free to modify the group's share on each cpu
8349 * w/o tripping rebalance_share or load_balance_fair.
8351 tg
->shares
= shares
;
8352 for_each_possible_cpu(i
) {
8356 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8357 set_se_shares(tg
->se
[i
], shares
);
8361 * Enable load balance activity on this group, by inserting it back on
8362 * each cpu's rq->leaf_cfs_rq_list.
8364 spin_lock_irqsave(&task_group_lock
, flags
);
8365 for_each_possible_cpu(i
)
8366 register_fair_sched_group(tg
, i
);
8367 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8368 spin_unlock_irqrestore(&task_group_lock
, flags
);
8370 mutex_unlock(&shares_mutex
);
8374 unsigned long sched_group_shares(struct task_group
*tg
)
8380 #ifdef CONFIG_RT_GROUP_SCHED
8382 * Ensure that the real time constraints are schedulable.
8384 static DEFINE_MUTEX(rt_constraints_mutex
);
8386 static unsigned long to_ratio(u64 period
, u64 runtime
)
8388 if (runtime
== RUNTIME_INF
)
8391 return div64_u64(runtime
<< 20, period
);
8394 /* Must be called with tasklist_lock held */
8395 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8397 struct task_struct
*g
, *p
;
8399 do_each_thread(g
, p
) {
8400 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8402 } while_each_thread(g
, p
);
8407 struct rt_schedulable_data
{
8408 struct task_group
*tg
;
8413 static int tg_schedulable(struct task_group
*tg
, void *data
)
8415 struct rt_schedulable_data
*d
= data
;
8416 struct task_group
*child
;
8417 unsigned long total
, sum
= 0;
8418 u64 period
, runtime
;
8420 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8421 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8424 period
= d
->rt_period
;
8425 runtime
= d
->rt_runtime
;
8429 * Cannot have more runtime than the period.
8431 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8435 * Ensure we don't starve existing RT tasks.
8437 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8440 total
= to_ratio(period
, runtime
);
8443 * Nobody can have more than the global setting allows.
8445 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8449 * The sum of our children's runtime should not exceed our own.
8451 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8452 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8453 runtime
= child
->rt_bandwidth
.rt_runtime
;
8455 if (child
== d
->tg
) {
8456 period
= d
->rt_period
;
8457 runtime
= d
->rt_runtime
;
8460 sum
+= to_ratio(period
, runtime
);
8469 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8471 struct rt_schedulable_data data
= {
8473 .rt_period
= period
,
8474 .rt_runtime
= runtime
,
8477 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8480 static int tg_set_bandwidth(struct task_group
*tg
,
8481 u64 rt_period
, u64 rt_runtime
)
8485 mutex_lock(&rt_constraints_mutex
);
8486 read_lock(&tasklist_lock
);
8487 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8491 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8492 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8493 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8495 for_each_possible_cpu(i
) {
8496 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8498 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8499 rt_rq
->rt_runtime
= rt_runtime
;
8500 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8502 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8504 read_unlock(&tasklist_lock
);
8505 mutex_unlock(&rt_constraints_mutex
);
8510 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8512 u64 rt_runtime
, rt_period
;
8514 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8515 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8516 if (rt_runtime_us
< 0)
8517 rt_runtime
= RUNTIME_INF
;
8519 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8522 long sched_group_rt_runtime(struct task_group
*tg
)
8526 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8529 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8530 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8531 return rt_runtime_us
;
8534 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8536 u64 rt_runtime
, rt_period
;
8538 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8539 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8544 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8547 long sched_group_rt_period(struct task_group
*tg
)
8551 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8552 do_div(rt_period_us
, NSEC_PER_USEC
);
8553 return rt_period_us
;
8556 static int sched_rt_global_constraints(void)
8558 u64 runtime
, period
;
8561 if (sysctl_sched_rt_period
<= 0)
8564 runtime
= global_rt_runtime();
8565 period
= global_rt_period();
8568 * Sanity check on the sysctl variables.
8570 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8573 mutex_lock(&rt_constraints_mutex
);
8574 read_lock(&tasklist_lock
);
8575 ret
= __rt_schedulable(NULL
, 0, 0);
8576 read_unlock(&tasklist_lock
);
8577 mutex_unlock(&rt_constraints_mutex
);
8582 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8584 /* Don't accept realtime tasks when there is no way for them to run */
8585 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8591 #else /* !CONFIG_RT_GROUP_SCHED */
8592 static int sched_rt_global_constraints(void)
8594 unsigned long flags
;
8597 if (sysctl_sched_rt_period
<= 0)
8601 * There's always some RT tasks in the root group
8602 * -- migration, kstopmachine etc..
8604 if (sysctl_sched_rt_runtime
== 0)
8607 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8608 for_each_possible_cpu(i
) {
8609 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8611 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8612 rt_rq
->rt_runtime
= global_rt_runtime();
8613 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8615 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8619 #endif /* CONFIG_RT_GROUP_SCHED */
8621 int sched_rt_handler(struct ctl_table
*table
, int write
,
8622 void __user
*buffer
, size_t *lenp
,
8626 int old_period
, old_runtime
;
8627 static DEFINE_MUTEX(mutex
);
8630 old_period
= sysctl_sched_rt_period
;
8631 old_runtime
= sysctl_sched_rt_runtime
;
8633 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8635 if (!ret
&& write
) {
8636 ret
= sched_rt_global_constraints();
8638 sysctl_sched_rt_period
= old_period
;
8639 sysctl_sched_rt_runtime
= old_runtime
;
8641 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8642 def_rt_bandwidth
.rt_period
=
8643 ns_to_ktime(global_rt_period());
8646 mutex_unlock(&mutex
);
8651 #ifdef CONFIG_CGROUP_SCHED
8653 /* return corresponding task_group object of a cgroup */
8654 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8656 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8657 struct task_group
, css
);
8660 static struct cgroup_subsys_state
*
8661 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8663 struct task_group
*tg
, *parent
;
8665 if (!cgrp
->parent
) {
8666 /* This is early initialization for the top cgroup */
8667 return &init_task_group
.css
;
8670 parent
= cgroup_tg(cgrp
->parent
);
8671 tg
= sched_create_group(parent
);
8673 return ERR_PTR(-ENOMEM
);
8679 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8681 struct task_group
*tg
= cgroup_tg(cgrp
);
8683 sched_destroy_group(tg
);
8687 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8689 #ifdef CONFIG_RT_GROUP_SCHED
8690 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8693 /* We don't support RT-tasks being in separate groups */
8694 if (tsk
->sched_class
!= &fair_sched_class
)
8701 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8702 struct task_struct
*tsk
, bool threadgroup
)
8704 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8708 struct task_struct
*c
;
8710 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8711 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8723 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8724 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8727 sched_move_task(tsk
);
8729 struct task_struct
*c
;
8731 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8738 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8742 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8745 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8747 struct task_group
*tg
= cgroup_tg(cgrp
);
8749 return (u64
) tg
->shares
;
8751 #endif /* CONFIG_FAIR_GROUP_SCHED */
8753 #ifdef CONFIG_RT_GROUP_SCHED
8754 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8757 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8760 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8762 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8765 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8768 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8771 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8773 return sched_group_rt_period(cgroup_tg(cgrp
));
8775 #endif /* CONFIG_RT_GROUP_SCHED */
8777 static struct cftype cpu_files
[] = {
8778 #ifdef CONFIG_FAIR_GROUP_SCHED
8781 .read_u64
= cpu_shares_read_u64
,
8782 .write_u64
= cpu_shares_write_u64
,
8785 #ifdef CONFIG_RT_GROUP_SCHED
8787 .name
= "rt_runtime_us",
8788 .read_s64
= cpu_rt_runtime_read
,
8789 .write_s64
= cpu_rt_runtime_write
,
8792 .name
= "rt_period_us",
8793 .read_u64
= cpu_rt_period_read_uint
,
8794 .write_u64
= cpu_rt_period_write_uint
,
8799 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8801 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8804 struct cgroup_subsys cpu_cgroup_subsys
= {
8806 .create
= cpu_cgroup_create
,
8807 .destroy
= cpu_cgroup_destroy
,
8808 .can_attach
= cpu_cgroup_can_attach
,
8809 .attach
= cpu_cgroup_attach
,
8810 .populate
= cpu_cgroup_populate
,
8811 .subsys_id
= cpu_cgroup_subsys_id
,
8815 #endif /* CONFIG_CGROUP_SCHED */
8817 #ifdef CONFIG_CGROUP_CPUACCT
8820 * CPU accounting code for task groups.
8822 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8823 * (balbir@in.ibm.com).
8826 /* track cpu usage of a group of tasks and its child groups */
8828 struct cgroup_subsys_state css
;
8829 /* cpuusage holds pointer to a u64-type object on every cpu */
8830 u64 __percpu
*cpuusage
;
8831 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8832 struct cpuacct
*parent
;
8835 struct cgroup_subsys cpuacct_subsys
;
8837 /* return cpu accounting group corresponding to this container */
8838 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8840 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8841 struct cpuacct
, css
);
8844 /* return cpu accounting group to which this task belongs */
8845 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8847 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8848 struct cpuacct
, css
);
8851 /* create a new cpu accounting group */
8852 static struct cgroup_subsys_state
*cpuacct_create(
8853 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8855 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8861 ca
->cpuusage
= alloc_percpu(u64
);
8865 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8866 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8867 goto out_free_counters
;
8870 ca
->parent
= cgroup_ca(cgrp
->parent
);
8876 percpu_counter_destroy(&ca
->cpustat
[i
]);
8877 free_percpu(ca
->cpuusage
);
8881 return ERR_PTR(-ENOMEM
);
8884 /* destroy an existing cpu accounting group */
8886 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8888 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8891 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8892 percpu_counter_destroy(&ca
->cpustat
[i
]);
8893 free_percpu(ca
->cpuusage
);
8897 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8899 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8902 #ifndef CONFIG_64BIT
8904 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8906 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8908 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8916 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8918 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8920 #ifndef CONFIG_64BIT
8922 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8924 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8926 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8932 /* return total cpu usage (in nanoseconds) of a group */
8933 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8935 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8936 u64 totalcpuusage
= 0;
8939 for_each_present_cpu(i
)
8940 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8942 return totalcpuusage
;
8945 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8948 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8957 for_each_present_cpu(i
)
8958 cpuacct_cpuusage_write(ca
, i
, 0);
8964 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8967 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8971 for_each_present_cpu(i
) {
8972 percpu
= cpuacct_cpuusage_read(ca
, i
);
8973 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8975 seq_printf(m
, "\n");
8979 static const char *cpuacct_stat_desc
[] = {
8980 [CPUACCT_STAT_USER
] = "user",
8981 [CPUACCT_STAT_SYSTEM
] = "system",
8984 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8985 struct cgroup_map_cb
*cb
)
8987 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8990 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8991 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8992 val
= cputime64_to_clock_t(val
);
8993 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8998 static struct cftype files
[] = {
9001 .read_u64
= cpuusage_read
,
9002 .write_u64
= cpuusage_write
,
9005 .name
= "usage_percpu",
9006 .read_seq_string
= cpuacct_percpu_seq_read
,
9010 .read_map
= cpuacct_stats_show
,
9014 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9016 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9020 * charge this task's execution time to its accounting group.
9022 * called with rq->lock held.
9024 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9029 if (unlikely(!cpuacct_subsys
.active
))
9032 cpu
= task_cpu(tsk
);
9038 for (; ca
; ca
= ca
->parent
) {
9039 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9040 *cpuusage
+= cputime
;
9047 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9048 * in cputime_t units. As a result, cpuacct_update_stats calls
9049 * percpu_counter_add with values large enough to always overflow the
9050 * per cpu batch limit causing bad SMP scalability.
9052 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9053 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9054 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9057 #define CPUACCT_BATCH \
9058 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9060 #define CPUACCT_BATCH 0
9064 * Charge the system/user time to the task's accounting group.
9066 static void cpuacct_update_stats(struct task_struct
*tsk
,
9067 enum cpuacct_stat_index idx
, cputime_t val
)
9070 int batch
= CPUACCT_BATCH
;
9072 if (unlikely(!cpuacct_subsys
.active
))
9079 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9085 struct cgroup_subsys cpuacct_subsys
= {
9087 .create
= cpuacct_create
,
9088 .destroy
= cpuacct_destroy
,
9089 .populate
= cpuacct_populate
,
9090 .subsys_id
= cpuacct_subsys_id
,
9092 #endif /* CONFIG_CGROUP_CPUACCT */
9096 int rcu_expedited_torture_stats(char *page
)
9100 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9102 void synchronize_sched_expedited(void)
9105 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9107 #else /* #ifndef CONFIG_SMP */
9109 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9110 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9112 #define RCU_EXPEDITED_STATE_POST -2
9113 #define RCU_EXPEDITED_STATE_IDLE -1
9115 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9117 int rcu_expedited_torture_stats(char *page
)
9122 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9123 for_each_online_cpu(cpu
) {
9124 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9125 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9127 cnt
+= sprintf(&page
[cnt
], "\n");
9130 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9132 static long synchronize_sched_expedited_count
;
9135 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9136 * approach to force grace period to end quickly. This consumes
9137 * significant time on all CPUs, and is thus not recommended for
9138 * any sort of common-case code.
9140 * Note that it is illegal to call this function while holding any
9141 * lock that is acquired by a CPU-hotplug notifier. Failing to
9142 * observe this restriction will result in deadlock.
9144 void synchronize_sched_expedited(void)
9147 unsigned long flags
;
9148 bool need_full_sync
= 0;
9150 struct migration_req
*req
;
9154 smp_mb(); /* ensure prior mod happens before capturing snap. */
9155 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9157 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9159 if (trycount
++ < 10)
9160 udelay(trycount
* num_online_cpus());
9162 synchronize_sched();
9165 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9166 smp_mb(); /* ensure test happens before caller kfree */
9171 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9172 for_each_online_cpu(cpu
) {
9174 req
= &per_cpu(rcu_migration_req
, cpu
);
9175 init_completion(&req
->done
);
9177 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9178 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9179 list_add(&req
->list
, &rq
->migration_queue
);
9180 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9181 wake_up_process(rq
->migration_thread
);
9183 for_each_online_cpu(cpu
) {
9184 rcu_expedited_state
= cpu
;
9185 req
= &per_cpu(rcu_migration_req
, cpu
);
9187 wait_for_completion(&req
->done
);
9188 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9189 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9191 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9192 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9194 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9195 synchronize_sched_expedited_count
++;
9196 mutex_unlock(&rcu_sched_expedited_mutex
);
9199 synchronize_sched();
9201 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9203 #endif /* #else #ifndef CONFIG_SMP */