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_counter.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/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.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)
125 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
133 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
142 sg
->__cpu_power
+= val
;
143 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
147 static inline int rt_policy(int policy
)
149 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
154 static inline int task_has_rt_policy(struct task_struct
*p
)
156 return rt_policy(p
->policy
);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array
{
163 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
164 struct list_head queue
[MAX_RT_PRIO
];
167 struct rt_bandwidth
{
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock
;
172 struct hrtimer rt_period_timer
;
175 static struct rt_bandwidth def_rt_bandwidth
;
177 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
179 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
181 struct rt_bandwidth
*rt_b
=
182 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
188 now
= hrtimer_cb_get_time(timer
);
189 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
194 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
197 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
201 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
203 rt_b
->rt_period
= ns_to_ktime(period
);
204 rt_b
->rt_runtime
= runtime
;
206 spin_lock_init(&rt_b
->rt_runtime_lock
);
208 hrtimer_init(&rt_b
->rt_period_timer
,
209 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
210 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime
>= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
222 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
225 if (hrtimer_active(&rt_b
->rt_period_timer
))
228 spin_lock(&rt_b
->rt_runtime_lock
);
233 if (hrtimer_active(&rt_b
->rt_period_timer
))
236 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
237 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
239 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
240 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
241 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
242 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
243 HRTIMER_MODE_ABS_PINNED
, 0);
245 spin_unlock(&rt_b
->rt_runtime_lock
);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
251 hrtimer_cancel(&rt_b
->rt_period_timer
);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex
);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups
);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css
;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity
**se
;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq
**cfs_rq
;
284 unsigned long shares
;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity
**rt_se
;
289 struct rt_rq
**rt_rq
;
291 struct rt_bandwidth rt_bandwidth
;
295 struct list_head list
;
297 struct task_group
*parent
;
298 struct list_head siblings
;
299 struct list_head children
;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct
*user
)
307 user
->tg
->uid
= user
->uid
;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group
;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq
, init_tg_cfs_rq
) ____cacheline_aligned_in_smp
;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
326 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock
);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group
.children
);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group
;
370 /* return group to which a task belongs */
371 static inline struct task_group
*task_group(struct task_struct
*p
)
373 struct task_group
*tg
;
375 #ifdef CONFIG_USER_SCHED
377 tg
= __task_cred(p
)->user
->tg
;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
381 struct task_group
, css
);
383 tg
= &init_task_group
;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
393 p
->se
.parent
= task_group(p
)->se
[cpu
];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
398 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
412 static inline struct task_group
*task_group(struct task_struct
*p
)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load
;
422 unsigned long nr_running
;
427 struct rb_root tasks_timeline
;
428 struct rb_node
*rb_leftmost
;
430 struct list_head tasks
;
431 struct list_head
*balance_iterator
;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity
*curr
, *next
, *last
;
439 unsigned int nr_spread_over
;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list
;
453 struct task_group
*tg
; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight
;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load
;
470 * this cpu's part of tg->shares
472 unsigned long shares
;
475 * load.weight at the time we set shares
477 unsigned long rq_weight
;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active
;
485 unsigned long rt_nr_running
;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr
; /* highest queued rt task prio */
490 int next
; /* next highest */
495 unsigned long rt_nr_migratory
;
496 unsigned long rt_nr_total
;
498 struct plist_head pushable_tasks
;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock
;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted
;
510 struct list_head leaf_rt_rq_list
;
511 struct task_group
*tg
;
512 struct sched_rt_entity
*rt_se
;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online
;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask
;
538 struct cpupri cpupri
;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu
;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain
;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running
;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
577 unsigned long last_tick_seen
;
578 unsigned char in_nohz_recently
;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load
;
582 unsigned long nr_load_updates
;
584 u64 nr_migrations_in
;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list
;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list
;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible
;
605 struct task_struct
*curr
, *idle
;
606 unsigned long next_balance
;
607 struct mm_struct
*prev_mm
;
614 struct root_domain
*rd
;
615 struct sched_domain
*sd
;
617 unsigned char idle_at_tick
;
618 /* For active balancing */
622 /* cpu of this runqueue: */
626 unsigned long avg_load_per_task
;
628 struct task_struct
*migration_thread
;
629 struct list_head migration_queue
;
635 /* calc_load related fields */
636 unsigned long calc_load_update
;
637 long calc_load_active
;
639 #ifdef CONFIG_SCHED_HRTICK
641 int hrtick_csd_pending
;
642 struct call_single_data hrtick_csd
;
644 struct hrtimer hrtick_timer
;
647 #ifdef CONFIG_SCHEDSTATS
649 struct sched_info rq_sched_info
;
650 unsigned long long rq_cpu_time
;
651 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
653 /* sys_sched_yield() stats */
654 unsigned int yld_count
;
656 /* schedule() stats */
657 unsigned int sched_switch
;
658 unsigned int sched_count
;
659 unsigned int sched_goidle
;
661 /* try_to_wake_up() stats */
662 unsigned int ttwu_count
;
663 unsigned int ttwu_local
;
666 unsigned int bkl_count
;
670 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
672 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
674 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
677 static inline int cpu_of(struct rq
*rq
)
687 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
688 * See detach_destroy_domains: synchronize_sched for details.
690 * The domain tree of any CPU may only be accessed from within
691 * preempt-disabled sections.
693 #define for_each_domain(cpu, __sd) \
694 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
696 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
697 #define this_rq() (&__get_cpu_var(runqueues))
698 #define task_rq(p) cpu_rq(task_cpu(p))
699 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
700 #define raw_rq() (&__raw_get_cpu_var(runqueues))
702 inline void update_rq_clock(struct rq
*rq
)
704 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
708 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
710 #ifdef CONFIG_SCHED_DEBUG
711 # define const_debug __read_mostly
713 # define const_debug static const
719 * Returns true if the current cpu runqueue is locked.
720 * This interface allows printk to be called with the runqueue lock
721 * held and know whether or not it is OK to wake up the klogd.
723 int runqueue_is_locked(void)
726 struct rq
*rq
= cpu_rq(cpu
);
729 ret
= spin_is_locked(&rq
->lock
);
735 * Debugging: various feature bits
738 #define SCHED_FEAT(name, enabled) \
739 __SCHED_FEAT_##name ,
742 #include "sched_features.h"
747 #define SCHED_FEAT(name, enabled) \
748 (1UL << __SCHED_FEAT_##name) * enabled |
750 const_debug
unsigned int sysctl_sched_features
=
751 #include "sched_features.h"
756 #ifdef CONFIG_SCHED_DEBUG
757 #define SCHED_FEAT(name, enabled) \
760 static __read_mostly
char *sched_feat_names
[] = {
761 #include "sched_features.h"
767 static int sched_feat_show(struct seq_file
*m
, void *v
)
771 for (i
= 0; sched_feat_names
[i
]; i
++) {
772 if (!(sysctl_sched_features
& (1UL << i
)))
774 seq_printf(m
, "%s ", sched_feat_names
[i
]);
782 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
783 size_t cnt
, loff_t
*ppos
)
793 if (copy_from_user(&buf
, ubuf
, cnt
))
798 if (strncmp(buf
, "NO_", 3) == 0) {
803 for (i
= 0; sched_feat_names
[i
]; i
++) {
804 int len
= strlen(sched_feat_names
[i
]);
806 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
808 sysctl_sched_features
&= ~(1UL << i
);
810 sysctl_sched_features
|= (1UL << i
);
815 if (!sched_feat_names
[i
])
823 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
825 return single_open(filp
, sched_feat_show
, NULL
);
828 static struct file_operations sched_feat_fops
= {
829 .open
= sched_feat_open
,
830 .write
= sched_feat_write
,
833 .release
= single_release
,
836 static __init
int sched_init_debug(void)
838 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
843 late_initcall(sched_init_debug
);
847 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
850 * Number of tasks to iterate in a single balance run.
851 * Limited because this is done with IRQs disabled.
853 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
856 * ratelimit for updating the group shares.
859 unsigned int sysctl_sched_shares_ratelimit
= 250000;
862 * Inject some fuzzyness into changing the per-cpu group shares
863 * this avoids remote rq-locks at the expense of fairness.
866 unsigned int sysctl_sched_shares_thresh
= 4;
869 * period over which we average the RT time consumption, measured
874 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
877 * period over which we measure -rt task cpu usage in us.
880 unsigned int sysctl_sched_rt_period
= 1000000;
882 static __read_mostly
int scheduler_running
;
885 * part of the period that we allow rt tasks to run in us.
888 int sysctl_sched_rt_runtime
= 950000;
890 static inline u64
global_rt_period(void)
892 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
895 static inline u64
global_rt_runtime(void)
897 if (sysctl_sched_rt_runtime
< 0)
900 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
903 #ifndef prepare_arch_switch
904 # define prepare_arch_switch(next) do { } while (0)
906 #ifndef finish_arch_switch
907 # define finish_arch_switch(prev) do { } while (0)
910 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
912 return rq
->curr
== p
;
915 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
916 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
918 return task_current(rq
, p
);
921 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
925 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
927 #ifdef CONFIG_DEBUG_SPINLOCK
928 /* this is a valid case when another task releases the spinlock */
929 rq
->lock
.owner
= current
;
932 * If we are tracking spinlock dependencies then we have to
933 * fix up the runqueue lock - which gets 'carried over' from
936 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
938 spin_unlock_irq(&rq
->lock
);
941 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
942 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
947 return task_current(rq
, p
);
951 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
955 * We can optimise this out completely for !SMP, because the
956 * SMP rebalancing from interrupt is the only thing that cares
961 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
962 spin_unlock_irq(&rq
->lock
);
964 spin_unlock(&rq
->lock
);
968 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
972 * After ->oncpu is cleared, the task can be moved to a different CPU.
973 * We must ensure this doesn't happen until the switch is completely
979 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
983 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
986 * __task_rq_lock - lock the runqueue a given task resides on.
987 * Must be called interrupts disabled.
989 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
993 struct rq
*rq
= task_rq(p
);
994 spin_lock(&rq
->lock
);
995 if (likely(rq
== task_rq(p
)))
997 spin_unlock(&rq
->lock
);
1002 * task_rq_lock - lock the runqueue a given task resides on and disable
1003 * interrupts. Note the ordering: we can safely lookup the task_rq without
1004 * explicitly disabling preemption.
1006 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1007 __acquires(rq
->lock
)
1012 local_irq_save(*flags
);
1014 spin_lock(&rq
->lock
);
1015 if (likely(rq
== task_rq(p
)))
1017 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1021 void task_rq_unlock_wait(struct task_struct
*p
)
1023 struct rq
*rq
= task_rq(p
);
1025 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1026 spin_unlock_wait(&rq
->lock
);
1029 static void __task_rq_unlock(struct rq
*rq
)
1030 __releases(rq
->lock
)
1032 spin_unlock(&rq
->lock
);
1035 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1036 __releases(rq
->lock
)
1038 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1042 * this_rq_lock - lock this runqueue and disable interrupts.
1044 static struct rq
*this_rq_lock(void)
1045 __acquires(rq
->lock
)
1049 local_irq_disable();
1051 spin_lock(&rq
->lock
);
1056 #ifdef CONFIG_SCHED_HRTICK
1058 * Use HR-timers to deliver accurate preemption points.
1060 * Its all a bit involved since we cannot program an hrt while holding the
1061 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1064 * When we get rescheduled we reprogram the hrtick_timer outside of the
1070 * - enabled by features
1071 * - hrtimer is actually high res
1073 static inline int hrtick_enabled(struct rq
*rq
)
1075 if (!sched_feat(HRTICK
))
1077 if (!cpu_active(cpu_of(rq
)))
1079 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1082 static void hrtick_clear(struct rq
*rq
)
1084 if (hrtimer_active(&rq
->hrtick_timer
))
1085 hrtimer_cancel(&rq
->hrtick_timer
);
1089 * High-resolution timer tick.
1090 * Runs from hardirq context with interrupts disabled.
1092 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1094 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1096 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1098 spin_lock(&rq
->lock
);
1099 update_rq_clock(rq
);
1100 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1101 spin_unlock(&rq
->lock
);
1103 return HRTIMER_NORESTART
;
1108 * called from hardirq (IPI) context
1110 static void __hrtick_start(void *arg
)
1112 struct rq
*rq
= arg
;
1114 spin_lock(&rq
->lock
);
1115 hrtimer_restart(&rq
->hrtick_timer
);
1116 rq
->hrtick_csd_pending
= 0;
1117 spin_unlock(&rq
->lock
);
1121 * Called to set the hrtick timer state.
1123 * called with rq->lock held and irqs disabled
1125 static void hrtick_start(struct rq
*rq
, u64 delay
)
1127 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1128 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1130 hrtimer_set_expires(timer
, time
);
1132 if (rq
== this_rq()) {
1133 hrtimer_restart(timer
);
1134 } else if (!rq
->hrtick_csd_pending
) {
1135 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1136 rq
->hrtick_csd_pending
= 1;
1141 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1143 int cpu
= (int)(long)hcpu
;
1146 case CPU_UP_CANCELED
:
1147 case CPU_UP_CANCELED_FROZEN
:
1148 case CPU_DOWN_PREPARE
:
1149 case CPU_DOWN_PREPARE_FROZEN
:
1151 case CPU_DEAD_FROZEN
:
1152 hrtick_clear(cpu_rq(cpu
));
1159 static __init
void init_hrtick(void)
1161 hotcpu_notifier(hotplug_hrtick
, 0);
1165 * Called to set the hrtick timer state.
1167 * called with rq->lock held and irqs disabled
1169 static void hrtick_start(struct rq
*rq
, u64 delay
)
1171 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1172 HRTIMER_MODE_REL_PINNED
, 0);
1175 static inline void init_hrtick(void)
1178 #endif /* CONFIG_SMP */
1180 static void init_rq_hrtick(struct rq
*rq
)
1183 rq
->hrtick_csd_pending
= 0;
1185 rq
->hrtick_csd
.flags
= 0;
1186 rq
->hrtick_csd
.func
= __hrtick_start
;
1187 rq
->hrtick_csd
.info
= rq
;
1190 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1191 rq
->hrtick_timer
.function
= hrtick
;
1193 #else /* CONFIG_SCHED_HRTICK */
1194 static inline void hrtick_clear(struct rq
*rq
)
1198 static inline void init_rq_hrtick(struct rq
*rq
)
1202 static inline void init_hrtick(void)
1205 #endif /* CONFIG_SCHED_HRTICK */
1208 * resched_task - mark a task 'to be rescheduled now'.
1210 * On UP this means the setting of the need_resched flag, on SMP it
1211 * might also involve a cross-CPU call to trigger the scheduler on
1216 #ifndef tsk_is_polling
1217 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1220 static void resched_task(struct task_struct
*p
)
1224 assert_spin_locked(&task_rq(p
)->lock
);
1226 if (test_tsk_need_resched(p
))
1229 set_tsk_need_resched(p
);
1232 if (cpu
== smp_processor_id())
1235 /* NEED_RESCHED must be visible before we test polling */
1237 if (!tsk_is_polling(p
))
1238 smp_send_reschedule(cpu
);
1241 static void resched_cpu(int cpu
)
1243 struct rq
*rq
= cpu_rq(cpu
);
1244 unsigned long flags
;
1246 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1248 resched_task(cpu_curr(cpu
));
1249 spin_unlock_irqrestore(&rq
->lock
, flags
);
1254 * When add_timer_on() enqueues a timer into the timer wheel of an
1255 * idle CPU then this timer might expire before the next timer event
1256 * which is scheduled to wake up that CPU. In case of a completely
1257 * idle system the next event might even be infinite time into the
1258 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1259 * leaves the inner idle loop so the newly added timer is taken into
1260 * account when the CPU goes back to idle and evaluates the timer
1261 * wheel for the next timer event.
1263 void wake_up_idle_cpu(int cpu
)
1265 struct rq
*rq
= cpu_rq(cpu
);
1267 if (cpu
== smp_processor_id())
1271 * This is safe, as this function is called with the timer
1272 * wheel base lock of (cpu) held. When the CPU is on the way
1273 * to idle and has not yet set rq->curr to idle then it will
1274 * be serialized on the timer wheel base lock and take the new
1275 * timer into account automatically.
1277 if (rq
->curr
!= rq
->idle
)
1281 * We can set TIF_RESCHED on the idle task of the other CPU
1282 * lockless. The worst case is that the other CPU runs the
1283 * idle task through an additional NOOP schedule()
1285 set_tsk_need_resched(rq
->idle
);
1287 /* NEED_RESCHED must be visible before we test polling */
1289 if (!tsk_is_polling(rq
->idle
))
1290 smp_send_reschedule(cpu
);
1292 #endif /* CONFIG_NO_HZ */
1294 static u64
sched_avg_period(void)
1296 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1299 static void sched_avg_update(struct rq
*rq
)
1301 s64 period
= sched_avg_period();
1303 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1304 rq
->age_stamp
+= period
;
1309 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1311 rq
->rt_avg
+= rt_delta
;
1312 sched_avg_update(rq
);
1315 #else /* !CONFIG_SMP */
1316 static void resched_task(struct task_struct
*p
)
1318 assert_spin_locked(&task_rq(p
)->lock
);
1319 set_tsk_need_resched(p
);
1322 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1325 #endif /* CONFIG_SMP */
1327 #if BITS_PER_LONG == 32
1328 # define WMULT_CONST (~0UL)
1330 # define WMULT_CONST (1UL << 32)
1333 #define WMULT_SHIFT 32
1336 * Shift right and round:
1338 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1341 * delta *= weight / lw
1343 static unsigned long
1344 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1345 struct load_weight
*lw
)
1349 if (!lw
->inv_weight
) {
1350 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1353 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1357 tmp
= (u64
)delta_exec
* weight
;
1359 * Check whether we'd overflow the 64-bit multiplication:
1361 if (unlikely(tmp
> WMULT_CONST
))
1362 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1365 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1367 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1370 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1376 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1383 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1384 * of tasks with abnormal "nice" values across CPUs the contribution that
1385 * each task makes to its run queue's load is weighted according to its
1386 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1387 * scaled version of the new time slice allocation that they receive on time
1391 #define WEIGHT_IDLEPRIO 3
1392 #define WMULT_IDLEPRIO 1431655765
1395 * Nice levels are multiplicative, with a gentle 10% change for every
1396 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1397 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1398 * that remained on nice 0.
1400 * The "10% effect" is relative and cumulative: from _any_ nice level,
1401 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1402 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1403 * If a task goes up by ~10% and another task goes down by ~10% then
1404 * the relative distance between them is ~25%.)
1406 static const int prio_to_weight
[40] = {
1407 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1408 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1409 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1410 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1411 /* 0 */ 1024, 820, 655, 526, 423,
1412 /* 5 */ 335, 272, 215, 172, 137,
1413 /* 10 */ 110, 87, 70, 56, 45,
1414 /* 15 */ 36, 29, 23, 18, 15,
1418 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1420 * In cases where the weight does not change often, we can use the
1421 * precalculated inverse to speed up arithmetics by turning divisions
1422 * into multiplications:
1424 static const u32 prio_to_wmult
[40] = {
1425 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1426 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1427 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1428 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1429 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1430 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1431 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1432 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1435 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1438 * runqueue iterator, to support SMP load-balancing between different
1439 * scheduling classes, without having to expose their internal data
1440 * structures to the load-balancing proper:
1442 struct rq_iterator
{
1444 struct task_struct
*(*start
)(void *);
1445 struct task_struct
*(*next
)(void *);
1449 static unsigned long
1450 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1451 unsigned long max_load_move
, struct sched_domain
*sd
,
1452 enum cpu_idle_type idle
, int *all_pinned
,
1453 int *this_best_prio
, struct rq_iterator
*iterator
);
1456 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1457 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1458 struct rq_iterator
*iterator
);
1461 /* Time spent by the tasks of the cpu accounting group executing in ... */
1462 enum cpuacct_stat_index
{
1463 CPUACCT_STAT_USER
, /* ... user mode */
1464 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1466 CPUACCT_STAT_NSTATS
,
1469 #ifdef CONFIG_CGROUP_CPUACCT
1470 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1471 static void cpuacct_update_stats(struct task_struct
*tsk
,
1472 enum cpuacct_stat_index idx
, cputime_t val
);
1474 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1475 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1476 enum cpuacct_stat_index idx
, cputime_t val
) {}
1479 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1481 update_load_add(&rq
->load
, load
);
1484 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1486 update_load_sub(&rq
->load
, load
);
1489 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1490 typedef int (*tg_visitor
)(struct task_group
*, void *);
1493 * Iterate the full tree, calling @down when first entering a node and @up when
1494 * leaving it for the final time.
1496 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1498 struct task_group
*parent
, *child
;
1502 parent
= &root_task_group
;
1504 ret
= (*down
)(parent
, data
);
1507 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1514 ret
= (*up
)(parent
, data
);
1519 parent
= parent
->parent
;
1528 static int tg_nop(struct task_group
*tg
, void *data
)
1535 static unsigned long source_load(int cpu
, int type
);
1536 static unsigned long target_load(int cpu
, int type
);
1537 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1539 static unsigned long cpu_avg_load_per_task(int cpu
)
1541 struct rq
*rq
= cpu_rq(cpu
);
1542 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1545 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1547 rq
->avg_load_per_task
= 0;
1549 return rq
->avg_load_per_task
;
1552 #ifdef CONFIG_FAIR_GROUP_SCHED
1554 struct update_shares_data
{
1555 unsigned long rq_weight
[NR_CPUS
];
1558 static DEFINE_PER_CPU(struct update_shares_data
, update_shares_data
);
1560 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1563 * Calculate and set the cpu's group shares.
1565 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1566 unsigned long sd_shares
,
1567 unsigned long sd_rq_weight
,
1568 struct update_shares_data
*usd
)
1570 unsigned long shares
, rq_weight
;
1573 rq_weight
= usd
->rq_weight
[cpu
];
1576 rq_weight
= NICE_0_LOAD
;
1580 * \Sum_j shares_j * rq_weight_i
1581 * shares_i = -----------------------------
1582 * \Sum_j rq_weight_j
1584 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1585 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1587 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1588 sysctl_sched_shares_thresh
) {
1589 struct rq
*rq
= cpu_rq(cpu
);
1590 unsigned long flags
;
1592 spin_lock_irqsave(&rq
->lock
, flags
);
1593 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1594 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1595 __set_se_shares(tg
->se
[cpu
], shares
);
1596 spin_unlock_irqrestore(&rq
->lock
, flags
);
1601 * Re-compute the task group their per cpu shares over the given domain.
1602 * This needs to be done in a bottom-up fashion because the rq weight of a
1603 * parent group depends on the shares of its child groups.
1605 static int tg_shares_up(struct task_group
*tg
, void *data
)
1607 unsigned long weight
, rq_weight
= 0, shares
= 0;
1608 struct update_shares_data
*usd
;
1609 struct sched_domain
*sd
= data
;
1610 unsigned long flags
;
1616 local_irq_save(flags
);
1617 usd
= &__get_cpu_var(update_shares_data
);
1619 for_each_cpu(i
, sched_domain_span(sd
)) {
1620 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1621 usd
->rq_weight
[i
] = weight
;
1624 * If there are currently no tasks on the cpu pretend there
1625 * is one of average load so that when a new task gets to
1626 * run here it will not get delayed by group starvation.
1629 weight
= NICE_0_LOAD
;
1631 rq_weight
+= weight
;
1632 shares
+= tg
->cfs_rq
[i
]->shares
;
1635 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1636 shares
= tg
->shares
;
1638 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1639 shares
= tg
->shares
;
1641 for_each_cpu(i
, sched_domain_span(sd
))
1642 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd
);
1644 local_irq_restore(flags
);
1650 * Compute the cpu's hierarchical load factor for each task group.
1651 * This needs to be done in a top-down fashion because the load of a child
1652 * group is a fraction of its parents load.
1654 static int tg_load_down(struct task_group
*tg
, void *data
)
1657 long cpu
= (long)data
;
1660 load
= cpu_rq(cpu
)->load
.weight
;
1662 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1663 load
*= tg
->cfs_rq
[cpu
]->shares
;
1664 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1667 tg
->cfs_rq
[cpu
]->h_load
= load
;
1672 static void update_shares(struct sched_domain
*sd
)
1677 if (root_task_group_empty())
1680 now
= cpu_clock(raw_smp_processor_id());
1681 elapsed
= now
- sd
->last_update
;
1683 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1684 sd
->last_update
= now
;
1685 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1689 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1691 if (root_task_group_empty())
1694 spin_unlock(&rq
->lock
);
1696 spin_lock(&rq
->lock
);
1699 static void update_h_load(long cpu
)
1701 if (root_task_group_empty())
1704 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1709 static inline void update_shares(struct sched_domain
*sd
)
1713 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1719 #ifdef CONFIG_PREEMPT
1722 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1723 * way at the expense of forcing extra atomic operations in all
1724 * invocations. This assures that the double_lock is acquired using the
1725 * same underlying policy as the spinlock_t on this architecture, which
1726 * reduces latency compared to the unfair variant below. However, it
1727 * also adds more overhead and therefore may reduce throughput.
1729 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1730 __releases(this_rq
->lock
)
1731 __acquires(busiest
->lock
)
1732 __acquires(this_rq
->lock
)
1734 spin_unlock(&this_rq
->lock
);
1735 double_rq_lock(this_rq
, busiest
);
1742 * Unfair double_lock_balance: Optimizes throughput at the expense of
1743 * latency by eliminating extra atomic operations when the locks are
1744 * already in proper order on entry. This favors lower cpu-ids and will
1745 * grant the double lock to lower cpus over higher ids under contention,
1746 * regardless of entry order into the function.
1748 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1749 __releases(this_rq
->lock
)
1750 __acquires(busiest
->lock
)
1751 __acquires(this_rq
->lock
)
1755 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1756 if (busiest
< this_rq
) {
1757 spin_unlock(&this_rq
->lock
);
1758 spin_lock(&busiest
->lock
);
1759 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1762 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1767 #endif /* CONFIG_PREEMPT */
1770 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1772 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1774 if (unlikely(!irqs_disabled())) {
1775 /* printk() doesn't work good under rq->lock */
1776 spin_unlock(&this_rq
->lock
);
1780 return _double_lock_balance(this_rq
, busiest
);
1783 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1784 __releases(busiest
->lock
)
1786 spin_unlock(&busiest
->lock
);
1787 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1791 #ifdef CONFIG_FAIR_GROUP_SCHED
1792 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1795 cfs_rq
->shares
= shares
;
1800 static void calc_load_account_active(struct rq
*this_rq
);
1802 #include "sched_stats.h"
1803 #include "sched_idletask.c"
1804 #include "sched_fair.c"
1805 #include "sched_rt.c"
1806 #ifdef CONFIG_SCHED_DEBUG
1807 # include "sched_debug.c"
1810 #define sched_class_highest (&rt_sched_class)
1811 #define for_each_class(class) \
1812 for (class = sched_class_highest; class; class = class->next)
1814 static void inc_nr_running(struct rq
*rq
)
1819 static void dec_nr_running(struct rq
*rq
)
1824 static void set_load_weight(struct task_struct
*p
)
1826 if (task_has_rt_policy(p
)) {
1827 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1828 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1833 * SCHED_IDLE tasks get minimal weight:
1835 if (p
->policy
== SCHED_IDLE
) {
1836 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1837 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1841 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1842 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1845 static void update_avg(u64
*avg
, u64 sample
)
1847 s64 diff
= sample
- *avg
;
1851 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1854 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1856 sched_info_queued(p
);
1857 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1861 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1864 if (p
->se
.last_wakeup
) {
1865 update_avg(&p
->se
.avg_overlap
,
1866 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1867 p
->se
.last_wakeup
= 0;
1869 update_avg(&p
->se
.avg_wakeup
,
1870 sysctl_sched_wakeup_granularity
);
1874 sched_info_dequeued(p
);
1875 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1880 * __normal_prio - return the priority that is based on the static prio
1882 static inline int __normal_prio(struct task_struct
*p
)
1884 return p
->static_prio
;
1888 * Calculate the expected normal priority: i.e. priority
1889 * without taking RT-inheritance into account. Might be
1890 * boosted by interactivity modifiers. Changes upon fork,
1891 * setprio syscalls, and whenever the interactivity
1892 * estimator recalculates.
1894 static inline int normal_prio(struct task_struct
*p
)
1898 if (task_has_rt_policy(p
))
1899 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1901 prio
= __normal_prio(p
);
1906 * Calculate the current priority, i.e. the priority
1907 * taken into account by the scheduler. This value might
1908 * be boosted by RT tasks, or might be boosted by
1909 * interactivity modifiers. Will be RT if the task got
1910 * RT-boosted. If not then it returns p->normal_prio.
1912 static int effective_prio(struct task_struct
*p
)
1914 p
->normal_prio
= normal_prio(p
);
1916 * If we are RT tasks or we were boosted to RT priority,
1917 * keep the priority unchanged. Otherwise, update priority
1918 * to the normal priority:
1920 if (!rt_prio(p
->prio
))
1921 return p
->normal_prio
;
1926 * activate_task - move a task to the runqueue.
1928 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1930 if (task_contributes_to_load(p
))
1931 rq
->nr_uninterruptible
--;
1933 enqueue_task(rq
, p
, wakeup
);
1938 * deactivate_task - remove a task from the runqueue.
1940 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1942 if (task_contributes_to_load(p
))
1943 rq
->nr_uninterruptible
++;
1945 dequeue_task(rq
, p
, sleep
);
1950 * task_curr - is this task currently executing on a CPU?
1951 * @p: the task in question.
1953 inline int task_curr(const struct task_struct
*p
)
1955 return cpu_curr(task_cpu(p
)) == p
;
1958 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1960 set_task_rq(p
, cpu
);
1963 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1964 * successfuly executed on another CPU. We must ensure that updates of
1965 * per-task data have been completed by this moment.
1968 task_thread_info(p
)->cpu
= cpu
;
1972 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1973 const struct sched_class
*prev_class
,
1974 int oldprio
, int running
)
1976 if (prev_class
!= p
->sched_class
) {
1977 if (prev_class
->switched_from
)
1978 prev_class
->switched_from(rq
, p
, running
);
1979 p
->sched_class
->switched_to(rq
, p
, running
);
1981 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1986 /* Used instead of source_load when we know the type == 0 */
1987 static unsigned long weighted_cpuload(const int cpu
)
1989 return cpu_rq(cpu
)->load
.weight
;
1993 * Is this task likely cache-hot:
1996 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2001 * Buddy candidates are cache hot:
2003 if (sched_feat(CACHE_HOT_BUDDY
) &&
2004 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2005 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2008 if (p
->sched_class
!= &fair_sched_class
)
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
;
2022 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2024 int old_cpu
= task_cpu(p
);
2025 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2026 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2027 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2030 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2032 trace_sched_migrate_task(p
, new_cpu
);
2034 #ifdef CONFIG_SCHEDSTATS
2035 if (p
->se
.wait_start
)
2036 p
->se
.wait_start
-= clock_offset
;
2037 if (p
->se
.sleep_start
)
2038 p
->se
.sleep_start
-= clock_offset
;
2039 if (p
->se
.block_start
)
2040 p
->se
.block_start
-= clock_offset
;
2042 if (old_cpu
!= new_cpu
) {
2043 p
->se
.nr_migrations
++;
2044 new_rq
->nr_migrations_in
++;
2045 #ifdef CONFIG_SCHEDSTATS
2046 if (task_hot(p
, old_rq
->clock
, NULL
))
2047 schedstat_inc(p
, se
.nr_forced2_migrations
);
2049 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2052 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2053 new_cfsrq
->min_vruntime
;
2055 __set_task_cpu(p
, new_cpu
);
2058 struct migration_req
{
2059 struct list_head list
;
2061 struct task_struct
*task
;
2064 struct completion done
;
2068 * The task's runqueue lock must be held.
2069 * Returns true if you have to wait for migration thread.
2072 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2074 struct rq
*rq
= task_rq(p
);
2077 * If the task is not on a runqueue (and not running), then
2078 * it is sufficient to simply update the task's cpu field.
2080 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2081 set_task_cpu(p
, dest_cpu
);
2085 init_completion(&req
->done
);
2087 req
->dest_cpu
= dest_cpu
;
2088 list_add(&req
->list
, &rq
->migration_queue
);
2094 * wait_task_context_switch - wait for a thread to complete at least one
2097 * @p must not be current.
2099 void wait_task_context_switch(struct task_struct
*p
)
2101 unsigned long nvcsw
, nivcsw
, flags
;
2109 * The runqueue is assigned before the actual context
2110 * switch. We need to take the runqueue lock.
2112 * We could check initially without the lock but it is
2113 * very likely that we need to take the lock in every
2116 rq
= task_rq_lock(p
, &flags
);
2117 running
= task_running(rq
, p
);
2118 task_rq_unlock(rq
, &flags
);
2120 if (likely(!running
))
2123 * The switch count is incremented before the actual
2124 * context switch. We thus wait for two switches to be
2125 * sure at least one completed.
2127 if ((p
->nvcsw
- nvcsw
) > 1)
2129 if ((p
->nivcsw
- nivcsw
) > 1)
2137 * wait_task_inactive - wait for a thread to unschedule.
2139 * If @match_state is nonzero, it's the @p->state value just checked and
2140 * not expected to change. If it changes, i.e. @p might have woken up,
2141 * then return zero. When we succeed in waiting for @p to be off its CPU,
2142 * we return a positive number (its total switch count). If a second call
2143 * a short while later returns the same number, the caller can be sure that
2144 * @p has remained unscheduled the whole time.
2146 * The caller must ensure that the task *will* unschedule sometime soon,
2147 * else this function might spin for a *long* time. This function can't
2148 * be called with interrupts off, or it may introduce deadlock with
2149 * smp_call_function() if an IPI is sent by the same process we are
2150 * waiting to become inactive.
2152 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2154 unsigned long flags
;
2161 * We do the initial early heuristics without holding
2162 * any task-queue locks at all. We'll only try to get
2163 * the runqueue lock when things look like they will
2169 * If the task is actively running on another CPU
2170 * still, just relax and busy-wait without holding
2173 * NOTE! Since we don't hold any locks, it's not
2174 * even sure that "rq" stays as the right runqueue!
2175 * But we don't care, since "task_running()" will
2176 * return false if the runqueue has changed and p
2177 * is actually now running somewhere else!
2179 while (task_running(rq
, p
)) {
2180 if (match_state
&& unlikely(p
->state
!= match_state
))
2186 * Ok, time to look more closely! We need the rq
2187 * lock now, to be *sure*. If we're wrong, we'll
2188 * just go back and repeat.
2190 rq
= task_rq_lock(p
, &flags
);
2191 trace_sched_wait_task(rq
, p
);
2192 running
= task_running(rq
, p
);
2193 on_rq
= p
->se
.on_rq
;
2195 if (!match_state
|| p
->state
== match_state
)
2196 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2197 task_rq_unlock(rq
, &flags
);
2200 * If it changed from the expected state, bail out now.
2202 if (unlikely(!ncsw
))
2206 * Was it really running after all now that we
2207 * checked with the proper locks actually held?
2209 * Oops. Go back and try again..
2211 if (unlikely(running
)) {
2217 * It's not enough that it's not actively running,
2218 * it must be off the runqueue _entirely_, and not
2221 * So if it was still runnable (but just not actively
2222 * running right now), it's preempted, and we should
2223 * yield - it could be a while.
2225 if (unlikely(on_rq
)) {
2226 schedule_timeout_uninterruptible(1);
2231 * Ahh, all good. It wasn't running, and it wasn't
2232 * runnable, which means that it will never become
2233 * running in the future either. We're all done!
2242 * kick_process - kick a running thread to enter/exit the kernel
2243 * @p: the to-be-kicked thread
2245 * Cause a process which is running on another CPU to enter
2246 * kernel-mode, without any delay. (to get signals handled.)
2248 * NOTE: this function doesnt have to take the runqueue lock,
2249 * because all it wants to ensure is that the remote task enters
2250 * the kernel. If the IPI races and the task has been migrated
2251 * to another CPU then no harm is done and the purpose has been
2254 void kick_process(struct task_struct
*p
)
2260 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2261 smp_send_reschedule(cpu
);
2264 EXPORT_SYMBOL_GPL(kick_process
);
2267 * Return a low guess at the load of a migration-source cpu weighted
2268 * according to the scheduling class and "nice" value.
2270 * We want to under-estimate the load of migration sources, to
2271 * balance conservatively.
2273 static unsigned long source_load(int cpu
, int type
)
2275 struct rq
*rq
= cpu_rq(cpu
);
2276 unsigned long total
= weighted_cpuload(cpu
);
2278 if (type
== 0 || !sched_feat(LB_BIAS
))
2281 return min(rq
->cpu_load
[type
-1], total
);
2285 * Return a high guess at the load of a migration-target cpu weighted
2286 * according to the scheduling class and "nice" value.
2288 static unsigned long target_load(int cpu
, int type
)
2290 struct rq
*rq
= cpu_rq(cpu
);
2291 unsigned long total
= weighted_cpuload(cpu
);
2293 if (type
== 0 || !sched_feat(LB_BIAS
))
2296 return max(rq
->cpu_load
[type
-1], total
);
2300 * find_idlest_group finds and returns the least busy CPU group within the
2303 static struct sched_group
*
2304 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2306 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2307 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2308 int load_idx
= sd
->forkexec_idx
;
2309 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2312 unsigned long load
, avg_load
;
2316 /* Skip over this group if it has no CPUs allowed */
2317 if (!cpumask_intersects(sched_group_cpus(group
),
2321 local_group
= cpumask_test_cpu(this_cpu
,
2322 sched_group_cpus(group
));
2324 /* Tally up the load of all CPUs in the group */
2327 for_each_cpu(i
, sched_group_cpus(group
)) {
2328 /* Bias balancing toward cpus of our domain */
2330 load
= source_load(i
, load_idx
);
2332 load
= target_load(i
, load_idx
);
2337 /* Adjust by relative CPU power of the group */
2338 avg_load
= sg_div_cpu_power(group
,
2339 avg_load
* SCHED_LOAD_SCALE
);
2342 this_load
= avg_load
;
2344 } else if (avg_load
< min_load
) {
2345 min_load
= avg_load
;
2348 } while (group
= group
->next
, group
!= sd
->groups
);
2350 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2356 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2359 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2361 unsigned long load
, min_load
= ULONG_MAX
;
2365 /* Traverse only the allowed CPUs */
2366 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2367 load
= weighted_cpuload(i
);
2369 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2379 * sched_balance_self: balance the current task (running on cpu) in domains
2380 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2383 * Balance, ie. select the least loaded group.
2385 * Returns the target CPU number, or the same CPU if no balancing is needed.
2387 * preempt must be disabled.
2389 static int sched_balance_self(int cpu
, int flag
)
2391 struct task_struct
*t
= current
;
2392 struct sched_domain
*tmp
, *sd
= NULL
;
2394 for_each_domain(cpu
, tmp
) {
2396 * If power savings logic is enabled for a domain, stop there.
2398 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2400 if (tmp
->flags
& flag
)
2408 struct sched_group
*group
;
2409 int new_cpu
, weight
;
2411 if (!(sd
->flags
& flag
)) {
2416 group
= find_idlest_group(sd
, t
, cpu
);
2422 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2423 if (new_cpu
== -1 || new_cpu
== cpu
) {
2424 /* Now try balancing at a lower domain level of cpu */
2429 /* Now try balancing at a lower domain level of new_cpu */
2431 weight
= cpumask_weight(sched_domain_span(sd
));
2433 for_each_domain(cpu
, tmp
) {
2434 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2436 if (tmp
->flags
& flag
)
2439 /* while loop will break here if sd == NULL */
2445 #endif /* CONFIG_SMP */
2448 * task_oncpu_function_call - call a function on the cpu on which a task runs
2449 * @p: the task to evaluate
2450 * @func: the function to be called
2451 * @info: the function call argument
2453 * Calls the function @func when the task is currently running. This might
2454 * be on the current CPU, which just calls the function directly
2456 void task_oncpu_function_call(struct task_struct
*p
,
2457 void (*func
) (void *info
), void *info
)
2464 smp_call_function_single(cpu
, func
, info
, 1);
2469 * try_to_wake_up - wake up a thread
2470 * @p: the to-be-woken-up thread
2471 * @state: the mask of task states that can be woken
2472 * @sync: do a synchronous wakeup?
2474 * Put it on the run-queue if it's not already there. The "current"
2475 * thread is always on the run-queue (except when the actual
2476 * re-schedule is in progress), and as such you're allowed to do
2477 * the simpler "current->state = TASK_RUNNING" to mark yourself
2478 * runnable without the overhead of this.
2480 * returns failure only if the task is already active.
2482 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2484 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2485 unsigned long flags
;
2489 if (!sched_feat(SYNC_WAKEUPS
))
2493 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2494 struct sched_domain
*sd
;
2496 this_cpu
= raw_smp_processor_id();
2499 for_each_domain(this_cpu
, sd
) {
2500 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2509 rq
= task_rq_lock(p
, &flags
);
2510 update_rq_clock(rq
);
2511 old_state
= p
->state
;
2512 if (!(old_state
& state
))
2520 this_cpu
= smp_processor_id();
2523 if (unlikely(task_running(rq
, p
)))
2526 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2527 if (cpu
!= orig_cpu
) {
2528 set_task_cpu(p
, cpu
);
2529 task_rq_unlock(rq
, &flags
);
2530 /* might preempt at this point */
2531 rq
= task_rq_lock(p
, &flags
);
2532 old_state
= p
->state
;
2533 if (!(old_state
& state
))
2538 this_cpu
= smp_processor_id();
2542 #ifdef CONFIG_SCHEDSTATS
2543 schedstat_inc(rq
, ttwu_count
);
2544 if (cpu
== this_cpu
)
2545 schedstat_inc(rq
, ttwu_local
);
2547 struct sched_domain
*sd
;
2548 for_each_domain(this_cpu
, sd
) {
2549 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2550 schedstat_inc(sd
, ttwu_wake_remote
);
2555 #endif /* CONFIG_SCHEDSTATS */
2558 #endif /* CONFIG_SMP */
2559 schedstat_inc(p
, se
.nr_wakeups
);
2561 schedstat_inc(p
, se
.nr_wakeups_sync
);
2562 if (orig_cpu
!= cpu
)
2563 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2564 if (cpu
== this_cpu
)
2565 schedstat_inc(p
, se
.nr_wakeups_local
);
2567 schedstat_inc(p
, se
.nr_wakeups_remote
);
2568 activate_task(rq
, p
, 1);
2572 * Only attribute actual wakeups done by this task.
2574 if (!in_interrupt()) {
2575 struct sched_entity
*se
= ¤t
->se
;
2576 u64 sample
= se
->sum_exec_runtime
;
2578 if (se
->last_wakeup
)
2579 sample
-= se
->last_wakeup
;
2581 sample
-= se
->start_runtime
;
2582 update_avg(&se
->avg_wakeup
, sample
);
2584 se
->last_wakeup
= se
->sum_exec_runtime
;
2588 trace_sched_wakeup(rq
, p
, success
);
2589 check_preempt_curr(rq
, p
, sync
);
2591 p
->state
= TASK_RUNNING
;
2593 if (p
->sched_class
->task_wake_up
)
2594 p
->sched_class
->task_wake_up(rq
, p
);
2597 task_rq_unlock(rq
, &flags
);
2603 * wake_up_process - Wake up a specific process
2604 * @p: The process to be woken up.
2606 * Attempt to wake up the nominated process and move it to the set of runnable
2607 * processes. Returns 1 if the process was woken up, 0 if it was already
2610 * It may be assumed that this function implies a write memory barrier before
2611 * changing the task state if and only if any tasks are woken up.
2613 int wake_up_process(struct task_struct
*p
)
2615 return try_to_wake_up(p
, TASK_ALL
, 0);
2617 EXPORT_SYMBOL(wake_up_process
);
2619 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2621 return try_to_wake_up(p
, state
, 0);
2625 * Perform scheduler related setup for a newly forked process p.
2626 * p is forked by current.
2628 * __sched_fork() is basic setup used by init_idle() too:
2630 static void __sched_fork(struct task_struct
*p
)
2632 p
->se
.exec_start
= 0;
2633 p
->se
.sum_exec_runtime
= 0;
2634 p
->se
.prev_sum_exec_runtime
= 0;
2635 p
->se
.nr_migrations
= 0;
2636 p
->se
.last_wakeup
= 0;
2637 p
->se
.avg_overlap
= 0;
2638 p
->se
.start_runtime
= 0;
2639 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2641 #ifdef CONFIG_SCHEDSTATS
2642 p
->se
.wait_start
= 0;
2644 p
->se
.wait_count
= 0;
2647 p
->se
.sleep_start
= 0;
2648 p
->se
.sleep_max
= 0;
2649 p
->se
.sum_sleep_runtime
= 0;
2651 p
->se
.block_start
= 0;
2652 p
->se
.block_max
= 0;
2654 p
->se
.slice_max
= 0;
2656 p
->se
.nr_migrations_cold
= 0;
2657 p
->se
.nr_failed_migrations_affine
= 0;
2658 p
->se
.nr_failed_migrations_running
= 0;
2659 p
->se
.nr_failed_migrations_hot
= 0;
2660 p
->se
.nr_forced_migrations
= 0;
2661 p
->se
.nr_forced2_migrations
= 0;
2663 p
->se
.nr_wakeups
= 0;
2664 p
->se
.nr_wakeups_sync
= 0;
2665 p
->se
.nr_wakeups_migrate
= 0;
2666 p
->se
.nr_wakeups_local
= 0;
2667 p
->se
.nr_wakeups_remote
= 0;
2668 p
->se
.nr_wakeups_affine
= 0;
2669 p
->se
.nr_wakeups_affine_attempts
= 0;
2670 p
->se
.nr_wakeups_passive
= 0;
2671 p
->se
.nr_wakeups_idle
= 0;
2675 INIT_LIST_HEAD(&p
->rt
.run_list
);
2677 INIT_LIST_HEAD(&p
->se
.group_node
);
2679 #ifdef CONFIG_PREEMPT_NOTIFIERS
2680 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2684 * We mark the process as running here, but have not actually
2685 * inserted it onto the runqueue yet. This guarantees that
2686 * nobody will actually run it, and a signal or other external
2687 * event cannot wake it up and insert it on the runqueue either.
2689 p
->state
= TASK_RUNNING
;
2693 * fork()/clone()-time setup:
2695 void sched_fork(struct task_struct
*p
, int clone_flags
)
2697 int cpu
= get_cpu();
2702 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2704 set_task_cpu(p
, cpu
);
2707 * Make sure we do not leak PI boosting priority to the child.
2709 p
->prio
= current
->normal_prio
;
2712 * Revert to default priority/policy on fork if requested.
2714 if (unlikely(p
->sched_reset_on_fork
)) {
2715 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
)
2716 p
->policy
= SCHED_NORMAL
;
2718 if (p
->normal_prio
< DEFAULT_PRIO
)
2719 p
->prio
= DEFAULT_PRIO
;
2721 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2722 p
->static_prio
= NICE_TO_PRIO(0);
2727 * We don't need the reset flag anymore after the fork. It has
2728 * fulfilled its duty:
2730 p
->sched_reset_on_fork
= 0;
2733 if (!rt_prio(p
->prio
))
2734 p
->sched_class
= &fair_sched_class
;
2736 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2737 if (likely(sched_info_on()))
2738 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2740 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2743 #ifdef CONFIG_PREEMPT
2744 /* Want to start with kernel preemption disabled. */
2745 task_thread_info(p
)->preempt_count
= 1;
2747 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2753 * wake_up_new_task - wake up a newly created task for the first time.
2755 * This function will do some initial scheduler statistics housekeeping
2756 * that must be done for every newly created context, then puts the task
2757 * on the runqueue and wakes it.
2759 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2761 unsigned long flags
;
2764 rq
= task_rq_lock(p
, &flags
);
2765 BUG_ON(p
->state
!= TASK_RUNNING
);
2766 update_rq_clock(rq
);
2768 p
->prio
= effective_prio(p
);
2770 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2771 activate_task(rq
, p
, 0);
2774 * Let the scheduling class do new task startup
2775 * management (if any):
2777 p
->sched_class
->task_new(rq
, p
);
2780 trace_sched_wakeup_new(rq
, p
, 1);
2781 check_preempt_curr(rq
, p
, 0);
2783 if (p
->sched_class
->task_wake_up
)
2784 p
->sched_class
->task_wake_up(rq
, p
);
2786 task_rq_unlock(rq
, &flags
);
2789 #ifdef CONFIG_PREEMPT_NOTIFIERS
2792 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2793 * @notifier: notifier struct to register
2795 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2797 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2799 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2802 * preempt_notifier_unregister - no longer interested in preemption notifications
2803 * @notifier: notifier struct to unregister
2805 * This is safe to call from within a preemption notifier.
2807 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2809 hlist_del(¬ifier
->link
);
2811 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2813 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2815 struct preempt_notifier
*notifier
;
2816 struct hlist_node
*node
;
2818 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2819 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2823 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2824 struct task_struct
*next
)
2826 struct preempt_notifier
*notifier
;
2827 struct hlist_node
*node
;
2829 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2830 notifier
->ops
->sched_out(notifier
, next
);
2833 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2835 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2840 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2841 struct task_struct
*next
)
2845 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2848 * prepare_task_switch - prepare to switch tasks
2849 * @rq: the runqueue preparing to switch
2850 * @prev: the current task that is being switched out
2851 * @next: the task we are going to switch to.
2853 * This is called with the rq lock held and interrupts off. It must
2854 * be paired with a subsequent finish_task_switch after the context
2857 * prepare_task_switch sets up locking and calls architecture specific
2861 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2862 struct task_struct
*next
)
2864 fire_sched_out_preempt_notifiers(prev
, next
);
2865 prepare_lock_switch(rq
, next
);
2866 prepare_arch_switch(next
);
2870 * finish_task_switch - clean up after a task-switch
2871 * @rq: runqueue associated with task-switch
2872 * @prev: the thread we just switched away from.
2874 * finish_task_switch must be called after the context switch, paired
2875 * with a prepare_task_switch call before the context switch.
2876 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2877 * and do any other architecture-specific cleanup actions.
2879 * Note that we may have delayed dropping an mm in context_switch(). If
2880 * so, we finish that here outside of the runqueue lock. (Doing it
2881 * with the lock held can cause deadlocks; see schedule() for
2884 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2885 __releases(rq
->lock
)
2887 struct mm_struct
*mm
= rq
->prev_mm
;
2893 * A task struct has one reference for the use as "current".
2894 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2895 * schedule one last time. The schedule call will never return, and
2896 * the scheduled task must drop that reference.
2897 * The test for TASK_DEAD must occur while the runqueue locks are
2898 * still held, otherwise prev could be scheduled on another cpu, die
2899 * there before we look at prev->state, and then the reference would
2901 * Manfred Spraul <manfred@colorfullife.com>
2903 prev_state
= prev
->state
;
2904 finish_arch_switch(prev
);
2905 perf_counter_task_sched_in(current
, cpu_of(rq
));
2906 finish_lock_switch(rq
, prev
);
2908 fire_sched_in_preempt_notifiers(current
);
2911 if (unlikely(prev_state
== TASK_DEAD
)) {
2913 * Remove function-return probe instances associated with this
2914 * task and put them back on the free list.
2916 kprobe_flush_task(prev
);
2917 put_task_struct(prev
);
2923 /* assumes rq->lock is held */
2924 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2926 if (prev
->sched_class
->pre_schedule
)
2927 prev
->sched_class
->pre_schedule(rq
, prev
);
2930 /* rq->lock is NOT held, but preemption is disabled */
2931 static inline void post_schedule(struct rq
*rq
)
2933 if (rq
->post_schedule
) {
2934 unsigned long flags
;
2936 spin_lock_irqsave(&rq
->lock
, flags
);
2937 if (rq
->curr
->sched_class
->post_schedule
)
2938 rq
->curr
->sched_class
->post_schedule(rq
);
2939 spin_unlock_irqrestore(&rq
->lock
, flags
);
2941 rq
->post_schedule
= 0;
2947 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2951 static inline void post_schedule(struct rq
*rq
)
2958 * schedule_tail - first thing a freshly forked thread must call.
2959 * @prev: the thread we just switched away from.
2961 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2962 __releases(rq
->lock
)
2964 struct rq
*rq
= this_rq();
2966 finish_task_switch(rq
, prev
);
2969 * FIXME: do we need to worry about rq being invalidated by the
2974 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2975 /* In this case, finish_task_switch does not reenable preemption */
2978 if (current
->set_child_tid
)
2979 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2983 * context_switch - switch to the new MM and the new
2984 * thread's register state.
2987 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2988 struct task_struct
*next
)
2990 struct mm_struct
*mm
, *oldmm
;
2992 prepare_task_switch(rq
, prev
, next
);
2993 trace_sched_switch(rq
, prev
, next
);
2995 oldmm
= prev
->active_mm
;
2997 * For paravirt, this is coupled with an exit in switch_to to
2998 * combine the page table reload and the switch backend into
3001 arch_start_context_switch(prev
);
3003 if (unlikely(!mm
)) {
3004 next
->active_mm
= oldmm
;
3005 atomic_inc(&oldmm
->mm_count
);
3006 enter_lazy_tlb(oldmm
, next
);
3008 switch_mm(oldmm
, mm
, next
);
3010 if (unlikely(!prev
->mm
)) {
3011 prev
->active_mm
= NULL
;
3012 rq
->prev_mm
= oldmm
;
3015 * Since the runqueue lock will be released by the next
3016 * task (which is an invalid locking op but in the case
3017 * of the scheduler it's an obvious special-case), so we
3018 * do an early lockdep release here:
3020 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3021 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3024 /* Here we just switch the register state and the stack. */
3025 switch_to(prev
, next
, prev
);
3029 * this_rq must be evaluated again because prev may have moved
3030 * CPUs since it called schedule(), thus the 'rq' on its stack
3031 * frame will be invalid.
3033 finish_task_switch(this_rq(), prev
);
3037 * nr_running, nr_uninterruptible and nr_context_switches:
3039 * externally visible scheduler statistics: current number of runnable
3040 * threads, current number of uninterruptible-sleeping threads, total
3041 * number of context switches performed since bootup.
3043 unsigned long nr_running(void)
3045 unsigned long i
, sum
= 0;
3047 for_each_online_cpu(i
)
3048 sum
+= cpu_rq(i
)->nr_running
;
3053 unsigned long nr_uninterruptible(void)
3055 unsigned long i
, sum
= 0;
3057 for_each_possible_cpu(i
)
3058 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3061 * Since we read the counters lockless, it might be slightly
3062 * inaccurate. Do not allow it to go below zero though:
3064 if (unlikely((long)sum
< 0))
3070 unsigned long long nr_context_switches(void)
3073 unsigned long long sum
= 0;
3075 for_each_possible_cpu(i
)
3076 sum
+= cpu_rq(i
)->nr_switches
;
3081 unsigned long nr_iowait(void)
3083 unsigned long i
, sum
= 0;
3085 for_each_possible_cpu(i
)
3086 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3091 /* Variables and functions for calc_load */
3092 static atomic_long_t calc_load_tasks
;
3093 static unsigned long calc_load_update
;
3094 unsigned long avenrun
[3];
3095 EXPORT_SYMBOL(avenrun
);
3098 * get_avenrun - get the load average array
3099 * @loads: pointer to dest load array
3100 * @offset: offset to add
3101 * @shift: shift count to shift the result left
3103 * These values are estimates at best, so no need for locking.
3105 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3107 loads
[0] = (avenrun
[0] + offset
) << shift
;
3108 loads
[1] = (avenrun
[1] + offset
) << shift
;
3109 loads
[2] = (avenrun
[2] + offset
) << shift
;
3112 static unsigned long
3113 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3116 load
+= active
* (FIXED_1
- exp
);
3117 return load
>> FSHIFT
;
3121 * calc_load - update the avenrun load estimates 10 ticks after the
3122 * CPUs have updated calc_load_tasks.
3124 void calc_global_load(void)
3126 unsigned long upd
= calc_load_update
+ 10;
3129 if (time_before(jiffies
, upd
))
3132 active
= atomic_long_read(&calc_load_tasks
);
3133 active
= active
> 0 ? active
* FIXED_1
: 0;
3135 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3136 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3137 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3139 calc_load_update
+= LOAD_FREQ
;
3143 * Either called from update_cpu_load() or from a cpu going idle
3145 static void calc_load_account_active(struct rq
*this_rq
)
3147 long nr_active
, delta
;
3149 nr_active
= this_rq
->nr_running
;
3150 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3152 if (nr_active
!= this_rq
->calc_load_active
) {
3153 delta
= nr_active
- this_rq
->calc_load_active
;
3154 this_rq
->calc_load_active
= nr_active
;
3155 atomic_long_add(delta
, &calc_load_tasks
);
3160 * Externally visible per-cpu scheduler statistics:
3161 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3163 u64
cpu_nr_migrations(int cpu
)
3165 return cpu_rq(cpu
)->nr_migrations_in
;
3169 * Update rq->cpu_load[] statistics. This function is usually called every
3170 * scheduler tick (TICK_NSEC).
3172 static void update_cpu_load(struct rq
*this_rq
)
3174 unsigned long this_load
= this_rq
->load
.weight
;
3177 this_rq
->nr_load_updates
++;
3179 /* Update our load: */
3180 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3181 unsigned long old_load
, new_load
;
3183 /* scale is effectively 1 << i now, and >> i divides by scale */
3185 old_load
= this_rq
->cpu_load
[i
];
3186 new_load
= this_load
;
3188 * Round up the averaging division if load is increasing. This
3189 * prevents us from getting stuck on 9 if the load is 10, for
3192 if (new_load
> old_load
)
3193 new_load
+= scale
-1;
3194 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3197 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3198 this_rq
->calc_load_update
+= LOAD_FREQ
;
3199 calc_load_account_active(this_rq
);
3206 * double_rq_lock - safely lock two runqueues
3208 * Note this does not disable interrupts like task_rq_lock,
3209 * you need to do so manually before calling.
3211 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3212 __acquires(rq1
->lock
)
3213 __acquires(rq2
->lock
)
3215 BUG_ON(!irqs_disabled());
3217 spin_lock(&rq1
->lock
);
3218 __acquire(rq2
->lock
); /* Fake it out ;) */
3221 spin_lock(&rq1
->lock
);
3222 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3224 spin_lock(&rq2
->lock
);
3225 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3228 update_rq_clock(rq1
);
3229 update_rq_clock(rq2
);
3233 * double_rq_unlock - safely unlock two runqueues
3235 * Note this does not restore interrupts like task_rq_unlock,
3236 * you need to do so manually after calling.
3238 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3239 __releases(rq1
->lock
)
3240 __releases(rq2
->lock
)
3242 spin_unlock(&rq1
->lock
);
3244 spin_unlock(&rq2
->lock
);
3246 __release(rq2
->lock
);
3250 * If dest_cpu is allowed for this process, migrate the task to it.
3251 * This is accomplished by forcing the cpu_allowed mask to only
3252 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3253 * the cpu_allowed mask is restored.
3255 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3257 struct migration_req req
;
3258 unsigned long flags
;
3261 rq
= task_rq_lock(p
, &flags
);
3262 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3263 || unlikely(!cpu_active(dest_cpu
)))
3266 /* force the process onto the specified CPU */
3267 if (migrate_task(p
, dest_cpu
, &req
)) {
3268 /* Need to wait for migration thread (might exit: take ref). */
3269 struct task_struct
*mt
= rq
->migration_thread
;
3271 get_task_struct(mt
);
3272 task_rq_unlock(rq
, &flags
);
3273 wake_up_process(mt
);
3274 put_task_struct(mt
);
3275 wait_for_completion(&req
.done
);
3280 task_rq_unlock(rq
, &flags
);
3284 * sched_exec - execve() is a valuable balancing opportunity, because at
3285 * this point the task has the smallest effective memory and cache footprint.
3287 void sched_exec(void)
3289 int new_cpu
, this_cpu
= get_cpu();
3290 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3292 if (new_cpu
!= this_cpu
)
3293 sched_migrate_task(current
, new_cpu
);
3297 * pull_task - move a task from a remote runqueue to the local runqueue.
3298 * Both runqueues must be locked.
3300 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3301 struct rq
*this_rq
, int this_cpu
)
3303 deactivate_task(src_rq
, p
, 0);
3304 set_task_cpu(p
, this_cpu
);
3305 activate_task(this_rq
, p
, 0);
3307 * Note that idle threads have a prio of MAX_PRIO, for this test
3308 * to be always true for them.
3310 check_preempt_curr(this_rq
, p
, 0);
3314 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3317 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3318 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3321 int tsk_cache_hot
= 0;
3323 * We do not migrate tasks that are:
3324 * 1) running (obviously), or
3325 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3326 * 3) are cache-hot on their current CPU.
3328 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3329 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3334 if (task_running(rq
, p
)) {
3335 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3340 * Aggressive migration if:
3341 * 1) task is cache cold, or
3342 * 2) too many balance attempts have failed.
3345 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3346 if (!tsk_cache_hot
||
3347 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3348 #ifdef CONFIG_SCHEDSTATS
3349 if (tsk_cache_hot
) {
3350 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3351 schedstat_inc(p
, se
.nr_forced_migrations
);
3357 if (tsk_cache_hot
) {
3358 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3364 static unsigned long
3365 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3366 unsigned long max_load_move
, struct sched_domain
*sd
,
3367 enum cpu_idle_type idle
, int *all_pinned
,
3368 int *this_best_prio
, struct rq_iterator
*iterator
)
3370 int loops
= 0, pulled
= 0, pinned
= 0;
3371 struct task_struct
*p
;
3372 long rem_load_move
= max_load_move
;
3374 if (max_load_move
== 0)
3380 * Start the load-balancing iterator:
3382 p
= iterator
->start(iterator
->arg
);
3384 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3387 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3388 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3389 p
= iterator
->next(iterator
->arg
);
3393 pull_task(busiest
, p
, this_rq
, this_cpu
);
3395 rem_load_move
-= p
->se
.load
.weight
;
3397 #ifdef CONFIG_PREEMPT
3399 * NEWIDLE balancing is a source of latency, so preemptible kernels
3400 * will stop after the first task is pulled to minimize the critical
3403 if (idle
== CPU_NEWLY_IDLE
)
3408 * We only want to steal up to the prescribed amount of weighted load.
3410 if (rem_load_move
> 0) {
3411 if (p
->prio
< *this_best_prio
)
3412 *this_best_prio
= p
->prio
;
3413 p
= iterator
->next(iterator
->arg
);
3418 * Right now, this is one of only two places pull_task() is called,
3419 * so we can safely collect pull_task() stats here rather than
3420 * inside pull_task().
3422 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3425 *all_pinned
= pinned
;
3427 return max_load_move
- rem_load_move
;
3431 * move_tasks tries to move up to max_load_move weighted load from busiest to
3432 * this_rq, as part of a balancing operation within domain "sd".
3433 * Returns 1 if successful and 0 otherwise.
3435 * Called with both runqueues locked.
3437 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3438 unsigned long max_load_move
,
3439 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3442 const struct sched_class
*class = sched_class_highest
;
3443 unsigned long total_load_moved
= 0;
3444 int this_best_prio
= this_rq
->curr
->prio
;
3448 class->load_balance(this_rq
, this_cpu
, busiest
,
3449 max_load_move
- total_load_moved
,
3450 sd
, idle
, all_pinned
, &this_best_prio
);
3451 class = class->next
;
3453 #ifdef CONFIG_PREEMPT
3455 * NEWIDLE balancing is a source of latency, so preemptible
3456 * kernels will stop after the first task is pulled to minimize
3457 * the critical section.
3459 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3462 } while (class && max_load_move
> total_load_moved
);
3464 return total_load_moved
> 0;
3468 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3469 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3470 struct rq_iterator
*iterator
)
3472 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3476 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3477 pull_task(busiest
, p
, this_rq
, this_cpu
);
3479 * Right now, this is only the second place pull_task()
3480 * is called, so we can safely collect pull_task()
3481 * stats here rather than inside pull_task().
3483 schedstat_inc(sd
, lb_gained
[idle
]);
3487 p
= iterator
->next(iterator
->arg
);
3494 * move_one_task tries to move exactly one task from busiest to this_rq, as
3495 * part of active balancing operations within "domain".
3496 * Returns 1 if successful and 0 otherwise.
3498 * Called with both runqueues locked.
3500 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3501 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3503 const struct sched_class
*class;
3505 for_each_class(class) {
3506 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3512 /********** Helpers for find_busiest_group ************************/
3514 * sd_lb_stats - Structure to store the statistics of a sched_domain
3515 * during load balancing.
3517 struct sd_lb_stats
{
3518 struct sched_group
*busiest
; /* Busiest group in this sd */
3519 struct sched_group
*this; /* Local group in this sd */
3520 unsigned long total_load
; /* Total load of all groups in sd */
3521 unsigned long total_pwr
; /* Total power of all groups in sd */
3522 unsigned long avg_load
; /* Average load across all groups in sd */
3524 /** Statistics of this group */
3525 unsigned long this_load
;
3526 unsigned long this_load_per_task
;
3527 unsigned long this_nr_running
;
3529 /* Statistics of the busiest group */
3530 unsigned long max_load
;
3531 unsigned long busiest_load_per_task
;
3532 unsigned long busiest_nr_running
;
3534 int group_imb
; /* Is there imbalance in this sd */
3535 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3536 int power_savings_balance
; /* Is powersave balance needed for this sd */
3537 struct sched_group
*group_min
; /* Least loaded group in sd */
3538 struct sched_group
*group_leader
; /* Group which relieves group_min */
3539 unsigned long min_load_per_task
; /* load_per_task in group_min */
3540 unsigned long leader_nr_running
; /* Nr running of group_leader */
3541 unsigned long min_nr_running
; /* Nr running of group_min */
3546 * sg_lb_stats - stats of a sched_group required for load_balancing
3548 struct sg_lb_stats
{
3549 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3550 unsigned long group_load
; /* Total load over the CPUs of the group */
3551 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3552 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3553 unsigned long group_capacity
;
3554 int group_imb
; /* Is there an imbalance in the group ? */
3558 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3559 * @group: The group whose first cpu is to be returned.
3561 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3563 return cpumask_first(sched_group_cpus(group
));
3567 * get_sd_load_idx - Obtain the load index for a given sched domain.
3568 * @sd: The sched_domain whose load_idx is to be obtained.
3569 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3571 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3572 enum cpu_idle_type idle
)
3578 load_idx
= sd
->busy_idx
;
3581 case CPU_NEWLY_IDLE
:
3582 load_idx
= sd
->newidle_idx
;
3585 load_idx
= sd
->idle_idx
;
3593 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3595 * init_sd_power_savings_stats - Initialize power savings statistics for
3596 * the given sched_domain, during load balancing.
3598 * @sd: Sched domain whose power-savings statistics are to be initialized.
3599 * @sds: Variable containing the statistics for sd.
3600 * @idle: Idle status of the CPU at which we're performing load-balancing.
3602 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3603 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3606 * Busy processors will not participate in power savings
3609 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3610 sds
->power_savings_balance
= 0;
3612 sds
->power_savings_balance
= 1;
3613 sds
->min_nr_running
= ULONG_MAX
;
3614 sds
->leader_nr_running
= 0;
3619 * update_sd_power_savings_stats - Update the power saving stats for a
3620 * sched_domain while performing load balancing.
3622 * @group: sched_group belonging to the sched_domain under consideration.
3623 * @sds: Variable containing the statistics of the sched_domain
3624 * @local_group: Does group contain the CPU for which we're performing
3626 * @sgs: Variable containing the statistics of the group.
3628 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3629 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3632 if (!sds
->power_savings_balance
)
3636 * If the local group is idle or completely loaded
3637 * no need to do power savings balance at this domain
3639 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3640 !sds
->this_nr_running
))
3641 sds
->power_savings_balance
= 0;
3644 * If a group is already running at full capacity or idle,
3645 * don't include that group in power savings calculations
3647 if (!sds
->power_savings_balance
||
3648 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3649 !sgs
->sum_nr_running
)
3653 * Calculate the group which has the least non-idle load.
3654 * This is the group from where we need to pick up the load
3657 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3658 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3659 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3660 sds
->group_min
= group
;
3661 sds
->min_nr_running
= sgs
->sum_nr_running
;
3662 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3663 sgs
->sum_nr_running
;
3667 * Calculate the group which is almost near its
3668 * capacity but still has some space to pick up some load
3669 * from other group and save more power
3671 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3674 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3675 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3676 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3677 sds
->group_leader
= group
;
3678 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3683 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3684 * @sds: Variable containing the statistics of the sched_domain
3685 * under consideration.
3686 * @this_cpu: Cpu at which we're currently performing load-balancing.
3687 * @imbalance: Variable to store the imbalance.
3690 * Check if we have potential to perform some power-savings balance.
3691 * If yes, set the busiest group to be the least loaded group in the
3692 * sched_domain, so that it's CPUs can be put to idle.
3694 * Returns 1 if there is potential to perform power-savings balance.
3697 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3698 int this_cpu
, unsigned long *imbalance
)
3700 if (!sds
->power_savings_balance
)
3703 if (sds
->this != sds
->group_leader
||
3704 sds
->group_leader
== sds
->group_min
)
3707 *imbalance
= sds
->min_load_per_task
;
3708 sds
->busiest
= sds
->group_min
;
3710 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3711 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3712 group_first_cpu(sds
->group_leader
);
3718 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3719 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3720 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3725 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3726 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3731 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3732 int this_cpu
, unsigned long *imbalance
)
3736 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3738 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3740 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3741 unsigned long smt_gain
= sd
->smt_gain
;
3748 unsigned long scale_rt_power(int cpu
)
3750 struct rq
*rq
= cpu_rq(cpu
);
3751 u64 total
, available
;
3753 sched_avg_update(rq
);
3755 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3756 available
= total
- rq
->rt_avg
;
3758 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3759 total
= SCHED_LOAD_SCALE
;
3761 total
>>= SCHED_LOAD_SHIFT
;
3763 return div_u64(available
, total
);
3766 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3768 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3769 unsigned long power
= SCHED_LOAD_SCALE
;
3770 struct sched_group
*sdg
= sd
->groups
;
3771 unsigned long old
= sdg
->__cpu_power
;
3773 /* here we could scale based on cpufreq */
3775 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3776 power
*= arch_scale_smt_power(sd
, cpu
);
3777 power
>>= SCHED_LOAD_SHIFT
;
3780 power
*= scale_rt_power(cpu
);
3781 power
>>= SCHED_LOAD_SHIFT
;
3787 sdg
->__cpu_power
= power
;
3788 sdg
->reciprocal_cpu_power
= reciprocal_value(power
);
3792 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3794 struct sched_domain
*child
= sd
->child
;
3795 struct sched_group
*group
, *sdg
= sd
->groups
;
3796 unsigned long power
= sdg
->__cpu_power
;
3799 update_cpu_power(sd
, cpu
);
3803 sdg
->__cpu_power
= 0;
3805 group
= child
->groups
;
3807 sdg
->__cpu_power
+= group
->__cpu_power
;
3808 group
= group
->next
;
3809 } while (group
!= child
->groups
);
3811 if (power
!= sdg
->__cpu_power
)
3812 sdg
->reciprocal_cpu_power
= reciprocal_value(sdg
->__cpu_power
);
3816 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3817 * @group: sched_group whose statistics are to be updated.
3818 * @this_cpu: Cpu for which load balance is currently performed.
3819 * @idle: Idle status of this_cpu
3820 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3821 * @sd_idle: Idle status of the sched_domain containing group.
3822 * @local_group: Does group contain this_cpu.
3823 * @cpus: Set of cpus considered for load balancing.
3824 * @balance: Should we balance.
3825 * @sgs: variable to hold the statistics for this group.
3827 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3828 struct sched_group
*group
, int this_cpu
,
3829 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3830 int local_group
, const struct cpumask
*cpus
,
3831 int *balance
, struct sg_lb_stats
*sgs
)
3833 unsigned long load
, max_cpu_load
, min_cpu_load
;
3835 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3836 unsigned long sum_avg_load_per_task
;
3837 unsigned long avg_load_per_task
;
3840 balance_cpu
= group_first_cpu(group
);
3841 if (balance_cpu
== this_cpu
)
3842 update_group_power(sd
, this_cpu
);
3845 /* Tally up the load of all CPUs in the group */
3846 sum_avg_load_per_task
= avg_load_per_task
= 0;
3848 min_cpu_load
= ~0UL;
3850 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3851 struct rq
*rq
= cpu_rq(i
);
3853 if (*sd_idle
&& rq
->nr_running
)
3856 /* Bias balancing toward cpus of our domain */
3858 if (idle_cpu(i
) && !first_idle_cpu
) {
3863 load
= target_load(i
, load_idx
);
3865 load
= source_load(i
, load_idx
);
3866 if (load
> max_cpu_load
)
3867 max_cpu_load
= load
;
3868 if (min_cpu_load
> load
)
3869 min_cpu_load
= load
;
3872 sgs
->group_load
+= load
;
3873 sgs
->sum_nr_running
+= rq
->nr_running
;
3874 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3876 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3880 * First idle cpu or the first cpu(busiest) in this sched group
3881 * is eligible for doing load balancing at this and above
3882 * domains. In the newly idle case, we will allow all the cpu's
3883 * to do the newly idle load balance.
3885 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3886 balance_cpu
!= this_cpu
&& balance
) {
3891 /* Adjust by relative CPU power of the group */
3892 sgs
->avg_load
= sg_div_cpu_power(group
,
3893 sgs
->group_load
* SCHED_LOAD_SCALE
);
3897 * Consider the group unbalanced when the imbalance is larger
3898 * than the average weight of two tasks.
3900 * APZ: with cgroup the avg task weight can vary wildly and
3901 * might not be a suitable number - should we keep a
3902 * normalized nr_running number somewhere that negates
3905 avg_load_per_task
= sg_div_cpu_power(group
,
3906 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3908 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3911 sgs
->group_capacity
=
3912 DIV_ROUND_CLOSEST(group
->__cpu_power
, SCHED_LOAD_SCALE
);
3916 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3917 * @sd: sched_domain whose statistics are to be updated.
3918 * @this_cpu: Cpu for which load balance is currently performed.
3919 * @idle: Idle status of this_cpu
3920 * @sd_idle: Idle status of the sched_domain containing group.
3921 * @cpus: Set of cpus considered for load balancing.
3922 * @balance: Should we balance.
3923 * @sds: variable to hold the statistics for this sched_domain.
3925 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3926 enum cpu_idle_type idle
, int *sd_idle
,
3927 const struct cpumask
*cpus
, int *balance
,
3928 struct sd_lb_stats
*sds
)
3930 struct sched_domain
*child
= sd
->child
;
3931 struct sched_group
*group
= sd
->groups
;
3932 struct sg_lb_stats sgs
;
3933 int load_idx
, prefer_sibling
= 0;
3935 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3938 init_sd_power_savings_stats(sd
, sds
, idle
);
3939 load_idx
= get_sd_load_idx(sd
, idle
);
3944 local_group
= cpumask_test_cpu(this_cpu
,
3945 sched_group_cpus(group
));
3946 memset(&sgs
, 0, sizeof(sgs
));
3947 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3948 local_group
, cpus
, balance
, &sgs
);
3950 if (local_group
&& balance
&& !(*balance
))
3953 sds
->total_load
+= sgs
.group_load
;
3954 sds
->total_pwr
+= group
->__cpu_power
;
3957 * In case the child domain prefers tasks go to siblings
3958 * first, lower the group capacity to one so that we'll try
3959 * and move all the excess tasks away.
3962 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3965 sds
->this_load
= sgs
.avg_load
;
3967 sds
->this_nr_running
= sgs
.sum_nr_running
;
3968 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3969 } else if (sgs
.avg_load
> sds
->max_load
&&
3970 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3972 sds
->max_load
= sgs
.avg_load
;
3973 sds
->busiest
= group
;
3974 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3975 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3976 sds
->group_imb
= sgs
.group_imb
;
3979 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3980 group
= group
->next
;
3981 } while (group
!= sd
->groups
);
3985 * fix_small_imbalance - Calculate the minor imbalance that exists
3986 * amongst the groups of a sched_domain, during
3988 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3989 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3990 * @imbalance: Variable to store the imbalance.
3992 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3993 int this_cpu
, unsigned long *imbalance
)
3995 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3996 unsigned int imbn
= 2;
3998 if (sds
->this_nr_running
) {
3999 sds
->this_load_per_task
/= sds
->this_nr_running
;
4000 if (sds
->busiest_load_per_task
>
4001 sds
->this_load_per_task
)
4004 sds
->this_load_per_task
=
4005 cpu_avg_load_per_task(this_cpu
);
4007 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
4008 sds
->busiest_load_per_task
* imbn
) {
4009 *imbalance
= sds
->busiest_load_per_task
;
4014 * OK, we don't have enough imbalance to justify moving tasks,
4015 * however we may be able to increase total CPU power used by
4019 pwr_now
+= sds
->busiest
->__cpu_power
*
4020 min(sds
->busiest_load_per_task
, sds
->max_load
);
4021 pwr_now
+= sds
->this->__cpu_power
*
4022 min(sds
->this_load_per_task
, sds
->this_load
);
4023 pwr_now
/= SCHED_LOAD_SCALE
;
4025 /* Amount of load we'd subtract */
4026 tmp
= sg_div_cpu_power(sds
->busiest
,
4027 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
4028 if (sds
->max_load
> tmp
)
4029 pwr_move
+= sds
->busiest
->__cpu_power
*
4030 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4032 /* Amount of load we'd add */
4033 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
4034 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
4035 tmp
= sg_div_cpu_power(sds
->this,
4036 sds
->max_load
* sds
->busiest
->__cpu_power
);
4038 tmp
= sg_div_cpu_power(sds
->this,
4039 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
4040 pwr_move
+= sds
->this->__cpu_power
*
4041 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4042 pwr_move
/= SCHED_LOAD_SCALE
;
4044 /* Move if we gain throughput */
4045 if (pwr_move
> pwr_now
)
4046 *imbalance
= sds
->busiest_load_per_task
;
4050 * calculate_imbalance - Calculate the amount of imbalance present within the
4051 * groups of a given sched_domain during load balance.
4052 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4053 * @this_cpu: Cpu for which currently load balance is being performed.
4054 * @imbalance: The variable to store the imbalance.
4056 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
4057 unsigned long *imbalance
)
4059 unsigned long max_pull
;
4061 * In the presence of smp nice balancing, certain scenarios can have
4062 * max load less than avg load(as we skip the groups at or below
4063 * its cpu_power, while calculating max_load..)
4065 if (sds
->max_load
< sds
->avg_load
) {
4067 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4070 /* Don't want to pull so many tasks that a group would go idle */
4071 max_pull
= min(sds
->max_load
- sds
->avg_load
,
4072 sds
->max_load
- sds
->busiest_load_per_task
);
4074 /* How much load to actually move to equalise the imbalance */
4075 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
4076 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
4080 * if *imbalance is less than the average load per runnable task
4081 * there is no gaurantee that any tasks will be moved so we'll have
4082 * a think about bumping its value to force at least one task to be
4085 if (*imbalance
< sds
->busiest_load_per_task
)
4086 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4089 /******* find_busiest_group() helpers end here *********************/
4092 * find_busiest_group - Returns the busiest group within the sched_domain
4093 * if there is an imbalance. If there isn't an imbalance, and
4094 * the user has opted for power-savings, it returns a group whose
4095 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4096 * such a group exists.
4098 * Also calculates the amount of weighted load which should be moved
4099 * to restore balance.
4101 * @sd: The sched_domain whose busiest group is to be returned.
4102 * @this_cpu: The cpu for which load balancing is currently being performed.
4103 * @imbalance: Variable which stores amount of weighted load which should
4104 * be moved to restore balance/put a group to idle.
4105 * @idle: The idle status of this_cpu.
4106 * @sd_idle: The idleness of sd
4107 * @cpus: The set of CPUs under consideration for load-balancing.
4108 * @balance: Pointer to a variable indicating if this_cpu
4109 * is the appropriate cpu to perform load balancing at this_level.
4111 * Returns: - the busiest group if imbalance exists.
4112 * - If no imbalance and user has opted for power-savings balance,
4113 * return the least loaded group whose CPUs can be
4114 * put to idle by rebalancing its tasks onto our group.
4116 static struct sched_group
*
4117 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4118 unsigned long *imbalance
, enum cpu_idle_type idle
,
4119 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4121 struct sd_lb_stats sds
;
4123 memset(&sds
, 0, sizeof(sds
));
4126 * Compute the various statistics relavent for load balancing at
4129 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4132 /* Cases where imbalance does not exist from POV of this_cpu */
4133 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4135 * 2) There is no busy sibling group to pull from.
4136 * 3) This group is the busiest group.
4137 * 4) This group is more busy than the avg busieness at this
4139 * 5) The imbalance is within the specified limit.
4140 * 6) Any rebalance would lead to ping-pong
4142 if (balance
&& !(*balance
))
4145 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4148 if (sds
.this_load
>= sds
.max_load
)
4151 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4153 if (sds
.this_load
>= sds
.avg_load
)
4156 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4159 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4161 sds
.busiest_load_per_task
=
4162 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4165 * We're trying to get all the cpus to the average_load, so we don't
4166 * want to push ourselves above the average load, nor do we wish to
4167 * reduce the max loaded cpu below the average load, as either of these
4168 * actions would just result in more rebalancing later, and ping-pong
4169 * tasks around. Thus we look for the minimum possible imbalance.
4170 * Negative imbalances (*we* are more loaded than anyone else) will
4171 * be counted as no imbalance for these purposes -- we can't fix that
4172 * by pulling tasks to us. Be careful of negative numbers as they'll
4173 * appear as very large values with unsigned longs.
4175 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4178 /* Looks like there is an imbalance. Compute it */
4179 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4184 * There is no obvious imbalance. But check if we can do some balancing
4187 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4194 static struct sched_group
*group_of(int cpu
)
4196 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
4204 static unsigned long power_of(int cpu
)
4206 struct sched_group
*group
= group_of(cpu
);
4209 return SCHED_LOAD_SCALE
;
4211 return group
->__cpu_power
;
4215 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4218 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4219 unsigned long imbalance
, const struct cpumask
*cpus
)
4221 struct rq
*busiest
= NULL
, *rq
;
4222 unsigned long max_load
= 0;
4225 for_each_cpu(i
, sched_group_cpus(group
)) {
4226 unsigned long power
= power_of(i
);
4227 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4230 if (!cpumask_test_cpu(i
, cpus
))
4234 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4237 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4240 if (wl
> max_load
) {
4250 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4251 * so long as it is large enough.
4253 #define MAX_PINNED_INTERVAL 512
4255 /* Working cpumask for load_balance and load_balance_newidle. */
4256 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4259 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4260 * tasks if there is an imbalance.
4262 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4263 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4266 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4267 struct sched_group
*group
;
4268 unsigned long imbalance
;
4270 unsigned long flags
;
4271 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4273 cpumask_setall(cpus
);
4276 * When power savings policy is enabled for the parent domain, idle
4277 * sibling can pick up load irrespective of busy siblings. In this case,
4278 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4279 * portraying it as CPU_NOT_IDLE.
4281 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4282 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4285 schedstat_inc(sd
, lb_count
[idle
]);
4289 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4296 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4300 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4302 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4306 BUG_ON(busiest
== this_rq
);
4308 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4311 if (busiest
->nr_running
> 1) {
4313 * Attempt to move tasks. If find_busiest_group has found
4314 * an imbalance but busiest->nr_running <= 1, the group is
4315 * still unbalanced. ld_moved simply stays zero, so it is
4316 * correctly treated as an imbalance.
4318 local_irq_save(flags
);
4319 double_rq_lock(this_rq
, busiest
);
4320 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4321 imbalance
, sd
, idle
, &all_pinned
);
4322 double_rq_unlock(this_rq
, busiest
);
4323 local_irq_restore(flags
);
4326 * some other cpu did the load balance for us.
4328 if (ld_moved
&& this_cpu
!= smp_processor_id())
4329 resched_cpu(this_cpu
);
4331 /* All tasks on this runqueue were pinned by CPU affinity */
4332 if (unlikely(all_pinned
)) {
4333 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4334 if (!cpumask_empty(cpus
))
4341 schedstat_inc(sd
, lb_failed
[idle
]);
4342 sd
->nr_balance_failed
++;
4344 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4346 spin_lock_irqsave(&busiest
->lock
, flags
);
4348 /* don't kick the migration_thread, if the curr
4349 * task on busiest cpu can't be moved to this_cpu
4351 if (!cpumask_test_cpu(this_cpu
,
4352 &busiest
->curr
->cpus_allowed
)) {
4353 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4355 goto out_one_pinned
;
4358 if (!busiest
->active_balance
) {
4359 busiest
->active_balance
= 1;
4360 busiest
->push_cpu
= this_cpu
;
4363 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4365 wake_up_process(busiest
->migration_thread
);
4368 * We've kicked active balancing, reset the failure
4371 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4374 sd
->nr_balance_failed
= 0;
4376 if (likely(!active_balance
)) {
4377 /* We were unbalanced, so reset the balancing interval */
4378 sd
->balance_interval
= sd
->min_interval
;
4381 * If we've begun active balancing, start to back off. This
4382 * case may not be covered by the all_pinned logic if there
4383 * is only 1 task on the busy runqueue (because we don't call
4386 if (sd
->balance_interval
< sd
->max_interval
)
4387 sd
->balance_interval
*= 2;
4390 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4391 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4397 schedstat_inc(sd
, lb_balanced
[idle
]);
4399 sd
->nr_balance_failed
= 0;
4402 /* tune up the balancing interval */
4403 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4404 (sd
->balance_interval
< sd
->max_interval
))
4405 sd
->balance_interval
*= 2;
4407 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4408 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4419 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4420 * tasks if there is an imbalance.
4422 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4423 * this_rq is locked.
4426 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4428 struct sched_group
*group
;
4429 struct rq
*busiest
= NULL
;
4430 unsigned long imbalance
;
4434 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4436 cpumask_setall(cpus
);
4439 * When power savings policy is enabled for the parent domain, idle
4440 * sibling can pick up load irrespective of busy siblings. In this case,
4441 * let the state of idle sibling percolate up as IDLE, instead of
4442 * portraying it as CPU_NOT_IDLE.
4444 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4445 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4448 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4450 update_shares_locked(this_rq
, sd
);
4451 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4452 &sd_idle
, cpus
, NULL
);
4454 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4458 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4460 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4464 BUG_ON(busiest
== this_rq
);
4466 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4469 if (busiest
->nr_running
> 1) {
4470 /* Attempt to move tasks */
4471 double_lock_balance(this_rq
, busiest
);
4472 /* this_rq->clock is already updated */
4473 update_rq_clock(busiest
);
4474 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4475 imbalance
, sd
, CPU_NEWLY_IDLE
,
4477 double_unlock_balance(this_rq
, busiest
);
4479 if (unlikely(all_pinned
)) {
4480 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4481 if (!cpumask_empty(cpus
))
4487 int active_balance
= 0;
4489 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4490 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4491 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4494 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4497 if (sd
->nr_balance_failed
++ < 2)
4501 * The only task running in a non-idle cpu can be moved to this
4502 * cpu in an attempt to completely freeup the other CPU
4503 * package. The same method used to move task in load_balance()
4504 * have been extended for load_balance_newidle() to speedup
4505 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4507 * The package power saving logic comes from
4508 * find_busiest_group(). If there are no imbalance, then
4509 * f_b_g() will return NULL. However when sched_mc={1,2} then
4510 * f_b_g() will select a group from which a running task may be
4511 * pulled to this cpu in order to make the other package idle.
4512 * If there is no opportunity to make a package idle and if
4513 * there are no imbalance, then f_b_g() will return NULL and no
4514 * action will be taken in load_balance_newidle().
4516 * Under normal task pull operation due to imbalance, there
4517 * will be more than one task in the source run queue and
4518 * move_tasks() will succeed. ld_moved will be true and this
4519 * active balance code will not be triggered.
4522 /* Lock busiest in correct order while this_rq is held */
4523 double_lock_balance(this_rq
, busiest
);
4526 * don't kick the migration_thread, if the curr
4527 * task on busiest cpu can't be moved to this_cpu
4529 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4530 double_unlock_balance(this_rq
, busiest
);
4535 if (!busiest
->active_balance
) {
4536 busiest
->active_balance
= 1;
4537 busiest
->push_cpu
= this_cpu
;
4541 double_unlock_balance(this_rq
, busiest
);
4543 * Should not call ttwu while holding a rq->lock
4545 spin_unlock(&this_rq
->lock
);
4547 wake_up_process(busiest
->migration_thread
);
4548 spin_lock(&this_rq
->lock
);
4551 sd
->nr_balance_failed
= 0;
4553 update_shares_locked(this_rq
, sd
);
4557 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4558 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4559 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4561 sd
->nr_balance_failed
= 0;
4567 * idle_balance is called by schedule() if this_cpu is about to become
4568 * idle. Attempts to pull tasks from other CPUs.
4570 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4572 struct sched_domain
*sd
;
4573 int pulled_task
= 0;
4574 unsigned long next_balance
= jiffies
+ HZ
;
4576 for_each_domain(this_cpu
, sd
) {
4577 unsigned long interval
;
4579 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4582 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4583 /* If we've pulled tasks over stop searching: */
4584 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4587 interval
= msecs_to_jiffies(sd
->balance_interval
);
4588 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4589 next_balance
= sd
->last_balance
+ interval
;
4593 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4595 * We are going idle. next_balance may be set based on
4596 * a busy processor. So reset next_balance.
4598 this_rq
->next_balance
= next_balance
;
4603 * active_load_balance is run by migration threads. It pushes running tasks
4604 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4605 * running on each physical CPU where possible, and avoids physical /
4606 * logical imbalances.
4608 * Called with busiest_rq locked.
4610 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4612 int target_cpu
= busiest_rq
->push_cpu
;
4613 struct sched_domain
*sd
;
4614 struct rq
*target_rq
;
4616 /* Is there any task to move? */
4617 if (busiest_rq
->nr_running
<= 1)
4620 target_rq
= cpu_rq(target_cpu
);
4623 * This condition is "impossible", if it occurs
4624 * we need to fix it. Originally reported by
4625 * Bjorn Helgaas on a 128-cpu setup.
4627 BUG_ON(busiest_rq
== target_rq
);
4629 /* move a task from busiest_rq to target_rq */
4630 double_lock_balance(busiest_rq
, target_rq
);
4631 update_rq_clock(busiest_rq
);
4632 update_rq_clock(target_rq
);
4634 /* Search for an sd spanning us and the target CPU. */
4635 for_each_domain(target_cpu
, sd
) {
4636 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4637 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4642 schedstat_inc(sd
, alb_count
);
4644 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4646 schedstat_inc(sd
, alb_pushed
);
4648 schedstat_inc(sd
, alb_failed
);
4650 double_unlock_balance(busiest_rq
, target_rq
);
4655 atomic_t load_balancer
;
4656 cpumask_var_t cpu_mask
;
4657 cpumask_var_t ilb_grp_nohz_mask
;
4658 } nohz ____cacheline_aligned
= {
4659 .load_balancer
= ATOMIC_INIT(-1),
4662 int get_nohz_load_balancer(void)
4664 return atomic_read(&nohz
.load_balancer
);
4667 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4669 * lowest_flag_domain - Return lowest sched_domain containing flag.
4670 * @cpu: The cpu whose lowest level of sched domain is to
4672 * @flag: The flag to check for the lowest sched_domain
4673 * for the given cpu.
4675 * Returns the lowest sched_domain of a cpu which contains the given flag.
4677 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4679 struct sched_domain
*sd
;
4681 for_each_domain(cpu
, sd
)
4682 if (sd
&& (sd
->flags
& flag
))
4689 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4690 * @cpu: The cpu whose domains we're iterating over.
4691 * @sd: variable holding the value of the power_savings_sd
4693 * @flag: The flag to filter the sched_domains to be iterated.
4695 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4696 * set, starting from the lowest sched_domain to the highest.
4698 #define for_each_flag_domain(cpu, sd, flag) \
4699 for (sd = lowest_flag_domain(cpu, flag); \
4700 (sd && (sd->flags & flag)); sd = sd->parent)
4703 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4704 * @ilb_group: group to be checked for semi-idleness
4706 * Returns: 1 if the group is semi-idle. 0 otherwise.
4708 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4709 * and atleast one non-idle CPU. This helper function checks if the given
4710 * sched_group is semi-idle or not.
4712 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4714 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4715 sched_group_cpus(ilb_group
));
4718 * A sched_group is semi-idle when it has atleast one busy cpu
4719 * and atleast one idle cpu.
4721 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4724 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4730 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4731 * @cpu: The cpu which is nominating a new idle_load_balancer.
4733 * Returns: Returns the id of the idle load balancer if it exists,
4734 * Else, returns >= nr_cpu_ids.
4736 * This algorithm picks the idle load balancer such that it belongs to a
4737 * semi-idle powersavings sched_domain. The idea is to try and avoid
4738 * completely idle packages/cores just for the purpose of idle load balancing
4739 * when there are other idle cpu's which are better suited for that job.
4741 static int find_new_ilb(int cpu
)
4743 struct sched_domain
*sd
;
4744 struct sched_group
*ilb_group
;
4747 * Have idle load balancer selection from semi-idle packages only
4748 * when power-aware load balancing is enabled
4750 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4754 * Optimize for the case when we have no idle CPUs or only one
4755 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4757 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4760 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4761 ilb_group
= sd
->groups
;
4764 if (is_semi_idle_group(ilb_group
))
4765 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4767 ilb_group
= ilb_group
->next
;
4769 } while (ilb_group
!= sd
->groups
);
4773 return cpumask_first(nohz
.cpu_mask
);
4775 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4776 static inline int find_new_ilb(int call_cpu
)
4778 return cpumask_first(nohz
.cpu_mask
);
4783 * This routine will try to nominate the ilb (idle load balancing)
4784 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4785 * load balancing on behalf of all those cpus. If all the cpus in the system
4786 * go into this tickless mode, then there will be no ilb owner (as there is
4787 * no need for one) and all the cpus will sleep till the next wakeup event
4790 * For the ilb owner, tick is not stopped. And this tick will be used
4791 * for idle load balancing. ilb owner will still be part of
4794 * While stopping the tick, this cpu will become the ilb owner if there
4795 * is no other owner. And will be the owner till that cpu becomes busy
4796 * or if all cpus in the system stop their ticks at which point
4797 * there is no need for ilb owner.
4799 * When the ilb owner becomes busy, it nominates another owner, during the
4800 * next busy scheduler_tick()
4802 int select_nohz_load_balancer(int stop_tick
)
4804 int cpu
= smp_processor_id();
4807 cpu_rq(cpu
)->in_nohz_recently
= 1;
4809 if (!cpu_active(cpu
)) {
4810 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4814 * If we are going offline and still the leader,
4817 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4823 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4825 /* time for ilb owner also to sleep */
4826 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4827 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4828 atomic_set(&nohz
.load_balancer
, -1);
4832 if (atomic_read(&nohz
.load_balancer
) == -1) {
4833 /* make me the ilb owner */
4834 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4836 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4839 if (!(sched_smt_power_savings
||
4840 sched_mc_power_savings
))
4843 * Check to see if there is a more power-efficient
4846 new_ilb
= find_new_ilb(cpu
);
4847 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4848 atomic_set(&nohz
.load_balancer
, -1);
4849 resched_cpu(new_ilb
);
4855 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4858 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4860 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4861 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4868 static DEFINE_SPINLOCK(balancing
);
4871 * It checks each scheduling domain to see if it is due to be balanced,
4872 * and initiates a balancing operation if so.
4874 * Balancing parameters are set up in arch_init_sched_domains.
4876 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4879 struct rq
*rq
= cpu_rq(cpu
);
4880 unsigned long interval
;
4881 struct sched_domain
*sd
;
4882 /* Earliest time when we have to do rebalance again */
4883 unsigned long next_balance
= jiffies
+ 60*HZ
;
4884 int update_next_balance
= 0;
4887 for_each_domain(cpu
, sd
) {
4888 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4891 interval
= sd
->balance_interval
;
4892 if (idle
!= CPU_IDLE
)
4893 interval
*= sd
->busy_factor
;
4895 /* scale ms to jiffies */
4896 interval
= msecs_to_jiffies(interval
);
4897 if (unlikely(!interval
))
4899 if (interval
> HZ
*NR_CPUS
/10)
4900 interval
= HZ
*NR_CPUS
/10;
4902 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4904 if (need_serialize
) {
4905 if (!spin_trylock(&balancing
))
4909 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4910 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4912 * We've pulled tasks over so either we're no
4913 * longer idle, or one of our SMT siblings is
4916 idle
= CPU_NOT_IDLE
;
4918 sd
->last_balance
= jiffies
;
4921 spin_unlock(&balancing
);
4923 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4924 next_balance
= sd
->last_balance
+ interval
;
4925 update_next_balance
= 1;
4929 * Stop the load balance at this level. There is another
4930 * CPU in our sched group which is doing load balancing more
4938 * next_balance will be updated only when there is a need.
4939 * When the cpu is attached to null domain for ex, it will not be
4942 if (likely(update_next_balance
))
4943 rq
->next_balance
= next_balance
;
4947 * run_rebalance_domains is triggered when needed from the scheduler tick.
4948 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4949 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4951 static void run_rebalance_domains(struct softirq_action
*h
)
4953 int this_cpu
= smp_processor_id();
4954 struct rq
*this_rq
= cpu_rq(this_cpu
);
4955 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4956 CPU_IDLE
: CPU_NOT_IDLE
;
4958 rebalance_domains(this_cpu
, idle
);
4962 * If this cpu is the owner for idle load balancing, then do the
4963 * balancing on behalf of the other idle cpus whose ticks are
4966 if (this_rq
->idle_at_tick
&&
4967 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4971 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4972 if (balance_cpu
== this_cpu
)
4976 * If this cpu gets work to do, stop the load balancing
4977 * work being done for other cpus. Next load
4978 * balancing owner will pick it up.
4983 rebalance_domains(balance_cpu
, CPU_IDLE
);
4985 rq
= cpu_rq(balance_cpu
);
4986 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4987 this_rq
->next_balance
= rq
->next_balance
;
4993 static inline int on_null_domain(int cpu
)
4995 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4999 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5001 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
5002 * idle load balancing owner or decide to stop the periodic load balancing,
5003 * if the whole system is idle.
5005 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
5009 * If we were in the nohz mode recently and busy at the current
5010 * scheduler tick, then check if we need to nominate new idle
5013 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
5014 rq
->in_nohz_recently
= 0;
5016 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
5017 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
5018 atomic_set(&nohz
.load_balancer
, -1);
5021 if (atomic_read(&nohz
.load_balancer
) == -1) {
5022 int ilb
= find_new_ilb(cpu
);
5024 if (ilb
< nr_cpu_ids
)
5030 * If this cpu is idle and doing idle load balancing for all the
5031 * cpus with ticks stopped, is it time for that to stop?
5033 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
5034 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
5040 * If this cpu is idle and the idle load balancing is done by
5041 * someone else, then no need raise the SCHED_SOFTIRQ
5043 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
5044 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
5047 /* Don't need to rebalance while attached to NULL domain */
5048 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5049 likely(!on_null_domain(cpu
)))
5050 raise_softirq(SCHED_SOFTIRQ
);
5053 #else /* CONFIG_SMP */
5056 * on UP we do not need to balance between CPUs:
5058 static inline void idle_balance(int cpu
, struct rq
*rq
)
5064 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
5066 EXPORT_PER_CPU_SYMBOL(kstat
);
5069 * Return any ns on the sched_clock that have not yet been accounted in
5070 * @p in case that task is currently running.
5072 * Called with task_rq_lock() held on @rq.
5074 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
5078 if (task_current(rq
, p
)) {
5079 update_rq_clock(rq
);
5080 ns
= rq
->clock
- p
->se
.exec_start
;
5088 unsigned long long task_delta_exec(struct task_struct
*p
)
5090 unsigned long flags
;
5094 rq
= task_rq_lock(p
, &flags
);
5095 ns
= do_task_delta_exec(p
, rq
);
5096 task_rq_unlock(rq
, &flags
);
5102 * Return accounted runtime for the task.
5103 * In case the task is currently running, return the runtime plus current's
5104 * pending runtime that have not been accounted yet.
5106 unsigned long long task_sched_runtime(struct task_struct
*p
)
5108 unsigned long flags
;
5112 rq
= task_rq_lock(p
, &flags
);
5113 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5114 task_rq_unlock(rq
, &flags
);
5120 * Return sum_exec_runtime for the thread group.
5121 * In case the task is currently running, return the sum plus current's
5122 * pending runtime that have not been accounted yet.
5124 * Note that the thread group might have other running tasks as well,
5125 * so the return value not includes other pending runtime that other
5126 * running tasks might have.
5128 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5130 struct task_cputime totals
;
5131 unsigned long flags
;
5135 rq
= task_rq_lock(p
, &flags
);
5136 thread_group_cputime(p
, &totals
);
5137 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5138 task_rq_unlock(rq
, &flags
);
5144 * Account user cpu time to a process.
5145 * @p: the process that the cpu time gets accounted to
5146 * @cputime: the cpu time spent in user space since the last update
5147 * @cputime_scaled: cputime scaled by cpu frequency
5149 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5150 cputime_t cputime_scaled
)
5152 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5155 /* Add user time to process. */
5156 p
->utime
= cputime_add(p
->utime
, cputime
);
5157 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5158 account_group_user_time(p
, cputime
);
5160 /* Add user time to cpustat. */
5161 tmp
= cputime_to_cputime64(cputime
);
5162 if (TASK_NICE(p
) > 0)
5163 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5165 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5167 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5168 /* Account for user time used */
5169 acct_update_integrals(p
);
5173 * Account guest cpu time to a process.
5174 * @p: the process that the cpu time gets accounted to
5175 * @cputime: the cpu time spent in virtual machine since the last update
5176 * @cputime_scaled: cputime scaled by cpu frequency
5178 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5179 cputime_t cputime_scaled
)
5182 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5184 tmp
= cputime_to_cputime64(cputime
);
5186 /* Add guest time to process. */
5187 p
->utime
= cputime_add(p
->utime
, cputime
);
5188 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5189 account_group_user_time(p
, cputime
);
5190 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5192 /* Add guest time to cpustat. */
5193 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5194 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5198 * Account system cpu time to a process.
5199 * @p: the process that the cpu time gets accounted to
5200 * @hardirq_offset: the offset to subtract from hardirq_count()
5201 * @cputime: the cpu time spent in kernel space since the last update
5202 * @cputime_scaled: cputime scaled by cpu frequency
5204 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5205 cputime_t cputime
, cputime_t cputime_scaled
)
5207 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5210 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5211 account_guest_time(p
, cputime
, cputime_scaled
);
5215 /* Add system time to process. */
5216 p
->stime
= cputime_add(p
->stime
, cputime
);
5217 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5218 account_group_system_time(p
, cputime
);
5220 /* Add system time to cpustat. */
5221 tmp
= cputime_to_cputime64(cputime
);
5222 if (hardirq_count() - hardirq_offset
)
5223 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5224 else if (softirq_count())
5225 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5227 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5229 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5231 /* Account for system time used */
5232 acct_update_integrals(p
);
5236 * Account for involuntary wait time.
5237 * @steal: the cpu time spent in involuntary wait
5239 void account_steal_time(cputime_t cputime
)
5241 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5242 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5244 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5248 * Account for idle time.
5249 * @cputime: the cpu time spent in idle wait
5251 void account_idle_time(cputime_t cputime
)
5253 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5254 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5255 struct rq
*rq
= this_rq();
5257 if (atomic_read(&rq
->nr_iowait
) > 0)
5258 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5260 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5263 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5266 * Account a single tick of cpu time.
5267 * @p: the process that the cpu time gets accounted to
5268 * @user_tick: indicates if the tick is a user or a system tick
5270 void account_process_tick(struct task_struct
*p
, int user_tick
)
5272 cputime_t one_jiffy
= jiffies_to_cputime(1);
5273 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5274 struct rq
*rq
= this_rq();
5277 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5278 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5279 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5282 account_idle_time(one_jiffy
);
5286 * Account multiple ticks of steal time.
5287 * @p: the process from which the cpu time has been stolen
5288 * @ticks: number of stolen ticks
5290 void account_steal_ticks(unsigned long ticks
)
5292 account_steal_time(jiffies_to_cputime(ticks
));
5296 * Account multiple ticks of idle time.
5297 * @ticks: number of stolen ticks
5299 void account_idle_ticks(unsigned long ticks
)
5301 account_idle_time(jiffies_to_cputime(ticks
));
5307 * Use precise platform statistics if available:
5309 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5310 cputime_t
task_utime(struct task_struct
*p
)
5315 cputime_t
task_stime(struct task_struct
*p
)
5320 cputime_t
task_utime(struct task_struct
*p
)
5322 clock_t utime
= cputime_to_clock_t(p
->utime
),
5323 total
= utime
+ cputime_to_clock_t(p
->stime
);
5327 * Use CFS's precise accounting:
5329 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5333 do_div(temp
, total
);
5335 utime
= (clock_t)temp
;
5337 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5338 return p
->prev_utime
;
5341 cputime_t
task_stime(struct task_struct
*p
)
5346 * Use CFS's precise accounting. (we subtract utime from
5347 * the total, to make sure the total observed by userspace
5348 * grows monotonically - apps rely on that):
5350 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5351 cputime_to_clock_t(task_utime(p
));
5354 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5356 return p
->prev_stime
;
5360 inline cputime_t
task_gtime(struct task_struct
*p
)
5366 * This function gets called by the timer code, with HZ frequency.
5367 * We call it with interrupts disabled.
5369 * It also gets called by the fork code, when changing the parent's
5372 void scheduler_tick(void)
5374 int cpu
= smp_processor_id();
5375 struct rq
*rq
= cpu_rq(cpu
);
5376 struct task_struct
*curr
= rq
->curr
;
5380 spin_lock(&rq
->lock
);
5381 update_rq_clock(rq
);
5382 update_cpu_load(rq
);
5383 curr
->sched_class
->task_tick(rq
, curr
, 0);
5384 spin_unlock(&rq
->lock
);
5386 perf_counter_task_tick(curr
, cpu
);
5389 rq
->idle_at_tick
= idle_cpu(cpu
);
5390 trigger_load_balance(rq
, cpu
);
5394 notrace
unsigned long get_parent_ip(unsigned long addr
)
5396 if (in_lock_functions(addr
)) {
5397 addr
= CALLER_ADDR2
;
5398 if (in_lock_functions(addr
))
5399 addr
= CALLER_ADDR3
;
5404 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5405 defined(CONFIG_PREEMPT_TRACER))
5407 void __kprobes
add_preempt_count(int val
)
5409 #ifdef CONFIG_DEBUG_PREEMPT
5413 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5416 preempt_count() += val
;
5417 #ifdef CONFIG_DEBUG_PREEMPT
5419 * Spinlock count overflowing soon?
5421 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5424 if (preempt_count() == val
)
5425 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5427 EXPORT_SYMBOL(add_preempt_count
);
5429 void __kprobes
sub_preempt_count(int val
)
5431 #ifdef CONFIG_DEBUG_PREEMPT
5435 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5438 * Is the spinlock portion underflowing?
5440 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5441 !(preempt_count() & PREEMPT_MASK
)))
5445 if (preempt_count() == val
)
5446 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5447 preempt_count() -= val
;
5449 EXPORT_SYMBOL(sub_preempt_count
);
5454 * Print scheduling while atomic bug:
5456 static noinline
void __schedule_bug(struct task_struct
*prev
)
5458 struct pt_regs
*regs
= get_irq_regs();
5460 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5461 prev
->comm
, prev
->pid
, preempt_count());
5463 debug_show_held_locks(prev
);
5465 if (irqs_disabled())
5466 print_irqtrace_events(prev
);
5475 * Various schedule()-time debugging checks and statistics:
5477 static inline void schedule_debug(struct task_struct
*prev
)
5480 * Test if we are atomic. Since do_exit() needs to call into
5481 * schedule() atomically, we ignore that path for now.
5482 * Otherwise, whine if we are scheduling when we should not be.
5484 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5485 __schedule_bug(prev
);
5487 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5489 schedstat_inc(this_rq(), sched_count
);
5490 #ifdef CONFIG_SCHEDSTATS
5491 if (unlikely(prev
->lock_depth
>= 0)) {
5492 schedstat_inc(this_rq(), bkl_count
);
5493 schedstat_inc(prev
, sched_info
.bkl_count
);
5498 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5500 if (prev
->state
== TASK_RUNNING
) {
5501 u64 runtime
= prev
->se
.sum_exec_runtime
;
5503 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5504 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5507 * In order to avoid avg_overlap growing stale when we are
5508 * indeed overlapping and hence not getting put to sleep, grow
5509 * the avg_overlap on preemption.
5511 * We use the average preemption runtime because that
5512 * correlates to the amount of cache footprint a task can
5515 update_avg(&prev
->se
.avg_overlap
, runtime
);
5517 prev
->sched_class
->put_prev_task(rq
, prev
);
5521 * Pick up the highest-prio task:
5523 static inline struct task_struct
*
5524 pick_next_task(struct rq
*rq
)
5526 const struct sched_class
*class;
5527 struct task_struct
*p
;
5530 * Optimization: we know that if all tasks are in
5531 * the fair class we can call that function directly:
5533 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5534 p
= fair_sched_class
.pick_next_task(rq
);
5539 class = sched_class_highest
;
5541 p
= class->pick_next_task(rq
);
5545 * Will never be NULL as the idle class always
5546 * returns a non-NULL p:
5548 class = class->next
;
5553 * schedule() is the main scheduler function.
5555 asmlinkage
void __sched
schedule(void)
5557 struct task_struct
*prev
, *next
;
5558 unsigned long *switch_count
;
5564 cpu
= smp_processor_id();
5568 switch_count
= &prev
->nivcsw
;
5570 release_kernel_lock(prev
);
5571 need_resched_nonpreemptible
:
5573 schedule_debug(prev
);
5575 if (sched_feat(HRTICK
))
5578 spin_lock_irq(&rq
->lock
);
5579 update_rq_clock(rq
);
5580 clear_tsk_need_resched(prev
);
5582 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5583 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5584 prev
->state
= TASK_RUNNING
;
5586 deactivate_task(rq
, prev
, 1);
5587 switch_count
= &prev
->nvcsw
;
5590 pre_schedule(rq
, prev
);
5592 if (unlikely(!rq
->nr_running
))
5593 idle_balance(cpu
, rq
);
5595 put_prev_task(rq
, prev
);
5596 next
= pick_next_task(rq
);
5598 if (likely(prev
!= next
)) {
5599 sched_info_switch(prev
, next
);
5600 perf_counter_task_sched_out(prev
, next
, cpu
);
5606 context_switch(rq
, prev
, next
); /* unlocks the rq */
5608 * the context switch might have flipped the stack from under
5609 * us, hence refresh the local variables.
5611 cpu
= smp_processor_id();
5614 spin_unlock_irq(&rq
->lock
);
5618 if (unlikely(reacquire_kernel_lock(current
) < 0))
5619 goto need_resched_nonpreemptible
;
5621 preempt_enable_no_resched();
5625 EXPORT_SYMBOL(schedule
);
5629 * Look out! "owner" is an entirely speculative pointer
5630 * access and not reliable.
5632 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5637 if (!sched_feat(OWNER_SPIN
))
5640 #ifdef CONFIG_DEBUG_PAGEALLOC
5642 * Need to access the cpu field knowing that
5643 * DEBUG_PAGEALLOC could have unmapped it if
5644 * the mutex owner just released it and exited.
5646 if (probe_kernel_address(&owner
->cpu
, cpu
))
5653 * Even if the access succeeded (likely case),
5654 * the cpu field may no longer be valid.
5656 if (cpu
>= nr_cpumask_bits
)
5660 * We need to validate that we can do a
5661 * get_cpu() and that we have the percpu area.
5663 if (!cpu_online(cpu
))
5670 * Owner changed, break to re-assess state.
5672 if (lock
->owner
!= owner
)
5676 * Is that owner really running on that cpu?
5678 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5688 #ifdef CONFIG_PREEMPT
5690 * this is the entry point to schedule() from in-kernel preemption
5691 * off of preempt_enable. Kernel preemptions off return from interrupt
5692 * occur there and call schedule directly.
5694 asmlinkage
void __sched
preempt_schedule(void)
5696 struct thread_info
*ti
= current_thread_info();
5699 * If there is a non-zero preempt_count or interrupts are disabled,
5700 * we do not want to preempt the current task. Just return..
5702 if (likely(ti
->preempt_count
|| irqs_disabled()))
5706 add_preempt_count(PREEMPT_ACTIVE
);
5708 sub_preempt_count(PREEMPT_ACTIVE
);
5711 * Check again in case we missed a preemption opportunity
5712 * between schedule and now.
5715 } while (need_resched());
5717 EXPORT_SYMBOL(preempt_schedule
);
5720 * this is the entry point to schedule() from kernel preemption
5721 * off of irq context.
5722 * Note, that this is called and return with irqs disabled. This will
5723 * protect us against recursive calling from irq.
5725 asmlinkage
void __sched
preempt_schedule_irq(void)
5727 struct thread_info
*ti
= current_thread_info();
5729 /* Catch callers which need to be fixed */
5730 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5733 add_preempt_count(PREEMPT_ACTIVE
);
5736 local_irq_disable();
5737 sub_preempt_count(PREEMPT_ACTIVE
);
5740 * Check again in case we missed a preemption opportunity
5741 * between schedule and now.
5744 } while (need_resched());
5747 #endif /* CONFIG_PREEMPT */
5749 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5752 return try_to_wake_up(curr
->private, mode
, sync
);
5754 EXPORT_SYMBOL(default_wake_function
);
5757 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5758 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5759 * number) then we wake all the non-exclusive tasks and one exclusive task.
5761 * There are circumstances in which we can try to wake a task which has already
5762 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5763 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5765 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5766 int nr_exclusive
, int sync
, void *key
)
5768 wait_queue_t
*curr
, *next
;
5770 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5771 unsigned flags
= curr
->flags
;
5773 if (curr
->func(curr
, mode
, sync
, key
) &&
5774 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5780 * __wake_up - wake up threads blocked on a waitqueue.
5782 * @mode: which threads
5783 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5784 * @key: is directly passed to the wakeup function
5786 * It may be assumed that this function implies a write memory barrier before
5787 * changing the task state if and only if any tasks are woken up.
5789 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5790 int nr_exclusive
, void *key
)
5792 unsigned long flags
;
5794 spin_lock_irqsave(&q
->lock
, flags
);
5795 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5796 spin_unlock_irqrestore(&q
->lock
, flags
);
5798 EXPORT_SYMBOL(__wake_up
);
5801 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5803 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5805 __wake_up_common(q
, mode
, 1, 0, NULL
);
5808 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5810 __wake_up_common(q
, mode
, 1, 0, key
);
5814 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5816 * @mode: which threads
5817 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5818 * @key: opaque value to be passed to wakeup targets
5820 * The sync wakeup differs that the waker knows that it will schedule
5821 * away soon, so while the target thread will be woken up, it will not
5822 * be migrated to another CPU - ie. the two threads are 'synchronized'
5823 * with each other. This can prevent needless bouncing between CPUs.
5825 * On UP it can prevent extra preemption.
5827 * It may be assumed that this function implies a write memory barrier before
5828 * changing the task state if and only if any tasks are woken up.
5830 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5831 int nr_exclusive
, void *key
)
5833 unsigned long flags
;
5839 if (unlikely(!nr_exclusive
))
5842 spin_lock_irqsave(&q
->lock
, flags
);
5843 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5844 spin_unlock_irqrestore(&q
->lock
, flags
);
5846 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5849 * __wake_up_sync - see __wake_up_sync_key()
5851 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5853 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5855 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5858 * complete: - signals a single thread waiting on this completion
5859 * @x: holds the state of this particular completion
5861 * This will wake up a single thread waiting on this completion. Threads will be
5862 * awakened in the same order in which they were queued.
5864 * See also complete_all(), wait_for_completion() and related routines.
5866 * It may be assumed that this function implies a write memory barrier before
5867 * changing the task state if and only if any tasks are woken up.
5869 void complete(struct completion
*x
)
5871 unsigned long flags
;
5873 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5875 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5876 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5878 EXPORT_SYMBOL(complete
);
5881 * complete_all: - signals all threads waiting on this completion
5882 * @x: holds the state of this particular completion
5884 * This will wake up all threads waiting on this particular completion event.
5886 * It may be assumed that this function implies a write memory barrier before
5887 * changing the task state if and only if any tasks are woken up.
5889 void complete_all(struct completion
*x
)
5891 unsigned long flags
;
5893 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5894 x
->done
+= UINT_MAX
/2;
5895 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5896 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5898 EXPORT_SYMBOL(complete_all
);
5900 static inline long __sched
5901 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5904 DECLARE_WAITQUEUE(wait
, current
);
5906 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5907 __add_wait_queue_tail(&x
->wait
, &wait
);
5909 if (signal_pending_state(state
, current
)) {
5910 timeout
= -ERESTARTSYS
;
5913 __set_current_state(state
);
5914 spin_unlock_irq(&x
->wait
.lock
);
5915 timeout
= schedule_timeout(timeout
);
5916 spin_lock_irq(&x
->wait
.lock
);
5917 } while (!x
->done
&& timeout
);
5918 __remove_wait_queue(&x
->wait
, &wait
);
5923 return timeout
?: 1;
5927 wait_for_common(struct completion
*x
, long timeout
, int state
)
5931 spin_lock_irq(&x
->wait
.lock
);
5932 timeout
= do_wait_for_common(x
, timeout
, state
);
5933 spin_unlock_irq(&x
->wait
.lock
);
5938 * wait_for_completion: - waits for completion of a task
5939 * @x: holds the state of this particular completion
5941 * This waits to be signaled for completion of a specific task. It is NOT
5942 * interruptible and there is no timeout.
5944 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5945 * and interrupt capability. Also see complete().
5947 void __sched
wait_for_completion(struct completion
*x
)
5949 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5951 EXPORT_SYMBOL(wait_for_completion
);
5954 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5955 * @x: holds the state of this particular completion
5956 * @timeout: timeout value in jiffies
5958 * This waits for either a completion of a specific task to be signaled or for a
5959 * specified timeout to expire. The timeout is in jiffies. It is not
5962 unsigned long __sched
5963 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5965 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5967 EXPORT_SYMBOL(wait_for_completion_timeout
);
5970 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5971 * @x: holds the state of this particular completion
5973 * This waits for completion of a specific task to be signaled. It is
5976 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5978 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5979 if (t
== -ERESTARTSYS
)
5983 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5986 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5987 * @x: holds the state of this particular completion
5988 * @timeout: timeout value in jiffies
5990 * This waits for either a completion of a specific task to be signaled or for a
5991 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5993 unsigned long __sched
5994 wait_for_completion_interruptible_timeout(struct completion
*x
,
5995 unsigned long timeout
)
5997 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5999 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
6002 * wait_for_completion_killable: - waits for completion of a task (killable)
6003 * @x: holds the state of this particular completion
6005 * This waits to be signaled for completion of a specific task. It can be
6006 * interrupted by a kill signal.
6008 int __sched
wait_for_completion_killable(struct completion
*x
)
6010 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
6011 if (t
== -ERESTARTSYS
)
6015 EXPORT_SYMBOL(wait_for_completion_killable
);
6018 * try_wait_for_completion - try to decrement a completion without blocking
6019 * @x: completion structure
6021 * Returns: 0 if a decrement cannot be done without blocking
6022 * 1 if a decrement succeeded.
6024 * If a completion is being used as a counting completion,
6025 * attempt to decrement the counter without blocking. This
6026 * enables us to avoid waiting if the resource the completion
6027 * is protecting is not available.
6029 bool try_wait_for_completion(struct completion
*x
)
6033 spin_lock_irq(&x
->wait
.lock
);
6038 spin_unlock_irq(&x
->wait
.lock
);
6041 EXPORT_SYMBOL(try_wait_for_completion
);
6044 * completion_done - Test to see if a completion has any waiters
6045 * @x: completion structure
6047 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6048 * 1 if there are no waiters.
6051 bool completion_done(struct completion
*x
)
6055 spin_lock_irq(&x
->wait
.lock
);
6058 spin_unlock_irq(&x
->wait
.lock
);
6061 EXPORT_SYMBOL(completion_done
);
6064 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
6066 unsigned long flags
;
6069 init_waitqueue_entry(&wait
, current
);
6071 __set_current_state(state
);
6073 spin_lock_irqsave(&q
->lock
, flags
);
6074 __add_wait_queue(q
, &wait
);
6075 spin_unlock(&q
->lock
);
6076 timeout
= schedule_timeout(timeout
);
6077 spin_lock_irq(&q
->lock
);
6078 __remove_wait_queue(q
, &wait
);
6079 spin_unlock_irqrestore(&q
->lock
, flags
);
6084 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
6086 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6088 EXPORT_SYMBOL(interruptible_sleep_on
);
6091 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6093 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
6095 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6097 void __sched
sleep_on(wait_queue_head_t
*q
)
6099 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6101 EXPORT_SYMBOL(sleep_on
);
6103 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6105 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6107 EXPORT_SYMBOL(sleep_on_timeout
);
6109 #ifdef CONFIG_RT_MUTEXES
6112 * rt_mutex_setprio - set the current priority of a task
6114 * @prio: prio value (kernel-internal form)
6116 * This function changes the 'effective' priority of a task. It does
6117 * not touch ->normal_prio like __setscheduler().
6119 * Used by the rt_mutex code to implement priority inheritance logic.
6121 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6123 unsigned long flags
;
6124 int oldprio
, on_rq
, running
;
6126 const struct sched_class
*prev_class
= p
->sched_class
;
6128 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6130 rq
= task_rq_lock(p
, &flags
);
6131 update_rq_clock(rq
);
6134 on_rq
= p
->se
.on_rq
;
6135 running
= task_current(rq
, p
);
6137 dequeue_task(rq
, p
, 0);
6139 p
->sched_class
->put_prev_task(rq
, p
);
6142 p
->sched_class
= &rt_sched_class
;
6144 p
->sched_class
= &fair_sched_class
;
6149 p
->sched_class
->set_curr_task(rq
);
6151 enqueue_task(rq
, p
, 0);
6153 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6155 task_rq_unlock(rq
, &flags
);
6160 void set_user_nice(struct task_struct
*p
, long nice
)
6162 int old_prio
, delta
, on_rq
;
6163 unsigned long flags
;
6166 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6169 * We have to be careful, if called from sys_setpriority(),
6170 * the task might be in the middle of scheduling on another CPU.
6172 rq
= task_rq_lock(p
, &flags
);
6173 update_rq_clock(rq
);
6175 * The RT priorities are set via sched_setscheduler(), but we still
6176 * allow the 'normal' nice value to be set - but as expected
6177 * it wont have any effect on scheduling until the task is
6178 * SCHED_FIFO/SCHED_RR:
6180 if (task_has_rt_policy(p
)) {
6181 p
->static_prio
= NICE_TO_PRIO(nice
);
6184 on_rq
= p
->se
.on_rq
;
6186 dequeue_task(rq
, p
, 0);
6188 p
->static_prio
= NICE_TO_PRIO(nice
);
6191 p
->prio
= effective_prio(p
);
6192 delta
= p
->prio
- old_prio
;
6195 enqueue_task(rq
, p
, 0);
6197 * If the task increased its priority or is running and
6198 * lowered its priority, then reschedule its CPU:
6200 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6201 resched_task(rq
->curr
);
6204 task_rq_unlock(rq
, &flags
);
6206 EXPORT_SYMBOL(set_user_nice
);
6209 * can_nice - check if a task can reduce its nice value
6213 int can_nice(const struct task_struct
*p
, const int nice
)
6215 /* convert nice value [19,-20] to rlimit style value [1,40] */
6216 int nice_rlim
= 20 - nice
;
6218 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6219 capable(CAP_SYS_NICE
));
6222 #ifdef __ARCH_WANT_SYS_NICE
6225 * sys_nice - change the priority of the current process.
6226 * @increment: priority increment
6228 * sys_setpriority is a more generic, but much slower function that
6229 * does similar things.
6231 SYSCALL_DEFINE1(nice
, int, increment
)
6236 * Setpriority might change our priority at the same moment.
6237 * We don't have to worry. Conceptually one call occurs first
6238 * and we have a single winner.
6240 if (increment
< -40)
6245 nice
= TASK_NICE(current
) + increment
;
6251 if (increment
< 0 && !can_nice(current
, nice
))
6254 retval
= security_task_setnice(current
, nice
);
6258 set_user_nice(current
, nice
);
6265 * task_prio - return the priority value of a given task.
6266 * @p: the task in question.
6268 * This is the priority value as seen by users in /proc.
6269 * RT tasks are offset by -200. Normal tasks are centered
6270 * around 0, value goes from -16 to +15.
6272 int task_prio(const struct task_struct
*p
)
6274 return p
->prio
- MAX_RT_PRIO
;
6278 * task_nice - return the nice value of a given task.
6279 * @p: the task in question.
6281 int task_nice(const struct task_struct
*p
)
6283 return TASK_NICE(p
);
6285 EXPORT_SYMBOL(task_nice
);
6288 * idle_cpu - is a given cpu idle currently?
6289 * @cpu: the processor in question.
6291 int idle_cpu(int cpu
)
6293 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6297 * idle_task - return the idle task for a given cpu.
6298 * @cpu: the processor in question.
6300 struct task_struct
*idle_task(int cpu
)
6302 return cpu_rq(cpu
)->idle
;
6306 * find_process_by_pid - find a process with a matching PID value.
6307 * @pid: the pid in question.
6309 static struct task_struct
*find_process_by_pid(pid_t pid
)
6311 return pid
? find_task_by_vpid(pid
) : current
;
6314 /* Actually do priority change: must hold rq lock. */
6316 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6318 BUG_ON(p
->se
.on_rq
);
6321 switch (p
->policy
) {
6325 p
->sched_class
= &fair_sched_class
;
6329 p
->sched_class
= &rt_sched_class
;
6333 p
->rt_priority
= prio
;
6334 p
->normal_prio
= normal_prio(p
);
6335 /* we are holding p->pi_lock already */
6336 p
->prio
= rt_mutex_getprio(p
);
6341 * check the target process has a UID that matches the current process's
6343 static bool check_same_owner(struct task_struct
*p
)
6345 const struct cred
*cred
= current_cred(), *pcred
;
6349 pcred
= __task_cred(p
);
6350 match
= (cred
->euid
== pcred
->euid
||
6351 cred
->euid
== pcred
->uid
);
6356 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6357 struct sched_param
*param
, bool user
)
6359 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6360 unsigned long flags
;
6361 const struct sched_class
*prev_class
= p
->sched_class
;
6365 /* may grab non-irq protected spin_locks */
6366 BUG_ON(in_interrupt());
6368 /* double check policy once rq lock held */
6370 reset_on_fork
= p
->sched_reset_on_fork
;
6371 policy
= oldpolicy
= p
->policy
;
6373 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6374 policy
&= ~SCHED_RESET_ON_FORK
;
6376 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6377 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6378 policy
!= SCHED_IDLE
)
6383 * Valid priorities for SCHED_FIFO and SCHED_RR are
6384 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6385 * SCHED_BATCH and SCHED_IDLE is 0.
6387 if (param
->sched_priority
< 0 ||
6388 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6389 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6391 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6395 * Allow unprivileged RT tasks to decrease priority:
6397 if (user
&& !capable(CAP_SYS_NICE
)) {
6398 if (rt_policy(policy
)) {
6399 unsigned long rlim_rtprio
;
6401 if (!lock_task_sighand(p
, &flags
))
6403 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6404 unlock_task_sighand(p
, &flags
);
6406 /* can't set/change the rt policy */
6407 if (policy
!= p
->policy
&& !rlim_rtprio
)
6410 /* can't increase priority */
6411 if (param
->sched_priority
> p
->rt_priority
&&
6412 param
->sched_priority
> rlim_rtprio
)
6416 * Like positive nice levels, dont allow tasks to
6417 * move out of SCHED_IDLE either:
6419 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6422 /* can't change other user's priorities */
6423 if (!check_same_owner(p
))
6426 /* Normal users shall not reset the sched_reset_on_fork flag */
6427 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6432 #ifdef CONFIG_RT_GROUP_SCHED
6434 * Do not allow realtime tasks into groups that have no runtime
6437 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6438 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6442 retval
= security_task_setscheduler(p
, policy
, param
);
6448 * make sure no PI-waiters arrive (or leave) while we are
6449 * changing the priority of the task:
6451 spin_lock_irqsave(&p
->pi_lock
, flags
);
6453 * To be able to change p->policy safely, the apropriate
6454 * runqueue lock must be held.
6456 rq
= __task_rq_lock(p
);
6457 /* recheck policy now with rq lock held */
6458 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6459 policy
= oldpolicy
= -1;
6460 __task_rq_unlock(rq
);
6461 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6464 update_rq_clock(rq
);
6465 on_rq
= p
->se
.on_rq
;
6466 running
= task_current(rq
, p
);
6468 deactivate_task(rq
, p
, 0);
6470 p
->sched_class
->put_prev_task(rq
, p
);
6472 p
->sched_reset_on_fork
= reset_on_fork
;
6475 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6478 p
->sched_class
->set_curr_task(rq
);
6480 activate_task(rq
, p
, 0);
6482 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6484 __task_rq_unlock(rq
);
6485 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6487 rt_mutex_adjust_pi(p
);
6493 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6494 * @p: the task in question.
6495 * @policy: new policy.
6496 * @param: structure containing the new RT priority.
6498 * NOTE that the task may be already dead.
6500 int sched_setscheduler(struct task_struct
*p
, int policy
,
6501 struct sched_param
*param
)
6503 return __sched_setscheduler(p
, policy
, param
, true);
6505 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6508 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6509 * @p: the task in question.
6510 * @policy: new policy.
6511 * @param: structure containing the new RT priority.
6513 * Just like sched_setscheduler, only don't bother checking if the
6514 * current context has permission. For example, this is needed in
6515 * stop_machine(): we create temporary high priority worker threads,
6516 * but our caller might not have that capability.
6518 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6519 struct sched_param
*param
)
6521 return __sched_setscheduler(p
, policy
, param
, false);
6525 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6527 struct sched_param lparam
;
6528 struct task_struct
*p
;
6531 if (!param
|| pid
< 0)
6533 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6538 p
= find_process_by_pid(pid
);
6540 retval
= sched_setscheduler(p
, policy
, &lparam
);
6547 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6548 * @pid: the pid in question.
6549 * @policy: new policy.
6550 * @param: structure containing the new RT priority.
6552 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6553 struct sched_param __user
*, param
)
6555 /* negative values for policy are not valid */
6559 return do_sched_setscheduler(pid
, policy
, param
);
6563 * sys_sched_setparam - set/change the RT priority of a thread
6564 * @pid: the pid in question.
6565 * @param: structure containing the new RT priority.
6567 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6569 return do_sched_setscheduler(pid
, -1, param
);
6573 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6574 * @pid: the pid in question.
6576 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6578 struct task_struct
*p
;
6585 read_lock(&tasklist_lock
);
6586 p
= find_process_by_pid(pid
);
6588 retval
= security_task_getscheduler(p
);
6591 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6593 read_unlock(&tasklist_lock
);
6598 * sys_sched_getparam - get the RT priority of a thread
6599 * @pid: the pid in question.
6600 * @param: structure containing the RT priority.
6602 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6604 struct sched_param lp
;
6605 struct task_struct
*p
;
6608 if (!param
|| pid
< 0)
6611 read_lock(&tasklist_lock
);
6612 p
= find_process_by_pid(pid
);
6617 retval
= security_task_getscheduler(p
);
6621 lp
.sched_priority
= p
->rt_priority
;
6622 read_unlock(&tasklist_lock
);
6625 * This one might sleep, we cannot do it with a spinlock held ...
6627 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6632 read_unlock(&tasklist_lock
);
6636 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6638 cpumask_var_t cpus_allowed
, new_mask
;
6639 struct task_struct
*p
;
6643 read_lock(&tasklist_lock
);
6645 p
= find_process_by_pid(pid
);
6647 read_unlock(&tasklist_lock
);
6653 * It is not safe to call set_cpus_allowed with the
6654 * tasklist_lock held. We will bump the task_struct's
6655 * usage count and then drop tasklist_lock.
6658 read_unlock(&tasklist_lock
);
6660 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6664 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6666 goto out_free_cpus_allowed
;
6669 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6672 retval
= security_task_setscheduler(p
, 0, NULL
);
6676 cpuset_cpus_allowed(p
, cpus_allowed
);
6677 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6679 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6682 cpuset_cpus_allowed(p
, cpus_allowed
);
6683 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6685 * We must have raced with a concurrent cpuset
6686 * update. Just reset the cpus_allowed to the
6687 * cpuset's cpus_allowed
6689 cpumask_copy(new_mask
, cpus_allowed
);
6694 free_cpumask_var(new_mask
);
6695 out_free_cpus_allowed
:
6696 free_cpumask_var(cpus_allowed
);
6703 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6704 struct cpumask
*new_mask
)
6706 if (len
< cpumask_size())
6707 cpumask_clear(new_mask
);
6708 else if (len
> cpumask_size())
6709 len
= cpumask_size();
6711 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6715 * sys_sched_setaffinity - set the cpu affinity of a process
6716 * @pid: pid of the process
6717 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6718 * @user_mask_ptr: user-space pointer to the new cpu mask
6720 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6721 unsigned long __user
*, user_mask_ptr
)
6723 cpumask_var_t new_mask
;
6726 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6729 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6731 retval
= sched_setaffinity(pid
, new_mask
);
6732 free_cpumask_var(new_mask
);
6736 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6738 struct task_struct
*p
;
6742 read_lock(&tasklist_lock
);
6745 p
= find_process_by_pid(pid
);
6749 retval
= security_task_getscheduler(p
);
6753 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6756 read_unlock(&tasklist_lock
);
6763 * sys_sched_getaffinity - get the cpu affinity of a process
6764 * @pid: pid of the process
6765 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6766 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6768 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6769 unsigned long __user
*, user_mask_ptr
)
6774 if (len
< cpumask_size())
6777 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6780 ret
= sched_getaffinity(pid
, mask
);
6782 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6785 ret
= cpumask_size();
6787 free_cpumask_var(mask
);
6793 * sys_sched_yield - yield the current processor to other threads.
6795 * This function yields the current CPU to other tasks. If there are no
6796 * other threads running on this CPU then this function will return.
6798 SYSCALL_DEFINE0(sched_yield
)
6800 struct rq
*rq
= this_rq_lock();
6802 schedstat_inc(rq
, yld_count
);
6803 current
->sched_class
->yield_task(rq
);
6806 * Since we are going to call schedule() anyway, there's
6807 * no need to preempt or enable interrupts:
6809 __release(rq
->lock
);
6810 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6811 _raw_spin_unlock(&rq
->lock
);
6812 preempt_enable_no_resched();
6819 static inline int should_resched(void)
6821 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6824 static void __cond_resched(void)
6826 add_preempt_count(PREEMPT_ACTIVE
);
6828 sub_preempt_count(PREEMPT_ACTIVE
);
6831 int __sched
_cond_resched(void)
6833 if (should_resched()) {
6839 EXPORT_SYMBOL(_cond_resched
);
6842 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6843 * call schedule, and on return reacquire the lock.
6845 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6846 * operations here to prevent schedule() from being called twice (once via
6847 * spin_unlock(), once by hand).
6849 int __cond_resched_lock(spinlock_t
*lock
)
6851 int resched
= should_resched();
6854 if (spin_needbreak(lock
) || resched
) {
6865 EXPORT_SYMBOL(__cond_resched_lock
);
6867 int __sched
__cond_resched_softirq(void)
6869 BUG_ON(!in_softirq());
6871 if (should_resched()) {
6879 EXPORT_SYMBOL(__cond_resched_softirq
);
6882 * yield - yield the current processor to other threads.
6884 * This is a shortcut for kernel-space yielding - it marks the
6885 * thread runnable and calls sys_sched_yield().
6887 void __sched
yield(void)
6889 set_current_state(TASK_RUNNING
);
6892 EXPORT_SYMBOL(yield
);
6895 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6896 * that process accounting knows that this is a task in IO wait state.
6898 * But don't do that if it is a deliberate, throttling IO wait (this task
6899 * has set its backing_dev_info: the queue against which it should throttle)
6901 void __sched
io_schedule(void)
6903 struct rq
*rq
= raw_rq();
6905 delayacct_blkio_start();
6906 atomic_inc(&rq
->nr_iowait
);
6907 current
->in_iowait
= 1;
6909 current
->in_iowait
= 0;
6910 atomic_dec(&rq
->nr_iowait
);
6911 delayacct_blkio_end();
6913 EXPORT_SYMBOL(io_schedule
);
6915 long __sched
io_schedule_timeout(long timeout
)
6917 struct rq
*rq
= raw_rq();
6920 delayacct_blkio_start();
6921 atomic_inc(&rq
->nr_iowait
);
6922 current
->in_iowait
= 1;
6923 ret
= schedule_timeout(timeout
);
6924 current
->in_iowait
= 0;
6925 atomic_dec(&rq
->nr_iowait
);
6926 delayacct_blkio_end();
6931 * sys_sched_get_priority_max - return maximum RT priority.
6932 * @policy: scheduling class.
6934 * this syscall returns the maximum rt_priority that can be used
6935 * by a given scheduling class.
6937 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6944 ret
= MAX_USER_RT_PRIO
-1;
6956 * sys_sched_get_priority_min - return minimum RT priority.
6957 * @policy: scheduling class.
6959 * this syscall returns the minimum rt_priority that can be used
6960 * by a given scheduling class.
6962 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6980 * sys_sched_rr_get_interval - return the default timeslice of a process.
6981 * @pid: pid of the process.
6982 * @interval: userspace pointer to the timeslice value.
6984 * this syscall writes the default timeslice value of a given process
6985 * into the user-space timespec buffer. A value of '0' means infinity.
6987 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6988 struct timespec __user
*, interval
)
6990 struct task_struct
*p
;
6991 unsigned int time_slice
;
6999 read_lock(&tasklist_lock
);
7000 p
= find_process_by_pid(pid
);
7004 retval
= security_task_getscheduler(p
);
7009 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
7010 * tasks that are on an otherwise idle runqueue:
7013 if (p
->policy
== SCHED_RR
) {
7014 time_slice
= DEF_TIMESLICE
;
7015 } else if (p
->policy
!= SCHED_FIFO
) {
7016 struct sched_entity
*se
= &p
->se
;
7017 unsigned long flags
;
7020 rq
= task_rq_lock(p
, &flags
);
7021 if (rq
->cfs
.load
.weight
)
7022 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
7023 task_rq_unlock(rq
, &flags
);
7025 read_unlock(&tasklist_lock
);
7026 jiffies_to_timespec(time_slice
, &t
);
7027 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
7031 read_unlock(&tasklist_lock
);
7035 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
7037 void sched_show_task(struct task_struct
*p
)
7039 unsigned long free
= 0;
7042 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
7043 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
7044 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
7045 #if BITS_PER_LONG == 32
7046 if (state
== TASK_RUNNING
)
7047 printk(KERN_CONT
" running ");
7049 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
7051 if (state
== TASK_RUNNING
)
7052 printk(KERN_CONT
" running task ");
7054 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
7056 #ifdef CONFIG_DEBUG_STACK_USAGE
7057 free
= stack_not_used(p
);
7059 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
7060 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
7061 (unsigned long)task_thread_info(p
)->flags
);
7063 show_stack(p
, NULL
);
7066 void show_state_filter(unsigned long state_filter
)
7068 struct task_struct
*g
, *p
;
7070 #if BITS_PER_LONG == 32
7072 " task PC stack pid father\n");
7075 " task PC stack pid father\n");
7077 read_lock(&tasklist_lock
);
7078 do_each_thread(g
, p
) {
7080 * reset the NMI-timeout, listing all files on a slow
7081 * console might take alot of time:
7083 touch_nmi_watchdog();
7084 if (!state_filter
|| (p
->state
& state_filter
))
7086 } while_each_thread(g
, p
);
7088 touch_all_softlockup_watchdogs();
7090 #ifdef CONFIG_SCHED_DEBUG
7091 sysrq_sched_debug_show();
7093 read_unlock(&tasklist_lock
);
7095 * Only show locks if all tasks are dumped:
7097 if (state_filter
== -1)
7098 debug_show_all_locks();
7101 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
7103 idle
->sched_class
= &idle_sched_class
;
7107 * init_idle - set up an idle thread for a given CPU
7108 * @idle: task in question
7109 * @cpu: cpu the idle task belongs to
7111 * NOTE: this function does not set the idle thread's NEED_RESCHED
7112 * flag, to make booting more robust.
7114 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7116 struct rq
*rq
= cpu_rq(cpu
);
7117 unsigned long flags
;
7119 spin_lock_irqsave(&rq
->lock
, flags
);
7122 idle
->se
.exec_start
= sched_clock();
7124 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
7125 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7126 __set_task_cpu(idle
, cpu
);
7128 rq
->curr
= rq
->idle
= idle
;
7129 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7132 spin_unlock_irqrestore(&rq
->lock
, flags
);
7134 /* Set the preempt count _outside_ the spinlocks! */
7135 #if defined(CONFIG_PREEMPT)
7136 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7138 task_thread_info(idle
)->preempt_count
= 0;
7141 * The idle tasks have their own, simple scheduling class:
7143 idle
->sched_class
= &idle_sched_class
;
7144 ftrace_graph_init_task(idle
);
7148 * In a system that switches off the HZ timer nohz_cpu_mask
7149 * indicates which cpus entered this state. This is used
7150 * in the rcu update to wait only for active cpus. For system
7151 * which do not switch off the HZ timer nohz_cpu_mask should
7152 * always be CPU_BITS_NONE.
7154 cpumask_var_t nohz_cpu_mask
;
7157 * Increase the granularity value when there are more CPUs,
7158 * because with more CPUs the 'effective latency' as visible
7159 * to users decreases. But the relationship is not linear,
7160 * so pick a second-best guess by going with the log2 of the
7163 * This idea comes from the SD scheduler of Con Kolivas:
7165 static inline void sched_init_granularity(void)
7167 unsigned int factor
= 1 + ilog2(num_online_cpus());
7168 const unsigned long limit
= 200000000;
7170 sysctl_sched_min_granularity
*= factor
;
7171 if (sysctl_sched_min_granularity
> limit
)
7172 sysctl_sched_min_granularity
= limit
;
7174 sysctl_sched_latency
*= factor
;
7175 if (sysctl_sched_latency
> limit
)
7176 sysctl_sched_latency
= limit
;
7178 sysctl_sched_wakeup_granularity
*= factor
;
7180 sysctl_sched_shares_ratelimit
*= factor
;
7185 * This is how migration works:
7187 * 1) we queue a struct migration_req structure in the source CPU's
7188 * runqueue and wake up that CPU's migration thread.
7189 * 2) we down() the locked semaphore => thread blocks.
7190 * 3) migration thread wakes up (implicitly it forces the migrated
7191 * thread off the CPU)
7192 * 4) it gets the migration request and checks whether the migrated
7193 * task is still in the wrong runqueue.
7194 * 5) if it's in the wrong runqueue then the migration thread removes
7195 * it and puts it into the right queue.
7196 * 6) migration thread up()s the semaphore.
7197 * 7) we wake up and the migration is done.
7201 * Change a given task's CPU affinity. Migrate the thread to a
7202 * proper CPU and schedule it away if the CPU it's executing on
7203 * is removed from the allowed bitmask.
7205 * NOTE: the caller must have a valid reference to the task, the
7206 * task must not exit() & deallocate itself prematurely. The
7207 * call is not atomic; no spinlocks may be held.
7209 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7211 struct migration_req req
;
7212 unsigned long flags
;
7216 rq
= task_rq_lock(p
, &flags
);
7217 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7222 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7223 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7228 if (p
->sched_class
->set_cpus_allowed
)
7229 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7231 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7232 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7235 /* Can the task run on the task's current CPU? If so, we're done */
7236 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7239 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7240 /* Need help from migration thread: drop lock and wait. */
7241 struct task_struct
*mt
= rq
->migration_thread
;
7243 get_task_struct(mt
);
7244 task_rq_unlock(rq
, &flags
);
7245 wake_up_process(rq
->migration_thread
);
7246 put_task_struct(mt
);
7247 wait_for_completion(&req
.done
);
7248 tlb_migrate_finish(p
->mm
);
7252 task_rq_unlock(rq
, &flags
);
7256 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7259 * Move (not current) task off this cpu, onto dest cpu. We're doing
7260 * this because either it can't run here any more (set_cpus_allowed()
7261 * away from this CPU, or CPU going down), or because we're
7262 * attempting to rebalance this task on exec (sched_exec).
7264 * So we race with normal scheduler movements, but that's OK, as long
7265 * as the task is no longer on this CPU.
7267 * Returns non-zero if task was successfully migrated.
7269 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7271 struct rq
*rq_dest
, *rq_src
;
7274 if (unlikely(!cpu_active(dest_cpu
)))
7277 rq_src
= cpu_rq(src_cpu
);
7278 rq_dest
= cpu_rq(dest_cpu
);
7280 double_rq_lock(rq_src
, rq_dest
);
7281 /* Already moved. */
7282 if (task_cpu(p
) != src_cpu
)
7284 /* Affinity changed (again). */
7285 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7288 on_rq
= p
->se
.on_rq
;
7290 deactivate_task(rq_src
, p
, 0);
7292 set_task_cpu(p
, dest_cpu
);
7294 activate_task(rq_dest
, p
, 0);
7295 check_preempt_curr(rq_dest
, p
, 0);
7300 double_rq_unlock(rq_src
, rq_dest
);
7305 * migration_thread - this is a highprio system thread that performs
7306 * thread migration by bumping thread off CPU then 'pushing' onto
7309 static int migration_thread(void *data
)
7311 int cpu
= (long)data
;
7315 BUG_ON(rq
->migration_thread
!= current
);
7317 set_current_state(TASK_INTERRUPTIBLE
);
7318 while (!kthread_should_stop()) {
7319 struct migration_req
*req
;
7320 struct list_head
*head
;
7322 spin_lock_irq(&rq
->lock
);
7324 if (cpu_is_offline(cpu
)) {
7325 spin_unlock_irq(&rq
->lock
);
7329 if (rq
->active_balance
) {
7330 active_load_balance(rq
, cpu
);
7331 rq
->active_balance
= 0;
7334 head
= &rq
->migration_queue
;
7336 if (list_empty(head
)) {
7337 spin_unlock_irq(&rq
->lock
);
7339 set_current_state(TASK_INTERRUPTIBLE
);
7342 req
= list_entry(head
->next
, struct migration_req
, list
);
7343 list_del_init(head
->next
);
7345 spin_unlock(&rq
->lock
);
7346 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7349 complete(&req
->done
);
7351 __set_current_state(TASK_RUNNING
);
7356 #ifdef CONFIG_HOTPLUG_CPU
7358 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7362 local_irq_disable();
7363 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7369 * Figure out where task on dead CPU should go, use force if necessary.
7371 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7374 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7377 /* Look for allowed, online CPU in same node. */
7378 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7379 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7382 /* Any allowed, online CPU? */
7383 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7384 if (dest_cpu
< nr_cpu_ids
)
7387 /* No more Mr. Nice Guy. */
7388 if (dest_cpu
>= nr_cpu_ids
) {
7389 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7390 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7393 * Don't tell them about moving exiting tasks or
7394 * kernel threads (both mm NULL), since they never
7397 if (p
->mm
&& printk_ratelimit()) {
7398 printk(KERN_INFO
"process %d (%s) no "
7399 "longer affine to cpu%d\n",
7400 task_pid_nr(p
), p
->comm
, dead_cpu
);
7405 /* It can have affinity changed while we were choosing. */
7406 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7411 * While a dead CPU has no uninterruptible tasks queued at this point,
7412 * it might still have a nonzero ->nr_uninterruptible counter, because
7413 * for performance reasons the counter is not stricly tracking tasks to
7414 * their home CPUs. So we just add the counter to another CPU's counter,
7415 * to keep the global sum constant after CPU-down:
7417 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7419 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7420 unsigned long flags
;
7422 local_irq_save(flags
);
7423 double_rq_lock(rq_src
, rq_dest
);
7424 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7425 rq_src
->nr_uninterruptible
= 0;
7426 double_rq_unlock(rq_src
, rq_dest
);
7427 local_irq_restore(flags
);
7430 /* Run through task list and migrate tasks from the dead cpu. */
7431 static void migrate_live_tasks(int src_cpu
)
7433 struct task_struct
*p
, *t
;
7435 read_lock(&tasklist_lock
);
7437 do_each_thread(t
, p
) {
7441 if (task_cpu(p
) == src_cpu
)
7442 move_task_off_dead_cpu(src_cpu
, p
);
7443 } while_each_thread(t
, p
);
7445 read_unlock(&tasklist_lock
);
7449 * Schedules idle task to be the next runnable task on current CPU.
7450 * It does so by boosting its priority to highest possible.
7451 * Used by CPU offline code.
7453 void sched_idle_next(void)
7455 int this_cpu
= smp_processor_id();
7456 struct rq
*rq
= cpu_rq(this_cpu
);
7457 struct task_struct
*p
= rq
->idle
;
7458 unsigned long flags
;
7460 /* cpu has to be offline */
7461 BUG_ON(cpu_online(this_cpu
));
7464 * Strictly not necessary since rest of the CPUs are stopped by now
7465 * and interrupts disabled on the current cpu.
7467 spin_lock_irqsave(&rq
->lock
, flags
);
7469 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7471 update_rq_clock(rq
);
7472 activate_task(rq
, p
, 0);
7474 spin_unlock_irqrestore(&rq
->lock
, flags
);
7478 * Ensures that the idle task is using init_mm right before its cpu goes
7481 void idle_task_exit(void)
7483 struct mm_struct
*mm
= current
->active_mm
;
7485 BUG_ON(cpu_online(smp_processor_id()));
7488 switch_mm(mm
, &init_mm
, current
);
7492 /* called under rq->lock with disabled interrupts */
7493 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7495 struct rq
*rq
= cpu_rq(dead_cpu
);
7497 /* Must be exiting, otherwise would be on tasklist. */
7498 BUG_ON(!p
->exit_state
);
7500 /* Cannot have done final schedule yet: would have vanished. */
7501 BUG_ON(p
->state
== TASK_DEAD
);
7506 * Drop lock around migration; if someone else moves it,
7507 * that's OK. No task can be added to this CPU, so iteration is
7510 spin_unlock_irq(&rq
->lock
);
7511 move_task_off_dead_cpu(dead_cpu
, p
);
7512 spin_lock_irq(&rq
->lock
);
7517 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7518 static void migrate_dead_tasks(unsigned int dead_cpu
)
7520 struct rq
*rq
= cpu_rq(dead_cpu
);
7521 struct task_struct
*next
;
7524 if (!rq
->nr_running
)
7526 update_rq_clock(rq
);
7527 next
= pick_next_task(rq
);
7530 next
->sched_class
->put_prev_task(rq
, next
);
7531 migrate_dead(dead_cpu
, next
);
7537 * remove the tasks which were accounted by rq from calc_load_tasks.
7539 static void calc_global_load_remove(struct rq
*rq
)
7541 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7542 rq
->calc_load_active
= 0;
7544 #endif /* CONFIG_HOTPLUG_CPU */
7546 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7548 static struct ctl_table sd_ctl_dir
[] = {
7550 .procname
= "sched_domain",
7556 static struct ctl_table sd_ctl_root
[] = {
7558 .ctl_name
= CTL_KERN
,
7559 .procname
= "kernel",
7561 .child
= sd_ctl_dir
,
7566 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7568 struct ctl_table
*entry
=
7569 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7574 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7576 struct ctl_table
*entry
;
7579 * In the intermediate directories, both the child directory and
7580 * procname are dynamically allocated and could fail but the mode
7581 * will always be set. In the lowest directory the names are
7582 * static strings and all have proc handlers.
7584 for (entry
= *tablep
; entry
->mode
; entry
++) {
7586 sd_free_ctl_entry(&entry
->child
);
7587 if (entry
->proc_handler
== NULL
)
7588 kfree(entry
->procname
);
7596 set_table_entry(struct ctl_table
*entry
,
7597 const char *procname
, void *data
, int maxlen
,
7598 mode_t mode
, proc_handler
*proc_handler
)
7600 entry
->procname
= procname
;
7602 entry
->maxlen
= maxlen
;
7604 entry
->proc_handler
= proc_handler
;
7607 static struct ctl_table
*
7608 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7610 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7615 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7616 sizeof(long), 0644, proc_doulongvec_minmax
);
7617 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7618 sizeof(long), 0644, proc_doulongvec_minmax
);
7619 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7620 sizeof(int), 0644, proc_dointvec_minmax
);
7621 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7622 sizeof(int), 0644, proc_dointvec_minmax
);
7623 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7624 sizeof(int), 0644, proc_dointvec_minmax
);
7625 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7626 sizeof(int), 0644, proc_dointvec_minmax
);
7627 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7628 sizeof(int), 0644, proc_dointvec_minmax
);
7629 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7630 sizeof(int), 0644, proc_dointvec_minmax
);
7631 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7632 sizeof(int), 0644, proc_dointvec_minmax
);
7633 set_table_entry(&table
[9], "cache_nice_tries",
7634 &sd
->cache_nice_tries
,
7635 sizeof(int), 0644, proc_dointvec_minmax
);
7636 set_table_entry(&table
[10], "flags", &sd
->flags
,
7637 sizeof(int), 0644, proc_dointvec_minmax
);
7638 set_table_entry(&table
[11], "name", sd
->name
,
7639 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7640 /* &table[12] is terminator */
7645 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7647 struct ctl_table
*entry
, *table
;
7648 struct sched_domain
*sd
;
7649 int domain_num
= 0, i
;
7652 for_each_domain(cpu
, sd
)
7654 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7659 for_each_domain(cpu
, sd
) {
7660 snprintf(buf
, 32, "domain%d", i
);
7661 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7663 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7670 static struct ctl_table_header
*sd_sysctl_header
;
7671 static void register_sched_domain_sysctl(void)
7673 int i
, cpu_num
= num_online_cpus();
7674 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7677 WARN_ON(sd_ctl_dir
[0].child
);
7678 sd_ctl_dir
[0].child
= entry
;
7683 for_each_online_cpu(i
) {
7684 snprintf(buf
, 32, "cpu%d", i
);
7685 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7687 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7691 WARN_ON(sd_sysctl_header
);
7692 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7695 /* may be called multiple times per register */
7696 static void unregister_sched_domain_sysctl(void)
7698 if (sd_sysctl_header
)
7699 unregister_sysctl_table(sd_sysctl_header
);
7700 sd_sysctl_header
= NULL
;
7701 if (sd_ctl_dir
[0].child
)
7702 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7705 static void register_sched_domain_sysctl(void)
7708 static void unregister_sched_domain_sysctl(void)
7713 static void set_rq_online(struct rq
*rq
)
7716 const struct sched_class
*class;
7718 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7721 for_each_class(class) {
7722 if (class->rq_online
)
7723 class->rq_online(rq
);
7728 static void set_rq_offline(struct rq
*rq
)
7731 const struct sched_class
*class;
7733 for_each_class(class) {
7734 if (class->rq_offline
)
7735 class->rq_offline(rq
);
7738 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7744 * migration_call - callback that gets triggered when a CPU is added.
7745 * Here we can start up the necessary migration thread for the new CPU.
7747 static int __cpuinit
7748 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7750 struct task_struct
*p
;
7751 int cpu
= (long)hcpu
;
7752 unsigned long flags
;
7757 case CPU_UP_PREPARE
:
7758 case CPU_UP_PREPARE_FROZEN
:
7759 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7762 kthread_bind(p
, cpu
);
7763 /* Must be high prio: stop_machine expects to yield to it. */
7764 rq
= task_rq_lock(p
, &flags
);
7765 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7766 task_rq_unlock(rq
, &flags
);
7768 cpu_rq(cpu
)->migration_thread
= p
;
7769 rq
->calc_load_update
= calc_load_update
;
7773 case CPU_ONLINE_FROZEN
:
7774 /* Strictly unnecessary, as first user will wake it. */
7775 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7777 /* Update our root-domain */
7779 spin_lock_irqsave(&rq
->lock
, flags
);
7781 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7785 spin_unlock_irqrestore(&rq
->lock
, flags
);
7788 #ifdef CONFIG_HOTPLUG_CPU
7789 case CPU_UP_CANCELED
:
7790 case CPU_UP_CANCELED_FROZEN
:
7791 if (!cpu_rq(cpu
)->migration_thread
)
7793 /* Unbind it from offline cpu so it can run. Fall thru. */
7794 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7795 cpumask_any(cpu_online_mask
));
7796 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7797 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7798 cpu_rq(cpu
)->migration_thread
= NULL
;
7802 case CPU_DEAD_FROZEN
:
7803 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7804 migrate_live_tasks(cpu
);
7806 kthread_stop(rq
->migration_thread
);
7807 put_task_struct(rq
->migration_thread
);
7808 rq
->migration_thread
= NULL
;
7809 /* Idle task back to normal (off runqueue, low prio) */
7810 spin_lock_irq(&rq
->lock
);
7811 update_rq_clock(rq
);
7812 deactivate_task(rq
, rq
->idle
, 0);
7813 rq
->idle
->static_prio
= MAX_PRIO
;
7814 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7815 rq
->idle
->sched_class
= &idle_sched_class
;
7816 migrate_dead_tasks(cpu
);
7817 spin_unlock_irq(&rq
->lock
);
7819 migrate_nr_uninterruptible(rq
);
7820 BUG_ON(rq
->nr_running
!= 0);
7821 calc_global_load_remove(rq
);
7823 * No need to migrate the tasks: it was best-effort if
7824 * they didn't take sched_hotcpu_mutex. Just wake up
7827 spin_lock_irq(&rq
->lock
);
7828 while (!list_empty(&rq
->migration_queue
)) {
7829 struct migration_req
*req
;
7831 req
= list_entry(rq
->migration_queue
.next
,
7832 struct migration_req
, list
);
7833 list_del_init(&req
->list
);
7834 spin_unlock_irq(&rq
->lock
);
7835 complete(&req
->done
);
7836 spin_lock_irq(&rq
->lock
);
7838 spin_unlock_irq(&rq
->lock
);
7842 case CPU_DYING_FROZEN
:
7843 /* Update our root-domain */
7845 spin_lock_irqsave(&rq
->lock
, flags
);
7847 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7850 spin_unlock_irqrestore(&rq
->lock
, flags
);
7858 * Register at high priority so that task migration (migrate_all_tasks)
7859 * happens before everything else. This has to be lower priority than
7860 * the notifier in the perf_counter subsystem, though.
7862 static struct notifier_block __cpuinitdata migration_notifier
= {
7863 .notifier_call
= migration_call
,
7867 static int __init
migration_init(void)
7869 void *cpu
= (void *)(long)smp_processor_id();
7872 /* Start one for the boot CPU: */
7873 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7874 BUG_ON(err
== NOTIFY_BAD
);
7875 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7876 register_cpu_notifier(&migration_notifier
);
7880 early_initcall(migration_init
);
7885 #ifdef CONFIG_SCHED_DEBUG
7887 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7888 struct cpumask
*groupmask
)
7890 struct sched_group
*group
= sd
->groups
;
7893 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7894 cpumask_clear(groupmask
);
7896 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7898 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7899 printk("does not load-balance\n");
7901 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7906 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7908 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7909 printk(KERN_ERR
"ERROR: domain->span does not contain "
7912 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7913 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7917 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7921 printk(KERN_ERR
"ERROR: group is NULL\n");
7925 if (!group
->__cpu_power
) {
7926 printk(KERN_CONT
"\n");
7927 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7932 if (!cpumask_weight(sched_group_cpus(group
))) {
7933 printk(KERN_CONT
"\n");
7934 printk(KERN_ERR
"ERROR: empty group\n");
7938 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7939 printk(KERN_CONT
"\n");
7940 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7944 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7946 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7948 printk(KERN_CONT
" %s", str
);
7949 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7950 printk(KERN_CONT
" (__cpu_power = %d)",
7951 group
->__cpu_power
);
7954 group
= group
->next
;
7955 } while (group
!= sd
->groups
);
7956 printk(KERN_CONT
"\n");
7958 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7959 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7962 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7963 printk(KERN_ERR
"ERROR: parent span is not a superset "
7964 "of domain->span\n");
7968 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7970 cpumask_var_t groupmask
;
7974 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7978 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7980 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7981 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7986 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7993 free_cpumask_var(groupmask
);
7995 #else /* !CONFIG_SCHED_DEBUG */
7996 # define sched_domain_debug(sd, cpu) do { } while (0)
7997 #endif /* CONFIG_SCHED_DEBUG */
7999 static int sd_degenerate(struct sched_domain
*sd
)
8001 if (cpumask_weight(sched_domain_span(sd
)) == 1)
8004 /* Following flags need at least 2 groups */
8005 if (sd
->flags
& (SD_LOAD_BALANCE
|
8006 SD_BALANCE_NEWIDLE
|
8010 SD_SHARE_PKG_RESOURCES
)) {
8011 if (sd
->groups
!= sd
->groups
->next
)
8015 /* Following flags don't use groups */
8016 if (sd
->flags
& (SD_WAKE_IDLE
|
8025 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
8027 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
8029 if (sd_degenerate(parent
))
8032 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
8035 /* Does parent contain flags not in child? */
8036 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
8037 if (cflags
& SD_WAKE_AFFINE
)
8038 pflags
&= ~SD_WAKE_BALANCE
;
8039 /* Flags needing groups don't count if only 1 group in parent */
8040 if (parent
->groups
== parent
->groups
->next
) {
8041 pflags
&= ~(SD_LOAD_BALANCE
|
8042 SD_BALANCE_NEWIDLE
|
8046 SD_SHARE_PKG_RESOURCES
);
8047 if (nr_node_ids
== 1)
8048 pflags
&= ~SD_SERIALIZE
;
8050 if (~cflags
& pflags
)
8056 static void free_rootdomain(struct root_domain
*rd
)
8058 cpupri_cleanup(&rd
->cpupri
);
8060 free_cpumask_var(rd
->rto_mask
);
8061 free_cpumask_var(rd
->online
);
8062 free_cpumask_var(rd
->span
);
8066 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8068 struct root_domain
*old_rd
= NULL
;
8069 unsigned long flags
;
8071 spin_lock_irqsave(&rq
->lock
, flags
);
8076 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8079 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8082 * If we dont want to free the old_rt yet then
8083 * set old_rd to NULL to skip the freeing later
8086 if (!atomic_dec_and_test(&old_rd
->refcount
))
8090 atomic_inc(&rd
->refcount
);
8093 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8094 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8097 spin_unlock_irqrestore(&rq
->lock
, flags
);
8100 free_rootdomain(old_rd
);
8103 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8105 gfp_t gfp
= GFP_KERNEL
;
8107 memset(rd
, 0, sizeof(*rd
));
8112 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8114 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8116 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8119 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8124 free_cpumask_var(rd
->rto_mask
);
8126 free_cpumask_var(rd
->online
);
8128 free_cpumask_var(rd
->span
);
8133 static void init_defrootdomain(void)
8135 init_rootdomain(&def_root_domain
, true);
8137 atomic_set(&def_root_domain
.refcount
, 1);
8140 static struct root_domain
*alloc_rootdomain(void)
8142 struct root_domain
*rd
;
8144 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8148 if (init_rootdomain(rd
, false) != 0) {
8157 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8158 * hold the hotplug lock.
8161 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8163 struct rq
*rq
= cpu_rq(cpu
);
8164 struct sched_domain
*tmp
;
8166 /* Remove the sched domains which do not contribute to scheduling. */
8167 for (tmp
= sd
; tmp
; ) {
8168 struct sched_domain
*parent
= tmp
->parent
;
8172 if (sd_parent_degenerate(tmp
, parent
)) {
8173 tmp
->parent
= parent
->parent
;
8175 parent
->parent
->child
= tmp
;
8180 if (sd
&& sd_degenerate(sd
)) {
8186 sched_domain_debug(sd
, cpu
);
8188 rq_attach_root(rq
, rd
);
8189 rcu_assign_pointer(rq
->sd
, sd
);
8192 /* cpus with isolated domains */
8193 static cpumask_var_t cpu_isolated_map
;
8195 /* Setup the mask of cpus configured for isolated domains */
8196 static int __init
isolated_cpu_setup(char *str
)
8198 cpulist_parse(str
, cpu_isolated_map
);
8202 __setup("isolcpus=", isolated_cpu_setup
);
8205 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8206 * to a function which identifies what group(along with sched group) a CPU
8207 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8208 * (due to the fact that we keep track of groups covered with a struct cpumask).
8210 * init_sched_build_groups will build a circular linked list of the groups
8211 * covered by the given span, and will set each group's ->cpumask correctly,
8212 * and ->cpu_power to 0.
8215 init_sched_build_groups(const struct cpumask
*span
,
8216 const struct cpumask
*cpu_map
,
8217 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8218 struct sched_group
**sg
,
8219 struct cpumask
*tmpmask
),
8220 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8222 struct sched_group
*first
= NULL
, *last
= NULL
;
8225 cpumask_clear(covered
);
8227 for_each_cpu(i
, span
) {
8228 struct sched_group
*sg
;
8229 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8232 if (cpumask_test_cpu(i
, covered
))
8235 cpumask_clear(sched_group_cpus(sg
));
8236 sg
->__cpu_power
= 0;
8238 for_each_cpu(j
, span
) {
8239 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8242 cpumask_set_cpu(j
, covered
);
8243 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8254 #define SD_NODES_PER_DOMAIN 16
8259 * find_next_best_node - find the next node to include in a sched_domain
8260 * @node: node whose sched_domain we're building
8261 * @used_nodes: nodes already in the sched_domain
8263 * Find the next node to include in a given scheduling domain. Simply
8264 * finds the closest node not already in the @used_nodes map.
8266 * Should use nodemask_t.
8268 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8270 int i
, n
, val
, min_val
, best_node
= 0;
8274 for (i
= 0; i
< nr_node_ids
; i
++) {
8275 /* Start at @node */
8276 n
= (node
+ i
) % nr_node_ids
;
8278 if (!nr_cpus_node(n
))
8281 /* Skip already used nodes */
8282 if (node_isset(n
, *used_nodes
))
8285 /* Simple min distance search */
8286 val
= node_distance(node
, n
);
8288 if (val
< min_val
) {
8294 node_set(best_node
, *used_nodes
);
8299 * sched_domain_node_span - get a cpumask for a node's sched_domain
8300 * @node: node whose cpumask we're constructing
8301 * @span: resulting cpumask
8303 * Given a node, construct a good cpumask for its sched_domain to span. It
8304 * should be one that prevents unnecessary balancing, but also spreads tasks
8307 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8309 nodemask_t used_nodes
;
8312 cpumask_clear(span
);
8313 nodes_clear(used_nodes
);
8315 cpumask_or(span
, span
, cpumask_of_node(node
));
8316 node_set(node
, used_nodes
);
8318 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8319 int next_node
= find_next_best_node(node
, &used_nodes
);
8321 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8324 #endif /* CONFIG_NUMA */
8326 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8329 * The cpus mask in sched_group and sched_domain hangs off the end.
8331 * ( See the the comments in include/linux/sched.h:struct sched_group
8332 * and struct sched_domain. )
8334 struct static_sched_group
{
8335 struct sched_group sg
;
8336 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8339 struct static_sched_domain
{
8340 struct sched_domain sd
;
8341 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8347 cpumask_var_t domainspan
;
8348 cpumask_var_t covered
;
8349 cpumask_var_t notcovered
;
8351 cpumask_var_t nodemask
;
8352 cpumask_var_t this_sibling_map
;
8353 cpumask_var_t this_core_map
;
8354 cpumask_var_t send_covered
;
8355 cpumask_var_t tmpmask
;
8356 struct sched_group
**sched_group_nodes
;
8357 struct root_domain
*rd
;
8361 sa_sched_groups
= 0,
8366 sa_this_sibling_map
,
8368 sa_sched_group_nodes
,
8378 * SMT sched-domains:
8380 #ifdef CONFIG_SCHED_SMT
8381 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8382 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8385 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8386 struct sched_group
**sg
, struct cpumask
*unused
)
8389 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8392 #endif /* CONFIG_SCHED_SMT */
8395 * multi-core sched-domains:
8397 #ifdef CONFIG_SCHED_MC
8398 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8399 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8400 #endif /* CONFIG_SCHED_MC */
8402 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8404 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8405 struct sched_group
**sg
, struct cpumask
*mask
)
8409 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8410 group
= cpumask_first(mask
);
8412 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8415 #elif defined(CONFIG_SCHED_MC)
8417 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8418 struct sched_group
**sg
, struct cpumask
*unused
)
8421 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8426 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8427 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8430 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8431 struct sched_group
**sg
, struct cpumask
*mask
)
8434 #ifdef CONFIG_SCHED_MC
8435 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8436 group
= cpumask_first(mask
);
8437 #elif defined(CONFIG_SCHED_SMT)
8438 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8439 group
= cpumask_first(mask
);
8444 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8450 * The init_sched_build_groups can't handle what we want to do with node
8451 * groups, so roll our own. Now each node has its own list of groups which
8452 * gets dynamically allocated.
8454 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8455 static struct sched_group
***sched_group_nodes_bycpu
;
8457 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8458 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8460 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8461 struct sched_group
**sg
,
8462 struct cpumask
*nodemask
)
8466 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8467 group
= cpumask_first(nodemask
);
8470 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8474 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8476 struct sched_group
*sg
= group_head
;
8482 for_each_cpu(j
, sched_group_cpus(sg
)) {
8483 struct sched_domain
*sd
;
8485 sd
= &per_cpu(phys_domains
, j
).sd
;
8486 if (j
!= group_first_cpu(sd
->groups
)) {
8488 * Only add "power" once for each
8494 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
8497 } while (sg
!= group_head
);
8500 static int build_numa_sched_groups(struct s_data
*d
,
8501 const struct cpumask
*cpu_map
, int num
)
8503 struct sched_domain
*sd
;
8504 struct sched_group
*sg
, *prev
;
8507 cpumask_clear(d
->covered
);
8508 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8509 if (cpumask_empty(d
->nodemask
)) {
8510 d
->sched_group_nodes
[num
] = NULL
;
8514 sched_domain_node_span(num
, d
->domainspan
);
8515 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8517 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8520 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8524 d
->sched_group_nodes
[num
] = sg
;
8526 for_each_cpu(j
, d
->nodemask
) {
8527 sd
= &per_cpu(node_domains
, j
).sd
;
8531 sg
->__cpu_power
= 0;
8532 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8534 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8537 for (j
= 0; j
< nr_node_ids
; j
++) {
8538 n
= (num
+ j
) % nr_node_ids
;
8539 cpumask_complement(d
->notcovered
, d
->covered
);
8540 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8541 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8542 if (cpumask_empty(d
->tmpmask
))
8544 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8545 if (cpumask_empty(d
->tmpmask
))
8547 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8551 "Can not alloc domain group for node %d\n", j
);
8554 sg
->__cpu_power
= 0;
8555 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8556 sg
->next
= prev
->next
;
8557 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8564 #endif /* CONFIG_NUMA */
8567 /* Free memory allocated for various sched_group structures */
8568 static void free_sched_groups(const struct cpumask
*cpu_map
,
8569 struct cpumask
*nodemask
)
8573 for_each_cpu(cpu
, cpu_map
) {
8574 struct sched_group
**sched_group_nodes
8575 = sched_group_nodes_bycpu
[cpu
];
8577 if (!sched_group_nodes
)
8580 for (i
= 0; i
< nr_node_ids
; i
++) {
8581 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8583 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8584 if (cpumask_empty(nodemask
))
8594 if (oldsg
!= sched_group_nodes
[i
])
8597 kfree(sched_group_nodes
);
8598 sched_group_nodes_bycpu
[cpu
] = NULL
;
8601 #else /* !CONFIG_NUMA */
8602 static void free_sched_groups(const struct cpumask
*cpu_map
,
8603 struct cpumask
*nodemask
)
8606 #endif /* CONFIG_NUMA */
8609 * Initialize sched groups cpu_power.
8611 * cpu_power indicates the capacity of sched group, which is used while
8612 * distributing the load between different sched groups in a sched domain.
8613 * Typically cpu_power for all the groups in a sched domain will be same unless
8614 * there are asymmetries in the topology. If there are asymmetries, group
8615 * having more cpu_power will pickup more load compared to the group having
8618 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8620 struct sched_domain
*child
;
8621 struct sched_group
*group
;
8625 WARN_ON(!sd
|| !sd
->groups
);
8627 if (cpu
!= group_first_cpu(sd
->groups
))
8632 sd
->groups
->__cpu_power
= 0;
8635 power
= SCHED_LOAD_SCALE
;
8636 weight
= cpumask_weight(sched_domain_span(sd
));
8638 * SMT siblings share the power of a single core.
8639 * Usually multiple threads get a better yield out of
8640 * that one core than a single thread would have,
8641 * reflect that in sd->smt_gain.
8643 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8644 power
*= sd
->smt_gain
;
8646 power
>>= SCHED_LOAD_SHIFT
;
8648 sg_inc_cpu_power(sd
->groups
, power
);
8653 * Add cpu_power of each child group to this groups cpu_power.
8655 group
= child
->groups
;
8657 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8658 group
= group
->next
;
8659 } while (group
!= child
->groups
);
8663 * Initializers for schedule domains
8664 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8667 #ifdef CONFIG_SCHED_DEBUG
8668 # define SD_INIT_NAME(sd, type) sd->name = #type
8670 # define SD_INIT_NAME(sd, type) do { } while (0)
8673 #define SD_INIT(sd, type) sd_init_##type(sd)
8675 #define SD_INIT_FUNC(type) \
8676 static noinline void sd_init_##type(struct sched_domain *sd) \
8678 memset(sd, 0, sizeof(*sd)); \
8679 *sd = SD_##type##_INIT; \
8680 sd->level = SD_LV_##type; \
8681 SD_INIT_NAME(sd, type); \
8686 SD_INIT_FUNC(ALLNODES
)
8689 #ifdef CONFIG_SCHED_SMT
8690 SD_INIT_FUNC(SIBLING
)
8692 #ifdef CONFIG_SCHED_MC
8696 static int default_relax_domain_level
= -1;
8698 static int __init
setup_relax_domain_level(char *str
)
8702 val
= simple_strtoul(str
, NULL
, 0);
8703 if (val
< SD_LV_MAX
)
8704 default_relax_domain_level
= val
;
8708 __setup("relax_domain_level=", setup_relax_domain_level
);
8710 static void set_domain_attribute(struct sched_domain
*sd
,
8711 struct sched_domain_attr
*attr
)
8715 if (!attr
|| attr
->relax_domain_level
< 0) {
8716 if (default_relax_domain_level
< 0)
8719 request
= default_relax_domain_level
;
8721 request
= attr
->relax_domain_level
;
8722 if (request
< sd
->level
) {
8723 /* turn off idle balance on this domain */
8724 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8726 /* turn on idle balance on this domain */
8727 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8731 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8732 const struct cpumask
*cpu_map
)
8735 case sa_sched_groups
:
8736 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8737 d
->sched_group_nodes
= NULL
;
8739 free_rootdomain(d
->rd
); /* fall through */
8741 free_cpumask_var(d
->tmpmask
); /* fall through */
8742 case sa_send_covered
:
8743 free_cpumask_var(d
->send_covered
); /* fall through */
8744 case sa_this_core_map
:
8745 free_cpumask_var(d
->this_core_map
); /* fall through */
8746 case sa_this_sibling_map
:
8747 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8749 free_cpumask_var(d
->nodemask
); /* fall through */
8750 case sa_sched_group_nodes
:
8752 kfree(d
->sched_group_nodes
); /* fall through */
8754 free_cpumask_var(d
->notcovered
); /* fall through */
8756 free_cpumask_var(d
->covered
); /* fall through */
8758 free_cpumask_var(d
->domainspan
); /* fall through */
8765 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8766 const struct cpumask
*cpu_map
)
8769 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8771 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8772 return sa_domainspan
;
8773 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8775 /* Allocate the per-node list of sched groups */
8776 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8777 sizeof(struct sched_group
*), GFP_KERNEL
);
8778 if (!d
->sched_group_nodes
) {
8779 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8780 return sa_notcovered
;
8782 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8784 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8785 return sa_sched_group_nodes
;
8786 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8788 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8789 return sa_this_sibling_map
;
8790 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8791 return sa_this_core_map
;
8792 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8793 return sa_send_covered
;
8794 d
->rd
= alloc_rootdomain();
8796 printk(KERN_WARNING
"Cannot alloc root domain\n");
8799 return sa_rootdomain
;
8802 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8803 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8805 struct sched_domain
*sd
= NULL
;
8807 struct sched_domain
*parent
;
8810 if (cpumask_weight(cpu_map
) >
8811 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8812 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8813 SD_INIT(sd
, ALLNODES
);
8814 set_domain_attribute(sd
, attr
);
8815 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8816 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8821 sd
= &per_cpu(node_domains
, i
).sd
;
8823 set_domain_attribute(sd
, attr
);
8824 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8825 sd
->parent
= parent
;
8828 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8833 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8834 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8835 struct sched_domain
*parent
, int i
)
8837 struct sched_domain
*sd
;
8838 sd
= &per_cpu(phys_domains
, i
).sd
;
8840 set_domain_attribute(sd
, attr
);
8841 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8842 sd
->parent
= parent
;
8845 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8849 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8850 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8851 struct sched_domain
*parent
, int i
)
8853 struct sched_domain
*sd
= parent
;
8854 #ifdef CONFIG_SCHED_MC
8855 sd
= &per_cpu(core_domains
, i
).sd
;
8857 set_domain_attribute(sd
, attr
);
8858 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8859 sd
->parent
= parent
;
8861 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8866 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8867 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8868 struct sched_domain
*parent
, int i
)
8870 struct sched_domain
*sd
= parent
;
8871 #ifdef CONFIG_SCHED_SMT
8872 sd
= &per_cpu(cpu_domains
, i
).sd
;
8873 SD_INIT(sd
, SIBLING
);
8874 set_domain_attribute(sd
, attr
);
8875 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8876 sd
->parent
= parent
;
8878 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8883 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8884 const struct cpumask
*cpu_map
, int cpu
)
8887 #ifdef CONFIG_SCHED_SMT
8888 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8889 cpumask_and(d
->this_sibling_map
, cpu_map
,
8890 topology_thread_cpumask(cpu
));
8891 if (cpu
== cpumask_first(d
->this_sibling_map
))
8892 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8894 d
->send_covered
, d
->tmpmask
);
8897 #ifdef CONFIG_SCHED_MC
8898 case SD_LV_MC
: /* set up multi-core groups */
8899 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8900 if (cpu
== cpumask_first(d
->this_core_map
))
8901 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8903 d
->send_covered
, d
->tmpmask
);
8906 case SD_LV_CPU
: /* set up physical groups */
8907 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8908 if (!cpumask_empty(d
->nodemask
))
8909 init_sched_build_groups(d
->nodemask
, cpu_map
,
8911 d
->send_covered
, d
->tmpmask
);
8914 case SD_LV_ALLNODES
:
8915 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8916 d
->send_covered
, d
->tmpmask
);
8925 * Build sched domains for a given set of cpus and attach the sched domains
8926 * to the individual cpus
8928 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8929 struct sched_domain_attr
*attr
)
8931 enum s_alloc alloc_state
= sa_none
;
8933 struct sched_domain
*sd
;
8939 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8940 if (alloc_state
!= sa_rootdomain
)
8942 alloc_state
= sa_sched_groups
;
8945 * Set up domains for cpus specified by the cpu_map.
8947 for_each_cpu(i
, cpu_map
) {
8948 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8951 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8952 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8953 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8954 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8957 for_each_cpu(i
, cpu_map
) {
8958 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8959 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8962 /* Set up physical groups */
8963 for (i
= 0; i
< nr_node_ids
; i
++)
8964 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8967 /* Set up node groups */
8969 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8971 for (i
= 0; i
< nr_node_ids
; i
++)
8972 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8976 /* Calculate CPU power for physical packages and nodes */
8977 #ifdef CONFIG_SCHED_SMT
8978 for_each_cpu(i
, cpu_map
) {
8979 sd
= &per_cpu(cpu_domains
, i
).sd
;
8980 init_sched_groups_power(i
, sd
);
8983 #ifdef CONFIG_SCHED_MC
8984 for_each_cpu(i
, cpu_map
) {
8985 sd
= &per_cpu(core_domains
, i
).sd
;
8986 init_sched_groups_power(i
, sd
);
8990 for_each_cpu(i
, cpu_map
) {
8991 sd
= &per_cpu(phys_domains
, i
).sd
;
8992 init_sched_groups_power(i
, sd
);
8996 for (i
= 0; i
< nr_node_ids
; i
++)
8997 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8999 if (d
.sd_allnodes
) {
9000 struct sched_group
*sg
;
9002 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
9004 init_numa_sched_groups_power(sg
);
9008 /* Attach the domains */
9009 for_each_cpu(i
, cpu_map
) {
9010 #ifdef CONFIG_SCHED_SMT
9011 sd
= &per_cpu(cpu_domains
, i
).sd
;
9012 #elif defined(CONFIG_SCHED_MC)
9013 sd
= &per_cpu(core_domains
, i
).sd
;
9015 sd
= &per_cpu(phys_domains
, i
).sd
;
9017 cpu_attach_domain(sd
, d
.rd
, i
);
9020 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
9021 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
9025 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
9029 static int build_sched_domains(const struct cpumask
*cpu_map
)
9031 return __build_sched_domains(cpu_map
, NULL
);
9034 static struct cpumask
*doms_cur
; /* current sched domains */
9035 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
9036 static struct sched_domain_attr
*dattr_cur
;
9037 /* attribues of custom domains in 'doms_cur' */
9040 * Special case: If a kmalloc of a doms_cur partition (array of
9041 * cpumask) fails, then fallback to a single sched domain,
9042 * as determined by the single cpumask fallback_doms.
9044 static cpumask_var_t fallback_doms
;
9047 * arch_update_cpu_topology lets virtualized architectures update the
9048 * cpu core maps. It is supposed to return 1 if the topology changed
9049 * or 0 if it stayed the same.
9051 int __attribute__((weak
)) arch_update_cpu_topology(void)
9057 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9058 * For now this just excludes isolated cpus, but could be used to
9059 * exclude other special cases in the future.
9061 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9065 arch_update_cpu_topology();
9067 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
9069 doms_cur
= fallback_doms
;
9070 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
9072 err
= build_sched_domains(doms_cur
);
9073 register_sched_domain_sysctl();
9078 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9079 struct cpumask
*tmpmask
)
9081 free_sched_groups(cpu_map
, tmpmask
);
9085 * Detach sched domains from a group of cpus specified in cpu_map
9086 * These cpus will now be attached to the NULL domain
9088 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9090 /* Save because hotplug lock held. */
9091 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9094 for_each_cpu(i
, cpu_map
)
9095 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9096 synchronize_sched();
9097 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9100 /* handle null as "default" */
9101 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9102 struct sched_domain_attr
*new, int idx_new
)
9104 struct sched_domain_attr tmp
;
9111 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9112 new ? (new + idx_new
) : &tmp
,
9113 sizeof(struct sched_domain_attr
));
9117 * Partition sched domains as specified by the 'ndoms_new'
9118 * cpumasks in the array doms_new[] of cpumasks. This compares
9119 * doms_new[] to the current sched domain partitioning, doms_cur[].
9120 * It destroys each deleted domain and builds each new domain.
9122 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9123 * The masks don't intersect (don't overlap.) We should setup one
9124 * sched domain for each mask. CPUs not in any of the cpumasks will
9125 * not be load balanced. If the same cpumask appears both in the
9126 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9129 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9130 * ownership of it and will kfree it when done with it. If the caller
9131 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9132 * ndoms_new == 1, and partition_sched_domains() will fallback to
9133 * the single partition 'fallback_doms', it also forces the domains
9136 * If doms_new == NULL it will be replaced with cpu_online_mask.
9137 * ndoms_new == 0 is a special case for destroying existing domains,
9138 * and it will not create the default domain.
9140 * Call with hotplug lock held
9142 /* FIXME: Change to struct cpumask *doms_new[] */
9143 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9144 struct sched_domain_attr
*dattr_new
)
9149 mutex_lock(&sched_domains_mutex
);
9151 /* always unregister in case we don't destroy any domains */
9152 unregister_sched_domain_sysctl();
9154 /* Let architecture update cpu core mappings. */
9155 new_topology
= arch_update_cpu_topology();
9157 n
= doms_new
? ndoms_new
: 0;
9159 /* Destroy deleted domains */
9160 for (i
= 0; i
< ndoms_cur
; i
++) {
9161 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9162 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9163 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9166 /* no match - a current sched domain not in new doms_new[] */
9167 detach_destroy_domains(doms_cur
+ i
);
9172 if (doms_new
== NULL
) {
9174 doms_new
= fallback_doms
;
9175 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
9176 WARN_ON_ONCE(dattr_new
);
9179 /* Build new domains */
9180 for (i
= 0; i
< ndoms_new
; i
++) {
9181 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9182 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9183 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9186 /* no match - add a new doms_new */
9187 __build_sched_domains(doms_new
+ i
,
9188 dattr_new
? dattr_new
+ i
: NULL
);
9193 /* Remember the new sched domains */
9194 if (doms_cur
!= fallback_doms
)
9196 kfree(dattr_cur
); /* kfree(NULL) is safe */
9197 doms_cur
= doms_new
;
9198 dattr_cur
= dattr_new
;
9199 ndoms_cur
= ndoms_new
;
9201 register_sched_domain_sysctl();
9203 mutex_unlock(&sched_domains_mutex
);
9206 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9207 static void arch_reinit_sched_domains(void)
9211 /* Destroy domains first to force the rebuild */
9212 partition_sched_domains(0, NULL
, NULL
);
9214 rebuild_sched_domains();
9218 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9220 unsigned int level
= 0;
9222 if (sscanf(buf
, "%u", &level
) != 1)
9226 * level is always be positive so don't check for
9227 * level < POWERSAVINGS_BALANCE_NONE which is 0
9228 * What happens on 0 or 1 byte write,
9229 * need to check for count as well?
9232 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9236 sched_smt_power_savings
= level
;
9238 sched_mc_power_savings
= level
;
9240 arch_reinit_sched_domains();
9245 #ifdef CONFIG_SCHED_MC
9246 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9249 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9251 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9252 const char *buf
, size_t count
)
9254 return sched_power_savings_store(buf
, count
, 0);
9256 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9257 sched_mc_power_savings_show
,
9258 sched_mc_power_savings_store
);
9261 #ifdef CONFIG_SCHED_SMT
9262 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9265 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9267 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9268 const char *buf
, size_t count
)
9270 return sched_power_savings_store(buf
, count
, 1);
9272 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9273 sched_smt_power_savings_show
,
9274 sched_smt_power_savings_store
);
9277 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9281 #ifdef CONFIG_SCHED_SMT
9283 err
= sysfs_create_file(&cls
->kset
.kobj
,
9284 &attr_sched_smt_power_savings
.attr
);
9286 #ifdef CONFIG_SCHED_MC
9287 if (!err
&& mc_capable())
9288 err
= sysfs_create_file(&cls
->kset
.kobj
,
9289 &attr_sched_mc_power_savings
.attr
);
9293 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9295 #ifndef CONFIG_CPUSETS
9297 * Add online and remove offline CPUs from the scheduler domains.
9298 * When cpusets are enabled they take over this function.
9300 static int update_sched_domains(struct notifier_block
*nfb
,
9301 unsigned long action
, void *hcpu
)
9305 case CPU_ONLINE_FROZEN
:
9307 case CPU_DEAD_FROZEN
:
9308 partition_sched_domains(1, NULL
, NULL
);
9317 static int update_runtime(struct notifier_block
*nfb
,
9318 unsigned long action
, void *hcpu
)
9320 int cpu
= (int)(long)hcpu
;
9323 case CPU_DOWN_PREPARE
:
9324 case CPU_DOWN_PREPARE_FROZEN
:
9325 disable_runtime(cpu_rq(cpu
));
9328 case CPU_DOWN_FAILED
:
9329 case CPU_DOWN_FAILED_FROZEN
:
9331 case CPU_ONLINE_FROZEN
:
9332 enable_runtime(cpu_rq(cpu
));
9340 void __init
sched_init_smp(void)
9342 cpumask_var_t non_isolated_cpus
;
9344 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9346 #if defined(CONFIG_NUMA)
9347 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9349 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9352 mutex_lock(&sched_domains_mutex
);
9353 arch_init_sched_domains(cpu_online_mask
);
9354 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9355 if (cpumask_empty(non_isolated_cpus
))
9356 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9357 mutex_unlock(&sched_domains_mutex
);
9360 #ifndef CONFIG_CPUSETS
9361 /* XXX: Theoretical race here - CPU may be hotplugged now */
9362 hotcpu_notifier(update_sched_domains
, 0);
9365 /* RT runtime code needs to handle some hotplug events */
9366 hotcpu_notifier(update_runtime
, 0);
9370 /* Move init over to a non-isolated CPU */
9371 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9373 sched_init_granularity();
9374 free_cpumask_var(non_isolated_cpus
);
9376 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9377 init_sched_rt_class();
9380 void __init
sched_init_smp(void)
9382 sched_init_granularity();
9384 #endif /* CONFIG_SMP */
9386 const_debug
unsigned int sysctl_timer_migration
= 1;
9388 int in_sched_functions(unsigned long addr
)
9390 return in_lock_functions(addr
) ||
9391 (addr
>= (unsigned long)__sched_text_start
9392 && addr
< (unsigned long)__sched_text_end
);
9395 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9397 cfs_rq
->tasks_timeline
= RB_ROOT
;
9398 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9399 #ifdef CONFIG_FAIR_GROUP_SCHED
9402 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9405 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9407 struct rt_prio_array
*array
;
9410 array
= &rt_rq
->active
;
9411 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9412 INIT_LIST_HEAD(array
->queue
+ i
);
9413 __clear_bit(i
, array
->bitmap
);
9415 /* delimiter for bitsearch: */
9416 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9418 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9419 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9421 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9425 rt_rq
->rt_nr_migratory
= 0;
9426 rt_rq
->overloaded
= 0;
9427 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9431 rt_rq
->rt_throttled
= 0;
9432 rt_rq
->rt_runtime
= 0;
9433 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9435 #ifdef CONFIG_RT_GROUP_SCHED
9436 rt_rq
->rt_nr_boosted
= 0;
9441 #ifdef CONFIG_FAIR_GROUP_SCHED
9442 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9443 struct sched_entity
*se
, int cpu
, int add
,
9444 struct sched_entity
*parent
)
9446 struct rq
*rq
= cpu_rq(cpu
);
9447 tg
->cfs_rq
[cpu
] = cfs_rq
;
9448 init_cfs_rq(cfs_rq
, rq
);
9451 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9454 /* se could be NULL for init_task_group */
9459 se
->cfs_rq
= &rq
->cfs
;
9461 se
->cfs_rq
= parent
->my_q
;
9464 se
->load
.weight
= tg
->shares
;
9465 se
->load
.inv_weight
= 0;
9466 se
->parent
= parent
;
9470 #ifdef CONFIG_RT_GROUP_SCHED
9471 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9472 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9473 struct sched_rt_entity
*parent
)
9475 struct rq
*rq
= cpu_rq(cpu
);
9477 tg
->rt_rq
[cpu
] = rt_rq
;
9478 init_rt_rq(rt_rq
, rq
);
9480 rt_rq
->rt_se
= rt_se
;
9481 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9483 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9485 tg
->rt_se
[cpu
] = rt_se
;
9490 rt_se
->rt_rq
= &rq
->rt
;
9492 rt_se
->rt_rq
= parent
->my_q
;
9494 rt_se
->my_q
= rt_rq
;
9495 rt_se
->parent
= parent
;
9496 INIT_LIST_HEAD(&rt_se
->run_list
);
9500 void __init
sched_init(void)
9503 unsigned long alloc_size
= 0, ptr
;
9505 #ifdef CONFIG_FAIR_GROUP_SCHED
9506 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9508 #ifdef CONFIG_RT_GROUP_SCHED
9509 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9511 #ifdef CONFIG_USER_SCHED
9514 #ifdef CONFIG_CPUMASK_OFFSTACK
9515 alloc_size
+= num_possible_cpus() * cpumask_size();
9518 * As sched_init() is called before page_alloc is setup,
9519 * we use alloc_bootmem().
9522 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9524 #ifdef CONFIG_FAIR_GROUP_SCHED
9525 init_task_group
.se
= (struct sched_entity
**)ptr
;
9526 ptr
+= nr_cpu_ids
* sizeof(void **);
9528 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9529 ptr
+= nr_cpu_ids
* sizeof(void **);
9531 #ifdef CONFIG_USER_SCHED
9532 root_task_group
.se
= (struct sched_entity
**)ptr
;
9533 ptr
+= nr_cpu_ids
* sizeof(void **);
9535 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9536 ptr
+= nr_cpu_ids
* sizeof(void **);
9537 #endif /* CONFIG_USER_SCHED */
9538 #endif /* CONFIG_FAIR_GROUP_SCHED */
9539 #ifdef CONFIG_RT_GROUP_SCHED
9540 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9541 ptr
+= nr_cpu_ids
* sizeof(void **);
9543 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9544 ptr
+= nr_cpu_ids
* sizeof(void **);
9546 #ifdef CONFIG_USER_SCHED
9547 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9548 ptr
+= nr_cpu_ids
* sizeof(void **);
9550 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9551 ptr
+= nr_cpu_ids
* sizeof(void **);
9552 #endif /* CONFIG_USER_SCHED */
9553 #endif /* CONFIG_RT_GROUP_SCHED */
9554 #ifdef CONFIG_CPUMASK_OFFSTACK
9555 for_each_possible_cpu(i
) {
9556 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9557 ptr
+= cpumask_size();
9559 #endif /* CONFIG_CPUMASK_OFFSTACK */
9563 init_defrootdomain();
9566 init_rt_bandwidth(&def_rt_bandwidth
,
9567 global_rt_period(), global_rt_runtime());
9569 #ifdef CONFIG_RT_GROUP_SCHED
9570 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9571 global_rt_period(), global_rt_runtime());
9572 #ifdef CONFIG_USER_SCHED
9573 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9574 global_rt_period(), RUNTIME_INF
);
9575 #endif /* CONFIG_USER_SCHED */
9576 #endif /* CONFIG_RT_GROUP_SCHED */
9578 #ifdef CONFIG_GROUP_SCHED
9579 list_add(&init_task_group
.list
, &task_groups
);
9580 INIT_LIST_HEAD(&init_task_group
.children
);
9582 #ifdef CONFIG_USER_SCHED
9583 INIT_LIST_HEAD(&root_task_group
.children
);
9584 init_task_group
.parent
= &root_task_group
;
9585 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9586 #endif /* CONFIG_USER_SCHED */
9587 #endif /* CONFIG_GROUP_SCHED */
9589 for_each_possible_cpu(i
) {
9593 spin_lock_init(&rq
->lock
);
9595 rq
->calc_load_active
= 0;
9596 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9597 init_cfs_rq(&rq
->cfs
, rq
);
9598 init_rt_rq(&rq
->rt
, rq
);
9599 #ifdef CONFIG_FAIR_GROUP_SCHED
9600 init_task_group
.shares
= init_task_group_load
;
9601 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9602 #ifdef CONFIG_CGROUP_SCHED
9604 * How much cpu bandwidth does init_task_group get?
9606 * In case of task-groups formed thr' the cgroup filesystem, it
9607 * gets 100% of the cpu resources in the system. This overall
9608 * system cpu resource is divided among the tasks of
9609 * init_task_group and its child task-groups in a fair manner,
9610 * based on each entity's (task or task-group's) weight
9611 * (se->load.weight).
9613 * In other words, if init_task_group has 10 tasks of weight
9614 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9615 * then A0's share of the cpu resource is:
9617 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9619 * We achieve this by letting init_task_group's tasks sit
9620 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9622 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9623 #elif defined CONFIG_USER_SCHED
9624 root_task_group
.shares
= NICE_0_LOAD
;
9625 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9627 * In case of task-groups formed thr' the user id of tasks,
9628 * init_task_group represents tasks belonging to root user.
9629 * Hence it forms a sibling of all subsequent groups formed.
9630 * In this case, init_task_group gets only a fraction of overall
9631 * system cpu resource, based on the weight assigned to root
9632 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9633 * by letting tasks of init_task_group sit in a separate cfs_rq
9634 * (init_tg_cfs_rq) and having one entity represent this group of
9635 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9637 init_tg_cfs_entry(&init_task_group
,
9638 &per_cpu(init_tg_cfs_rq
, i
),
9639 &per_cpu(init_sched_entity
, i
), i
, 1,
9640 root_task_group
.se
[i
]);
9643 #endif /* CONFIG_FAIR_GROUP_SCHED */
9645 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9646 #ifdef CONFIG_RT_GROUP_SCHED
9647 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9648 #ifdef CONFIG_CGROUP_SCHED
9649 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9650 #elif defined CONFIG_USER_SCHED
9651 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9652 init_tg_rt_entry(&init_task_group
,
9653 &per_cpu(init_rt_rq
, i
),
9654 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9655 root_task_group
.rt_se
[i
]);
9659 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9660 rq
->cpu_load
[j
] = 0;
9664 rq
->post_schedule
= 0;
9665 rq
->active_balance
= 0;
9666 rq
->next_balance
= jiffies
;
9670 rq
->migration_thread
= NULL
;
9671 INIT_LIST_HEAD(&rq
->migration_queue
);
9672 rq_attach_root(rq
, &def_root_domain
);
9675 atomic_set(&rq
->nr_iowait
, 0);
9678 set_load_weight(&init_task
);
9680 #ifdef CONFIG_PREEMPT_NOTIFIERS
9681 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9685 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9688 #ifdef CONFIG_RT_MUTEXES
9689 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9693 * The boot idle thread does lazy MMU switching as well:
9695 atomic_inc(&init_mm
.mm_count
);
9696 enter_lazy_tlb(&init_mm
, current
);
9699 * Make us the idle thread. Technically, schedule() should not be
9700 * called from this thread, however somewhere below it might be,
9701 * but because we are the idle thread, we just pick up running again
9702 * when this runqueue becomes "idle".
9704 init_idle(current
, smp_processor_id());
9706 calc_load_update
= jiffies
+ LOAD_FREQ
;
9709 * During early bootup we pretend to be a normal task:
9711 current
->sched_class
= &fair_sched_class
;
9713 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9714 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9717 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9718 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9720 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9723 perf_counter_init();
9725 scheduler_running
= 1;
9728 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9729 static inline int preempt_count_equals(int preempt_offset
)
9731 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9733 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9736 void __might_sleep(char *file
, int line
, int preempt_offset
)
9739 static unsigned long prev_jiffy
; /* ratelimiting */
9741 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9742 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9744 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9746 prev_jiffy
= jiffies
;
9749 "BUG: sleeping function called from invalid context at %s:%d\n",
9752 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9753 in_atomic(), irqs_disabled(),
9754 current
->pid
, current
->comm
);
9756 debug_show_held_locks(current
);
9757 if (irqs_disabled())
9758 print_irqtrace_events(current
);
9762 EXPORT_SYMBOL(__might_sleep
);
9765 #ifdef CONFIG_MAGIC_SYSRQ
9766 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9770 update_rq_clock(rq
);
9771 on_rq
= p
->se
.on_rq
;
9773 deactivate_task(rq
, p
, 0);
9774 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9776 activate_task(rq
, p
, 0);
9777 resched_task(rq
->curr
);
9781 void normalize_rt_tasks(void)
9783 struct task_struct
*g
, *p
;
9784 unsigned long flags
;
9787 read_lock_irqsave(&tasklist_lock
, flags
);
9788 do_each_thread(g
, p
) {
9790 * Only normalize user tasks:
9795 p
->se
.exec_start
= 0;
9796 #ifdef CONFIG_SCHEDSTATS
9797 p
->se
.wait_start
= 0;
9798 p
->se
.sleep_start
= 0;
9799 p
->se
.block_start
= 0;
9804 * Renice negative nice level userspace
9807 if (TASK_NICE(p
) < 0 && p
->mm
)
9808 set_user_nice(p
, 0);
9812 spin_lock(&p
->pi_lock
);
9813 rq
= __task_rq_lock(p
);
9815 normalize_task(rq
, p
);
9817 __task_rq_unlock(rq
);
9818 spin_unlock(&p
->pi_lock
);
9819 } while_each_thread(g
, p
);
9821 read_unlock_irqrestore(&tasklist_lock
, flags
);
9824 #endif /* CONFIG_MAGIC_SYSRQ */
9828 * These functions are only useful for the IA64 MCA handling.
9830 * They can only be called when the whole system has been
9831 * stopped - every CPU needs to be quiescent, and no scheduling
9832 * activity can take place. Using them for anything else would
9833 * be a serious bug, and as a result, they aren't even visible
9834 * under any other configuration.
9838 * curr_task - return the current task for a given cpu.
9839 * @cpu: the processor in question.
9841 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9843 struct task_struct
*curr_task(int cpu
)
9845 return cpu_curr(cpu
);
9849 * set_curr_task - set the current task for a given cpu.
9850 * @cpu: the processor in question.
9851 * @p: the task pointer to set.
9853 * Description: This function must only be used when non-maskable interrupts
9854 * are serviced on a separate stack. It allows the architecture to switch the
9855 * notion of the current task on a cpu in a non-blocking manner. This function
9856 * must be called with all CPU's synchronized, and interrupts disabled, the
9857 * and caller must save the original value of the current task (see
9858 * curr_task() above) and restore that value before reenabling interrupts and
9859 * re-starting the system.
9861 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9863 void set_curr_task(int cpu
, struct task_struct
*p
)
9870 #ifdef CONFIG_FAIR_GROUP_SCHED
9871 static void free_fair_sched_group(struct task_group
*tg
)
9875 for_each_possible_cpu(i
) {
9877 kfree(tg
->cfs_rq
[i
]);
9887 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9889 struct cfs_rq
*cfs_rq
;
9890 struct sched_entity
*se
;
9894 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9897 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9901 tg
->shares
= NICE_0_LOAD
;
9903 for_each_possible_cpu(i
) {
9906 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9907 GFP_KERNEL
, cpu_to_node(i
));
9911 se
= kzalloc_node(sizeof(struct sched_entity
),
9912 GFP_KERNEL
, cpu_to_node(i
));
9916 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9925 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9927 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9928 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9931 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9933 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9935 #else /* !CONFG_FAIR_GROUP_SCHED */
9936 static inline void free_fair_sched_group(struct task_group
*tg
)
9941 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9946 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9950 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9953 #endif /* CONFIG_FAIR_GROUP_SCHED */
9955 #ifdef CONFIG_RT_GROUP_SCHED
9956 static void free_rt_sched_group(struct task_group
*tg
)
9960 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9962 for_each_possible_cpu(i
) {
9964 kfree(tg
->rt_rq
[i
]);
9966 kfree(tg
->rt_se
[i
]);
9974 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9976 struct rt_rq
*rt_rq
;
9977 struct sched_rt_entity
*rt_se
;
9981 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9984 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9988 init_rt_bandwidth(&tg
->rt_bandwidth
,
9989 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9991 for_each_possible_cpu(i
) {
9994 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9995 GFP_KERNEL
, cpu_to_node(i
));
9999 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
10000 GFP_KERNEL
, cpu_to_node(i
));
10004 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
10013 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10015 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
10016 &cpu_rq(cpu
)->leaf_rt_rq_list
);
10019 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10021 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
10023 #else /* !CONFIG_RT_GROUP_SCHED */
10024 static inline void free_rt_sched_group(struct task_group
*tg
)
10029 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10034 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10038 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10041 #endif /* CONFIG_RT_GROUP_SCHED */
10043 #ifdef CONFIG_GROUP_SCHED
10044 static void free_sched_group(struct task_group
*tg
)
10046 free_fair_sched_group(tg
);
10047 free_rt_sched_group(tg
);
10051 /* allocate runqueue etc for a new task group */
10052 struct task_group
*sched_create_group(struct task_group
*parent
)
10054 struct task_group
*tg
;
10055 unsigned long flags
;
10058 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10060 return ERR_PTR(-ENOMEM
);
10062 if (!alloc_fair_sched_group(tg
, parent
))
10065 if (!alloc_rt_sched_group(tg
, parent
))
10068 spin_lock_irqsave(&task_group_lock
, flags
);
10069 for_each_possible_cpu(i
) {
10070 register_fair_sched_group(tg
, i
);
10071 register_rt_sched_group(tg
, i
);
10073 list_add_rcu(&tg
->list
, &task_groups
);
10075 WARN_ON(!parent
); /* root should already exist */
10077 tg
->parent
= parent
;
10078 INIT_LIST_HEAD(&tg
->children
);
10079 list_add_rcu(&tg
->siblings
, &parent
->children
);
10080 spin_unlock_irqrestore(&task_group_lock
, flags
);
10085 free_sched_group(tg
);
10086 return ERR_PTR(-ENOMEM
);
10089 /* rcu callback to free various structures associated with a task group */
10090 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10092 /* now it should be safe to free those cfs_rqs */
10093 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10096 /* Destroy runqueue etc associated with a task group */
10097 void sched_destroy_group(struct task_group
*tg
)
10099 unsigned long flags
;
10102 spin_lock_irqsave(&task_group_lock
, flags
);
10103 for_each_possible_cpu(i
) {
10104 unregister_fair_sched_group(tg
, i
);
10105 unregister_rt_sched_group(tg
, i
);
10107 list_del_rcu(&tg
->list
);
10108 list_del_rcu(&tg
->siblings
);
10109 spin_unlock_irqrestore(&task_group_lock
, flags
);
10111 /* wait for possible concurrent references to cfs_rqs complete */
10112 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10115 /* change task's runqueue when it moves between groups.
10116 * The caller of this function should have put the task in its new group
10117 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10118 * reflect its new group.
10120 void sched_move_task(struct task_struct
*tsk
)
10122 int on_rq
, running
;
10123 unsigned long flags
;
10126 rq
= task_rq_lock(tsk
, &flags
);
10128 update_rq_clock(rq
);
10130 running
= task_current(rq
, tsk
);
10131 on_rq
= tsk
->se
.on_rq
;
10134 dequeue_task(rq
, tsk
, 0);
10135 if (unlikely(running
))
10136 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10138 set_task_rq(tsk
, task_cpu(tsk
));
10140 #ifdef CONFIG_FAIR_GROUP_SCHED
10141 if (tsk
->sched_class
->moved_group
)
10142 tsk
->sched_class
->moved_group(tsk
);
10145 if (unlikely(running
))
10146 tsk
->sched_class
->set_curr_task(rq
);
10148 enqueue_task(rq
, tsk
, 0);
10150 task_rq_unlock(rq
, &flags
);
10152 #endif /* CONFIG_GROUP_SCHED */
10154 #ifdef CONFIG_FAIR_GROUP_SCHED
10155 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10157 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10162 dequeue_entity(cfs_rq
, se
, 0);
10164 se
->load
.weight
= shares
;
10165 se
->load
.inv_weight
= 0;
10168 enqueue_entity(cfs_rq
, se
, 0);
10171 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10173 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10174 struct rq
*rq
= cfs_rq
->rq
;
10175 unsigned long flags
;
10177 spin_lock_irqsave(&rq
->lock
, flags
);
10178 __set_se_shares(se
, shares
);
10179 spin_unlock_irqrestore(&rq
->lock
, flags
);
10182 static DEFINE_MUTEX(shares_mutex
);
10184 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10187 unsigned long flags
;
10190 * We can't change the weight of the root cgroup.
10195 if (shares
< MIN_SHARES
)
10196 shares
= MIN_SHARES
;
10197 else if (shares
> MAX_SHARES
)
10198 shares
= MAX_SHARES
;
10200 mutex_lock(&shares_mutex
);
10201 if (tg
->shares
== shares
)
10204 spin_lock_irqsave(&task_group_lock
, flags
);
10205 for_each_possible_cpu(i
)
10206 unregister_fair_sched_group(tg
, i
);
10207 list_del_rcu(&tg
->siblings
);
10208 spin_unlock_irqrestore(&task_group_lock
, flags
);
10210 /* wait for any ongoing reference to this group to finish */
10211 synchronize_sched();
10214 * Now we are free to modify the group's share on each cpu
10215 * w/o tripping rebalance_share or load_balance_fair.
10217 tg
->shares
= shares
;
10218 for_each_possible_cpu(i
) {
10220 * force a rebalance
10222 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10223 set_se_shares(tg
->se
[i
], shares
);
10227 * Enable load balance activity on this group, by inserting it back on
10228 * each cpu's rq->leaf_cfs_rq_list.
10230 spin_lock_irqsave(&task_group_lock
, flags
);
10231 for_each_possible_cpu(i
)
10232 register_fair_sched_group(tg
, i
);
10233 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10234 spin_unlock_irqrestore(&task_group_lock
, flags
);
10236 mutex_unlock(&shares_mutex
);
10240 unsigned long sched_group_shares(struct task_group
*tg
)
10246 #ifdef CONFIG_RT_GROUP_SCHED
10248 * Ensure that the real time constraints are schedulable.
10250 static DEFINE_MUTEX(rt_constraints_mutex
);
10252 static unsigned long to_ratio(u64 period
, u64 runtime
)
10254 if (runtime
== RUNTIME_INF
)
10257 return div64_u64(runtime
<< 20, period
);
10260 /* Must be called with tasklist_lock held */
10261 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10263 struct task_struct
*g
, *p
;
10265 do_each_thread(g
, p
) {
10266 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10268 } while_each_thread(g
, p
);
10273 struct rt_schedulable_data
{
10274 struct task_group
*tg
;
10279 static int tg_schedulable(struct task_group
*tg
, void *data
)
10281 struct rt_schedulable_data
*d
= data
;
10282 struct task_group
*child
;
10283 unsigned long total
, sum
= 0;
10284 u64 period
, runtime
;
10286 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10287 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10290 period
= d
->rt_period
;
10291 runtime
= d
->rt_runtime
;
10294 #ifdef CONFIG_USER_SCHED
10295 if (tg
== &root_task_group
) {
10296 period
= global_rt_period();
10297 runtime
= global_rt_runtime();
10302 * Cannot have more runtime than the period.
10304 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10308 * Ensure we don't starve existing RT tasks.
10310 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10313 total
= to_ratio(period
, runtime
);
10316 * Nobody can have more than the global setting allows.
10318 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10322 * The sum of our children's runtime should not exceed our own.
10324 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10325 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10326 runtime
= child
->rt_bandwidth
.rt_runtime
;
10328 if (child
== d
->tg
) {
10329 period
= d
->rt_period
;
10330 runtime
= d
->rt_runtime
;
10333 sum
+= to_ratio(period
, runtime
);
10342 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10344 struct rt_schedulable_data data
= {
10346 .rt_period
= period
,
10347 .rt_runtime
= runtime
,
10350 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10353 static int tg_set_bandwidth(struct task_group
*tg
,
10354 u64 rt_period
, u64 rt_runtime
)
10358 mutex_lock(&rt_constraints_mutex
);
10359 read_lock(&tasklist_lock
);
10360 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10364 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10365 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10366 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10368 for_each_possible_cpu(i
) {
10369 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10371 spin_lock(&rt_rq
->rt_runtime_lock
);
10372 rt_rq
->rt_runtime
= rt_runtime
;
10373 spin_unlock(&rt_rq
->rt_runtime_lock
);
10375 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10377 read_unlock(&tasklist_lock
);
10378 mutex_unlock(&rt_constraints_mutex
);
10383 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10385 u64 rt_runtime
, rt_period
;
10387 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10388 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10389 if (rt_runtime_us
< 0)
10390 rt_runtime
= RUNTIME_INF
;
10392 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10395 long sched_group_rt_runtime(struct task_group
*tg
)
10399 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10402 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10403 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10404 return rt_runtime_us
;
10407 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10409 u64 rt_runtime
, rt_period
;
10411 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10412 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10414 if (rt_period
== 0)
10417 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10420 long sched_group_rt_period(struct task_group
*tg
)
10424 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10425 do_div(rt_period_us
, NSEC_PER_USEC
);
10426 return rt_period_us
;
10429 static int sched_rt_global_constraints(void)
10431 u64 runtime
, period
;
10434 if (sysctl_sched_rt_period
<= 0)
10437 runtime
= global_rt_runtime();
10438 period
= global_rt_period();
10441 * Sanity check on the sysctl variables.
10443 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10446 mutex_lock(&rt_constraints_mutex
);
10447 read_lock(&tasklist_lock
);
10448 ret
= __rt_schedulable(NULL
, 0, 0);
10449 read_unlock(&tasklist_lock
);
10450 mutex_unlock(&rt_constraints_mutex
);
10455 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10457 /* Don't accept realtime tasks when there is no way for them to run */
10458 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10464 #else /* !CONFIG_RT_GROUP_SCHED */
10465 static int sched_rt_global_constraints(void)
10467 unsigned long flags
;
10470 if (sysctl_sched_rt_period
<= 0)
10474 * There's always some RT tasks in the root group
10475 * -- migration, kstopmachine etc..
10477 if (sysctl_sched_rt_runtime
== 0)
10480 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10481 for_each_possible_cpu(i
) {
10482 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10484 spin_lock(&rt_rq
->rt_runtime_lock
);
10485 rt_rq
->rt_runtime
= global_rt_runtime();
10486 spin_unlock(&rt_rq
->rt_runtime_lock
);
10488 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10492 #endif /* CONFIG_RT_GROUP_SCHED */
10494 int sched_rt_handler(struct ctl_table
*table
, int write
,
10495 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10499 int old_period
, old_runtime
;
10500 static DEFINE_MUTEX(mutex
);
10502 mutex_lock(&mutex
);
10503 old_period
= sysctl_sched_rt_period
;
10504 old_runtime
= sysctl_sched_rt_runtime
;
10506 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10508 if (!ret
&& write
) {
10509 ret
= sched_rt_global_constraints();
10511 sysctl_sched_rt_period
= old_period
;
10512 sysctl_sched_rt_runtime
= old_runtime
;
10514 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10515 def_rt_bandwidth
.rt_period
=
10516 ns_to_ktime(global_rt_period());
10519 mutex_unlock(&mutex
);
10524 #ifdef CONFIG_CGROUP_SCHED
10526 /* return corresponding task_group object of a cgroup */
10527 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10529 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10530 struct task_group
, css
);
10533 static struct cgroup_subsys_state
*
10534 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10536 struct task_group
*tg
, *parent
;
10538 if (!cgrp
->parent
) {
10539 /* This is early initialization for the top cgroup */
10540 return &init_task_group
.css
;
10543 parent
= cgroup_tg(cgrp
->parent
);
10544 tg
= sched_create_group(parent
);
10546 return ERR_PTR(-ENOMEM
);
10552 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10554 struct task_group
*tg
= cgroup_tg(cgrp
);
10556 sched_destroy_group(tg
);
10560 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10561 struct task_struct
*tsk
)
10563 #ifdef CONFIG_RT_GROUP_SCHED
10564 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10567 /* We don't support RT-tasks being in separate groups */
10568 if (tsk
->sched_class
!= &fair_sched_class
)
10576 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10577 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10579 sched_move_task(tsk
);
10582 #ifdef CONFIG_FAIR_GROUP_SCHED
10583 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10586 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10589 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10591 struct task_group
*tg
= cgroup_tg(cgrp
);
10593 return (u64
) tg
->shares
;
10595 #endif /* CONFIG_FAIR_GROUP_SCHED */
10597 #ifdef CONFIG_RT_GROUP_SCHED
10598 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10601 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10604 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10606 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10609 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10612 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10615 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10617 return sched_group_rt_period(cgroup_tg(cgrp
));
10619 #endif /* CONFIG_RT_GROUP_SCHED */
10621 static struct cftype cpu_files
[] = {
10622 #ifdef CONFIG_FAIR_GROUP_SCHED
10625 .read_u64
= cpu_shares_read_u64
,
10626 .write_u64
= cpu_shares_write_u64
,
10629 #ifdef CONFIG_RT_GROUP_SCHED
10631 .name
= "rt_runtime_us",
10632 .read_s64
= cpu_rt_runtime_read
,
10633 .write_s64
= cpu_rt_runtime_write
,
10636 .name
= "rt_period_us",
10637 .read_u64
= cpu_rt_period_read_uint
,
10638 .write_u64
= cpu_rt_period_write_uint
,
10643 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10645 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10648 struct cgroup_subsys cpu_cgroup_subsys
= {
10650 .create
= cpu_cgroup_create
,
10651 .destroy
= cpu_cgroup_destroy
,
10652 .can_attach
= cpu_cgroup_can_attach
,
10653 .attach
= cpu_cgroup_attach
,
10654 .populate
= cpu_cgroup_populate
,
10655 .subsys_id
= cpu_cgroup_subsys_id
,
10659 #endif /* CONFIG_CGROUP_SCHED */
10661 #ifdef CONFIG_CGROUP_CPUACCT
10664 * CPU accounting code for task groups.
10666 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10667 * (balbir@in.ibm.com).
10670 /* track cpu usage of a group of tasks and its child groups */
10672 struct cgroup_subsys_state css
;
10673 /* cpuusage holds pointer to a u64-type object on every cpu */
10675 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10676 struct cpuacct
*parent
;
10679 struct cgroup_subsys cpuacct_subsys
;
10681 /* return cpu accounting group corresponding to this container */
10682 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10684 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10685 struct cpuacct
, css
);
10688 /* return cpu accounting group to which this task belongs */
10689 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10691 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10692 struct cpuacct
, css
);
10695 /* create a new cpu accounting group */
10696 static struct cgroup_subsys_state
*cpuacct_create(
10697 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10699 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10705 ca
->cpuusage
= alloc_percpu(u64
);
10709 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10710 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10711 goto out_free_counters
;
10714 ca
->parent
= cgroup_ca(cgrp
->parent
);
10720 percpu_counter_destroy(&ca
->cpustat
[i
]);
10721 free_percpu(ca
->cpuusage
);
10725 return ERR_PTR(-ENOMEM
);
10728 /* destroy an existing cpu accounting group */
10730 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10732 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10735 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10736 percpu_counter_destroy(&ca
->cpustat
[i
]);
10737 free_percpu(ca
->cpuusage
);
10741 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10743 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10746 #ifndef CONFIG_64BIT
10748 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10750 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10752 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10760 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10762 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10764 #ifndef CONFIG_64BIT
10766 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10768 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10770 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10776 /* return total cpu usage (in nanoseconds) of a group */
10777 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10779 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10780 u64 totalcpuusage
= 0;
10783 for_each_present_cpu(i
)
10784 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10786 return totalcpuusage
;
10789 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10792 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10801 for_each_present_cpu(i
)
10802 cpuacct_cpuusage_write(ca
, i
, 0);
10808 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10809 struct seq_file
*m
)
10811 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10815 for_each_present_cpu(i
) {
10816 percpu
= cpuacct_cpuusage_read(ca
, i
);
10817 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10819 seq_printf(m
, "\n");
10823 static const char *cpuacct_stat_desc
[] = {
10824 [CPUACCT_STAT_USER
] = "user",
10825 [CPUACCT_STAT_SYSTEM
] = "system",
10828 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10829 struct cgroup_map_cb
*cb
)
10831 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10834 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10835 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10836 val
= cputime64_to_clock_t(val
);
10837 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10842 static struct cftype files
[] = {
10845 .read_u64
= cpuusage_read
,
10846 .write_u64
= cpuusage_write
,
10849 .name
= "usage_percpu",
10850 .read_seq_string
= cpuacct_percpu_seq_read
,
10854 .read_map
= cpuacct_stats_show
,
10858 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10860 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10864 * charge this task's execution time to its accounting group.
10866 * called with rq->lock held.
10868 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10870 struct cpuacct
*ca
;
10873 if (unlikely(!cpuacct_subsys
.active
))
10876 cpu
= task_cpu(tsk
);
10882 for (; ca
; ca
= ca
->parent
) {
10883 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10884 *cpuusage
+= cputime
;
10891 * Charge the system/user time to the task's accounting group.
10893 static void cpuacct_update_stats(struct task_struct
*tsk
,
10894 enum cpuacct_stat_index idx
, cputime_t val
)
10896 struct cpuacct
*ca
;
10898 if (unlikely(!cpuacct_subsys
.active
))
10905 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10911 struct cgroup_subsys cpuacct_subsys
= {
10913 .create
= cpuacct_create
,
10914 .destroy
= cpuacct_destroy
,
10915 .populate
= cpuacct_populate
,
10916 .subsys_id
= cpuacct_subsys_id
,
10918 #endif /* CONFIG_CGROUP_CPUACCT */