4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity
**se
;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq
**cfs_rq
;
259 unsigned long shares
;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity
**rt_se
;
264 struct rt_rq
**rt_rq
;
266 struct rt_bandwidth rt_bandwidth
;
270 struct list_head list
;
272 struct task_group
*parent
;
273 struct list_head siblings
;
274 struct list_head children
;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct
*user
)
282 user
->tg
->uid
= user
->uid
;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group
;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq
, init_tg_cfs_rq
);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq
, init_rt_rq
);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock
);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group
.children
);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group
;
345 /* return group to which a task belongs */
346 static inline struct task_group
*task_group(struct task_struct
*p
)
348 struct task_group
*tg
;
350 #ifdef CONFIG_USER_SCHED
352 tg
= __task_cred(p
)->user
->tg
;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
356 struct task_group
, css
);
358 tg
= &init_task_group
;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
368 p
->se
.parent
= task_group(p
)->se
[cpu
];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
373 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
379 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
380 static inline struct task_group
*task_group(struct task_struct
*p
)
385 #endif /* CONFIG_GROUP_SCHED */
387 /* CFS-related fields in a runqueue */
389 struct load_weight load
;
390 unsigned long nr_running
;
395 struct rb_root tasks_timeline
;
396 struct rb_node
*rb_leftmost
;
398 struct list_head tasks
;
399 struct list_head
*balance_iterator
;
402 * 'curr' points to currently running entity on this cfs_rq.
403 * It is set to NULL otherwise (i.e when none are currently running).
405 struct sched_entity
*curr
, *next
, *last
;
407 unsigned int nr_spread_over
;
409 #ifdef CONFIG_FAIR_GROUP_SCHED
410 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
420 struct list_head leaf_cfs_rq_list
;
421 struct task_group
*tg
; /* group that "owns" this runqueue */
425 * the part of load.weight contributed by tasks
427 unsigned long task_weight
;
430 * h_load = weight * f(tg)
432 * Where f(tg) is the recursive weight fraction assigned to
435 unsigned long h_load
;
438 * this cpu's part of tg->shares
440 unsigned long shares
;
443 * load.weight at the time we set shares
445 unsigned long rq_weight
;
450 /* Real-Time classes' related field in a runqueue: */
452 struct rt_prio_array active
;
453 unsigned long rt_nr_running
;
454 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int curr
; /* highest queued rt task prio */
458 int next
; /* next highest */
463 unsigned long rt_nr_migratory
;
464 unsigned long rt_nr_total
;
466 struct plist_head pushable_tasks
;
471 /* Nests inside the rq lock: */
472 spinlock_t rt_runtime_lock
;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 unsigned long rt_nr_boosted
;
478 struct list_head leaf_rt_rq_list
;
479 struct task_group
*tg
;
480 struct sched_rt_entity
*rt_se
;
487 * We add the notion of a root-domain which will be used to define per-domain
488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
490 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online
;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask
;
506 struct cpupri cpupri
;
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
514 static struct root_domain def_root_domain
;
519 * This is the main, per-CPU runqueue data structure.
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
533 unsigned long nr_running
;
534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
537 unsigned long last_tick_seen
;
538 unsigned char in_nohz_recently
;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load
;
542 unsigned long nr_load_updates
;
544 u64 nr_migrations_in
;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list
;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list
;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible
;
565 struct task_struct
*curr
, *idle
;
566 unsigned long next_balance
;
567 struct mm_struct
*prev_mm
;
574 struct root_domain
*rd
;
575 struct sched_domain
*sd
;
577 unsigned char idle_at_tick
;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task
;
588 struct task_struct
*migration_thread
;
589 struct list_head migration_queue
;
595 /* calc_load related fields */
596 unsigned long calc_load_update
;
597 long calc_load_active
;
599 #ifdef CONFIG_SCHED_HRTICK
601 int hrtick_csd_pending
;
602 struct call_single_data hrtick_csd
;
604 struct hrtimer hrtick_timer
;
607 #ifdef CONFIG_SCHEDSTATS
609 struct sched_info rq_sched_info
;
610 unsigned long long rq_cpu_time
;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
613 /* sys_sched_yield() stats */
614 unsigned int yld_count
;
616 /* schedule() stats */
617 unsigned int sched_switch
;
618 unsigned int sched_count
;
619 unsigned int sched_goidle
;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count
;
623 unsigned int ttwu_local
;
626 unsigned int bkl_count
;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
633 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
635 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
638 static inline int cpu_of(struct rq
*rq
)
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 #define raw_rq() (&__raw_get_cpu_var(runqueues))
663 inline void update_rq_clock(struct rq
*rq
)
665 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 #ifdef CONFIG_SCHED_DEBUG
672 # define const_debug __read_mostly
674 # define const_debug static const
680 * Returns true if the current cpu runqueue is locked.
681 * This interface allows printk to be called with the runqueue lock
682 * held and know whether or not it is OK to wake up the klogd.
684 int runqueue_is_locked(int cpu
)
686 return spin_is_locked(&cpu_rq(cpu
)->lock
);
690 * Debugging: various feature bits
693 #define SCHED_FEAT(name, enabled) \
694 __SCHED_FEAT_##name ,
697 #include "sched_features.h"
702 #define SCHED_FEAT(name, enabled) \
703 (1UL << __SCHED_FEAT_##name) * enabled |
705 const_debug
unsigned int sysctl_sched_features
=
706 #include "sched_features.h"
711 #ifdef CONFIG_SCHED_DEBUG
712 #define SCHED_FEAT(name, enabled) \
715 static __read_mostly
char *sched_feat_names
[] = {
716 #include "sched_features.h"
722 static int sched_feat_show(struct seq_file
*m
, void *v
)
726 for (i
= 0; sched_feat_names
[i
]; i
++) {
727 if (!(sysctl_sched_features
& (1UL << i
)))
729 seq_printf(m
, "%s ", sched_feat_names
[i
]);
737 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
738 size_t cnt
, loff_t
*ppos
)
748 if (copy_from_user(&buf
, ubuf
, cnt
))
753 if (strncmp(buf
, "NO_", 3) == 0) {
758 for (i
= 0; sched_feat_names
[i
]; i
++) {
759 int len
= strlen(sched_feat_names
[i
]);
761 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
763 sysctl_sched_features
&= ~(1UL << i
);
765 sysctl_sched_features
|= (1UL << i
);
770 if (!sched_feat_names
[i
])
778 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
780 return single_open(filp
, sched_feat_show
, NULL
);
783 static const struct file_operations sched_feat_fops
= {
784 .open
= sched_feat_open
,
785 .write
= sched_feat_write
,
788 .release
= single_release
,
791 static __init
int sched_init_debug(void)
793 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
798 late_initcall(sched_init_debug
);
802 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
805 * Number of tasks to iterate in a single balance run.
806 * Limited because this is done with IRQs disabled.
808 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
811 * ratelimit for updating the group shares.
814 unsigned int sysctl_sched_shares_ratelimit
= 250000;
817 * Inject some fuzzyness into changing the per-cpu group shares
818 * this avoids remote rq-locks at the expense of fairness.
821 unsigned int sysctl_sched_shares_thresh
= 4;
824 * period over which we average the RT time consumption, measured
829 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
832 * period over which we measure -rt task cpu usage in us.
835 unsigned int sysctl_sched_rt_period
= 1000000;
837 static __read_mostly
int scheduler_running
;
840 * part of the period that we allow rt tasks to run in us.
843 int sysctl_sched_rt_runtime
= 950000;
845 static inline u64
global_rt_period(void)
847 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
850 static inline u64
global_rt_runtime(void)
852 if (sysctl_sched_rt_runtime
< 0)
855 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
858 #ifndef prepare_arch_switch
859 # define prepare_arch_switch(next) do { } while (0)
861 #ifndef finish_arch_switch
862 # define finish_arch_switch(prev) do { } while (0)
865 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
867 return rq
->curr
== p
;
870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
871 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
873 return task_current(rq
, p
);
876 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
880 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
882 #ifdef CONFIG_DEBUG_SPINLOCK
883 /* this is a valid case when another task releases the spinlock */
884 rq
->lock
.owner
= current
;
887 * If we are tracking spinlock dependencies then we have to
888 * fix up the runqueue lock - which gets 'carried over' from
891 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
893 spin_unlock_irq(&rq
->lock
);
896 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
897 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
902 return task_current(rq
, p
);
906 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 spin_unlock_irq(&rq
->lock
);
919 spin_unlock(&rq
->lock
);
923 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
927 * After ->oncpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * __task_rq_lock - lock the runqueue a given task resides on.
942 * Must be called interrupts disabled.
944 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
948 struct rq
*rq
= task_rq(p
);
949 spin_lock(&rq
->lock
);
950 if (likely(rq
== task_rq(p
)))
952 spin_unlock(&rq
->lock
);
957 * task_rq_lock - lock the runqueue a given task resides on and disable
958 * interrupts. Note the ordering: we can safely lookup the task_rq without
959 * explicitly disabling preemption.
961 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
967 local_irq_save(*flags
);
969 spin_lock(&rq
->lock
);
970 if (likely(rq
== task_rq(p
)))
972 spin_unlock_irqrestore(&rq
->lock
, *flags
);
976 void task_rq_unlock_wait(struct task_struct
*p
)
978 struct rq
*rq
= task_rq(p
);
980 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
981 spin_unlock_wait(&rq
->lock
);
984 static void __task_rq_unlock(struct rq
*rq
)
987 spin_unlock(&rq
->lock
);
990 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
993 spin_unlock_irqrestore(&rq
->lock
, *flags
);
997 * this_rq_lock - lock this runqueue and disable interrupts.
999 static struct rq
*this_rq_lock(void)
1000 __acquires(rq
->lock
)
1004 local_irq_disable();
1006 spin_lock(&rq
->lock
);
1011 #ifdef CONFIG_SCHED_HRTICK
1013 * Use HR-timers to deliver accurate preemption points.
1015 * Its all a bit involved since we cannot program an hrt while holding the
1016 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1019 * When we get rescheduled we reprogram the hrtick_timer outside of the
1025 * - enabled by features
1026 * - hrtimer is actually high res
1028 static inline int hrtick_enabled(struct rq
*rq
)
1030 if (!sched_feat(HRTICK
))
1032 if (!cpu_active(cpu_of(rq
)))
1034 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1037 static void hrtick_clear(struct rq
*rq
)
1039 if (hrtimer_active(&rq
->hrtick_timer
))
1040 hrtimer_cancel(&rq
->hrtick_timer
);
1044 * High-resolution timer tick.
1045 * Runs from hardirq context with interrupts disabled.
1047 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1049 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1051 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1053 spin_lock(&rq
->lock
);
1054 update_rq_clock(rq
);
1055 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1056 spin_unlock(&rq
->lock
);
1058 return HRTIMER_NORESTART
;
1063 * called from hardirq (IPI) context
1065 static void __hrtick_start(void *arg
)
1067 struct rq
*rq
= arg
;
1069 spin_lock(&rq
->lock
);
1070 hrtimer_restart(&rq
->hrtick_timer
);
1071 rq
->hrtick_csd_pending
= 0;
1072 spin_unlock(&rq
->lock
);
1076 * Called to set the hrtick timer state.
1078 * called with rq->lock held and irqs disabled
1080 static void hrtick_start(struct rq
*rq
, u64 delay
)
1082 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1083 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1085 hrtimer_set_expires(timer
, time
);
1087 if (rq
== this_rq()) {
1088 hrtimer_restart(timer
);
1089 } else if (!rq
->hrtick_csd_pending
) {
1090 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1091 rq
->hrtick_csd_pending
= 1;
1096 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1098 int cpu
= (int)(long)hcpu
;
1101 case CPU_UP_CANCELED
:
1102 case CPU_UP_CANCELED_FROZEN
:
1103 case CPU_DOWN_PREPARE
:
1104 case CPU_DOWN_PREPARE_FROZEN
:
1106 case CPU_DEAD_FROZEN
:
1107 hrtick_clear(cpu_rq(cpu
));
1114 static __init
void init_hrtick(void)
1116 hotcpu_notifier(hotplug_hrtick
, 0);
1120 * Called to set the hrtick timer state.
1122 * called with rq->lock held and irqs disabled
1124 static void hrtick_start(struct rq
*rq
, u64 delay
)
1126 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1127 HRTIMER_MODE_REL_PINNED
, 0);
1130 static inline void init_hrtick(void)
1133 #endif /* CONFIG_SMP */
1135 static void init_rq_hrtick(struct rq
*rq
)
1138 rq
->hrtick_csd_pending
= 0;
1140 rq
->hrtick_csd
.flags
= 0;
1141 rq
->hrtick_csd
.func
= __hrtick_start
;
1142 rq
->hrtick_csd
.info
= rq
;
1145 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1146 rq
->hrtick_timer
.function
= hrtick
;
1148 #else /* CONFIG_SCHED_HRTICK */
1149 static inline void hrtick_clear(struct rq
*rq
)
1153 static inline void init_rq_hrtick(struct rq
*rq
)
1157 static inline void init_hrtick(void)
1160 #endif /* CONFIG_SCHED_HRTICK */
1163 * resched_task - mark a task 'to be rescheduled now'.
1165 * On UP this means the setting of the need_resched flag, on SMP it
1166 * might also involve a cross-CPU call to trigger the scheduler on
1171 #ifndef tsk_is_polling
1172 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1175 static void resched_task(struct task_struct
*p
)
1179 assert_spin_locked(&task_rq(p
)->lock
);
1181 if (test_tsk_need_resched(p
))
1184 set_tsk_need_resched(p
);
1187 if (cpu
== smp_processor_id())
1190 /* NEED_RESCHED must be visible before we test polling */
1192 if (!tsk_is_polling(p
))
1193 smp_send_reschedule(cpu
);
1196 static void resched_cpu(int cpu
)
1198 struct rq
*rq
= cpu_rq(cpu
);
1199 unsigned long flags
;
1201 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1203 resched_task(cpu_curr(cpu
));
1204 spin_unlock_irqrestore(&rq
->lock
, flags
);
1209 * When add_timer_on() enqueues a timer into the timer wheel of an
1210 * idle CPU then this timer might expire before the next timer event
1211 * which is scheduled to wake up that CPU. In case of a completely
1212 * idle system the next event might even be infinite time into the
1213 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1214 * leaves the inner idle loop so the newly added timer is taken into
1215 * account when the CPU goes back to idle and evaluates the timer
1216 * wheel for the next timer event.
1218 void wake_up_idle_cpu(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1222 if (cpu
== smp_processor_id())
1226 * This is safe, as this function is called with the timer
1227 * wheel base lock of (cpu) held. When the CPU is on the way
1228 * to idle and has not yet set rq->curr to idle then it will
1229 * be serialized on the timer wheel base lock and take the new
1230 * timer into account automatically.
1232 if (rq
->curr
!= rq
->idle
)
1236 * We can set TIF_RESCHED on the idle task of the other CPU
1237 * lockless. The worst case is that the other CPU runs the
1238 * idle task through an additional NOOP schedule()
1240 set_tsk_need_resched(rq
->idle
);
1242 /* NEED_RESCHED must be visible before we test polling */
1244 if (!tsk_is_polling(rq
->idle
))
1245 smp_send_reschedule(cpu
);
1247 #endif /* CONFIG_NO_HZ */
1249 static u64
sched_avg_period(void)
1251 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1254 static void sched_avg_update(struct rq
*rq
)
1256 s64 period
= sched_avg_period();
1258 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1259 rq
->age_stamp
+= period
;
1264 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1266 rq
->rt_avg
+= rt_delta
;
1267 sched_avg_update(rq
);
1270 #else /* !CONFIG_SMP */
1271 static void resched_task(struct task_struct
*p
)
1273 assert_spin_locked(&task_rq(p
)->lock
);
1274 set_tsk_need_resched(p
);
1277 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1280 #endif /* CONFIG_SMP */
1282 #if BITS_PER_LONG == 32
1283 # define WMULT_CONST (~0UL)
1285 # define WMULT_CONST (1UL << 32)
1288 #define WMULT_SHIFT 32
1291 * Shift right and round:
1293 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1296 * delta *= weight / lw
1298 static unsigned long
1299 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1300 struct load_weight
*lw
)
1304 if (!lw
->inv_weight
) {
1305 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1308 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1312 tmp
= (u64
)delta_exec
* weight
;
1314 * Check whether we'd overflow the 64-bit multiplication:
1316 if (unlikely(tmp
> WMULT_CONST
))
1317 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1320 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1322 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1325 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1331 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1338 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1339 * of tasks with abnormal "nice" values across CPUs the contribution that
1340 * each task makes to its run queue's load is weighted according to its
1341 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1342 * scaled version of the new time slice allocation that they receive on time
1346 #define WEIGHT_IDLEPRIO 3
1347 #define WMULT_IDLEPRIO 1431655765
1350 * Nice levels are multiplicative, with a gentle 10% change for every
1351 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1352 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1353 * that remained on nice 0.
1355 * The "10% effect" is relative and cumulative: from _any_ nice level,
1356 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1357 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1358 * If a task goes up by ~10% and another task goes down by ~10% then
1359 * the relative distance between them is ~25%.)
1361 static const int prio_to_weight
[40] = {
1362 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1363 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1364 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1365 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1366 /* 0 */ 1024, 820, 655, 526, 423,
1367 /* 5 */ 335, 272, 215, 172, 137,
1368 /* 10 */ 110, 87, 70, 56, 45,
1369 /* 15 */ 36, 29, 23, 18, 15,
1373 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1375 * In cases where the weight does not change often, we can use the
1376 * precalculated inverse to speed up arithmetics by turning divisions
1377 * into multiplications:
1379 static const u32 prio_to_wmult
[40] = {
1380 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1381 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1382 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1383 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1384 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1385 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1386 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1387 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1390 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1393 * runqueue iterator, to support SMP load-balancing between different
1394 * scheduling classes, without having to expose their internal data
1395 * structures to the load-balancing proper:
1397 struct rq_iterator
{
1399 struct task_struct
*(*start
)(void *);
1400 struct task_struct
*(*next
)(void *);
1404 static unsigned long
1405 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1406 unsigned long max_load_move
, struct sched_domain
*sd
,
1407 enum cpu_idle_type idle
, int *all_pinned
,
1408 int *this_best_prio
, struct rq_iterator
*iterator
);
1411 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1412 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1413 struct rq_iterator
*iterator
);
1416 /* Time spent by the tasks of the cpu accounting group executing in ... */
1417 enum cpuacct_stat_index
{
1418 CPUACCT_STAT_USER
, /* ... user mode */
1419 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1421 CPUACCT_STAT_NSTATS
,
1424 #ifdef CONFIG_CGROUP_CPUACCT
1425 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1426 static void cpuacct_update_stats(struct task_struct
*tsk
,
1427 enum cpuacct_stat_index idx
, cputime_t val
);
1429 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1430 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1431 enum cpuacct_stat_index idx
, cputime_t val
) {}
1434 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1436 update_load_add(&rq
->load
, load
);
1439 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1441 update_load_sub(&rq
->load
, load
);
1444 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1445 typedef int (*tg_visitor
)(struct task_group
*, void *);
1448 * Iterate the full tree, calling @down when first entering a node and @up when
1449 * leaving it for the final time.
1451 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1453 struct task_group
*parent
, *child
;
1457 parent
= &root_task_group
;
1459 ret
= (*down
)(parent
, data
);
1462 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1469 ret
= (*up
)(parent
, data
);
1474 parent
= parent
->parent
;
1483 static int tg_nop(struct task_group
*tg
, void *data
)
1490 /* Used instead of source_load when we know the type == 0 */
1491 static unsigned long weighted_cpuload(const int cpu
)
1493 return cpu_rq(cpu
)->load
.weight
;
1497 * Return a low guess at the load of a migration-source cpu weighted
1498 * according to the scheduling class and "nice" value.
1500 * We want to under-estimate the load of migration sources, to
1501 * balance conservatively.
1503 static unsigned long source_load(int cpu
, int type
)
1505 struct rq
*rq
= cpu_rq(cpu
);
1506 unsigned long total
= weighted_cpuload(cpu
);
1508 if (type
== 0 || !sched_feat(LB_BIAS
))
1511 return min(rq
->cpu_load
[type
-1], total
);
1515 * Return a high guess at the load of a migration-target cpu weighted
1516 * according to the scheduling class and "nice" value.
1518 static unsigned long target_load(int cpu
, int type
)
1520 struct rq
*rq
= cpu_rq(cpu
);
1521 unsigned long total
= weighted_cpuload(cpu
);
1523 if (type
== 0 || !sched_feat(LB_BIAS
))
1526 return max(rq
->cpu_load
[type
-1], total
);
1529 static struct sched_group
*group_of(int cpu
)
1531 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1539 static unsigned long power_of(int cpu
)
1541 struct sched_group
*group
= group_of(cpu
);
1544 return SCHED_LOAD_SCALE
;
1546 return group
->cpu_power
;
1549 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1551 static unsigned long cpu_avg_load_per_task(int cpu
)
1553 struct rq
*rq
= cpu_rq(cpu
);
1554 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1557 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1559 rq
->avg_load_per_task
= 0;
1561 return rq
->avg_load_per_task
;
1564 #ifdef CONFIG_FAIR_GROUP_SCHED
1566 struct update_shares_data
{
1567 unsigned long rq_weight
[NR_CPUS
];
1570 static DEFINE_PER_CPU(struct update_shares_data
, update_shares_data
);
1572 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1575 * Calculate and set the cpu's group shares.
1577 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1578 unsigned long sd_shares
,
1579 unsigned long sd_rq_weight
,
1580 struct update_shares_data
*usd
)
1582 unsigned long shares
, rq_weight
;
1585 rq_weight
= usd
->rq_weight
[cpu
];
1588 rq_weight
= NICE_0_LOAD
;
1592 * \Sum_j shares_j * rq_weight_i
1593 * shares_i = -----------------------------
1594 * \Sum_j rq_weight_j
1596 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1597 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1599 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1600 sysctl_sched_shares_thresh
) {
1601 struct rq
*rq
= cpu_rq(cpu
);
1602 unsigned long flags
;
1604 spin_lock_irqsave(&rq
->lock
, flags
);
1605 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1606 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1607 __set_se_shares(tg
->se
[cpu
], shares
);
1608 spin_unlock_irqrestore(&rq
->lock
, flags
);
1613 * Re-compute the task group their per cpu shares over the given domain.
1614 * This needs to be done in a bottom-up fashion because the rq weight of a
1615 * parent group depends on the shares of its child groups.
1617 static int tg_shares_up(struct task_group
*tg
, void *data
)
1619 unsigned long weight
, rq_weight
= 0, shares
= 0;
1620 struct update_shares_data
*usd
;
1621 struct sched_domain
*sd
= data
;
1622 unsigned long flags
;
1628 local_irq_save(flags
);
1629 usd
= &__get_cpu_var(update_shares_data
);
1631 for_each_cpu(i
, sched_domain_span(sd
)) {
1632 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1633 usd
->rq_weight
[i
] = weight
;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1641 weight
= NICE_0_LOAD
;
1643 rq_weight
+= weight
;
1644 shares
+= tg
->cfs_rq
[i
]->shares
;
1647 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1648 shares
= tg
->shares
;
1650 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1651 shares
= tg
->shares
;
1653 for_each_cpu(i
, sched_domain_span(sd
))
1654 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd
);
1656 local_irq_restore(flags
);
1662 * Compute the cpu's hierarchical load factor for each task group.
1663 * This needs to be done in a top-down fashion because the load of a child
1664 * group is a fraction of its parents load.
1666 static int tg_load_down(struct task_group
*tg
, void *data
)
1669 long cpu
= (long)data
;
1672 load
= cpu_rq(cpu
)->load
.weight
;
1674 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1675 load
*= tg
->cfs_rq
[cpu
]->shares
;
1676 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1679 tg
->cfs_rq
[cpu
]->h_load
= load
;
1684 static void update_shares(struct sched_domain
*sd
)
1689 if (root_task_group_empty())
1692 now
= cpu_clock(raw_smp_processor_id());
1693 elapsed
= now
- sd
->last_update
;
1695 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1696 sd
->last_update
= now
;
1697 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1701 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1703 if (root_task_group_empty())
1706 spin_unlock(&rq
->lock
);
1708 spin_lock(&rq
->lock
);
1711 static void update_h_load(long cpu
)
1713 if (root_task_group_empty())
1716 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1721 static inline void update_shares(struct sched_domain
*sd
)
1725 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1731 #ifdef CONFIG_PREEMPT
1733 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1736 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1737 * way at the expense of forcing extra atomic operations in all
1738 * invocations. This assures that the double_lock is acquired using the
1739 * same underlying policy as the spinlock_t on this architecture, which
1740 * reduces latency compared to the unfair variant below. However, it
1741 * also adds more overhead and therefore may reduce throughput.
1743 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1744 __releases(this_rq
->lock
)
1745 __acquires(busiest
->lock
)
1746 __acquires(this_rq
->lock
)
1748 spin_unlock(&this_rq
->lock
);
1749 double_rq_lock(this_rq
, busiest
);
1756 * Unfair double_lock_balance: Optimizes throughput at the expense of
1757 * latency by eliminating extra atomic operations when the locks are
1758 * already in proper order on entry. This favors lower cpu-ids and will
1759 * grant the double lock to lower cpus over higher ids under contention,
1760 * regardless of entry order into the function.
1762 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1763 __releases(this_rq
->lock
)
1764 __acquires(busiest
->lock
)
1765 __acquires(this_rq
->lock
)
1769 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1770 if (busiest
< this_rq
) {
1771 spin_unlock(&this_rq
->lock
);
1772 spin_lock(&busiest
->lock
);
1773 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1776 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1781 #endif /* CONFIG_PREEMPT */
1784 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1786 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1788 if (unlikely(!irqs_disabled())) {
1789 /* printk() doesn't work good under rq->lock */
1790 spin_unlock(&this_rq
->lock
);
1794 return _double_lock_balance(this_rq
, busiest
);
1797 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1798 __releases(busiest
->lock
)
1800 spin_unlock(&busiest
->lock
);
1801 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1805 #ifdef CONFIG_FAIR_GROUP_SCHED
1806 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1809 cfs_rq
->shares
= shares
;
1814 static void calc_load_account_active(struct rq
*this_rq
);
1816 #include "sched_stats.h"
1817 #include "sched_idletask.c"
1818 #include "sched_fair.c"
1819 #include "sched_rt.c"
1820 #ifdef CONFIG_SCHED_DEBUG
1821 # include "sched_debug.c"
1824 #define sched_class_highest (&rt_sched_class)
1825 #define for_each_class(class) \
1826 for (class = sched_class_highest; class; class = class->next)
1828 static void inc_nr_running(struct rq
*rq
)
1833 static void dec_nr_running(struct rq
*rq
)
1838 static void set_load_weight(struct task_struct
*p
)
1840 if (task_has_rt_policy(p
)) {
1841 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1842 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1847 * SCHED_IDLE tasks get minimal weight:
1849 if (p
->policy
== SCHED_IDLE
) {
1850 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1851 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1855 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1856 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1859 static void update_avg(u64
*avg
, u64 sample
)
1861 s64 diff
= sample
- *avg
;
1865 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1868 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1870 sched_info_queued(p
);
1871 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1875 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1878 if (p
->se
.last_wakeup
) {
1879 update_avg(&p
->se
.avg_overlap
,
1880 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1881 p
->se
.last_wakeup
= 0;
1883 update_avg(&p
->se
.avg_wakeup
,
1884 sysctl_sched_wakeup_granularity
);
1888 sched_info_dequeued(p
);
1889 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1894 * __normal_prio - return the priority that is based on the static prio
1896 static inline int __normal_prio(struct task_struct
*p
)
1898 return p
->static_prio
;
1902 * Calculate the expected normal priority: i.e. priority
1903 * without taking RT-inheritance into account. Might be
1904 * boosted by interactivity modifiers. Changes upon fork,
1905 * setprio syscalls, and whenever the interactivity
1906 * estimator recalculates.
1908 static inline int normal_prio(struct task_struct
*p
)
1912 if (task_has_rt_policy(p
))
1913 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1915 prio
= __normal_prio(p
);
1920 * Calculate the current priority, i.e. the priority
1921 * taken into account by the scheduler. This value might
1922 * be boosted by RT tasks, or might be boosted by
1923 * interactivity modifiers. Will be RT if the task got
1924 * RT-boosted. If not then it returns p->normal_prio.
1926 static int effective_prio(struct task_struct
*p
)
1928 p
->normal_prio
= normal_prio(p
);
1930 * If we are RT tasks or we were boosted to RT priority,
1931 * keep the priority unchanged. Otherwise, update priority
1932 * to the normal priority:
1934 if (!rt_prio(p
->prio
))
1935 return p
->normal_prio
;
1940 * activate_task - move a task to the runqueue.
1942 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1944 if (task_contributes_to_load(p
))
1945 rq
->nr_uninterruptible
--;
1947 enqueue_task(rq
, p
, wakeup
);
1952 * deactivate_task - remove a task from the runqueue.
1954 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1956 if (task_contributes_to_load(p
))
1957 rq
->nr_uninterruptible
++;
1959 dequeue_task(rq
, p
, sleep
);
1964 * task_curr - is this task currently executing on a CPU?
1965 * @p: the task in question.
1967 inline int task_curr(const struct task_struct
*p
)
1969 return cpu_curr(task_cpu(p
)) == p
;
1972 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1974 set_task_rq(p
, cpu
);
1977 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1978 * successfuly executed on another CPU. We must ensure that updates of
1979 * per-task data have been completed by this moment.
1982 task_thread_info(p
)->cpu
= cpu
;
1986 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1987 const struct sched_class
*prev_class
,
1988 int oldprio
, int running
)
1990 if (prev_class
!= p
->sched_class
) {
1991 if (prev_class
->switched_from
)
1992 prev_class
->switched_from(rq
, p
, running
);
1993 p
->sched_class
->switched_to(rq
, p
, running
);
1995 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2000 * Is this task likely cache-hot:
2003 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY
) &&
2011 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2012 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2015 if (p
->sched_class
!= &fair_sched_class
)
2018 if (sysctl_sched_migration_cost
== -1)
2020 if (sysctl_sched_migration_cost
== 0)
2023 delta
= now
- p
->se
.exec_start
;
2025 return delta
< (s64
)sysctl_sched_migration_cost
;
2029 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2031 int old_cpu
= task_cpu(p
);
2032 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2033 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2034 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2037 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2039 trace_sched_migrate_task(p
, new_cpu
);
2041 #ifdef CONFIG_SCHEDSTATS
2042 if (p
->se
.wait_start
)
2043 p
->se
.wait_start
-= clock_offset
;
2044 if (p
->se
.sleep_start
)
2045 p
->se
.sleep_start
-= clock_offset
;
2046 if (p
->se
.block_start
)
2047 p
->se
.block_start
-= clock_offset
;
2049 if (old_cpu
!= new_cpu
) {
2050 p
->se
.nr_migrations
++;
2051 new_rq
->nr_migrations_in
++;
2052 #ifdef CONFIG_SCHEDSTATS
2053 if (task_hot(p
, old_rq
->clock
, NULL
))
2054 schedstat_inc(p
, se
.nr_forced2_migrations
);
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2059 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2060 new_cfsrq
->min_vruntime
;
2062 __set_task_cpu(p
, new_cpu
);
2065 struct migration_req
{
2066 struct list_head list
;
2068 struct task_struct
*task
;
2071 struct completion done
;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2081 struct rq
*rq
= task_rq(p
);
2084 * If the task is not on a runqueue (and not running), then
2085 * it is sufficient to simply update the task's cpu field.
2087 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2088 set_task_cpu(p
, dest_cpu
);
2092 init_completion(&req
->done
);
2094 req
->dest_cpu
= dest_cpu
;
2095 list_add(&req
->list
, &rq
->migration_queue
);
2101 * wait_task_context_switch - wait for a thread to complete at least one
2104 * @p must not be current.
2106 void wait_task_context_switch(struct task_struct
*p
)
2108 unsigned long nvcsw
, nivcsw
, flags
;
2116 * The runqueue is assigned before the actual context
2117 * switch. We need to take the runqueue lock.
2119 * We could check initially without the lock but it is
2120 * very likely that we need to take the lock in every
2123 rq
= task_rq_lock(p
, &flags
);
2124 running
= task_running(rq
, p
);
2125 task_rq_unlock(rq
, &flags
);
2127 if (likely(!running
))
2130 * The switch count is incremented before the actual
2131 * context switch. We thus wait for two switches to be
2132 * sure at least one completed.
2134 if ((p
->nvcsw
- nvcsw
) > 1)
2136 if ((p
->nivcsw
- nivcsw
) > 1)
2144 * wait_task_inactive - wait for a thread to unschedule.
2146 * If @match_state is nonzero, it's the @p->state value just checked and
2147 * not expected to change. If it changes, i.e. @p might have woken up,
2148 * then return zero. When we succeed in waiting for @p to be off its CPU,
2149 * we return a positive number (its total switch count). If a second call
2150 * a short while later returns the same number, the caller can be sure that
2151 * @p has remained unscheduled the whole time.
2153 * The caller must ensure that the task *will* unschedule sometime soon,
2154 * else this function might spin for a *long* time. This function can't
2155 * be called with interrupts off, or it may introduce deadlock with
2156 * smp_call_function() if an IPI is sent by the same process we are
2157 * waiting to become inactive.
2159 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2161 unsigned long flags
;
2168 * We do the initial early heuristics without holding
2169 * any task-queue locks at all. We'll only try to get
2170 * the runqueue lock when things look like they will
2176 * If the task is actively running on another CPU
2177 * still, just relax and busy-wait without holding
2180 * NOTE! Since we don't hold any locks, it's not
2181 * even sure that "rq" stays as the right runqueue!
2182 * But we don't care, since "task_running()" will
2183 * return false if the runqueue has changed and p
2184 * is actually now running somewhere else!
2186 while (task_running(rq
, p
)) {
2187 if (match_state
&& unlikely(p
->state
!= match_state
))
2193 * Ok, time to look more closely! We need the rq
2194 * lock now, to be *sure*. If we're wrong, we'll
2195 * just go back and repeat.
2197 rq
= task_rq_lock(p
, &flags
);
2198 trace_sched_wait_task(rq
, p
);
2199 running
= task_running(rq
, p
);
2200 on_rq
= p
->se
.on_rq
;
2202 if (!match_state
|| p
->state
== match_state
)
2203 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2204 task_rq_unlock(rq
, &flags
);
2207 * If it changed from the expected state, bail out now.
2209 if (unlikely(!ncsw
))
2213 * Was it really running after all now that we
2214 * checked with the proper locks actually held?
2216 * Oops. Go back and try again..
2218 if (unlikely(running
)) {
2224 * It's not enough that it's not actively running,
2225 * it must be off the runqueue _entirely_, and not
2228 * So if it was still runnable (but just not actively
2229 * running right now), it's preempted, and we should
2230 * yield - it could be a while.
2232 if (unlikely(on_rq
)) {
2233 schedule_timeout_uninterruptible(1);
2238 * Ahh, all good. It wasn't running, and it wasn't
2239 * runnable, which means that it will never become
2240 * running in the future either. We're all done!
2249 * kick_process - kick a running thread to enter/exit the kernel
2250 * @p: the to-be-kicked thread
2252 * Cause a process which is running on another CPU to enter
2253 * kernel-mode, without any delay. (to get signals handled.)
2255 * NOTE: this function doesnt have to take the runqueue lock,
2256 * because all it wants to ensure is that the remote task enters
2257 * the kernel. If the IPI races and the task has been migrated
2258 * to another CPU then no harm is done and the purpose has been
2261 void kick_process(struct task_struct
*p
)
2267 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2268 smp_send_reschedule(cpu
);
2271 EXPORT_SYMBOL_GPL(kick_process
);
2272 #endif /* CONFIG_SMP */
2275 * task_oncpu_function_call - call a function on the cpu on which a task runs
2276 * @p: the task to evaluate
2277 * @func: the function to be called
2278 * @info: the function call argument
2280 * Calls the function @func when the task is currently running. This might
2281 * be on the current CPU, which just calls the function directly
2283 void task_oncpu_function_call(struct task_struct
*p
,
2284 void (*func
) (void *info
), void *info
)
2291 smp_call_function_single(cpu
, func
, info
, 1);
2296 * try_to_wake_up - wake up a thread
2297 * @p: the to-be-woken-up thread
2298 * @state: the mask of task states that can be woken
2299 * @sync: do a synchronous wakeup?
2301 * Put it on the run-queue if it's not already there. The "current"
2302 * thread is always on the run-queue (except when the actual
2303 * re-schedule is in progress), and as such you're allowed to do
2304 * the simpler "current->state = TASK_RUNNING" to mark yourself
2305 * runnable without the overhead of this.
2307 * returns failure only if the task is already active.
2309 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2312 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2313 unsigned long flags
;
2316 if (!sched_feat(SYNC_WAKEUPS
))
2317 wake_flags
&= ~WF_SYNC
;
2319 this_cpu
= get_cpu();
2322 rq
= task_rq_lock(p
, &flags
);
2323 update_rq_clock(rq
);
2324 if (!(p
->state
& state
))
2334 if (unlikely(task_running(rq
, p
)))
2338 * In order to handle concurrent wakeups and release the rq->lock
2339 * we put the task in TASK_WAKING state.
2341 * First fix up the nr_uninterruptible count:
2343 if (task_contributes_to_load(p
))
2344 rq
->nr_uninterruptible
--;
2345 p
->state
= TASK_WAKING
;
2346 task_rq_unlock(rq
, &flags
);
2348 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2349 if (cpu
!= orig_cpu
)
2350 set_task_cpu(p
, cpu
);
2352 rq
= task_rq_lock(p
, &flags
);
2353 WARN_ON(p
->state
!= TASK_WAKING
);
2356 #ifdef CONFIG_SCHEDSTATS
2357 schedstat_inc(rq
, ttwu_count
);
2358 if (cpu
== this_cpu
)
2359 schedstat_inc(rq
, ttwu_local
);
2361 struct sched_domain
*sd
;
2362 for_each_domain(this_cpu
, sd
) {
2363 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2364 schedstat_inc(sd
, ttwu_wake_remote
);
2369 #endif /* CONFIG_SCHEDSTATS */
2372 #endif /* CONFIG_SMP */
2373 schedstat_inc(p
, se
.nr_wakeups
);
2374 if (wake_flags
& WF_SYNC
)
2375 schedstat_inc(p
, se
.nr_wakeups_sync
);
2376 if (orig_cpu
!= cpu
)
2377 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2378 if (cpu
== this_cpu
)
2379 schedstat_inc(p
, se
.nr_wakeups_local
);
2381 schedstat_inc(p
, se
.nr_wakeups_remote
);
2382 activate_task(rq
, p
, 1);
2386 * Only attribute actual wakeups done by this task.
2388 if (!in_interrupt()) {
2389 struct sched_entity
*se
= ¤t
->se
;
2390 u64 sample
= se
->sum_exec_runtime
;
2392 if (se
->last_wakeup
)
2393 sample
-= se
->last_wakeup
;
2395 sample
-= se
->start_runtime
;
2396 update_avg(&se
->avg_wakeup
, sample
);
2398 se
->last_wakeup
= se
->sum_exec_runtime
;
2402 trace_sched_wakeup(rq
, p
, success
);
2403 check_preempt_curr(rq
, p
, wake_flags
);
2405 p
->state
= TASK_RUNNING
;
2407 if (p
->sched_class
->task_wake_up
)
2408 p
->sched_class
->task_wake_up(rq
, p
);
2411 task_rq_unlock(rq
, &flags
);
2418 * wake_up_process - Wake up a specific process
2419 * @p: The process to be woken up.
2421 * Attempt to wake up the nominated process and move it to the set of runnable
2422 * processes. Returns 1 if the process was woken up, 0 if it was already
2425 * It may be assumed that this function implies a write memory barrier before
2426 * changing the task state if and only if any tasks are woken up.
2428 int wake_up_process(struct task_struct
*p
)
2430 return try_to_wake_up(p
, TASK_ALL
, 0);
2432 EXPORT_SYMBOL(wake_up_process
);
2434 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2436 return try_to_wake_up(p
, state
, 0);
2440 * Perform scheduler related setup for a newly forked process p.
2441 * p is forked by current.
2443 * __sched_fork() is basic setup used by init_idle() too:
2445 static void __sched_fork(struct task_struct
*p
)
2447 p
->se
.exec_start
= 0;
2448 p
->se
.sum_exec_runtime
= 0;
2449 p
->se
.prev_sum_exec_runtime
= 0;
2450 p
->se
.nr_migrations
= 0;
2451 p
->se
.last_wakeup
= 0;
2452 p
->se
.avg_overlap
= 0;
2453 p
->se
.start_runtime
= 0;
2454 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2455 p
->se
.avg_running
= 0;
2457 #ifdef CONFIG_SCHEDSTATS
2458 p
->se
.wait_start
= 0;
2460 p
->se
.wait_count
= 0;
2463 p
->se
.sleep_start
= 0;
2464 p
->se
.sleep_max
= 0;
2465 p
->se
.sum_sleep_runtime
= 0;
2467 p
->se
.block_start
= 0;
2468 p
->se
.block_max
= 0;
2470 p
->se
.slice_max
= 0;
2472 p
->se
.nr_migrations_cold
= 0;
2473 p
->se
.nr_failed_migrations_affine
= 0;
2474 p
->se
.nr_failed_migrations_running
= 0;
2475 p
->se
.nr_failed_migrations_hot
= 0;
2476 p
->se
.nr_forced_migrations
= 0;
2477 p
->se
.nr_forced2_migrations
= 0;
2479 p
->se
.nr_wakeups
= 0;
2480 p
->se
.nr_wakeups_sync
= 0;
2481 p
->se
.nr_wakeups_migrate
= 0;
2482 p
->se
.nr_wakeups_local
= 0;
2483 p
->se
.nr_wakeups_remote
= 0;
2484 p
->se
.nr_wakeups_affine
= 0;
2485 p
->se
.nr_wakeups_affine_attempts
= 0;
2486 p
->se
.nr_wakeups_passive
= 0;
2487 p
->se
.nr_wakeups_idle
= 0;
2491 INIT_LIST_HEAD(&p
->rt
.run_list
);
2493 INIT_LIST_HEAD(&p
->se
.group_node
);
2495 #ifdef CONFIG_PREEMPT_NOTIFIERS
2496 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2500 * We mark the process as running here, but have not actually
2501 * inserted it onto the runqueue yet. This guarantees that
2502 * nobody will actually run it, and a signal or other external
2503 * event cannot wake it up and insert it on the runqueue either.
2505 p
->state
= TASK_RUNNING
;
2509 * fork()/clone()-time setup:
2511 void sched_fork(struct task_struct
*p
, int clone_flags
)
2513 int cpu
= get_cpu();
2518 * Revert to default priority/policy on fork if requested.
2520 if (unlikely(p
->sched_reset_on_fork
)) {
2521 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2522 p
->policy
= SCHED_NORMAL
;
2523 p
->normal_prio
= p
->static_prio
;
2526 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2527 p
->static_prio
= NICE_TO_PRIO(0);
2528 p
->normal_prio
= p
->static_prio
;
2533 * We don't need the reset flag anymore after the fork. It has
2534 * fulfilled its duty:
2536 p
->sched_reset_on_fork
= 0;
2540 * Make sure we do not leak PI boosting priority to the child.
2542 p
->prio
= current
->normal_prio
;
2544 if (!rt_prio(p
->prio
))
2545 p
->sched_class
= &fair_sched_class
;
2548 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2550 set_task_cpu(p
, cpu
);
2552 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2553 if (likely(sched_info_on()))
2554 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2556 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2559 #ifdef CONFIG_PREEMPT
2560 /* Want to start with kernel preemption disabled. */
2561 task_thread_info(p
)->preempt_count
= 1;
2563 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2569 * wake_up_new_task - wake up a newly created task for the first time.
2571 * This function will do some initial scheduler statistics housekeeping
2572 * that must be done for every newly created context, then puts the task
2573 * on the runqueue and wakes it.
2575 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2577 unsigned long flags
;
2580 rq
= task_rq_lock(p
, &flags
);
2581 BUG_ON(p
->state
!= TASK_RUNNING
);
2582 update_rq_clock(rq
);
2584 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2585 activate_task(rq
, p
, 0);
2588 * Let the scheduling class do new task startup
2589 * management (if any):
2591 p
->sched_class
->task_new(rq
, p
);
2594 trace_sched_wakeup_new(rq
, p
, 1);
2595 check_preempt_curr(rq
, p
, WF_FORK
);
2597 if (p
->sched_class
->task_wake_up
)
2598 p
->sched_class
->task_wake_up(rq
, p
);
2600 task_rq_unlock(rq
, &flags
);
2603 #ifdef CONFIG_PREEMPT_NOTIFIERS
2606 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2607 * @notifier: notifier struct to register
2609 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2611 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2613 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2616 * preempt_notifier_unregister - no longer interested in preemption notifications
2617 * @notifier: notifier struct to unregister
2619 * This is safe to call from within a preemption notifier.
2621 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2623 hlist_del(¬ifier
->link
);
2625 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2627 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2629 struct preempt_notifier
*notifier
;
2630 struct hlist_node
*node
;
2632 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2633 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2637 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2638 struct task_struct
*next
)
2640 struct preempt_notifier
*notifier
;
2641 struct hlist_node
*node
;
2643 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2644 notifier
->ops
->sched_out(notifier
, next
);
2647 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2649 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2654 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2655 struct task_struct
*next
)
2659 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2662 * prepare_task_switch - prepare to switch tasks
2663 * @rq: the runqueue preparing to switch
2664 * @prev: the current task that is being switched out
2665 * @next: the task we are going to switch to.
2667 * This is called with the rq lock held and interrupts off. It must
2668 * be paired with a subsequent finish_task_switch after the context
2671 * prepare_task_switch sets up locking and calls architecture specific
2675 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2676 struct task_struct
*next
)
2678 fire_sched_out_preempt_notifiers(prev
, next
);
2679 prepare_lock_switch(rq
, next
);
2680 prepare_arch_switch(next
);
2684 * finish_task_switch - clean up after a task-switch
2685 * @rq: runqueue associated with task-switch
2686 * @prev: the thread we just switched away from.
2688 * finish_task_switch must be called after the context switch, paired
2689 * with a prepare_task_switch call before the context switch.
2690 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2691 * and do any other architecture-specific cleanup actions.
2693 * Note that we may have delayed dropping an mm in context_switch(). If
2694 * so, we finish that here outside of the runqueue lock. (Doing it
2695 * with the lock held can cause deadlocks; see schedule() for
2698 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2699 __releases(rq
->lock
)
2701 struct mm_struct
*mm
= rq
->prev_mm
;
2707 * A task struct has one reference for the use as "current".
2708 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2709 * schedule one last time. The schedule call will never return, and
2710 * the scheduled task must drop that reference.
2711 * The test for TASK_DEAD must occur while the runqueue locks are
2712 * still held, otherwise prev could be scheduled on another cpu, die
2713 * there before we look at prev->state, and then the reference would
2715 * Manfred Spraul <manfred@colorfullife.com>
2717 prev_state
= prev
->state
;
2718 finish_arch_switch(prev
);
2719 perf_event_task_sched_in(current
, cpu_of(rq
));
2720 finish_lock_switch(rq
, prev
);
2722 fire_sched_in_preempt_notifiers(current
);
2725 if (unlikely(prev_state
== TASK_DEAD
)) {
2727 * Remove function-return probe instances associated with this
2728 * task and put them back on the free list.
2730 kprobe_flush_task(prev
);
2731 put_task_struct(prev
);
2737 /* assumes rq->lock is held */
2738 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2740 if (prev
->sched_class
->pre_schedule
)
2741 prev
->sched_class
->pre_schedule(rq
, prev
);
2744 /* rq->lock is NOT held, but preemption is disabled */
2745 static inline void post_schedule(struct rq
*rq
)
2747 if (rq
->post_schedule
) {
2748 unsigned long flags
;
2750 spin_lock_irqsave(&rq
->lock
, flags
);
2751 if (rq
->curr
->sched_class
->post_schedule
)
2752 rq
->curr
->sched_class
->post_schedule(rq
);
2753 spin_unlock_irqrestore(&rq
->lock
, flags
);
2755 rq
->post_schedule
= 0;
2761 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2765 static inline void post_schedule(struct rq
*rq
)
2772 * schedule_tail - first thing a freshly forked thread must call.
2773 * @prev: the thread we just switched away from.
2775 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2776 __releases(rq
->lock
)
2778 struct rq
*rq
= this_rq();
2780 finish_task_switch(rq
, prev
);
2783 * FIXME: do we need to worry about rq being invalidated by the
2788 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2789 /* In this case, finish_task_switch does not reenable preemption */
2792 if (current
->set_child_tid
)
2793 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2797 * context_switch - switch to the new MM and the new
2798 * thread's register state.
2801 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2802 struct task_struct
*next
)
2804 struct mm_struct
*mm
, *oldmm
;
2806 prepare_task_switch(rq
, prev
, next
);
2807 trace_sched_switch(rq
, prev
, next
);
2809 oldmm
= prev
->active_mm
;
2811 * For paravirt, this is coupled with an exit in switch_to to
2812 * combine the page table reload and the switch backend into
2815 arch_start_context_switch(prev
);
2817 if (unlikely(!mm
)) {
2818 next
->active_mm
= oldmm
;
2819 atomic_inc(&oldmm
->mm_count
);
2820 enter_lazy_tlb(oldmm
, next
);
2822 switch_mm(oldmm
, mm
, next
);
2824 if (unlikely(!prev
->mm
)) {
2825 prev
->active_mm
= NULL
;
2826 rq
->prev_mm
= oldmm
;
2829 * Since the runqueue lock will be released by the next
2830 * task (which is an invalid locking op but in the case
2831 * of the scheduler it's an obvious special-case), so we
2832 * do an early lockdep release here:
2834 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2835 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2838 /* Here we just switch the register state and the stack. */
2839 switch_to(prev
, next
, prev
);
2843 * this_rq must be evaluated again because prev may have moved
2844 * CPUs since it called schedule(), thus the 'rq' on its stack
2845 * frame will be invalid.
2847 finish_task_switch(this_rq(), prev
);
2851 * nr_running, nr_uninterruptible and nr_context_switches:
2853 * externally visible scheduler statistics: current number of runnable
2854 * threads, current number of uninterruptible-sleeping threads, total
2855 * number of context switches performed since bootup.
2857 unsigned long nr_running(void)
2859 unsigned long i
, sum
= 0;
2861 for_each_online_cpu(i
)
2862 sum
+= cpu_rq(i
)->nr_running
;
2867 unsigned long nr_uninterruptible(void)
2869 unsigned long i
, sum
= 0;
2871 for_each_possible_cpu(i
)
2872 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2875 * Since we read the counters lockless, it might be slightly
2876 * inaccurate. Do not allow it to go below zero though:
2878 if (unlikely((long)sum
< 0))
2884 unsigned long long nr_context_switches(void)
2887 unsigned long long sum
= 0;
2889 for_each_possible_cpu(i
)
2890 sum
+= cpu_rq(i
)->nr_switches
;
2895 unsigned long nr_iowait(void)
2897 unsigned long i
, sum
= 0;
2899 for_each_possible_cpu(i
)
2900 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2905 unsigned long nr_iowait_cpu(void)
2907 struct rq
*this = this_rq();
2908 return atomic_read(&this->nr_iowait
);
2911 unsigned long this_cpu_load(void)
2913 struct rq
*this = this_rq();
2914 return this->cpu_load
[0];
2918 /* Variables and functions for calc_load */
2919 static atomic_long_t calc_load_tasks
;
2920 static unsigned long calc_load_update
;
2921 unsigned long avenrun
[3];
2922 EXPORT_SYMBOL(avenrun
);
2925 * get_avenrun - get the load average array
2926 * @loads: pointer to dest load array
2927 * @offset: offset to add
2928 * @shift: shift count to shift the result left
2930 * These values are estimates at best, so no need for locking.
2932 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2934 loads
[0] = (avenrun
[0] + offset
) << shift
;
2935 loads
[1] = (avenrun
[1] + offset
) << shift
;
2936 loads
[2] = (avenrun
[2] + offset
) << shift
;
2939 static unsigned long
2940 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2943 load
+= active
* (FIXED_1
- exp
);
2944 return load
>> FSHIFT
;
2948 * calc_load - update the avenrun load estimates 10 ticks after the
2949 * CPUs have updated calc_load_tasks.
2951 void calc_global_load(void)
2953 unsigned long upd
= calc_load_update
+ 10;
2956 if (time_before(jiffies
, upd
))
2959 active
= atomic_long_read(&calc_load_tasks
);
2960 active
= active
> 0 ? active
* FIXED_1
: 0;
2962 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2963 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2964 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2966 calc_load_update
+= LOAD_FREQ
;
2970 * Either called from update_cpu_load() or from a cpu going idle
2972 static void calc_load_account_active(struct rq
*this_rq
)
2974 long nr_active
, delta
;
2976 nr_active
= this_rq
->nr_running
;
2977 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2979 if (nr_active
!= this_rq
->calc_load_active
) {
2980 delta
= nr_active
- this_rq
->calc_load_active
;
2981 this_rq
->calc_load_active
= nr_active
;
2982 atomic_long_add(delta
, &calc_load_tasks
);
2987 * Externally visible per-cpu scheduler statistics:
2988 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2990 u64
cpu_nr_migrations(int cpu
)
2992 return cpu_rq(cpu
)->nr_migrations_in
;
2996 * Update rq->cpu_load[] statistics. This function is usually called every
2997 * scheduler tick (TICK_NSEC).
2999 static void update_cpu_load(struct rq
*this_rq
)
3001 unsigned long this_load
= this_rq
->load
.weight
;
3004 this_rq
->nr_load_updates
++;
3006 /* Update our load: */
3007 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3008 unsigned long old_load
, new_load
;
3010 /* scale is effectively 1 << i now, and >> i divides by scale */
3012 old_load
= this_rq
->cpu_load
[i
];
3013 new_load
= this_load
;
3015 * Round up the averaging division if load is increasing. This
3016 * prevents us from getting stuck on 9 if the load is 10, for
3019 if (new_load
> old_load
)
3020 new_load
+= scale
-1;
3021 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3024 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3025 this_rq
->calc_load_update
+= LOAD_FREQ
;
3026 calc_load_account_active(this_rq
);
3033 * double_rq_lock - safely lock two runqueues
3035 * Note this does not disable interrupts like task_rq_lock,
3036 * you need to do so manually before calling.
3038 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3039 __acquires(rq1
->lock
)
3040 __acquires(rq2
->lock
)
3042 BUG_ON(!irqs_disabled());
3044 spin_lock(&rq1
->lock
);
3045 __acquire(rq2
->lock
); /* Fake it out ;) */
3048 spin_lock(&rq1
->lock
);
3049 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3051 spin_lock(&rq2
->lock
);
3052 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3055 update_rq_clock(rq1
);
3056 update_rq_clock(rq2
);
3060 * double_rq_unlock - safely unlock two runqueues
3062 * Note this does not restore interrupts like task_rq_unlock,
3063 * you need to do so manually after calling.
3065 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3066 __releases(rq1
->lock
)
3067 __releases(rq2
->lock
)
3069 spin_unlock(&rq1
->lock
);
3071 spin_unlock(&rq2
->lock
);
3073 __release(rq2
->lock
);
3077 * If dest_cpu is allowed for this process, migrate the task to it.
3078 * This is accomplished by forcing the cpu_allowed mask to only
3079 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3080 * the cpu_allowed mask is restored.
3082 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3084 struct migration_req req
;
3085 unsigned long flags
;
3088 rq
= task_rq_lock(p
, &flags
);
3089 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3090 || unlikely(!cpu_active(dest_cpu
)))
3093 /* force the process onto the specified CPU */
3094 if (migrate_task(p
, dest_cpu
, &req
)) {
3095 /* Need to wait for migration thread (might exit: take ref). */
3096 struct task_struct
*mt
= rq
->migration_thread
;
3098 get_task_struct(mt
);
3099 task_rq_unlock(rq
, &flags
);
3100 wake_up_process(mt
);
3101 put_task_struct(mt
);
3102 wait_for_completion(&req
.done
);
3107 task_rq_unlock(rq
, &flags
);
3111 * sched_exec - execve() is a valuable balancing opportunity, because at
3112 * this point the task has the smallest effective memory and cache footprint.
3114 void sched_exec(void)
3116 int new_cpu
, this_cpu
= get_cpu();
3117 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3119 if (new_cpu
!= this_cpu
)
3120 sched_migrate_task(current
, new_cpu
);
3124 * pull_task - move a task from a remote runqueue to the local runqueue.
3125 * Both runqueues must be locked.
3127 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3128 struct rq
*this_rq
, int this_cpu
)
3130 deactivate_task(src_rq
, p
, 0);
3131 set_task_cpu(p
, this_cpu
);
3132 activate_task(this_rq
, p
, 0);
3134 * Note that idle threads have a prio of MAX_PRIO, for this test
3135 * to be always true for them.
3137 check_preempt_curr(this_rq
, p
, 0);
3141 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3144 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3145 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3148 int tsk_cache_hot
= 0;
3150 * We do not migrate tasks that are:
3151 * 1) running (obviously), or
3152 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3153 * 3) are cache-hot on their current CPU.
3155 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3156 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3161 if (task_running(rq
, p
)) {
3162 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3167 * Aggressive migration if:
3168 * 1) task is cache cold, or
3169 * 2) too many balance attempts have failed.
3172 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3173 if (!tsk_cache_hot
||
3174 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3175 #ifdef CONFIG_SCHEDSTATS
3176 if (tsk_cache_hot
) {
3177 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3178 schedstat_inc(p
, se
.nr_forced_migrations
);
3184 if (tsk_cache_hot
) {
3185 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3191 static unsigned long
3192 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3193 unsigned long max_load_move
, struct sched_domain
*sd
,
3194 enum cpu_idle_type idle
, int *all_pinned
,
3195 int *this_best_prio
, struct rq_iterator
*iterator
)
3197 int loops
= 0, pulled
= 0, pinned
= 0;
3198 struct task_struct
*p
;
3199 long rem_load_move
= max_load_move
;
3201 if (max_load_move
== 0)
3207 * Start the load-balancing iterator:
3209 p
= iterator
->start(iterator
->arg
);
3211 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3214 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3215 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3216 p
= iterator
->next(iterator
->arg
);
3220 pull_task(busiest
, p
, this_rq
, this_cpu
);
3222 rem_load_move
-= p
->se
.load
.weight
;
3224 #ifdef CONFIG_PREEMPT
3226 * NEWIDLE balancing is a source of latency, so preemptible kernels
3227 * will stop after the first task is pulled to minimize the critical
3230 if (idle
== CPU_NEWLY_IDLE
)
3235 * We only want to steal up to the prescribed amount of weighted load.
3237 if (rem_load_move
> 0) {
3238 if (p
->prio
< *this_best_prio
)
3239 *this_best_prio
= p
->prio
;
3240 p
= iterator
->next(iterator
->arg
);
3245 * Right now, this is one of only two places pull_task() is called,
3246 * so we can safely collect pull_task() stats here rather than
3247 * inside pull_task().
3249 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3252 *all_pinned
= pinned
;
3254 return max_load_move
- rem_load_move
;
3258 * move_tasks tries to move up to max_load_move weighted load from busiest to
3259 * this_rq, as part of a balancing operation within domain "sd".
3260 * Returns 1 if successful and 0 otherwise.
3262 * Called with both runqueues locked.
3264 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3265 unsigned long max_load_move
,
3266 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3269 const struct sched_class
*class = sched_class_highest
;
3270 unsigned long total_load_moved
= 0;
3271 int this_best_prio
= this_rq
->curr
->prio
;
3275 class->load_balance(this_rq
, this_cpu
, busiest
,
3276 max_load_move
- total_load_moved
,
3277 sd
, idle
, all_pinned
, &this_best_prio
);
3278 class = class->next
;
3280 #ifdef CONFIG_PREEMPT
3282 * NEWIDLE balancing is a source of latency, so preemptible
3283 * kernels will stop after the first task is pulled to minimize
3284 * the critical section.
3286 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3289 } while (class && max_load_move
> total_load_moved
);
3291 return total_load_moved
> 0;
3295 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3296 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3297 struct rq_iterator
*iterator
)
3299 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3303 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3304 pull_task(busiest
, p
, this_rq
, this_cpu
);
3306 * Right now, this is only the second place pull_task()
3307 * is called, so we can safely collect pull_task()
3308 * stats here rather than inside pull_task().
3310 schedstat_inc(sd
, lb_gained
[idle
]);
3314 p
= iterator
->next(iterator
->arg
);
3321 * move_one_task tries to move exactly one task from busiest to this_rq, as
3322 * part of active balancing operations within "domain".
3323 * Returns 1 if successful and 0 otherwise.
3325 * Called with both runqueues locked.
3327 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3328 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3330 const struct sched_class
*class;
3332 for_each_class(class) {
3333 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3339 /********** Helpers for find_busiest_group ************************/
3341 * sd_lb_stats - Structure to store the statistics of a sched_domain
3342 * during load balancing.
3344 struct sd_lb_stats
{
3345 struct sched_group
*busiest
; /* Busiest group in this sd */
3346 struct sched_group
*this; /* Local group in this sd */
3347 unsigned long total_load
; /* Total load of all groups in sd */
3348 unsigned long total_pwr
; /* Total power of all groups in sd */
3349 unsigned long avg_load
; /* Average load across all groups in sd */
3351 /** Statistics of this group */
3352 unsigned long this_load
;
3353 unsigned long this_load_per_task
;
3354 unsigned long this_nr_running
;
3356 /* Statistics of the busiest group */
3357 unsigned long max_load
;
3358 unsigned long busiest_load_per_task
;
3359 unsigned long busiest_nr_running
;
3361 int group_imb
; /* Is there imbalance in this sd */
3362 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3363 int power_savings_balance
; /* Is powersave balance needed for this sd */
3364 struct sched_group
*group_min
; /* Least loaded group in sd */
3365 struct sched_group
*group_leader
; /* Group which relieves group_min */
3366 unsigned long min_load_per_task
; /* load_per_task in group_min */
3367 unsigned long leader_nr_running
; /* Nr running of group_leader */
3368 unsigned long min_nr_running
; /* Nr running of group_min */
3373 * sg_lb_stats - stats of a sched_group required for load_balancing
3375 struct sg_lb_stats
{
3376 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3377 unsigned long group_load
; /* Total load over the CPUs of the group */
3378 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3379 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3380 unsigned long group_capacity
;
3381 int group_imb
; /* Is there an imbalance in the group ? */
3385 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3386 * @group: The group whose first cpu is to be returned.
3388 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3390 return cpumask_first(sched_group_cpus(group
));
3394 * get_sd_load_idx - Obtain the load index for a given sched domain.
3395 * @sd: The sched_domain whose load_idx is to be obtained.
3396 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3398 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3399 enum cpu_idle_type idle
)
3405 load_idx
= sd
->busy_idx
;
3408 case CPU_NEWLY_IDLE
:
3409 load_idx
= sd
->newidle_idx
;
3412 load_idx
= sd
->idle_idx
;
3420 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3422 * init_sd_power_savings_stats - Initialize power savings statistics for
3423 * the given sched_domain, during load balancing.
3425 * @sd: Sched domain whose power-savings statistics are to be initialized.
3426 * @sds: Variable containing the statistics for sd.
3427 * @idle: Idle status of the CPU at which we're performing load-balancing.
3429 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3430 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3433 * Busy processors will not participate in power savings
3436 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3437 sds
->power_savings_balance
= 0;
3439 sds
->power_savings_balance
= 1;
3440 sds
->min_nr_running
= ULONG_MAX
;
3441 sds
->leader_nr_running
= 0;
3446 * update_sd_power_savings_stats - Update the power saving stats for a
3447 * sched_domain while performing load balancing.
3449 * @group: sched_group belonging to the sched_domain under consideration.
3450 * @sds: Variable containing the statistics of the sched_domain
3451 * @local_group: Does group contain the CPU for which we're performing
3453 * @sgs: Variable containing the statistics of the group.
3455 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3456 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3459 if (!sds
->power_savings_balance
)
3463 * If the local group is idle or completely loaded
3464 * no need to do power savings balance at this domain
3466 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3467 !sds
->this_nr_running
))
3468 sds
->power_savings_balance
= 0;
3471 * If a group is already running at full capacity or idle,
3472 * don't include that group in power savings calculations
3474 if (!sds
->power_savings_balance
||
3475 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3476 !sgs
->sum_nr_running
)
3480 * Calculate the group which has the least non-idle load.
3481 * This is the group from where we need to pick up the load
3484 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3485 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3486 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3487 sds
->group_min
= group
;
3488 sds
->min_nr_running
= sgs
->sum_nr_running
;
3489 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3490 sgs
->sum_nr_running
;
3494 * Calculate the group which is almost near its
3495 * capacity but still has some space to pick up some load
3496 * from other group and save more power
3498 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3501 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3502 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3503 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3504 sds
->group_leader
= group
;
3505 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3510 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3511 * @sds: Variable containing the statistics of the sched_domain
3512 * under consideration.
3513 * @this_cpu: Cpu at which we're currently performing load-balancing.
3514 * @imbalance: Variable to store the imbalance.
3517 * Check if we have potential to perform some power-savings balance.
3518 * If yes, set the busiest group to be the least loaded group in the
3519 * sched_domain, so that it's CPUs can be put to idle.
3521 * Returns 1 if there is potential to perform power-savings balance.
3524 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3525 int this_cpu
, unsigned long *imbalance
)
3527 if (!sds
->power_savings_balance
)
3530 if (sds
->this != sds
->group_leader
||
3531 sds
->group_leader
== sds
->group_min
)
3534 *imbalance
= sds
->min_load_per_task
;
3535 sds
->busiest
= sds
->group_min
;
3540 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3541 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3542 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3547 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3548 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3553 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3554 int this_cpu
, unsigned long *imbalance
)
3558 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3561 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3563 return SCHED_LOAD_SCALE
;
3566 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3568 return default_scale_freq_power(sd
, cpu
);
3571 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3573 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3574 unsigned long smt_gain
= sd
->smt_gain
;
3581 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3583 return default_scale_smt_power(sd
, cpu
);
3586 unsigned long scale_rt_power(int cpu
)
3588 struct rq
*rq
= cpu_rq(cpu
);
3589 u64 total
, available
;
3591 sched_avg_update(rq
);
3593 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3594 available
= total
- rq
->rt_avg
;
3596 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3597 total
= SCHED_LOAD_SCALE
;
3599 total
>>= SCHED_LOAD_SHIFT
;
3601 return div_u64(available
, total
);
3604 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3606 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3607 unsigned long power
= SCHED_LOAD_SCALE
;
3608 struct sched_group
*sdg
= sd
->groups
;
3610 if (sched_feat(ARCH_POWER
))
3611 power
*= arch_scale_freq_power(sd
, cpu
);
3613 power
*= default_scale_freq_power(sd
, cpu
);
3615 power
>>= SCHED_LOAD_SHIFT
;
3617 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3618 if (sched_feat(ARCH_POWER
))
3619 power
*= arch_scale_smt_power(sd
, cpu
);
3621 power
*= default_scale_smt_power(sd
, cpu
);
3623 power
>>= SCHED_LOAD_SHIFT
;
3626 power
*= scale_rt_power(cpu
);
3627 power
>>= SCHED_LOAD_SHIFT
;
3632 sdg
->cpu_power
= power
;
3635 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3637 struct sched_domain
*child
= sd
->child
;
3638 struct sched_group
*group
, *sdg
= sd
->groups
;
3639 unsigned long power
;
3642 update_cpu_power(sd
, cpu
);
3648 group
= child
->groups
;
3650 power
+= group
->cpu_power
;
3651 group
= group
->next
;
3652 } while (group
!= child
->groups
);
3654 sdg
->cpu_power
= power
;
3658 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3659 * @group: sched_group whose statistics are to be updated.
3660 * @this_cpu: Cpu for which load balance is currently performed.
3661 * @idle: Idle status of this_cpu
3662 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3663 * @sd_idle: Idle status of the sched_domain containing group.
3664 * @local_group: Does group contain this_cpu.
3665 * @cpus: Set of cpus considered for load balancing.
3666 * @balance: Should we balance.
3667 * @sgs: variable to hold the statistics for this group.
3669 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3670 struct sched_group
*group
, int this_cpu
,
3671 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3672 int local_group
, const struct cpumask
*cpus
,
3673 int *balance
, struct sg_lb_stats
*sgs
)
3675 unsigned long load
, max_cpu_load
, min_cpu_load
;
3677 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3678 unsigned long sum_avg_load_per_task
;
3679 unsigned long avg_load_per_task
;
3682 balance_cpu
= group_first_cpu(group
);
3683 if (balance_cpu
== this_cpu
)
3684 update_group_power(sd
, this_cpu
);
3687 /* Tally up the load of all CPUs in the group */
3688 sum_avg_load_per_task
= avg_load_per_task
= 0;
3690 min_cpu_load
= ~0UL;
3692 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3693 struct rq
*rq
= cpu_rq(i
);
3695 if (*sd_idle
&& rq
->nr_running
)
3698 /* Bias balancing toward cpus of our domain */
3700 if (idle_cpu(i
) && !first_idle_cpu
) {
3705 load
= target_load(i
, load_idx
);
3707 load
= source_load(i
, load_idx
);
3708 if (load
> max_cpu_load
)
3709 max_cpu_load
= load
;
3710 if (min_cpu_load
> load
)
3711 min_cpu_load
= load
;
3714 sgs
->group_load
+= load
;
3715 sgs
->sum_nr_running
+= rq
->nr_running
;
3716 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3718 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3722 * First idle cpu or the first cpu(busiest) in this sched group
3723 * is eligible for doing load balancing at this and above
3724 * domains. In the newly idle case, we will allow all the cpu's
3725 * to do the newly idle load balance.
3727 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3728 balance_cpu
!= this_cpu
&& balance
) {
3733 /* Adjust by relative CPU power of the group */
3734 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3738 * Consider the group unbalanced when the imbalance is larger
3739 * than the average weight of two tasks.
3741 * APZ: with cgroup the avg task weight can vary wildly and
3742 * might not be a suitable number - should we keep a
3743 * normalized nr_running number somewhere that negates
3746 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3749 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3752 sgs
->group_capacity
=
3753 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3757 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3758 * @sd: sched_domain whose statistics are to be updated.
3759 * @this_cpu: Cpu for which load balance is currently performed.
3760 * @idle: Idle status of this_cpu
3761 * @sd_idle: Idle status of the sched_domain containing group.
3762 * @cpus: Set of cpus considered for load balancing.
3763 * @balance: Should we balance.
3764 * @sds: variable to hold the statistics for this sched_domain.
3766 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3767 enum cpu_idle_type idle
, int *sd_idle
,
3768 const struct cpumask
*cpus
, int *balance
,
3769 struct sd_lb_stats
*sds
)
3771 struct sched_domain
*child
= sd
->child
;
3772 struct sched_group
*group
= sd
->groups
;
3773 struct sg_lb_stats sgs
;
3774 int load_idx
, prefer_sibling
= 0;
3776 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3779 init_sd_power_savings_stats(sd
, sds
, idle
);
3780 load_idx
= get_sd_load_idx(sd
, idle
);
3785 local_group
= cpumask_test_cpu(this_cpu
,
3786 sched_group_cpus(group
));
3787 memset(&sgs
, 0, sizeof(sgs
));
3788 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3789 local_group
, cpus
, balance
, &sgs
);
3791 if (local_group
&& balance
&& !(*balance
))
3794 sds
->total_load
+= sgs
.group_load
;
3795 sds
->total_pwr
+= group
->cpu_power
;
3798 * In case the child domain prefers tasks go to siblings
3799 * first, lower the group capacity to one so that we'll try
3800 * and move all the excess tasks away.
3803 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3806 sds
->this_load
= sgs
.avg_load
;
3808 sds
->this_nr_running
= sgs
.sum_nr_running
;
3809 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3810 } else if (sgs
.avg_load
> sds
->max_load
&&
3811 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3813 sds
->max_load
= sgs
.avg_load
;
3814 sds
->busiest
= group
;
3815 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3816 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3817 sds
->group_imb
= sgs
.group_imb
;
3820 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3821 group
= group
->next
;
3822 } while (group
!= sd
->groups
);
3826 * fix_small_imbalance - Calculate the minor imbalance that exists
3827 * amongst the groups of a sched_domain, during
3829 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3830 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3831 * @imbalance: Variable to store the imbalance.
3833 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3834 int this_cpu
, unsigned long *imbalance
)
3836 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3837 unsigned int imbn
= 2;
3839 if (sds
->this_nr_running
) {
3840 sds
->this_load_per_task
/= sds
->this_nr_running
;
3841 if (sds
->busiest_load_per_task
>
3842 sds
->this_load_per_task
)
3845 sds
->this_load_per_task
=
3846 cpu_avg_load_per_task(this_cpu
);
3848 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3849 sds
->busiest_load_per_task
* imbn
) {
3850 *imbalance
= sds
->busiest_load_per_task
;
3855 * OK, we don't have enough imbalance to justify moving tasks,
3856 * however we may be able to increase total CPU power used by
3860 pwr_now
+= sds
->busiest
->cpu_power
*
3861 min(sds
->busiest_load_per_task
, sds
->max_load
);
3862 pwr_now
+= sds
->this->cpu_power
*
3863 min(sds
->this_load_per_task
, sds
->this_load
);
3864 pwr_now
/= SCHED_LOAD_SCALE
;
3866 /* Amount of load we'd subtract */
3867 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3868 sds
->busiest
->cpu_power
;
3869 if (sds
->max_load
> tmp
)
3870 pwr_move
+= sds
->busiest
->cpu_power
*
3871 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3873 /* Amount of load we'd add */
3874 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3875 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3876 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3877 sds
->this->cpu_power
;
3879 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3880 sds
->this->cpu_power
;
3881 pwr_move
+= sds
->this->cpu_power
*
3882 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3883 pwr_move
/= SCHED_LOAD_SCALE
;
3885 /* Move if we gain throughput */
3886 if (pwr_move
> pwr_now
)
3887 *imbalance
= sds
->busiest_load_per_task
;
3891 * calculate_imbalance - Calculate the amount of imbalance present within the
3892 * groups of a given sched_domain during load balance.
3893 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3894 * @this_cpu: Cpu for which currently load balance is being performed.
3895 * @imbalance: The variable to store the imbalance.
3897 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3898 unsigned long *imbalance
)
3900 unsigned long max_pull
;
3902 * In the presence of smp nice balancing, certain scenarios can have
3903 * max load less than avg load(as we skip the groups at or below
3904 * its cpu_power, while calculating max_load..)
3906 if (sds
->max_load
< sds
->avg_load
) {
3908 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3911 /* Don't want to pull so many tasks that a group would go idle */
3912 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3913 sds
->max_load
- sds
->busiest_load_per_task
);
3915 /* How much load to actually move to equalise the imbalance */
3916 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3917 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3921 * if *imbalance is less than the average load per runnable task
3922 * there is no gaurantee that any tasks will be moved so we'll have
3923 * a think about bumping its value to force at least one task to be
3926 if (*imbalance
< sds
->busiest_load_per_task
)
3927 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3930 /******* find_busiest_group() helpers end here *********************/
3933 * find_busiest_group - Returns the busiest group within the sched_domain
3934 * if there is an imbalance. If there isn't an imbalance, and
3935 * the user has opted for power-savings, it returns a group whose
3936 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3937 * such a group exists.
3939 * Also calculates the amount of weighted load which should be moved
3940 * to restore balance.
3942 * @sd: The sched_domain whose busiest group is to be returned.
3943 * @this_cpu: The cpu for which load balancing is currently being performed.
3944 * @imbalance: Variable which stores amount of weighted load which should
3945 * be moved to restore balance/put a group to idle.
3946 * @idle: The idle status of this_cpu.
3947 * @sd_idle: The idleness of sd
3948 * @cpus: The set of CPUs under consideration for load-balancing.
3949 * @balance: Pointer to a variable indicating if this_cpu
3950 * is the appropriate cpu to perform load balancing at this_level.
3952 * Returns: - the busiest group if imbalance exists.
3953 * - If no imbalance and user has opted for power-savings balance,
3954 * return the least loaded group whose CPUs can be
3955 * put to idle by rebalancing its tasks onto our group.
3957 static struct sched_group
*
3958 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3959 unsigned long *imbalance
, enum cpu_idle_type idle
,
3960 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3962 struct sd_lb_stats sds
;
3964 memset(&sds
, 0, sizeof(sds
));
3967 * Compute the various statistics relavent for load balancing at
3970 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3973 /* Cases where imbalance does not exist from POV of this_cpu */
3974 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3976 * 2) There is no busy sibling group to pull from.
3977 * 3) This group is the busiest group.
3978 * 4) This group is more busy than the avg busieness at this
3980 * 5) The imbalance is within the specified limit.
3981 * 6) Any rebalance would lead to ping-pong
3983 if (balance
&& !(*balance
))
3986 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3989 if (sds
.this_load
>= sds
.max_load
)
3992 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3994 if (sds
.this_load
>= sds
.avg_load
)
3997 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4000 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4002 sds
.busiest_load_per_task
=
4003 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4006 * We're trying to get all the cpus to the average_load, so we don't
4007 * want to push ourselves above the average load, nor do we wish to
4008 * reduce the max loaded cpu below the average load, as either of these
4009 * actions would just result in more rebalancing later, and ping-pong
4010 * tasks around. Thus we look for the minimum possible imbalance.
4011 * Negative imbalances (*we* are more loaded than anyone else) will
4012 * be counted as no imbalance for these purposes -- we can't fix that
4013 * by pulling tasks to us. Be careful of negative numbers as they'll
4014 * appear as very large values with unsigned longs.
4016 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4019 /* Looks like there is an imbalance. Compute it */
4020 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4025 * There is no obvious imbalance. But check if we can do some balancing
4028 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4036 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4039 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4040 unsigned long imbalance
, const struct cpumask
*cpus
)
4042 struct rq
*busiest
= NULL
, *rq
;
4043 unsigned long max_load
= 0;
4046 for_each_cpu(i
, sched_group_cpus(group
)) {
4047 unsigned long power
= power_of(i
);
4048 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4051 if (!cpumask_test_cpu(i
, cpus
))
4055 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4058 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4061 if (wl
> max_load
) {
4071 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4072 * so long as it is large enough.
4074 #define MAX_PINNED_INTERVAL 512
4076 /* Working cpumask for load_balance and load_balance_newidle. */
4077 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4080 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4081 * tasks if there is an imbalance.
4083 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4084 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4087 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4088 struct sched_group
*group
;
4089 unsigned long imbalance
;
4091 unsigned long flags
;
4092 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4094 cpumask_setall(cpus
);
4097 * When power savings policy is enabled for the parent domain, idle
4098 * sibling can pick up load irrespective of busy siblings. In this case,
4099 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4100 * portraying it as CPU_NOT_IDLE.
4102 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4103 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4106 schedstat_inc(sd
, lb_count
[idle
]);
4110 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4117 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4121 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4123 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4127 BUG_ON(busiest
== this_rq
);
4129 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4132 if (busiest
->nr_running
> 1) {
4134 * Attempt to move tasks. If find_busiest_group has found
4135 * an imbalance but busiest->nr_running <= 1, the group is
4136 * still unbalanced. ld_moved simply stays zero, so it is
4137 * correctly treated as an imbalance.
4139 local_irq_save(flags
);
4140 double_rq_lock(this_rq
, busiest
);
4141 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4142 imbalance
, sd
, idle
, &all_pinned
);
4143 double_rq_unlock(this_rq
, busiest
);
4144 local_irq_restore(flags
);
4147 * some other cpu did the load balance for us.
4149 if (ld_moved
&& this_cpu
!= smp_processor_id())
4150 resched_cpu(this_cpu
);
4152 /* All tasks on this runqueue were pinned by CPU affinity */
4153 if (unlikely(all_pinned
)) {
4154 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4155 if (!cpumask_empty(cpus
))
4162 schedstat_inc(sd
, lb_failed
[idle
]);
4163 sd
->nr_balance_failed
++;
4165 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4167 spin_lock_irqsave(&busiest
->lock
, flags
);
4169 /* don't kick the migration_thread, if the curr
4170 * task on busiest cpu can't be moved to this_cpu
4172 if (!cpumask_test_cpu(this_cpu
,
4173 &busiest
->curr
->cpus_allowed
)) {
4174 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4176 goto out_one_pinned
;
4179 if (!busiest
->active_balance
) {
4180 busiest
->active_balance
= 1;
4181 busiest
->push_cpu
= this_cpu
;
4184 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4186 wake_up_process(busiest
->migration_thread
);
4189 * We've kicked active balancing, reset the failure
4192 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4195 sd
->nr_balance_failed
= 0;
4197 if (likely(!active_balance
)) {
4198 /* We were unbalanced, so reset the balancing interval */
4199 sd
->balance_interval
= sd
->min_interval
;
4202 * If we've begun active balancing, start to back off. This
4203 * case may not be covered by the all_pinned logic if there
4204 * is only 1 task on the busy runqueue (because we don't call
4207 if (sd
->balance_interval
< sd
->max_interval
)
4208 sd
->balance_interval
*= 2;
4211 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4212 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4218 schedstat_inc(sd
, lb_balanced
[idle
]);
4220 sd
->nr_balance_failed
= 0;
4223 /* tune up the balancing interval */
4224 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4225 (sd
->balance_interval
< sd
->max_interval
))
4226 sd
->balance_interval
*= 2;
4228 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4229 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4240 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4241 * tasks if there is an imbalance.
4243 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4244 * this_rq is locked.
4247 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4249 struct sched_group
*group
;
4250 struct rq
*busiest
= NULL
;
4251 unsigned long imbalance
;
4255 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4257 cpumask_setall(cpus
);
4260 * When power savings policy is enabled for the parent domain, idle
4261 * sibling can pick up load irrespective of busy siblings. In this case,
4262 * let the state of idle sibling percolate up as IDLE, instead of
4263 * portraying it as CPU_NOT_IDLE.
4265 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4266 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4269 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4271 update_shares_locked(this_rq
, sd
);
4272 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4273 &sd_idle
, cpus
, NULL
);
4275 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4279 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4281 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4285 BUG_ON(busiest
== this_rq
);
4287 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4290 if (busiest
->nr_running
> 1) {
4291 /* Attempt to move tasks */
4292 double_lock_balance(this_rq
, busiest
);
4293 /* this_rq->clock is already updated */
4294 update_rq_clock(busiest
);
4295 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4296 imbalance
, sd
, CPU_NEWLY_IDLE
,
4298 double_unlock_balance(this_rq
, busiest
);
4300 if (unlikely(all_pinned
)) {
4301 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4302 if (!cpumask_empty(cpus
))
4308 int active_balance
= 0;
4310 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4311 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4312 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4315 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4318 if (sd
->nr_balance_failed
++ < 2)
4322 * The only task running in a non-idle cpu can be moved to this
4323 * cpu in an attempt to completely freeup the other CPU
4324 * package. The same method used to move task in load_balance()
4325 * have been extended for load_balance_newidle() to speedup
4326 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4328 * The package power saving logic comes from
4329 * find_busiest_group(). If there are no imbalance, then
4330 * f_b_g() will return NULL. However when sched_mc={1,2} then
4331 * f_b_g() will select a group from which a running task may be
4332 * pulled to this cpu in order to make the other package idle.
4333 * If there is no opportunity to make a package idle and if
4334 * there are no imbalance, then f_b_g() will return NULL and no
4335 * action will be taken in load_balance_newidle().
4337 * Under normal task pull operation due to imbalance, there
4338 * will be more than one task in the source run queue and
4339 * move_tasks() will succeed. ld_moved will be true and this
4340 * active balance code will not be triggered.
4343 /* Lock busiest in correct order while this_rq is held */
4344 double_lock_balance(this_rq
, busiest
);
4347 * don't kick the migration_thread, if the curr
4348 * task on busiest cpu can't be moved to this_cpu
4350 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4351 double_unlock_balance(this_rq
, busiest
);
4356 if (!busiest
->active_balance
) {
4357 busiest
->active_balance
= 1;
4358 busiest
->push_cpu
= this_cpu
;
4362 double_unlock_balance(this_rq
, busiest
);
4364 * Should not call ttwu while holding a rq->lock
4366 spin_unlock(&this_rq
->lock
);
4368 wake_up_process(busiest
->migration_thread
);
4369 spin_lock(&this_rq
->lock
);
4372 sd
->nr_balance_failed
= 0;
4374 update_shares_locked(this_rq
, sd
);
4378 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4379 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4380 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4382 sd
->nr_balance_failed
= 0;
4388 * idle_balance is called by schedule() if this_cpu is about to become
4389 * idle. Attempts to pull tasks from other CPUs.
4391 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4393 struct sched_domain
*sd
;
4394 int pulled_task
= 0;
4395 unsigned long next_balance
= jiffies
+ HZ
;
4397 for_each_domain(this_cpu
, sd
) {
4398 unsigned long interval
;
4400 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4403 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4404 /* If we've pulled tasks over stop searching: */
4405 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4408 interval
= msecs_to_jiffies(sd
->balance_interval
);
4409 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4410 next_balance
= sd
->last_balance
+ interval
;
4414 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4416 * We are going idle. next_balance may be set based on
4417 * a busy processor. So reset next_balance.
4419 this_rq
->next_balance
= next_balance
;
4424 * active_load_balance is run by migration threads. It pushes running tasks
4425 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4426 * running on each physical CPU where possible, and avoids physical /
4427 * logical imbalances.
4429 * Called with busiest_rq locked.
4431 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4433 int target_cpu
= busiest_rq
->push_cpu
;
4434 struct sched_domain
*sd
;
4435 struct rq
*target_rq
;
4437 /* Is there any task to move? */
4438 if (busiest_rq
->nr_running
<= 1)
4441 target_rq
= cpu_rq(target_cpu
);
4444 * This condition is "impossible", if it occurs
4445 * we need to fix it. Originally reported by
4446 * Bjorn Helgaas on a 128-cpu setup.
4448 BUG_ON(busiest_rq
== target_rq
);
4450 /* move a task from busiest_rq to target_rq */
4451 double_lock_balance(busiest_rq
, target_rq
);
4452 update_rq_clock(busiest_rq
);
4453 update_rq_clock(target_rq
);
4455 /* Search for an sd spanning us and the target CPU. */
4456 for_each_domain(target_cpu
, sd
) {
4457 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4458 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4463 schedstat_inc(sd
, alb_count
);
4465 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4467 schedstat_inc(sd
, alb_pushed
);
4469 schedstat_inc(sd
, alb_failed
);
4471 double_unlock_balance(busiest_rq
, target_rq
);
4476 atomic_t load_balancer
;
4477 cpumask_var_t cpu_mask
;
4478 cpumask_var_t ilb_grp_nohz_mask
;
4479 } nohz ____cacheline_aligned
= {
4480 .load_balancer
= ATOMIC_INIT(-1),
4483 int get_nohz_load_balancer(void)
4485 return atomic_read(&nohz
.load_balancer
);
4488 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4490 * lowest_flag_domain - Return lowest sched_domain containing flag.
4491 * @cpu: The cpu whose lowest level of sched domain is to
4493 * @flag: The flag to check for the lowest sched_domain
4494 * for the given cpu.
4496 * Returns the lowest sched_domain of a cpu which contains the given flag.
4498 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4500 struct sched_domain
*sd
;
4502 for_each_domain(cpu
, sd
)
4503 if (sd
&& (sd
->flags
& flag
))
4510 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4511 * @cpu: The cpu whose domains we're iterating over.
4512 * @sd: variable holding the value of the power_savings_sd
4514 * @flag: The flag to filter the sched_domains to be iterated.
4516 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4517 * set, starting from the lowest sched_domain to the highest.
4519 #define for_each_flag_domain(cpu, sd, flag) \
4520 for (sd = lowest_flag_domain(cpu, flag); \
4521 (sd && (sd->flags & flag)); sd = sd->parent)
4524 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4525 * @ilb_group: group to be checked for semi-idleness
4527 * Returns: 1 if the group is semi-idle. 0 otherwise.
4529 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4530 * and atleast one non-idle CPU. This helper function checks if the given
4531 * sched_group is semi-idle or not.
4533 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4535 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4536 sched_group_cpus(ilb_group
));
4539 * A sched_group is semi-idle when it has atleast one busy cpu
4540 * and atleast one idle cpu.
4542 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4545 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4551 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4552 * @cpu: The cpu which is nominating a new idle_load_balancer.
4554 * Returns: Returns the id of the idle load balancer if it exists,
4555 * Else, returns >= nr_cpu_ids.
4557 * This algorithm picks the idle load balancer such that it belongs to a
4558 * semi-idle powersavings sched_domain. The idea is to try and avoid
4559 * completely idle packages/cores just for the purpose of idle load balancing
4560 * when there are other idle cpu's which are better suited for that job.
4562 static int find_new_ilb(int cpu
)
4564 struct sched_domain
*sd
;
4565 struct sched_group
*ilb_group
;
4568 * Have idle load balancer selection from semi-idle packages only
4569 * when power-aware load balancing is enabled
4571 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4575 * Optimize for the case when we have no idle CPUs or only one
4576 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4578 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4581 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4582 ilb_group
= sd
->groups
;
4585 if (is_semi_idle_group(ilb_group
))
4586 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4588 ilb_group
= ilb_group
->next
;
4590 } while (ilb_group
!= sd
->groups
);
4594 return cpumask_first(nohz
.cpu_mask
);
4596 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4597 static inline int find_new_ilb(int call_cpu
)
4599 return cpumask_first(nohz
.cpu_mask
);
4604 * This routine will try to nominate the ilb (idle load balancing)
4605 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4606 * load balancing on behalf of all those cpus. If all the cpus in the system
4607 * go into this tickless mode, then there will be no ilb owner (as there is
4608 * no need for one) and all the cpus will sleep till the next wakeup event
4611 * For the ilb owner, tick is not stopped. And this tick will be used
4612 * for idle load balancing. ilb owner will still be part of
4615 * While stopping the tick, this cpu will become the ilb owner if there
4616 * is no other owner. And will be the owner till that cpu becomes busy
4617 * or if all cpus in the system stop their ticks at which point
4618 * there is no need for ilb owner.
4620 * When the ilb owner becomes busy, it nominates another owner, during the
4621 * next busy scheduler_tick()
4623 int select_nohz_load_balancer(int stop_tick
)
4625 int cpu
= smp_processor_id();
4628 cpu_rq(cpu
)->in_nohz_recently
= 1;
4630 if (!cpu_active(cpu
)) {
4631 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4635 * If we are going offline and still the leader,
4638 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4644 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4646 /* time for ilb owner also to sleep */
4647 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4648 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4649 atomic_set(&nohz
.load_balancer
, -1);
4653 if (atomic_read(&nohz
.load_balancer
) == -1) {
4654 /* make me the ilb owner */
4655 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4657 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4660 if (!(sched_smt_power_savings
||
4661 sched_mc_power_savings
))
4664 * Check to see if there is a more power-efficient
4667 new_ilb
= find_new_ilb(cpu
);
4668 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4669 atomic_set(&nohz
.load_balancer
, -1);
4670 resched_cpu(new_ilb
);
4676 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4679 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4681 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4682 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4689 static DEFINE_SPINLOCK(balancing
);
4692 * It checks each scheduling domain to see if it is due to be balanced,
4693 * and initiates a balancing operation if so.
4695 * Balancing parameters are set up in arch_init_sched_domains.
4697 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4700 struct rq
*rq
= cpu_rq(cpu
);
4701 unsigned long interval
;
4702 struct sched_domain
*sd
;
4703 /* Earliest time when we have to do rebalance again */
4704 unsigned long next_balance
= jiffies
+ 60*HZ
;
4705 int update_next_balance
= 0;
4708 for_each_domain(cpu
, sd
) {
4709 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4712 interval
= sd
->balance_interval
;
4713 if (idle
!= CPU_IDLE
)
4714 interval
*= sd
->busy_factor
;
4716 /* scale ms to jiffies */
4717 interval
= msecs_to_jiffies(interval
);
4718 if (unlikely(!interval
))
4720 if (interval
> HZ
*NR_CPUS
/10)
4721 interval
= HZ
*NR_CPUS
/10;
4723 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4725 if (need_serialize
) {
4726 if (!spin_trylock(&balancing
))
4730 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4731 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4733 * We've pulled tasks over so either we're no
4734 * longer idle, or one of our SMT siblings is
4737 idle
= CPU_NOT_IDLE
;
4739 sd
->last_balance
= jiffies
;
4742 spin_unlock(&balancing
);
4744 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4745 next_balance
= sd
->last_balance
+ interval
;
4746 update_next_balance
= 1;
4750 * Stop the load balance at this level. There is another
4751 * CPU in our sched group which is doing load balancing more
4759 * next_balance will be updated only when there is a need.
4760 * When the cpu is attached to null domain for ex, it will not be
4763 if (likely(update_next_balance
))
4764 rq
->next_balance
= next_balance
;
4768 * run_rebalance_domains is triggered when needed from the scheduler tick.
4769 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4770 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4772 static void run_rebalance_domains(struct softirq_action
*h
)
4774 int this_cpu
= smp_processor_id();
4775 struct rq
*this_rq
= cpu_rq(this_cpu
);
4776 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4777 CPU_IDLE
: CPU_NOT_IDLE
;
4779 rebalance_domains(this_cpu
, idle
);
4783 * If this cpu is the owner for idle load balancing, then do the
4784 * balancing on behalf of the other idle cpus whose ticks are
4787 if (this_rq
->idle_at_tick
&&
4788 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4792 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4793 if (balance_cpu
== this_cpu
)
4797 * If this cpu gets work to do, stop the load balancing
4798 * work being done for other cpus. Next load
4799 * balancing owner will pick it up.
4804 rebalance_domains(balance_cpu
, CPU_IDLE
);
4806 rq
= cpu_rq(balance_cpu
);
4807 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4808 this_rq
->next_balance
= rq
->next_balance
;
4814 static inline int on_null_domain(int cpu
)
4816 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4820 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4822 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4823 * idle load balancing owner or decide to stop the periodic load balancing,
4824 * if the whole system is idle.
4826 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4830 * If we were in the nohz mode recently and busy at the current
4831 * scheduler tick, then check if we need to nominate new idle
4834 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4835 rq
->in_nohz_recently
= 0;
4837 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4838 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4839 atomic_set(&nohz
.load_balancer
, -1);
4842 if (atomic_read(&nohz
.load_balancer
) == -1) {
4843 int ilb
= find_new_ilb(cpu
);
4845 if (ilb
< nr_cpu_ids
)
4851 * If this cpu is idle and doing idle load balancing for all the
4852 * cpus with ticks stopped, is it time for that to stop?
4854 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4855 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4861 * If this cpu is idle and the idle load balancing is done by
4862 * someone else, then no need raise the SCHED_SOFTIRQ
4864 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4865 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4868 /* Don't need to rebalance while attached to NULL domain */
4869 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4870 likely(!on_null_domain(cpu
)))
4871 raise_softirq(SCHED_SOFTIRQ
);
4874 #else /* CONFIG_SMP */
4877 * on UP we do not need to balance between CPUs:
4879 static inline void idle_balance(int cpu
, struct rq
*rq
)
4885 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4887 EXPORT_PER_CPU_SYMBOL(kstat
);
4890 * Return any ns on the sched_clock that have not yet been accounted in
4891 * @p in case that task is currently running.
4893 * Called with task_rq_lock() held on @rq.
4895 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4899 if (task_current(rq
, p
)) {
4900 update_rq_clock(rq
);
4901 ns
= rq
->clock
- p
->se
.exec_start
;
4909 unsigned long long task_delta_exec(struct task_struct
*p
)
4911 unsigned long flags
;
4915 rq
= task_rq_lock(p
, &flags
);
4916 ns
= do_task_delta_exec(p
, rq
);
4917 task_rq_unlock(rq
, &flags
);
4923 * Return accounted runtime for the task.
4924 * In case the task is currently running, return the runtime plus current's
4925 * pending runtime that have not been accounted yet.
4927 unsigned long long task_sched_runtime(struct task_struct
*p
)
4929 unsigned long flags
;
4933 rq
= task_rq_lock(p
, &flags
);
4934 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4935 task_rq_unlock(rq
, &flags
);
4941 * Return sum_exec_runtime for the thread group.
4942 * In case the task is currently running, return the sum plus current's
4943 * pending runtime that have not been accounted yet.
4945 * Note that the thread group might have other running tasks as well,
4946 * so the return value not includes other pending runtime that other
4947 * running tasks might have.
4949 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4951 struct task_cputime totals
;
4952 unsigned long flags
;
4956 rq
= task_rq_lock(p
, &flags
);
4957 thread_group_cputime(p
, &totals
);
4958 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4959 task_rq_unlock(rq
, &flags
);
4965 * Account user cpu time to a process.
4966 * @p: the process that the cpu time gets accounted to
4967 * @cputime: the cpu time spent in user space since the last update
4968 * @cputime_scaled: cputime scaled by cpu frequency
4970 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4971 cputime_t cputime_scaled
)
4973 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4976 /* Add user time to process. */
4977 p
->utime
= cputime_add(p
->utime
, cputime
);
4978 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4979 account_group_user_time(p
, cputime
);
4981 /* Add user time to cpustat. */
4982 tmp
= cputime_to_cputime64(cputime
);
4983 if (TASK_NICE(p
) > 0)
4984 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4986 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4988 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4989 /* Account for user time used */
4990 acct_update_integrals(p
);
4994 * Account guest cpu time to a process.
4995 * @p: the process that the cpu time gets accounted to
4996 * @cputime: the cpu time spent in virtual machine since the last update
4997 * @cputime_scaled: cputime scaled by cpu frequency
4999 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5000 cputime_t cputime_scaled
)
5003 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5005 tmp
= cputime_to_cputime64(cputime
);
5007 /* Add guest time to process. */
5008 p
->utime
= cputime_add(p
->utime
, cputime
);
5009 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5010 account_group_user_time(p
, cputime
);
5011 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5013 /* Add guest time to cpustat. */
5014 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5015 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5019 * Account system cpu time to a process.
5020 * @p: the process that the cpu time gets accounted to
5021 * @hardirq_offset: the offset to subtract from hardirq_count()
5022 * @cputime: the cpu time spent in kernel space since the last update
5023 * @cputime_scaled: cputime scaled by cpu frequency
5025 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5026 cputime_t cputime
, cputime_t cputime_scaled
)
5028 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5031 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5032 account_guest_time(p
, cputime
, cputime_scaled
);
5036 /* Add system time to process. */
5037 p
->stime
= cputime_add(p
->stime
, cputime
);
5038 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5039 account_group_system_time(p
, cputime
);
5041 /* Add system time to cpustat. */
5042 tmp
= cputime_to_cputime64(cputime
);
5043 if (hardirq_count() - hardirq_offset
)
5044 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5045 else if (softirq_count())
5046 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5048 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5050 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5052 /* Account for system time used */
5053 acct_update_integrals(p
);
5057 * Account for involuntary wait time.
5058 * @steal: the cpu time spent in involuntary wait
5060 void account_steal_time(cputime_t cputime
)
5062 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5063 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5065 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5069 * Account for idle time.
5070 * @cputime: the cpu time spent in idle wait
5072 void account_idle_time(cputime_t cputime
)
5074 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5075 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5076 struct rq
*rq
= this_rq();
5078 if (atomic_read(&rq
->nr_iowait
) > 0)
5079 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5081 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5084 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5087 * Account a single tick of cpu time.
5088 * @p: the process that the cpu time gets accounted to
5089 * @user_tick: indicates if the tick is a user or a system tick
5091 void account_process_tick(struct task_struct
*p
, int user_tick
)
5093 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5094 struct rq
*rq
= this_rq();
5097 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5098 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5099 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5102 account_idle_time(cputime_one_jiffy
);
5106 * Account multiple ticks of steal time.
5107 * @p: the process from which the cpu time has been stolen
5108 * @ticks: number of stolen ticks
5110 void account_steal_ticks(unsigned long ticks
)
5112 account_steal_time(jiffies_to_cputime(ticks
));
5116 * Account multiple ticks of idle time.
5117 * @ticks: number of stolen ticks
5119 void account_idle_ticks(unsigned long ticks
)
5121 account_idle_time(jiffies_to_cputime(ticks
));
5127 * Use precise platform statistics if available:
5129 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5130 cputime_t
task_utime(struct task_struct
*p
)
5135 cputime_t
task_stime(struct task_struct
*p
)
5140 cputime_t
task_utime(struct task_struct
*p
)
5142 clock_t utime
= cputime_to_clock_t(p
->utime
),
5143 total
= utime
+ cputime_to_clock_t(p
->stime
);
5147 * Use CFS's precise accounting:
5149 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5153 do_div(temp
, total
);
5155 utime
= (clock_t)temp
;
5157 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5158 return p
->prev_utime
;
5161 cputime_t
task_stime(struct task_struct
*p
)
5166 * Use CFS's precise accounting. (we subtract utime from
5167 * the total, to make sure the total observed by userspace
5168 * grows monotonically - apps rely on that):
5170 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5171 cputime_to_clock_t(task_utime(p
));
5174 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5176 return p
->prev_stime
;
5180 inline cputime_t
task_gtime(struct task_struct
*p
)
5186 * This function gets called by the timer code, with HZ frequency.
5187 * We call it with interrupts disabled.
5189 * It also gets called by the fork code, when changing the parent's
5192 void scheduler_tick(void)
5194 int cpu
= smp_processor_id();
5195 struct rq
*rq
= cpu_rq(cpu
);
5196 struct task_struct
*curr
= rq
->curr
;
5200 spin_lock(&rq
->lock
);
5201 update_rq_clock(rq
);
5202 update_cpu_load(rq
);
5203 curr
->sched_class
->task_tick(rq
, curr
, 0);
5204 spin_unlock(&rq
->lock
);
5206 perf_event_task_tick(curr
, cpu
);
5209 rq
->idle_at_tick
= idle_cpu(cpu
);
5210 trigger_load_balance(rq
, cpu
);
5214 notrace
unsigned long get_parent_ip(unsigned long addr
)
5216 if (in_lock_functions(addr
)) {
5217 addr
= CALLER_ADDR2
;
5218 if (in_lock_functions(addr
))
5219 addr
= CALLER_ADDR3
;
5224 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5225 defined(CONFIG_PREEMPT_TRACER))
5227 void __kprobes
add_preempt_count(int val
)
5229 #ifdef CONFIG_DEBUG_PREEMPT
5233 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5236 preempt_count() += val
;
5237 #ifdef CONFIG_DEBUG_PREEMPT
5239 * Spinlock count overflowing soon?
5241 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5244 if (preempt_count() == val
)
5245 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5247 EXPORT_SYMBOL(add_preempt_count
);
5249 void __kprobes
sub_preempt_count(int val
)
5251 #ifdef CONFIG_DEBUG_PREEMPT
5255 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5258 * Is the spinlock portion underflowing?
5260 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5261 !(preempt_count() & PREEMPT_MASK
)))
5265 if (preempt_count() == val
)
5266 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5267 preempt_count() -= val
;
5269 EXPORT_SYMBOL(sub_preempt_count
);
5274 * Print scheduling while atomic bug:
5276 static noinline
void __schedule_bug(struct task_struct
*prev
)
5278 struct pt_regs
*regs
= get_irq_regs();
5280 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5281 prev
->comm
, prev
->pid
, preempt_count());
5283 debug_show_held_locks(prev
);
5285 if (irqs_disabled())
5286 print_irqtrace_events(prev
);
5295 * Various schedule()-time debugging checks and statistics:
5297 static inline void schedule_debug(struct task_struct
*prev
)
5300 * Test if we are atomic. Since do_exit() needs to call into
5301 * schedule() atomically, we ignore that path for now.
5302 * Otherwise, whine if we are scheduling when we should not be.
5304 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5305 __schedule_bug(prev
);
5307 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5309 schedstat_inc(this_rq(), sched_count
);
5310 #ifdef CONFIG_SCHEDSTATS
5311 if (unlikely(prev
->lock_depth
>= 0)) {
5312 schedstat_inc(this_rq(), bkl_count
);
5313 schedstat_inc(prev
, sched_info
.bkl_count
);
5318 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5320 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5322 update_avg(&p
->se
.avg_running
, runtime
);
5324 if (p
->state
== TASK_RUNNING
) {
5326 * In order to avoid avg_overlap growing stale when we are
5327 * indeed overlapping and hence not getting put to sleep, grow
5328 * the avg_overlap on preemption.
5330 * We use the average preemption runtime because that
5331 * correlates to the amount of cache footprint a task can
5334 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5335 update_avg(&p
->se
.avg_overlap
, runtime
);
5337 update_avg(&p
->se
.avg_running
, 0);
5339 p
->sched_class
->put_prev_task(rq
, p
);
5343 * Pick up the highest-prio task:
5345 static inline struct task_struct
*
5346 pick_next_task(struct rq
*rq
)
5348 const struct sched_class
*class;
5349 struct task_struct
*p
;
5352 * Optimization: we know that if all tasks are in
5353 * the fair class we can call that function directly:
5355 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5356 p
= fair_sched_class
.pick_next_task(rq
);
5361 class = sched_class_highest
;
5363 p
= class->pick_next_task(rq
);
5367 * Will never be NULL as the idle class always
5368 * returns a non-NULL p:
5370 class = class->next
;
5375 * schedule() is the main scheduler function.
5377 asmlinkage
void __sched
schedule(void)
5379 struct task_struct
*prev
, *next
;
5380 unsigned long *switch_count
;
5386 cpu
= smp_processor_id();
5390 switch_count
= &prev
->nivcsw
;
5392 release_kernel_lock(prev
);
5393 need_resched_nonpreemptible
:
5395 schedule_debug(prev
);
5397 if (sched_feat(HRTICK
))
5400 spin_lock_irq(&rq
->lock
);
5401 update_rq_clock(rq
);
5402 clear_tsk_need_resched(prev
);
5404 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5405 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5406 prev
->state
= TASK_RUNNING
;
5408 deactivate_task(rq
, prev
, 1);
5409 switch_count
= &prev
->nvcsw
;
5412 pre_schedule(rq
, prev
);
5414 if (unlikely(!rq
->nr_running
))
5415 idle_balance(cpu
, rq
);
5417 put_prev_task(rq
, prev
);
5418 next
= pick_next_task(rq
);
5420 if (likely(prev
!= next
)) {
5421 sched_info_switch(prev
, next
);
5422 perf_event_task_sched_out(prev
, next
, cpu
);
5428 context_switch(rq
, prev
, next
); /* unlocks the rq */
5430 * the context switch might have flipped the stack from under
5431 * us, hence refresh the local variables.
5433 cpu
= smp_processor_id();
5436 spin_unlock_irq(&rq
->lock
);
5440 if (unlikely(reacquire_kernel_lock(current
) < 0))
5441 goto need_resched_nonpreemptible
;
5443 preempt_enable_no_resched();
5447 EXPORT_SYMBOL(schedule
);
5451 * Look out! "owner" is an entirely speculative pointer
5452 * access and not reliable.
5454 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5459 if (!sched_feat(OWNER_SPIN
))
5462 #ifdef CONFIG_DEBUG_PAGEALLOC
5464 * Need to access the cpu field knowing that
5465 * DEBUG_PAGEALLOC could have unmapped it if
5466 * the mutex owner just released it and exited.
5468 if (probe_kernel_address(&owner
->cpu
, cpu
))
5475 * Even if the access succeeded (likely case),
5476 * the cpu field may no longer be valid.
5478 if (cpu
>= nr_cpumask_bits
)
5482 * We need to validate that we can do a
5483 * get_cpu() and that we have the percpu area.
5485 if (!cpu_online(cpu
))
5492 * Owner changed, break to re-assess state.
5494 if (lock
->owner
!= owner
)
5498 * Is that owner really running on that cpu?
5500 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5510 #ifdef CONFIG_PREEMPT
5512 * this is the entry point to schedule() from in-kernel preemption
5513 * off of preempt_enable. Kernel preemptions off return from interrupt
5514 * occur there and call schedule directly.
5516 asmlinkage
void __sched
preempt_schedule(void)
5518 struct thread_info
*ti
= current_thread_info();
5521 * If there is a non-zero preempt_count or interrupts are disabled,
5522 * we do not want to preempt the current task. Just return..
5524 if (likely(ti
->preempt_count
|| irqs_disabled()))
5528 add_preempt_count(PREEMPT_ACTIVE
);
5530 sub_preempt_count(PREEMPT_ACTIVE
);
5533 * Check again in case we missed a preemption opportunity
5534 * between schedule and now.
5537 } while (need_resched());
5539 EXPORT_SYMBOL(preempt_schedule
);
5542 * this is the entry point to schedule() from kernel preemption
5543 * off of irq context.
5544 * Note, that this is called and return with irqs disabled. This will
5545 * protect us against recursive calling from irq.
5547 asmlinkage
void __sched
preempt_schedule_irq(void)
5549 struct thread_info
*ti
= current_thread_info();
5551 /* Catch callers which need to be fixed */
5552 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5555 add_preempt_count(PREEMPT_ACTIVE
);
5558 local_irq_disable();
5559 sub_preempt_count(PREEMPT_ACTIVE
);
5562 * Check again in case we missed a preemption opportunity
5563 * between schedule and now.
5566 } while (need_resched());
5569 #endif /* CONFIG_PREEMPT */
5571 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5574 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5576 EXPORT_SYMBOL(default_wake_function
);
5579 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5580 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5581 * number) then we wake all the non-exclusive tasks and one exclusive task.
5583 * There are circumstances in which we can try to wake a task which has already
5584 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5585 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5587 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5588 int nr_exclusive
, int wake_flags
, void *key
)
5590 wait_queue_t
*curr
, *next
;
5592 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5593 unsigned flags
= curr
->flags
;
5595 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5596 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5602 * __wake_up - wake up threads blocked on a waitqueue.
5604 * @mode: which threads
5605 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5606 * @key: is directly passed to the wakeup function
5608 * It may be assumed that this function implies a write memory barrier before
5609 * changing the task state if and only if any tasks are woken up.
5611 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5612 int nr_exclusive
, void *key
)
5614 unsigned long flags
;
5616 spin_lock_irqsave(&q
->lock
, flags
);
5617 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5618 spin_unlock_irqrestore(&q
->lock
, flags
);
5620 EXPORT_SYMBOL(__wake_up
);
5623 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5625 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5627 __wake_up_common(q
, mode
, 1, 0, NULL
);
5630 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5632 __wake_up_common(q
, mode
, 1, 0, key
);
5636 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5638 * @mode: which threads
5639 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5640 * @key: opaque value to be passed to wakeup targets
5642 * The sync wakeup differs that the waker knows that it will schedule
5643 * away soon, so while the target thread will be woken up, it will not
5644 * be migrated to another CPU - ie. the two threads are 'synchronized'
5645 * with each other. This can prevent needless bouncing between CPUs.
5647 * On UP it can prevent extra preemption.
5649 * It may be assumed that this function implies a write memory barrier before
5650 * changing the task state if and only if any tasks are woken up.
5652 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5653 int nr_exclusive
, void *key
)
5655 unsigned long flags
;
5656 int wake_flags
= WF_SYNC
;
5661 if (unlikely(!nr_exclusive
))
5664 spin_lock_irqsave(&q
->lock
, flags
);
5665 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5666 spin_unlock_irqrestore(&q
->lock
, flags
);
5668 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5671 * __wake_up_sync - see __wake_up_sync_key()
5673 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5675 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5677 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5680 * complete: - signals a single thread waiting on this completion
5681 * @x: holds the state of this particular completion
5683 * This will wake up a single thread waiting on this completion. Threads will be
5684 * awakened in the same order in which they were queued.
5686 * See also complete_all(), wait_for_completion() and related routines.
5688 * It may be assumed that this function implies a write memory barrier before
5689 * changing the task state if and only if any tasks are woken up.
5691 void complete(struct completion
*x
)
5693 unsigned long flags
;
5695 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5697 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5698 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5700 EXPORT_SYMBOL(complete
);
5703 * complete_all: - signals all threads waiting on this completion
5704 * @x: holds the state of this particular completion
5706 * This will wake up all threads waiting on this particular completion event.
5708 * It may be assumed that this function implies a write memory barrier before
5709 * changing the task state if and only if any tasks are woken up.
5711 void complete_all(struct completion
*x
)
5713 unsigned long flags
;
5715 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5716 x
->done
+= UINT_MAX
/2;
5717 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5718 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5720 EXPORT_SYMBOL(complete_all
);
5722 static inline long __sched
5723 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5726 DECLARE_WAITQUEUE(wait
, current
);
5728 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5729 __add_wait_queue_tail(&x
->wait
, &wait
);
5731 if (signal_pending_state(state
, current
)) {
5732 timeout
= -ERESTARTSYS
;
5735 __set_current_state(state
);
5736 spin_unlock_irq(&x
->wait
.lock
);
5737 timeout
= schedule_timeout(timeout
);
5738 spin_lock_irq(&x
->wait
.lock
);
5739 } while (!x
->done
&& timeout
);
5740 __remove_wait_queue(&x
->wait
, &wait
);
5745 return timeout
?: 1;
5749 wait_for_common(struct completion
*x
, long timeout
, int state
)
5753 spin_lock_irq(&x
->wait
.lock
);
5754 timeout
= do_wait_for_common(x
, timeout
, state
);
5755 spin_unlock_irq(&x
->wait
.lock
);
5760 * wait_for_completion: - waits for completion of a task
5761 * @x: holds the state of this particular completion
5763 * This waits to be signaled for completion of a specific task. It is NOT
5764 * interruptible and there is no timeout.
5766 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5767 * and interrupt capability. Also see complete().
5769 void __sched
wait_for_completion(struct completion
*x
)
5771 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5773 EXPORT_SYMBOL(wait_for_completion
);
5776 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5777 * @x: holds the state of this particular completion
5778 * @timeout: timeout value in jiffies
5780 * This waits for either a completion of a specific task to be signaled or for a
5781 * specified timeout to expire. The timeout is in jiffies. It is not
5784 unsigned long __sched
5785 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5787 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5789 EXPORT_SYMBOL(wait_for_completion_timeout
);
5792 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5793 * @x: holds the state of this particular completion
5795 * This waits for completion of a specific task to be signaled. It is
5798 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5800 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5801 if (t
== -ERESTARTSYS
)
5805 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5808 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5809 * @x: holds the state of this particular completion
5810 * @timeout: timeout value in jiffies
5812 * This waits for either a completion of a specific task to be signaled or for a
5813 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5815 unsigned long __sched
5816 wait_for_completion_interruptible_timeout(struct completion
*x
,
5817 unsigned long timeout
)
5819 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5821 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5824 * wait_for_completion_killable: - waits for completion of a task (killable)
5825 * @x: holds the state of this particular completion
5827 * This waits to be signaled for completion of a specific task. It can be
5828 * interrupted by a kill signal.
5830 int __sched
wait_for_completion_killable(struct completion
*x
)
5832 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5833 if (t
== -ERESTARTSYS
)
5837 EXPORT_SYMBOL(wait_for_completion_killable
);
5840 * try_wait_for_completion - try to decrement a completion without blocking
5841 * @x: completion structure
5843 * Returns: 0 if a decrement cannot be done without blocking
5844 * 1 if a decrement succeeded.
5846 * If a completion is being used as a counting completion,
5847 * attempt to decrement the counter without blocking. This
5848 * enables us to avoid waiting if the resource the completion
5849 * is protecting is not available.
5851 bool try_wait_for_completion(struct completion
*x
)
5855 spin_lock_irq(&x
->wait
.lock
);
5860 spin_unlock_irq(&x
->wait
.lock
);
5863 EXPORT_SYMBOL(try_wait_for_completion
);
5866 * completion_done - Test to see if a completion has any waiters
5867 * @x: completion structure
5869 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5870 * 1 if there are no waiters.
5873 bool completion_done(struct completion
*x
)
5877 spin_lock_irq(&x
->wait
.lock
);
5880 spin_unlock_irq(&x
->wait
.lock
);
5883 EXPORT_SYMBOL(completion_done
);
5886 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5888 unsigned long flags
;
5891 init_waitqueue_entry(&wait
, current
);
5893 __set_current_state(state
);
5895 spin_lock_irqsave(&q
->lock
, flags
);
5896 __add_wait_queue(q
, &wait
);
5897 spin_unlock(&q
->lock
);
5898 timeout
= schedule_timeout(timeout
);
5899 spin_lock_irq(&q
->lock
);
5900 __remove_wait_queue(q
, &wait
);
5901 spin_unlock_irqrestore(&q
->lock
, flags
);
5906 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5908 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5910 EXPORT_SYMBOL(interruptible_sleep_on
);
5913 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5915 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5917 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5919 void __sched
sleep_on(wait_queue_head_t
*q
)
5921 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5923 EXPORT_SYMBOL(sleep_on
);
5925 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5927 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5929 EXPORT_SYMBOL(sleep_on_timeout
);
5931 #ifdef CONFIG_RT_MUTEXES
5934 * rt_mutex_setprio - set the current priority of a task
5936 * @prio: prio value (kernel-internal form)
5938 * This function changes the 'effective' priority of a task. It does
5939 * not touch ->normal_prio like __setscheduler().
5941 * Used by the rt_mutex code to implement priority inheritance logic.
5943 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5945 unsigned long flags
;
5946 int oldprio
, on_rq
, running
;
5948 const struct sched_class
*prev_class
= p
->sched_class
;
5950 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5952 rq
= task_rq_lock(p
, &flags
);
5953 update_rq_clock(rq
);
5956 on_rq
= p
->se
.on_rq
;
5957 running
= task_current(rq
, p
);
5959 dequeue_task(rq
, p
, 0);
5961 p
->sched_class
->put_prev_task(rq
, p
);
5964 p
->sched_class
= &rt_sched_class
;
5966 p
->sched_class
= &fair_sched_class
;
5971 p
->sched_class
->set_curr_task(rq
);
5973 enqueue_task(rq
, p
, 0);
5975 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5977 task_rq_unlock(rq
, &flags
);
5982 void set_user_nice(struct task_struct
*p
, long nice
)
5984 int old_prio
, delta
, on_rq
;
5985 unsigned long flags
;
5988 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5991 * We have to be careful, if called from sys_setpriority(),
5992 * the task might be in the middle of scheduling on another CPU.
5994 rq
= task_rq_lock(p
, &flags
);
5995 update_rq_clock(rq
);
5997 * The RT priorities are set via sched_setscheduler(), but we still
5998 * allow the 'normal' nice value to be set - but as expected
5999 * it wont have any effect on scheduling until the task is
6000 * SCHED_FIFO/SCHED_RR:
6002 if (task_has_rt_policy(p
)) {
6003 p
->static_prio
= NICE_TO_PRIO(nice
);
6006 on_rq
= p
->se
.on_rq
;
6008 dequeue_task(rq
, p
, 0);
6010 p
->static_prio
= NICE_TO_PRIO(nice
);
6013 p
->prio
= effective_prio(p
);
6014 delta
= p
->prio
- old_prio
;
6017 enqueue_task(rq
, p
, 0);
6019 * If the task increased its priority or is running and
6020 * lowered its priority, then reschedule its CPU:
6022 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6023 resched_task(rq
->curr
);
6026 task_rq_unlock(rq
, &flags
);
6028 EXPORT_SYMBOL(set_user_nice
);
6031 * can_nice - check if a task can reduce its nice value
6035 int can_nice(const struct task_struct
*p
, const int nice
)
6037 /* convert nice value [19,-20] to rlimit style value [1,40] */
6038 int nice_rlim
= 20 - nice
;
6040 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6041 capable(CAP_SYS_NICE
));
6044 #ifdef __ARCH_WANT_SYS_NICE
6047 * sys_nice - change the priority of the current process.
6048 * @increment: priority increment
6050 * sys_setpriority is a more generic, but much slower function that
6051 * does similar things.
6053 SYSCALL_DEFINE1(nice
, int, increment
)
6058 * Setpriority might change our priority at the same moment.
6059 * We don't have to worry. Conceptually one call occurs first
6060 * and we have a single winner.
6062 if (increment
< -40)
6067 nice
= TASK_NICE(current
) + increment
;
6073 if (increment
< 0 && !can_nice(current
, nice
))
6076 retval
= security_task_setnice(current
, nice
);
6080 set_user_nice(current
, nice
);
6087 * task_prio - return the priority value of a given task.
6088 * @p: the task in question.
6090 * This is the priority value as seen by users in /proc.
6091 * RT tasks are offset by -200. Normal tasks are centered
6092 * around 0, value goes from -16 to +15.
6094 int task_prio(const struct task_struct
*p
)
6096 return p
->prio
- MAX_RT_PRIO
;
6100 * task_nice - return the nice value of a given task.
6101 * @p: the task in question.
6103 int task_nice(const struct task_struct
*p
)
6105 return TASK_NICE(p
);
6107 EXPORT_SYMBOL(task_nice
);
6110 * idle_cpu - is a given cpu idle currently?
6111 * @cpu: the processor in question.
6113 int idle_cpu(int cpu
)
6115 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6119 * idle_task - return the idle task for a given cpu.
6120 * @cpu: the processor in question.
6122 struct task_struct
*idle_task(int cpu
)
6124 return cpu_rq(cpu
)->idle
;
6128 * find_process_by_pid - find a process with a matching PID value.
6129 * @pid: the pid in question.
6131 static struct task_struct
*find_process_by_pid(pid_t pid
)
6133 return pid
? find_task_by_vpid(pid
) : current
;
6136 /* Actually do priority change: must hold rq lock. */
6138 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6140 BUG_ON(p
->se
.on_rq
);
6143 switch (p
->policy
) {
6147 p
->sched_class
= &fair_sched_class
;
6151 p
->sched_class
= &rt_sched_class
;
6155 p
->rt_priority
= prio
;
6156 p
->normal_prio
= normal_prio(p
);
6157 /* we are holding p->pi_lock already */
6158 p
->prio
= rt_mutex_getprio(p
);
6163 * check the target process has a UID that matches the current process's
6165 static bool check_same_owner(struct task_struct
*p
)
6167 const struct cred
*cred
= current_cred(), *pcred
;
6171 pcred
= __task_cred(p
);
6172 match
= (cred
->euid
== pcred
->euid
||
6173 cred
->euid
== pcred
->uid
);
6178 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6179 struct sched_param
*param
, bool user
)
6181 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6182 unsigned long flags
;
6183 const struct sched_class
*prev_class
= p
->sched_class
;
6187 /* may grab non-irq protected spin_locks */
6188 BUG_ON(in_interrupt());
6190 /* double check policy once rq lock held */
6192 reset_on_fork
= p
->sched_reset_on_fork
;
6193 policy
= oldpolicy
= p
->policy
;
6195 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6196 policy
&= ~SCHED_RESET_ON_FORK
;
6198 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6199 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6200 policy
!= SCHED_IDLE
)
6205 * Valid priorities for SCHED_FIFO and SCHED_RR are
6206 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6207 * SCHED_BATCH and SCHED_IDLE is 0.
6209 if (param
->sched_priority
< 0 ||
6210 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6211 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6213 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6217 * Allow unprivileged RT tasks to decrease priority:
6219 if (user
&& !capable(CAP_SYS_NICE
)) {
6220 if (rt_policy(policy
)) {
6221 unsigned long rlim_rtprio
;
6223 if (!lock_task_sighand(p
, &flags
))
6225 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6226 unlock_task_sighand(p
, &flags
);
6228 /* can't set/change the rt policy */
6229 if (policy
!= p
->policy
&& !rlim_rtprio
)
6232 /* can't increase priority */
6233 if (param
->sched_priority
> p
->rt_priority
&&
6234 param
->sched_priority
> rlim_rtprio
)
6238 * Like positive nice levels, dont allow tasks to
6239 * move out of SCHED_IDLE either:
6241 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6244 /* can't change other user's priorities */
6245 if (!check_same_owner(p
))
6248 /* Normal users shall not reset the sched_reset_on_fork flag */
6249 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6254 #ifdef CONFIG_RT_GROUP_SCHED
6256 * Do not allow realtime tasks into groups that have no runtime
6259 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6260 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6264 retval
= security_task_setscheduler(p
, policy
, param
);
6270 * make sure no PI-waiters arrive (or leave) while we are
6271 * changing the priority of the task:
6273 spin_lock_irqsave(&p
->pi_lock
, flags
);
6275 * To be able to change p->policy safely, the apropriate
6276 * runqueue lock must be held.
6278 rq
= __task_rq_lock(p
);
6279 /* recheck policy now with rq lock held */
6280 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6281 policy
= oldpolicy
= -1;
6282 __task_rq_unlock(rq
);
6283 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6286 update_rq_clock(rq
);
6287 on_rq
= p
->se
.on_rq
;
6288 running
= task_current(rq
, p
);
6290 deactivate_task(rq
, p
, 0);
6292 p
->sched_class
->put_prev_task(rq
, p
);
6294 p
->sched_reset_on_fork
= reset_on_fork
;
6297 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6300 p
->sched_class
->set_curr_task(rq
);
6302 activate_task(rq
, p
, 0);
6304 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6306 __task_rq_unlock(rq
);
6307 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6309 rt_mutex_adjust_pi(p
);
6315 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6316 * @p: the task in question.
6317 * @policy: new policy.
6318 * @param: structure containing the new RT priority.
6320 * NOTE that the task may be already dead.
6322 int sched_setscheduler(struct task_struct
*p
, int policy
,
6323 struct sched_param
*param
)
6325 return __sched_setscheduler(p
, policy
, param
, true);
6327 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6330 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6331 * @p: the task in question.
6332 * @policy: new policy.
6333 * @param: structure containing the new RT priority.
6335 * Just like sched_setscheduler, only don't bother checking if the
6336 * current context has permission. For example, this is needed in
6337 * stop_machine(): we create temporary high priority worker threads,
6338 * but our caller might not have that capability.
6340 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6341 struct sched_param
*param
)
6343 return __sched_setscheduler(p
, policy
, param
, false);
6347 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6349 struct sched_param lparam
;
6350 struct task_struct
*p
;
6353 if (!param
|| pid
< 0)
6355 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6360 p
= find_process_by_pid(pid
);
6362 retval
= sched_setscheduler(p
, policy
, &lparam
);
6369 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6370 * @pid: the pid in question.
6371 * @policy: new policy.
6372 * @param: structure containing the new RT priority.
6374 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6375 struct sched_param __user
*, param
)
6377 /* negative values for policy are not valid */
6381 return do_sched_setscheduler(pid
, policy
, param
);
6385 * sys_sched_setparam - set/change the RT priority of a thread
6386 * @pid: the pid in question.
6387 * @param: structure containing the new RT priority.
6389 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6391 return do_sched_setscheduler(pid
, -1, param
);
6395 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6396 * @pid: the pid in question.
6398 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6400 struct task_struct
*p
;
6407 read_lock(&tasklist_lock
);
6408 p
= find_process_by_pid(pid
);
6410 retval
= security_task_getscheduler(p
);
6413 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6415 read_unlock(&tasklist_lock
);
6420 * sys_sched_getparam - get the RT priority of a thread
6421 * @pid: the pid in question.
6422 * @param: structure containing the RT priority.
6424 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6426 struct sched_param lp
;
6427 struct task_struct
*p
;
6430 if (!param
|| pid
< 0)
6433 read_lock(&tasklist_lock
);
6434 p
= find_process_by_pid(pid
);
6439 retval
= security_task_getscheduler(p
);
6443 lp
.sched_priority
= p
->rt_priority
;
6444 read_unlock(&tasklist_lock
);
6447 * This one might sleep, we cannot do it with a spinlock held ...
6449 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6454 read_unlock(&tasklist_lock
);
6458 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6460 cpumask_var_t cpus_allowed
, new_mask
;
6461 struct task_struct
*p
;
6465 read_lock(&tasklist_lock
);
6467 p
= find_process_by_pid(pid
);
6469 read_unlock(&tasklist_lock
);
6475 * It is not safe to call set_cpus_allowed with the
6476 * tasklist_lock held. We will bump the task_struct's
6477 * usage count and then drop tasklist_lock.
6480 read_unlock(&tasklist_lock
);
6482 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6486 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6488 goto out_free_cpus_allowed
;
6491 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6494 retval
= security_task_setscheduler(p
, 0, NULL
);
6498 cpuset_cpus_allowed(p
, cpus_allowed
);
6499 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6501 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6504 cpuset_cpus_allowed(p
, cpus_allowed
);
6505 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6507 * We must have raced with a concurrent cpuset
6508 * update. Just reset the cpus_allowed to the
6509 * cpuset's cpus_allowed
6511 cpumask_copy(new_mask
, cpus_allowed
);
6516 free_cpumask_var(new_mask
);
6517 out_free_cpus_allowed
:
6518 free_cpumask_var(cpus_allowed
);
6525 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6526 struct cpumask
*new_mask
)
6528 if (len
< cpumask_size())
6529 cpumask_clear(new_mask
);
6530 else if (len
> cpumask_size())
6531 len
= cpumask_size();
6533 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6537 * sys_sched_setaffinity - set the cpu affinity of a process
6538 * @pid: pid of the process
6539 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6540 * @user_mask_ptr: user-space pointer to the new cpu mask
6542 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6543 unsigned long __user
*, user_mask_ptr
)
6545 cpumask_var_t new_mask
;
6548 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6551 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6553 retval
= sched_setaffinity(pid
, new_mask
);
6554 free_cpumask_var(new_mask
);
6558 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6560 struct task_struct
*p
;
6564 read_lock(&tasklist_lock
);
6567 p
= find_process_by_pid(pid
);
6571 retval
= security_task_getscheduler(p
);
6575 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6578 read_unlock(&tasklist_lock
);
6585 * sys_sched_getaffinity - get the cpu affinity of a process
6586 * @pid: pid of the process
6587 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6588 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6590 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6591 unsigned long __user
*, user_mask_ptr
)
6596 if (len
< cpumask_size())
6599 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6602 ret
= sched_getaffinity(pid
, mask
);
6604 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6607 ret
= cpumask_size();
6609 free_cpumask_var(mask
);
6615 * sys_sched_yield - yield the current processor to other threads.
6617 * This function yields the current CPU to other tasks. If there are no
6618 * other threads running on this CPU then this function will return.
6620 SYSCALL_DEFINE0(sched_yield
)
6622 struct rq
*rq
= this_rq_lock();
6624 schedstat_inc(rq
, yld_count
);
6625 current
->sched_class
->yield_task(rq
);
6628 * Since we are going to call schedule() anyway, there's
6629 * no need to preempt or enable interrupts:
6631 __release(rq
->lock
);
6632 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6633 _raw_spin_unlock(&rq
->lock
);
6634 preempt_enable_no_resched();
6641 static inline int should_resched(void)
6643 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6646 static void __cond_resched(void)
6648 add_preempt_count(PREEMPT_ACTIVE
);
6650 sub_preempt_count(PREEMPT_ACTIVE
);
6653 int __sched
_cond_resched(void)
6655 if (should_resched()) {
6661 EXPORT_SYMBOL(_cond_resched
);
6664 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6665 * call schedule, and on return reacquire the lock.
6667 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6668 * operations here to prevent schedule() from being called twice (once via
6669 * spin_unlock(), once by hand).
6671 int __cond_resched_lock(spinlock_t
*lock
)
6673 int resched
= should_resched();
6676 lockdep_assert_held(lock
);
6678 if (spin_needbreak(lock
) || resched
) {
6689 EXPORT_SYMBOL(__cond_resched_lock
);
6691 int __sched
__cond_resched_softirq(void)
6693 BUG_ON(!in_softirq());
6695 if (should_resched()) {
6703 EXPORT_SYMBOL(__cond_resched_softirq
);
6706 * yield - yield the current processor to other threads.
6708 * This is a shortcut for kernel-space yielding - it marks the
6709 * thread runnable and calls sys_sched_yield().
6711 void __sched
yield(void)
6713 set_current_state(TASK_RUNNING
);
6716 EXPORT_SYMBOL(yield
);
6719 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6720 * that process accounting knows that this is a task in IO wait state.
6722 void __sched
io_schedule(void)
6724 struct rq
*rq
= raw_rq();
6726 delayacct_blkio_start();
6727 atomic_inc(&rq
->nr_iowait
);
6728 current
->in_iowait
= 1;
6730 current
->in_iowait
= 0;
6731 atomic_dec(&rq
->nr_iowait
);
6732 delayacct_blkio_end();
6734 EXPORT_SYMBOL(io_schedule
);
6736 long __sched
io_schedule_timeout(long timeout
)
6738 struct rq
*rq
= raw_rq();
6741 delayacct_blkio_start();
6742 atomic_inc(&rq
->nr_iowait
);
6743 current
->in_iowait
= 1;
6744 ret
= schedule_timeout(timeout
);
6745 current
->in_iowait
= 0;
6746 atomic_dec(&rq
->nr_iowait
);
6747 delayacct_blkio_end();
6752 * sys_sched_get_priority_max - return maximum RT priority.
6753 * @policy: scheduling class.
6755 * this syscall returns the maximum rt_priority that can be used
6756 * by a given scheduling class.
6758 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6765 ret
= MAX_USER_RT_PRIO
-1;
6777 * sys_sched_get_priority_min - return minimum RT priority.
6778 * @policy: scheduling class.
6780 * this syscall returns the minimum rt_priority that can be used
6781 * by a given scheduling class.
6783 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6801 * sys_sched_rr_get_interval - return the default timeslice of a process.
6802 * @pid: pid of the process.
6803 * @interval: userspace pointer to the timeslice value.
6805 * this syscall writes the default timeslice value of a given process
6806 * into the user-space timespec buffer. A value of '0' means infinity.
6808 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6809 struct timespec __user
*, interval
)
6811 struct task_struct
*p
;
6812 unsigned int time_slice
;
6820 read_lock(&tasklist_lock
);
6821 p
= find_process_by_pid(pid
);
6825 retval
= security_task_getscheduler(p
);
6829 time_slice
= p
->sched_class
->get_rr_interval(p
);
6831 read_unlock(&tasklist_lock
);
6832 jiffies_to_timespec(time_slice
, &t
);
6833 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6837 read_unlock(&tasklist_lock
);
6841 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6843 void sched_show_task(struct task_struct
*p
)
6845 unsigned long free
= 0;
6848 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6849 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6850 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6851 #if BITS_PER_LONG == 32
6852 if (state
== TASK_RUNNING
)
6853 printk(KERN_CONT
" running ");
6855 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6857 if (state
== TASK_RUNNING
)
6858 printk(KERN_CONT
" running task ");
6860 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6862 #ifdef CONFIG_DEBUG_STACK_USAGE
6863 free
= stack_not_used(p
);
6865 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6866 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6867 (unsigned long)task_thread_info(p
)->flags
);
6869 show_stack(p
, NULL
);
6872 void show_state_filter(unsigned long state_filter
)
6874 struct task_struct
*g
, *p
;
6876 #if BITS_PER_LONG == 32
6878 " task PC stack pid father\n");
6881 " task PC stack pid father\n");
6883 read_lock(&tasklist_lock
);
6884 do_each_thread(g
, p
) {
6886 * reset the NMI-timeout, listing all files on a slow
6887 * console might take alot of time:
6889 touch_nmi_watchdog();
6890 if (!state_filter
|| (p
->state
& state_filter
))
6892 } while_each_thread(g
, p
);
6894 touch_all_softlockup_watchdogs();
6896 #ifdef CONFIG_SCHED_DEBUG
6897 sysrq_sched_debug_show();
6899 read_unlock(&tasklist_lock
);
6901 * Only show locks if all tasks are dumped:
6903 if (state_filter
== -1)
6904 debug_show_all_locks();
6907 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6909 idle
->sched_class
= &idle_sched_class
;
6913 * init_idle - set up an idle thread for a given CPU
6914 * @idle: task in question
6915 * @cpu: cpu the idle task belongs to
6917 * NOTE: this function does not set the idle thread's NEED_RESCHED
6918 * flag, to make booting more robust.
6920 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6922 struct rq
*rq
= cpu_rq(cpu
);
6923 unsigned long flags
;
6925 spin_lock_irqsave(&rq
->lock
, flags
);
6928 idle
->se
.exec_start
= sched_clock();
6930 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6931 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6932 __set_task_cpu(idle
, cpu
);
6934 rq
->curr
= rq
->idle
= idle
;
6935 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6938 spin_unlock_irqrestore(&rq
->lock
, flags
);
6940 /* Set the preempt count _outside_ the spinlocks! */
6941 #if defined(CONFIG_PREEMPT)
6942 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6944 task_thread_info(idle
)->preempt_count
= 0;
6947 * The idle tasks have their own, simple scheduling class:
6949 idle
->sched_class
= &idle_sched_class
;
6950 ftrace_graph_init_task(idle
);
6954 * In a system that switches off the HZ timer nohz_cpu_mask
6955 * indicates which cpus entered this state. This is used
6956 * in the rcu update to wait only for active cpus. For system
6957 * which do not switch off the HZ timer nohz_cpu_mask should
6958 * always be CPU_BITS_NONE.
6960 cpumask_var_t nohz_cpu_mask
;
6963 * Increase the granularity value when there are more CPUs,
6964 * because with more CPUs the 'effective latency' as visible
6965 * to users decreases. But the relationship is not linear,
6966 * so pick a second-best guess by going with the log2 of the
6969 * This idea comes from the SD scheduler of Con Kolivas:
6971 static inline void sched_init_granularity(void)
6973 unsigned int factor
= 1 + ilog2(num_online_cpus());
6974 const unsigned long limit
= 200000000;
6976 sysctl_sched_min_granularity
*= factor
;
6977 if (sysctl_sched_min_granularity
> limit
)
6978 sysctl_sched_min_granularity
= limit
;
6980 sysctl_sched_latency
*= factor
;
6981 if (sysctl_sched_latency
> limit
)
6982 sysctl_sched_latency
= limit
;
6984 sysctl_sched_wakeup_granularity
*= factor
;
6986 sysctl_sched_shares_ratelimit
*= factor
;
6991 * This is how migration works:
6993 * 1) we queue a struct migration_req structure in the source CPU's
6994 * runqueue and wake up that CPU's migration thread.
6995 * 2) we down() the locked semaphore => thread blocks.
6996 * 3) migration thread wakes up (implicitly it forces the migrated
6997 * thread off the CPU)
6998 * 4) it gets the migration request and checks whether the migrated
6999 * task is still in the wrong runqueue.
7000 * 5) if it's in the wrong runqueue then the migration thread removes
7001 * it and puts it into the right queue.
7002 * 6) migration thread up()s the semaphore.
7003 * 7) we wake up and the migration is done.
7007 * Change a given task's CPU affinity. Migrate the thread to a
7008 * proper CPU and schedule it away if the CPU it's executing on
7009 * is removed from the allowed bitmask.
7011 * NOTE: the caller must have a valid reference to the task, the
7012 * task must not exit() & deallocate itself prematurely. The
7013 * call is not atomic; no spinlocks may be held.
7015 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7017 struct migration_req req
;
7018 unsigned long flags
;
7022 rq
= task_rq_lock(p
, &flags
);
7023 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7028 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7029 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7034 if (p
->sched_class
->set_cpus_allowed
)
7035 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7037 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7038 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7041 /* Can the task run on the task's current CPU? If so, we're done */
7042 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7045 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7046 /* Need help from migration thread: drop lock and wait. */
7047 struct task_struct
*mt
= rq
->migration_thread
;
7049 get_task_struct(mt
);
7050 task_rq_unlock(rq
, &flags
);
7051 wake_up_process(rq
->migration_thread
);
7052 put_task_struct(mt
);
7053 wait_for_completion(&req
.done
);
7054 tlb_migrate_finish(p
->mm
);
7058 task_rq_unlock(rq
, &flags
);
7062 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7065 * Move (not current) task off this cpu, onto dest cpu. We're doing
7066 * this because either it can't run here any more (set_cpus_allowed()
7067 * away from this CPU, or CPU going down), or because we're
7068 * attempting to rebalance this task on exec (sched_exec).
7070 * So we race with normal scheduler movements, but that's OK, as long
7071 * as the task is no longer on this CPU.
7073 * Returns non-zero if task was successfully migrated.
7075 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7077 struct rq
*rq_dest
, *rq_src
;
7080 if (unlikely(!cpu_active(dest_cpu
)))
7083 rq_src
= cpu_rq(src_cpu
);
7084 rq_dest
= cpu_rq(dest_cpu
);
7086 double_rq_lock(rq_src
, rq_dest
);
7087 /* Already moved. */
7088 if (task_cpu(p
) != src_cpu
)
7090 /* Affinity changed (again). */
7091 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7094 on_rq
= p
->se
.on_rq
;
7096 deactivate_task(rq_src
, p
, 0);
7098 set_task_cpu(p
, dest_cpu
);
7100 activate_task(rq_dest
, p
, 0);
7101 check_preempt_curr(rq_dest
, p
, 0);
7106 double_rq_unlock(rq_src
, rq_dest
);
7110 #define RCU_MIGRATION_IDLE 0
7111 #define RCU_MIGRATION_NEED_QS 1
7112 #define RCU_MIGRATION_GOT_QS 2
7113 #define RCU_MIGRATION_MUST_SYNC 3
7116 * migration_thread - this is a highprio system thread that performs
7117 * thread migration by bumping thread off CPU then 'pushing' onto
7120 static int migration_thread(void *data
)
7123 int cpu
= (long)data
;
7127 BUG_ON(rq
->migration_thread
!= current
);
7129 set_current_state(TASK_INTERRUPTIBLE
);
7130 while (!kthread_should_stop()) {
7131 struct migration_req
*req
;
7132 struct list_head
*head
;
7134 spin_lock_irq(&rq
->lock
);
7136 if (cpu_is_offline(cpu
)) {
7137 spin_unlock_irq(&rq
->lock
);
7141 if (rq
->active_balance
) {
7142 active_load_balance(rq
, cpu
);
7143 rq
->active_balance
= 0;
7146 head
= &rq
->migration_queue
;
7148 if (list_empty(head
)) {
7149 spin_unlock_irq(&rq
->lock
);
7151 set_current_state(TASK_INTERRUPTIBLE
);
7154 req
= list_entry(head
->next
, struct migration_req
, list
);
7155 list_del_init(head
->next
);
7157 if (req
->task
!= NULL
) {
7158 spin_unlock(&rq
->lock
);
7159 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7160 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7161 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7162 spin_unlock(&rq
->lock
);
7164 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7165 spin_unlock(&rq
->lock
);
7166 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7170 complete(&req
->done
);
7172 __set_current_state(TASK_RUNNING
);
7177 #ifdef CONFIG_HOTPLUG_CPU
7179 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7183 local_irq_disable();
7184 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7190 * Figure out where task on dead CPU should go, use force if necessary.
7192 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7195 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7198 /* Look for allowed, online CPU in same node. */
7199 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7200 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7203 /* Any allowed, online CPU? */
7204 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7205 if (dest_cpu
< nr_cpu_ids
)
7208 /* No more Mr. Nice Guy. */
7209 if (dest_cpu
>= nr_cpu_ids
) {
7210 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7211 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7214 * Don't tell them about moving exiting tasks or
7215 * kernel threads (both mm NULL), since they never
7218 if (p
->mm
&& printk_ratelimit()) {
7219 printk(KERN_INFO
"process %d (%s) no "
7220 "longer affine to cpu%d\n",
7221 task_pid_nr(p
), p
->comm
, dead_cpu
);
7226 /* It can have affinity changed while we were choosing. */
7227 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7232 * While a dead CPU has no uninterruptible tasks queued at this point,
7233 * it might still have a nonzero ->nr_uninterruptible counter, because
7234 * for performance reasons the counter is not stricly tracking tasks to
7235 * their home CPUs. So we just add the counter to another CPU's counter,
7236 * to keep the global sum constant after CPU-down:
7238 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7240 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7241 unsigned long flags
;
7243 local_irq_save(flags
);
7244 double_rq_lock(rq_src
, rq_dest
);
7245 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7246 rq_src
->nr_uninterruptible
= 0;
7247 double_rq_unlock(rq_src
, rq_dest
);
7248 local_irq_restore(flags
);
7251 /* Run through task list and migrate tasks from the dead cpu. */
7252 static void migrate_live_tasks(int src_cpu
)
7254 struct task_struct
*p
, *t
;
7256 read_lock(&tasklist_lock
);
7258 do_each_thread(t
, p
) {
7262 if (task_cpu(p
) == src_cpu
)
7263 move_task_off_dead_cpu(src_cpu
, p
);
7264 } while_each_thread(t
, p
);
7266 read_unlock(&tasklist_lock
);
7270 * Schedules idle task to be the next runnable task on current CPU.
7271 * It does so by boosting its priority to highest possible.
7272 * Used by CPU offline code.
7274 void sched_idle_next(void)
7276 int this_cpu
= smp_processor_id();
7277 struct rq
*rq
= cpu_rq(this_cpu
);
7278 struct task_struct
*p
= rq
->idle
;
7279 unsigned long flags
;
7281 /* cpu has to be offline */
7282 BUG_ON(cpu_online(this_cpu
));
7285 * Strictly not necessary since rest of the CPUs are stopped by now
7286 * and interrupts disabled on the current cpu.
7288 spin_lock_irqsave(&rq
->lock
, flags
);
7290 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7292 update_rq_clock(rq
);
7293 activate_task(rq
, p
, 0);
7295 spin_unlock_irqrestore(&rq
->lock
, flags
);
7299 * Ensures that the idle task is using init_mm right before its cpu goes
7302 void idle_task_exit(void)
7304 struct mm_struct
*mm
= current
->active_mm
;
7306 BUG_ON(cpu_online(smp_processor_id()));
7309 switch_mm(mm
, &init_mm
, current
);
7313 /* called under rq->lock with disabled interrupts */
7314 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7316 struct rq
*rq
= cpu_rq(dead_cpu
);
7318 /* Must be exiting, otherwise would be on tasklist. */
7319 BUG_ON(!p
->exit_state
);
7321 /* Cannot have done final schedule yet: would have vanished. */
7322 BUG_ON(p
->state
== TASK_DEAD
);
7327 * Drop lock around migration; if someone else moves it,
7328 * that's OK. No task can be added to this CPU, so iteration is
7331 spin_unlock_irq(&rq
->lock
);
7332 move_task_off_dead_cpu(dead_cpu
, p
);
7333 spin_lock_irq(&rq
->lock
);
7338 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7339 static void migrate_dead_tasks(unsigned int dead_cpu
)
7341 struct rq
*rq
= cpu_rq(dead_cpu
);
7342 struct task_struct
*next
;
7345 if (!rq
->nr_running
)
7347 update_rq_clock(rq
);
7348 next
= pick_next_task(rq
);
7351 next
->sched_class
->put_prev_task(rq
, next
);
7352 migrate_dead(dead_cpu
, next
);
7358 * remove the tasks which were accounted by rq from calc_load_tasks.
7360 static void calc_global_load_remove(struct rq
*rq
)
7362 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7363 rq
->calc_load_active
= 0;
7365 #endif /* CONFIG_HOTPLUG_CPU */
7367 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7369 static struct ctl_table sd_ctl_dir
[] = {
7371 .procname
= "sched_domain",
7377 static struct ctl_table sd_ctl_root
[] = {
7379 .ctl_name
= CTL_KERN
,
7380 .procname
= "kernel",
7382 .child
= sd_ctl_dir
,
7387 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7389 struct ctl_table
*entry
=
7390 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7395 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7397 struct ctl_table
*entry
;
7400 * In the intermediate directories, both the child directory and
7401 * procname are dynamically allocated and could fail but the mode
7402 * will always be set. In the lowest directory the names are
7403 * static strings and all have proc handlers.
7405 for (entry
= *tablep
; entry
->mode
; entry
++) {
7407 sd_free_ctl_entry(&entry
->child
);
7408 if (entry
->proc_handler
== NULL
)
7409 kfree(entry
->procname
);
7417 set_table_entry(struct ctl_table
*entry
,
7418 const char *procname
, void *data
, int maxlen
,
7419 mode_t mode
, proc_handler
*proc_handler
)
7421 entry
->procname
= procname
;
7423 entry
->maxlen
= maxlen
;
7425 entry
->proc_handler
= proc_handler
;
7428 static struct ctl_table
*
7429 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7431 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7436 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7437 sizeof(long), 0644, proc_doulongvec_minmax
);
7438 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7439 sizeof(long), 0644, proc_doulongvec_minmax
);
7440 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7441 sizeof(int), 0644, proc_dointvec_minmax
);
7442 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7443 sizeof(int), 0644, proc_dointvec_minmax
);
7444 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7445 sizeof(int), 0644, proc_dointvec_minmax
);
7446 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7447 sizeof(int), 0644, proc_dointvec_minmax
);
7448 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7449 sizeof(int), 0644, proc_dointvec_minmax
);
7450 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7451 sizeof(int), 0644, proc_dointvec_minmax
);
7452 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7453 sizeof(int), 0644, proc_dointvec_minmax
);
7454 set_table_entry(&table
[9], "cache_nice_tries",
7455 &sd
->cache_nice_tries
,
7456 sizeof(int), 0644, proc_dointvec_minmax
);
7457 set_table_entry(&table
[10], "flags", &sd
->flags
,
7458 sizeof(int), 0644, proc_dointvec_minmax
);
7459 set_table_entry(&table
[11], "name", sd
->name
,
7460 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7461 /* &table[12] is terminator */
7466 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7468 struct ctl_table
*entry
, *table
;
7469 struct sched_domain
*sd
;
7470 int domain_num
= 0, i
;
7473 for_each_domain(cpu
, sd
)
7475 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7480 for_each_domain(cpu
, sd
) {
7481 snprintf(buf
, 32, "domain%d", i
);
7482 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7484 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7491 static struct ctl_table_header
*sd_sysctl_header
;
7492 static void register_sched_domain_sysctl(void)
7494 int i
, cpu_num
= num_online_cpus();
7495 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7498 WARN_ON(sd_ctl_dir
[0].child
);
7499 sd_ctl_dir
[0].child
= entry
;
7504 for_each_online_cpu(i
) {
7505 snprintf(buf
, 32, "cpu%d", i
);
7506 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7508 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7512 WARN_ON(sd_sysctl_header
);
7513 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7516 /* may be called multiple times per register */
7517 static void unregister_sched_domain_sysctl(void)
7519 if (sd_sysctl_header
)
7520 unregister_sysctl_table(sd_sysctl_header
);
7521 sd_sysctl_header
= NULL
;
7522 if (sd_ctl_dir
[0].child
)
7523 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7526 static void register_sched_domain_sysctl(void)
7529 static void unregister_sched_domain_sysctl(void)
7534 static void set_rq_online(struct rq
*rq
)
7537 const struct sched_class
*class;
7539 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7542 for_each_class(class) {
7543 if (class->rq_online
)
7544 class->rq_online(rq
);
7549 static void set_rq_offline(struct rq
*rq
)
7552 const struct sched_class
*class;
7554 for_each_class(class) {
7555 if (class->rq_offline
)
7556 class->rq_offline(rq
);
7559 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7565 * migration_call - callback that gets triggered when a CPU is added.
7566 * Here we can start up the necessary migration thread for the new CPU.
7568 static int __cpuinit
7569 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7571 struct task_struct
*p
;
7572 int cpu
= (long)hcpu
;
7573 unsigned long flags
;
7578 case CPU_UP_PREPARE
:
7579 case CPU_UP_PREPARE_FROZEN
:
7580 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7583 kthread_bind(p
, cpu
);
7584 /* Must be high prio: stop_machine expects to yield to it. */
7585 rq
= task_rq_lock(p
, &flags
);
7586 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7587 task_rq_unlock(rq
, &flags
);
7589 cpu_rq(cpu
)->migration_thread
= p
;
7590 rq
->calc_load_update
= calc_load_update
;
7594 case CPU_ONLINE_FROZEN
:
7595 /* Strictly unnecessary, as first user will wake it. */
7596 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7598 /* Update our root-domain */
7600 spin_lock_irqsave(&rq
->lock
, flags
);
7602 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7606 spin_unlock_irqrestore(&rq
->lock
, flags
);
7609 #ifdef CONFIG_HOTPLUG_CPU
7610 case CPU_UP_CANCELED
:
7611 case CPU_UP_CANCELED_FROZEN
:
7612 if (!cpu_rq(cpu
)->migration_thread
)
7614 /* Unbind it from offline cpu so it can run. Fall thru. */
7615 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7616 cpumask_any(cpu_online_mask
));
7617 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7618 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7619 cpu_rq(cpu
)->migration_thread
= NULL
;
7623 case CPU_DEAD_FROZEN
:
7624 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7625 migrate_live_tasks(cpu
);
7627 kthread_stop(rq
->migration_thread
);
7628 put_task_struct(rq
->migration_thread
);
7629 rq
->migration_thread
= NULL
;
7630 /* Idle task back to normal (off runqueue, low prio) */
7631 spin_lock_irq(&rq
->lock
);
7632 update_rq_clock(rq
);
7633 deactivate_task(rq
, rq
->idle
, 0);
7634 rq
->idle
->static_prio
= MAX_PRIO
;
7635 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7636 rq
->idle
->sched_class
= &idle_sched_class
;
7637 migrate_dead_tasks(cpu
);
7638 spin_unlock_irq(&rq
->lock
);
7640 migrate_nr_uninterruptible(rq
);
7641 BUG_ON(rq
->nr_running
!= 0);
7642 calc_global_load_remove(rq
);
7644 * No need to migrate the tasks: it was best-effort if
7645 * they didn't take sched_hotcpu_mutex. Just wake up
7648 spin_lock_irq(&rq
->lock
);
7649 while (!list_empty(&rq
->migration_queue
)) {
7650 struct migration_req
*req
;
7652 req
= list_entry(rq
->migration_queue
.next
,
7653 struct migration_req
, list
);
7654 list_del_init(&req
->list
);
7655 spin_unlock_irq(&rq
->lock
);
7656 complete(&req
->done
);
7657 spin_lock_irq(&rq
->lock
);
7659 spin_unlock_irq(&rq
->lock
);
7663 case CPU_DYING_FROZEN
:
7664 /* Update our root-domain */
7666 spin_lock_irqsave(&rq
->lock
, flags
);
7668 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7671 spin_unlock_irqrestore(&rq
->lock
, flags
);
7679 * Register at high priority so that task migration (migrate_all_tasks)
7680 * happens before everything else. This has to be lower priority than
7681 * the notifier in the perf_event subsystem, though.
7683 static struct notifier_block __cpuinitdata migration_notifier
= {
7684 .notifier_call
= migration_call
,
7688 static int __init
migration_init(void)
7690 void *cpu
= (void *)(long)smp_processor_id();
7693 /* Start one for the boot CPU: */
7694 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7695 BUG_ON(err
== NOTIFY_BAD
);
7696 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7697 register_cpu_notifier(&migration_notifier
);
7701 early_initcall(migration_init
);
7706 #ifdef CONFIG_SCHED_DEBUG
7708 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7709 struct cpumask
*groupmask
)
7711 struct sched_group
*group
= sd
->groups
;
7714 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7715 cpumask_clear(groupmask
);
7717 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7719 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7720 printk("does not load-balance\n");
7722 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7727 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7729 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7730 printk(KERN_ERR
"ERROR: domain->span does not contain "
7733 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7734 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7738 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7742 printk(KERN_ERR
"ERROR: group is NULL\n");
7746 if (!group
->cpu_power
) {
7747 printk(KERN_CONT
"\n");
7748 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7753 if (!cpumask_weight(sched_group_cpus(group
))) {
7754 printk(KERN_CONT
"\n");
7755 printk(KERN_ERR
"ERROR: empty group\n");
7759 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7760 printk(KERN_CONT
"\n");
7761 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7765 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7767 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7769 printk(KERN_CONT
" %s", str
);
7770 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7771 printk(KERN_CONT
" (cpu_power = %d)",
7775 group
= group
->next
;
7776 } while (group
!= sd
->groups
);
7777 printk(KERN_CONT
"\n");
7779 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7780 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7783 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7784 printk(KERN_ERR
"ERROR: parent span is not a superset "
7785 "of domain->span\n");
7789 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7791 cpumask_var_t groupmask
;
7795 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7799 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7801 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7802 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7807 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7814 free_cpumask_var(groupmask
);
7816 #else /* !CONFIG_SCHED_DEBUG */
7817 # define sched_domain_debug(sd, cpu) do { } while (0)
7818 #endif /* CONFIG_SCHED_DEBUG */
7820 static int sd_degenerate(struct sched_domain
*sd
)
7822 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7825 /* Following flags need at least 2 groups */
7826 if (sd
->flags
& (SD_LOAD_BALANCE
|
7827 SD_BALANCE_NEWIDLE
|
7831 SD_SHARE_PKG_RESOURCES
)) {
7832 if (sd
->groups
!= sd
->groups
->next
)
7836 /* Following flags don't use groups */
7837 if (sd
->flags
& (SD_WAKE_AFFINE
))
7844 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7846 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7848 if (sd_degenerate(parent
))
7851 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7854 /* Flags needing groups don't count if only 1 group in parent */
7855 if (parent
->groups
== parent
->groups
->next
) {
7856 pflags
&= ~(SD_LOAD_BALANCE
|
7857 SD_BALANCE_NEWIDLE
|
7861 SD_SHARE_PKG_RESOURCES
);
7862 if (nr_node_ids
== 1)
7863 pflags
&= ~SD_SERIALIZE
;
7865 if (~cflags
& pflags
)
7871 static void free_rootdomain(struct root_domain
*rd
)
7873 cpupri_cleanup(&rd
->cpupri
);
7875 free_cpumask_var(rd
->rto_mask
);
7876 free_cpumask_var(rd
->online
);
7877 free_cpumask_var(rd
->span
);
7881 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7883 struct root_domain
*old_rd
= NULL
;
7884 unsigned long flags
;
7886 spin_lock_irqsave(&rq
->lock
, flags
);
7891 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7894 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7897 * If we dont want to free the old_rt yet then
7898 * set old_rd to NULL to skip the freeing later
7901 if (!atomic_dec_and_test(&old_rd
->refcount
))
7905 atomic_inc(&rd
->refcount
);
7908 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7909 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
7912 spin_unlock_irqrestore(&rq
->lock
, flags
);
7915 free_rootdomain(old_rd
);
7918 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7920 gfp_t gfp
= GFP_KERNEL
;
7922 memset(rd
, 0, sizeof(*rd
));
7927 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7929 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7931 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7934 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7939 free_cpumask_var(rd
->rto_mask
);
7941 free_cpumask_var(rd
->online
);
7943 free_cpumask_var(rd
->span
);
7948 static void init_defrootdomain(void)
7950 init_rootdomain(&def_root_domain
, true);
7952 atomic_set(&def_root_domain
.refcount
, 1);
7955 static struct root_domain
*alloc_rootdomain(void)
7957 struct root_domain
*rd
;
7959 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7963 if (init_rootdomain(rd
, false) != 0) {
7972 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7973 * hold the hotplug lock.
7976 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7978 struct rq
*rq
= cpu_rq(cpu
);
7979 struct sched_domain
*tmp
;
7981 /* Remove the sched domains which do not contribute to scheduling. */
7982 for (tmp
= sd
; tmp
; ) {
7983 struct sched_domain
*parent
= tmp
->parent
;
7987 if (sd_parent_degenerate(tmp
, parent
)) {
7988 tmp
->parent
= parent
->parent
;
7990 parent
->parent
->child
= tmp
;
7995 if (sd
&& sd_degenerate(sd
)) {
8001 sched_domain_debug(sd
, cpu
);
8003 rq_attach_root(rq
, rd
);
8004 rcu_assign_pointer(rq
->sd
, sd
);
8007 /* cpus with isolated domains */
8008 static cpumask_var_t cpu_isolated_map
;
8010 /* Setup the mask of cpus configured for isolated domains */
8011 static int __init
isolated_cpu_setup(char *str
)
8013 cpulist_parse(str
, cpu_isolated_map
);
8017 __setup("isolcpus=", isolated_cpu_setup
);
8020 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8021 * to a function which identifies what group(along with sched group) a CPU
8022 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8023 * (due to the fact that we keep track of groups covered with a struct cpumask).
8025 * init_sched_build_groups will build a circular linked list of the groups
8026 * covered by the given span, and will set each group's ->cpumask correctly,
8027 * and ->cpu_power to 0.
8030 init_sched_build_groups(const struct cpumask
*span
,
8031 const struct cpumask
*cpu_map
,
8032 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8033 struct sched_group
**sg
,
8034 struct cpumask
*tmpmask
),
8035 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8037 struct sched_group
*first
= NULL
, *last
= NULL
;
8040 cpumask_clear(covered
);
8042 for_each_cpu(i
, span
) {
8043 struct sched_group
*sg
;
8044 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8047 if (cpumask_test_cpu(i
, covered
))
8050 cpumask_clear(sched_group_cpus(sg
));
8053 for_each_cpu(j
, span
) {
8054 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8057 cpumask_set_cpu(j
, covered
);
8058 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8069 #define SD_NODES_PER_DOMAIN 16
8074 * find_next_best_node - find the next node to include in a sched_domain
8075 * @node: node whose sched_domain we're building
8076 * @used_nodes: nodes already in the sched_domain
8078 * Find the next node to include in a given scheduling domain. Simply
8079 * finds the closest node not already in the @used_nodes map.
8081 * Should use nodemask_t.
8083 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8085 int i
, n
, val
, min_val
, best_node
= 0;
8089 for (i
= 0; i
< nr_node_ids
; i
++) {
8090 /* Start at @node */
8091 n
= (node
+ i
) % nr_node_ids
;
8093 if (!nr_cpus_node(n
))
8096 /* Skip already used nodes */
8097 if (node_isset(n
, *used_nodes
))
8100 /* Simple min distance search */
8101 val
= node_distance(node
, n
);
8103 if (val
< min_val
) {
8109 node_set(best_node
, *used_nodes
);
8114 * sched_domain_node_span - get a cpumask for a node's sched_domain
8115 * @node: node whose cpumask we're constructing
8116 * @span: resulting cpumask
8118 * Given a node, construct a good cpumask for its sched_domain to span. It
8119 * should be one that prevents unnecessary balancing, but also spreads tasks
8122 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8124 nodemask_t used_nodes
;
8127 cpumask_clear(span
);
8128 nodes_clear(used_nodes
);
8130 cpumask_or(span
, span
, cpumask_of_node(node
));
8131 node_set(node
, used_nodes
);
8133 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8134 int next_node
= find_next_best_node(node
, &used_nodes
);
8136 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8139 #endif /* CONFIG_NUMA */
8141 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8144 * The cpus mask in sched_group and sched_domain hangs off the end.
8146 * ( See the the comments in include/linux/sched.h:struct sched_group
8147 * and struct sched_domain. )
8149 struct static_sched_group
{
8150 struct sched_group sg
;
8151 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8154 struct static_sched_domain
{
8155 struct sched_domain sd
;
8156 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8162 cpumask_var_t domainspan
;
8163 cpumask_var_t covered
;
8164 cpumask_var_t notcovered
;
8166 cpumask_var_t nodemask
;
8167 cpumask_var_t this_sibling_map
;
8168 cpumask_var_t this_core_map
;
8169 cpumask_var_t send_covered
;
8170 cpumask_var_t tmpmask
;
8171 struct sched_group
**sched_group_nodes
;
8172 struct root_domain
*rd
;
8176 sa_sched_groups
= 0,
8181 sa_this_sibling_map
,
8183 sa_sched_group_nodes
,
8193 * SMT sched-domains:
8195 #ifdef CONFIG_SCHED_SMT
8196 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8197 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8200 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8201 struct sched_group
**sg
, struct cpumask
*unused
)
8204 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8207 #endif /* CONFIG_SCHED_SMT */
8210 * multi-core sched-domains:
8212 #ifdef CONFIG_SCHED_MC
8213 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8214 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8215 #endif /* CONFIG_SCHED_MC */
8217 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8219 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8220 struct sched_group
**sg
, struct cpumask
*mask
)
8224 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8225 group
= cpumask_first(mask
);
8227 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8230 #elif defined(CONFIG_SCHED_MC)
8232 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8233 struct sched_group
**sg
, struct cpumask
*unused
)
8236 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8241 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8242 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8245 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8246 struct sched_group
**sg
, struct cpumask
*mask
)
8249 #ifdef CONFIG_SCHED_MC
8250 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8251 group
= cpumask_first(mask
);
8252 #elif defined(CONFIG_SCHED_SMT)
8253 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8254 group
= cpumask_first(mask
);
8259 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8265 * The init_sched_build_groups can't handle what we want to do with node
8266 * groups, so roll our own. Now each node has its own list of groups which
8267 * gets dynamically allocated.
8269 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8270 static struct sched_group
***sched_group_nodes_bycpu
;
8272 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8273 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8275 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8276 struct sched_group
**sg
,
8277 struct cpumask
*nodemask
)
8281 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8282 group
= cpumask_first(nodemask
);
8285 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8289 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8291 struct sched_group
*sg
= group_head
;
8297 for_each_cpu(j
, sched_group_cpus(sg
)) {
8298 struct sched_domain
*sd
;
8300 sd
= &per_cpu(phys_domains
, j
).sd
;
8301 if (j
!= group_first_cpu(sd
->groups
)) {
8303 * Only add "power" once for each
8309 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8312 } while (sg
!= group_head
);
8315 static int build_numa_sched_groups(struct s_data
*d
,
8316 const struct cpumask
*cpu_map
, int num
)
8318 struct sched_domain
*sd
;
8319 struct sched_group
*sg
, *prev
;
8322 cpumask_clear(d
->covered
);
8323 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8324 if (cpumask_empty(d
->nodemask
)) {
8325 d
->sched_group_nodes
[num
] = NULL
;
8329 sched_domain_node_span(num
, d
->domainspan
);
8330 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8332 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8335 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8339 d
->sched_group_nodes
[num
] = sg
;
8341 for_each_cpu(j
, d
->nodemask
) {
8342 sd
= &per_cpu(node_domains
, j
).sd
;
8347 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8349 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8352 for (j
= 0; j
< nr_node_ids
; j
++) {
8353 n
= (num
+ j
) % nr_node_ids
;
8354 cpumask_complement(d
->notcovered
, d
->covered
);
8355 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8356 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8357 if (cpumask_empty(d
->tmpmask
))
8359 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8360 if (cpumask_empty(d
->tmpmask
))
8362 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8366 "Can not alloc domain group for node %d\n", j
);
8370 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8371 sg
->next
= prev
->next
;
8372 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8379 #endif /* CONFIG_NUMA */
8382 /* Free memory allocated for various sched_group structures */
8383 static void free_sched_groups(const struct cpumask
*cpu_map
,
8384 struct cpumask
*nodemask
)
8388 for_each_cpu(cpu
, cpu_map
) {
8389 struct sched_group
**sched_group_nodes
8390 = sched_group_nodes_bycpu
[cpu
];
8392 if (!sched_group_nodes
)
8395 for (i
= 0; i
< nr_node_ids
; i
++) {
8396 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8398 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8399 if (cpumask_empty(nodemask
))
8409 if (oldsg
!= sched_group_nodes
[i
])
8412 kfree(sched_group_nodes
);
8413 sched_group_nodes_bycpu
[cpu
] = NULL
;
8416 #else /* !CONFIG_NUMA */
8417 static void free_sched_groups(const struct cpumask
*cpu_map
,
8418 struct cpumask
*nodemask
)
8421 #endif /* CONFIG_NUMA */
8424 * Initialize sched groups cpu_power.
8426 * cpu_power indicates the capacity of sched group, which is used while
8427 * distributing the load between different sched groups in a sched domain.
8428 * Typically cpu_power for all the groups in a sched domain will be same unless
8429 * there are asymmetries in the topology. If there are asymmetries, group
8430 * having more cpu_power will pickup more load compared to the group having
8433 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8435 struct sched_domain
*child
;
8436 struct sched_group
*group
;
8440 WARN_ON(!sd
|| !sd
->groups
);
8442 if (cpu
!= group_first_cpu(sd
->groups
))
8447 sd
->groups
->cpu_power
= 0;
8450 power
= SCHED_LOAD_SCALE
;
8451 weight
= cpumask_weight(sched_domain_span(sd
));
8453 * SMT siblings share the power of a single core.
8454 * Usually multiple threads get a better yield out of
8455 * that one core than a single thread would have,
8456 * reflect that in sd->smt_gain.
8458 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8459 power
*= sd
->smt_gain
;
8461 power
>>= SCHED_LOAD_SHIFT
;
8463 sd
->groups
->cpu_power
+= power
;
8468 * Add cpu_power of each child group to this groups cpu_power.
8470 group
= child
->groups
;
8472 sd
->groups
->cpu_power
+= group
->cpu_power
;
8473 group
= group
->next
;
8474 } while (group
!= child
->groups
);
8478 * Initializers for schedule domains
8479 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8482 #ifdef CONFIG_SCHED_DEBUG
8483 # define SD_INIT_NAME(sd, type) sd->name = #type
8485 # define SD_INIT_NAME(sd, type) do { } while (0)
8488 #define SD_INIT(sd, type) sd_init_##type(sd)
8490 #define SD_INIT_FUNC(type) \
8491 static noinline void sd_init_##type(struct sched_domain *sd) \
8493 memset(sd, 0, sizeof(*sd)); \
8494 *sd = SD_##type##_INIT; \
8495 sd->level = SD_LV_##type; \
8496 SD_INIT_NAME(sd, type); \
8501 SD_INIT_FUNC(ALLNODES
)
8504 #ifdef CONFIG_SCHED_SMT
8505 SD_INIT_FUNC(SIBLING
)
8507 #ifdef CONFIG_SCHED_MC
8511 static int default_relax_domain_level
= -1;
8513 static int __init
setup_relax_domain_level(char *str
)
8517 val
= simple_strtoul(str
, NULL
, 0);
8518 if (val
< SD_LV_MAX
)
8519 default_relax_domain_level
= val
;
8523 __setup("relax_domain_level=", setup_relax_domain_level
);
8525 static void set_domain_attribute(struct sched_domain
*sd
,
8526 struct sched_domain_attr
*attr
)
8530 if (!attr
|| attr
->relax_domain_level
< 0) {
8531 if (default_relax_domain_level
< 0)
8534 request
= default_relax_domain_level
;
8536 request
= attr
->relax_domain_level
;
8537 if (request
< sd
->level
) {
8538 /* turn off idle balance on this domain */
8539 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8541 /* turn on idle balance on this domain */
8542 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8546 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8547 const struct cpumask
*cpu_map
)
8550 case sa_sched_groups
:
8551 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8552 d
->sched_group_nodes
= NULL
;
8554 free_rootdomain(d
->rd
); /* fall through */
8556 free_cpumask_var(d
->tmpmask
); /* fall through */
8557 case sa_send_covered
:
8558 free_cpumask_var(d
->send_covered
); /* fall through */
8559 case sa_this_core_map
:
8560 free_cpumask_var(d
->this_core_map
); /* fall through */
8561 case sa_this_sibling_map
:
8562 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8564 free_cpumask_var(d
->nodemask
); /* fall through */
8565 case sa_sched_group_nodes
:
8567 kfree(d
->sched_group_nodes
); /* fall through */
8569 free_cpumask_var(d
->notcovered
); /* fall through */
8571 free_cpumask_var(d
->covered
); /* fall through */
8573 free_cpumask_var(d
->domainspan
); /* fall through */
8580 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8581 const struct cpumask
*cpu_map
)
8584 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8586 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8587 return sa_domainspan
;
8588 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8590 /* Allocate the per-node list of sched groups */
8591 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8592 sizeof(struct sched_group
*), GFP_KERNEL
);
8593 if (!d
->sched_group_nodes
) {
8594 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8595 return sa_notcovered
;
8597 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8599 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8600 return sa_sched_group_nodes
;
8601 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8603 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8604 return sa_this_sibling_map
;
8605 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8606 return sa_this_core_map
;
8607 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8608 return sa_send_covered
;
8609 d
->rd
= alloc_rootdomain();
8611 printk(KERN_WARNING
"Cannot alloc root domain\n");
8614 return sa_rootdomain
;
8617 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8618 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8620 struct sched_domain
*sd
= NULL
;
8622 struct sched_domain
*parent
;
8625 if (cpumask_weight(cpu_map
) >
8626 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8627 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8628 SD_INIT(sd
, ALLNODES
);
8629 set_domain_attribute(sd
, attr
);
8630 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8631 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8636 sd
= &per_cpu(node_domains
, i
).sd
;
8638 set_domain_attribute(sd
, attr
);
8639 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8640 sd
->parent
= parent
;
8643 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8648 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8649 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8650 struct sched_domain
*parent
, int i
)
8652 struct sched_domain
*sd
;
8653 sd
= &per_cpu(phys_domains
, i
).sd
;
8655 set_domain_attribute(sd
, attr
);
8656 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8657 sd
->parent
= parent
;
8660 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8664 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8665 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8666 struct sched_domain
*parent
, int i
)
8668 struct sched_domain
*sd
= parent
;
8669 #ifdef CONFIG_SCHED_MC
8670 sd
= &per_cpu(core_domains
, i
).sd
;
8672 set_domain_attribute(sd
, attr
);
8673 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8674 sd
->parent
= parent
;
8676 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8681 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8682 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8683 struct sched_domain
*parent
, int i
)
8685 struct sched_domain
*sd
= parent
;
8686 #ifdef CONFIG_SCHED_SMT
8687 sd
= &per_cpu(cpu_domains
, i
).sd
;
8688 SD_INIT(sd
, SIBLING
);
8689 set_domain_attribute(sd
, attr
);
8690 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8691 sd
->parent
= parent
;
8693 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8698 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8699 const struct cpumask
*cpu_map
, int cpu
)
8702 #ifdef CONFIG_SCHED_SMT
8703 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8704 cpumask_and(d
->this_sibling_map
, cpu_map
,
8705 topology_thread_cpumask(cpu
));
8706 if (cpu
== cpumask_first(d
->this_sibling_map
))
8707 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8709 d
->send_covered
, d
->tmpmask
);
8712 #ifdef CONFIG_SCHED_MC
8713 case SD_LV_MC
: /* set up multi-core groups */
8714 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8715 if (cpu
== cpumask_first(d
->this_core_map
))
8716 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8718 d
->send_covered
, d
->tmpmask
);
8721 case SD_LV_CPU
: /* set up physical groups */
8722 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8723 if (!cpumask_empty(d
->nodemask
))
8724 init_sched_build_groups(d
->nodemask
, cpu_map
,
8726 d
->send_covered
, d
->tmpmask
);
8729 case SD_LV_ALLNODES
:
8730 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8731 d
->send_covered
, d
->tmpmask
);
8740 * Build sched domains for a given set of cpus and attach the sched domains
8741 * to the individual cpus
8743 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8744 struct sched_domain_attr
*attr
)
8746 enum s_alloc alloc_state
= sa_none
;
8748 struct sched_domain
*sd
;
8754 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8755 if (alloc_state
!= sa_rootdomain
)
8757 alloc_state
= sa_sched_groups
;
8760 * Set up domains for cpus specified by the cpu_map.
8762 for_each_cpu(i
, cpu_map
) {
8763 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8766 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8767 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8768 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8769 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8772 for_each_cpu(i
, cpu_map
) {
8773 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8774 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8777 /* Set up physical groups */
8778 for (i
= 0; i
< nr_node_ids
; i
++)
8779 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8782 /* Set up node groups */
8784 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8786 for (i
= 0; i
< nr_node_ids
; i
++)
8787 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8791 /* Calculate CPU power for physical packages and nodes */
8792 #ifdef CONFIG_SCHED_SMT
8793 for_each_cpu(i
, cpu_map
) {
8794 sd
= &per_cpu(cpu_domains
, i
).sd
;
8795 init_sched_groups_power(i
, sd
);
8798 #ifdef CONFIG_SCHED_MC
8799 for_each_cpu(i
, cpu_map
) {
8800 sd
= &per_cpu(core_domains
, i
).sd
;
8801 init_sched_groups_power(i
, sd
);
8805 for_each_cpu(i
, cpu_map
) {
8806 sd
= &per_cpu(phys_domains
, i
).sd
;
8807 init_sched_groups_power(i
, sd
);
8811 for (i
= 0; i
< nr_node_ids
; i
++)
8812 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8814 if (d
.sd_allnodes
) {
8815 struct sched_group
*sg
;
8817 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8819 init_numa_sched_groups_power(sg
);
8823 /* Attach the domains */
8824 for_each_cpu(i
, cpu_map
) {
8825 #ifdef CONFIG_SCHED_SMT
8826 sd
= &per_cpu(cpu_domains
, i
).sd
;
8827 #elif defined(CONFIG_SCHED_MC)
8828 sd
= &per_cpu(core_domains
, i
).sd
;
8830 sd
= &per_cpu(phys_domains
, i
).sd
;
8832 cpu_attach_domain(sd
, d
.rd
, i
);
8835 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8836 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8840 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8844 static int build_sched_domains(const struct cpumask
*cpu_map
)
8846 return __build_sched_domains(cpu_map
, NULL
);
8849 static struct cpumask
*doms_cur
; /* current sched domains */
8850 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8851 static struct sched_domain_attr
*dattr_cur
;
8852 /* attribues of custom domains in 'doms_cur' */
8855 * Special case: If a kmalloc of a doms_cur partition (array of
8856 * cpumask) fails, then fallback to a single sched domain,
8857 * as determined by the single cpumask fallback_doms.
8859 static cpumask_var_t fallback_doms
;
8862 * arch_update_cpu_topology lets virtualized architectures update the
8863 * cpu core maps. It is supposed to return 1 if the topology changed
8864 * or 0 if it stayed the same.
8866 int __attribute__((weak
)) arch_update_cpu_topology(void)
8872 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8873 * For now this just excludes isolated cpus, but could be used to
8874 * exclude other special cases in the future.
8876 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8880 arch_update_cpu_topology();
8882 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8884 doms_cur
= fallback_doms
;
8885 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8887 err
= build_sched_domains(doms_cur
);
8888 register_sched_domain_sysctl();
8893 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8894 struct cpumask
*tmpmask
)
8896 free_sched_groups(cpu_map
, tmpmask
);
8900 * Detach sched domains from a group of cpus specified in cpu_map
8901 * These cpus will now be attached to the NULL domain
8903 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8905 /* Save because hotplug lock held. */
8906 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8909 for_each_cpu(i
, cpu_map
)
8910 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8911 synchronize_sched();
8912 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8915 /* handle null as "default" */
8916 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8917 struct sched_domain_attr
*new, int idx_new
)
8919 struct sched_domain_attr tmp
;
8926 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8927 new ? (new + idx_new
) : &tmp
,
8928 sizeof(struct sched_domain_attr
));
8932 * Partition sched domains as specified by the 'ndoms_new'
8933 * cpumasks in the array doms_new[] of cpumasks. This compares
8934 * doms_new[] to the current sched domain partitioning, doms_cur[].
8935 * It destroys each deleted domain and builds each new domain.
8937 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8938 * The masks don't intersect (don't overlap.) We should setup one
8939 * sched domain for each mask. CPUs not in any of the cpumasks will
8940 * not be load balanced. If the same cpumask appears both in the
8941 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8944 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8945 * ownership of it and will kfree it when done with it. If the caller
8946 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8947 * ndoms_new == 1, and partition_sched_domains() will fallback to
8948 * the single partition 'fallback_doms', it also forces the domains
8951 * If doms_new == NULL it will be replaced with cpu_online_mask.
8952 * ndoms_new == 0 is a special case for destroying existing domains,
8953 * and it will not create the default domain.
8955 * Call with hotplug lock held
8957 /* FIXME: Change to struct cpumask *doms_new[] */
8958 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8959 struct sched_domain_attr
*dattr_new
)
8964 mutex_lock(&sched_domains_mutex
);
8966 /* always unregister in case we don't destroy any domains */
8967 unregister_sched_domain_sysctl();
8969 /* Let architecture update cpu core mappings. */
8970 new_topology
= arch_update_cpu_topology();
8972 n
= doms_new
? ndoms_new
: 0;
8974 /* Destroy deleted domains */
8975 for (i
= 0; i
< ndoms_cur
; i
++) {
8976 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8977 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8978 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8981 /* no match - a current sched domain not in new doms_new[] */
8982 detach_destroy_domains(doms_cur
+ i
);
8987 if (doms_new
== NULL
) {
8989 doms_new
= fallback_doms
;
8990 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8991 WARN_ON_ONCE(dattr_new
);
8994 /* Build new domains */
8995 for (i
= 0; i
< ndoms_new
; i
++) {
8996 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8997 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8998 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9001 /* no match - add a new doms_new */
9002 __build_sched_domains(doms_new
+ i
,
9003 dattr_new
? dattr_new
+ i
: NULL
);
9008 /* Remember the new sched domains */
9009 if (doms_cur
!= fallback_doms
)
9011 kfree(dattr_cur
); /* kfree(NULL) is safe */
9012 doms_cur
= doms_new
;
9013 dattr_cur
= dattr_new
;
9014 ndoms_cur
= ndoms_new
;
9016 register_sched_domain_sysctl();
9018 mutex_unlock(&sched_domains_mutex
);
9021 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9022 static void arch_reinit_sched_domains(void)
9026 /* Destroy domains first to force the rebuild */
9027 partition_sched_domains(0, NULL
, NULL
);
9029 rebuild_sched_domains();
9033 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9035 unsigned int level
= 0;
9037 if (sscanf(buf
, "%u", &level
) != 1)
9041 * level is always be positive so don't check for
9042 * level < POWERSAVINGS_BALANCE_NONE which is 0
9043 * What happens on 0 or 1 byte write,
9044 * need to check for count as well?
9047 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9051 sched_smt_power_savings
= level
;
9053 sched_mc_power_savings
= level
;
9055 arch_reinit_sched_domains();
9060 #ifdef CONFIG_SCHED_MC
9061 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9064 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9066 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9067 const char *buf
, size_t count
)
9069 return sched_power_savings_store(buf
, count
, 0);
9071 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9072 sched_mc_power_savings_show
,
9073 sched_mc_power_savings_store
);
9076 #ifdef CONFIG_SCHED_SMT
9077 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9080 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9082 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9083 const char *buf
, size_t count
)
9085 return sched_power_savings_store(buf
, count
, 1);
9087 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9088 sched_smt_power_savings_show
,
9089 sched_smt_power_savings_store
);
9092 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9096 #ifdef CONFIG_SCHED_SMT
9098 err
= sysfs_create_file(&cls
->kset
.kobj
,
9099 &attr_sched_smt_power_savings
.attr
);
9101 #ifdef CONFIG_SCHED_MC
9102 if (!err
&& mc_capable())
9103 err
= sysfs_create_file(&cls
->kset
.kobj
,
9104 &attr_sched_mc_power_savings
.attr
);
9108 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9110 #ifndef CONFIG_CPUSETS
9112 * Add online and remove offline CPUs from the scheduler domains.
9113 * When cpusets are enabled they take over this function.
9115 static int update_sched_domains(struct notifier_block
*nfb
,
9116 unsigned long action
, void *hcpu
)
9120 case CPU_ONLINE_FROZEN
:
9122 case CPU_DEAD_FROZEN
:
9123 partition_sched_domains(1, NULL
, NULL
);
9132 static int update_runtime(struct notifier_block
*nfb
,
9133 unsigned long action
, void *hcpu
)
9135 int cpu
= (int)(long)hcpu
;
9138 case CPU_DOWN_PREPARE
:
9139 case CPU_DOWN_PREPARE_FROZEN
:
9140 disable_runtime(cpu_rq(cpu
));
9143 case CPU_DOWN_FAILED
:
9144 case CPU_DOWN_FAILED_FROZEN
:
9146 case CPU_ONLINE_FROZEN
:
9147 enable_runtime(cpu_rq(cpu
));
9155 void __init
sched_init_smp(void)
9157 cpumask_var_t non_isolated_cpus
;
9159 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9160 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9162 #if defined(CONFIG_NUMA)
9163 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9165 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9168 mutex_lock(&sched_domains_mutex
);
9169 arch_init_sched_domains(cpu_online_mask
);
9170 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9171 if (cpumask_empty(non_isolated_cpus
))
9172 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9173 mutex_unlock(&sched_domains_mutex
);
9176 #ifndef CONFIG_CPUSETS
9177 /* XXX: Theoretical race here - CPU may be hotplugged now */
9178 hotcpu_notifier(update_sched_domains
, 0);
9181 /* RT runtime code needs to handle some hotplug events */
9182 hotcpu_notifier(update_runtime
, 0);
9186 /* Move init over to a non-isolated CPU */
9187 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9189 sched_init_granularity();
9190 free_cpumask_var(non_isolated_cpus
);
9192 init_sched_rt_class();
9195 void __init
sched_init_smp(void)
9197 sched_init_granularity();
9199 #endif /* CONFIG_SMP */
9201 const_debug
unsigned int sysctl_timer_migration
= 1;
9203 int in_sched_functions(unsigned long addr
)
9205 return in_lock_functions(addr
) ||
9206 (addr
>= (unsigned long)__sched_text_start
9207 && addr
< (unsigned long)__sched_text_end
);
9210 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9212 cfs_rq
->tasks_timeline
= RB_ROOT
;
9213 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9214 #ifdef CONFIG_FAIR_GROUP_SCHED
9217 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9220 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9222 struct rt_prio_array
*array
;
9225 array
= &rt_rq
->active
;
9226 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9227 INIT_LIST_HEAD(array
->queue
+ i
);
9228 __clear_bit(i
, array
->bitmap
);
9230 /* delimiter for bitsearch: */
9231 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9233 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9234 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9236 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9240 rt_rq
->rt_nr_migratory
= 0;
9241 rt_rq
->overloaded
= 0;
9242 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9246 rt_rq
->rt_throttled
= 0;
9247 rt_rq
->rt_runtime
= 0;
9248 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9250 #ifdef CONFIG_RT_GROUP_SCHED
9251 rt_rq
->rt_nr_boosted
= 0;
9256 #ifdef CONFIG_FAIR_GROUP_SCHED
9257 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9258 struct sched_entity
*se
, int cpu
, int add
,
9259 struct sched_entity
*parent
)
9261 struct rq
*rq
= cpu_rq(cpu
);
9262 tg
->cfs_rq
[cpu
] = cfs_rq
;
9263 init_cfs_rq(cfs_rq
, rq
);
9266 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9269 /* se could be NULL for init_task_group */
9274 se
->cfs_rq
= &rq
->cfs
;
9276 se
->cfs_rq
= parent
->my_q
;
9279 se
->load
.weight
= tg
->shares
;
9280 se
->load
.inv_weight
= 0;
9281 se
->parent
= parent
;
9285 #ifdef CONFIG_RT_GROUP_SCHED
9286 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9287 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9288 struct sched_rt_entity
*parent
)
9290 struct rq
*rq
= cpu_rq(cpu
);
9292 tg
->rt_rq
[cpu
] = rt_rq
;
9293 init_rt_rq(rt_rq
, rq
);
9295 rt_rq
->rt_se
= rt_se
;
9296 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9298 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9300 tg
->rt_se
[cpu
] = rt_se
;
9305 rt_se
->rt_rq
= &rq
->rt
;
9307 rt_se
->rt_rq
= parent
->my_q
;
9309 rt_se
->my_q
= rt_rq
;
9310 rt_se
->parent
= parent
;
9311 INIT_LIST_HEAD(&rt_se
->run_list
);
9315 void __init
sched_init(void)
9318 unsigned long alloc_size
= 0, ptr
;
9320 #ifdef CONFIG_FAIR_GROUP_SCHED
9321 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9323 #ifdef CONFIG_RT_GROUP_SCHED
9324 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9326 #ifdef CONFIG_USER_SCHED
9329 #ifdef CONFIG_CPUMASK_OFFSTACK
9330 alloc_size
+= num_possible_cpus() * cpumask_size();
9333 * As sched_init() is called before page_alloc is setup,
9334 * we use alloc_bootmem().
9337 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9339 #ifdef CONFIG_FAIR_GROUP_SCHED
9340 init_task_group
.se
= (struct sched_entity
**)ptr
;
9341 ptr
+= nr_cpu_ids
* sizeof(void **);
9343 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9344 ptr
+= nr_cpu_ids
* sizeof(void **);
9346 #ifdef CONFIG_USER_SCHED
9347 root_task_group
.se
= (struct sched_entity
**)ptr
;
9348 ptr
+= nr_cpu_ids
* sizeof(void **);
9350 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9351 ptr
+= nr_cpu_ids
* sizeof(void **);
9352 #endif /* CONFIG_USER_SCHED */
9353 #endif /* CONFIG_FAIR_GROUP_SCHED */
9354 #ifdef CONFIG_RT_GROUP_SCHED
9355 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9356 ptr
+= nr_cpu_ids
* sizeof(void **);
9358 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9359 ptr
+= nr_cpu_ids
* sizeof(void **);
9361 #ifdef CONFIG_USER_SCHED
9362 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9363 ptr
+= nr_cpu_ids
* sizeof(void **);
9365 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9366 ptr
+= nr_cpu_ids
* sizeof(void **);
9367 #endif /* CONFIG_USER_SCHED */
9368 #endif /* CONFIG_RT_GROUP_SCHED */
9369 #ifdef CONFIG_CPUMASK_OFFSTACK
9370 for_each_possible_cpu(i
) {
9371 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9372 ptr
+= cpumask_size();
9374 #endif /* CONFIG_CPUMASK_OFFSTACK */
9378 init_defrootdomain();
9381 init_rt_bandwidth(&def_rt_bandwidth
,
9382 global_rt_period(), global_rt_runtime());
9384 #ifdef CONFIG_RT_GROUP_SCHED
9385 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9386 global_rt_period(), global_rt_runtime());
9387 #ifdef CONFIG_USER_SCHED
9388 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9389 global_rt_period(), RUNTIME_INF
);
9390 #endif /* CONFIG_USER_SCHED */
9391 #endif /* CONFIG_RT_GROUP_SCHED */
9393 #ifdef CONFIG_GROUP_SCHED
9394 list_add(&init_task_group
.list
, &task_groups
);
9395 INIT_LIST_HEAD(&init_task_group
.children
);
9397 #ifdef CONFIG_USER_SCHED
9398 INIT_LIST_HEAD(&root_task_group
.children
);
9399 init_task_group
.parent
= &root_task_group
;
9400 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9401 #endif /* CONFIG_USER_SCHED */
9402 #endif /* CONFIG_GROUP_SCHED */
9404 for_each_possible_cpu(i
) {
9408 spin_lock_init(&rq
->lock
);
9410 rq
->calc_load_active
= 0;
9411 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9412 init_cfs_rq(&rq
->cfs
, rq
);
9413 init_rt_rq(&rq
->rt
, rq
);
9414 #ifdef CONFIG_FAIR_GROUP_SCHED
9415 init_task_group
.shares
= init_task_group_load
;
9416 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9417 #ifdef CONFIG_CGROUP_SCHED
9419 * How much cpu bandwidth does init_task_group get?
9421 * In case of task-groups formed thr' the cgroup filesystem, it
9422 * gets 100% of the cpu resources in the system. This overall
9423 * system cpu resource is divided among the tasks of
9424 * init_task_group and its child task-groups in a fair manner,
9425 * based on each entity's (task or task-group's) weight
9426 * (se->load.weight).
9428 * In other words, if init_task_group has 10 tasks of weight
9429 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9430 * then A0's share of the cpu resource is:
9432 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9434 * We achieve this by letting init_task_group's tasks sit
9435 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9437 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9438 #elif defined CONFIG_USER_SCHED
9439 root_task_group
.shares
= NICE_0_LOAD
;
9440 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9442 * In case of task-groups formed thr' the user id of tasks,
9443 * init_task_group represents tasks belonging to root user.
9444 * Hence it forms a sibling of all subsequent groups formed.
9445 * In this case, init_task_group gets only a fraction of overall
9446 * system cpu resource, based on the weight assigned to root
9447 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9448 * by letting tasks of init_task_group sit in a separate cfs_rq
9449 * (init_tg_cfs_rq) and having one entity represent this group of
9450 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9452 init_tg_cfs_entry(&init_task_group
,
9453 &per_cpu(init_tg_cfs_rq
, i
),
9454 &per_cpu(init_sched_entity
, i
), i
, 1,
9455 root_task_group
.se
[i
]);
9458 #endif /* CONFIG_FAIR_GROUP_SCHED */
9460 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9461 #ifdef CONFIG_RT_GROUP_SCHED
9462 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9463 #ifdef CONFIG_CGROUP_SCHED
9464 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9465 #elif defined CONFIG_USER_SCHED
9466 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9467 init_tg_rt_entry(&init_task_group
,
9468 &per_cpu(init_rt_rq
, i
),
9469 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9470 root_task_group
.rt_se
[i
]);
9474 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9475 rq
->cpu_load
[j
] = 0;
9479 rq
->post_schedule
= 0;
9480 rq
->active_balance
= 0;
9481 rq
->next_balance
= jiffies
;
9485 rq
->migration_thread
= NULL
;
9486 INIT_LIST_HEAD(&rq
->migration_queue
);
9487 rq_attach_root(rq
, &def_root_domain
);
9490 atomic_set(&rq
->nr_iowait
, 0);
9493 set_load_weight(&init_task
);
9495 #ifdef CONFIG_PREEMPT_NOTIFIERS
9496 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9500 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9503 #ifdef CONFIG_RT_MUTEXES
9504 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9508 * The boot idle thread does lazy MMU switching as well:
9510 atomic_inc(&init_mm
.mm_count
);
9511 enter_lazy_tlb(&init_mm
, current
);
9514 * Make us the idle thread. Technically, schedule() should not be
9515 * called from this thread, however somewhere below it might be,
9516 * but because we are the idle thread, we just pick up running again
9517 * when this runqueue becomes "idle".
9519 init_idle(current
, smp_processor_id());
9521 calc_load_update
= jiffies
+ LOAD_FREQ
;
9524 * During early bootup we pretend to be a normal task:
9526 current
->sched_class
= &fair_sched_class
;
9528 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9529 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9532 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9533 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9535 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9540 scheduler_running
= 1;
9543 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9544 static inline int preempt_count_equals(int preempt_offset
)
9546 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9548 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9551 void __might_sleep(char *file
, int line
, int preempt_offset
)
9554 static unsigned long prev_jiffy
; /* ratelimiting */
9556 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9557 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9559 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9561 prev_jiffy
= jiffies
;
9564 "BUG: sleeping function called from invalid context at %s:%d\n",
9567 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9568 in_atomic(), irqs_disabled(),
9569 current
->pid
, current
->comm
);
9571 debug_show_held_locks(current
);
9572 if (irqs_disabled())
9573 print_irqtrace_events(current
);
9577 EXPORT_SYMBOL(__might_sleep
);
9580 #ifdef CONFIG_MAGIC_SYSRQ
9581 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9585 update_rq_clock(rq
);
9586 on_rq
= p
->se
.on_rq
;
9588 deactivate_task(rq
, p
, 0);
9589 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9591 activate_task(rq
, p
, 0);
9592 resched_task(rq
->curr
);
9596 void normalize_rt_tasks(void)
9598 struct task_struct
*g
, *p
;
9599 unsigned long flags
;
9602 read_lock_irqsave(&tasklist_lock
, flags
);
9603 do_each_thread(g
, p
) {
9605 * Only normalize user tasks:
9610 p
->se
.exec_start
= 0;
9611 #ifdef CONFIG_SCHEDSTATS
9612 p
->se
.wait_start
= 0;
9613 p
->se
.sleep_start
= 0;
9614 p
->se
.block_start
= 0;
9619 * Renice negative nice level userspace
9622 if (TASK_NICE(p
) < 0 && p
->mm
)
9623 set_user_nice(p
, 0);
9627 spin_lock(&p
->pi_lock
);
9628 rq
= __task_rq_lock(p
);
9630 normalize_task(rq
, p
);
9632 __task_rq_unlock(rq
);
9633 spin_unlock(&p
->pi_lock
);
9634 } while_each_thread(g
, p
);
9636 read_unlock_irqrestore(&tasklist_lock
, flags
);
9639 #endif /* CONFIG_MAGIC_SYSRQ */
9643 * These functions are only useful for the IA64 MCA handling.
9645 * They can only be called when the whole system has been
9646 * stopped - every CPU needs to be quiescent, and no scheduling
9647 * activity can take place. Using them for anything else would
9648 * be a serious bug, and as a result, they aren't even visible
9649 * under any other configuration.
9653 * curr_task - return the current task for a given cpu.
9654 * @cpu: the processor in question.
9656 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9658 struct task_struct
*curr_task(int cpu
)
9660 return cpu_curr(cpu
);
9664 * set_curr_task - set the current task for a given cpu.
9665 * @cpu: the processor in question.
9666 * @p: the task pointer to set.
9668 * Description: This function must only be used when non-maskable interrupts
9669 * are serviced on a separate stack. It allows the architecture to switch the
9670 * notion of the current task on a cpu in a non-blocking manner. This function
9671 * must be called with all CPU's synchronized, and interrupts disabled, the
9672 * and caller must save the original value of the current task (see
9673 * curr_task() above) and restore that value before reenabling interrupts and
9674 * re-starting the system.
9676 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9678 void set_curr_task(int cpu
, struct task_struct
*p
)
9685 #ifdef CONFIG_FAIR_GROUP_SCHED
9686 static void free_fair_sched_group(struct task_group
*tg
)
9690 for_each_possible_cpu(i
) {
9692 kfree(tg
->cfs_rq
[i
]);
9702 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9704 struct cfs_rq
*cfs_rq
;
9705 struct sched_entity
*se
;
9709 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9712 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9716 tg
->shares
= NICE_0_LOAD
;
9718 for_each_possible_cpu(i
) {
9721 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9722 GFP_KERNEL
, cpu_to_node(i
));
9726 se
= kzalloc_node(sizeof(struct sched_entity
),
9727 GFP_KERNEL
, cpu_to_node(i
));
9731 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9740 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9742 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9743 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9746 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9748 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9750 #else /* !CONFG_FAIR_GROUP_SCHED */
9751 static inline void free_fair_sched_group(struct task_group
*tg
)
9756 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9761 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9765 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9768 #endif /* CONFIG_FAIR_GROUP_SCHED */
9770 #ifdef CONFIG_RT_GROUP_SCHED
9771 static void free_rt_sched_group(struct task_group
*tg
)
9775 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9777 for_each_possible_cpu(i
) {
9779 kfree(tg
->rt_rq
[i
]);
9781 kfree(tg
->rt_se
[i
]);
9789 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9791 struct rt_rq
*rt_rq
;
9792 struct sched_rt_entity
*rt_se
;
9796 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9799 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9803 init_rt_bandwidth(&tg
->rt_bandwidth
,
9804 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9806 for_each_possible_cpu(i
) {
9809 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9810 GFP_KERNEL
, cpu_to_node(i
));
9814 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9815 GFP_KERNEL
, cpu_to_node(i
));
9819 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9828 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9830 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9831 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9834 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9836 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9838 #else /* !CONFIG_RT_GROUP_SCHED */
9839 static inline void free_rt_sched_group(struct task_group
*tg
)
9844 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9849 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9853 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9856 #endif /* CONFIG_RT_GROUP_SCHED */
9858 #ifdef CONFIG_GROUP_SCHED
9859 static void free_sched_group(struct task_group
*tg
)
9861 free_fair_sched_group(tg
);
9862 free_rt_sched_group(tg
);
9866 /* allocate runqueue etc for a new task group */
9867 struct task_group
*sched_create_group(struct task_group
*parent
)
9869 struct task_group
*tg
;
9870 unsigned long flags
;
9873 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9875 return ERR_PTR(-ENOMEM
);
9877 if (!alloc_fair_sched_group(tg
, parent
))
9880 if (!alloc_rt_sched_group(tg
, parent
))
9883 spin_lock_irqsave(&task_group_lock
, flags
);
9884 for_each_possible_cpu(i
) {
9885 register_fair_sched_group(tg
, i
);
9886 register_rt_sched_group(tg
, i
);
9888 list_add_rcu(&tg
->list
, &task_groups
);
9890 WARN_ON(!parent
); /* root should already exist */
9892 tg
->parent
= parent
;
9893 INIT_LIST_HEAD(&tg
->children
);
9894 list_add_rcu(&tg
->siblings
, &parent
->children
);
9895 spin_unlock_irqrestore(&task_group_lock
, flags
);
9900 free_sched_group(tg
);
9901 return ERR_PTR(-ENOMEM
);
9904 /* rcu callback to free various structures associated with a task group */
9905 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9907 /* now it should be safe to free those cfs_rqs */
9908 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9911 /* Destroy runqueue etc associated with a task group */
9912 void sched_destroy_group(struct task_group
*tg
)
9914 unsigned long flags
;
9917 spin_lock_irqsave(&task_group_lock
, flags
);
9918 for_each_possible_cpu(i
) {
9919 unregister_fair_sched_group(tg
, i
);
9920 unregister_rt_sched_group(tg
, i
);
9922 list_del_rcu(&tg
->list
);
9923 list_del_rcu(&tg
->siblings
);
9924 spin_unlock_irqrestore(&task_group_lock
, flags
);
9926 /* wait for possible concurrent references to cfs_rqs complete */
9927 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9930 /* change task's runqueue when it moves between groups.
9931 * The caller of this function should have put the task in its new group
9932 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9933 * reflect its new group.
9935 void sched_move_task(struct task_struct
*tsk
)
9938 unsigned long flags
;
9941 rq
= task_rq_lock(tsk
, &flags
);
9943 update_rq_clock(rq
);
9945 running
= task_current(rq
, tsk
);
9946 on_rq
= tsk
->se
.on_rq
;
9949 dequeue_task(rq
, tsk
, 0);
9950 if (unlikely(running
))
9951 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9953 set_task_rq(tsk
, task_cpu(tsk
));
9955 #ifdef CONFIG_FAIR_GROUP_SCHED
9956 if (tsk
->sched_class
->moved_group
)
9957 tsk
->sched_class
->moved_group(tsk
);
9960 if (unlikely(running
))
9961 tsk
->sched_class
->set_curr_task(rq
);
9963 enqueue_task(rq
, tsk
, 0);
9965 task_rq_unlock(rq
, &flags
);
9967 #endif /* CONFIG_GROUP_SCHED */
9969 #ifdef CONFIG_FAIR_GROUP_SCHED
9970 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9972 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9977 dequeue_entity(cfs_rq
, se
, 0);
9979 se
->load
.weight
= shares
;
9980 se
->load
.inv_weight
= 0;
9983 enqueue_entity(cfs_rq
, se
, 0);
9986 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9988 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9989 struct rq
*rq
= cfs_rq
->rq
;
9990 unsigned long flags
;
9992 spin_lock_irqsave(&rq
->lock
, flags
);
9993 __set_se_shares(se
, shares
);
9994 spin_unlock_irqrestore(&rq
->lock
, flags
);
9997 static DEFINE_MUTEX(shares_mutex
);
9999 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10002 unsigned long flags
;
10005 * We can't change the weight of the root cgroup.
10010 if (shares
< MIN_SHARES
)
10011 shares
= MIN_SHARES
;
10012 else if (shares
> MAX_SHARES
)
10013 shares
= MAX_SHARES
;
10015 mutex_lock(&shares_mutex
);
10016 if (tg
->shares
== shares
)
10019 spin_lock_irqsave(&task_group_lock
, flags
);
10020 for_each_possible_cpu(i
)
10021 unregister_fair_sched_group(tg
, i
);
10022 list_del_rcu(&tg
->siblings
);
10023 spin_unlock_irqrestore(&task_group_lock
, flags
);
10025 /* wait for any ongoing reference to this group to finish */
10026 synchronize_sched();
10029 * Now we are free to modify the group's share on each cpu
10030 * w/o tripping rebalance_share or load_balance_fair.
10032 tg
->shares
= shares
;
10033 for_each_possible_cpu(i
) {
10035 * force a rebalance
10037 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10038 set_se_shares(tg
->se
[i
], shares
);
10042 * Enable load balance activity on this group, by inserting it back on
10043 * each cpu's rq->leaf_cfs_rq_list.
10045 spin_lock_irqsave(&task_group_lock
, flags
);
10046 for_each_possible_cpu(i
)
10047 register_fair_sched_group(tg
, i
);
10048 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10049 spin_unlock_irqrestore(&task_group_lock
, flags
);
10051 mutex_unlock(&shares_mutex
);
10055 unsigned long sched_group_shares(struct task_group
*tg
)
10061 #ifdef CONFIG_RT_GROUP_SCHED
10063 * Ensure that the real time constraints are schedulable.
10065 static DEFINE_MUTEX(rt_constraints_mutex
);
10067 static unsigned long to_ratio(u64 period
, u64 runtime
)
10069 if (runtime
== RUNTIME_INF
)
10072 return div64_u64(runtime
<< 20, period
);
10075 /* Must be called with tasklist_lock held */
10076 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10078 struct task_struct
*g
, *p
;
10080 do_each_thread(g
, p
) {
10081 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10083 } while_each_thread(g
, p
);
10088 struct rt_schedulable_data
{
10089 struct task_group
*tg
;
10094 static int tg_schedulable(struct task_group
*tg
, void *data
)
10096 struct rt_schedulable_data
*d
= data
;
10097 struct task_group
*child
;
10098 unsigned long total
, sum
= 0;
10099 u64 period
, runtime
;
10101 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10102 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10105 period
= d
->rt_period
;
10106 runtime
= d
->rt_runtime
;
10109 #ifdef CONFIG_USER_SCHED
10110 if (tg
== &root_task_group
) {
10111 period
= global_rt_period();
10112 runtime
= global_rt_runtime();
10117 * Cannot have more runtime than the period.
10119 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10123 * Ensure we don't starve existing RT tasks.
10125 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10128 total
= to_ratio(period
, runtime
);
10131 * Nobody can have more than the global setting allows.
10133 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10137 * The sum of our children's runtime should not exceed our own.
10139 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10140 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10141 runtime
= child
->rt_bandwidth
.rt_runtime
;
10143 if (child
== d
->tg
) {
10144 period
= d
->rt_period
;
10145 runtime
= d
->rt_runtime
;
10148 sum
+= to_ratio(period
, runtime
);
10157 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10159 struct rt_schedulable_data data
= {
10161 .rt_period
= period
,
10162 .rt_runtime
= runtime
,
10165 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10168 static int tg_set_bandwidth(struct task_group
*tg
,
10169 u64 rt_period
, u64 rt_runtime
)
10173 mutex_lock(&rt_constraints_mutex
);
10174 read_lock(&tasklist_lock
);
10175 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10179 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10180 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10181 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10183 for_each_possible_cpu(i
) {
10184 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10186 spin_lock(&rt_rq
->rt_runtime_lock
);
10187 rt_rq
->rt_runtime
= rt_runtime
;
10188 spin_unlock(&rt_rq
->rt_runtime_lock
);
10190 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10192 read_unlock(&tasklist_lock
);
10193 mutex_unlock(&rt_constraints_mutex
);
10198 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10200 u64 rt_runtime
, rt_period
;
10202 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10203 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10204 if (rt_runtime_us
< 0)
10205 rt_runtime
= RUNTIME_INF
;
10207 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10210 long sched_group_rt_runtime(struct task_group
*tg
)
10214 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10217 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10218 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10219 return rt_runtime_us
;
10222 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10224 u64 rt_runtime
, rt_period
;
10226 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10227 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10229 if (rt_period
== 0)
10232 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10235 long sched_group_rt_period(struct task_group
*tg
)
10239 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10240 do_div(rt_period_us
, NSEC_PER_USEC
);
10241 return rt_period_us
;
10244 static int sched_rt_global_constraints(void)
10246 u64 runtime
, period
;
10249 if (sysctl_sched_rt_period
<= 0)
10252 runtime
= global_rt_runtime();
10253 period
= global_rt_period();
10256 * Sanity check on the sysctl variables.
10258 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10261 mutex_lock(&rt_constraints_mutex
);
10262 read_lock(&tasklist_lock
);
10263 ret
= __rt_schedulable(NULL
, 0, 0);
10264 read_unlock(&tasklist_lock
);
10265 mutex_unlock(&rt_constraints_mutex
);
10270 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10272 /* Don't accept realtime tasks when there is no way for them to run */
10273 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10279 #else /* !CONFIG_RT_GROUP_SCHED */
10280 static int sched_rt_global_constraints(void)
10282 unsigned long flags
;
10285 if (sysctl_sched_rt_period
<= 0)
10289 * There's always some RT tasks in the root group
10290 * -- migration, kstopmachine etc..
10292 if (sysctl_sched_rt_runtime
== 0)
10295 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10296 for_each_possible_cpu(i
) {
10297 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10299 spin_lock(&rt_rq
->rt_runtime_lock
);
10300 rt_rq
->rt_runtime
= global_rt_runtime();
10301 spin_unlock(&rt_rq
->rt_runtime_lock
);
10303 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10307 #endif /* CONFIG_RT_GROUP_SCHED */
10309 int sched_rt_handler(struct ctl_table
*table
, int write
,
10310 void __user
*buffer
, size_t *lenp
,
10314 int old_period
, old_runtime
;
10315 static DEFINE_MUTEX(mutex
);
10317 mutex_lock(&mutex
);
10318 old_period
= sysctl_sched_rt_period
;
10319 old_runtime
= sysctl_sched_rt_runtime
;
10321 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10323 if (!ret
&& write
) {
10324 ret
= sched_rt_global_constraints();
10326 sysctl_sched_rt_period
= old_period
;
10327 sysctl_sched_rt_runtime
= old_runtime
;
10329 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10330 def_rt_bandwidth
.rt_period
=
10331 ns_to_ktime(global_rt_period());
10334 mutex_unlock(&mutex
);
10339 #ifdef CONFIG_CGROUP_SCHED
10341 /* return corresponding task_group object of a cgroup */
10342 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10344 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10345 struct task_group
, css
);
10348 static struct cgroup_subsys_state
*
10349 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10351 struct task_group
*tg
, *parent
;
10353 if (!cgrp
->parent
) {
10354 /* This is early initialization for the top cgroup */
10355 return &init_task_group
.css
;
10358 parent
= cgroup_tg(cgrp
->parent
);
10359 tg
= sched_create_group(parent
);
10361 return ERR_PTR(-ENOMEM
);
10367 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10369 struct task_group
*tg
= cgroup_tg(cgrp
);
10371 sched_destroy_group(tg
);
10375 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10377 #ifdef CONFIG_RT_GROUP_SCHED
10378 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10381 /* We don't support RT-tasks being in separate groups */
10382 if (tsk
->sched_class
!= &fair_sched_class
)
10389 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10390 struct task_struct
*tsk
, bool threadgroup
)
10392 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10396 struct task_struct
*c
;
10398 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10399 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10411 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10412 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10415 sched_move_task(tsk
);
10417 struct task_struct
*c
;
10419 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10420 sched_move_task(c
);
10426 #ifdef CONFIG_FAIR_GROUP_SCHED
10427 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10430 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10433 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10435 struct task_group
*tg
= cgroup_tg(cgrp
);
10437 return (u64
) tg
->shares
;
10439 #endif /* CONFIG_FAIR_GROUP_SCHED */
10441 #ifdef CONFIG_RT_GROUP_SCHED
10442 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10445 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10448 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10450 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10453 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10456 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10459 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10461 return sched_group_rt_period(cgroup_tg(cgrp
));
10463 #endif /* CONFIG_RT_GROUP_SCHED */
10465 static struct cftype cpu_files
[] = {
10466 #ifdef CONFIG_FAIR_GROUP_SCHED
10469 .read_u64
= cpu_shares_read_u64
,
10470 .write_u64
= cpu_shares_write_u64
,
10473 #ifdef CONFIG_RT_GROUP_SCHED
10475 .name
= "rt_runtime_us",
10476 .read_s64
= cpu_rt_runtime_read
,
10477 .write_s64
= cpu_rt_runtime_write
,
10480 .name
= "rt_period_us",
10481 .read_u64
= cpu_rt_period_read_uint
,
10482 .write_u64
= cpu_rt_period_write_uint
,
10487 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10489 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10492 struct cgroup_subsys cpu_cgroup_subsys
= {
10494 .create
= cpu_cgroup_create
,
10495 .destroy
= cpu_cgroup_destroy
,
10496 .can_attach
= cpu_cgroup_can_attach
,
10497 .attach
= cpu_cgroup_attach
,
10498 .populate
= cpu_cgroup_populate
,
10499 .subsys_id
= cpu_cgroup_subsys_id
,
10503 #endif /* CONFIG_CGROUP_SCHED */
10505 #ifdef CONFIG_CGROUP_CPUACCT
10508 * CPU accounting code for task groups.
10510 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10511 * (balbir@in.ibm.com).
10514 /* track cpu usage of a group of tasks and its child groups */
10516 struct cgroup_subsys_state css
;
10517 /* cpuusage holds pointer to a u64-type object on every cpu */
10519 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10520 struct cpuacct
*parent
;
10523 struct cgroup_subsys cpuacct_subsys
;
10525 /* return cpu accounting group corresponding to this container */
10526 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10528 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10529 struct cpuacct
, css
);
10532 /* return cpu accounting group to which this task belongs */
10533 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10535 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10536 struct cpuacct
, css
);
10539 /* create a new cpu accounting group */
10540 static struct cgroup_subsys_state
*cpuacct_create(
10541 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10543 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10549 ca
->cpuusage
= alloc_percpu(u64
);
10553 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10554 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10555 goto out_free_counters
;
10558 ca
->parent
= cgroup_ca(cgrp
->parent
);
10564 percpu_counter_destroy(&ca
->cpustat
[i
]);
10565 free_percpu(ca
->cpuusage
);
10569 return ERR_PTR(-ENOMEM
);
10572 /* destroy an existing cpu accounting group */
10574 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10576 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10579 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10580 percpu_counter_destroy(&ca
->cpustat
[i
]);
10581 free_percpu(ca
->cpuusage
);
10585 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10587 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10590 #ifndef CONFIG_64BIT
10592 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10594 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10596 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10604 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10606 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10608 #ifndef CONFIG_64BIT
10610 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10612 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10614 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10620 /* return total cpu usage (in nanoseconds) of a group */
10621 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10623 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10624 u64 totalcpuusage
= 0;
10627 for_each_present_cpu(i
)
10628 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10630 return totalcpuusage
;
10633 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10636 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10645 for_each_present_cpu(i
)
10646 cpuacct_cpuusage_write(ca
, i
, 0);
10652 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10653 struct seq_file
*m
)
10655 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10659 for_each_present_cpu(i
) {
10660 percpu
= cpuacct_cpuusage_read(ca
, i
);
10661 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10663 seq_printf(m
, "\n");
10667 static const char *cpuacct_stat_desc
[] = {
10668 [CPUACCT_STAT_USER
] = "user",
10669 [CPUACCT_STAT_SYSTEM
] = "system",
10672 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10673 struct cgroup_map_cb
*cb
)
10675 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10678 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10679 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10680 val
= cputime64_to_clock_t(val
);
10681 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10686 static struct cftype files
[] = {
10689 .read_u64
= cpuusage_read
,
10690 .write_u64
= cpuusage_write
,
10693 .name
= "usage_percpu",
10694 .read_seq_string
= cpuacct_percpu_seq_read
,
10698 .read_map
= cpuacct_stats_show
,
10702 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10704 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10708 * charge this task's execution time to its accounting group.
10710 * called with rq->lock held.
10712 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10714 struct cpuacct
*ca
;
10717 if (unlikely(!cpuacct_subsys
.active
))
10720 cpu
= task_cpu(tsk
);
10726 for (; ca
; ca
= ca
->parent
) {
10727 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10728 *cpuusage
+= cputime
;
10735 * Charge the system/user time to the task's accounting group.
10737 static void cpuacct_update_stats(struct task_struct
*tsk
,
10738 enum cpuacct_stat_index idx
, cputime_t val
)
10740 struct cpuacct
*ca
;
10742 if (unlikely(!cpuacct_subsys
.active
))
10749 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10755 struct cgroup_subsys cpuacct_subsys
= {
10757 .create
= cpuacct_create
,
10758 .destroy
= cpuacct_destroy
,
10759 .populate
= cpuacct_populate
,
10760 .subsys_id
= cpuacct_subsys_id
,
10762 #endif /* CONFIG_CGROUP_CPUACCT */
10766 int rcu_expedited_torture_stats(char *page
)
10770 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10772 void synchronize_sched_expedited(void)
10775 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10777 #else /* #ifndef CONFIG_SMP */
10779 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10780 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10782 #define RCU_EXPEDITED_STATE_POST -2
10783 #define RCU_EXPEDITED_STATE_IDLE -1
10785 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10787 int rcu_expedited_torture_stats(char *page
)
10792 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10793 for_each_online_cpu(cpu
) {
10794 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10795 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10797 cnt
+= sprintf(&page
[cnt
], "\n");
10800 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10802 static long synchronize_sched_expedited_count
;
10805 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10806 * approach to force grace period to end quickly. This consumes
10807 * significant time on all CPUs, and is thus not recommended for
10808 * any sort of common-case code.
10810 * Note that it is illegal to call this function while holding any
10811 * lock that is acquired by a CPU-hotplug notifier. Failing to
10812 * observe this restriction will result in deadlock.
10814 void synchronize_sched_expedited(void)
10817 unsigned long flags
;
10818 bool need_full_sync
= 0;
10820 struct migration_req
*req
;
10824 smp_mb(); /* ensure prior mod happens before capturing snap. */
10825 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10827 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10829 if (trycount
++ < 10)
10830 udelay(trycount
* num_online_cpus());
10832 synchronize_sched();
10835 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10836 smp_mb(); /* ensure test happens before caller kfree */
10841 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10842 for_each_online_cpu(cpu
) {
10844 req
= &per_cpu(rcu_migration_req
, cpu
);
10845 init_completion(&req
->done
);
10847 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10848 spin_lock_irqsave(&rq
->lock
, flags
);
10849 list_add(&req
->list
, &rq
->migration_queue
);
10850 spin_unlock_irqrestore(&rq
->lock
, flags
);
10851 wake_up_process(rq
->migration_thread
);
10853 for_each_online_cpu(cpu
) {
10854 rcu_expedited_state
= cpu
;
10855 req
= &per_cpu(rcu_migration_req
, cpu
);
10857 wait_for_completion(&req
->done
);
10858 spin_lock_irqsave(&rq
->lock
, flags
);
10859 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10860 need_full_sync
= 1;
10861 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10862 spin_unlock_irqrestore(&rq
->lock
, flags
);
10864 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10865 mutex_unlock(&rcu_sched_expedited_mutex
);
10867 if (need_full_sync
)
10868 synchronize_sched();
10870 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10872 #endif /* #else #ifndef CONFIG_SMP */