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
);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group
.children
);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group
;
346 /* return group to which a task belongs */
347 static inline struct task_group
*task_group(struct task_struct
*p
)
349 struct task_group
*tg
;
351 #ifdef CONFIG_USER_SCHED
353 tg
= __task_cred(p
)->user
->tg
;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
357 struct task_group
, css
);
359 tg
= &init_task_group
;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
369 p
->se
.parent
= task_group(p
)->se
[cpu
];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
374 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
380 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
381 static inline struct task_group
*task_group(struct task_struct
*p
)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load
;
391 unsigned long nr_running
;
396 struct rb_root tasks_timeline
;
397 struct rb_node
*rb_leftmost
;
399 struct list_head tasks
;
400 struct list_head
*balance_iterator
;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity
*curr
, *next
, *last
;
408 unsigned int nr_spread_over
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list
;
422 struct task_group
*tg
; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight
;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load
;
439 * this cpu's part of tg->shares
441 unsigned long shares
;
444 * load.weight at the time we set shares
446 unsigned long rq_weight
;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active
;
454 unsigned long rt_nr_running
;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr
; /* highest queued rt task prio */
459 int next
; /* next highest */
464 unsigned long rt_nr_migratory
;
465 unsigned long rt_nr_total
;
467 struct plist_head pushable_tasks
;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock
;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted
;
479 struct list_head leaf_rt_rq_list
;
480 struct task_group
*tg
;
481 struct sched_rt_entity
*rt_se
;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online
;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask
;
507 struct cpupri cpupri
;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain
;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running
;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
538 unsigned long last_tick_seen
;
539 unsigned char in_nohz_recently
;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load
;
543 unsigned long nr_load_updates
;
545 u64 nr_migrations_in
;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list
;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list
;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible
;
566 struct task_struct
*curr
, *idle
;
567 unsigned long next_balance
;
568 struct mm_struct
*prev_mm
;
575 struct root_domain
*rd
;
576 struct sched_domain
*sd
;
578 unsigned char idle_at_tick
;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task
;
589 struct task_struct
*migration_thread
;
590 struct list_head migration_queue
;
598 /* calc_load related fields */
599 unsigned long calc_load_update
;
600 long calc_load_active
;
602 #ifdef CONFIG_SCHED_HRTICK
604 int hrtick_csd_pending
;
605 struct call_single_data hrtick_csd
;
607 struct hrtimer hrtick_timer
;
610 #ifdef CONFIG_SCHEDSTATS
612 struct sched_info rq_sched_info
;
613 unsigned long long rq_cpu_time
;
614 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
616 /* sys_sched_yield() stats */
617 unsigned int yld_count
;
619 /* schedule() stats */
620 unsigned int sched_switch
;
621 unsigned int sched_count
;
622 unsigned int sched_goidle
;
624 /* try_to_wake_up() stats */
625 unsigned int ttwu_count
;
626 unsigned int ttwu_local
;
629 unsigned int bkl_count
;
633 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
636 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
638 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
641 static inline int cpu_of(struct rq
*rq
)
651 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
652 * See detach_destroy_domains: synchronize_sched for details.
654 * The domain tree of any CPU may only be accessed from within
655 * preempt-disabled sections.
657 #define for_each_domain(cpu, __sd) \
658 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
660 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
661 #define this_rq() (&__get_cpu_var(runqueues))
662 #define task_rq(p) cpu_rq(task_cpu(p))
663 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
664 #define raw_rq() (&__raw_get_cpu_var(runqueues))
666 inline void update_rq_clock(struct rq
*rq
)
668 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
672 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 #ifdef CONFIG_SCHED_DEBUG
675 # define const_debug __read_mostly
677 # define const_debug static const
682 * @cpu: the processor in question.
684 * Returns true if the current cpu runqueue is locked.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu
)
690 return spin_is_locked(&cpu_rq(cpu
)->lock
);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug
unsigned int sysctl_sched_features
=
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly
char *sched_feat_names
[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file
*m
, void *v
)
730 for (i
= 0; sched_feat_names
[i
]; i
++) {
731 if (!(sysctl_sched_features
& (1UL << i
)))
733 seq_printf(m
, "%s ", sched_feat_names
[i
]);
741 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
742 size_t cnt
, loff_t
*ppos
)
752 if (copy_from_user(&buf
, ubuf
, cnt
))
757 if (strncmp(buf
, "NO_", 3) == 0) {
762 for (i
= 0; sched_feat_names
[i
]; i
++) {
763 int len
= strlen(sched_feat_names
[i
]);
765 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
767 sysctl_sched_features
&= ~(1UL << i
);
769 sysctl_sched_features
|= (1UL << i
);
774 if (!sched_feat_names
[i
])
782 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
784 return single_open(filp
, sched_feat_show
, NULL
);
787 static const struct file_operations sched_feat_fops
= {
788 .open
= sched_feat_open
,
789 .write
= sched_feat_write
,
792 .release
= single_release
,
795 static __init
int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
802 late_initcall(sched_init_debug
);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit
= 250000;
821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
825 unsigned int sysctl_sched_shares_thresh
= 4;
828 * period over which we average the RT time consumption, measured
833 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
836 * period over which we measure -rt task cpu usage in us.
839 unsigned int sysctl_sched_rt_period
= 1000000;
841 static __read_mostly
int scheduler_running
;
844 * part of the period that we allow rt tasks to run in us.
847 int sysctl_sched_rt_runtime
= 950000;
849 static inline u64
global_rt_period(void)
851 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
854 static inline u64
global_rt_runtime(void)
856 if (sysctl_sched_rt_runtime
< 0)
859 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
862 #ifndef prepare_arch_switch
863 # define prepare_arch_switch(next) do { } while (0)
865 #ifndef finish_arch_switch
866 # define finish_arch_switch(prev) do { } while (0)
869 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
871 return rq
->curr
== p
;
874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
875 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
877 return task_current(rq
, p
);
880 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
884 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq
->lock
.owner
= current
;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
897 spin_unlock_irq(&rq
->lock
);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
906 return task_current(rq
, p
);
910 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
914 * We can optimise this out completely for !SMP, because the
915 * SMP rebalancing from interrupt is the only thing that cares
920 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 spin_unlock_irq(&rq
->lock
);
923 spin_unlock(&rq
->lock
);
927 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
931 * After ->oncpu is cleared, the task can be moved to a different CPU.
932 * We must ensure this doesn't happen until the switch is completely
938 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
942 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
945 * __task_rq_lock - lock the runqueue a given task resides on.
946 * Must be called interrupts disabled.
948 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
952 struct rq
*rq
= task_rq(p
);
953 spin_lock(&rq
->lock
);
954 if (likely(rq
== task_rq(p
)))
956 spin_unlock(&rq
->lock
);
961 * task_rq_lock - lock the runqueue a given task resides on and disable
962 * interrupts. Note the ordering: we can safely lookup the task_rq without
963 * explicitly disabling preemption.
965 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
971 local_irq_save(*flags
);
973 spin_lock(&rq
->lock
);
974 if (likely(rq
== task_rq(p
)))
976 spin_unlock_irqrestore(&rq
->lock
, *flags
);
980 void task_rq_unlock_wait(struct task_struct
*p
)
982 struct rq
*rq
= task_rq(p
);
984 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
985 spin_unlock_wait(&rq
->lock
);
988 static void __task_rq_unlock(struct rq
*rq
)
991 spin_unlock(&rq
->lock
);
994 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
997 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1001 * this_rq_lock - lock this runqueue and disable interrupts.
1003 static struct rq
*this_rq_lock(void)
1004 __acquires(rq
->lock
)
1008 local_irq_disable();
1010 spin_lock(&rq
->lock
);
1015 #ifdef CONFIG_SCHED_HRTICK
1017 * Use HR-timers to deliver accurate preemption points.
1019 * Its all a bit involved since we cannot program an hrt while holding the
1020 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1023 * When we get rescheduled we reprogram the hrtick_timer outside of the
1029 * - enabled by features
1030 * - hrtimer is actually high res
1032 static inline int hrtick_enabled(struct rq
*rq
)
1034 if (!sched_feat(HRTICK
))
1036 if (!cpu_active(cpu_of(rq
)))
1038 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1041 static void hrtick_clear(struct rq
*rq
)
1043 if (hrtimer_active(&rq
->hrtick_timer
))
1044 hrtimer_cancel(&rq
->hrtick_timer
);
1048 * High-resolution timer tick.
1049 * Runs from hardirq context with interrupts disabled.
1051 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1053 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1055 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1057 spin_lock(&rq
->lock
);
1058 update_rq_clock(rq
);
1059 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1060 spin_unlock(&rq
->lock
);
1062 return HRTIMER_NORESTART
;
1067 * called from hardirq (IPI) context
1069 static void __hrtick_start(void *arg
)
1071 struct rq
*rq
= arg
;
1073 spin_lock(&rq
->lock
);
1074 hrtimer_restart(&rq
->hrtick_timer
);
1075 rq
->hrtick_csd_pending
= 0;
1076 spin_unlock(&rq
->lock
);
1080 * Called to set the hrtick timer state.
1082 * called with rq->lock held and irqs disabled
1084 static void hrtick_start(struct rq
*rq
, u64 delay
)
1086 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1087 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1089 hrtimer_set_expires(timer
, time
);
1091 if (rq
== this_rq()) {
1092 hrtimer_restart(timer
);
1093 } else if (!rq
->hrtick_csd_pending
) {
1094 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1095 rq
->hrtick_csd_pending
= 1;
1100 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1102 int cpu
= (int)(long)hcpu
;
1105 case CPU_UP_CANCELED
:
1106 case CPU_UP_CANCELED_FROZEN
:
1107 case CPU_DOWN_PREPARE
:
1108 case CPU_DOWN_PREPARE_FROZEN
:
1110 case CPU_DEAD_FROZEN
:
1111 hrtick_clear(cpu_rq(cpu
));
1118 static __init
void init_hrtick(void)
1120 hotcpu_notifier(hotplug_hrtick
, 0);
1124 * Called to set the hrtick timer state.
1126 * called with rq->lock held and irqs disabled
1128 static void hrtick_start(struct rq
*rq
, u64 delay
)
1130 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1131 HRTIMER_MODE_REL_PINNED
, 0);
1134 static inline void init_hrtick(void)
1137 #endif /* CONFIG_SMP */
1139 static void init_rq_hrtick(struct rq
*rq
)
1142 rq
->hrtick_csd_pending
= 0;
1144 rq
->hrtick_csd
.flags
= 0;
1145 rq
->hrtick_csd
.func
= __hrtick_start
;
1146 rq
->hrtick_csd
.info
= rq
;
1149 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1150 rq
->hrtick_timer
.function
= hrtick
;
1152 #else /* CONFIG_SCHED_HRTICK */
1153 static inline void hrtick_clear(struct rq
*rq
)
1157 static inline void init_rq_hrtick(struct rq
*rq
)
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SCHED_HRTICK */
1167 * resched_task - mark a task 'to be rescheduled now'.
1169 * On UP this means the setting of the need_resched flag, on SMP it
1170 * might also involve a cross-CPU call to trigger the scheduler on
1175 #ifndef tsk_is_polling
1176 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1179 static void resched_task(struct task_struct
*p
)
1183 assert_spin_locked(&task_rq(p
)->lock
);
1185 if (test_tsk_need_resched(p
))
1188 set_tsk_need_resched(p
);
1191 if (cpu
== smp_processor_id())
1194 /* NEED_RESCHED must be visible before we test polling */
1196 if (!tsk_is_polling(p
))
1197 smp_send_reschedule(cpu
);
1200 static void resched_cpu(int cpu
)
1202 struct rq
*rq
= cpu_rq(cpu
);
1203 unsigned long flags
;
1205 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1207 resched_task(cpu_curr(cpu
));
1208 spin_unlock_irqrestore(&rq
->lock
, flags
);
1213 * When add_timer_on() enqueues a timer into the timer wheel of an
1214 * idle CPU then this timer might expire before the next timer event
1215 * which is scheduled to wake up that CPU. In case of a completely
1216 * idle system the next event might even be infinite time into the
1217 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1218 * leaves the inner idle loop so the newly added timer is taken into
1219 * account when the CPU goes back to idle and evaluates the timer
1220 * wheel for the next timer event.
1222 void wake_up_idle_cpu(int cpu
)
1224 struct rq
*rq
= cpu_rq(cpu
);
1226 if (cpu
== smp_processor_id())
1230 * This is safe, as this function is called with the timer
1231 * wheel base lock of (cpu) held. When the CPU is on the way
1232 * to idle and has not yet set rq->curr to idle then it will
1233 * be serialized on the timer wheel base lock and take the new
1234 * timer into account automatically.
1236 if (rq
->curr
!= rq
->idle
)
1240 * We can set TIF_RESCHED on the idle task of the other CPU
1241 * lockless. The worst case is that the other CPU runs the
1242 * idle task through an additional NOOP schedule()
1244 set_tsk_need_resched(rq
->idle
);
1246 /* NEED_RESCHED must be visible before we test polling */
1248 if (!tsk_is_polling(rq
->idle
))
1249 smp_send_reschedule(cpu
);
1251 #endif /* CONFIG_NO_HZ */
1253 static u64
sched_avg_period(void)
1255 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1258 static void sched_avg_update(struct rq
*rq
)
1260 s64 period
= sched_avg_period();
1262 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1263 rq
->age_stamp
+= period
;
1268 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1270 rq
->rt_avg
+= rt_delta
;
1271 sched_avg_update(rq
);
1274 #else /* !CONFIG_SMP */
1275 static void resched_task(struct task_struct
*p
)
1277 assert_spin_locked(&task_rq(p
)->lock
);
1278 set_tsk_need_resched(p
);
1281 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1284 #endif /* CONFIG_SMP */
1286 #if BITS_PER_LONG == 32
1287 # define WMULT_CONST (~0UL)
1289 # define WMULT_CONST (1UL << 32)
1292 #define WMULT_SHIFT 32
1295 * Shift right and round:
1297 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1300 * delta *= weight / lw
1302 static unsigned long
1303 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1304 struct load_weight
*lw
)
1308 if (!lw
->inv_weight
) {
1309 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1312 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1316 tmp
= (u64
)delta_exec
* weight
;
1318 * Check whether we'd overflow the 64-bit multiplication:
1320 if (unlikely(tmp
> WMULT_CONST
))
1321 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1324 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1326 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1329 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1335 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1342 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1343 * of tasks with abnormal "nice" values across CPUs the contribution that
1344 * each task makes to its run queue's load is weighted according to its
1345 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1346 * scaled version of the new time slice allocation that they receive on time
1350 #define WEIGHT_IDLEPRIO 3
1351 #define WMULT_IDLEPRIO 1431655765
1354 * Nice levels are multiplicative, with a gentle 10% change for every
1355 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1356 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1357 * that remained on nice 0.
1359 * The "10% effect" is relative and cumulative: from _any_ nice level,
1360 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1361 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1362 * If a task goes up by ~10% and another task goes down by ~10% then
1363 * the relative distance between them is ~25%.)
1365 static const int prio_to_weight
[40] = {
1366 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1367 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1368 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1369 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1370 /* 0 */ 1024, 820, 655, 526, 423,
1371 /* 5 */ 335, 272, 215, 172, 137,
1372 /* 10 */ 110, 87, 70, 56, 45,
1373 /* 15 */ 36, 29, 23, 18, 15,
1377 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1379 * In cases where the weight does not change often, we can use the
1380 * precalculated inverse to speed up arithmetics by turning divisions
1381 * into multiplications:
1383 static const u32 prio_to_wmult
[40] = {
1384 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1385 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1386 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1387 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1388 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1389 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1390 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1391 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1394 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1397 * runqueue iterator, to support SMP load-balancing between different
1398 * scheduling classes, without having to expose their internal data
1399 * structures to the load-balancing proper:
1401 struct rq_iterator
{
1403 struct task_struct
*(*start
)(void *);
1404 struct task_struct
*(*next
)(void *);
1408 static unsigned long
1409 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1410 unsigned long max_load_move
, struct sched_domain
*sd
,
1411 enum cpu_idle_type idle
, int *all_pinned
,
1412 int *this_best_prio
, struct rq_iterator
*iterator
);
1415 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1416 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1417 struct rq_iterator
*iterator
);
1420 /* Time spent by the tasks of the cpu accounting group executing in ... */
1421 enum cpuacct_stat_index
{
1422 CPUACCT_STAT_USER
, /* ... user mode */
1423 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1425 CPUACCT_STAT_NSTATS
,
1428 #ifdef CONFIG_CGROUP_CPUACCT
1429 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1430 static void cpuacct_update_stats(struct task_struct
*tsk
,
1431 enum cpuacct_stat_index idx
, cputime_t val
);
1433 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1434 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1435 enum cpuacct_stat_index idx
, cputime_t val
) {}
1438 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1440 update_load_add(&rq
->load
, load
);
1443 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1445 update_load_sub(&rq
->load
, load
);
1448 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1449 typedef int (*tg_visitor
)(struct task_group
*, void *);
1452 * Iterate the full tree, calling @down when first entering a node and @up when
1453 * leaving it for the final time.
1455 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1457 struct task_group
*parent
, *child
;
1461 parent
= &root_task_group
;
1463 ret
= (*down
)(parent
, data
);
1466 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1473 ret
= (*up
)(parent
, data
);
1478 parent
= parent
->parent
;
1487 static int tg_nop(struct task_group
*tg
, void *data
)
1494 /* Used instead of source_load when we know the type == 0 */
1495 static unsigned long weighted_cpuload(const int cpu
)
1497 return cpu_rq(cpu
)->load
.weight
;
1501 * Return a low guess at the load of a migration-source cpu weighted
1502 * according to the scheduling class and "nice" value.
1504 * We want to under-estimate the load of migration sources, to
1505 * balance conservatively.
1507 static unsigned long source_load(int cpu
, int type
)
1509 struct rq
*rq
= cpu_rq(cpu
);
1510 unsigned long total
= weighted_cpuload(cpu
);
1512 if (type
== 0 || !sched_feat(LB_BIAS
))
1515 return min(rq
->cpu_load
[type
-1], total
);
1519 * Return a high guess at the load of a migration-target cpu weighted
1520 * according to the scheduling class and "nice" value.
1522 static unsigned long target_load(int cpu
, int type
)
1524 struct rq
*rq
= cpu_rq(cpu
);
1525 unsigned long total
= weighted_cpuload(cpu
);
1527 if (type
== 0 || !sched_feat(LB_BIAS
))
1530 return max(rq
->cpu_load
[type
-1], total
);
1533 static struct sched_group
*group_of(int cpu
)
1535 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1543 static unsigned long power_of(int cpu
)
1545 struct sched_group
*group
= group_of(cpu
);
1548 return SCHED_LOAD_SCALE
;
1550 return group
->cpu_power
;
1553 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1555 static unsigned long cpu_avg_load_per_task(int cpu
)
1557 struct rq
*rq
= cpu_rq(cpu
);
1558 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1561 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1563 rq
->avg_load_per_task
= 0;
1565 return rq
->avg_load_per_task
;
1568 #ifdef CONFIG_FAIR_GROUP_SCHED
1570 static __read_mostly
unsigned long *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 unsigned long *usd_rq_weight
)
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 unsigned long *usd_rq_weight
;
1621 struct sched_domain
*sd
= data
;
1622 unsigned long flags
;
1628 local_irq_save(flags
);
1629 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
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_rq_weight
);
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
);
1999 * kthread_bind - bind a just-created kthread to a cpu.
2000 * @p: thread created by kthread_create().
2001 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2003 * Description: This function is equivalent to set_cpus_allowed(),
2004 * except that @cpu doesn't need to be online, and the thread must be
2005 * stopped (i.e., just returned from kthread_create()).
2007 * Function lives here instead of kthread.c because it messes with
2008 * scheduler internals which require locking.
2010 void kthread_bind(struct task_struct
*p
, unsigned int cpu
)
2012 struct rq
*rq
= cpu_rq(cpu
);
2013 unsigned long flags
;
2015 /* Must have done schedule() in kthread() before we set_task_cpu */
2016 if (!wait_task_inactive(p
, TASK_UNINTERRUPTIBLE
)) {
2021 spin_lock_irqsave(&rq
->lock
, flags
);
2022 set_task_cpu(p
, cpu
);
2023 p
->cpus_allowed
= cpumask_of_cpu(cpu
);
2024 p
->rt
.nr_cpus_allowed
= 1;
2025 p
->flags
|= PF_THREAD_BOUND
;
2026 spin_unlock_irqrestore(&rq
->lock
, flags
);
2028 EXPORT_SYMBOL(kthread_bind
);
2032 * Is this task likely cache-hot:
2035 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2040 * Buddy candidates are cache hot:
2042 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2043 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2044 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2047 if (p
->sched_class
!= &fair_sched_class
)
2050 if (sysctl_sched_migration_cost
== -1)
2052 if (sysctl_sched_migration_cost
== 0)
2055 delta
= now
- p
->se
.exec_start
;
2057 return delta
< (s64
)sysctl_sched_migration_cost
;
2061 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2063 int old_cpu
= task_cpu(p
);
2064 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2065 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2066 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2069 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2071 trace_sched_migrate_task(p
, new_cpu
);
2073 #ifdef CONFIG_SCHEDSTATS
2074 if (p
->se
.wait_start
)
2075 p
->se
.wait_start
-= clock_offset
;
2076 if (p
->se
.sleep_start
)
2077 p
->se
.sleep_start
-= clock_offset
;
2078 if (p
->se
.block_start
)
2079 p
->se
.block_start
-= clock_offset
;
2081 if (old_cpu
!= new_cpu
) {
2082 p
->se
.nr_migrations
++;
2083 new_rq
->nr_migrations_in
++;
2084 #ifdef CONFIG_SCHEDSTATS
2085 if (task_hot(p
, old_rq
->clock
, NULL
))
2086 schedstat_inc(p
, se
.nr_forced2_migrations
);
2088 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2091 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2092 new_cfsrq
->min_vruntime
;
2094 __set_task_cpu(p
, new_cpu
);
2097 struct migration_req
{
2098 struct list_head list
;
2100 struct task_struct
*task
;
2103 struct completion done
;
2107 * The task's runqueue lock must be held.
2108 * Returns true if you have to wait for migration thread.
2111 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2113 struct rq
*rq
= task_rq(p
);
2116 * If the task is not on a runqueue (and not running), then
2117 * it is sufficient to simply update the task's cpu field.
2119 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2120 set_task_cpu(p
, dest_cpu
);
2124 init_completion(&req
->done
);
2126 req
->dest_cpu
= dest_cpu
;
2127 list_add(&req
->list
, &rq
->migration_queue
);
2133 * wait_task_context_switch - wait for a thread to complete at least one
2136 * @p must not be current.
2138 void wait_task_context_switch(struct task_struct
*p
)
2140 unsigned long nvcsw
, nivcsw
, flags
;
2148 * The runqueue is assigned before the actual context
2149 * switch. We need to take the runqueue lock.
2151 * We could check initially without the lock but it is
2152 * very likely that we need to take the lock in every
2155 rq
= task_rq_lock(p
, &flags
);
2156 running
= task_running(rq
, p
);
2157 task_rq_unlock(rq
, &flags
);
2159 if (likely(!running
))
2162 * The switch count is incremented before the actual
2163 * context switch. We thus wait for two switches to be
2164 * sure at least one completed.
2166 if ((p
->nvcsw
- nvcsw
) > 1)
2168 if ((p
->nivcsw
- nivcsw
) > 1)
2176 * wait_task_inactive - wait for a thread to unschedule.
2178 * If @match_state is nonzero, it's the @p->state value just checked and
2179 * not expected to change. If it changes, i.e. @p might have woken up,
2180 * then return zero. When we succeed in waiting for @p to be off its CPU,
2181 * we return a positive number (its total switch count). If a second call
2182 * a short while later returns the same number, the caller can be sure that
2183 * @p has remained unscheduled the whole time.
2185 * The caller must ensure that the task *will* unschedule sometime soon,
2186 * else this function might spin for a *long* time. This function can't
2187 * be called with interrupts off, or it may introduce deadlock with
2188 * smp_call_function() if an IPI is sent by the same process we are
2189 * waiting to become inactive.
2191 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2193 unsigned long flags
;
2200 * We do the initial early heuristics without holding
2201 * any task-queue locks at all. We'll only try to get
2202 * the runqueue lock when things look like they will
2208 * If the task is actively running on another CPU
2209 * still, just relax and busy-wait without holding
2212 * NOTE! Since we don't hold any locks, it's not
2213 * even sure that "rq" stays as the right runqueue!
2214 * But we don't care, since "task_running()" will
2215 * return false if the runqueue has changed and p
2216 * is actually now running somewhere else!
2218 while (task_running(rq
, p
)) {
2219 if (match_state
&& unlikely(p
->state
!= match_state
))
2225 * Ok, time to look more closely! We need the rq
2226 * lock now, to be *sure*. If we're wrong, we'll
2227 * just go back and repeat.
2229 rq
= task_rq_lock(p
, &flags
);
2230 trace_sched_wait_task(rq
, p
);
2231 running
= task_running(rq
, p
);
2232 on_rq
= p
->se
.on_rq
;
2234 if (!match_state
|| p
->state
== match_state
)
2235 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2236 task_rq_unlock(rq
, &flags
);
2239 * If it changed from the expected state, bail out now.
2241 if (unlikely(!ncsw
))
2245 * Was it really running after all now that we
2246 * checked with the proper locks actually held?
2248 * Oops. Go back and try again..
2250 if (unlikely(running
)) {
2256 * It's not enough that it's not actively running,
2257 * it must be off the runqueue _entirely_, and not
2260 * So if it was still runnable (but just not actively
2261 * running right now), it's preempted, and we should
2262 * yield - it could be a while.
2264 if (unlikely(on_rq
)) {
2265 schedule_timeout_uninterruptible(1);
2270 * Ahh, all good. It wasn't running, and it wasn't
2271 * runnable, which means that it will never become
2272 * running in the future either. We're all done!
2281 * kick_process - kick a running thread to enter/exit the kernel
2282 * @p: the to-be-kicked thread
2284 * Cause a process which is running on another CPU to enter
2285 * kernel-mode, without any delay. (to get signals handled.)
2287 * NOTE: this function doesnt have to take the runqueue lock,
2288 * because all it wants to ensure is that the remote task enters
2289 * the kernel. If the IPI races and the task has been migrated
2290 * to another CPU then no harm is done and the purpose has been
2293 void kick_process(struct task_struct
*p
)
2299 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2300 smp_send_reschedule(cpu
);
2303 EXPORT_SYMBOL_GPL(kick_process
);
2304 #endif /* CONFIG_SMP */
2307 * task_oncpu_function_call - call a function on the cpu on which a task runs
2308 * @p: the task to evaluate
2309 * @func: the function to be called
2310 * @info: the function call argument
2312 * Calls the function @func when the task is currently running. This might
2313 * be on the current CPU, which just calls the function directly
2315 void task_oncpu_function_call(struct task_struct
*p
,
2316 void (*func
) (void *info
), void *info
)
2323 smp_call_function_single(cpu
, func
, info
, 1);
2328 * try_to_wake_up - wake up a thread
2329 * @p: the to-be-woken-up thread
2330 * @state: the mask of task states that can be woken
2331 * @sync: do a synchronous wakeup?
2333 * Put it on the run-queue if it's not already there. The "current"
2334 * thread is always on the run-queue (except when the actual
2335 * re-schedule is in progress), and as such you're allowed to do
2336 * the simpler "current->state = TASK_RUNNING" to mark yourself
2337 * runnable without the overhead of this.
2339 * returns failure only if the task is already active.
2341 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2344 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2345 unsigned long flags
;
2346 struct rq
*rq
, *orig_rq
;
2348 if (!sched_feat(SYNC_WAKEUPS
))
2349 wake_flags
&= ~WF_SYNC
;
2351 this_cpu
= get_cpu();
2354 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2355 update_rq_clock(rq
);
2356 if (!(p
->state
& state
))
2366 if (unlikely(task_running(rq
, p
)))
2370 * In order to handle concurrent wakeups and release the rq->lock
2371 * we put the task in TASK_WAKING state.
2373 * First fix up the nr_uninterruptible count:
2375 if (task_contributes_to_load(p
))
2376 rq
->nr_uninterruptible
--;
2377 p
->state
= TASK_WAKING
;
2378 task_rq_unlock(rq
, &flags
);
2380 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2381 if (cpu
!= orig_cpu
)
2382 set_task_cpu(p
, cpu
);
2384 rq
= task_rq_lock(p
, &flags
);
2387 update_rq_clock(rq
);
2389 WARN_ON(p
->state
!= TASK_WAKING
);
2392 #ifdef CONFIG_SCHEDSTATS
2393 schedstat_inc(rq
, ttwu_count
);
2394 if (cpu
== this_cpu
)
2395 schedstat_inc(rq
, ttwu_local
);
2397 struct sched_domain
*sd
;
2398 for_each_domain(this_cpu
, sd
) {
2399 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2400 schedstat_inc(sd
, ttwu_wake_remote
);
2405 #endif /* CONFIG_SCHEDSTATS */
2408 #endif /* CONFIG_SMP */
2409 schedstat_inc(p
, se
.nr_wakeups
);
2410 if (wake_flags
& WF_SYNC
)
2411 schedstat_inc(p
, se
.nr_wakeups_sync
);
2412 if (orig_cpu
!= cpu
)
2413 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2414 if (cpu
== this_cpu
)
2415 schedstat_inc(p
, se
.nr_wakeups_local
);
2417 schedstat_inc(p
, se
.nr_wakeups_remote
);
2418 activate_task(rq
, p
, 1);
2422 * Only attribute actual wakeups done by this task.
2424 if (!in_interrupt()) {
2425 struct sched_entity
*se
= ¤t
->se
;
2426 u64 sample
= se
->sum_exec_runtime
;
2428 if (se
->last_wakeup
)
2429 sample
-= se
->last_wakeup
;
2431 sample
-= se
->start_runtime
;
2432 update_avg(&se
->avg_wakeup
, sample
);
2434 se
->last_wakeup
= se
->sum_exec_runtime
;
2438 trace_sched_wakeup(rq
, p
, success
);
2439 check_preempt_curr(rq
, p
, wake_flags
);
2441 p
->state
= TASK_RUNNING
;
2443 if (p
->sched_class
->task_wake_up
)
2444 p
->sched_class
->task_wake_up(rq
, p
);
2446 if (unlikely(rq
->idle_stamp
)) {
2447 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2448 u64 max
= 2*sysctl_sched_migration_cost
;
2453 update_avg(&rq
->avg_idle
, delta
);
2458 task_rq_unlock(rq
, &flags
);
2465 * wake_up_process - Wake up a specific process
2466 * @p: The process to be woken up.
2468 * Attempt to wake up the nominated process and move it to the set of runnable
2469 * processes. Returns 1 if the process was woken up, 0 if it was already
2472 * It may be assumed that this function implies a write memory barrier before
2473 * changing the task state if and only if any tasks are woken up.
2475 int wake_up_process(struct task_struct
*p
)
2477 return try_to_wake_up(p
, TASK_ALL
, 0);
2479 EXPORT_SYMBOL(wake_up_process
);
2481 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2483 return try_to_wake_up(p
, state
, 0);
2487 * Perform scheduler related setup for a newly forked process p.
2488 * p is forked by current.
2490 * __sched_fork() is basic setup used by init_idle() too:
2492 static void __sched_fork(struct task_struct
*p
)
2494 p
->se
.exec_start
= 0;
2495 p
->se
.sum_exec_runtime
= 0;
2496 p
->se
.prev_sum_exec_runtime
= 0;
2497 p
->se
.nr_migrations
= 0;
2498 p
->se
.last_wakeup
= 0;
2499 p
->se
.avg_overlap
= 0;
2500 p
->se
.start_runtime
= 0;
2501 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2502 p
->se
.avg_running
= 0;
2504 #ifdef CONFIG_SCHEDSTATS
2505 p
->se
.wait_start
= 0;
2507 p
->se
.wait_count
= 0;
2510 p
->se
.sleep_start
= 0;
2511 p
->se
.sleep_max
= 0;
2512 p
->se
.sum_sleep_runtime
= 0;
2514 p
->se
.block_start
= 0;
2515 p
->se
.block_max
= 0;
2517 p
->se
.slice_max
= 0;
2519 p
->se
.nr_migrations_cold
= 0;
2520 p
->se
.nr_failed_migrations_affine
= 0;
2521 p
->se
.nr_failed_migrations_running
= 0;
2522 p
->se
.nr_failed_migrations_hot
= 0;
2523 p
->se
.nr_forced_migrations
= 0;
2524 p
->se
.nr_forced2_migrations
= 0;
2526 p
->se
.nr_wakeups
= 0;
2527 p
->se
.nr_wakeups_sync
= 0;
2528 p
->se
.nr_wakeups_migrate
= 0;
2529 p
->se
.nr_wakeups_local
= 0;
2530 p
->se
.nr_wakeups_remote
= 0;
2531 p
->se
.nr_wakeups_affine
= 0;
2532 p
->se
.nr_wakeups_affine_attempts
= 0;
2533 p
->se
.nr_wakeups_passive
= 0;
2534 p
->se
.nr_wakeups_idle
= 0;
2538 INIT_LIST_HEAD(&p
->rt
.run_list
);
2540 INIT_LIST_HEAD(&p
->se
.group_node
);
2542 #ifdef CONFIG_PREEMPT_NOTIFIERS
2543 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2547 * We mark the process as running here, but have not actually
2548 * inserted it onto the runqueue yet. This guarantees that
2549 * nobody will actually run it, and a signal or other external
2550 * event cannot wake it up and insert it on the runqueue either.
2552 p
->state
= TASK_RUNNING
;
2556 * fork()/clone()-time setup:
2558 void sched_fork(struct task_struct
*p
, int clone_flags
)
2560 int cpu
= get_cpu();
2565 * Revert to default priority/policy on fork if requested.
2567 if (unlikely(p
->sched_reset_on_fork
)) {
2568 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2569 p
->policy
= SCHED_NORMAL
;
2570 p
->normal_prio
= p
->static_prio
;
2573 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2574 p
->static_prio
= NICE_TO_PRIO(0);
2575 p
->normal_prio
= p
->static_prio
;
2580 * We don't need the reset flag anymore after the fork. It has
2581 * fulfilled its duty:
2583 p
->sched_reset_on_fork
= 0;
2587 * Make sure we do not leak PI boosting priority to the child.
2589 p
->prio
= current
->normal_prio
;
2591 if (!rt_prio(p
->prio
))
2592 p
->sched_class
= &fair_sched_class
;
2595 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2597 set_task_cpu(p
, cpu
);
2599 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2600 if (likely(sched_info_on()))
2601 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2603 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2606 #ifdef CONFIG_PREEMPT
2607 /* Want to start with kernel preemption disabled. */
2608 task_thread_info(p
)->preempt_count
= 1;
2610 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2616 * wake_up_new_task - wake up a newly created task for the first time.
2618 * This function will do some initial scheduler statistics housekeeping
2619 * that must be done for every newly created context, then puts the task
2620 * on the runqueue and wakes it.
2622 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2624 unsigned long flags
;
2627 rq
= task_rq_lock(p
, &flags
);
2628 BUG_ON(p
->state
!= TASK_RUNNING
);
2629 update_rq_clock(rq
);
2631 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2632 activate_task(rq
, p
, 0);
2635 * Let the scheduling class do new task startup
2636 * management (if any):
2638 p
->sched_class
->task_new(rq
, p
);
2641 trace_sched_wakeup_new(rq
, p
, 1);
2642 check_preempt_curr(rq
, p
, WF_FORK
);
2644 if (p
->sched_class
->task_wake_up
)
2645 p
->sched_class
->task_wake_up(rq
, p
);
2647 task_rq_unlock(rq
, &flags
);
2650 #ifdef CONFIG_PREEMPT_NOTIFIERS
2653 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2654 * @notifier: notifier struct to register
2656 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2658 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2660 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2663 * preempt_notifier_unregister - no longer interested in preemption notifications
2664 * @notifier: notifier struct to unregister
2666 * This is safe to call from within a preemption notifier.
2668 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2670 hlist_del(¬ifier
->link
);
2672 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2674 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2676 struct preempt_notifier
*notifier
;
2677 struct hlist_node
*node
;
2679 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2680 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2684 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2685 struct task_struct
*next
)
2687 struct preempt_notifier
*notifier
;
2688 struct hlist_node
*node
;
2690 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2691 notifier
->ops
->sched_out(notifier
, next
);
2694 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2696 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2701 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2702 struct task_struct
*next
)
2706 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2709 * prepare_task_switch - prepare to switch tasks
2710 * @rq: the runqueue preparing to switch
2711 * @prev: the current task that is being switched out
2712 * @next: the task we are going to switch to.
2714 * This is called with the rq lock held and interrupts off. It must
2715 * be paired with a subsequent finish_task_switch after the context
2718 * prepare_task_switch sets up locking and calls architecture specific
2722 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2723 struct task_struct
*next
)
2725 fire_sched_out_preempt_notifiers(prev
, next
);
2726 prepare_lock_switch(rq
, next
);
2727 prepare_arch_switch(next
);
2731 * finish_task_switch - clean up after a task-switch
2732 * @rq: runqueue associated with task-switch
2733 * @prev: the thread we just switched away from.
2735 * finish_task_switch must be called after the context switch, paired
2736 * with a prepare_task_switch call before the context switch.
2737 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2738 * and do any other architecture-specific cleanup actions.
2740 * Note that we may have delayed dropping an mm in context_switch(). If
2741 * so, we finish that here outside of the runqueue lock. (Doing it
2742 * with the lock held can cause deadlocks; see schedule() for
2745 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2746 __releases(rq
->lock
)
2748 struct mm_struct
*mm
= rq
->prev_mm
;
2754 * A task struct has one reference for the use as "current".
2755 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2756 * schedule one last time. The schedule call will never return, and
2757 * the scheduled task must drop that reference.
2758 * The test for TASK_DEAD must occur while the runqueue locks are
2759 * still held, otherwise prev could be scheduled on another cpu, die
2760 * there before we look at prev->state, and then the reference would
2762 * Manfred Spraul <manfred@colorfullife.com>
2764 prev_state
= prev
->state
;
2765 finish_arch_switch(prev
);
2766 perf_event_task_sched_in(current
, cpu_of(rq
));
2767 finish_lock_switch(rq
, prev
);
2769 fire_sched_in_preempt_notifiers(current
);
2772 if (unlikely(prev_state
== TASK_DEAD
)) {
2774 * Remove function-return probe instances associated with this
2775 * task and put them back on the free list.
2777 kprobe_flush_task(prev
);
2778 put_task_struct(prev
);
2784 /* assumes rq->lock is held */
2785 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2787 if (prev
->sched_class
->pre_schedule
)
2788 prev
->sched_class
->pre_schedule(rq
, prev
);
2791 /* rq->lock is NOT held, but preemption is disabled */
2792 static inline void post_schedule(struct rq
*rq
)
2794 if (rq
->post_schedule
) {
2795 unsigned long flags
;
2797 spin_lock_irqsave(&rq
->lock
, flags
);
2798 if (rq
->curr
->sched_class
->post_schedule
)
2799 rq
->curr
->sched_class
->post_schedule(rq
);
2800 spin_unlock_irqrestore(&rq
->lock
, flags
);
2802 rq
->post_schedule
= 0;
2808 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2812 static inline void post_schedule(struct rq
*rq
)
2819 * schedule_tail - first thing a freshly forked thread must call.
2820 * @prev: the thread we just switched away from.
2822 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2823 __releases(rq
->lock
)
2825 struct rq
*rq
= this_rq();
2827 finish_task_switch(rq
, prev
);
2830 * FIXME: do we need to worry about rq being invalidated by the
2835 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2836 /* In this case, finish_task_switch does not reenable preemption */
2839 if (current
->set_child_tid
)
2840 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2844 * context_switch - switch to the new MM and the new
2845 * thread's register state.
2848 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2849 struct task_struct
*next
)
2851 struct mm_struct
*mm
, *oldmm
;
2853 prepare_task_switch(rq
, prev
, next
);
2854 trace_sched_switch(rq
, prev
, next
);
2856 oldmm
= prev
->active_mm
;
2858 * For paravirt, this is coupled with an exit in switch_to to
2859 * combine the page table reload and the switch backend into
2862 arch_start_context_switch(prev
);
2864 if (unlikely(!mm
)) {
2865 next
->active_mm
= oldmm
;
2866 atomic_inc(&oldmm
->mm_count
);
2867 enter_lazy_tlb(oldmm
, next
);
2869 switch_mm(oldmm
, mm
, next
);
2871 if (unlikely(!prev
->mm
)) {
2872 prev
->active_mm
= NULL
;
2873 rq
->prev_mm
= oldmm
;
2876 * Since the runqueue lock will be released by the next
2877 * task (which is an invalid locking op but in the case
2878 * of the scheduler it's an obvious special-case), so we
2879 * do an early lockdep release here:
2881 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2882 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2885 /* Here we just switch the register state and the stack. */
2886 switch_to(prev
, next
, prev
);
2890 * this_rq must be evaluated again because prev may have moved
2891 * CPUs since it called schedule(), thus the 'rq' on its stack
2892 * frame will be invalid.
2894 finish_task_switch(this_rq(), prev
);
2898 * nr_running, nr_uninterruptible and nr_context_switches:
2900 * externally visible scheduler statistics: current number of runnable
2901 * threads, current number of uninterruptible-sleeping threads, total
2902 * number of context switches performed since bootup.
2904 unsigned long nr_running(void)
2906 unsigned long i
, sum
= 0;
2908 for_each_online_cpu(i
)
2909 sum
+= cpu_rq(i
)->nr_running
;
2914 unsigned long nr_uninterruptible(void)
2916 unsigned long i
, sum
= 0;
2918 for_each_possible_cpu(i
)
2919 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2922 * Since we read the counters lockless, it might be slightly
2923 * inaccurate. Do not allow it to go below zero though:
2925 if (unlikely((long)sum
< 0))
2931 unsigned long long nr_context_switches(void)
2934 unsigned long long sum
= 0;
2936 for_each_possible_cpu(i
)
2937 sum
+= cpu_rq(i
)->nr_switches
;
2942 unsigned long nr_iowait(void)
2944 unsigned long i
, sum
= 0;
2946 for_each_possible_cpu(i
)
2947 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2952 unsigned long nr_iowait_cpu(void)
2954 struct rq
*this = this_rq();
2955 return atomic_read(&this->nr_iowait
);
2958 unsigned long this_cpu_load(void)
2960 struct rq
*this = this_rq();
2961 return this->cpu_load
[0];
2965 /* Variables and functions for calc_load */
2966 static atomic_long_t calc_load_tasks
;
2967 static unsigned long calc_load_update
;
2968 unsigned long avenrun
[3];
2969 EXPORT_SYMBOL(avenrun
);
2972 * get_avenrun - get the load average array
2973 * @loads: pointer to dest load array
2974 * @offset: offset to add
2975 * @shift: shift count to shift the result left
2977 * These values are estimates at best, so no need for locking.
2979 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2981 loads
[0] = (avenrun
[0] + offset
) << shift
;
2982 loads
[1] = (avenrun
[1] + offset
) << shift
;
2983 loads
[2] = (avenrun
[2] + offset
) << shift
;
2986 static unsigned long
2987 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2990 load
+= active
* (FIXED_1
- exp
);
2991 return load
>> FSHIFT
;
2995 * calc_load - update the avenrun load estimates 10 ticks after the
2996 * CPUs have updated calc_load_tasks.
2998 void calc_global_load(void)
3000 unsigned long upd
= calc_load_update
+ 10;
3003 if (time_before(jiffies
, upd
))
3006 active
= atomic_long_read(&calc_load_tasks
);
3007 active
= active
> 0 ? active
* FIXED_1
: 0;
3009 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3010 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3011 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3013 calc_load_update
+= LOAD_FREQ
;
3017 * Either called from update_cpu_load() or from a cpu going idle
3019 static void calc_load_account_active(struct rq
*this_rq
)
3021 long nr_active
, delta
;
3023 nr_active
= this_rq
->nr_running
;
3024 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3026 if (nr_active
!= this_rq
->calc_load_active
) {
3027 delta
= nr_active
- this_rq
->calc_load_active
;
3028 this_rq
->calc_load_active
= nr_active
;
3029 atomic_long_add(delta
, &calc_load_tasks
);
3034 * Externally visible per-cpu scheduler statistics:
3035 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3037 u64
cpu_nr_migrations(int cpu
)
3039 return cpu_rq(cpu
)->nr_migrations_in
;
3043 * Update rq->cpu_load[] statistics. This function is usually called every
3044 * scheduler tick (TICK_NSEC).
3046 static void update_cpu_load(struct rq
*this_rq
)
3048 unsigned long this_load
= this_rq
->load
.weight
;
3051 this_rq
->nr_load_updates
++;
3053 /* Update our load: */
3054 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3055 unsigned long old_load
, new_load
;
3057 /* scale is effectively 1 << i now, and >> i divides by scale */
3059 old_load
= this_rq
->cpu_load
[i
];
3060 new_load
= this_load
;
3062 * Round up the averaging division if load is increasing. This
3063 * prevents us from getting stuck on 9 if the load is 10, for
3066 if (new_load
> old_load
)
3067 new_load
+= scale
-1;
3068 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3071 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3072 this_rq
->calc_load_update
+= LOAD_FREQ
;
3073 calc_load_account_active(this_rq
);
3080 * double_rq_lock - safely lock two runqueues
3082 * Note this does not disable interrupts like task_rq_lock,
3083 * you need to do so manually before calling.
3085 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3086 __acquires(rq1
->lock
)
3087 __acquires(rq2
->lock
)
3089 BUG_ON(!irqs_disabled());
3091 spin_lock(&rq1
->lock
);
3092 __acquire(rq2
->lock
); /* Fake it out ;) */
3095 spin_lock(&rq1
->lock
);
3096 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3098 spin_lock(&rq2
->lock
);
3099 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3102 update_rq_clock(rq1
);
3103 update_rq_clock(rq2
);
3107 * double_rq_unlock - safely unlock two runqueues
3109 * Note this does not restore interrupts like task_rq_unlock,
3110 * you need to do so manually after calling.
3112 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3113 __releases(rq1
->lock
)
3114 __releases(rq2
->lock
)
3116 spin_unlock(&rq1
->lock
);
3118 spin_unlock(&rq2
->lock
);
3120 __release(rq2
->lock
);
3124 * If dest_cpu is allowed for this process, migrate the task to it.
3125 * This is accomplished by forcing the cpu_allowed mask to only
3126 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3127 * the cpu_allowed mask is restored.
3129 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3131 struct migration_req req
;
3132 unsigned long flags
;
3135 rq
= task_rq_lock(p
, &flags
);
3136 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3137 || unlikely(!cpu_active(dest_cpu
)))
3140 /* force the process onto the specified CPU */
3141 if (migrate_task(p
, dest_cpu
, &req
)) {
3142 /* Need to wait for migration thread (might exit: take ref). */
3143 struct task_struct
*mt
= rq
->migration_thread
;
3145 get_task_struct(mt
);
3146 task_rq_unlock(rq
, &flags
);
3147 wake_up_process(mt
);
3148 put_task_struct(mt
);
3149 wait_for_completion(&req
.done
);
3154 task_rq_unlock(rq
, &flags
);
3158 * sched_exec - execve() is a valuable balancing opportunity, because at
3159 * this point the task has the smallest effective memory and cache footprint.
3161 void sched_exec(void)
3163 int new_cpu
, this_cpu
= get_cpu();
3164 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3166 if (new_cpu
!= this_cpu
)
3167 sched_migrate_task(current
, new_cpu
);
3171 * pull_task - move a task from a remote runqueue to the local runqueue.
3172 * Both runqueues must be locked.
3174 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3175 struct rq
*this_rq
, int this_cpu
)
3177 deactivate_task(src_rq
, p
, 0);
3178 set_task_cpu(p
, this_cpu
);
3179 activate_task(this_rq
, p
, 0);
3181 * Note that idle threads have a prio of MAX_PRIO, for this test
3182 * to be always true for them.
3184 check_preempt_curr(this_rq
, p
, 0);
3188 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3191 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3192 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3195 int tsk_cache_hot
= 0;
3197 * We do not migrate tasks that are:
3198 * 1) running (obviously), or
3199 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3200 * 3) are cache-hot on their current CPU.
3202 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3203 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3208 if (task_running(rq
, p
)) {
3209 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3214 * Aggressive migration if:
3215 * 1) task is cache cold, or
3216 * 2) too many balance attempts have failed.
3219 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3220 if (!tsk_cache_hot
||
3221 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3222 #ifdef CONFIG_SCHEDSTATS
3223 if (tsk_cache_hot
) {
3224 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3225 schedstat_inc(p
, se
.nr_forced_migrations
);
3231 if (tsk_cache_hot
) {
3232 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3238 static unsigned long
3239 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3240 unsigned long max_load_move
, struct sched_domain
*sd
,
3241 enum cpu_idle_type idle
, int *all_pinned
,
3242 int *this_best_prio
, struct rq_iterator
*iterator
)
3244 int loops
= 0, pulled
= 0, pinned
= 0;
3245 struct task_struct
*p
;
3246 long rem_load_move
= max_load_move
;
3248 if (max_load_move
== 0)
3254 * Start the load-balancing iterator:
3256 p
= iterator
->start(iterator
->arg
);
3258 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3261 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3262 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3263 p
= iterator
->next(iterator
->arg
);
3267 pull_task(busiest
, p
, this_rq
, this_cpu
);
3269 rem_load_move
-= p
->se
.load
.weight
;
3271 #ifdef CONFIG_PREEMPT
3273 * NEWIDLE balancing is a source of latency, so preemptible kernels
3274 * will stop after the first task is pulled to minimize the critical
3277 if (idle
== CPU_NEWLY_IDLE
)
3282 * We only want to steal up to the prescribed amount of weighted load.
3284 if (rem_load_move
> 0) {
3285 if (p
->prio
< *this_best_prio
)
3286 *this_best_prio
= p
->prio
;
3287 p
= iterator
->next(iterator
->arg
);
3292 * Right now, this is one of only two places pull_task() is called,
3293 * so we can safely collect pull_task() stats here rather than
3294 * inside pull_task().
3296 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3299 *all_pinned
= pinned
;
3301 return max_load_move
- rem_load_move
;
3305 * move_tasks tries to move up to max_load_move weighted load from busiest to
3306 * this_rq, as part of a balancing operation within domain "sd".
3307 * Returns 1 if successful and 0 otherwise.
3309 * Called with both runqueues locked.
3311 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3312 unsigned long max_load_move
,
3313 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3316 const struct sched_class
*class = sched_class_highest
;
3317 unsigned long total_load_moved
= 0;
3318 int this_best_prio
= this_rq
->curr
->prio
;
3322 class->load_balance(this_rq
, this_cpu
, busiest
,
3323 max_load_move
- total_load_moved
,
3324 sd
, idle
, all_pinned
, &this_best_prio
);
3325 class = class->next
;
3327 #ifdef CONFIG_PREEMPT
3329 * NEWIDLE balancing is a source of latency, so preemptible
3330 * kernels will stop after the first task is pulled to minimize
3331 * the critical section.
3333 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3336 } while (class && max_load_move
> total_load_moved
);
3338 return total_load_moved
> 0;
3342 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3343 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3344 struct rq_iterator
*iterator
)
3346 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3350 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3351 pull_task(busiest
, p
, this_rq
, this_cpu
);
3353 * Right now, this is only the second place pull_task()
3354 * is called, so we can safely collect pull_task()
3355 * stats here rather than inside pull_task().
3357 schedstat_inc(sd
, lb_gained
[idle
]);
3361 p
= iterator
->next(iterator
->arg
);
3368 * move_one_task tries to move exactly one task from busiest to this_rq, as
3369 * part of active balancing operations within "domain".
3370 * Returns 1 if successful and 0 otherwise.
3372 * Called with both runqueues locked.
3374 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3375 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3377 const struct sched_class
*class;
3379 for_each_class(class) {
3380 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3386 /********** Helpers for find_busiest_group ************************/
3388 * sd_lb_stats - Structure to store the statistics of a sched_domain
3389 * during load balancing.
3391 struct sd_lb_stats
{
3392 struct sched_group
*busiest
; /* Busiest group in this sd */
3393 struct sched_group
*this; /* Local group in this sd */
3394 unsigned long total_load
; /* Total load of all groups in sd */
3395 unsigned long total_pwr
; /* Total power of all groups in sd */
3396 unsigned long avg_load
; /* Average load across all groups in sd */
3398 /** Statistics of this group */
3399 unsigned long this_load
;
3400 unsigned long this_load_per_task
;
3401 unsigned long this_nr_running
;
3403 /* Statistics of the busiest group */
3404 unsigned long max_load
;
3405 unsigned long busiest_load_per_task
;
3406 unsigned long busiest_nr_running
;
3408 int group_imb
; /* Is there imbalance in this sd */
3409 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3410 int power_savings_balance
; /* Is powersave balance needed for this sd */
3411 struct sched_group
*group_min
; /* Least loaded group in sd */
3412 struct sched_group
*group_leader
; /* Group which relieves group_min */
3413 unsigned long min_load_per_task
; /* load_per_task in group_min */
3414 unsigned long leader_nr_running
; /* Nr running of group_leader */
3415 unsigned long min_nr_running
; /* Nr running of group_min */
3420 * sg_lb_stats - stats of a sched_group required for load_balancing
3422 struct sg_lb_stats
{
3423 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3424 unsigned long group_load
; /* Total load over the CPUs of the group */
3425 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3426 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3427 unsigned long group_capacity
;
3428 int group_imb
; /* Is there an imbalance in the group ? */
3432 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3433 * @group: The group whose first cpu is to be returned.
3435 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3437 return cpumask_first(sched_group_cpus(group
));
3441 * get_sd_load_idx - Obtain the load index for a given sched domain.
3442 * @sd: The sched_domain whose load_idx is to be obtained.
3443 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3445 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3446 enum cpu_idle_type idle
)
3452 load_idx
= sd
->busy_idx
;
3455 case CPU_NEWLY_IDLE
:
3456 load_idx
= sd
->newidle_idx
;
3459 load_idx
= sd
->idle_idx
;
3467 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3469 * init_sd_power_savings_stats - Initialize power savings statistics for
3470 * the given sched_domain, during load balancing.
3472 * @sd: Sched domain whose power-savings statistics are to be initialized.
3473 * @sds: Variable containing the statistics for sd.
3474 * @idle: Idle status of the CPU at which we're performing load-balancing.
3476 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3477 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3480 * Busy processors will not participate in power savings
3483 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3484 sds
->power_savings_balance
= 0;
3486 sds
->power_savings_balance
= 1;
3487 sds
->min_nr_running
= ULONG_MAX
;
3488 sds
->leader_nr_running
= 0;
3493 * update_sd_power_savings_stats - Update the power saving stats for a
3494 * sched_domain while performing load balancing.
3496 * @group: sched_group belonging to the sched_domain under consideration.
3497 * @sds: Variable containing the statistics of the sched_domain
3498 * @local_group: Does group contain the CPU for which we're performing
3500 * @sgs: Variable containing the statistics of the group.
3502 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3503 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3506 if (!sds
->power_savings_balance
)
3510 * If the local group is idle or completely loaded
3511 * no need to do power savings balance at this domain
3513 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3514 !sds
->this_nr_running
))
3515 sds
->power_savings_balance
= 0;
3518 * If a group is already running at full capacity or idle,
3519 * don't include that group in power savings calculations
3521 if (!sds
->power_savings_balance
||
3522 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3523 !sgs
->sum_nr_running
)
3527 * Calculate the group which has the least non-idle load.
3528 * This is the group from where we need to pick up the load
3531 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3532 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3533 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3534 sds
->group_min
= group
;
3535 sds
->min_nr_running
= sgs
->sum_nr_running
;
3536 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3537 sgs
->sum_nr_running
;
3541 * Calculate the group which is almost near its
3542 * capacity but still has some space to pick up some load
3543 * from other group and save more power
3545 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3548 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3549 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3550 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3551 sds
->group_leader
= group
;
3552 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3557 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3558 * @sds: Variable containing the statistics of the sched_domain
3559 * under consideration.
3560 * @this_cpu: Cpu at which we're currently performing load-balancing.
3561 * @imbalance: Variable to store the imbalance.
3564 * Check if we have potential to perform some power-savings balance.
3565 * If yes, set the busiest group to be the least loaded group in the
3566 * sched_domain, so that it's CPUs can be put to idle.
3568 * Returns 1 if there is potential to perform power-savings balance.
3571 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3572 int this_cpu
, unsigned long *imbalance
)
3574 if (!sds
->power_savings_balance
)
3577 if (sds
->this != sds
->group_leader
||
3578 sds
->group_leader
== sds
->group_min
)
3581 *imbalance
= sds
->min_load_per_task
;
3582 sds
->busiest
= sds
->group_min
;
3587 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3588 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3589 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3594 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3595 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3600 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3601 int this_cpu
, unsigned long *imbalance
)
3605 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3608 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3610 return SCHED_LOAD_SCALE
;
3613 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3615 return default_scale_freq_power(sd
, cpu
);
3618 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3620 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3621 unsigned long smt_gain
= sd
->smt_gain
;
3628 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3630 return default_scale_smt_power(sd
, cpu
);
3633 unsigned long scale_rt_power(int cpu
)
3635 struct rq
*rq
= cpu_rq(cpu
);
3636 u64 total
, available
;
3638 sched_avg_update(rq
);
3640 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3641 available
= total
- rq
->rt_avg
;
3643 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3644 total
= SCHED_LOAD_SCALE
;
3646 total
>>= SCHED_LOAD_SHIFT
;
3648 return div_u64(available
, total
);
3651 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3653 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3654 unsigned long power
= SCHED_LOAD_SCALE
;
3655 struct sched_group
*sdg
= sd
->groups
;
3657 if (sched_feat(ARCH_POWER
))
3658 power
*= arch_scale_freq_power(sd
, cpu
);
3660 power
*= default_scale_freq_power(sd
, cpu
);
3662 power
>>= SCHED_LOAD_SHIFT
;
3664 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3665 if (sched_feat(ARCH_POWER
))
3666 power
*= arch_scale_smt_power(sd
, cpu
);
3668 power
*= default_scale_smt_power(sd
, cpu
);
3670 power
>>= SCHED_LOAD_SHIFT
;
3673 power
*= scale_rt_power(cpu
);
3674 power
>>= SCHED_LOAD_SHIFT
;
3679 sdg
->cpu_power
= power
;
3682 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3684 struct sched_domain
*child
= sd
->child
;
3685 struct sched_group
*group
, *sdg
= sd
->groups
;
3686 unsigned long power
;
3689 update_cpu_power(sd
, cpu
);
3695 group
= child
->groups
;
3697 power
+= group
->cpu_power
;
3698 group
= group
->next
;
3699 } while (group
!= child
->groups
);
3701 sdg
->cpu_power
= power
;
3705 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3706 * @sd: The sched_domain whose statistics are to be updated.
3707 * @group: sched_group whose statistics are to be updated.
3708 * @this_cpu: Cpu for which load balance is currently performed.
3709 * @idle: Idle status of this_cpu
3710 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3711 * @sd_idle: Idle status of the sched_domain containing group.
3712 * @local_group: Does group contain this_cpu.
3713 * @cpus: Set of cpus considered for load balancing.
3714 * @balance: Should we balance.
3715 * @sgs: variable to hold the statistics for this group.
3717 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3718 struct sched_group
*group
, int this_cpu
,
3719 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3720 int local_group
, const struct cpumask
*cpus
,
3721 int *balance
, struct sg_lb_stats
*sgs
)
3723 unsigned long load
, max_cpu_load
, min_cpu_load
;
3725 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3726 unsigned long sum_avg_load_per_task
;
3727 unsigned long avg_load_per_task
;
3730 balance_cpu
= group_first_cpu(group
);
3731 if (balance_cpu
== this_cpu
)
3732 update_group_power(sd
, this_cpu
);
3735 /* Tally up the load of all CPUs in the group */
3736 sum_avg_load_per_task
= avg_load_per_task
= 0;
3738 min_cpu_load
= ~0UL;
3740 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3741 struct rq
*rq
= cpu_rq(i
);
3743 if (*sd_idle
&& rq
->nr_running
)
3746 /* Bias balancing toward cpus of our domain */
3748 if (idle_cpu(i
) && !first_idle_cpu
) {
3753 load
= target_load(i
, load_idx
);
3755 load
= source_load(i
, load_idx
);
3756 if (load
> max_cpu_load
)
3757 max_cpu_load
= load
;
3758 if (min_cpu_load
> load
)
3759 min_cpu_load
= load
;
3762 sgs
->group_load
+= load
;
3763 sgs
->sum_nr_running
+= rq
->nr_running
;
3764 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3766 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3770 * First idle cpu or the first cpu(busiest) in this sched group
3771 * is eligible for doing load balancing at this and above
3772 * domains. In the newly idle case, we will allow all the cpu's
3773 * to do the newly idle load balance.
3775 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3776 balance_cpu
!= this_cpu
&& balance
) {
3781 /* Adjust by relative CPU power of the group */
3782 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3786 * Consider the group unbalanced when the imbalance is larger
3787 * than the average weight of two tasks.
3789 * APZ: with cgroup the avg task weight can vary wildly and
3790 * might not be a suitable number - should we keep a
3791 * normalized nr_running number somewhere that negates
3794 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3797 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3800 sgs
->group_capacity
=
3801 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3805 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3806 * @sd: sched_domain whose statistics are to be updated.
3807 * @this_cpu: Cpu for which load balance is currently performed.
3808 * @idle: Idle status of this_cpu
3809 * @sd_idle: Idle status of the sched_domain containing group.
3810 * @cpus: Set of cpus considered for load balancing.
3811 * @balance: Should we balance.
3812 * @sds: variable to hold the statistics for this sched_domain.
3814 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3815 enum cpu_idle_type idle
, int *sd_idle
,
3816 const struct cpumask
*cpus
, int *balance
,
3817 struct sd_lb_stats
*sds
)
3819 struct sched_domain
*child
= sd
->child
;
3820 struct sched_group
*group
= sd
->groups
;
3821 struct sg_lb_stats sgs
;
3822 int load_idx
, prefer_sibling
= 0;
3824 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3827 init_sd_power_savings_stats(sd
, sds
, idle
);
3828 load_idx
= get_sd_load_idx(sd
, idle
);
3833 local_group
= cpumask_test_cpu(this_cpu
,
3834 sched_group_cpus(group
));
3835 memset(&sgs
, 0, sizeof(sgs
));
3836 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3837 local_group
, cpus
, balance
, &sgs
);
3839 if (local_group
&& balance
&& !(*balance
))
3842 sds
->total_load
+= sgs
.group_load
;
3843 sds
->total_pwr
+= group
->cpu_power
;
3846 * In case the child domain prefers tasks go to siblings
3847 * first, lower the group capacity to one so that we'll try
3848 * and move all the excess tasks away.
3851 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3854 sds
->this_load
= sgs
.avg_load
;
3856 sds
->this_nr_running
= sgs
.sum_nr_running
;
3857 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3858 } else if (sgs
.avg_load
> sds
->max_load
&&
3859 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3861 sds
->max_load
= sgs
.avg_load
;
3862 sds
->busiest
= group
;
3863 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3864 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3865 sds
->group_imb
= sgs
.group_imb
;
3868 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3869 group
= group
->next
;
3870 } while (group
!= sd
->groups
);
3874 * fix_small_imbalance - Calculate the minor imbalance that exists
3875 * amongst the groups of a sched_domain, during
3877 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3878 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3879 * @imbalance: Variable to store the imbalance.
3881 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3882 int this_cpu
, unsigned long *imbalance
)
3884 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3885 unsigned int imbn
= 2;
3887 if (sds
->this_nr_running
) {
3888 sds
->this_load_per_task
/= sds
->this_nr_running
;
3889 if (sds
->busiest_load_per_task
>
3890 sds
->this_load_per_task
)
3893 sds
->this_load_per_task
=
3894 cpu_avg_load_per_task(this_cpu
);
3896 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3897 sds
->busiest_load_per_task
* imbn
) {
3898 *imbalance
= sds
->busiest_load_per_task
;
3903 * OK, we don't have enough imbalance to justify moving tasks,
3904 * however we may be able to increase total CPU power used by
3908 pwr_now
+= sds
->busiest
->cpu_power
*
3909 min(sds
->busiest_load_per_task
, sds
->max_load
);
3910 pwr_now
+= sds
->this->cpu_power
*
3911 min(sds
->this_load_per_task
, sds
->this_load
);
3912 pwr_now
/= SCHED_LOAD_SCALE
;
3914 /* Amount of load we'd subtract */
3915 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3916 sds
->busiest
->cpu_power
;
3917 if (sds
->max_load
> tmp
)
3918 pwr_move
+= sds
->busiest
->cpu_power
*
3919 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3921 /* Amount of load we'd add */
3922 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3923 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3924 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3925 sds
->this->cpu_power
;
3927 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3928 sds
->this->cpu_power
;
3929 pwr_move
+= sds
->this->cpu_power
*
3930 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3931 pwr_move
/= SCHED_LOAD_SCALE
;
3933 /* Move if we gain throughput */
3934 if (pwr_move
> pwr_now
)
3935 *imbalance
= sds
->busiest_load_per_task
;
3939 * calculate_imbalance - Calculate the amount of imbalance present within the
3940 * groups of a given sched_domain during load balance.
3941 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3942 * @this_cpu: Cpu for which currently load balance is being performed.
3943 * @imbalance: The variable to store the imbalance.
3945 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3946 unsigned long *imbalance
)
3948 unsigned long max_pull
;
3950 * In the presence of smp nice balancing, certain scenarios can have
3951 * max load less than avg load(as we skip the groups at or below
3952 * its cpu_power, while calculating max_load..)
3954 if (sds
->max_load
< sds
->avg_load
) {
3956 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3959 /* Don't want to pull so many tasks that a group would go idle */
3960 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3961 sds
->max_load
- sds
->busiest_load_per_task
);
3963 /* How much load to actually move to equalise the imbalance */
3964 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3965 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3969 * if *imbalance is less than the average load per runnable task
3970 * there is no gaurantee that any tasks will be moved so we'll have
3971 * a think about bumping its value to force at least one task to be
3974 if (*imbalance
< sds
->busiest_load_per_task
)
3975 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3978 /******* find_busiest_group() helpers end here *********************/
3981 * find_busiest_group - Returns the busiest group within the sched_domain
3982 * if there is an imbalance. If there isn't an imbalance, and
3983 * the user has opted for power-savings, it returns a group whose
3984 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3985 * such a group exists.
3987 * Also calculates the amount of weighted load which should be moved
3988 * to restore balance.
3990 * @sd: The sched_domain whose busiest group is to be returned.
3991 * @this_cpu: The cpu for which load balancing is currently being performed.
3992 * @imbalance: Variable which stores amount of weighted load which should
3993 * be moved to restore balance/put a group to idle.
3994 * @idle: The idle status of this_cpu.
3995 * @sd_idle: The idleness of sd
3996 * @cpus: The set of CPUs under consideration for load-balancing.
3997 * @balance: Pointer to a variable indicating if this_cpu
3998 * is the appropriate cpu to perform load balancing at this_level.
4000 * Returns: - the busiest group if imbalance exists.
4001 * - If no imbalance and user has opted for power-savings balance,
4002 * return the least loaded group whose CPUs can be
4003 * put to idle by rebalancing its tasks onto our group.
4005 static struct sched_group
*
4006 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4007 unsigned long *imbalance
, enum cpu_idle_type idle
,
4008 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4010 struct sd_lb_stats sds
;
4012 memset(&sds
, 0, sizeof(sds
));
4015 * Compute the various statistics relavent for load balancing at
4018 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4021 /* Cases where imbalance does not exist from POV of this_cpu */
4022 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4024 * 2) There is no busy sibling group to pull from.
4025 * 3) This group is the busiest group.
4026 * 4) This group is more busy than the avg busieness at this
4028 * 5) The imbalance is within the specified limit.
4029 * 6) Any rebalance would lead to ping-pong
4031 if (balance
&& !(*balance
))
4034 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4037 if (sds
.this_load
>= sds
.max_load
)
4040 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4042 if (sds
.this_load
>= sds
.avg_load
)
4045 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4048 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4050 sds
.busiest_load_per_task
=
4051 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4054 * We're trying to get all the cpus to the average_load, so we don't
4055 * want to push ourselves above the average load, nor do we wish to
4056 * reduce the max loaded cpu below the average load, as either of these
4057 * actions would just result in more rebalancing later, and ping-pong
4058 * tasks around. Thus we look for the minimum possible imbalance.
4059 * Negative imbalances (*we* are more loaded than anyone else) will
4060 * be counted as no imbalance for these purposes -- we can't fix that
4061 * by pulling tasks to us. Be careful of negative numbers as they'll
4062 * appear as very large values with unsigned longs.
4064 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4067 /* Looks like there is an imbalance. Compute it */
4068 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4073 * There is no obvious imbalance. But check if we can do some balancing
4076 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4084 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4087 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4088 unsigned long imbalance
, const struct cpumask
*cpus
)
4090 struct rq
*busiest
= NULL
, *rq
;
4091 unsigned long max_load
= 0;
4094 for_each_cpu(i
, sched_group_cpus(group
)) {
4095 unsigned long power
= power_of(i
);
4096 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4099 if (!cpumask_test_cpu(i
, cpus
))
4103 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4106 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4109 if (wl
> max_load
) {
4119 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4120 * so long as it is large enough.
4122 #define MAX_PINNED_INTERVAL 512
4124 /* Working cpumask for load_balance and load_balance_newidle. */
4125 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4128 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4129 * tasks if there is an imbalance.
4131 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4132 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4135 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4136 struct sched_group
*group
;
4137 unsigned long imbalance
;
4139 unsigned long flags
;
4140 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4142 cpumask_copy(cpus
, cpu_online_mask
);
4145 * When power savings policy is enabled for the parent domain, idle
4146 * sibling can pick up load irrespective of busy siblings. In this case,
4147 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4148 * portraying it as CPU_NOT_IDLE.
4150 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4151 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4154 schedstat_inc(sd
, lb_count
[idle
]);
4158 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4165 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4169 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4171 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4175 BUG_ON(busiest
== this_rq
);
4177 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4180 if (busiest
->nr_running
> 1) {
4182 * Attempt to move tasks. If find_busiest_group has found
4183 * an imbalance but busiest->nr_running <= 1, the group is
4184 * still unbalanced. ld_moved simply stays zero, so it is
4185 * correctly treated as an imbalance.
4187 local_irq_save(flags
);
4188 double_rq_lock(this_rq
, busiest
);
4189 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4190 imbalance
, sd
, idle
, &all_pinned
);
4191 double_rq_unlock(this_rq
, busiest
);
4192 local_irq_restore(flags
);
4195 * some other cpu did the load balance for us.
4197 if (ld_moved
&& this_cpu
!= smp_processor_id())
4198 resched_cpu(this_cpu
);
4200 /* All tasks on this runqueue were pinned by CPU affinity */
4201 if (unlikely(all_pinned
)) {
4202 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4203 if (!cpumask_empty(cpus
))
4210 schedstat_inc(sd
, lb_failed
[idle
]);
4211 sd
->nr_balance_failed
++;
4213 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4215 spin_lock_irqsave(&busiest
->lock
, flags
);
4217 /* don't kick the migration_thread, if the curr
4218 * task on busiest cpu can't be moved to this_cpu
4220 if (!cpumask_test_cpu(this_cpu
,
4221 &busiest
->curr
->cpus_allowed
)) {
4222 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4224 goto out_one_pinned
;
4227 if (!busiest
->active_balance
) {
4228 busiest
->active_balance
= 1;
4229 busiest
->push_cpu
= this_cpu
;
4232 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4234 wake_up_process(busiest
->migration_thread
);
4237 * We've kicked active balancing, reset the failure
4240 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4243 sd
->nr_balance_failed
= 0;
4245 if (likely(!active_balance
)) {
4246 /* We were unbalanced, so reset the balancing interval */
4247 sd
->balance_interval
= sd
->min_interval
;
4250 * If we've begun active balancing, start to back off. This
4251 * case may not be covered by the all_pinned logic if there
4252 * is only 1 task on the busy runqueue (because we don't call
4255 if (sd
->balance_interval
< sd
->max_interval
)
4256 sd
->balance_interval
*= 2;
4259 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4260 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4266 schedstat_inc(sd
, lb_balanced
[idle
]);
4268 sd
->nr_balance_failed
= 0;
4271 /* tune up the balancing interval */
4272 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4273 (sd
->balance_interval
< sd
->max_interval
))
4274 sd
->balance_interval
*= 2;
4276 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4277 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4288 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4289 * tasks if there is an imbalance.
4291 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4292 * this_rq is locked.
4295 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4297 struct sched_group
*group
;
4298 struct rq
*busiest
= NULL
;
4299 unsigned long imbalance
;
4303 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4305 cpumask_copy(cpus
, cpu_online_mask
);
4308 * When power savings policy is enabled for the parent domain, idle
4309 * sibling can pick up load irrespective of busy siblings. In this case,
4310 * let the state of idle sibling percolate up as IDLE, instead of
4311 * portraying it as CPU_NOT_IDLE.
4313 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4314 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4317 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4319 update_shares_locked(this_rq
, sd
);
4320 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4321 &sd_idle
, cpus
, NULL
);
4323 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4327 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4329 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4333 BUG_ON(busiest
== this_rq
);
4335 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4338 if (busiest
->nr_running
> 1) {
4339 /* Attempt to move tasks */
4340 double_lock_balance(this_rq
, busiest
);
4341 /* this_rq->clock is already updated */
4342 update_rq_clock(busiest
);
4343 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4344 imbalance
, sd
, CPU_NEWLY_IDLE
,
4346 double_unlock_balance(this_rq
, busiest
);
4348 if (unlikely(all_pinned
)) {
4349 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4350 if (!cpumask_empty(cpus
))
4356 int active_balance
= 0;
4358 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4359 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4360 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4363 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4366 if (sd
->nr_balance_failed
++ < 2)
4370 * The only task running in a non-idle cpu can be moved to this
4371 * cpu in an attempt to completely freeup the other CPU
4372 * package. The same method used to move task in load_balance()
4373 * have been extended for load_balance_newidle() to speedup
4374 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4376 * The package power saving logic comes from
4377 * find_busiest_group(). If there are no imbalance, then
4378 * f_b_g() will return NULL. However when sched_mc={1,2} then
4379 * f_b_g() will select a group from which a running task may be
4380 * pulled to this cpu in order to make the other package idle.
4381 * If there is no opportunity to make a package idle and if
4382 * there are no imbalance, then f_b_g() will return NULL and no
4383 * action will be taken in load_balance_newidle().
4385 * Under normal task pull operation due to imbalance, there
4386 * will be more than one task in the source run queue and
4387 * move_tasks() will succeed. ld_moved will be true and this
4388 * active balance code will not be triggered.
4391 /* Lock busiest in correct order while this_rq is held */
4392 double_lock_balance(this_rq
, busiest
);
4395 * don't kick the migration_thread, if the curr
4396 * task on busiest cpu can't be moved to this_cpu
4398 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4399 double_unlock_balance(this_rq
, busiest
);
4404 if (!busiest
->active_balance
) {
4405 busiest
->active_balance
= 1;
4406 busiest
->push_cpu
= this_cpu
;
4410 double_unlock_balance(this_rq
, busiest
);
4412 * Should not call ttwu while holding a rq->lock
4414 spin_unlock(&this_rq
->lock
);
4416 wake_up_process(busiest
->migration_thread
);
4417 spin_lock(&this_rq
->lock
);
4420 sd
->nr_balance_failed
= 0;
4422 update_shares_locked(this_rq
, sd
);
4426 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4427 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4428 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4430 sd
->nr_balance_failed
= 0;
4436 * idle_balance is called by schedule() if this_cpu is about to become
4437 * idle. Attempts to pull tasks from other CPUs.
4439 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4441 struct sched_domain
*sd
;
4442 int pulled_task
= 0;
4443 unsigned long next_balance
= jiffies
+ HZ
;
4445 this_rq
->idle_stamp
= this_rq
->clock
;
4447 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4450 for_each_domain(this_cpu
, sd
) {
4451 unsigned long interval
;
4453 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4456 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4457 /* If we've pulled tasks over stop searching: */
4458 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4461 interval
= msecs_to_jiffies(sd
->balance_interval
);
4462 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4463 next_balance
= sd
->last_balance
+ interval
;
4465 this_rq
->idle_stamp
= 0;
4469 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4471 * We are going idle. next_balance may be set based on
4472 * a busy processor. So reset next_balance.
4474 this_rq
->next_balance
= next_balance
;
4479 * active_load_balance is run by migration threads. It pushes running tasks
4480 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4481 * running on each physical CPU where possible, and avoids physical /
4482 * logical imbalances.
4484 * Called with busiest_rq locked.
4486 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4488 int target_cpu
= busiest_rq
->push_cpu
;
4489 struct sched_domain
*sd
;
4490 struct rq
*target_rq
;
4492 /* Is there any task to move? */
4493 if (busiest_rq
->nr_running
<= 1)
4496 target_rq
= cpu_rq(target_cpu
);
4499 * This condition is "impossible", if it occurs
4500 * we need to fix it. Originally reported by
4501 * Bjorn Helgaas on a 128-cpu setup.
4503 BUG_ON(busiest_rq
== target_rq
);
4505 /* move a task from busiest_rq to target_rq */
4506 double_lock_balance(busiest_rq
, target_rq
);
4507 update_rq_clock(busiest_rq
);
4508 update_rq_clock(target_rq
);
4510 /* Search for an sd spanning us and the target CPU. */
4511 for_each_domain(target_cpu
, sd
) {
4512 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4513 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4518 schedstat_inc(sd
, alb_count
);
4520 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4522 schedstat_inc(sd
, alb_pushed
);
4524 schedstat_inc(sd
, alb_failed
);
4526 double_unlock_balance(busiest_rq
, target_rq
);
4531 atomic_t load_balancer
;
4532 cpumask_var_t cpu_mask
;
4533 cpumask_var_t ilb_grp_nohz_mask
;
4534 } nohz ____cacheline_aligned
= {
4535 .load_balancer
= ATOMIC_INIT(-1),
4538 int get_nohz_load_balancer(void)
4540 return atomic_read(&nohz
.load_balancer
);
4543 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4545 * lowest_flag_domain - Return lowest sched_domain containing flag.
4546 * @cpu: The cpu whose lowest level of sched domain is to
4548 * @flag: The flag to check for the lowest sched_domain
4549 * for the given cpu.
4551 * Returns the lowest sched_domain of a cpu which contains the given flag.
4553 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4555 struct sched_domain
*sd
;
4557 for_each_domain(cpu
, sd
)
4558 if (sd
&& (sd
->flags
& flag
))
4565 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4566 * @cpu: The cpu whose domains we're iterating over.
4567 * @sd: variable holding the value of the power_savings_sd
4569 * @flag: The flag to filter the sched_domains to be iterated.
4571 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4572 * set, starting from the lowest sched_domain to the highest.
4574 #define for_each_flag_domain(cpu, sd, flag) \
4575 for (sd = lowest_flag_domain(cpu, flag); \
4576 (sd && (sd->flags & flag)); sd = sd->parent)
4579 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4580 * @ilb_group: group to be checked for semi-idleness
4582 * Returns: 1 if the group is semi-idle. 0 otherwise.
4584 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4585 * and atleast one non-idle CPU. This helper function checks if the given
4586 * sched_group is semi-idle or not.
4588 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4590 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4591 sched_group_cpus(ilb_group
));
4594 * A sched_group is semi-idle when it has atleast one busy cpu
4595 * and atleast one idle cpu.
4597 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4600 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4606 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4607 * @cpu: The cpu which is nominating a new idle_load_balancer.
4609 * Returns: Returns the id of the idle load balancer if it exists,
4610 * Else, returns >= nr_cpu_ids.
4612 * This algorithm picks the idle load balancer such that it belongs to a
4613 * semi-idle powersavings sched_domain. The idea is to try and avoid
4614 * completely idle packages/cores just for the purpose of idle load balancing
4615 * when there are other idle cpu's which are better suited for that job.
4617 static int find_new_ilb(int cpu
)
4619 struct sched_domain
*sd
;
4620 struct sched_group
*ilb_group
;
4623 * Have idle load balancer selection from semi-idle packages only
4624 * when power-aware load balancing is enabled
4626 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4630 * Optimize for the case when we have no idle CPUs or only one
4631 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4633 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4636 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4637 ilb_group
= sd
->groups
;
4640 if (is_semi_idle_group(ilb_group
))
4641 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4643 ilb_group
= ilb_group
->next
;
4645 } while (ilb_group
!= sd
->groups
);
4649 return cpumask_first(nohz
.cpu_mask
);
4651 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4652 static inline int find_new_ilb(int call_cpu
)
4654 return cpumask_first(nohz
.cpu_mask
);
4659 * This routine will try to nominate the ilb (idle load balancing)
4660 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4661 * load balancing on behalf of all those cpus. If all the cpus in the system
4662 * go into this tickless mode, then there will be no ilb owner (as there is
4663 * no need for one) and all the cpus will sleep till the next wakeup event
4666 * For the ilb owner, tick is not stopped. And this tick will be used
4667 * for idle load balancing. ilb owner will still be part of
4670 * While stopping the tick, this cpu will become the ilb owner if there
4671 * is no other owner. And will be the owner till that cpu becomes busy
4672 * or if all cpus in the system stop their ticks at which point
4673 * there is no need for ilb owner.
4675 * When the ilb owner becomes busy, it nominates another owner, during the
4676 * next busy scheduler_tick()
4678 int select_nohz_load_balancer(int stop_tick
)
4680 int cpu
= smp_processor_id();
4683 cpu_rq(cpu
)->in_nohz_recently
= 1;
4685 if (!cpu_active(cpu
)) {
4686 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4690 * If we are going offline and still the leader,
4693 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4699 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4701 /* time for ilb owner also to sleep */
4702 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4703 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4704 atomic_set(&nohz
.load_balancer
, -1);
4708 if (atomic_read(&nohz
.load_balancer
) == -1) {
4709 /* make me the ilb owner */
4710 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4712 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4715 if (!(sched_smt_power_savings
||
4716 sched_mc_power_savings
))
4719 * Check to see if there is a more power-efficient
4722 new_ilb
= find_new_ilb(cpu
);
4723 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4724 atomic_set(&nohz
.load_balancer
, -1);
4725 resched_cpu(new_ilb
);
4731 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4734 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4736 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4737 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4744 static DEFINE_SPINLOCK(balancing
);
4747 * It checks each scheduling domain to see if it is due to be balanced,
4748 * and initiates a balancing operation if so.
4750 * Balancing parameters are set up in arch_init_sched_domains.
4752 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4755 struct rq
*rq
= cpu_rq(cpu
);
4756 unsigned long interval
;
4757 struct sched_domain
*sd
;
4758 /* Earliest time when we have to do rebalance again */
4759 unsigned long next_balance
= jiffies
+ 60*HZ
;
4760 int update_next_balance
= 0;
4763 for_each_domain(cpu
, sd
) {
4764 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4767 interval
= sd
->balance_interval
;
4768 if (idle
!= CPU_IDLE
)
4769 interval
*= sd
->busy_factor
;
4771 /* scale ms to jiffies */
4772 interval
= msecs_to_jiffies(interval
);
4773 if (unlikely(!interval
))
4775 if (interval
> HZ
*NR_CPUS
/10)
4776 interval
= HZ
*NR_CPUS
/10;
4778 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4780 if (need_serialize
) {
4781 if (!spin_trylock(&balancing
))
4785 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4786 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4788 * We've pulled tasks over so either we're no
4789 * longer idle, or one of our SMT siblings is
4792 idle
= CPU_NOT_IDLE
;
4794 sd
->last_balance
= jiffies
;
4797 spin_unlock(&balancing
);
4799 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4800 next_balance
= sd
->last_balance
+ interval
;
4801 update_next_balance
= 1;
4805 * Stop the load balance at this level. There is another
4806 * CPU in our sched group which is doing load balancing more
4814 * next_balance will be updated only when there is a need.
4815 * When the cpu is attached to null domain for ex, it will not be
4818 if (likely(update_next_balance
))
4819 rq
->next_balance
= next_balance
;
4823 * run_rebalance_domains is triggered when needed from the scheduler tick.
4824 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4825 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4827 static void run_rebalance_domains(struct softirq_action
*h
)
4829 int this_cpu
= smp_processor_id();
4830 struct rq
*this_rq
= cpu_rq(this_cpu
);
4831 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4832 CPU_IDLE
: CPU_NOT_IDLE
;
4834 rebalance_domains(this_cpu
, idle
);
4838 * If this cpu is the owner for idle load balancing, then do the
4839 * balancing on behalf of the other idle cpus whose ticks are
4842 if (this_rq
->idle_at_tick
&&
4843 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4847 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4848 if (balance_cpu
== this_cpu
)
4852 * If this cpu gets work to do, stop the load balancing
4853 * work being done for other cpus. Next load
4854 * balancing owner will pick it up.
4859 rebalance_domains(balance_cpu
, CPU_IDLE
);
4861 rq
= cpu_rq(balance_cpu
);
4862 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4863 this_rq
->next_balance
= rq
->next_balance
;
4869 static inline int on_null_domain(int cpu
)
4871 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4875 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4877 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4878 * idle load balancing owner or decide to stop the periodic load balancing,
4879 * if the whole system is idle.
4881 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4885 * If we were in the nohz mode recently and busy at the current
4886 * scheduler tick, then check if we need to nominate new idle
4889 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4890 rq
->in_nohz_recently
= 0;
4892 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4893 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4894 atomic_set(&nohz
.load_balancer
, -1);
4897 if (atomic_read(&nohz
.load_balancer
) == -1) {
4898 int ilb
= find_new_ilb(cpu
);
4900 if (ilb
< nr_cpu_ids
)
4906 * If this cpu is idle and doing idle load balancing for all the
4907 * cpus with ticks stopped, is it time for that to stop?
4909 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4910 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4916 * If this cpu is idle and the idle load balancing is done by
4917 * someone else, then no need raise the SCHED_SOFTIRQ
4919 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4920 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4923 /* Don't need to rebalance while attached to NULL domain */
4924 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4925 likely(!on_null_domain(cpu
)))
4926 raise_softirq(SCHED_SOFTIRQ
);
4929 #else /* CONFIG_SMP */
4932 * on UP we do not need to balance between CPUs:
4934 static inline void idle_balance(int cpu
, struct rq
*rq
)
4940 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4942 EXPORT_PER_CPU_SYMBOL(kstat
);
4945 * Return any ns on the sched_clock that have not yet been accounted in
4946 * @p in case that task is currently running.
4948 * Called with task_rq_lock() held on @rq.
4950 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4954 if (task_current(rq
, p
)) {
4955 update_rq_clock(rq
);
4956 ns
= rq
->clock
- p
->se
.exec_start
;
4964 unsigned long long task_delta_exec(struct task_struct
*p
)
4966 unsigned long flags
;
4970 rq
= task_rq_lock(p
, &flags
);
4971 ns
= do_task_delta_exec(p
, rq
);
4972 task_rq_unlock(rq
, &flags
);
4978 * Return accounted runtime for the task.
4979 * In case the task is currently running, return the runtime plus current's
4980 * pending runtime that have not been accounted yet.
4982 unsigned long long task_sched_runtime(struct task_struct
*p
)
4984 unsigned long flags
;
4988 rq
= task_rq_lock(p
, &flags
);
4989 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4990 task_rq_unlock(rq
, &flags
);
4996 * Return sum_exec_runtime for the thread group.
4997 * In case the task is currently running, return the sum plus current's
4998 * pending runtime that have not been accounted yet.
5000 * Note that the thread group might have other running tasks as well,
5001 * so the return value not includes other pending runtime that other
5002 * running tasks might have.
5004 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5006 struct task_cputime totals
;
5007 unsigned long flags
;
5011 rq
= task_rq_lock(p
, &flags
);
5012 thread_group_cputime(p
, &totals
);
5013 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5014 task_rq_unlock(rq
, &flags
);
5020 * Account user cpu time to a process.
5021 * @p: the process that the cpu time gets accounted to
5022 * @cputime: the cpu time spent in user space since the last update
5023 * @cputime_scaled: cputime scaled by cpu frequency
5025 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5026 cputime_t cputime_scaled
)
5028 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5031 /* Add user time to process. */
5032 p
->utime
= cputime_add(p
->utime
, cputime
);
5033 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5034 account_group_user_time(p
, cputime
);
5036 /* Add user time to cpustat. */
5037 tmp
= cputime_to_cputime64(cputime
);
5038 if (TASK_NICE(p
) > 0)
5039 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5041 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5043 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5044 /* Account for user time used */
5045 acct_update_integrals(p
);
5049 * Account guest cpu time to a process.
5050 * @p: the process that the cpu time gets accounted to
5051 * @cputime: the cpu time spent in virtual machine since the last update
5052 * @cputime_scaled: cputime scaled by cpu frequency
5054 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5055 cputime_t cputime_scaled
)
5058 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5060 tmp
= cputime_to_cputime64(cputime
);
5062 /* Add guest time to process. */
5063 p
->utime
= cputime_add(p
->utime
, cputime
);
5064 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5065 account_group_user_time(p
, cputime
);
5066 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5068 /* Add guest time to cpustat. */
5069 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5070 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5074 * Account system cpu time to a process.
5075 * @p: the process that the cpu time gets accounted to
5076 * @hardirq_offset: the offset to subtract from hardirq_count()
5077 * @cputime: the cpu time spent in kernel space since the last update
5078 * @cputime_scaled: cputime scaled by cpu frequency
5080 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5081 cputime_t cputime
, cputime_t cputime_scaled
)
5083 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5086 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5087 account_guest_time(p
, cputime
, cputime_scaled
);
5091 /* Add system time to process. */
5092 p
->stime
= cputime_add(p
->stime
, cputime
);
5093 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5094 account_group_system_time(p
, cputime
);
5096 /* Add system time to cpustat. */
5097 tmp
= cputime_to_cputime64(cputime
);
5098 if (hardirq_count() - hardirq_offset
)
5099 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5100 else if (softirq_count())
5101 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5103 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5105 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5107 /* Account for system time used */
5108 acct_update_integrals(p
);
5112 * Account for involuntary wait time.
5113 * @steal: the cpu time spent in involuntary wait
5115 void account_steal_time(cputime_t cputime
)
5117 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5118 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5120 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5124 * Account for idle time.
5125 * @cputime: the cpu time spent in idle wait
5127 void account_idle_time(cputime_t cputime
)
5129 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5130 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5131 struct rq
*rq
= this_rq();
5133 if (atomic_read(&rq
->nr_iowait
) > 0)
5134 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5136 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5139 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5142 * Account a single tick of cpu time.
5143 * @p: the process that the cpu time gets accounted to
5144 * @user_tick: indicates if the tick is a user or a system tick
5146 void account_process_tick(struct task_struct
*p
, int user_tick
)
5148 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5149 struct rq
*rq
= this_rq();
5152 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5153 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5154 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5157 account_idle_time(cputime_one_jiffy
);
5161 * Account multiple ticks of steal time.
5162 * @p: the process from which the cpu time has been stolen
5163 * @ticks: number of stolen ticks
5165 void account_steal_ticks(unsigned long ticks
)
5167 account_steal_time(jiffies_to_cputime(ticks
));
5171 * Account multiple ticks of idle time.
5172 * @ticks: number of stolen ticks
5174 void account_idle_ticks(unsigned long ticks
)
5176 account_idle_time(jiffies_to_cputime(ticks
));
5182 * Use precise platform statistics if available:
5184 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5185 cputime_t
task_utime(struct task_struct
*p
)
5190 cputime_t
task_stime(struct task_struct
*p
)
5195 cputime_t
task_utime(struct task_struct
*p
)
5197 clock_t utime
= cputime_to_clock_t(p
->utime
),
5198 total
= utime
+ cputime_to_clock_t(p
->stime
);
5202 * Use CFS's precise accounting:
5204 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5208 do_div(temp
, total
);
5210 utime
= (clock_t)temp
;
5212 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5213 return p
->prev_utime
;
5216 cputime_t
task_stime(struct task_struct
*p
)
5221 * Use CFS's precise accounting. (we subtract utime from
5222 * the total, to make sure the total observed by userspace
5223 * grows monotonically - apps rely on that):
5225 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5226 cputime_to_clock_t(task_utime(p
));
5229 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5231 return p
->prev_stime
;
5235 inline cputime_t
task_gtime(struct task_struct
*p
)
5241 * This function gets called by the timer code, with HZ frequency.
5242 * We call it with interrupts disabled.
5244 * It also gets called by the fork code, when changing the parent's
5247 void scheduler_tick(void)
5249 int cpu
= smp_processor_id();
5250 struct rq
*rq
= cpu_rq(cpu
);
5251 struct task_struct
*curr
= rq
->curr
;
5255 spin_lock(&rq
->lock
);
5256 update_rq_clock(rq
);
5257 update_cpu_load(rq
);
5258 curr
->sched_class
->task_tick(rq
, curr
, 0);
5259 spin_unlock(&rq
->lock
);
5261 perf_event_task_tick(curr
, cpu
);
5264 rq
->idle_at_tick
= idle_cpu(cpu
);
5265 trigger_load_balance(rq
, cpu
);
5269 notrace
unsigned long get_parent_ip(unsigned long addr
)
5271 if (in_lock_functions(addr
)) {
5272 addr
= CALLER_ADDR2
;
5273 if (in_lock_functions(addr
))
5274 addr
= CALLER_ADDR3
;
5279 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5280 defined(CONFIG_PREEMPT_TRACER))
5282 void __kprobes
add_preempt_count(int val
)
5284 #ifdef CONFIG_DEBUG_PREEMPT
5288 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5291 preempt_count() += val
;
5292 #ifdef CONFIG_DEBUG_PREEMPT
5294 * Spinlock count overflowing soon?
5296 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5299 if (preempt_count() == val
)
5300 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5302 EXPORT_SYMBOL(add_preempt_count
);
5304 void __kprobes
sub_preempt_count(int val
)
5306 #ifdef CONFIG_DEBUG_PREEMPT
5310 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5313 * Is the spinlock portion underflowing?
5315 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5316 !(preempt_count() & PREEMPT_MASK
)))
5320 if (preempt_count() == val
)
5321 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5322 preempt_count() -= val
;
5324 EXPORT_SYMBOL(sub_preempt_count
);
5329 * Print scheduling while atomic bug:
5331 static noinline
void __schedule_bug(struct task_struct
*prev
)
5333 struct pt_regs
*regs
= get_irq_regs();
5335 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5336 prev
->comm
, prev
->pid
, preempt_count());
5338 debug_show_held_locks(prev
);
5340 if (irqs_disabled())
5341 print_irqtrace_events(prev
);
5350 * Various schedule()-time debugging checks and statistics:
5352 static inline void schedule_debug(struct task_struct
*prev
)
5355 * Test if we are atomic. Since do_exit() needs to call into
5356 * schedule() atomically, we ignore that path for now.
5357 * Otherwise, whine if we are scheduling when we should not be.
5359 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5360 __schedule_bug(prev
);
5362 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5364 schedstat_inc(this_rq(), sched_count
);
5365 #ifdef CONFIG_SCHEDSTATS
5366 if (unlikely(prev
->lock_depth
>= 0)) {
5367 schedstat_inc(this_rq(), bkl_count
);
5368 schedstat_inc(prev
, sched_info
.bkl_count
);
5373 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5375 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5377 update_avg(&p
->se
.avg_running
, runtime
);
5379 if (p
->state
== TASK_RUNNING
) {
5381 * In order to avoid avg_overlap growing stale when we are
5382 * indeed overlapping and hence not getting put to sleep, grow
5383 * the avg_overlap on preemption.
5385 * We use the average preemption runtime because that
5386 * correlates to the amount of cache footprint a task can
5389 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5390 update_avg(&p
->se
.avg_overlap
, runtime
);
5392 update_avg(&p
->se
.avg_running
, 0);
5394 p
->sched_class
->put_prev_task(rq
, p
);
5398 * Pick up the highest-prio task:
5400 static inline struct task_struct
*
5401 pick_next_task(struct rq
*rq
)
5403 const struct sched_class
*class;
5404 struct task_struct
*p
;
5407 * Optimization: we know that if all tasks are in
5408 * the fair class we can call that function directly:
5410 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5411 p
= fair_sched_class
.pick_next_task(rq
);
5416 class = sched_class_highest
;
5418 p
= class->pick_next_task(rq
);
5422 * Will never be NULL as the idle class always
5423 * returns a non-NULL p:
5425 class = class->next
;
5430 * schedule() is the main scheduler function.
5432 asmlinkage
void __sched
schedule(void)
5434 struct task_struct
*prev
, *next
;
5435 unsigned long *switch_count
;
5441 cpu
= smp_processor_id();
5445 switch_count
= &prev
->nivcsw
;
5447 release_kernel_lock(prev
);
5448 need_resched_nonpreemptible
:
5450 schedule_debug(prev
);
5452 if (sched_feat(HRTICK
))
5455 spin_lock_irq(&rq
->lock
);
5456 update_rq_clock(rq
);
5457 clear_tsk_need_resched(prev
);
5459 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5460 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5461 prev
->state
= TASK_RUNNING
;
5463 deactivate_task(rq
, prev
, 1);
5464 switch_count
= &prev
->nvcsw
;
5467 pre_schedule(rq
, prev
);
5469 if (unlikely(!rq
->nr_running
))
5470 idle_balance(cpu
, rq
);
5472 put_prev_task(rq
, prev
);
5473 next
= pick_next_task(rq
);
5475 if (likely(prev
!= next
)) {
5476 sched_info_switch(prev
, next
);
5477 perf_event_task_sched_out(prev
, next
, cpu
);
5483 context_switch(rq
, prev
, next
); /* unlocks the rq */
5485 * the context switch might have flipped the stack from under
5486 * us, hence refresh the local variables.
5488 cpu
= smp_processor_id();
5491 spin_unlock_irq(&rq
->lock
);
5495 if (unlikely(reacquire_kernel_lock(current
) < 0))
5496 goto need_resched_nonpreemptible
;
5498 preempt_enable_no_resched();
5502 EXPORT_SYMBOL(schedule
);
5506 * Look out! "owner" is an entirely speculative pointer
5507 * access and not reliable.
5509 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5514 if (!sched_feat(OWNER_SPIN
))
5517 #ifdef CONFIG_DEBUG_PAGEALLOC
5519 * Need to access the cpu field knowing that
5520 * DEBUG_PAGEALLOC could have unmapped it if
5521 * the mutex owner just released it and exited.
5523 if (probe_kernel_address(&owner
->cpu
, cpu
))
5530 * Even if the access succeeded (likely case),
5531 * the cpu field may no longer be valid.
5533 if (cpu
>= nr_cpumask_bits
)
5537 * We need to validate that we can do a
5538 * get_cpu() and that we have the percpu area.
5540 if (!cpu_online(cpu
))
5547 * Owner changed, break to re-assess state.
5549 if (lock
->owner
!= owner
)
5553 * Is that owner really running on that cpu?
5555 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5565 #ifdef CONFIG_PREEMPT
5567 * this is the entry point to schedule() from in-kernel preemption
5568 * off of preempt_enable. Kernel preemptions off return from interrupt
5569 * occur there and call schedule directly.
5571 asmlinkage
void __sched
preempt_schedule(void)
5573 struct thread_info
*ti
= current_thread_info();
5576 * If there is a non-zero preempt_count or interrupts are disabled,
5577 * we do not want to preempt the current task. Just return..
5579 if (likely(ti
->preempt_count
|| irqs_disabled()))
5583 add_preempt_count(PREEMPT_ACTIVE
);
5585 sub_preempt_count(PREEMPT_ACTIVE
);
5588 * Check again in case we missed a preemption opportunity
5589 * between schedule and now.
5592 } while (need_resched());
5594 EXPORT_SYMBOL(preempt_schedule
);
5597 * this is the entry point to schedule() from kernel preemption
5598 * off of irq context.
5599 * Note, that this is called and return with irqs disabled. This will
5600 * protect us against recursive calling from irq.
5602 asmlinkage
void __sched
preempt_schedule_irq(void)
5604 struct thread_info
*ti
= current_thread_info();
5606 /* Catch callers which need to be fixed */
5607 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5610 add_preempt_count(PREEMPT_ACTIVE
);
5613 local_irq_disable();
5614 sub_preempt_count(PREEMPT_ACTIVE
);
5617 * Check again in case we missed a preemption opportunity
5618 * between schedule and now.
5621 } while (need_resched());
5624 #endif /* CONFIG_PREEMPT */
5626 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5629 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5631 EXPORT_SYMBOL(default_wake_function
);
5634 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5635 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5636 * number) then we wake all the non-exclusive tasks and one exclusive task.
5638 * There are circumstances in which we can try to wake a task which has already
5639 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5640 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5642 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5643 int nr_exclusive
, int wake_flags
, void *key
)
5645 wait_queue_t
*curr
, *next
;
5647 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5648 unsigned flags
= curr
->flags
;
5650 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5651 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5657 * __wake_up - wake up threads blocked on a waitqueue.
5659 * @mode: which threads
5660 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5661 * @key: is directly passed to the wakeup function
5663 * It may be assumed that this function implies a write memory barrier before
5664 * changing the task state if and only if any tasks are woken up.
5666 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5667 int nr_exclusive
, void *key
)
5669 unsigned long flags
;
5671 spin_lock_irqsave(&q
->lock
, flags
);
5672 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5673 spin_unlock_irqrestore(&q
->lock
, flags
);
5675 EXPORT_SYMBOL(__wake_up
);
5678 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5680 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5682 __wake_up_common(q
, mode
, 1, 0, NULL
);
5685 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5687 __wake_up_common(q
, mode
, 1, 0, key
);
5691 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5693 * @mode: which threads
5694 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5695 * @key: opaque value to be passed to wakeup targets
5697 * The sync wakeup differs that the waker knows that it will schedule
5698 * away soon, so while the target thread will be woken up, it will not
5699 * be migrated to another CPU - ie. the two threads are 'synchronized'
5700 * with each other. This can prevent needless bouncing between CPUs.
5702 * On UP it can prevent extra preemption.
5704 * It may be assumed that this function implies a write memory barrier before
5705 * changing the task state if and only if any tasks are woken up.
5707 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5708 int nr_exclusive
, void *key
)
5710 unsigned long flags
;
5711 int wake_flags
= WF_SYNC
;
5716 if (unlikely(!nr_exclusive
))
5719 spin_lock_irqsave(&q
->lock
, flags
);
5720 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5721 spin_unlock_irqrestore(&q
->lock
, flags
);
5723 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5726 * __wake_up_sync - see __wake_up_sync_key()
5728 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5730 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5732 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5735 * complete: - signals a single thread waiting on this completion
5736 * @x: holds the state of this particular completion
5738 * This will wake up a single thread waiting on this completion. Threads will be
5739 * awakened in the same order in which they were queued.
5741 * See also complete_all(), wait_for_completion() and related routines.
5743 * It may be assumed that this function implies a write memory barrier before
5744 * changing the task state if and only if any tasks are woken up.
5746 void complete(struct completion
*x
)
5748 unsigned long flags
;
5750 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5752 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5753 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5755 EXPORT_SYMBOL(complete
);
5758 * complete_all: - signals all threads waiting on this completion
5759 * @x: holds the state of this particular completion
5761 * This will wake up all threads waiting on this particular completion event.
5763 * It may be assumed that this function implies a write memory barrier before
5764 * changing the task state if and only if any tasks are woken up.
5766 void complete_all(struct completion
*x
)
5768 unsigned long flags
;
5770 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5771 x
->done
+= UINT_MAX
/2;
5772 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5773 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5775 EXPORT_SYMBOL(complete_all
);
5777 static inline long __sched
5778 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5781 DECLARE_WAITQUEUE(wait
, current
);
5783 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5784 __add_wait_queue_tail(&x
->wait
, &wait
);
5786 if (signal_pending_state(state
, current
)) {
5787 timeout
= -ERESTARTSYS
;
5790 __set_current_state(state
);
5791 spin_unlock_irq(&x
->wait
.lock
);
5792 timeout
= schedule_timeout(timeout
);
5793 spin_lock_irq(&x
->wait
.lock
);
5794 } while (!x
->done
&& timeout
);
5795 __remove_wait_queue(&x
->wait
, &wait
);
5800 return timeout
?: 1;
5804 wait_for_common(struct completion
*x
, long timeout
, int state
)
5808 spin_lock_irq(&x
->wait
.lock
);
5809 timeout
= do_wait_for_common(x
, timeout
, state
);
5810 spin_unlock_irq(&x
->wait
.lock
);
5815 * wait_for_completion: - waits for completion of a task
5816 * @x: holds the state of this particular completion
5818 * This waits to be signaled for completion of a specific task. It is NOT
5819 * interruptible and there is no timeout.
5821 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5822 * and interrupt capability. Also see complete().
5824 void __sched
wait_for_completion(struct completion
*x
)
5826 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5828 EXPORT_SYMBOL(wait_for_completion
);
5831 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5832 * @x: holds the state of this particular completion
5833 * @timeout: timeout value in jiffies
5835 * This waits for either a completion of a specific task to be signaled or for a
5836 * specified timeout to expire. The timeout is in jiffies. It is not
5839 unsigned long __sched
5840 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5842 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5844 EXPORT_SYMBOL(wait_for_completion_timeout
);
5847 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5848 * @x: holds the state of this particular completion
5850 * This waits for completion of a specific task to be signaled. It is
5853 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5855 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5856 if (t
== -ERESTARTSYS
)
5860 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5863 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5864 * @x: holds the state of this particular completion
5865 * @timeout: timeout value in jiffies
5867 * This waits for either a completion of a specific task to be signaled or for a
5868 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5870 unsigned long __sched
5871 wait_for_completion_interruptible_timeout(struct completion
*x
,
5872 unsigned long timeout
)
5874 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5876 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5879 * wait_for_completion_killable: - waits for completion of a task (killable)
5880 * @x: holds the state of this particular completion
5882 * This waits to be signaled for completion of a specific task. It can be
5883 * interrupted by a kill signal.
5885 int __sched
wait_for_completion_killable(struct completion
*x
)
5887 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5888 if (t
== -ERESTARTSYS
)
5892 EXPORT_SYMBOL(wait_for_completion_killable
);
5895 * try_wait_for_completion - try to decrement a completion without blocking
5896 * @x: completion structure
5898 * Returns: 0 if a decrement cannot be done without blocking
5899 * 1 if a decrement succeeded.
5901 * If a completion is being used as a counting completion,
5902 * attempt to decrement the counter without blocking. This
5903 * enables us to avoid waiting if the resource the completion
5904 * is protecting is not available.
5906 bool try_wait_for_completion(struct completion
*x
)
5910 spin_lock_irq(&x
->wait
.lock
);
5915 spin_unlock_irq(&x
->wait
.lock
);
5918 EXPORT_SYMBOL(try_wait_for_completion
);
5921 * completion_done - Test to see if a completion has any waiters
5922 * @x: completion structure
5924 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5925 * 1 if there are no waiters.
5928 bool completion_done(struct completion
*x
)
5932 spin_lock_irq(&x
->wait
.lock
);
5935 spin_unlock_irq(&x
->wait
.lock
);
5938 EXPORT_SYMBOL(completion_done
);
5941 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5943 unsigned long flags
;
5946 init_waitqueue_entry(&wait
, current
);
5948 __set_current_state(state
);
5950 spin_lock_irqsave(&q
->lock
, flags
);
5951 __add_wait_queue(q
, &wait
);
5952 spin_unlock(&q
->lock
);
5953 timeout
= schedule_timeout(timeout
);
5954 spin_lock_irq(&q
->lock
);
5955 __remove_wait_queue(q
, &wait
);
5956 spin_unlock_irqrestore(&q
->lock
, flags
);
5961 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5963 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5965 EXPORT_SYMBOL(interruptible_sleep_on
);
5968 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5970 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5972 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5974 void __sched
sleep_on(wait_queue_head_t
*q
)
5976 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5978 EXPORT_SYMBOL(sleep_on
);
5980 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5982 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5984 EXPORT_SYMBOL(sleep_on_timeout
);
5986 #ifdef CONFIG_RT_MUTEXES
5989 * rt_mutex_setprio - set the current priority of a task
5991 * @prio: prio value (kernel-internal form)
5993 * This function changes the 'effective' priority of a task. It does
5994 * not touch ->normal_prio like __setscheduler().
5996 * Used by the rt_mutex code to implement priority inheritance logic.
5998 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6000 unsigned long flags
;
6001 int oldprio
, on_rq
, running
;
6003 const struct sched_class
*prev_class
= p
->sched_class
;
6005 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6007 rq
= task_rq_lock(p
, &flags
);
6008 update_rq_clock(rq
);
6011 on_rq
= p
->se
.on_rq
;
6012 running
= task_current(rq
, p
);
6014 dequeue_task(rq
, p
, 0);
6016 p
->sched_class
->put_prev_task(rq
, p
);
6019 p
->sched_class
= &rt_sched_class
;
6021 p
->sched_class
= &fair_sched_class
;
6026 p
->sched_class
->set_curr_task(rq
);
6028 enqueue_task(rq
, p
, 0);
6030 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6032 task_rq_unlock(rq
, &flags
);
6037 void set_user_nice(struct task_struct
*p
, long nice
)
6039 int old_prio
, delta
, on_rq
;
6040 unsigned long flags
;
6043 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6046 * We have to be careful, if called from sys_setpriority(),
6047 * the task might be in the middle of scheduling on another CPU.
6049 rq
= task_rq_lock(p
, &flags
);
6050 update_rq_clock(rq
);
6052 * The RT priorities are set via sched_setscheduler(), but we still
6053 * allow the 'normal' nice value to be set - but as expected
6054 * it wont have any effect on scheduling until the task is
6055 * SCHED_FIFO/SCHED_RR:
6057 if (task_has_rt_policy(p
)) {
6058 p
->static_prio
= NICE_TO_PRIO(nice
);
6061 on_rq
= p
->se
.on_rq
;
6063 dequeue_task(rq
, p
, 0);
6065 p
->static_prio
= NICE_TO_PRIO(nice
);
6068 p
->prio
= effective_prio(p
);
6069 delta
= p
->prio
- old_prio
;
6072 enqueue_task(rq
, p
, 0);
6074 * If the task increased its priority or is running and
6075 * lowered its priority, then reschedule its CPU:
6077 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6078 resched_task(rq
->curr
);
6081 task_rq_unlock(rq
, &flags
);
6083 EXPORT_SYMBOL(set_user_nice
);
6086 * can_nice - check if a task can reduce its nice value
6090 int can_nice(const struct task_struct
*p
, const int nice
)
6092 /* convert nice value [19,-20] to rlimit style value [1,40] */
6093 int nice_rlim
= 20 - nice
;
6095 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6096 capable(CAP_SYS_NICE
));
6099 #ifdef __ARCH_WANT_SYS_NICE
6102 * sys_nice - change the priority of the current process.
6103 * @increment: priority increment
6105 * sys_setpriority is a more generic, but much slower function that
6106 * does similar things.
6108 SYSCALL_DEFINE1(nice
, int, increment
)
6113 * Setpriority might change our priority at the same moment.
6114 * We don't have to worry. Conceptually one call occurs first
6115 * and we have a single winner.
6117 if (increment
< -40)
6122 nice
= TASK_NICE(current
) + increment
;
6128 if (increment
< 0 && !can_nice(current
, nice
))
6131 retval
= security_task_setnice(current
, nice
);
6135 set_user_nice(current
, nice
);
6142 * task_prio - return the priority value of a given task.
6143 * @p: the task in question.
6145 * This is the priority value as seen by users in /proc.
6146 * RT tasks are offset by -200. Normal tasks are centered
6147 * around 0, value goes from -16 to +15.
6149 int task_prio(const struct task_struct
*p
)
6151 return p
->prio
- MAX_RT_PRIO
;
6155 * task_nice - return the nice value of a given task.
6156 * @p: the task in question.
6158 int task_nice(const struct task_struct
*p
)
6160 return TASK_NICE(p
);
6162 EXPORT_SYMBOL(task_nice
);
6165 * idle_cpu - is a given cpu idle currently?
6166 * @cpu: the processor in question.
6168 int idle_cpu(int cpu
)
6170 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6174 * idle_task - return the idle task for a given cpu.
6175 * @cpu: the processor in question.
6177 struct task_struct
*idle_task(int cpu
)
6179 return cpu_rq(cpu
)->idle
;
6183 * find_process_by_pid - find a process with a matching PID value.
6184 * @pid: the pid in question.
6186 static struct task_struct
*find_process_by_pid(pid_t pid
)
6188 return pid
? find_task_by_vpid(pid
) : current
;
6191 /* Actually do priority change: must hold rq lock. */
6193 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6195 BUG_ON(p
->se
.on_rq
);
6198 switch (p
->policy
) {
6202 p
->sched_class
= &fair_sched_class
;
6206 p
->sched_class
= &rt_sched_class
;
6210 p
->rt_priority
= prio
;
6211 p
->normal_prio
= normal_prio(p
);
6212 /* we are holding p->pi_lock already */
6213 p
->prio
= rt_mutex_getprio(p
);
6218 * check the target process has a UID that matches the current process's
6220 static bool check_same_owner(struct task_struct
*p
)
6222 const struct cred
*cred
= current_cred(), *pcred
;
6226 pcred
= __task_cred(p
);
6227 match
= (cred
->euid
== pcred
->euid
||
6228 cred
->euid
== pcred
->uid
);
6233 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6234 struct sched_param
*param
, bool user
)
6236 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6237 unsigned long flags
;
6238 const struct sched_class
*prev_class
= p
->sched_class
;
6242 /* may grab non-irq protected spin_locks */
6243 BUG_ON(in_interrupt());
6245 /* double check policy once rq lock held */
6247 reset_on_fork
= p
->sched_reset_on_fork
;
6248 policy
= oldpolicy
= p
->policy
;
6250 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6251 policy
&= ~SCHED_RESET_ON_FORK
;
6253 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6254 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6255 policy
!= SCHED_IDLE
)
6260 * Valid priorities for SCHED_FIFO and SCHED_RR are
6261 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6262 * SCHED_BATCH and SCHED_IDLE is 0.
6264 if (param
->sched_priority
< 0 ||
6265 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6266 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6268 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6272 * Allow unprivileged RT tasks to decrease priority:
6274 if (user
&& !capable(CAP_SYS_NICE
)) {
6275 if (rt_policy(policy
)) {
6276 unsigned long rlim_rtprio
;
6278 if (!lock_task_sighand(p
, &flags
))
6280 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6281 unlock_task_sighand(p
, &flags
);
6283 /* can't set/change the rt policy */
6284 if (policy
!= p
->policy
&& !rlim_rtprio
)
6287 /* can't increase priority */
6288 if (param
->sched_priority
> p
->rt_priority
&&
6289 param
->sched_priority
> rlim_rtprio
)
6293 * Like positive nice levels, dont allow tasks to
6294 * move out of SCHED_IDLE either:
6296 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6299 /* can't change other user's priorities */
6300 if (!check_same_owner(p
))
6303 /* Normal users shall not reset the sched_reset_on_fork flag */
6304 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6309 #ifdef CONFIG_RT_GROUP_SCHED
6311 * Do not allow realtime tasks into groups that have no runtime
6314 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6315 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6319 retval
= security_task_setscheduler(p
, policy
, param
);
6325 * make sure no PI-waiters arrive (or leave) while we are
6326 * changing the priority of the task:
6328 spin_lock_irqsave(&p
->pi_lock
, flags
);
6330 * To be able to change p->policy safely, the apropriate
6331 * runqueue lock must be held.
6333 rq
= __task_rq_lock(p
);
6334 /* recheck policy now with rq lock held */
6335 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6336 policy
= oldpolicy
= -1;
6337 __task_rq_unlock(rq
);
6338 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6341 update_rq_clock(rq
);
6342 on_rq
= p
->se
.on_rq
;
6343 running
= task_current(rq
, p
);
6345 deactivate_task(rq
, p
, 0);
6347 p
->sched_class
->put_prev_task(rq
, p
);
6349 p
->sched_reset_on_fork
= reset_on_fork
;
6352 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6355 p
->sched_class
->set_curr_task(rq
);
6357 activate_task(rq
, p
, 0);
6359 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6361 __task_rq_unlock(rq
);
6362 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6364 rt_mutex_adjust_pi(p
);
6370 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6371 * @p: the task in question.
6372 * @policy: new policy.
6373 * @param: structure containing the new RT priority.
6375 * NOTE that the task may be already dead.
6377 int sched_setscheduler(struct task_struct
*p
, int policy
,
6378 struct sched_param
*param
)
6380 return __sched_setscheduler(p
, policy
, param
, true);
6382 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6385 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6386 * @p: the task in question.
6387 * @policy: new policy.
6388 * @param: structure containing the new RT priority.
6390 * Just like sched_setscheduler, only don't bother checking if the
6391 * current context has permission. For example, this is needed in
6392 * stop_machine(): we create temporary high priority worker threads,
6393 * but our caller might not have that capability.
6395 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6396 struct sched_param
*param
)
6398 return __sched_setscheduler(p
, policy
, param
, false);
6402 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6404 struct sched_param lparam
;
6405 struct task_struct
*p
;
6408 if (!param
|| pid
< 0)
6410 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6415 p
= find_process_by_pid(pid
);
6417 retval
= sched_setscheduler(p
, policy
, &lparam
);
6424 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6425 * @pid: the pid in question.
6426 * @policy: new policy.
6427 * @param: structure containing the new RT priority.
6429 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6430 struct sched_param __user
*, param
)
6432 /* negative values for policy are not valid */
6436 return do_sched_setscheduler(pid
, policy
, param
);
6440 * sys_sched_setparam - set/change the RT priority of a thread
6441 * @pid: the pid in question.
6442 * @param: structure containing the new RT priority.
6444 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6446 return do_sched_setscheduler(pid
, -1, param
);
6450 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6451 * @pid: the pid in question.
6453 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6455 struct task_struct
*p
;
6462 read_lock(&tasklist_lock
);
6463 p
= find_process_by_pid(pid
);
6465 retval
= security_task_getscheduler(p
);
6468 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6470 read_unlock(&tasklist_lock
);
6475 * sys_sched_getparam - get the RT priority of a thread
6476 * @pid: the pid in question.
6477 * @param: structure containing the RT priority.
6479 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6481 struct sched_param lp
;
6482 struct task_struct
*p
;
6485 if (!param
|| pid
< 0)
6488 read_lock(&tasklist_lock
);
6489 p
= find_process_by_pid(pid
);
6494 retval
= security_task_getscheduler(p
);
6498 lp
.sched_priority
= p
->rt_priority
;
6499 read_unlock(&tasklist_lock
);
6502 * This one might sleep, we cannot do it with a spinlock held ...
6504 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6509 read_unlock(&tasklist_lock
);
6513 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6515 cpumask_var_t cpus_allowed
, new_mask
;
6516 struct task_struct
*p
;
6520 read_lock(&tasklist_lock
);
6522 p
= find_process_by_pid(pid
);
6524 read_unlock(&tasklist_lock
);
6530 * It is not safe to call set_cpus_allowed with the
6531 * tasklist_lock held. We will bump the task_struct's
6532 * usage count and then drop tasklist_lock.
6535 read_unlock(&tasklist_lock
);
6537 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6541 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6543 goto out_free_cpus_allowed
;
6546 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6549 retval
= security_task_setscheduler(p
, 0, NULL
);
6553 cpuset_cpus_allowed(p
, cpus_allowed
);
6554 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6556 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6559 cpuset_cpus_allowed(p
, cpus_allowed
);
6560 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6562 * We must have raced with a concurrent cpuset
6563 * update. Just reset the cpus_allowed to the
6564 * cpuset's cpus_allowed
6566 cpumask_copy(new_mask
, cpus_allowed
);
6571 free_cpumask_var(new_mask
);
6572 out_free_cpus_allowed
:
6573 free_cpumask_var(cpus_allowed
);
6580 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6581 struct cpumask
*new_mask
)
6583 if (len
< cpumask_size())
6584 cpumask_clear(new_mask
);
6585 else if (len
> cpumask_size())
6586 len
= cpumask_size();
6588 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6592 * sys_sched_setaffinity - set the cpu affinity of a process
6593 * @pid: pid of the process
6594 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6595 * @user_mask_ptr: user-space pointer to the new cpu mask
6597 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6598 unsigned long __user
*, user_mask_ptr
)
6600 cpumask_var_t new_mask
;
6603 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6606 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6608 retval
= sched_setaffinity(pid
, new_mask
);
6609 free_cpumask_var(new_mask
);
6613 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6615 struct task_struct
*p
;
6619 read_lock(&tasklist_lock
);
6622 p
= find_process_by_pid(pid
);
6626 retval
= security_task_getscheduler(p
);
6630 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6633 read_unlock(&tasklist_lock
);
6640 * sys_sched_getaffinity - get the cpu affinity of a process
6641 * @pid: pid of the process
6642 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6643 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6645 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6646 unsigned long __user
*, user_mask_ptr
)
6651 if (len
< cpumask_size())
6654 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6657 ret
= sched_getaffinity(pid
, mask
);
6659 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6662 ret
= cpumask_size();
6664 free_cpumask_var(mask
);
6670 * sys_sched_yield - yield the current processor to other threads.
6672 * This function yields the current CPU to other tasks. If there are no
6673 * other threads running on this CPU then this function will return.
6675 SYSCALL_DEFINE0(sched_yield
)
6677 struct rq
*rq
= this_rq_lock();
6679 schedstat_inc(rq
, yld_count
);
6680 current
->sched_class
->yield_task(rq
);
6683 * Since we are going to call schedule() anyway, there's
6684 * no need to preempt or enable interrupts:
6686 __release(rq
->lock
);
6687 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6688 _raw_spin_unlock(&rq
->lock
);
6689 preempt_enable_no_resched();
6696 static inline int should_resched(void)
6698 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6701 static void __cond_resched(void)
6703 add_preempt_count(PREEMPT_ACTIVE
);
6705 sub_preempt_count(PREEMPT_ACTIVE
);
6708 int __sched
_cond_resched(void)
6710 if (should_resched()) {
6716 EXPORT_SYMBOL(_cond_resched
);
6719 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6720 * call schedule, and on return reacquire the lock.
6722 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6723 * operations here to prevent schedule() from being called twice (once via
6724 * spin_unlock(), once by hand).
6726 int __cond_resched_lock(spinlock_t
*lock
)
6728 int resched
= should_resched();
6731 lockdep_assert_held(lock
);
6733 if (spin_needbreak(lock
) || resched
) {
6744 EXPORT_SYMBOL(__cond_resched_lock
);
6746 int __sched
__cond_resched_softirq(void)
6748 BUG_ON(!in_softirq());
6750 if (should_resched()) {
6758 EXPORT_SYMBOL(__cond_resched_softirq
);
6761 * yield - yield the current processor to other threads.
6763 * This is a shortcut for kernel-space yielding - it marks the
6764 * thread runnable and calls sys_sched_yield().
6766 void __sched
yield(void)
6768 set_current_state(TASK_RUNNING
);
6771 EXPORT_SYMBOL(yield
);
6774 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6775 * that process accounting knows that this is a task in IO wait state.
6777 void __sched
io_schedule(void)
6779 struct rq
*rq
= raw_rq();
6781 delayacct_blkio_start();
6782 atomic_inc(&rq
->nr_iowait
);
6783 current
->in_iowait
= 1;
6785 current
->in_iowait
= 0;
6786 atomic_dec(&rq
->nr_iowait
);
6787 delayacct_blkio_end();
6789 EXPORT_SYMBOL(io_schedule
);
6791 long __sched
io_schedule_timeout(long timeout
)
6793 struct rq
*rq
= raw_rq();
6796 delayacct_blkio_start();
6797 atomic_inc(&rq
->nr_iowait
);
6798 current
->in_iowait
= 1;
6799 ret
= schedule_timeout(timeout
);
6800 current
->in_iowait
= 0;
6801 atomic_dec(&rq
->nr_iowait
);
6802 delayacct_blkio_end();
6807 * sys_sched_get_priority_max - return maximum RT priority.
6808 * @policy: scheduling class.
6810 * this syscall returns the maximum rt_priority that can be used
6811 * by a given scheduling class.
6813 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6820 ret
= MAX_USER_RT_PRIO
-1;
6832 * sys_sched_get_priority_min - return minimum RT priority.
6833 * @policy: scheduling class.
6835 * this syscall returns the minimum rt_priority that can be used
6836 * by a given scheduling class.
6838 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6856 * sys_sched_rr_get_interval - return the default timeslice of a process.
6857 * @pid: pid of the process.
6858 * @interval: userspace pointer to the timeslice value.
6860 * this syscall writes the default timeslice value of a given process
6861 * into the user-space timespec buffer. A value of '0' means infinity.
6863 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6864 struct timespec __user
*, interval
)
6866 struct task_struct
*p
;
6867 unsigned int time_slice
;
6875 read_lock(&tasklist_lock
);
6876 p
= find_process_by_pid(pid
);
6880 retval
= security_task_getscheduler(p
);
6884 time_slice
= p
->sched_class
->get_rr_interval(p
);
6886 read_unlock(&tasklist_lock
);
6887 jiffies_to_timespec(time_slice
, &t
);
6888 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6892 read_unlock(&tasklist_lock
);
6896 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6898 void sched_show_task(struct task_struct
*p
)
6900 unsigned long free
= 0;
6903 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6904 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6905 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6906 #if BITS_PER_LONG == 32
6907 if (state
== TASK_RUNNING
)
6908 printk(KERN_CONT
" running ");
6910 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6912 if (state
== TASK_RUNNING
)
6913 printk(KERN_CONT
" running task ");
6915 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6917 #ifdef CONFIG_DEBUG_STACK_USAGE
6918 free
= stack_not_used(p
);
6920 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6921 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6922 (unsigned long)task_thread_info(p
)->flags
);
6924 show_stack(p
, NULL
);
6927 void show_state_filter(unsigned long state_filter
)
6929 struct task_struct
*g
, *p
;
6931 #if BITS_PER_LONG == 32
6933 " task PC stack pid father\n");
6936 " task PC stack pid father\n");
6938 read_lock(&tasklist_lock
);
6939 do_each_thread(g
, p
) {
6941 * reset the NMI-timeout, listing all files on a slow
6942 * console might take alot of time:
6944 touch_nmi_watchdog();
6945 if (!state_filter
|| (p
->state
& state_filter
))
6947 } while_each_thread(g
, p
);
6949 touch_all_softlockup_watchdogs();
6951 #ifdef CONFIG_SCHED_DEBUG
6952 sysrq_sched_debug_show();
6954 read_unlock(&tasklist_lock
);
6956 * Only show locks if all tasks are dumped:
6958 if (state_filter
== -1)
6959 debug_show_all_locks();
6962 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6964 idle
->sched_class
= &idle_sched_class
;
6968 * init_idle - set up an idle thread for a given CPU
6969 * @idle: task in question
6970 * @cpu: cpu the idle task belongs to
6972 * NOTE: this function does not set the idle thread's NEED_RESCHED
6973 * flag, to make booting more robust.
6975 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6977 struct rq
*rq
= cpu_rq(cpu
);
6978 unsigned long flags
;
6980 spin_lock_irqsave(&rq
->lock
, flags
);
6983 idle
->se
.exec_start
= sched_clock();
6985 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6986 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6987 __set_task_cpu(idle
, cpu
);
6989 rq
->curr
= rq
->idle
= idle
;
6990 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6993 spin_unlock_irqrestore(&rq
->lock
, flags
);
6995 /* Set the preempt count _outside_ the spinlocks! */
6996 #if defined(CONFIG_PREEMPT)
6997 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6999 task_thread_info(idle
)->preempt_count
= 0;
7002 * The idle tasks have their own, simple scheduling class:
7004 idle
->sched_class
= &idle_sched_class
;
7005 ftrace_graph_init_task(idle
);
7009 * In a system that switches off the HZ timer nohz_cpu_mask
7010 * indicates which cpus entered this state. This is used
7011 * in the rcu update to wait only for active cpus. For system
7012 * which do not switch off the HZ timer nohz_cpu_mask should
7013 * always be CPU_BITS_NONE.
7015 cpumask_var_t nohz_cpu_mask
;
7018 * Increase the granularity value when there are more CPUs,
7019 * because with more CPUs the 'effective latency' as visible
7020 * to users decreases. But the relationship is not linear,
7021 * so pick a second-best guess by going with the log2 of the
7024 * This idea comes from the SD scheduler of Con Kolivas:
7026 static inline void sched_init_granularity(void)
7028 unsigned int factor
= 1 + ilog2(num_online_cpus());
7029 const unsigned long limit
= 200000000;
7031 sysctl_sched_min_granularity
*= factor
;
7032 if (sysctl_sched_min_granularity
> limit
)
7033 sysctl_sched_min_granularity
= limit
;
7035 sysctl_sched_latency
*= factor
;
7036 if (sysctl_sched_latency
> limit
)
7037 sysctl_sched_latency
= limit
;
7039 sysctl_sched_wakeup_granularity
*= factor
;
7041 sysctl_sched_shares_ratelimit
*= factor
;
7046 * This is how migration works:
7048 * 1) we queue a struct migration_req structure in the source CPU's
7049 * runqueue and wake up that CPU's migration thread.
7050 * 2) we down() the locked semaphore => thread blocks.
7051 * 3) migration thread wakes up (implicitly it forces the migrated
7052 * thread off the CPU)
7053 * 4) it gets the migration request and checks whether the migrated
7054 * task is still in the wrong runqueue.
7055 * 5) if it's in the wrong runqueue then the migration thread removes
7056 * it and puts it into the right queue.
7057 * 6) migration thread up()s the semaphore.
7058 * 7) we wake up and the migration is done.
7062 * Change a given task's CPU affinity. Migrate the thread to a
7063 * proper CPU and schedule it away if the CPU it's executing on
7064 * is removed from the allowed bitmask.
7066 * NOTE: the caller must have a valid reference to the task, the
7067 * task must not exit() & deallocate itself prematurely. The
7068 * call is not atomic; no spinlocks may be held.
7070 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7072 struct migration_req req
;
7073 unsigned long flags
;
7077 rq
= task_rq_lock(p
, &flags
);
7078 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7083 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7084 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7089 if (p
->sched_class
->set_cpus_allowed
)
7090 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7092 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7093 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7096 /* Can the task run on the task's current CPU? If so, we're done */
7097 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7100 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7101 /* Need help from migration thread: drop lock and wait. */
7102 struct task_struct
*mt
= rq
->migration_thread
;
7104 get_task_struct(mt
);
7105 task_rq_unlock(rq
, &flags
);
7106 wake_up_process(rq
->migration_thread
);
7107 put_task_struct(mt
);
7108 wait_for_completion(&req
.done
);
7109 tlb_migrate_finish(p
->mm
);
7113 task_rq_unlock(rq
, &flags
);
7117 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7120 * Move (not current) task off this cpu, onto dest cpu. We're doing
7121 * this because either it can't run here any more (set_cpus_allowed()
7122 * away from this CPU, or CPU going down), or because we're
7123 * attempting to rebalance this task on exec (sched_exec).
7125 * So we race with normal scheduler movements, but that's OK, as long
7126 * as the task is no longer on this CPU.
7128 * Returns non-zero if task was successfully migrated.
7130 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7132 struct rq
*rq_dest
, *rq_src
;
7135 if (unlikely(!cpu_active(dest_cpu
)))
7138 rq_src
= cpu_rq(src_cpu
);
7139 rq_dest
= cpu_rq(dest_cpu
);
7141 double_rq_lock(rq_src
, rq_dest
);
7142 /* Already moved. */
7143 if (task_cpu(p
) != src_cpu
)
7145 /* Affinity changed (again). */
7146 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7149 on_rq
= p
->se
.on_rq
;
7151 deactivate_task(rq_src
, p
, 0);
7153 set_task_cpu(p
, dest_cpu
);
7155 activate_task(rq_dest
, p
, 0);
7156 check_preempt_curr(rq_dest
, p
, 0);
7161 double_rq_unlock(rq_src
, rq_dest
);
7165 #define RCU_MIGRATION_IDLE 0
7166 #define RCU_MIGRATION_NEED_QS 1
7167 #define RCU_MIGRATION_GOT_QS 2
7168 #define RCU_MIGRATION_MUST_SYNC 3
7171 * migration_thread - this is a highprio system thread that performs
7172 * thread migration by bumping thread off CPU then 'pushing' onto
7175 static int migration_thread(void *data
)
7178 int cpu
= (long)data
;
7182 BUG_ON(rq
->migration_thread
!= current
);
7184 set_current_state(TASK_INTERRUPTIBLE
);
7185 while (!kthread_should_stop()) {
7186 struct migration_req
*req
;
7187 struct list_head
*head
;
7189 spin_lock_irq(&rq
->lock
);
7191 if (cpu_is_offline(cpu
)) {
7192 spin_unlock_irq(&rq
->lock
);
7196 if (rq
->active_balance
) {
7197 active_load_balance(rq
, cpu
);
7198 rq
->active_balance
= 0;
7201 head
= &rq
->migration_queue
;
7203 if (list_empty(head
)) {
7204 spin_unlock_irq(&rq
->lock
);
7206 set_current_state(TASK_INTERRUPTIBLE
);
7209 req
= list_entry(head
->next
, struct migration_req
, list
);
7210 list_del_init(head
->next
);
7212 if (req
->task
!= NULL
) {
7213 spin_unlock(&rq
->lock
);
7214 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7215 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7216 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7217 spin_unlock(&rq
->lock
);
7219 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7220 spin_unlock(&rq
->lock
);
7221 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7225 complete(&req
->done
);
7227 __set_current_state(TASK_RUNNING
);
7232 #ifdef CONFIG_HOTPLUG_CPU
7234 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7238 local_irq_disable();
7239 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7245 * Figure out where task on dead CPU should go, use force if necessary.
7247 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7250 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7253 /* Look for allowed, online CPU in same node. */
7254 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7255 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7258 /* Any allowed, online CPU? */
7259 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7260 if (dest_cpu
< nr_cpu_ids
)
7263 /* No more Mr. Nice Guy. */
7264 if (dest_cpu
>= nr_cpu_ids
) {
7265 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7266 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7269 * Don't tell them about moving exiting tasks or
7270 * kernel threads (both mm NULL), since they never
7273 if (p
->mm
&& printk_ratelimit()) {
7274 printk(KERN_INFO
"process %d (%s) no "
7275 "longer affine to cpu%d\n",
7276 task_pid_nr(p
), p
->comm
, dead_cpu
);
7281 /* It can have affinity changed while we were choosing. */
7282 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7287 * While a dead CPU has no uninterruptible tasks queued at this point,
7288 * it might still have a nonzero ->nr_uninterruptible counter, because
7289 * for performance reasons the counter is not stricly tracking tasks to
7290 * their home CPUs. So we just add the counter to another CPU's counter,
7291 * to keep the global sum constant after CPU-down:
7293 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7295 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7296 unsigned long flags
;
7298 local_irq_save(flags
);
7299 double_rq_lock(rq_src
, rq_dest
);
7300 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7301 rq_src
->nr_uninterruptible
= 0;
7302 double_rq_unlock(rq_src
, rq_dest
);
7303 local_irq_restore(flags
);
7306 /* Run through task list and migrate tasks from the dead cpu. */
7307 static void migrate_live_tasks(int src_cpu
)
7309 struct task_struct
*p
, *t
;
7311 read_lock(&tasklist_lock
);
7313 do_each_thread(t
, p
) {
7317 if (task_cpu(p
) == src_cpu
)
7318 move_task_off_dead_cpu(src_cpu
, p
);
7319 } while_each_thread(t
, p
);
7321 read_unlock(&tasklist_lock
);
7325 * Schedules idle task to be the next runnable task on current CPU.
7326 * It does so by boosting its priority to highest possible.
7327 * Used by CPU offline code.
7329 void sched_idle_next(void)
7331 int this_cpu
= smp_processor_id();
7332 struct rq
*rq
= cpu_rq(this_cpu
);
7333 struct task_struct
*p
= rq
->idle
;
7334 unsigned long flags
;
7336 /* cpu has to be offline */
7337 BUG_ON(cpu_online(this_cpu
));
7340 * Strictly not necessary since rest of the CPUs are stopped by now
7341 * and interrupts disabled on the current cpu.
7343 spin_lock_irqsave(&rq
->lock
, flags
);
7345 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7347 update_rq_clock(rq
);
7348 activate_task(rq
, p
, 0);
7350 spin_unlock_irqrestore(&rq
->lock
, flags
);
7354 * Ensures that the idle task is using init_mm right before its cpu goes
7357 void idle_task_exit(void)
7359 struct mm_struct
*mm
= current
->active_mm
;
7361 BUG_ON(cpu_online(smp_processor_id()));
7364 switch_mm(mm
, &init_mm
, current
);
7368 /* called under rq->lock with disabled interrupts */
7369 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7371 struct rq
*rq
= cpu_rq(dead_cpu
);
7373 /* Must be exiting, otherwise would be on tasklist. */
7374 BUG_ON(!p
->exit_state
);
7376 /* Cannot have done final schedule yet: would have vanished. */
7377 BUG_ON(p
->state
== TASK_DEAD
);
7382 * Drop lock around migration; if someone else moves it,
7383 * that's OK. No task can be added to this CPU, so iteration is
7386 spin_unlock_irq(&rq
->lock
);
7387 move_task_off_dead_cpu(dead_cpu
, p
);
7388 spin_lock_irq(&rq
->lock
);
7393 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7394 static void migrate_dead_tasks(unsigned int dead_cpu
)
7396 struct rq
*rq
= cpu_rq(dead_cpu
);
7397 struct task_struct
*next
;
7400 if (!rq
->nr_running
)
7402 update_rq_clock(rq
);
7403 next
= pick_next_task(rq
);
7406 next
->sched_class
->put_prev_task(rq
, next
);
7407 migrate_dead(dead_cpu
, next
);
7413 * remove the tasks which were accounted by rq from calc_load_tasks.
7415 static void calc_global_load_remove(struct rq
*rq
)
7417 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7418 rq
->calc_load_active
= 0;
7420 #endif /* CONFIG_HOTPLUG_CPU */
7422 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7424 static struct ctl_table sd_ctl_dir
[] = {
7426 .procname
= "sched_domain",
7432 static struct ctl_table sd_ctl_root
[] = {
7434 .ctl_name
= CTL_KERN
,
7435 .procname
= "kernel",
7437 .child
= sd_ctl_dir
,
7442 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7444 struct ctl_table
*entry
=
7445 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7450 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7452 struct ctl_table
*entry
;
7455 * In the intermediate directories, both the child directory and
7456 * procname are dynamically allocated and could fail but the mode
7457 * will always be set. In the lowest directory the names are
7458 * static strings and all have proc handlers.
7460 for (entry
= *tablep
; entry
->mode
; entry
++) {
7462 sd_free_ctl_entry(&entry
->child
);
7463 if (entry
->proc_handler
== NULL
)
7464 kfree(entry
->procname
);
7472 set_table_entry(struct ctl_table
*entry
,
7473 const char *procname
, void *data
, int maxlen
,
7474 mode_t mode
, proc_handler
*proc_handler
)
7476 entry
->procname
= procname
;
7478 entry
->maxlen
= maxlen
;
7480 entry
->proc_handler
= proc_handler
;
7483 static struct ctl_table
*
7484 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7486 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7491 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7492 sizeof(long), 0644, proc_doulongvec_minmax
);
7493 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7494 sizeof(long), 0644, proc_doulongvec_minmax
);
7495 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7496 sizeof(int), 0644, proc_dointvec_minmax
);
7497 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7498 sizeof(int), 0644, proc_dointvec_minmax
);
7499 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7500 sizeof(int), 0644, proc_dointvec_minmax
);
7501 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7502 sizeof(int), 0644, proc_dointvec_minmax
);
7503 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7504 sizeof(int), 0644, proc_dointvec_minmax
);
7505 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7506 sizeof(int), 0644, proc_dointvec_minmax
);
7507 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7508 sizeof(int), 0644, proc_dointvec_minmax
);
7509 set_table_entry(&table
[9], "cache_nice_tries",
7510 &sd
->cache_nice_tries
,
7511 sizeof(int), 0644, proc_dointvec_minmax
);
7512 set_table_entry(&table
[10], "flags", &sd
->flags
,
7513 sizeof(int), 0644, proc_dointvec_minmax
);
7514 set_table_entry(&table
[11], "name", sd
->name
,
7515 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7516 /* &table[12] is terminator */
7521 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7523 struct ctl_table
*entry
, *table
;
7524 struct sched_domain
*sd
;
7525 int domain_num
= 0, i
;
7528 for_each_domain(cpu
, sd
)
7530 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7535 for_each_domain(cpu
, sd
) {
7536 snprintf(buf
, 32, "domain%d", i
);
7537 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7539 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7546 static struct ctl_table_header
*sd_sysctl_header
;
7547 static void register_sched_domain_sysctl(void)
7549 int i
, cpu_num
= num_online_cpus();
7550 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7553 WARN_ON(sd_ctl_dir
[0].child
);
7554 sd_ctl_dir
[0].child
= entry
;
7559 for_each_online_cpu(i
) {
7560 snprintf(buf
, 32, "cpu%d", i
);
7561 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7563 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7567 WARN_ON(sd_sysctl_header
);
7568 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7571 /* may be called multiple times per register */
7572 static void unregister_sched_domain_sysctl(void)
7574 if (sd_sysctl_header
)
7575 unregister_sysctl_table(sd_sysctl_header
);
7576 sd_sysctl_header
= NULL
;
7577 if (sd_ctl_dir
[0].child
)
7578 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7581 static void register_sched_domain_sysctl(void)
7584 static void unregister_sched_domain_sysctl(void)
7589 static void set_rq_online(struct rq
*rq
)
7592 const struct sched_class
*class;
7594 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7597 for_each_class(class) {
7598 if (class->rq_online
)
7599 class->rq_online(rq
);
7604 static void set_rq_offline(struct rq
*rq
)
7607 const struct sched_class
*class;
7609 for_each_class(class) {
7610 if (class->rq_offline
)
7611 class->rq_offline(rq
);
7614 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7620 * migration_call - callback that gets triggered when a CPU is added.
7621 * Here we can start up the necessary migration thread for the new CPU.
7623 static int __cpuinit
7624 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7626 struct task_struct
*p
;
7627 int cpu
= (long)hcpu
;
7628 unsigned long flags
;
7633 case CPU_UP_PREPARE
:
7634 case CPU_UP_PREPARE_FROZEN
:
7635 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7638 kthread_bind(p
, cpu
);
7639 /* Must be high prio: stop_machine expects to yield to it. */
7640 rq
= task_rq_lock(p
, &flags
);
7641 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7642 task_rq_unlock(rq
, &flags
);
7644 cpu_rq(cpu
)->migration_thread
= p
;
7645 rq
->calc_load_update
= calc_load_update
;
7649 case CPU_ONLINE_FROZEN
:
7650 /* Strictly unnecessary, as first user will wake it. */
7651 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7653 /* Update our root-domain */
7655 spin_lock_irqsave(&rq
->lock
, flags
);
7657 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7661 spin_unlock_irqrestore(&rq
->lock
, flags
);
7664 #ifdef CONFIG_HOTPLUG_CPU
7665 case CPU_UP_CANCELED
:
7666 case CPU_UP_CANCELED_FROZEN
:
7667 if (!cpu_rq(cpu
)->migration_thread
)
7669 /* Unbind it from offline cpu so it can run. Fall thru. */
7670 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7671 cpumask_any(cpu_online_mask
));
7672 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7673 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7674 cpu_rq(cpu
)->migration_thread
= NULL
;
7678 case CPU_DEAD_FROZEN
:
7679 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7680 migrate_live_tasks(cpu
);
7682 kthread_stop(rq
->migration_thread
);
7683 put_task_struct(rq
->migration_thread
);
7684 rq
->migration_thread
= NULL
;
7685 /* Idle task back to normal (off runqueue, low prio) */
7686 spin_lock_irq(&rq
->lock
);
7687 update_rq_clock(rq
);
7688 deactivate_task(rq
, rq
->idle
, 0);
7689 rq
->idle
->static_prio
= MAX_PRIO
;
7690 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7691 rq
->idle
->sched_class
= &idle_sched_class
;
7692 migrate_dead_tasks(cpu
);
7693 spin_unlock_irq(&rq
->lock
);
7695 migrate_nr_uninterruptible(rq
);
7696 BUG_ON(rq
->nr_running
!= 0);
7697 calc_global_load_remove(rq
);
7699 * No need to migrate the tasks: it was best-effort if
7700 * they didn't take sched_hotcpu_mutex. Just wake up
7703 spin_lock_irq(&rq
->lock
);
7704 while (!list_empty(&rq
->migration_queue
)) {
7705 struct migration_req
*req
;
7707 req
= list_entry(rq
->migration_queue
.next
,
7708 struct migration_req
, list
);
7709 list_del_init(&req
->list
);
7710 spin_unlock_irq(&rq
->lock
);
7711 complete(&req
->done
);
7712 spin_lock_irq(&rq
->lock
);
7714 spin_unlock_irq(&rq
->lock
);
7718 case CPU_DYING_FROZEN
:
7719 /* Update our root-domain */
7721 spin_lock_irqsave(&rq
->lock
, flags
);
7723 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7726 spin_unlock_irqrestore(&rq
->lock
, flags
);
7734 * Register at high priority so that task migration (migrate_all_tasks)
7735 * happens before everything else. This has to be lower priority than
7736 * the notifier in the perf_event subsystem, though.
7738 static struct notifier_block __cpuinitdata migration_notifier
= {
7739 .notifier_call
= migration_call
,
7743 static int __init
migration_init(void)
7745 void *cpu
= (void *)(long)smp_processor_id();
7748 /* Start one for the boot CPU: */
7749 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7750 BUG_ON(err
== NOTIFY_BAD
);
7751 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7752 register_cpu_notifier(&migration_notifier
);
7756 early_initcall(migration_init
);
7761 #ifdef CONFIG_SCHED_DEBUG
7763 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7764 struct cpumask
*groupmask
)
7766 struct sched_group
*group
= sd
->groups
;
7769 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7770 cpumask_clear(groupmask
);
7772 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7774 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7775 printk("does not load-balance\n");
7777 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7782 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7784 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7785 printk(KERN_ERR
"ERROR: domain->span does not contain "
7788 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7789 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7793 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7797 printk(KERN_ERR
"ERROR: group is NULL\n");
7801 if (!group
->cpu_power
) {
7802 printk(KERN_CONT
"\n");
7803 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7808 if (!cpumask_weight(sched_group_cpus(group
))) {
7809 printk(KERN_CONT
"\n");
7810 printk(KERN_ERR
"ERROR: empty group\n");
7814 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7815 printk(KERN_CONT
"\n");
7816 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7820 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7822 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7824 printk(KERN_CONT
" %s", str
);
7825 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7826 printk(KERN_CONT
" (cpu_power = %d)",
7830 group
= group
->next
;
7831 } while (group
!= sd
->groups
);
7832 printk(KERN_CONT
"\n");
7834 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7835 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7838 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7839 printk(KERN_ERR
"ERROR: parent span is not a superset "
7840 "of domain->span\n");
7844 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7846 cpumask_var_t groupmask
;
7850 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7854 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7856 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7857 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7862 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7869 free_cpumask_var(groupmask
);
7871 #else /* !CONFIG_SCHED_DEBUG */
7872 # define sched_domain_debug(sd, cpu) do { } while (0)
7873 #endif /* CONFIG_SCHED_DEBUG */
7875 static int sd_degenerate(struct sched_domain
*sd
)
7877 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7880 /* Following flags need at least 2 groups */
7881 if (sd
->flags
& (SD_LOAD_BALANCE
|
7882 SD_BALANCE_NEWIDLE
|
7886 SD_SHARE_PKG_RESOURCES
)) {
7887 if (sd
->groups
!= sd
->groups
->next
)
7891 /* Following flags don't use groups */
7892 if (sd
->flags
& (SD_WAKE_AFFINE
))
7899 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7901 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7903 if (sd_degenerate(parent
))
7906 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7909 /* Flags needing groups don't count if only 1 group in parent */
7910 if (parent
->groups
== parent
->groups
->next
) {
7911 pflags
&= ~(SD_LOAD_BALANCE
|
7912 SD_BALANCE_NEWIDLE
|
7916 SD_SHARE_PKG_RESOURCES
);
7917 if (nr_node_ids
== 1)
7918 pflags
&= ~SD_SERIALIZE
;
7920 if (~cflags
& pflags
)
7926 static void free_rootdomain(struct root_domain
*rd
)
7928 cpupri_cleanup(&rd
->cpupri
);
7930 free_cpumask_var(rd
->rto_mask
);
7931 free_cpumask_var(rd
->online
);
7932 free_cpumask_var(rd
->span
);
7936 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7938 struct root_domain
*old_rd
= NULL
;
7939 unsigned long flags
;
7941 spin_lock_irqsave(&rq
->lock
, flags
);
7946 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7949 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7952 * If we dont want to free the old_rt yet then
7953 * set old_rd to NULL to skip the freeing later
7956 if (!atomic_dec_and_test(&old_rd
->refcount
))
7960 atomic_inc(&rd
->refcount
);
7963 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7964 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
7967 spin_unlock_irqrestore(&rq
->lock
, flags
);
7970 free_rootdomain(old_rd
);
7973 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7975 gfp_t gfp
= GFP_KERNEL
;
7977 memset(rd
, 0, sizeof(*rd
));
7982 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7984 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7986 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7989 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7994 free_cpumask_var(rd
->rto_mask
);
7996 free_cpumask_var(rd
->online
);
7998 free_cpumask_var(rd
->span
);
8003 static void init_defrootdomain(void)
8005 init_rootdomain(&def_root_domain
, true);
8007 atomic_set(&def_root_domain
.refcount
, 1);
8010 static struct root_domain
*alloc_rootdomain(void)
8012 struct root_domain
*rd
;
8014 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8018 if (init_rootdomain(rd
, false) != 0) {
8027 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8028 * hold the hotplug lock.
8031 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8033 struct rq
*rq
= cpu_rq(cpu
);
8034 struct sched_domain
*tmp
;
8036 /* Remove the sched domains which do not contribute to scheduling. */
8037 for (tmp
= sd
; tmp
; ) {
8038 struct sched_domain
*parent
= tmp
->parent
;
8042 if (sd_parent_degenerate(tmp
, parent
)) {
8043 tmp
->parent
= parent
->parent
;
8045 parent
->parent
->child
= tmp
;
8050 if (sd
&& sd_degenerate(sd
)) {
8056 sched_domain_debug(sd
, cpu
);
8058 rq_attach_root(rq
, rd
);
8059 rcu_assign_pointer(rq
->sd
, sd
);
8062 /* cpus with isolated domains */
8063 static cpumask_var_t cpu_isolated_map
;
8065 /* Setup the mask of cpus configured for isolated domains */
8066 static int __init
isolated_cpu_setup(char *str
)
8068 cpulist_parse(str
, cpu_isolated_map
);
8072 __setup("isolcpus=", isolated_cpu_setup
);
8075 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8076 * to a function which identifies what group(along with sched group) a CPU
8077 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8078 * (due to the fact that we keep track of groups covered with a struct cpumask).
8080 * init_sched_build_groups will build a circular linked list of the groups
8081 * covered by the given span, and will set each group's ->cpumask correctly,
8082 * and ->cpu_power to 0.
8085 init_sched_build_groups(const struct cpumask
*span
,
8086 const struct cpumask
*cpu_map
,
8087 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8088 struct sched_group
**sg
,
8089 struct cpumask
*tmpmask
),
8090 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8092 struct sched_group
*first
= NULL
, *last
= NULL
;
8095 cpumask_clear(covered
);
8097 for_each_cpu(i
, span
) {
8098 struct sched_group
*sg
;
8099 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8102 if (cpumask_test_cpu(i
, covered
))
8105 cpumask_clear(sched_group_cpus(sg
));
8108 for_each_cpu(j
, span
) {
8109 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8112 cpumask_set_cpu(j
, covered
);
8113 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8124 #define SD_NODES_PER_DOMAIN 16
8129 * find_next_best_node - find the next node to include in a sched_domain
8130 * @node: node whose sched_domain we're building
8131 * @used_nodes: nodes already in the sched_domain
8133 * Find the next node to include in a given scheduling domain. Simply
8134 * finds the closest node not already in the @used_nodes map.
8136 * Should use nodemask_t.
8138 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8140 int i
, n
, val
, min_val
, best_node
= 0;
8144 for (i
= 0; i
< nr_node_ids
; i
++) {
8145 /* Start at @node */
8146 n
= (node
+ i
) % nr_node_ids
;
8148 if (!nr_cpus_node(n
))
8151 /* Skip already used nodes */
8152 if (node_isset(n
, *used_nodes
))
8155 /* Simple min distance search */
8156 val
= node_distance(node
, n
);
8158 if (val
< min_val
) {
8164 node_set(best_node
, *used_nodes
);
8169 * sched_domain_node_span - get a cpumask for a node's sched_domain
8170 * @node: node whose cpumask we're constructing
8171 * @span: resulting cpumask
8173 * Given a node, construct a good cpumask for its sched_domain to span. It
8174 * should be one that prevents unnecessary balancing, but also spreads tasks
8177 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8179 nodemask_t used_nodes
;
8182 cpumask_clear(span
);
8183 nodes_clear(used_nodes
);
8185 cpumask_or(span
, span
, cpumask_of_node(node
));
8186 node_set(node
, used_nodes
);
8188 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8189 int next_node
= find_next_best_node(node
, &used_nodes
);
8191 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8194 #endif /* CONFIG_NUMA */
8196 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8199 * The cpus mask in sched_group and sched_domain hangs off the end.
8201 * ( See the the comments in include/linux/sched.h:struct sched_group
8202 * and struct sched_domain. )
8204 struct static_sched_group
{
8205 struct sched_group sg
;
8206 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8209 struct static_sched_domain
{
8210 struct sched_domain sd
;
8211 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8217 cpumask_var_t domainspan
;
8218 cpumask_var_t covered
;
8219 cpumask_var_t notcovered
;
8221 cpumask_var_t nodemask
;
8222 cpumask_var_t this_sibling_map
;
8223 cpumask_var_t this_core_map
;
8224 cpumask_var_t send_covered
;
8225 cpumask_var_t tmpmask
;
8226 struct sched_group
**sched_group_nodes
;
8227 struct root_domain
*rd
;
8231 sa_sched_groups
= 0,
8236 sa_this_sibling_map
,
8238 sa_sched_group_nodes
,
8248 * SMT sched-domains:
8250 #ifdef CONFIG_SCHED_SMT
8251 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8252 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8255 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8256 struct sched_group
**sg
, struct cpumask
*unused
)
8259 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8262 #endif /* CONFIG_SCHED_SMT */
8265 * multi-core sched-domains:
8267 #ifdef CONFIG_SCHED_MC
8268 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8269 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8270 #endif /* CONFIG_SCHED_MC */
8272 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8274 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8275 struct sched_group
**sg
, struct cpumask
*mask
)
8279 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8280 group
= cpumask_first(mask
);
8282 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8285 #elif defined(CONFIG_SCHED_MC)
8287 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8288 struct sched_group
**sg
, struct cpumask
*unused
)
8291 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8296 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8297 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8300 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8301 struct sched_group
**sg
, struct cpumask
*mask
)
8304 #ifdef CONFIG_SCHED_MC
8305 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8306 group
= cpumask_first(mask
);
8307 #elif defined(CONFIG_SCHED_SMT)
8308 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8309 group
= cpumask_first(mask
);
8314 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8320 * The init_sched_build_groups can't handle what we want to do with node
8321 * groups, so roll our own. Now each node has its own list of groups which
8322 * gets dynamically allocated.
8324 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8325 static struct sched_group
***sched_group_nodes_bycpu
;
8327 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8328 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8330 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8331 struct sched_group
**sg
,
8332 struct cpumask
*nodemask
)
8336 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8337 group
= cpumask_first(nodemask
);
8340 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8344 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8346 struct sched_group
*sg
= group_head
;
8352 for_each_cpu(j
, sched_group_cpus(sg
)) {
8353 struct sched_domain
*sd
;
8355 sd
= &per_cpu(phys_domains
, j
).sd
;
8356 if (j
!= group_first_cpu(sd
->groups
)) {
8358 * Only add "power" once for each
8364 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8367 } while (sg
!= group_head
);
8370 static int build_numa_sched_groups(struct s_data
*d
,
8371 const struct cpumask
*cpu_map
, int num
)
8373 struct sched_domain
*sd
;
8374 struct sched_group
*sg
, *prev
;
8377 cpumask_clear(d
->covered
);
8378 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8379 if (cpumask_empty(d
->nodemask
)) {
8380 d
->sched_group_nodes
[num
] = NULL
;
8384 sched_domain_node_span(num
, d
->domainspan
);
8385 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8387 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8390 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8394 d
->sched_group_nodes
[num
] = sg
;
8396 for_each_cpu(j
, d
->nodemask
) {
8397 sd
= &per_cpu(node_domains
, j
).sd
;
8402 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8404 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8407 for (j
= 0; j
< nr_node_ids
; j
++) {
8408 n
= (num
+ j
) % nr_node_ids
;
8409 cpumask_complement(d
->notcovered
, d
->covered
);
8410 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8411 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8412 if (cpumask_empty(d
->tmpmask
))
8414 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8415 if (cpumask_empty(d
->tmpmask
))
8417 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8421 "Can not alloc domain group for node %d\n", j
);
8425 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8426 sg
->next
= prev
->next
;
8427 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8434 #endif /* CONFIG_NUMA */
8437 /* Free memory allocated for various sched_group structures */
8438 static void free_sched_groups(const struct cpumask
*cpu_map
,
8439 struct cpumask
*nodemask
)
8443 for_each_cpu(cpu
, cpu_map
) {
8444 struct sched_group
**sched_group_nodes
8445 = sched_group_nodes_bycpu
[cpu
];
8447 if (!sched_group_nodes
)
8450 for (i
= 0; i
< nr_node_ids
; i
++) {
8451 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8453 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8454 if (cpumask_empty(nodemask
))
8464 if (oldsg
!= sched_group_nodes
[i
])
8467 kfree(sched_group_nodes
);
8468 sched_group_nodes_bycpu
[cpu
] = NULL
;
8471 #else /* !CONFIG_NUMA */
8472 static void free_sched_groups(const struct cpumask
*cpu_map
,
8473 struct cpumask
*nodemask
)
8476 #endif /* CONFIG_NUMA */
8479 * Initialize sched groups cpu_power.
8481 * cpu_power indicates the capacity of sched group, which is used while
8482 * distributing the load between different sched groups in a sched domain.
8483 * Typically cpu_power for all the groups in a sched domain will be same unless
8484 * there are asymmetries in the topology. If there are asymmetries, group
8485 * having more cpu_power will pickup more load compared to the group having
8488 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8490 struct sched_domain
*child
;
8491 struct sched_group
*group
;
8495 WARN_ON(!sd
|| !sd
->groups
);
8497 if (cpu
!= group_first_cpu(sd
->groups
))
8502 sd
->groups
->cpu_power
= 0;
8505 power
= SCHED_LOAD_SCALE
;
8506 weight
= cpumask_weight(sched_domain_span(sd
));
8508 * SMT siblings share the power of a single core.
8509 * Usually multiple threads get a better yield out of
8510 * that one core than a single thread would have,
8511 * reflect that in sd->smt_gain.
8513 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8514 power
*= sd
->smt_gain
;
8516 power
>>= SCHED_LOAD_SHIFT
;
8518 sd
->groups
->cpu_power
+= power
;
8523 * Add cpu_power of each child group to this groups cpu_power.
8525 group
= child
->groups
;
8527 sd
->groups
->cpu_power
+= group
->cpu_power
;
8528 group
= group
->next
;
8529 } while (group
!= child
->groups
);
8533 * Initializers for schedule domains
8534 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8537 #ifdef CONFIG_SCHED_DEBUG
8538 # define SD_INIT_NAME(sd, type) sd->name = #type
8540 # define SD_INIT_NAME(sd, type) do { } while (0)
8543 #define SD_INIT(sd, type) sd_init_##type(sd)
8545 #define SD_INIT_FUNC(type) \
8546 static noinline void sd_init_##type(struct sched_domain *sd) \
8548 memset(sd, 0, sizeof(*sd)); \
8549 *sd = SD_##type##_INIT; \
8550 sd->level = SD_LV_##type; \
8551 SD_INIT_NAME(sd, type); \
8556 SD_INIT_FUNC(ALLNODES
)
8559 #ifdef CONFIG_SCHED_SMT
8560 SD_INIT_FUNC(SIBLING
)
8562 #ifdef CONFIG_SCHED_MC
8566 static int default_relax_domain_level
= -1;
8568 static int __init
setup_relax_domain_level(char *str
)
8572 val
= simple_strtoul(str
, NULL
, 0);
8573 if (val
< SD_LV_MAX
)
8574 default_relax_domain_level
= val
;
8578 __setup("relax_domain_level=", setup_relax_domain_level
);
8580 static void set_domain_attribute(struct sched_domain
*sd
,
8581 struct sched_domain_attr
*attr
)
8585 if (!attr
|| attr
->relax_domain_level
< 0) {
8586 if (default_relax_domain_level
< 0)
8589 request
= default_relax_domain_level
;
8591 request
= attr
->relax_domain_level
;
8592 if (request
< sd
->level
) {
8593 /* turn off idle balance on this domain */
8594 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8596 /* turn on idle balance on this domain */
8597 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8601 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8602 const struct cpumask
*cpu_map
)
8605 case sa_sched_groups
:
8606 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8607 d
->sched_group_nodes
= NULL
;
8609 free_rootdomain(d
->rd
); /* fall through */
8611 free_cpumask_var(d
->tmpmask
); /* fall through */
8612 case sa_send_covered
:
8613 free_cpumask_var(d
->send_covered
); /* fall through */
8614 case sa_this_core_map
:
8615 free_cpumask_var(d
->this_core_map
); /* fall through */
8616 case sa_this_sibling_map
:
8617 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8619 free_cpumask_var(d
->nodemask
); /* fall through */
8620 case sa_sched_group_nodes
:
8622 kfree(d
->sched_group_nodes
); /* fall through */
8624 free_cpumask_var(d
->notcovered
); /* fall through */
8626 free_cpumask_var(d
->covered
); /* fall through */
8628 free_cpumask_var(d
->domainspan
); /* fall through */
8635 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8636 const struct cpumask
*cpu_map
)
8639 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8641 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8642 return sa_domainspan
;
8643 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8645 /* Allocate the per-node list of sched groups */
8646 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8647 sizeof(struct sched_group
*), GFP_KERNEL
);
8648 if (!d
->sched_group_nodes
) {
8649 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8650 return sa_notcovered
;
8652 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8654 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8655 return sa_sched_group_nodes
;
8656 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8658 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8659 return sa_this_sibling_map
;
8660 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8661 return sa_this_core_map
;
8662 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8663 return sa_send_covered
;
8664 d
->rd
= alloc_rootdomain();
8666 printk(KERN_WARNING
"Cannot alloc root domain\n");
8669 return sa_rootdomain
;
8672 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8673 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8675 struct sched_domain
*sd
= NULL
;
8677 struct sched_domain
*parent
;
8680 if (cpumask_weight(cpu_map
) >
8681 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8682 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8683 SD_INIT(sd
, ALLNODES
);
8684 set_domain_attribute(sd
, attr
);
8685 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8686 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8691 sd
= &per_cpu(node_domains
, i
).sd
;
8693 set_domain_attribute(sd
, attr
);
8694 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8695 sd
->parent
= parent
;
8698 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8703 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8704 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8705 struct sched_domain
*parent
, int i
)
8707 struct sched_domain
*sd
;
8708 sd
= &per_cpu(phys_domains
, i
).sd
;
8710 set_domain_attribute(sd
, attr
);
8711 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8712 sd
->parent
= parent
;
8715 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8719 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8720 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8721 struct sched_domain
*parent
, int i
)
8723 struct sched_domain
*sd
= parent
;
8724 #ifdef CONFIG_SCHED_MC
8725 sd
= &per_cpu(core_domains
, i
).sd
;
8727 set_domain_attribute(sd
, attr
);
8728 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8729 sd
->parent
= parent
;
8731 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8736 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8737 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8738 struct sched_domain
*parent
, int i
)
8740 struct sched_domain
*sd
= parent
;
8741 #ifdef CONFIG_SCHED_SMT
8742 sd
= &per_cpu(cpu_domains
, i
).sd
;
8743 SD_INIT(sd
, SIBLING
);
8744 set_domain_attribute(sd
, attr
);
8745 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8746 sd
->parent
= parent
;
8748 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8753 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8754 const struct cpumask
*cpu_map
, int cpu
)
8757 #ifdef CONFIG_SCHED_SMT
8758 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8759 cpumask_and(d
->this_sibling_map
, cpu_map
,
8760 topology_thread_cpumask(cpu
));
8761 if (cpu
== cpumask_first(d
->this_sibling_map
))
8762 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8764 d
->send_covered
, d
->tmpmask
);
8767 #ifdef CONFIG_SCHED_MC
8768 case SD_LV_MC
: /* set up multi-core groups */
8769 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8770 if (cpu
== cpumask_first(d
->this_core_map
))
8771 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8773 d
->send_covered
, d
->tmpmask
);
8776 case SD_LV_CPU
: /* set up physical groups */
8777 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8778 if (!cpumask_empty(d
->nodemask
))
8779 init_sched_build_groups(d
->nodemask
, cpu_map
,
8781 d
->send_covered
, d
->tmpmask
);
8784 case SD_LV_ALLNODES
:
8785 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8786 d
->send_covered
, d
->tmpmask
);
8795 * Build sched domains for a given set of cpus and attach the sched domains
8796 * to the individual cpus
8798 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8799 struct sched_domain_attr
*attr
)
8801 enum s_alloc alloc_state
= sa_none
;
8803 struct sched_domain
*sd
;
8809 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8810 if (alloc_state
!= sa_rootdomain
)
8812 alloc_state
= sa_sched_groups
;
8815 * Set up domains for cpus specified by the cpu_map.
8817 for_each_cpu(i
, cpu_map
) {
8818 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8821 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8822 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8823 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8824 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8827 for_each_cpu(i
, cpu_map
) {
8828 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8829 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8832 /* Set up physical groups */
8833 for (i
= 0; i
< nr_node_ids
; i
++)
8834 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8837 /* Set up node groups */
8839 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8841 for (i
= 0; i
< nr_node_ids
; i
++)
8842 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8846 /* Calculate CPU power for physical packages and nodes */
8847 #ifdef CONFIG_SCHED_SMT
8848 for_each_cpu(i
, cpu_map
) {
8849 sd
= &per_cpu(cpu_domains
, i
).sd
;
8850 init_sched_groups_power(i
, sd
);
8853 #ifdef CONFIG_SCHED_MC
8854 for_each_cpu(i
, cpu_map
) {
8855 sd
= &per_cpu(core_domains
, i
).sd
;
8856 init_sched_groups_power(i
, sd
);
8860 for_each_cpu(i
, cpu_map
) {
8861 sd
= &per_cpu(phys_domains
, i
).sd
;
8862 init_sched_groups_power(i
, sd
);
8866 for (i
= 0; i
< nr_node_ids
; i
++)
8867 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8869 if (d
.sd_allnodes
) {
8870 struct sched_group
*sg
;
8872 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8874 init_numa_sched_groups_power(sg
);
8878 /* Attach the domains */
8879 for_each_cpu(i
, cpu_map
) {
8880 #ifdef CONFIG_SCHED_SMT
8881 sd
= &per_cpu(cpu_domains
, i
).sd
;
8882 #elif defined(CONFIG_SCHED_MC)
8883 sd
= &per_cpu(core_domains
, i
).sd
;
8885 sd
= &per_cpu(phys_domains
, i
).sd
;
8887 cpu_attach_domain(sd
, d
.rd
, i
);
8890 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8891 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8895 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8899 static int build_sched_domains(const struct cpumask
*cpu_map
)
8901 return __build_sched_domains(cpu_map
, NULL
);
8904 static struct cpumask
*doms_cur
; /* current sched domains */
8905 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8906 static struct sched_domain_attr
*dattr_cur
;
8907 /* attribues of custom domains in 'doms_cur' */
8910 * Special case: If a kmalloc of a doms_cur partition (array of
8911 * cpumask) fails, then fallback to a single sched domain,
8912 * as determined by the single cpumask fallback_doms.
8914 static cpumask_var_t fallback_doms
;
8917 * arch_update_cpu_topology lets virtualized architectures update the
8918 * cpu core maps. It is supposed to return 1 if the topology changed
8919 * or 0 if it stayed the same.
8921 int __attribute__((weak
)) arch_update_cpu_topology(void)
8927 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8928 * For now this just excludes isolated cpus, but could be used to
8929 * exclude other special cases in the future.
8931 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8935 arch_update_cpu_topology();
8937 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8939 doms_cur
= fallback_doms
;
8940 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8942 err
= build_sched_domains(doms_cur
);
8943 register_sched_domain_sysctl();
8948 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8949 struct cpumask
*tmpmask
)
8951 free_sched_groups(cpu_map
, tmpmask
);
8955 * Detach sched domains from a group of cpus specified in cpu_map
8956 * These cpus will now be attached to the NULL domain
8958 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8960 /* Save because hotplug lock held. */
8961 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8964 for_each_cpu(i
, cpu_map
)
8965 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8966 synchronize_sched();
8967 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8970 /* handle null as "default" */
8971 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8972 struct sched_domain_attr
*new, int idx_new
)
8974 struct sched_domain_attr tmp
;
8981 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8982 new ? (new + idx_new
) : &tmp
,
8983 sizeof(struct sched_domain_attr
));
8987 * Partition sched domains as specified by the 'ndoms_new'
8988 * cpumasks in the array doms_new[] of cpumasks. This compares
8989 * doms_new[] to the current sched domain partitioning, doms_cur[].
8990 * It destroys each deleted domain and builds each new domain.
8992 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8993 * The masks don't intersect (don't overlap.) We should setup one
8994 * sched domain for each mask. CPUs not in any of the cpumasks will
8995 * not be load balanced. If the same cpumask appears both in the
8996 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8999 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9000 * ownership of it and will kfree it when done with it. If the caller
9001 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9002 * ndoms_new == 1, and partition_sched_domains() will fallback to
9003 * the single partition 'fallback_doms', it also forces the domains
9006 * If doms_new == NULL it will be replaced with cpu_online_mask.
9007 * ndoms_new == 0 is a special case for destroying existing domains,
9008 * and it will not create the default domain.
9010 * Call with hotplug lock held
9012 /* FIXME: Change to struct cpumask *doms_new[] */
9013 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9014 struct sched_domain_attr
*dattr_new
)
9019 mutex_lock(&sched_domains_mutex
);
9021 /* always unregister in case we don't destroy any domains */
9022 unregister_sched_domain_sysctl();
9024 /* Let architecture update cpu core mappings. */
9025 new_topology
= arch_update_cpu_topology();
9027 n
= doms_new
? ndoms_new
: 0;
9029 /* Destroy deleted domains */
9030 for (i
= 0; i
< ndoms_cur
; i
++) {
9031 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9032 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9033 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9036 /* no match - a current sched domain not in new doms_new[] */
9037 detach_destroy_domains(doms_cur
+ i
);
9042 if (doms_new
== NULL
) {
9044 doms_new
= fallback_doms
;
9045 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
9046 WARN_ON_ONCE(dattr_new
);
9049 /* Build new domains */
9050 for (i
= 0; i
< ndoms_new
; i
++) {
9051 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9052 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9053 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9056 /* no match - add a new doms_new */
9057 __build_sched_domains(doms_new
+ i
,
9058 dattr_new
? dattr_new
+ i
: NULL
);
9063 /* Remember the new sched domains */
9064 if (doms_cur
!= fallback_doms
)
9066 kfree(dattr_cur
); /* kfree(NULL) is safe */
9067 doms_cur
= doms_new
;
9068 dattr_cur
= dattr_new
;
9069 ndoms_cur
= ndoms_new
;
9071 register_sched_domain_sysctl();
9073 mutex_unlock(&sched_domains_mutex
);
9076 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9077 static void arch_reinit_sched_domains(void)
9081 /* Destroy domains first to force the rebuild */
9082 partition_sched_domains(0, NULL
, NULL
);
9084 rebuild_sched_domains();
9088 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9090 unsigned int level
= 0;
9092 if (sscanf(buf
, "%u", &level
) != 1)
9096 * level is always be positive so don't check for
9097 * level < POWERSAVINGS_BALANCE_NONE which is 0
9098 * What happens on 0 or 1 byte write,
9099 * need to check for count as well?
9102 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9106 sched_smt_power_savings
= level
;
9108 sched_mc_power_savings
= level
;
9110 arch_reinit_sched_domains();
9115 #ifdef CONFIG_SCHED_MC
9116 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9119 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9121 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9122 const char *buf
, size_t count
)
9124 return sched_power_savings_store(buf
, count
, 0);
9126 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9127 sched_mc_power_savings_show
,
9128 sched_mc_power_savings_store
);
9131 #ifdef CONFIG_SCHED_SMT
9132 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9135 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9137 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9138 const char *buf
, size_t count
)
9140 return sched_power_savings_store(buf
, count
, 1);
9142 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9143 sched_smt_power_savings_show
,
9144 sched_smt_power_savings_store
);
9147 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9151 #ifdef CONFIG_SCHED_SMT
9153 err
= sysfs_create_file(&cls
->kset
.kobj
,
9154 &attr_sched_smt_power_savings
.attr
);
9156 #ifdef CONFIG_SCHED_MC
9157 if (!err
&& mc_capable())
9158 err
= sysfs_create_file(&cls
->kset
.kobj
,
9159 &attr_sched_mc_power_savings
.attr
);
9163 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9165 #ifndef CONFIG_CPUSETS
9167 * Add online and remove offline CPUs from the scheduler domains.
9168 * When cpusets are enabled they take over this function.
9170 static int update_sched_domains(struct notifier_block
*nfb
,
9171 unsigned long action
, void *hcpu
)
9175 case CPU_ONLINE_FROZEN
:
9177 case CPU_DEAD_FROZEN
:
9178 partition_sched_domains(1, NULL
, NULL
);
9187 static int update_runtime(struct notifier_block
*nfb
,
9188 unsigned long action
, void *hcpu
)
9190 int cpu
= (int)(long)hcpu
;
9193 case CPU_DOWN_PREPARE
:
9194 case CPU_DOWN_PREPARE_FROZEN
:
9195 disable_runtime(cpu_rq(cpu
));
9198 case CPU_DOWN_FAILED
:
9199 case CPU_DOWN_FAILED_FROZEN
:
9201 case CPU_ONLINE_FROZEN
:
9202 enable_runtime(cpu_rq(cpu
));
9210 void __init
sched_init_smp(void)
9212 cpumask_var_t non_isolated_cpus
;
9214 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9215 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9217 #if defined(CONFIG_NUMA)
9218 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9220 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9223 mutex_lock(&sched_domains_mutex
);
9224 arch_init_sched_domains(cpu_online_mask
);
9225 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9226 if (cpumask_empty(non_isolated_cpus
))
9227 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9228 mutex_unlock(&sched_domains_mutex
);
9231 #ifndef CONFIG_CPUSETS
9232 /* XXX: Theoretical race here - CPU may be hotplugged now */
9233 hotcpu_notifier(update_sched_domains
, 0);
9236 /* RT runtime code needs to handle some hotplug events */
9237 hotcpu_notifier(update_runtime
, 0);
9241 /* Move init over to a non-isolated CPU */
9242 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9244 sched_init_granularity();
9245 free_cpumask_var(non_isolated_cpus
);
9247 init_sched_rt_class();
9250 void __init
sched_init_smp(void)
9252 sched_init_granularity();
9254 #endif /* CONFIG_SMP */
9256 const_debug
unsigned int sysctl_timer_migration
= 1;
9258 int in_sched_functions(unsigned long addr
)
9260 return in_lock_functions(addr
) ||
9261 (addr
>= (unsigned long)__sched_text_start
9262 && addr
< (unsigned long)__sched_text_end
);
9265 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9267 cfs_rq
->tasks_timeline
= RB_ROOT
;
9268 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9269 #ifdef CONFIG_FAIR_GROUP_SCHED
9272 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9275 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9277 struct rt_prio_array
*array
;
9280 array
= &rt_rq
->active
;
9281 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9282 INIT_LIST_HEAD(array
->queue
+ i
);
9283 __clear_bit(i
, array
->bitmap
);
9285 /* delimiter for bitsearch: */
9286 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9288 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9289 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9291 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9295 rt_rq
->rt_nr_migratory
= 0;
9296 rt_rq
->overloaded
= 0;
9297 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9301 rt_rq
->rt_throttled
= 0;
9302 rt_rq
->rt_runtime
= 0;
9303 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9305 #ifdef CONFIG_RT_GROUP_SCHED
9306 rt_rq
->rt_nr_boosted
= 0;
9311 #ifdef CONFIG_FAIR_GROUP_SCHED
9312 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9313 struct sched_entity
*se
, int cpu
, int add
,
9314 struct sched_entity
*parent
)
9316 struct rq
*rq
= cpu_rq(cpu
);
9317 tg
->cfs_rq
[cpu
] = cfs_rq
;
9318 init_cfs_rq(cfs_rq
, rq
);
9321 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9324 /* se could be NULL for init_task_group */
9329 se
->cfs_rq
= &rq
->cfs
;
9331 se
->cfs_rq
= parent
->my_q
;
9334 se
->load
.weight
= tg
->shares
;
9335 se
->load
.inv_weight
= 0;
9336 se
->parent
= parent
;
9340 #ifdef CONFIG_RT_GROUP_SCHED
9341 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9342 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9343 struct sched_rt_entity
*parent
)
9345 struct rq
*rq
= cpu_rq(cpu
);
9347 tg
->rt_rq
[cpu
] = rt_rq
;
9348 init_rt_rq(rt_rq
, rq
);
9350 rt_rq
->rt_se
= rt_se
;
9351 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9353 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9355 tg
->rt_se
[cpu
] = rt_se
;
9360 rt_se
->rt_rq
= &rq
->rt
;
9362 rt_se
->rt_rq
= parent
->my_q
;
9364 rt_se
->my_q
= rt_rq
;
9365 rt_se
->parent
= parent
;
9366 INIT_LIST_HEAD(&rt_se
->run_list
);
9370 void __init
sched_init(void)
9373 unsigned long alloc_size
= 0, ptr
;
9375 #ifdef CONFIG_FAIR_GROUP_SCHED
9376 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9378 #ifdef CONFIG_RT_GROUP_SCHED
9379 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9381 #ifdef CONFIG_USER_SCHED
9384 #ifdef CONFIG_CPUMASK_OFFSTACK
9385 alloc_size
+= num_possible_cpus() * cpumask_size();
9388 * As sched_init() is called before page_alloc is setup,
9389 * we use alloc_bootmem().
9392 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9394 #ifdef CONFIG_FAIR_GROUP_SCHED
9395 init_task_group
.se
= (struct sched_entity
**)ptr
;
9396 ptr
+= nr_cpu_ids
* sizeof(void **);
9398 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9399 ptr
+= nr_cpu_ids
* sizeof(void **);
9401 #ifdef CONFIG_USER_SCHED
9402 root_task_group
.se
= (struct sched_entity
**)ptr
;
9403 ptr
+= nr_cpu_ids
* sizeof(void **);
9405 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9406 ptr
+= nr_cpu_ids
* sizeof(void **);
9407 #endif /* CONFIG_USER_SCHED */
9408 #endif /* CONFIG_FAIR_GROUP_SCHED */
9409 #ifdef CONFIG_RT_GROUP_SCHED
9410 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9411 ptr
+= nr_cpu_ids
* sizeof(void **);
9413 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9414 ptr
+= nr_cpu_ids
* sizeof(void **);
9416 #ifdef CONFIG_USER_SCHED
9417 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9418 ptr
+= nr_cpu_ids
* sizeof(void **);
9420 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9421 ptr
+= nr_cpu_ids
* sizeof(void **);
9422 #endif /* CONFIG_USER_SCHED */
9423 #endif /* CONFIG_RT_GROUP_SCHED */
9424 #ifdef CONFIG_CPUMASK_OFFSTACK
9425 for_each_possible_cpu(i
) {
9426 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9427 ptr
+= cpumask_size();
9429 #endif /* CONFIG_CPUMASK_OFFSTACK */
9433 init_defrootdomain();
9436 init_rt_bandwidth(&def_rt_bandwidth
,
9437 global_rt_period(), global_rt_runtime());
9439 #ifdef CONFIG_RT_GROUP_SCHED
9440 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9441 global_rt_period(), global_rt_runtime());
9442 #ifdef CONFIG_USER_SCHED
9443 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9444 global_rt_period(), RUNTIME_INF
);
9445 #endif /* CONFIG_USER_SCHED */
9446 #endif /* CONFIG_RT_GROUP_SCHED */
9448 #ifdef CONFIG_GROUP_SCHED
9449 list_add(&init_task_group
.list
, &task_groups
);
9450 INIT_LIST_HEAD(&init_task_group
.children
);
9452 #ifdef CONFIG_USER_SCHED
9453 INIT_LIST_HEAD(&root_task_group
.children
);
9454 init_task_group
.parent
= &root_task_group
;
9455 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9456 #endif /* CONFIG_USER_SCHED */
9457 #endif /* CONFIG_GROUP_SCHED */
9459 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9460 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9461 __alignof__(unsigned long));
9463 for_each_possible_cpu(i
) {
9467 spin_lock_init(&rq
->lock
);
9469 rq
->calc_load_active
= 0;
9470 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9471 init_cfs_rq(&rq
->cfs
, rq
);
9472 init_rt_rq(&rq
->rt
, rq
);
9473 #ifdef CONFIG_FAIR_GROUP_SCHED
9474 init_task_group
.shares
= init_task_group_load
;
9475 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9476 #ifdef CONFIG_CGROUP_SCHED
9478 * How much cpu bandwidth does init_task_group get?
9480 * In case of task-groups formed thr' the cgroup filesystem, it
9481 * gets 100% of the cpu resources in the system. This overall
9482 * system cpu resource is divided among the tasks of
9483 * init_task_group and its child task-groups in a fair manner,
9484 * based on each entity's (task or task-group's) weight
9485 * (se->load.weight).
9487 * In other words, if init_task_group has 10 tasks of weight
9488 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9489 * then A0's share of the cpu resource is:
9491 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9493 * We achieve this by letting init_task_group's tasks sit
9494 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9496 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9497 #elif defined CONFIG_USER_SCHED
9498 root_task_group
.shares
= NICE_0_LOAD
;
9499 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9501 * In case of task-groups formed thr' the user id of tasks,
9502 * init_task_group represents tasks belonging to root user.
9503 * Hence it forms a sibling of all subsequent groups formed.
9504 * In this case, init_task_group gets only a fraction of overall
9505 * system cpu resource, based on the weight assigned to root
9506 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9507 * by letting tasks of init_task_group sit in a separate cfs_rq
9508 * (init_tg_cfs_rq) and having one entity represent this group of
9509 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9511 init_tg_cfs_entry(&init_task_group
,
9512 &per_cpu(init_tg_cfs_rq
, i
),
9513 &per_cpu(init_sched_entity
, i
), i
, 1,
9514 root_task_group
.se
[i
]);
9517 #endif /* CONFIG_FAIR_GROUP_SCHED */
9519 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9520 #ifdef CONFIG_RT_GROUP_SCHED
9521 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9522 #ifdef CONFIG_CGROUP_SCHED
9523 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9524 #elif defined CONFIG_USER_SCHED
9525 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9526 init_tg_rt_entry(&init_task_group
,
9527 &per_cpu(init_rt_rq
, i
),
9528 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9529 root_task_group
.rt_se
[i
]);
9533 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9534 rq
->cpu_load
[j
] = 0;
9538 rq
->post_schedule
= 0;
9539 rq
->active_balance
= 0;
9540 rq
->next_balance
= jiffies
;
9544 rq
->migration_thread
= NULL
;
9546 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9547 INIT_LIST_HEAD(&rq
->migration_queue
);
9548 rq_attach_root(rq
, &def_root_domain
);
9551 atomic_set(&rq
->nr_iowait
, 0);
9554 set_load_weight(&init_task
);
9556 #ifdef CONFIG_PREEMPT_NOTIFIERS
9557 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9561 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9564 #ifdef CONFIG_RT_MUTEXES
9565 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9569 * The boot idle thread does lazy MMU switching as well:
9571 atomic_inc(&init_mm
.mm_count
);
9572 enter_lazy_tlb(&init_mm
, current
);
9575 * Make us the idle thread. Technically, schedule() should not be
9576 * called from this thread, however somewhere below it might be,
9577 * but because we are the idle thread, we just pick up running again
9578 * when this runqueue becomes "idle".
9580 init_idle(current
, smp_processor_id());
9582 calc_load_update
= jiffies
+ LOAD_FREQ
;
9585 * During early bootup we pretend to be a normal task:
9587 current
->sched_class
= &fair_sched_class
;
9589 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9590 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9593 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9594 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9596 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9601 scheduler_running
= 1;
9604 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9605 static inline int preempt_count_equals(int preempt_offset
)
9607 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9609 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9612 void __might_sleep(char *file
, int line
, int preempt_offset
)
9615 static unsigned long prev_jiffy
; /* ratelimiting */
9617 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9618 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9620 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9622 prev_jiffy
= jiffies
;
9625 "BUG: sleeping function called from invalid context at %s:%d\n",
9628 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9629 in_atomic(), irqs_disabled(),
9630 current
->pid
, current
->comm
);
9632 debug_show_held_locks(current
);
9633 if (irqs_disabled())
9634 print_irqtrace_events(current
);
9638 EXPORT_SYMBOL(__might_sleep
);
9641 #ifdef CONFIG_MAGIC_SYSRQ
9642 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9646 update_rq_clock(rq
);
9647 on_rq
= p
->se
.on_rq
;
9649 deactivate_task(rq
, p
, 0);
9650 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9652 activate_task(rq
, p
, 0);
9653 resched_task(rq
->curr
);
9657 void normalize_rt_tasks(void)
9659 struct task_struct
*g
, *p
;
9660 unsigned long flags
;
9663 read_lock_irqsave(&tasklist_lock
, flags
);
9664 do_each_thread(g
, p
) {
9666 * Only normalize user tasks:
9671 p
->se
.exec_start
= 0;
9672 #ifdef CONFIG_SCHEDSTATS
9673 p
->se
.wait_start
= 0;
9674 p
->se
.sleep_start
= 0;
9675 p
->se
.block_start
= 0;
9680 * Renice negative nice level userspace
9683 if (TASK_NICE(p
) < 0 && p
->mm
)
9684 set_user_nice(p
, 0);
9688 spin_lock(&p
->pi_lock
);
9689 rq
= __task_rq_lock(p
);
9691 normalize_task(rq
, p
);
9693 __task_rq_unlock(rq
);
9694 spin_unlock(&p
->pi_lock
);
9695 } while_each_thread(g
, p
);
9697 read_unlock_irqrestore(&tasklist_lock
, flags
);
9700 #endif /* CONFIG_MAGIC_SYSRQ */
9704 * These functions are only useful for the IA64 MCA handling.
9706 * They can only be called when the whole system has been
9707 * stopped - every CPU needs to be quiescent, and no scheduling
9708 * activity can take place. Using them for anything else would
9709 * be a serious bug, and as a result, they aren't even visible
9710 * under any other configuration.
9714 * curr_task - return the current task for a given cpu.
9715 * @cpu: the processor in question.
9717 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9719 struct task_struct
*curr_task(int cpu
)
9721 return cpu_curr(cpu
);
9725 * set_curr_task - set the current task for a given cpu.
9726 * @cpu: the processor in question.
9727 * @p: the task pointer to set.
9729 * Description: This function must only be used when non-maskable interrupts
9730 * are serviced on a separate stack. It allows the architecture to switch the
9731 * notion of the current task on a cpu in a non-blocking manner. This function
9732 * must be called with all CPU's synchronized, and interrupts disabled, the
9733 * and caller must save the original value of the current task (see
9734 * curr_task() above) and restore that value before reenabling interrupts and
9735 * re-starting the system.
9737 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9739 void set_curr_task(int cpu
, struct task_struct
*p
)
9746 #ifdef CONFIG_FAIR_GROUP_SCHED
9747 static void free_fair_sched_group(struct task_group
*tg
)
9751 for_each_possible_cpu(i
) {
9753 kfree(tg
->cfs_rq
[i
]);
9763 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9765 struct cfs_rq
*cfs_rq
;
9766 struct sched_entity
*se
;
9770 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9773 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9777 tg
->shares
= NICE_0_LOAD
;
9779 for_each_possible_cpu(i
) {
9782 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9783 GFP_KERNEL
, cpu_to_node(i
));
9787 se
= kzalloc_node(sizeof(struct sched_entity
),
9788 GFP_KERNEL
, cpu_to_node(i
));
9792 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9801 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9803 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9804 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9807 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9809 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9811 #else /* !CONFG_FAIR_GROUP_SCHED */
9812 static inline void free_fair_sched_group(struct task_group
*tg
)
9817 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9822 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9826 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9829 #endif /* CONFIG_FAIR_GROUP_SCHED */
9831 #ifdef CONFIG_RT_GROUP_SCHED
9832 static void free_rt_sched_group(struct task_group
*tg
)
9836 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9838 for_each_possible_cpu(i
) {
9840 kfree(tg
->rt_rq
[i
]);
9842 kfree(tg
->rt_se
[i
]);
9850 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9852 struct rt_rq
*rt_rq
;
9853 struct sched_rt_entity
*rt_se
;
9857 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9860 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9864 init_rt_bandwidth(&tg
->rt_bandwidth
,
9865 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9867 for_each_possible_cpu(i
) {
9870 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9871 GFP_KERNEL
, cpu_to_node(i
));
9875 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9876 GFP_KERNEL
, cpu_to_node(i
));
9880 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9889 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9891 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9892 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9895 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9897 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9899 #else /* !CONFIG_RT_GROUP_SCHED */
9900 static inline void free_rt_sched_group(struct task_group
*tg
)
9905 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9910 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9914 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9917 #endif /* CONFIG_RT_GROUP_SCHED */
9919 #ifdef CONFIG_GROUP_SCHED
9920 static void free_sched_group(struct task_group
*tg
)
9922 free_fair_sched_group(tg
);
9923 free_rt_sched_group(tg
);
9927 /* allocate runqueue etc for a new task group */
9928 struct task_group
*sched_create_group(struct task_group
*parent
)
9930 struct task_group
*tg
;
9931 unsigned long flags
;
9934 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9936 return ERR_PTR(-ENOMEM
);
9938 if (!alloc_fair_sched_group(tg
, parent
))
9941 if (!alloc_rt_sched_group(tg
, parent
))
9944 spin_lock_irqsave(&task_group_lock
, flags
);
9945 for_each_possible_cpu(i
) {
9946 register_fair_sched_group(tg
, i
);
9947 register_rt_sched_group(tg
, i
);
9949 list_add_rcu(&tg
->list
, &task_groups
);
9951 WARN_ON(!parent
); /* root should already exist */
9953 tg
->parent
= parent
;
9954 INIT_LIST_HEAD(&tg
->children
);
9955 list_add_rcu(&tg
->siblings
, &parent
->children
);
9956 spin_unlock_irqrestore(&task_group_lock
, flags
);
9961 free_sched_group(tg
);
9962 return ERR_PTR(-ENOMEM
);
9965 /* rcu callback to free various structures associated with a task group */
9966 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9968 /* now it should be safe to free those cfs_rqs */
9969 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9972 /* Destroy runqueue etc associated with a task group */
9973 void sched_destroy_group(struct task_group
*tg
)
9975 unsigned long flags
;
9978 spin_lock_irqsave(&task_group_lock
, flags
);
9979 for_each_possible_cpu(i
) {
9980 unregister_fair_sched_group(tg
, i
);
9981 unregister_rt_sched_group(tg
, i
);
9983 list_del_rcu(&tg
->list
);
9984 list_del_rcu(&tg
->siblings
);
9985 spin_unlock_irqrestore(&task_group_lock
, flags
);
9987 /* wait for possible concurrent references to cfs_rqs complete */
9988 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9991 /* change task's runqueue when it moves between groups.
9992 * The caller of this function should have put the task in its new group
9993 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9994 * reflect its new group.
9996 void sched_move_task(struct task_struct
*tsk
)
9999 unsigned long flags
;
10002 rq
= task_rq_lock(tsk
, &flags
);
10004 update_rq_clock(rq
);
10006 running
= task_current(rq
, tsk
);
10007 on_rq
= tsk
->se
.on_rq
;
10010 dequeue_task(rq
, tsk
, 0);
10011 if (unlikely(running
))
10012 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10014 set_task_rq(tsk
, task_cpu(tsk
));
10016 #ifdef CONFIG_FAIR_GROUP_SCHED
10017 if (tsk
->sched_class
->moved_group
)
10018 tsk
->sched_class
->moved_group(tsk
);
10021 if (unlikely(running
))
10022 tsk
->sched_class
->set_curr_task(rq
);
10024 enqueue_task(rq
, tsk
, 0);
10026 task_rq_unlock(rq
, &flags
);
10028 #endif /* CONFIG_GROUP_SCHED */
10030 #ifdef CONFIG_FAIR_GROUP_SCHED
10031 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10033 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10038 dequeue_entity(cfs_rq
, se
, 0);
10040 se
->load
.weight
= shares
;
10041 se
->load
.inv_weight
= 0;
10044 enqueue_entity(cfs_rq
, se
, 0);
10047 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10049 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10050 struct rq
*rq
= cfs_rq
->rq
;
10051 unsigned long flags
;
10053 spin_lock_irqsave(&rq
->lock
, flags
);
10054 __set_se_shares(se
, shares
);
10055 spin_unlock_irqrestore(&rq
->lock
, flags
);
10058 static DEFINE_MUTEX(shares_mutex
);
10060 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10063 unsigned long flags
;
10066 * We can't change the weight of the root cgroup.
10071 if (shares
< MIN_SHARES
)
10072 shares
= MIN_SHARES
;
10073 else if (shares
> MAX_SHARES
)
10074 shares
= MAX_SHARES
;
10076 mutex_lock(&shares_mutex
);
10077 if (tg
->shares
== shares
)
10080 spin_lock_irqsave(&task_group_lock
, flags
);
10081 for_each_possible_cpu(i
)
10082 unregister_fair_sched_group(tg
, i
);
10083 list_del_rcu(&tg
->siblings
);
10084 spin_unlock_irqrestore(&task_group_lock
, flags
);
10086 /* wait for any ongoing reference to this group to finish */
10087 synchronize_sched();
10090 * Now we are free to modify the group's share on each cpu
10091 * w/o tripping rebalance_share or load_balance_fair.
10093 tg
->shares
= shares
;
10094 for_each_possible_cpu(i
) {
10096 * force a rebalance
10098 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10099 set_se_shares(tg
->se
[i
], shares
);
10103 * Enable load balance activity on this group, by inserting it back on
10104 * each cpu's rq->leaf_cfs_rq_list.
10106 spin_lock_irqsave(&task_group_lock
, flags
);
10107 for_each_possible_cpu(i
)
10108 register_fair_sched_group(tg
, i
);
10109 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10110 spin_unlock_irqrestore(&task_group_lock
, flags
);
10112 mutex_unlock(&shares_mutex
);
10116 unsigned long sched_group_shares(struct task_group
*tg
)
10122 #ifdef CONFIG_RT_GROUP_SCHED
10124 * Ensure that the real time constraints are schedulable.
10126 static DEFINE_MUTEX(rt_constraints_mutex
);
10128 static unsigned long to_ratio(u64 period
, u64 runtime
)
10130 if (runtime
== RUNTIME_INF
)
10133 return div64_u64(runtime
<< 20, period
);
10136 /* Must be called with tasklist_lock held */
10137 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10139 struct task_struct
*g
, *p
;
10141 do_each_thread(g
, p
) {
10142 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10144 } while_each_thread(g
, p
);
10149 struct rt_schedulable_data
{
10150 struct task_group
*tg
;
10155 static int tg_schedulable(struct task_group
*tg
, void *data
)
10157 struct rt_schedulable_data
*d
= data
;
10158 struct task_group
*child
;
10159 unsigned long total
, sum
= 0;
10160 u64 period
, runtime
;
10162 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10163 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10166 period
= d
->rt_period
;
10167 runtime
= d
->rt_runtime
;
10170 #ifdef CONFIG_USER_SCHED
10171 if (tg
== &root_task_group
) {
10172 period
= global_rt_period();
10173 runtime
= global_rt_runtime();
10178 * Cannot have more runtime than the period.
10180 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10184 * Ensure we don't starve existing RT tasks.
10186 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10189 total
= to_ratio(period
, runtime
);
10192 * Nobody can have more than the global setting allows.
10194 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10198 * The sum of our children's runtime should not exceed our own.
10200 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10201 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10202 runtime
= child
->rt_bandwidth
.rt_runtime
;
10204 if (child
== d
->tg
) {
10205 period
= d
->rt_period
;
10206 runtime
= d
->rt_runtime
;
10209 sum
+= to_ratio(period
, runtime
);
10218 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10220 struct rt_schedulable_data data
= {
10222 .rt_period
= period
,
10223 .rt_runtime
= runtime
,
10226 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10229 static int tg_set_bandwidth(struct task_group
*tg
,
10230 u64 rt_period
, u64 rt_runtime
)
10234 mutex_lock(&rt_constraints_mutex
);
10235 read_lock(&tasklist_lock
);
10236 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10240 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10241 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10242 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10244 for_each_possible_cpu(i
) {
10245 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10247 spin_lock(&rt_rq
->rt_runtime_lock
);
10248 rt_rq
->rt_runtime
= rt_runtime
;
10249 spin_unlock(&rt_rq
->rt_runtime_lock
);
10251 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10253 read_unlock(&tasklist_lock
);
10254 mutex_unlock(&rt_constraints_mutex
);
10259 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10261 u64 rt_runtime
, rt_period
;
10263 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10264 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10265 if (rt_runtime_us
< 0)
10266 rt_runtime
= RUNTIME_INF
;
10268 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10271 long sched_group_rt_runtime(struct task_group
*tg
)
10275 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10278 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10279 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10280 return rt_runtime_us
;
10283 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10285 u64 rt_runtime
, rt_period
;
10287 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10288 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10290 if (rt_period
== 0)
10293 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10296 long sched_group_rt_period(struct task_group
*tg
)
10300 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10301 do_div(rt_period_us
, NSEC_PER_USEC
);
10302 return rt_period_us
;
10305 static int sched_rt_global_constraints(void)
10307 u64 runtime
, period
;
10310 if (sysctl_sched_rt_period
<= 0)
10313 runtime
= global_rt_runtime();
10314 period
= global_rt_period();
10317 * Sanity check on the sysctl variables.
10319 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10322 mutex_lock(&rt_constraints_mutex
);
10323 read_lock(&tasklist_lock
);
10324 ret
= __rt_schedulable(NULL
, 0, 0);
10325 read_unlock(&tasklist_lock
);
10326 mutex_unlock(&rt_constraints_mutex
);
10331 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10333 /* Don't accept realtime tasks when there is no way for them to run */
10334 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10340 #else /* !CONFIG_RT_GROUP_SCHED */
10341 static int sched_rt_global_constraints(void)
10343 unsigned long flags
;
10346 if (sysctl_sched_rt_period
<= 0)
10350 * There's always some RT tasks in the root group
10351 * -- migration, kstopmachine etc..
10353 if (sysctl_sched_rt_runtime
== 0)
10356 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10357 for_each_possible_cpu(i
) {
10358 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10360 spin_lock(&rt_rq
->rt_runtime_lock
);
10361 rt_rq
->rt_runtime
= global_rt_runtime();
10362 spin_unlock(&rt_rq
->rt_runtime_lock
);
10364 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10368 #endif /* CONFIG_RT_GROUP_SCHED */
10370 int sched_rt_handler(struct ctl_table
*table
, int write
,
10371 void __user
*buffer
, size_t *lenp
,
10375 int old_period
, old_runtime
;
10376 static DEFINE_MUTEX(mutex
);
10378 mutex_lock(&mutex
);
10379 old_period
= sysctl_sched_rt_period
;
10380 old_runtime
= sysctl_sched_rt_runtime
;
10382 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10384 if (!ret
&& write
) {
10385 ret
= sched_rt_global_constraints();
10387 sysctl_sched_rt_period
= old_period
;
10388 sysctl_sched_rt_runtime
= old_runtime
;
10390 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10391 def_rt_bandwidth
.rt_period
=
10392 ns_to_ktime(global_rt_period());
10395 mutex_unlock(&mutex
);
10400 #ifdef CONFIG_CGROUP_SCHED
10402 /* return corresponding task_group object of a cgroup */
10403 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10405 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10406 struct task_group
, css
);
10409 static struct cgroup_subsys_state
*
10410 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10412 struct task_group
*tg
, *parent
;
10414 if (!cgrp
->parent
) {
10415 /* This is early initialization for the top cgroup */
10416 return &init_task_group
.css
;
10419 parent
= cgroup_tg(cgrp
->parent
);
10420 tg
= sched_create_group(parent
);
10422 return ERR_PTR(-ENOMEM
);
10428 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10430 struct task_group
*tg
= cgroup_tg(cgrp
);
10432 sched_destroy_group(tg
);
10436 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10438 #ifdef CONFIG_RT_GROUP_SCHED
10439 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10442 /* We don't support RT-tasks being in separate groups */
10443 if (tsk
->sched_class
!= &fair_sched_class
)
10450 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10451 struct task_struct
*tsk
, bool threadgroup
)
10453 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10457 struct task_struct
*c
;
10459 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10460 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10472 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10473 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10476 sched_move_task(tsk
);
10478 struct task_struct
*c
;
10480 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10481 sched_move_task(c
);
10487 #ifdef CONFIG_FAIR_GROUP_SCHED
10488 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10491 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10494 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10496 struct task_group
*tg
= cgroup_tg(cgrp
);
10498 return (u64
) tg
->shares
;
10500 #endif /* CONFIG_FAIR_GROUP_SCHED */
10502 #ifdef CONFIG_RT_GROUP_SCHED
10503 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10506 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10509 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10511 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10514 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10517 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10520 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10522 return sched_group_rt_period(cgroup_tg(cgrp
));
10524 #endif /* CONFIG_RT_GROUP_SCHED */
10526 static struct cftype cpu_files
[] = {
10527 #ifdef CONFIG_FAIR_GROUP_SCHED
10530 .read_u64
= cpu_shares_read_u64
,
10531 .write_u64
= cpu_shares_write_u64
,
10534 #ifdef CONFIG_RT_GROUP_SCHED
10536 .name
= "rt_runtime_us",
10537 .read_s64
= cpu_rt_runtime_read
,
10538 .write_s64
= cpu_rt_runtime_write
,
10541 .name
= "rt_period_us",
10542 .read_u64
= cpu_rt_period_read_uint
,
10543 .write_u64
= cpu_rt_period_write_uint
,
10548 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10550 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10553 struct cgroup_subsys cpu_cgroup_subsys
= {
10555 .create
= cpu_cgroup_create
,
10556 .destroy
= cpu_cgroup_destroy
,
10557 .can_attach
= cpu_cgroup_can_attach
,
10558 .attach
= cpu_cgroup_attach
,
10559 .populate
= cpu_cgroup_populate
,
10560 .subsys_id
= cpu_cgroup_subsys_id
,
10564 #endif /* CONFIG_CGROUP_SCHED */
10566 #ifdef CONFIG_CGROUP_CPUACCT
10569 * CPU accounting code for task groups.
10571 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10572 * (balbir@in.ibm.com).
10575 /* track cpu usage of a group of tasks and its child groups */
10577 struct cgroup_subsys_state css
;
10578 /* cpuusage holds pointer to a u64-type object on every cpu */
10580 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10581 struct cpuacct
*parent
;
10584 struct cgroup_subsys cpuacct_subsys
;
10586 /* return cpu accounting group corresponding to this container */
10587 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10589 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10590 struct cpuacct
, css
);
10593 /* return cpu accounting group to which this task belongs */
10594 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10596 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10597 struct cpuacct
, css
);
10600 /* create a new cpu accounting group */
10601 static struct cgroup_subsys_state
*cpuacct_create(
10602 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10604 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10610 ca
->cpuusage
= alloc_percpu(u64
);
10614 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10615 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10616 goto out_free_counters
;
10619 ca
->parent
= cgroup_ca(cgrp
->parent
);
10625 percpu_counter_destroy(&ca
->cpustat
[i
]);
10626 free_percpu(ca
->cpuusage
);
10630 return ERR_PTR(-ENOMEM
);
10633 /* destroy an existing cpu accounting group */
10635 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10637 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10640 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10641 percpu_counter_destroy(&ca
->cpustat
[i
]);
10642 free_percpu(ca
->cpuusage
);
10646 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10648 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10651 #ifndef CONFIG_64BIT
10653 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10655 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10657 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10665 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10667 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10669 #ifndef CONFIG_64BIT
10671 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10673 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10675 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10681 /* return total cpu usage (in nanoseconds) of a group */
10682 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10684 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10685 u64 totalcpuusage
= 0;
10688 for_each_present_cpu(i
)
10689 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10691 return totalcpuusage
;
10694 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10697 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10706 for_each_present_cpu(i
)
10707 cpuacct_cpuusage_write(ca
, i
, 0);
10713 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10714 struct seq_file
*m
)
10716 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10720 for_each_present_cpu(i
) {
10721 percpu
= cpuacct_cpuusage_read(ca
, i
);
10722 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10724 seq_printf(m
, "\n");
10728 static const char *cpuacct_stat_desc
[] = {
10729 [CPUACCT_STAT_USER
] = "user",
10730 [CPUACCT_STAT_SYSTEM
] = "system",
10733 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10734 struct cgroup_map_cb
*cb
)
10736 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10739 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10740 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10741 val
= cputime64_to_clock_t(val
);
10742 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10747 static struct cftype files
[] = {
10750 .read_u64
= cpuusage_read
,
10751 .write_u64
= cpuusage_write
,
10754 .name
= "usage_percpu",
10755 .read_seq_string
= cpuacct_percpu_seq_read
,
10759 .read_map
= cpuacct_stats_show
,
10763 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10765 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10769 * charge this task's execution time to its accounting group.
10771 * called with rq->lock held.
10773 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10775 struct cpuacct
*ca
;
10778 if (unlikely(!cpuacct_subsys
.active
))
10781 cpu
= task_cpu(tsk
);
10787 for (; ca
; ca
= ca
->parent
) {
10788 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10789 *cpuusage
+= cputime
;
10796 * Charge the system/user time to the task's accounting group.
10798 static void cpuacct_update_stats(struct task_struct
*tsk
,
10799 enum cpuacct_stat_index idx
, cputime_t val
)
10801 struct cpuacct
*ca
;
10803 if (unlikely(!cpuacct_subsys
.active
))
10810 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10816 struct cgroup_subsys cpuacct_subsys
= {
10818 .create
= cpuacct_create
,
10819 .destroy
= cpuacct_destroy
,
10820 .populate
= cpuacct_populate
,
10821 .subsys_id
= cpuacct_subsys_id
,
10823 #endif /* CONFIG_CGROUP_CPUACCT */
10827 int rcu_expedited_torture_stats(char *page
)
10831 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10833 void synchronize_sched_expedited(void)
10836 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10838 #else /* #ifndef CONFIG_SMP */
10840 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10841 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10843 #define RCU_EXPEDITED_STATE_POST -2
10844 #define RCU_EXPEDITED_STATE_IDLE -1
10846 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10848 int rcu_expedited_torture_stats(char *page
)
10853 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10854 for_each_online_cpu(cpu
) {
10855 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10856 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10858 cnt
+= sprintf(&page
[cnt
], "\n");
10861 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10863 static long synchronize_sched_expedited_count
;
10866 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10867 * approach to force grace period to end quickly. This consumes
10868 * significant time on all CPUs, and is thus not recommended for
10869 * any sort of common-case code.
10871 * Note that it is illegal to call this function while holding any
10872 * lock that is acquired by a CPU-hotplug notifier. Failing to
10873 * observe this restriction will result in deadlock.
10875 void synchronize_sched_expedited(void)
10878 unsigned long flags
;
10879 bool need_full_sync
= 0;
10881 struct migration_req
*req
;
10885 smp_mb(); /* ensure prior mod happens before capturing snap. */
10886 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10888 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10890 if (trycount
++ < 10)
10891 udelay(trycount
* num_online_cpus());
10893 synchronize_sched();
10896 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10897 smp_mb(); /* ensure test happens before caller kfree */
10902 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10903 for_each_online_cpu(cpu
) {
10905 req
= &per_cpu(rcu_migration_req
, cpu
);
10906 init_completion(&req
->done
);
10908 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10909 spin_lock_irqsave(&rq
->lock
, flags
);
10910 list_add(&req
->list
, &rq
->migration_queue
);
10911 spin_unlock_irqrestore(&rq
->lock
, flags
);
10912 wake_up_process(rq
->migration_thread
);
10914 for_each_online_cpu(cpu
) {
10915 rcu_expedited_state
= cpu
;
10916 req
= &per_cpu(rcu_migration_req
, cpu
);
10918 wait_for_completion(&req
->done
);
10919 spin_lock_irqsave(&rq
->lock
, flags
);
10920 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10921 need_full_sync
= 1;
10922 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10923 spin_unlock_irqrestore(&rq
->lock
, flags
);
10925 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10926 mutex_unlock(&rcu_sched_expedited_mutex
);
10928 if (need_full_sync
)
10929 synchronize_sched();
10931 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10933 #endif /* #else #ifndef CONFIG_SMP */