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 char in_nohz_recently
;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load
;
542 unsigned long nr_load_updates
;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list
;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list
;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible
;
564 struct task_struct
*curr
, *idle
;
565 unsigned long next_balance
;
566 struct mm_struct
*prev_mm
;
573 struct root_domain
*rd
;
574 struct sched_domain
*sd
;
576 unsigned char idle_at_tick
;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task
;
587 struct task_struct
*migration_thread
;
588 struct list_head migration_queue
;
596 /* calc_load related fields */
597 unsigned long calc_load_update
;
598 long calc_load_active
;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending
;
603 struct call_single_data hrtick_csd
;
605 struct hrtimer hrtick_timer
;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info
;
611 unsigned long long rq_cpu_time
;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count
;
617 /* schedule() stats */
618 unsigned int sched_switch
;
619 unsigned int sched_count
;
620 unsigned int sched_goidle
;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count
;
624 unsigned int ttwu_local
;
627 unsigned int bkl_count
;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
634 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
636 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
639 static inline int cpu_of(struct rq
*rq
)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq
*rq
)
666 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu
)
688 return spin_is_locked(&cpu_rq(cpu
)->lock
);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug
unsigned int sysctl_sched_features
=
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly
char *sched_feat_names
[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file
*m
, void *v
)
728 for (i
= 0; sched_feat_names
[i
]; i
++) {
729 if (!(sysctl_sched_features
& (1UL << i
)))
731 seq_printf(m
, "%s ", sched_feat_names
[i
]);
739 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
740 size_t cnt
, loff_t
*ppos
)
750 if (copy_from_user(&buf
, ubuf
, cnt
))
755 if (strncmp(buf
, "NO_", 3) == 0) {
760 for (i
= 0; sched_feat_names
[i
]; i
++) {
761 int len
= strlen(sched_feat_names
[i
]);
763 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
765 sysctl_sched_features
&= ~(1UL << i
);
767 sysctl_sched_features
|= (1UL << i
);
772 if (!sched_feat_names
[i
])
780 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
782 return single_open(filp
, sched_feat_show
, NULL
);
785 static const struct file_operations sched_feat_fops
= {
786 .open
= sched_feat_open
,
787 .write
= sched_feat_write
,
790 .release
= single_release
,
793 static __init
int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
800 late_initcall(sched_init_debug
);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit
= 250000;
819 * Inject some fuzzyness into changing the per-cpu group shares
820 * this avoids remote rq-locks at the expense of fairness.
823 unsigned int sysctl_sched_shares_thresh
= 4;
826 * period over which we average the RT time consumption, measured
831 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period
= 1000000;
839 static __read_mostly
int scheduler_running
;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime
= 950000;
847 static inline u64
global_rt_period(void)
849 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
852 static inline u64
global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime
< 0)
857 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
869 return rq
->curr
== p
;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
875 return task_current(rq
, p
);
878 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
882 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq
->lock
.owner
= current
;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
895 spin_unlock_irq(&rq
->lock
);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
904 return task_current(rq
, p
);
908 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq
->lock
);
921 spin_unlock(&rq
->lock
);
925 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
950 struct rq
*rq
= task_rq(p
);
951 spin_lock(&rq
->lock
);
952 if (likely(rq
== task_rq(p
)))
954 spin_unlock(&rq
->lock
);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
969 local_irq_save(*flags
);
971 spin_lock(&rq
->lock
);
972 if (likely(rq
== task_rq(p
)))
974 spin_unlock_irqrestore(&rq
->lock
, *flags
);
978 void task_rq_unlock_wait(struct task_struct
*p
)
980 struct rq
*rq
= task_rq(p
);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq
->lock
);
986 static void __task_rq_unlock(struct rq
*rq
)
989 spin_unlock(&rq
->lock
);
992 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
995 spin_unlock_irqrestore(&rq
->lock
, *flags
);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq
*this_rq_lock(void)
1002 __acquires(rq
->lock
)
1006 local_irq_disable();
1008 spin_lock(&rq
->lock
);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq
*rq
)
1032 if (!sched_feat(HRTICK
))
1034 if (!cpu_active(cpu_of(rq
)))
1036 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1039 static void hrtick_clear(struct rq
*rq
)
1041 if (hrtimer_active(&rq
->hrtick_timer
))
1042 hrtimer_cancel(&rq
->hrtick_timer
);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1051 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1053 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1055 spin_lock(&rq
->lock
);
1056 update_rq_clock(rq
);
1057 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1058 spin_unlock(&rq
->lock
);
1060 return HRTIMER_NORESTART
;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg
)
1069 struct rq
*rq
= arg
;
1071 spin_lock(&rq
->lock
);
1072 hrtimer_restart(&rq
->hrtick_timer
);
1073 rq
->hrtick_csd_pending
= 0;
1074 spin_unlock(&rq
->lock
);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq
*rq
, u64 delay
)
1084 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1085 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1087 hrtimer_set_expires(timer
, time
);
1089 if (rq
== this_rq()) {
1090 hrtimer_restart(timer
);
1091 } else if (!rq
->hrtick_csd_pending
) {
1092 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1093 rq
->hrtick_csd_pending
= 1;
1098 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1100 int cpu
= (int)(long)hcpu
;
1103 case CPU_UP_CANCELED
:
1104 case CPU_UP_CANCELED_FROZEN
:
1105 case CPU_DOWN_PREPARE
:
1106 case CPU_DOWN_PREPARE_FROZEN
:
1108 case CPU_DEAD_FROZEN
:
1109 hrtick_clear(cpu_rq(cpu
));
1116 static __init
void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick
, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq
*rq
, u64 delay
)
1128 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1129 HRTIMER_MODE_REL_PINNED
, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq
*rq
)
1140 rq
->hrtick_csd_pending
= 0;
1142 rq
->hrtick_csd
.flags
= 0;
1143 rq
->hrtick_csd
.func
= __hrtick_start
;
1144 rq
->hrtick_csd
.info
= rq
;
1147 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1148 rq
->hrtick_timer
.function
= hrtick
;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq
*rq
)
1155 static inline void init_rq_hrtick(struct rq
*rq
)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct
*p
)
1181 assert_spin_locked(&task_rq(p
)->lock
);
1183 if (test_tsk_need_resched(p
))
1186 set_tsk_need_resched(p
);
1189 if (cpu
== smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p
))
1195 smp_send_reschedule(cpu
);
1198 static void resched_cpu(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1201 unsigned long flags
;
1203 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1205 resched_task(cpu_curr(cpu
));
1206 spin_unlock_irqrestore(&rq
->lock
, flags
);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu
)
1222 struct rq
*rq
= cpu_rq(cpu
);
1224 if (cpu
== smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq
->curr
!= rq
->idle
)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq
->idle
);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq
->idle
))
1247 smp_send_reschedule(cpu
);
1249 #endif /* CONFIG_NO_HZ */
1251 static u64
sched_avg_period(void)
1253 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1256 static void sched_avg_update(struct rq
*rq
)
1258 s64 period
= sched_avg_period();
1260 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1261 rq
->age_stamp
+= period
;
1266 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1268 rq
->rt_avg
+= rt_delta
;
1269 sched_avg_update(rq
);
1272 #else /* !CONFIG_SMP */
1273 static void resched_task(struct task_struct
*p
)
1275 assert_spin_locked(&task_rq(p
)->lock
);
1276 set_tsk_need_resched(p
);
1279 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1282 #endif /* CONFIG_SMP */
1284 #if BITS_PER_LONG == 32
1285 # define WMULT_CONST (~0UL)
1287 # define WMULT_CONST (1UL << 32)
1290 #define WMULT_SHIFT 32
1293 * Shift right and round:
1295 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1298 * delta *= weight / lw
1300 static unsigned long
1301 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1302 struct load_weight
*lw
)
1306 if (!lw
->inv_weight
) {
1307 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1310 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1314 tmp
= (u64
)delta_exec
* weight
;
1316 * Check whether we'd overflow the 64-bit multiplication:
1318 if (unlikely(tmp
> WMULT_CONST
))
1319 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1322 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1324 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1327 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1333 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1340 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1341 * of tasks with abnormal "nice" values across CPUs the contribution that
1342 * each task makes to its run queue's load is weighted according to its
1343 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1344 * scaled version of the new time slice allocation that they receive on time
1348 #define WEIGHT_IDLEPRIO 3
1349 #define WMULT_IDLEPRIO 1431655765
1352 * Nice levels are multiplicative, with a gentle 10% change for every
1353 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1354 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1355 * that remained on nice 0.
1357 * The "10% effect" is relative and cumulative: from _any_ nice level,
1358 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1359 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1360 * If a task goes up by ~10% and another task goes down by ~10% then
1361 * the relative distance between them is ~25%.)
1363 static const int prio_to_weight
[40] = {
1364 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1365 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1366 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1367 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1368 /* 0 */ 1024, 820, 655, 526, 423,
1369 /* 5 */ 335, 272, 215, 172, 137,
1370 /* 10 */ 110, 87, 70, 56, 45,
1371 /* 15 */ 36, 29, 23, 18, 15,
1375 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 * In cases where the weight does not change often, we can use the
1378 * precalculated inverse to speed up arithmetics by turning divisions
1379 * into multiplications:
1381 static const u32 prio_to_wmult
[40] = {
1382 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1383 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1384 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1385 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1386 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1387 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1388 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1389 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1392 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1395 * runqueue iterator, to support SMP load-balancing between different
1396 * scheduling classes, without having to expose their internal data
1397 * structures to the load-balancing proper:
1399 struct rq_iterator
{
1401 struct task_struct
*(*start
)(void *);
1402 struct task_struct
*(*next
)(void *);
1406 static unsigned long
1407 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1408 unsigned long max_load_move
, struct sched_domain
*sd
,
1409 enum cpu_idle_type idle
, int *all_pinned
,
1410 int *this_best_prio
, struct rq_iterator
*iterator
);
1413 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1414 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1415 struct rq_iterator
*iterator
);
1418 /* Time spent by the tasks of the cpu accounting group executing in ... */
1419 enum cpuacct_stat_index
{
1420 CPUACCT_STAT_USER
, /* ... user mode */
1421 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1423 CPUACCT_STAT_NSTATS
,
1426 #ifdef CONFIG_CGROUP_CPUACCT
1427 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1428 static void cpuacct_update_stats(struct task_struct
*tsk
,
1429 enum cpuacct_stat_index idx
, cputime_t val
);
1431 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1432 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1433 enum cpuacct_stat_index idx
, cputime_t val
) {}
1436 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1438 update_load_add(&rq
->load
, load
);
1441 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1443 update_load_sub(&rq
->load
, load
);
1446 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1447 typedef int (*tg_visitor
)(struct task_group
*, void *);
1450 * Iterate the full tree, calling @down when first entering a node and @up when
1451 * leaving it for the final time.
1453 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1455 struct task_group
*parent
, *child
;
1459 parent
= &root_task_group
;
1461 ret
= (*down
)(parent
, data
);
1464 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1471 ret
= (*up
)(parent
, data
);
1476 parent
= parent
->parent
;
1485 static int tg_nop(struct task_group
*tg
, void *data
)
1492 /* Used instead of source_load when we know the type == 0 */
1493 static unsigned long weighted_cpuload(const int cpu
)
1495 return cpu_rq(cpu
)->load
.weight
;
1499 * Return a low guess at the load of a migration-source cpu weighted
1500 * according to the scheduling class and "nice" value.
1502 * We want to under-estimate the load of migration sources, to
1503 * balance conservatively.
1505 static unsigned long source_load(int cpu
, int type
)
1507 struct rq
*rq
= cpu_rq(cpu
);
1508 unsigned long total
= weighted_cpuload(cpu
);
1510 if (type
== 0 || !sched_feat(LB_BIAS
))
1513 return min(rq
->cpu_load
[type
-1], total
);
1517 * Return a high guess at the load of a migration-target cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 static unsigned long target_load(int cpu
, int type
)
1522 struct rq
*rq
= cpu_rq(cpu
);
1523 unsigned long total
= weighted_cpuload(cpu
);
1525 if (type
== 0 || !sched_feat(LB_BIAS
))
1528 return max(rq
->cpu_load
[type
-1], total
);
1531 static struct sched_group
*group_of(int cpu
)
1533 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1541 static unsigned long power_of(int cpu
)
1543 struct sched_group
*group
= group_of(cpu
);
1546 return SCHED_LOAD_SCALE
;
1548 return group
->cpu_power
;
1551 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1553 static unsigned long cpu_avg_load_per_task(int cpu
)
1555 struct rq
*rq
= cpu_rq(cpu
);
1556 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1559 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1561 rq
->avg_load_per_task
= 0;
1563 return rq
->avg_load_per_task
;
1566 #ifdef CONFIG_FAIR_GROUP_SCHED
1568 static __read_mostly
unsigned long *update_shares_data
;
1570 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1573 * Calculate and set the cpu's group shares.
1575 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1576 unsigned long sd_shares
,
1577 unsigned long sd_rq_weight
,
1578 unsigned long *usd_rq_weight
)
1580 unsigned long shares
, rq_weight
;
1583 rq_weight
= usd_rq_weight
[cpu
];
1586 rq_weight
= NICE_0_LOAD
;
1590 * \Sum_j shares_j * rq_weight_i
1591 * shares_i = -----------------------------
1592 * \Sum_j rq_weight_j
1594 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1595 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1597 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1598 sysctl_sched_shares_thresh
) {
1599 struct rq
*rq
= cpu_rq(cpu
);
1600 unsigned long flags
;
1602 spin_lock_irqsave(&rq
->lock
, flags
);
1603 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1604 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1605 __set_se_shares(tg
->se
[cpu
], shares
);
1606 spin_unlock_irqrestore(&rq
->lock
, flags
);
1611 * Re-compute the task group their per cpu shares over the given domain.
1612 * This needs to be done in a bottom-up fashion because the rq weight of a
1613 * parent group depends on the shares of its child groups.
1615 static int tg_shares_up(struct task_group
*tg
, void *data
)
1617 unsigned long weight
, rq_weight
= 0, shares
= 0;
1618 unsigned long *usd_rq_weight
;
1619 struct sched_domain
*sd
= data
;
1620 unsigned long flags
;
1626 local_irq_save(flags
);
1627 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1629 for_each_cpu(i
, sched_domain_span(sd
)) {
1630 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1631 usd_rq_weight
[i
] = weight
;
1634 * If there are currently no tasks on the cpu pretend there
1635 * is one of average load so that when a new task gets to
1636 * run here it will not get delayed by group starvation.
1639 weight
= NICE_0_LOAD
;
1641 rq_weight
+= weight
;
1642 shares
+= tg
->cfs_rq
[i
]->shares
;
1645 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1646 shares
= tg
->shares
;
1648 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1649 shares
= tg
->shares
;
1651 for_each_cpu(i
, sched_domain_span(sd
))
1652 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1654 local_irq_restore(flags
);
1660 * Compute the cpu's hierarchical load factor for each task group.
1661 * This needs to be done in a top-down fashion because the load of a child
1662 * group is a fraction of its parents load.
1664 static int tg_load_down(struct task_group
*tg
, void *data
)
1667 long cpu
= (long)data
;
1670 load
= cpu_rq(cpu
)->load
.weight
;
1672 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1673 load
*= tg
->cfs_rq
[cpu
]->shares
;
1674 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1677 tg
->cfs_rq
[cpu
]->h_load
= load
;
1682 static void update_shares(struct sched_domain
*sd
)
1687 if (root_task_group_empty())
1690 now
= cpu_clock(raw_smp_processor_id());
1691 elapsed
= now
- sd
->last_update
;
1693 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1694 sd
->last_update
= now
;
1695 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1699 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1701 if (root_task_group_empty())
1704 spin_unlock(&rq
->lock
);
1706 spin_lock(&rq
->lock
);
1709 static void update_h_load(long cpu
)
1711 if (root_task_group_empty())
1714 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1719 static inline void update_shares(struct sched_domain
*sd
)
1723 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1729 #ifdef CONFIG_PREEMPT
1731 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1734 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1735 * way at the expense of forcing extra atomic operations in all
1736 * invocations. This assures that the double_lock is acquired using the
1737 * same underlying policy as the spinlock_t on this architecture, which
1738 * reduces latency compared to the unfair variant below. However, it
1739 * also adds more overhead and therefore may reduce throughput.
1741 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1742 __releases(this_rq
->lock
)
1743 __acquires(busiest
->lock
)
1744 __acquires(this_rq
->lock
)
1746 spin_unlock(&this_rq
->lock
);
1747 double_rq_lock(this_rq
, busiest
);
1754 * Unfair double_lock_balance: Optimizes throughput at the expense of
1755 * latency by eliminating extra atomic operations when the locks are
1756 * already in proper order on entry. This favors lower cpu-ids and will
1757 * grant the double lock to lower cpus over higher ids under contention,
1758 * regardless of entry order into the function.
1760 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1761 __releases(this_rq
->lock
)
1762 __acquires(busiest
->lock
)
1763 __acquires(this_rq
->lock
)
1767 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1768 if (busiest
< this_rq
) {
1769 spin_unlock(&this_rq
->lock
);
1770 spin_lock(&busiest
->lock
);
1771 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1774 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1779 #endif /* CONFIG_PREEMPT */
1782 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1784 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1786 if (unlikely(!irqs_disabled())) {
1787 /* printk() doesn't work good under rq->lock */
1788 spin_unlock(&this_rq
->lock
);
1792 return _double_lock_balance(this_rq
, busiest
);
1795 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1796 __releases(busiest
->lock
)
1798 spin_unlock(&busiest
->lock
);
1799 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1803 #ifdef CONFIG_FAIR_GROUP_SCHED
1804 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1807 cfs_rq
->shares
= shares
;
1812 static void calc_load_account_active(struct rq
*this_rq
);
1814 #include "sched_stats.h"
1815 #include "sched_idletask.c"
1816 #include "sched_fair.c"
1817 #include "sched_rt.c"
1818 #ifdef CONFIG_SCHED_DEBUG
1819 # include "sched_debug.c"
1822 #define sched_class_highest (&rt_sched_class)
1823 #define for_each_class(class) \
1824 for (class = sched_class_highest; class; class = class->next)
1826 static void inc_nr_running(struct rq
*rq
)
1831 static void dec_nr_running(struct rq
*rq
)
1836 static void set_load_weight(struct task_struct
*p
)
1838 if (task_has_rt_policy(p
)) {
1839 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1840 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1845 * SCHED_IDLE tasks get minimal weight:
1847 if (p
->policy
== SCHED_IDLE
) {
1848 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1849 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1853 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1854 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1857 static void update_avg(u64
*avg
, u64 sample
)
1859 s64 diff
= sample
- *avg
;
1863 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1866 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1868 sched_info_queued(p
);
1869 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1873 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1876 if (p
->se
.last_wakeup
) {
1877 update_avg(&p
->se
.avg_overlap
,
1878 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1879 p
->se
.last_wakeup
= 0;
1881 update_avg(&p
->se
.avg_wakeup
,
1882 sysctl_sched_wakeup_granularity
);
1886 sched_info_dequeued(p
);
1887 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1892 * __normal_prio - return the priority that is based on the static prio
1894 static inline int __normal_prio(struct task_struct
*p
)
1896 return p
->static_prio
;
1900 * Calculate the expected normal priority: i.e. priority
1901 * without taking RT-inheritance into account. Might be
1902 * boosted by interactivity modifiers. Changes upon fork,
1903 * setprio syscalls, and whenever the interactivity
1904 * estimator recalculates.
1906 static inline int normal_prio(struct task_struct
*p
)
1910 if (task_has_rt_policy(p
))
1911 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1913 prio
= __normal_prio(p
);
1918 * Calculate the current priority, i.e. the priority
1919 * taken into account by the scheduler. This value might
1920 * be boosted by RT tasks, or might be boosted by
1921 * interactivity modifiers. Will be RT if the task got
1922 * RT-boosted. If not then it returns p->normal_prio.
1924 static int effective_prio(struct task_struct
*p
)
1926 p
->normal_prio
= normal_prio(p
);
1928 * If we are RT tasks or we were boosted to RT priority,
1929 * keep the priority unchanged. Otherwise, update priority
1930 * to the normal priority:
1932 if (!rt_prio(p
->prio
))
1933 return p
->normal_prio
;
1938 * activate_task - move a task to the runqueue.
1940 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1942 if (task_contributes_to_load(p
))
1943 rq
->nr_uninterruptible
--;
1945 enqueue_task(rq
, p
, wakeup
);
1950 * deactivate_task - remove a task from the runqueue.
1952 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1954 if (task_contributes_to_load(p
))
1955 rq
->nr_uninterruptible
++;
1957 dequeue_task(rq
, p
, sleep
);
1962 * task_curr - is this task currently executing on a CPU?
1963 * @p: the task in question.
1965 inline int task_curr(const struct task_struct
*p
)
1967 return cpu_curr(task_cpu(p
)) == p
;
1970 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1972 set_task_rq(p
, cpu
);
1975 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1976 * successfuly executed on another CPU. We must ensure that updates of
1977 * per-task data have been completed by this moment.
1980 task_thread_info(p
)->cpu
= cpu
;
1984 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1985 const struct sched_class
*prev_class
,
1986 int oldprio
, int running
)
1988 if (prev_class
!= p
->sched_class
) {
1989 if (prev_class
->switched_from
)
1990 prev_class
->switched_from(rq
, p
, running
);
1991 p
->sched_class
->switched_to(rq
, p
, running
);
1993 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1997 * kthread_bind - bind a just-created kthread to a cpu.
1998 * @p: thread created by kthread_create().
1999 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2001 * Description: This function is equivalent to set_cpus_allowed(),
2002 * except that @cpu doesn't need to be online, and the thread must be
2003 * stopped (i.e., just returned from kthread_create()).
2005 * Function lives here instead of kthread.c because it messes with
2006 * scheduler internals which require locking.
2008 void kthread_bind(struct task_struct
*p
, unsigned int cpu
)
2010 struct rq
*rq
= cpu_rq(cpu
);
2011 unsigned long flags
;
2013 /* Must have done schedule() in kthread() before we set_task_cpu */
2014 if (!wait_task_inactive(p
, TASK_UNINTERRUPTIBLE
)) {
2019 spin_lock_irqsave(&rq
->lock
, flags
);
2020 update_rq_clock(rq
);
2021 set_task_cpu(p
, cpu
);
2022 p
->cpus_allowed
= cpumask_of_cpu(cpu
);
2023 p
->rt
.nr_cpus_allowed
= 1;
2024 p
->flags
|= PF_THREAD_BOUND
;
2025 spin_unlock_irqrestore(&rq
->lock
, flags
);
2027 EXPORT_SYMBOL(kthread_bind
);
2031 * Is this task likely cache-hot:
2034 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2039 * Buddy candidates are cache hot:
2041 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2042 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2043 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2046 if (p
->sched_class
!= &fair_sched_class
)
2049 if (sysctl_sched_migration_cost
== -1)
2051 if (sysctl_sched_migration_cost
== 0)
2054 delta
= now
- p
->se
.exec_start
;
2056 return delta
< (s64
)sysctl_sched_migration_cost
;
2060 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2062 int old_cpu
= task_cpu(p
);
2063 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2064 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2065 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2068 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2070 trace_sched_migrate_task(p
, new_cpu
);
2072 #ifdef CONFIG_SCHEDSTATS
2073 if (p
->se
.wait_start
)
2074 p
->se
.wait_start
-= clock_offset
;
2075 if (p
->se
.sleep_start
)
2076 p
->se
.sleep_start
-= clock_offset
;
2077 if (p
->se
.block_start
)
2078 p
->se
.block_start
-= clock_offset
;
2080 if (old_cpu
!= new_cpu
) {
2081 p
->se
.nr_migrations
++;
2082 #ifdef CONFIG_SCHEDSTATS
2083 if (task_hot(p
, old_rq
->clock
, NULL
))
2084 schedstat_inc(p
, se
.nr_forced2_migrations
);
2086 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2089 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2090 new_cfsrq
->min_vruntime
;
2092 __set_task_cpu(p
, new_cpu
);
2095 struct migration_req
{
2096 struct list_head list
;
2098 struct task_struct
*task
;
2101 struct completion done
;
2105 * The task's runqueue lock must be held.
2106 * Returns true if you have to wait for migration thread.
2109 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2111 struct rq
*rq
= task_rq(p
);
2114 * If the task is not on a runqueue (and not running), then
2115 * it is sufficient to simply update the task's cpu field.
2117 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2118 update_rq_clock(rq
);
2119 set_task_cpu(p
, dest_cpu
);
2123 init_completion(&req
->done
);
2125 req
->dest_cpu
= dest_cpu
;
2126 list_add(&req
->list
, &rq
->migration_queue
);
2132 * wait_task_context_switch - wait for a thread to complete at least one
2135 * @p must not be current.
2137 void wait_task_context_switch(struct task_struct
*p
)
2139 unsigned long nvcsw
, nivcsw
, flags
;
2147 * The runqueue is assigned before the actual context
2148 * switch. We need to take the runqueue lock.
2150 * We could check initially without the lock but it is
2151 * very likely that we need to take the lock in every
2154 rq
= task_rq_lock(p
, &flags
);
2155 running
= task_running(rq
, p
);
2156 task_rq_unlock(rq
, &flags
);
2158 if (likely(!running
))
2161 * The switch count is incremented before the actual
2162 * context switch. We thus wait for two switches to be
2163 * sure at least one completed.
2165 if ((p
->nvcsw
- nvcsw
) > 1)
2167 if ((p
->nivcsw
- nivcsw
) > 1)
2175 * wait_task_inactive - wait for a thread to unschedule.
2177 * If @match_state is nonzero, it's the @p->state value just checked and
2178 * not expected to change. If it changes, i.e. @p might have woken up,
2179 * then return zero. When we succeed in waiting for @p to be off its CPU,
2180 * we return a positive number (its total switch count). If a second call
2181 * a short while later returns the same number, the caller can be sure that
2182 * @p has remained unscheduled the whole time.
2184 * The caller must ensure that the task *will* unschedule sometime soon,
2185 * else this function might spin for a *long* time. This function can't
2186 * be called with interrupts off, or it may introduce deadlock with
2187 * smp_call_function() if an IPI is sent by the same process we are
2188 * waiting to become inactive.
2190 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2192 unsigned long flags
;
2199 * We do the initial early heuristics without holding
2200 * any task-queue locks at all. We'll only try to get
2201 * the runqueue lock when things look like they will
2207 * If the task is actively running on another CPU
2208 * still, just relax and busy-wait without holding
2211 * NOTE! Since we don't hold any locks, it's not
2212 * even sure that "rq" stays as the right runqueue!
2213 * But we don't care, since "task_running()" will
2214 * return false if the runqueue has changed and p
2215 * is actually now running somewhere else!
2217 while (task_running(rq
, p
)) {
2218 if (match_state
&& unlikely(p
->state
!= match_state
))
2224 * Ok, time to look more closely! We need the rq
2225 * lock now, to be *sure*. If we're wrong, we'll
2226 * just go back and repeat.
2228 rq
= task_rq_lock(p
, &flags
);
2229 trace_sched_wait_task(rq
, p
);
2230 running
= task_running(rq
, p
);
2231 on_rq
= p
->se
.on_rq
;
2233 if (!match_state
|| p
->state
== match_state
)
2234 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2235 task_rq_unlock(rq
, &flags
);
2238 * If it changed from the expected state, bail out now.
2240 if (unlikely(!ncsw
))
2244 * Was it really running after all now that we
2245 * checked with the proper locks actually held?
2247 * Oops. Go back and try again..
2249 if (unlikely(running
)) {
2255 * It's not enough that it's not actively running,
2256 * it must be off the runqueue _entirely_, and not
2259 * So if it was still runnable (but just not actively
2260 * running right now), it's preempted, and we should
2261 * yield - it could be a while.
2263 if (unlikely(on_rq
)) {
2264 schedule_timeout_uninterruptible(1);
2269 * Ahh, all good. It wasn't running, and it wasn't
2270 * runnable, which means that it will never become
2271 * running in the future either. We're all done!
2280 * kick_process - kick a running thread to enter/exit the kernel
2281 * @p: the to-be-kicked thread
2283 * Cause a process which is running on another CPU to enter
2284 * kernel-mode, without any delay. (to get signals handled.)
2286 * NOTE: this function doesnt have to take the runqueue lock,
2287 * because all it wants to ensure is that the remote task enters
2288 * the kernel. If the IPI races and the task has been migrated
2289 * to another CPU then no harm is done and the purpose has been
2292 void kick_process(struct task_struct
*p
)
2298 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2299 smp_send_reschedule(cpu
);
2302 EXPORT_SYMBOL_GPL(kick_process
);
2303 #endif /* CONFIG_SMP */
2306 * task_oncpu_function_call - call a function on the cpu on which a task runs
2307 * @p: the task to evaluate
2308 * @func: the function to be called
2309 * @info: the function call argument
2311 * Calls the function @func when the task is currently running. This might
2312 * be on the current CPU, which just calls the function directly
2314 void task_oncpu_function_call(struct task_struct
*p
,
2315 void (*func
) (void *info
), void *info
)
2322 smp_call_function_single(cpu
, func
, info
, 1);
2327 * try_to_wake_up - wake up a thread
2328 * @p: the to-be-woken-up thread
2329 * @state: the mask of task states that can be woken
2330 * @sync: do a synchronous wakeup?
2332 * Put it on the run-queue if it's not already there. The "current"
2333 * thread is always on the run-queue (except when the actual
2334 * re-schedule is in progress), and as such you're allowed to do
2335 * the simpler "current->state = TASK_RUNNING" to mark yourself
2336 * runnable without the overhead of this.
2338 * returns failure only if the task is already active.
2340 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2343 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2344 unsigned long flags
;
2345 struct rq
*rq
, *orig_rq
;
2347 if (!sched_feat(SYNC_WAKEUPS
))
2348 wake_flags
&= ~WF_SYNC
;
2350 this_cpu
= get_cpu();
2353 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2354 update_rq_clock(rq
);
2355 if (!(p
->state
& state
))
2365 if (unlikely(task_running(rq
, p
)))
2369 * In order to handle concurrent wakeups and release the rq->lock
2370 * we put the task in TASK_WAKING state.
2372 * First fix up the nr_uninterruptible count:
2374 if (task_contributes_to_load(p
))
2375 rq
->nr_uninterruptible
--;
2376 p
->state
= TASK_WAKING
;
2377 task_rq_unlock(rq
, &flags
);
2379 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2380 if (cpu
!= orig_cpu
) {
2381 local_irq_save(flags
);
2383 update_rq_clock(rq
);
2384 set_task_cpu(p
, cpu
);
2385 local_irq_restore(flags
);
2387 rq
= task_rq_lock(p
, &flags
);
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();
2561 unsigned long flags
;
2566 * Revert to default priority/policy on fork if requested.
2568 if (unlikely(p
->sched_reset_on_fork
)) {
2569 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2570 p
->policy
= SCHED_NORMAL
;
2571 p
->normal_prio
= p
->static_prio
;
2574 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2575 p
->static_prio
= NICE_TO_PRIO(0);
2576 p
->normal_prio
= p
->static_prio
;
2581 * We don't need the reset flag anymore after the fork. It has
2582 * fulfilled its duty:
2584 p
->sched_reset_on_fork
= 0;
2588 * Make sure we do not leak PI boosting priority to the child.
2590 p
->prio
= current
->normal_prio
;
2592 if (!rt_prio(p
->prio
))
2593 p
->sched_class
= &fair_sched_class
;
2596 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2598 local_irq_save(flags
);
2599 update_rq_clock(cpu_rq(cpu
));
2600 set_task_cpu(p
, cpu
);
2601 local_irq_restore(flags
);
2603 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2604 if (likely(sched_info_on()))
2605 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2607 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2610 #ifdef CONFIG_PREEMPT
2611 /* Want to start with kernel preemption disabled. */
2612 task_thread_info(p
)->preempt_count
= 1;
2614 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2620 * wake_up_new_task - wake up a newly created task for the first time.
2622 * This function will do some initial scheduler statistics housekeeping
2623 * that must be done for every newly created context, then puts the task
2624 * on the runqueue and wakes it.
2626 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2628 unsigned long flags
;
2631 rq
= task_rq_lock(p
, &flags
);
2632 BUG_ON(p
->state
!= TASK_RUNNING
);
2633 update_rq_clock(rq
);
2635 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2636 activate_task(rq
, p
, 0);
2639 * Let the scheduling class do new task startup
2640 * management (if any):
2642 p
->sched_class
->task_new(rq
, p
);
2645 trace_sched_wakeup_new(rq
, p
, 1);
2646 check_preempt_curr(rq
, p
, WF_FORK
);
2648 if (p
->sched_class
->task_wake_up
)
2649 p
->sched_class
->task_wake_up(rq
, p
);
2651 task_rq_unlock(rq
, &flags
);
2654 #ifdef CONFIG_PREEMPT_NOTIFIERS
2657 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2658 * @notifier: notifier struct to register
2660 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2662 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2664 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2667 * preempt_notifier_unregister - no longer interested in preemption notifications
2668 * @notifier: notifier struct to unregister
2670 * This is safe to call from within a preemption notifier.
2672 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2674 hlist_del(¬ifier
->link
);
2676 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2678 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2680 struct preempt_notifier
*notifier
;
2681 struct hlist_node
*node
;
2683 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2684 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2688 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2689 struct task_struct
*next
)
2691 struct preempt_notifier
*notifier
;
2692 struct hlist_node
*node
;
2694 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2695 notifier
->ops
->sched_out(notifier
, next
);
2698 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2700 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2705 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2706 struct task_struct
*next
)
2710 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2713 * prepare_task_switch - prepare to switch tasks
2714 * @rq: the runqueue preparing to switch
2715 * @prev: the current task that is being switched out
2716 * @next: the task we are going to switch to.
2718 * This is called with the rq lock held and interrupts off. It must
2719 * be paired with a subsequent finish_task_switch after the context
2722 * prepare_task_switch sets up locking and calls architecture specific
2726 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2727 struct task_struct
*next
)
2729 fire_sched_out_preempt_notifiers(prev
, next
);
2730 prepare_lock_switch(rq
, next
);
2731 prepare_arch_switch(next
);
2735 * finish_task_switch - clean up after a task-switch
2736 * @rq: runqueue associated with task-switch
2737 * @prev: the thread we just switched away from.
2739 * finish_task_switch must be called after the context switch, paired
2740 * with a prepare_task_switch call before the context switch.
2741 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2742 * and do any other architecture-specific cleanup actions.
2744 * Note that we may have delayed dropping an mm in context_switch(). If
2745 * so, we finish that here outside of the runqueue lock. (Doing it
2746 * with the lock held can cause deadlocks; see schedule() for
2749 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2750 __releases(rq
->lock
)
2752 struct mm_struct
*mm
= rq
->prev_mm
;
2758 * A task struct has one reference for the use as "current".
2759 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2760 * schedule one last time. The schedule call will never return, and
2761 * the scheduled task must drop that reference.
2762 * The test for TASK_DEAD must occur while the runqueue locks are
2763 * still held, otherwise prev could be scheduled on another cpu, die
2764 * there before we look at prev->state, and then the reference would
2766 * Manfred Spraul <manfred@colorfullife.com>
2768 prev_state
= prev
->state
;
2769 finish_arch_switch(prev
);
2770 perf_event_task_sched_in(current
, cpu_of(rq
));
2771 finish_lock_switch(rq
, prev
);
2773 fire_sched_in_preempt_notifiers(current
);
2776 if (unlikely(prev_state
== TASK_DEAD
)) {
2778 * Remove function-return probe instances associated with this
2779 * task and put them back on the free list.
2781 kprobe_flush_task(prev
);
2782 put_task_struct(prev
);
2788 /* assumes rq->lock is held */
2789 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2791 if (prev
->sched_class
->pre_schedule
)
2792 prev
->sched_class
->pre_schedule(rq
, prev
);
2795 /* rq->lock is NOT held, but preemption is disabled */
2796 static inline void post_schedule(struct rq
*rq
)
2798 if (rq
->post_schedule
) {
2799 unsigned long flags
;
2801 spin_lock_irqsave(&rq
->lock
, flags
);
2802 if (rq
->curr
->sched_class
->post_schedule
)
2803 rq
->curr
->sched_class
->post_schedule(rq
);
2804 spin_unlock_irqrestore(&rq
->lock
, flags
);
2806 rq
->post_schedule
= 0;
2812 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2816 static inline void post_schedule(struct rq
*rq
)
2823 * schedule_tail - first thing a freshly forked thread must call.
2824 * @prev: the thread we just switched away from.
2826 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2827 __releases(rq
->lock
)
2829 struct rq
*rq
= this_rq();
2831 finish_task_switch(rq
, prev
);
2834 * FIXME: do we need to worry about rq being invalidated by the
2839 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2840 /* In this case, finish_task_switch does not reenable preemption */
2843 if (current
->set_child_tid
)
2844 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2848 * context_switch - switch to the new MM and the new
2849 * thread's register state.
2852 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2853 struct task_struct
*next
)
2855 struct mm_struct
*mm
, *oldmm
;
2857 prepare_task_switch(rq
, prev
, next
);
2858 trace_sched_switch(rq
, prev
, next
);
2860 oldmm
= prev
->active_mm
;
2862 * For paravirt, this is coupled with an exit in switch_to to
2863 * combine the page table reload and the switch backend into
2866 arch_start_context_switch(prev
);
2869 next
->active_mm
= oldmm
;
2870 atomic_inc(&oldmm
->mm_count
);
2871 enter_lazy_tlb(oldmm
, next
);
2873 switch_mm(oldmm
, mm
, next
);
2875 if (likely(!prev
->mm
)) {
2876 prev
->active_mm
= NULL
;
2877 rq
->prev_mm
= oldmm
;
2880 * Since the runqueue lock will be released by the next
2881 * task (which is an invalid locking op but in the case
2882 * of the scheduler it's an obvious special-case), so we
2883 * do an early lockdep release here:
2885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2886 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2889 /* Here we just switch the register state and the stack. */
2890 switch_to(prev
, next
, prev
);
2894 * this_rq must be evaluated again because prev may have moved
2895 * CPUs since it called schedule(), thus the 'rq' on its stack
2896 * frame will be invalid.
2898 finish_task_switch(this_rq(), prev
);
2902 * nr_running, nr_uninterruptible and nr_context_switches:
2904 * externally visible scheduler statistics: current number of runnable
2905 * threads, current number of uninterruptible-sleeping threads, total
2906 * number of context switches performed since bootup.
2908 unsigned long nr_running(void)
2910 unsigned long i
, sum
= 0;
2912 for_each_online_cpu(i
)
2913 sum
+= cpu_rq(i
)->nr_running
;
2918 unsigned long nr_uninterruptible(void)
2920 unsigned long i
, sum
= 0;
2922 for_each_possible_cpu(i
)
2923 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2926 * Since we read the counters lockless, it might be slightly
2927 * inaccurate. Do not allow it to go below zero though:
2929 if (unlikely((long)sum
< 0))
2935 unsigned long long nr_context_switches(void)
2938 unsigned long long sum
= 0;
2940 for_each_possible_cpu(i
)
2941 sum
+= cpu_rq(i
)->nr_switches
;
2946 unsigned long nr_iowait(void)
2948 unsigned long i
, sum
= 0;
2950 for_each_possible_cpu(i
)
2951 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2956 unsigned long nr_iowait_cpu(void)
2958 struct rq
*this = this_rq();
2959 return atomic_read(&this->nr_iowait
);
2962 unsigned long this_cpu_load(void)
2964 struct rq
*this = this_rq();
2965 return this->cpu_load
[0];
2969 /* Variables and functions for calc_load */
2970 static atomic_long_t calc_load_tasks
;
2971 static unsigned long calc_load_update
;
2972 unsigned long avenrun
[3];
2973 EXPORT_SYMBOL(avenrun
);
2976 * get_avenrun - get the load average array
2977 * @loads: pointer to dest load array
2978 * @offset: offset to add
2979 * @shift: shift count to shift the result left
2981 * These values are estimates at best, so no need for locking.
2983 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2985 loads
[0] = (avenrun
[0] + offset
) << shift
;
2986 loads
[1] = (avenrun
[1] + offset
) << shift
;
2987 loads
[2] = (avenrun
[2] + offset
) << shift
;
2990 static unsigned long
2991 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2994 load
+= active
* (FIXED_1
- exp
);
2995 return load
>> FSHIFT
;
2999 * calc_load - update the avenrun load estimates 10 ticks after the
3000 * CPUs have updated calc_load_tasks.
3002 void calc_global_load(void)
3004 unsigned long upd
= calc_load_update
+ 10;
3007 if (time_before(jiffies
, upd
))
3010 active
= atomic_long_read(&calc_load_tasks
);
3011 active
= active
> 0 ? active
* FIXED_1
: 0;
3013 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3014 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3015 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3017 calc_load_update
+= LOAD_FREQ
;
3021 * Either called from update_cpu_load() or from a cpu going idle
3023 static void calc_load_account_active(struct rq
*this_rq
)
3025 long nr_active
, delta
;
3027 nr_active
= this_rq
->nr_running
;
3028 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3030 if (nr_active
!= this_rq
->calc_load_active
) {
3031 delta
= nr_active
- this_rq
->calc_load_active
;
3032 this_rq
->calc_load_active
= nr_active
;
3033 atomic_long_add(delta
, &calc_load_tasks
);
3038 * Update rq->cpu_load[] statistics. This function is usually called every
3039 * scheduler tick (TICK_NSEC).
3041 static void update_cpu_load(struct rq
*this_rq
)
3043 unsigned long this_load
= this_rq
->load
.weight
;
3046 this_rq
->nr_load_updates
++;
3048 /* Update our load: */
3049 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3050 unsigned long old_load
, new_load
;
3052 /* scale is effectively 1 << i now, and >> i divides by scale */
3054 old_load
= this_rq
->cpu_load
[i
];
3055 new_load
= this_load
;
3057 * Round up the averaging division if load is increasing. This
3058 * prevents us from getting stuck on 9 if the load is 10, for
3061 if (new_load
> old_load
)
3062 new_load
+= scale
-1;
3063 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3066 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3067 this_rq
->calc_load_update
+= LOAD_FREQ
;
3068 calc_load_account_active(this_rq
);
3075 * double_rq_lock - safely lock two runqueues
3077 * Note this does not disable interrupts like task_rq_lock,
3078 * you need to do so manually before calling.
3080 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3081 __acquires(rq1
->lock
)
3082 __acquires(rq2
->lock
)
3084 BUG_ON(!irqs_disabled());
3086 spin_lock(&rq1
->lock
);
3087 __acquire(rq2
->lock
); /* Fake it out ;) */
3090 spin_lock(&rq1
->lock
);
3091 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3093 spin_lock(&rq2
->lock
);
3094 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3097 update_rq_clock(rq1
);
3098 update_rq_clock(rq2
);
3102 * double_rq_unlock - safely unlock two runqueues
3104 * Note this does not restore interrupts like task_rq_unlock,
3105 * you need to do so manually after calling.
3107 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3108 __releases(rq1
->lock
)
3109 __releases(rq2
->lock
)
3111 spin_unlock(&rq1
->lock
);
3113 spin_unlock(&rq2
->lock
);
3115 __release(rq2
->lock
);
3119 * If dest_cpu is allowed for this process, migrate the task to it.
3120 * This is accomplished by forcing the cpu_allowed mask to only
3121 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3122 * the cpu_allowed mask is restored.
3124 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3126 struct migration_req req
;
3127 unsigned long flags
;
3130 rq
= task_rq_lock(p
, &flags
);
3131 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3132 || unlikely(!cpu_active(dest_cpu
)))
3135 /* force the process onto the specified CPU */
3136 if (migrate_task(p
, dest_cpu
, &req
)) {
3137 /* Need to wait for migration thread (might exit: take ref). */
3138 struct task_struct
*mt
= rq
->migration_thread
;
3140 get_task_struct(mt
);
3141 task_rq_unlock(rq
, &flags
);
3142 wake_up_process(mt
);
3143 put_task_struct(mt
);
3144 wait_for_completion(&req
.done
);
3149 task_rq_unlock(rq
, &flags
);
3153 * sched_exec - execve() is a valuable balancing opportunity, because at
3154 * this point the task has the smallest effective memory and cache footprint.
3156 void sched_exec(void)
3158 int new_cpu
, this_cpu
= get_cpu();
3159 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3161 if (new_cpu
!= this_cpu
)
3162 sched_migrate_task(current
, new_cpu
);
3166 * pull_task - move a task from a remote runqueue to the local runqueue.
3167 * Both runqueues must be locked.
3169 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3170 struct rq
*this_rq
, int this_cpu
)
3172 deactivate_task(src_rq
, p
, 0);
3173 set_task_cpu(p
, this_cpu
);
3174 activate_task(this_rq
, p
, 0);
3176 * Note that idle threads have a prio of MAX_PRIO, for this test
3177 * to be always true for them.
3179 check_preempt_curr(this_rq
, p
, 0);
3183 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3186 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3187 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3190 int tsk_cache_hot
= 0;
3192 * We do not migrate tasks that are:
3193 * 1) running (obviously), or
3194 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3195 * 3) are cache-hot on their current CPU.
3197 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3198 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3203 if (task_running(rq
, p
)) {
3204 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3209 * Aggressive migration if:
3210 * 1) task is cache cold, or
3211 * 2) too many balance attempts have failed.
3214 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3215 if (!tsk_cache_hot
||
3216 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3217 #ifdef CONFIG_SCHEDSTATS
3218 if (tsk_cache_hot
) {
3219 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3220 schedstat_inc(p
, se
.nr_forced_migrations
);
3226 if (tsk_cache_hot
) {
3227 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3233 static unsigned long
3234 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3235 unsigned long max_load_move
, struct sched_domain
*sd
,
3236 enum cpu_idle_type idle
, int *all_pinned
,
3237 int *this_best_prio
, struct rq_iterator
*iterator
)
3239 int loops
= 0, pulled
= 0, pinned
= 0;
3240 struct task_struct
*p
;
3241 long rem_load_move
= max_load_move
;
3243 if (max_load_move
== 0)
3249 * Start the load-balancing iterator:
3251 p
= iterator
->start(iterator
->arg
);
3253 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3256 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3257 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3258 p
= iterator
->next(iterator
->arg
);
3262 pull_task(busiest
, p
, this_rq
, this_cpu
);
3264 rem_load_move
-= p
->se
.load
.weight
;
3266 #ifdef CONFIG_PREEMPT
3268 * NEWIDLE balancing is a source of latency, so preemptible kernels
3269 * will stop after the first task is pulled to minimize the critical
3272 if (idle
== CPU_NEWLY_IDLE
)
3277 * We only want to steal up to the prescribed amount of weighted load.
3279 if (rem_load_move
> 0) {
3280 if (p
->prio
< *this_best_prio
)
3281 *this_best_prio
= p
->prio
;
3282 p
= iterator
->next(iterator
->arg
);
3287 * Right now, this is one of only two places pull_task() is called,
3288 * so we can safely collect pull_task() stats here rather than
3289 * inside pull_task().
3291 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3294 *all_pinned
= pinned
;
3296 return max_load_move
- rem_load_move
;
3300 * move_tasks tries to move up to max_load_move weighted load from busiest to
3301 * this_rq, as part of a balancing operation within domain "sd".
3302 * Returns 1 if successful and 0 otherwise.
3304 * Called with both runqueues locked.
3306 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3307 unsigned long max_load_move
,
3308 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3311 const struct sched_class
*class = sched_class_highest
;
3312 unsigned long total_load_moved
= 0;
3313 int this_best_prio
= this_rq
->curr
->prio
;
3317 class->load_balance(this_rq
, this_cpu
, busiest
,
3318 max_load_move
- total_load_moved
,
3319 sd
, idle
, all_pinned
, &this_best_prio
);
3320 class = class->next
;
3322 #ifdef CONFIG_PREEMPT
3324 * NEWIDLE balancing is a source of latency, so preemptible
3325 * kernels will stop after the first task is pulled to minimize
3326 * the critical section.
3328 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3331 } while (class && max_load_move
> total_load_moved
);
3333 return total_load_moved
> 0;
3337 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3338 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3339 struct rq_iterator
*iterator
)
3341 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3345 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3346 pull_task(busiest
, p
, this_rq
, this_cpu
);
3348 * Right now, this is only the second place pull_task()
3349 * is called, so we can safely collect pull_task()
3350 * stats here rather than inside pull_task().
3352 schedstat_inc(sd
, lb_gained
[idle
]);
3356 p
= iterator
->next(iterator
->arg
);
3363 * move_one_task tries to move exactly one task from busiest to this_rq, as
3364 * part of active balancing operations within "domain".
3365 * Returns 1 if successful and 0 otherwise.
3367 * Called with both runqueues locked.
3369 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3370 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3372 const struct sched_class
*class;
3374 for_each_class(class) {
3375 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3381 /********** Helpers for find_busiest_group ************************/
3383 * sd_lb_stats - Structure to store the statistics of a sched_domain
3384 * during load balancing.
3386 struct sd_lb_stats
{
3387 struct sched_group
*busiest
; /* Busiest group in this sd */
3388 struct sched_group
*this; /* Local group in this sd */
3389 unsigned long total_load
; /* Total load of all groups in sd */
3390 unsigned long total_pwr
; /* Total power of all groups in sd */
3391 unsigned long avg_load
; /* Average load across all groups in sd */
3393 /** Statistics of this group */
3394 unsigned long this_load
;
3395 unsigned long this_load_per_task
;
3396 unsigned long this_nr_running
;
3398 /* Statistics of the busiest group */
3399 unsigned long max_load
;
3400 unsigned long busiest_load_per_task
;
3401 unsigned long busiest_nr_running
;
3403 int group_imb
; /* Is there imbalance in this sd */
3404 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3405 int power_savings_balance
; /* Is powersave balance needed for this sd */
3406 struct sched_group
*group_min
; /* Least loaded group in sd */
3407 struct sched_group
*group_leader
; /* Group which relieves group_min */
3408 unsigned long min_load_per_task
; /* load_per_task in group_min */
3409 unsigned long leader_nr_running
; /* Nr running of group_leader */
3410 unsigned long min_nr_running
; /* Nr running of group_min */
3415 * sg_lb_stats - stats of a sched_group required for load_balancing
3417 struct sg_lb_stats
{
3418 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3419 unsigned long group_load
; /* Total load over the CPUs of the group */
3420 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3421 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3422 unsigned long group_capacity
;
3423 int group_imb
; /* Is there an imbalance in the group ? */
3427 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3428 * @group: The group whose first cpu is to be returned.
3430 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3432 return cpumask_first(sched_group_cpus(group
));
3436 * get_sd_load_idx - Obtain the load index for a given sched domain.
3437 * @sd: The sched_domain whose load_idx is to be obtained.
3438 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3440 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3441 enum cpu_idle_type idle
)
3447 load_idx
= sd
->busy_idx
;
3450 case CPU_NEWLY_IDLE
:
3451 load_idx
= sd
->newidle_idx
;
3454 load_idx
= sd
->idle_idx
;
3462 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3464 * init_sd_power_savings_stats - Initialize power savings statistics for
3465 * the given sched_domain, during load balancing.
3467 * @sd: Sched domain whose power-savings statistics are to be initialized.
3468 * @sds: Variable containing the statistics for sd.
3469 * @idle: Idle status of the CPU at which we're performing load-balancing.
3471 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3472 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3475 * Busy processors will not participate in power savings
3478 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3479 sds
->power_savings_balance
= 0;
3481 sds
->power_savings_balance
= 1;
3482 sds
->min_nr_running
= ULONG_MAX
;
3483 sds
->leader_nr_running
= 0;
3488 * update_sd_power_savings_stats - Update the power saving stats for a
3489 * sched_domain while performing load balancing.
3491 * @group: sched_group belonging to the sched_domain under consideration.
3492 * @sds: Variable containing the statistics of the sched_domain
3493 * @local_group: Does group contain the CPU for which we're performing
3495 * @sgs: Variable containing the statistics of the group.
3497 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3498 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3501 if (!sds
->power_savings_balance
)
3505 * If the local group is idle or completely loaded
3506 * no need to do power savings balance at this domain
3508 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3509 !sds
->this_nr_running
))
3510 sds
->power_savings_balance
= 0;
3513 * If a group is already running at full capacity or idle,
3514 * don't include that group in power savings calculations
3516 if (!sds
->power_savings_balance
||
3517 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3518 !sgs
->sum_nr_running
)
3522 * Calculate the group which has the least non-idle load.
3523 * This is the group from where we need to pick up the load
3526 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3527 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3528 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3529 sds
->group_min
= group
;
3530 sds
->min_nr_running
= sgs
->sum_nr_running
;
3531 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3532 sgs
->sum_nr_running
;
3536 * Calculate the group which is almost near its
3537 * capacity but still has some space to pick up some load
3538 * from other group and save more power
3540 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3543 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3544 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3545 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3546 sds
->group_leader
= group
;
3547 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3552 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3553 * @sds: Variable containing the statistics of the sched_domain
3554 * under consideration.
3555 * @this_cpu: Cpu at which we're currently performing load-balancing.
3556 * @imbalance: Variable to store the imbalance.
3559 * Check if we have potential to perform some power-savings balance.
3560 * If yes, set the busiest group to be the least loaded group in the
3561 * sched_domain, so that it's CPUs can be put to idle.
3563 * Returns 1 if there is potential to perform power-savings balance.
3566 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3567 int this_cpu
, unsigned long *imbalance
)
3569 if (!sds
->power_savings_balance
)
3572 if (sds
->this != sds
->group_leader
||
3573 sds
->group_leader
== sds
->group_min
)
3576 *imbalance
= sds
->min_load_per_task
;
3577 sds
->busiest
= sds
->group_min
;
3582 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3583 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3584 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3589 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3590 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3595 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3596 int this_cpu
, unsigned long *imbalance
)
3600 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3603 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3605 return SCHED_LOAD_SCALE
;
3608 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3610 return default_scale_freq_power(sd
, cpu
);
3613 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3615 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3616 unsigned long smt_gain
= sd
->smt_gain
;
3623 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3625 return default_scale_smt_power(sd
, cpu
);
3628 unsigned long scale_rt_power(int cpu
)
3630 struct rq
*rq
= cpu_rq(cpu
);
3631 u64 total
, available
;
3633 sched_avg_update(rq
);
3635 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3636 available
= total
- rq
->rt_avg
;
3638 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3639 total
= SCHED_LOAD_SCALE
;
3641 total
>>= SCHED_LOAD_SHIFT
;
3643 return div_u64(available
, total
);
3646 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3648 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3649 unsigned long power
= SCHED_LOAD_SCALE
;
3650 struct sched_group
*sdg
= sd
->groups
;
3652 if (sched_feat(ARCH_POWER
))
3653 power
*= arch_scale_freq_power(sd
, cpu
);
3655 power
*= default_scale_freq_power(sd
, cpu
);
3657 power
>>= SCHED_LOAD_SHIFT
;
3659 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3660 if (sched_feat(ARCH_POWER
))
3661 power
*= arch_scale_smt_power(sd
, cpu
);
3663 power
*= default_scale_smt_power(sd
, cpu
);
3665 power
>>= SCHED_LOAD_SHIFT
;
3668 power
*= scale_rt_power(cpu
);
3669 power
>>= SCHED_LOAD_SHIFT
;
3674 sdg
->cpu_power
= power
;
3677 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3679 struct sched_domain
*child
= sd
->child
;
3680 struct sched_group
*group
, *sdg
= sd
->groups
;
3681 unsigned long power
;
3684 update_cpu_power(sd
, cpu
);
3690 group
= child
->groups
;
3692 power
+= group
->cpu_power
;
3693 group
= group
->next
;
3694 } while (group
!= child
->groups
);
3696 sdg
->cpu_power
= power
;
3700 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3701 * @sd: The sched_domain whose statistics are to be updated.
3702 * @group: sched_group whose statistics are to be updated.
3703 * @this_cpu: Cpu for which load balance is currently performed.
3704 * @idle: Idle status of this_cpu
3705 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3706 * @sd_idle: Idle status of the sched_domain containing group.
3707 * @local_group: Does group contain this_cpu.
3708 * @cpus: Set of cpus considered for load balancing.
3709 * @balance: Should we balance.
3710 * @sgs: variable to hold the statistics for this group.
3712 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3713 struct sched_group
*group
, int this_cpu
,
3714 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3715 int local_group
, const struct cpumask
*cpus
,
3716 int *balance
, struct sg_lb_stats
*sgs
)
3718 unsigned long load
, max_cpu_load
, min_cpu_load
;
3720 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3721 unsigned long sum_avg_load_per_task
;
3722 unsigned long avg_load_per_task
;
3725 balance_cpu
= group_first_cpu(group
);
3726 if (balance_cpu
== this_cpu
)
3727 update_group_power(sd
, this_cpu
);
3730 /* Tally up the load of all CPUs in the group */
3731 sum_avg_load_per_task
= avg_load_per_task
= 0;
3733 min_cpu_load
= ~0UL;
3735 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3736 struct rq
*rq
= cpu_rq(i
);
3738 if (*sd_idle
&& rq
->nr_running
)
3741 /* Bias balancing toward cpus of our domain */
3743 if (idle_cpu(i
) && !first_idle_cpu
) {
3748 load
= target_load(i
, load_idx
);
3750 load
= source_load(i
, load_idx
);
3751 if (load
> max_cpu_load
)
3752 max_cpu_load
= load
;
3753 if (min_cpu_load
> load
)
3754 min_cpu_load
= load
;
3757 sgs
->group_load
+= load
;
3758 sgs
->sum_nr_running
+= rq
->nr_running
;
3759 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3761 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3765 * First idle cpu or the first cpu(busiest) in this sched group
3766 * is eligible for doing load balancing at this and above
3767 * domains. In the newly idle case, we will allow all the cpu's
3768 * to do the newly idle load balance.
3770 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3771 balance_cpu
!= this_cpu
&& balance
) {
3776 /* Adjust by relative CPU power of the group */
3777 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3781 * Consider the group unbalanced when the imbalance is larger
3782 * than the average weight of two tasks.
3784 * APZ: with cgroup the avg task weight can vary wildly and
3785 * might not be a suitable number - should we keep a
3786 * normalized nr_running number somewhere that negates
3789 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3792 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3795 sgs
->group_capacity
=
3796 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3800 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3801 * @sd: sched_domain whose statistics are to be updated.
3802 * @this_cpu: Cpu for which load balance is currently performed.
3803 * @idle: Idle status of this_cpu
3804 * @sd_idle: Idle status of the sched_domain containing group.
3805 * @cpus: Set of cpus considered for load balancing.
3806 * @balance: Should we balance.
3807 * @sds: variable to hold the statistics for this sched_domain.
3809 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3810 enum cpu_idle_type idle
, int *sd_idle
,
3811 const struct cpumask
*cpus
, int *balance
,
3812 struct sd_lb_stats
*sds
)
3814 struct sched_domain
*child
= sd
->child
;
3815 struct sched_group
*group
= sd
->groups
;
3816 struct sg_lb_stats sgs
;
3817 int load_idx
, prefer_sibling
= 0;
3819 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3822 init_sd_power_savings_stats(sd
, sds
, idle
);
3823 load_idx
= get_sd_load_idx(sd
, idle
);
3828 local_group
= cpumask_test_cpu(this_cpu
,
3829 sched_group_cpus(group
));
3830 memset(&sgs
, 0, sizeof(sgs
));
3831 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3832 local_group
, cpus
, balance
, &sgs
);
3834 if (local_group
&& balance
&& !(*balance
))
3837 sds
->total_load
+= sgs
.group_load
;
3838 sds
->total_pwr
+= group
->cpu_power
;
3841 * In case the child domain prefers tasks go to siblings
3842 * first, lower the group capacity to one so that we'll try
3843 * and move all the excess tasks away.
3846 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3849 sds
->this_load
= sgs
.avg_load
;
3851 sds
->this_nr_running
= sgs
.sum_nr_running
;
3852 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3853 } else if (sgs
.avg_load
> sds
->max_load
&&
3854 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3856 sds
->max_load
= sgs
.avg_load
;
3857 sds
->busiest
= group
;
3858 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3859 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3860 sds
->group_imb
= sgs
.group_imb
;
3863 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3864 group
= group
->next
;
3865 } while (group
!= sd
->groups
);
3869 * fix_small_imbalance - Calculate the minor imbalance that exists
3870 * amongst the groups of a sched_domain, during
3872 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3873 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3874 * @imbalance: Variable to store the imbalance.
3876 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3877 int this_cpu
, unsigned long *imbalance
)
3879 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3880 unsigned int imbn
= 2;
3882 if (sds
->this_nr_running
) {
3883 sds
->this_load_per_task
/= sds
->this_nr_running
;
3884 if (sds
->busiest_load_per_task
>
3885 sds
->this_load_per_task
)
3888 sds
->this_load_per_task
=
3889 cpu_avg_load_per_task(this_cpu
);
3891 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3892 sds
->busiest_load_per_task
* imbn
) {
3893 *imbalance
= sds
->busiest_load_per_task
;
3898 * OK, we don't have enough imbalance to justify moving tasks,
3899 * however we may be able to increase total CPU power used by
3903 pwr_now
+= sds
->busiest
->cpu_power
*
3904 min(sds
->busiest_load_per_task
, sds
->max_load
);
3905 pwr_now
+= sds
->this->cpu_power
*
3906 min(sds
->this_load_per_task
, sds
->this_load
);
3907 pwr_now
/= SCHED_LOAD_SCALE
;
3909 /* Amount of load we'd subtract */
3910 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3911 sds
->busiest
->cpu_power
;
3912 if (sds
->max_load
> tmp
)
3913 pwr_move
+= sds
->busiest
->cpu_power
*
3914 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3916 /* Amount of load we'd add */
3917 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3918 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3919 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3920 sds
->this->cpu_power
;
3922 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3923 sds
->this->cpu_power
;
3924 pwr_move
+= sds
->this->cpu_power
*
3925 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3926 pwr_move
/= SCHED_LOAD_SCALE
;
3928 /* Move if we gain throughput */
3929 if (pwr_move
> pwr_now
)
3930 *imbalance
= sds
->busiest_load_per_task
;
3934 * calculate_imbalance - Calculate the amount of imbalance present within the
3935 * groups of a given sched_domain during load balance.
3936 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3937 * @this_cpu: Cpu for which currently load balance is being performed.
3938 * @imbalance: The variable to store the imbalance.
3940 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3941 unsigned long *imbalance
)
3943 unsigned long max_pull
;
3945 * In the presence of smp nice balancing, certain scenarios can have
3946 * max load less than avg load(as we skip the groups at or below
3947 * its cpu_power, while calculating max_load..)
3949 if (sds
->max_load
< sds
->avg_load
) {
3951 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3954 /* Don't want to pull so many tasks that a group would go idle */
3955 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3956 sds
->max_load
- sds
->busiest_load_per_task
);
3958 /* How much load to actually move to equalise the imbalance */
3959 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3960 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3964 * if *imbalance is less than the average load per runnable task
3965 * there is no gaurantee that any tasks will be moved so we'll have
3966 * a think about bumping its value to force at least one task to be
3969 if (*imbalance
< sds
->busiest_load_per_task
)
3970 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3973 /******* find_busiest_group() helpers end here *********************/
3976 * find_busiest_group - Returns the busiest group within the sched_domain
3977 * if there is an imbalance. If there isn't an imbalance, and
3978 * the user has opted for power-savings, it returns a group whose
3979 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3980 * such a group exists.
3982 * Also calculates the amount of weighted load which should be moved
3983 * to restore balance.
3985 * @sd: The sched_domain whose busiest group is to be returned.
3986 * @this_cpu: The cpu for which load balancing is currently being performed.
3987 * @imbalance: Variable which stores amount of weighted load which should
3988 * be moved to restore balance/put a group to idle.
3989 * @idle: The idle status of this_cpu.
3990 * @sd_idle: The idleness of sd
3991 * @cpus: The set of CPUs under consideration for load-balancing.
3992 * @balance: Pointer to a variable indicating if this_cpu
3993 * is the appropriate cpu to perform load balancing at this_level.
3995 * Returns: - the busiest group if imbalance exists.
3996 * - If no imbalance and user has opted for power-savings balance,
3997 * return the least loaded group whose CPUs can be
3998 * put to idle by rebalancing its tasks onto our group.
4000 static struct sched_group
*
4001 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4002 unsigned long *imbalance
, enum cpu_idle_type idle
,
4003 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4005 struct sd_lb_stats sds
;
4007 memset(&sds
, 0, sizeof(sds
));
4010 * Compute the various statistics relavent for load balancing at
4013 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4016 /* Cases where imbalance does not exist from POV of this_cpu */
4017 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4019 * 2) There is no busy sibling group to pull from.
4020 * 3) This group is the busiest group.
4021 * 4) This group is more busy than the avg busieness at this
4023 * 5) The imbalance is within the specified limit.
4024 * 6) Any rebalance would lead to ping-pong
4026 if (balance
&& !(*balance
))
4029 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4032 if (sds
.this_load
>= sds
.max_load
)
4035 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4037 if (sds
.this_load
>= sds
.avg_load
)
4040 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4043 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4045 sds
.busiest_load_per_task
=
4046 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4049 * We're trying to get all the cpus to the average_load, so we don't
4050 * want to push ourselves above the average load, nor do we wish to
4051 * reduce the max loaded cpu below the average load, as either of these
4052 * actions would just result in more rebalancing later, and ping-pong
4053 * tasks around. Thus we look for the minimum possible imbalance.
4054 * Negative imbalances (*we* are more loaded than anyone else) will
4055 * be counted as no imbalance for these purposes -- we can't fix that
4056 * by pulling tasks to us. Be careful of negative numbers as they'll
4057 * appear as very large values with unsigned longs.
4059 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4062 /* Looks like there is an imbalance. Compute it */
4063 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4068 * There is no obvious imbalance. But check if we can do some balancing
4071 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4079 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4082 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4083 unsigned long imbalance
, const struct cpumask
*cpus
)
4085 struct rq
*busiest
= NULL
, *rq
;
4086 unsigned long max_load
= 0;
4089 for_each_cpu(i
, sched_group_cpus(group
)) {
4090 unsigned long power
= power_of(i
);
4091 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4094 if (!cpumask_test_cpu(i
, cpus
))
4098 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4101 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4104 if (wl
> max_load
) {
4114 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4115 * so long as it is large enough.
4117 #define MAX_PINNED_INTERVAL 512
4119 /* Working cpumask for load_balance and load_balance_newidle. */
4120 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4123 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4124 * tasks if there is an imbalance.
4126 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4127 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4130 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4131 struct sched_group
*group
;
4132 unsigned long imbalance
;
4134 unsigned long flags
;
4135 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4137 cpumask_copy(cpus
, cpu_active_mask
);
4140 * When power savings policy is enabled for the parent domain, idle
4141 * sibling can pick up load irrespective of busy siblings. In this case,
4142 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4143 * portraying it as CPU_NOT_IDLE.
4145 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4146 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4149 schedstat_inc(sd
, lb_count
[idle
]);
4153 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4160 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4164 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4166 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4170 BUG_ON(busiest
== this_rq
);
4172 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4175 if (busiest
->nr_running
> 1) {
4177 * Attempt to move tasks. If find_busiest_group has found
4178 * an imbalance but busiest->nr_running <= 1, the group is
4179 * still unbalanced. ld_moved simply stays zero, so it is
4180 * correctly treated as an imbalance.
4182 local_irq_save(flags
);
4183 double_rq_lock(this_rq
, busiest
);
4184 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4185 imbalance
, sd
, idle
, &all_pinned
);
4186 double_rq_unlock(this_rq
, busiest
);
4187 local_irq_restore(flags
);
4190 * some other cpu did the load balance for us.
4192 if (ld_moved
&& this_cpu
!= smp_processor_id())
4193 resched_cpu(this_cpu
);
4195 /* All tasks on this runqueue were pinned by CPU affinity */
4196 if (unlikely(all_pinned
)) {
4197 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4198 if (!cpumask_empty(cpus
))
4205 schedstat_inc(sd
, lb_failed
[idle
]);
4206 sd
->nr_balance_failed
++;
4208 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4210 spin_lock_irqsave(&busiest
->lock
, flags
);
4212 /* don't kick the migration_thread, if the curr
4213 * task on busiest cpu can't be moved to this_cpu
4215 if (!cpumask_test_cpu(this_cpu
,
4216 &busiest
->curr
->cpus_allowed
)) {
4217 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4219 goto out_one_pinned
;
4222 if (!busiest
->active_balance
) {
4223 busiest
->active_balance
= 1;
4224 busiest
->push_cpu
= this_cpu
;
4227 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4229 wake_up_process(busiest
->migration_thread
);
4232 * We've kicked active balancing, reset the failure
4235 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4238 sd
->nr_balance_failed
= 0;
4240 if (likely(!active_balance
)) {
4241 /* We were unbalanced, so reset the balancing interval */
4242 sd
->balance_interval
= sd
->min_interval
;
4245 * If we've begun active balancing, start to back off. This
4246 * case may not be covered by the all_pinned logic if there
4247 * is only 1 task on the busy runqueue (because we don't call
4250 if (sd
->balance_interval
< sd
->max_interval
)
4251 sd
->balance_interval
*= 2;
4254 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4255 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4261 schedstat_inc(sd
, lb_balanced
[idle
]);
4263 sd
->nr_balance_failed
= 0;
4266 /* tune up the balancing interval */
4267 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4268 (sd
->balance_interval
< sd
->max_interval
))
4269 sd
->balance_interval
*= 2;
4271 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4272 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4283 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4284 * tasks if there is an imbalance.
4286 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4287 * this_rq is locked.
4290 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4292 struct sched_group
*group
;
4293 struct rq
*busiest
= NULL
;
4294 unsigned long imbalance
;
4298 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4300 cpumask_copy(cpus
, cpu_active_mask
);
4303 * When power savings policy is enabled for the parent domain, idle
4304 * sibling can pick up load irrespective of busy siblings. In this case,
4305 * let the state of idle sibling percolate up as IDLE, instead of
4306 * portraying it as CPU_NOT_IDLE.
4308 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4309 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4312 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4314 update_shares_locked(this_rq
, sd
);
4315 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4316 &sd_idle
, cpus
, NULL
);
4318 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4322 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4324 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4328 BUG_ON(busiest
== this_rq
);
4330 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4333 if (busiest
->nr_running
> 1) {
4334 /* Attempt to move tasks */
4335 double_lock_balance(this_rq
, busiest
);
4336 /* this_rq->clock is already updated */
4337 update_rq_clock(busiest
);
4338 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4339 imbalance
, sd
, CPU_NEWLY_IDLE
,
4341 double_unlock_balance(this_rq
, busiest
);
4343 if (unlikely(all_pinned
)) {
4344 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4345 if (!cpumask_empty(cpus
))
4351 int active_balance
= 0;
4353 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4354 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4355 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4358 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4361 if (sd
->nr_balance_failed
++ < 2)
4365 * The only task running in a non-idle cpu can be moved to this
4366 * cpu in an attempt to completely freeup the other CPU
4367 * package. The same method used to move task in load_balance()
4368 * have been extended for load_balance_newidle() to speedup
4369 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4371 * The package power saving logic comes from
4372 * find_busiest_group(). If there are no imbalance, then
4373 * f_b_g() will return NULL. However when sched_mc={1,2} then
4374 * f_b_g() will select a group from which a running task may be
4375 * pulled to this cpu in order to make the other package idle.
4376 * If there is no opportunity to make a package idle and if
4377 * there are no imbalance, then f_b_g() will return NULL and no
4378 * action will be taken in load_balance_newidle().
4380 * Under normal task pull operation due to imbalance, there
4381 * will be more than one task in the source run queue and
4382 * move_tasks() will succeed. ld_moved will be true and this
4383 * active balance code will not be triggered.
4386 /* Lock busiest in correct order while this_rq is held */
4387 double_lock_balance(this_rq
, busiest
);
4390 * don't kick the migration_thread, if the curr
4391 * task on busiest cpu can't be moved to this_cpu
4393 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4394 double_unlock_balance(this_rq
, busiest
);
4399 if (!busiest
->active_balance
) {
4400 busiest
->active_balance
= 1;
4401 busiest
->push_cpu
= this_cpu
;
4405 double_unlock_balance(this_rq
, busiest
);
4407 * Should not call ttwu while holding a rq->lock
4409 spin_unlock(&this_rq
->lock
);
4411 wake_up_process(busiest
->migration_thread
);
4412 spin_lock(&this_rq
->lock
);
4415 sd
->nr_balance_failed
= 0;
4417 update_shares_locked(this_rq
, sd
);
4421 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4422 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4423 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4425 sd
->nr_balance_failed
= 0;
4431 * idle_balance is called by schedule() if this_cpu is about to become
4432 * idle. Attempts to pull tasks from other CPUs.
4434 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4436 struct sched_domain
*sd
;
4437 int pulled_task
= 0;
4438 unsigned long next_balance
= jiffies
+ HZ
;
4440 this_rq
->idle_stamp
= this_rq
->clock
;
4442 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4445 for_each_domain(this_cpu
, sd
) {
4446 unsigned long interval
;
4448 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4451 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4452 /* If we've pulled tasks over stop searching: */
4453 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4456 interval
= msecs_to_jiffies(sd
->balance_interval
);
4457 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4458 next_balance
= sd
->last_balance
+ interval
;
4460 this_rq
->idle_stamp
= 0;
4464 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4466 * We are going idle. next_balance may be set based on
4467 * a busy processor. So reset next_balance.
4469 this_rq
->next_balance
= next_balance
;
4474 * active_load_balance is run by migration threads. It pushes running tasks
4475 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4476 * running on each physical CPU where possible, and avoids physical /
4477 * logical imbalances.
4479 * Called with busiest_rq locked.
4481 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4483 int target_cpu
= busiest_rq
->push_cpu
;
4484 struct sched_domain
*sd
;
4485 struct rq
*target_rq
;
4487 /* Is there any task to move? */
4488 if (busiest_rq
->nr_running
<= 1)
4491 target_rq
= cpu_rq(target_cpu
);
4494 * This condition is "impossible", if it occurs
4495 * we need to fix it. Originally reported by
4496 * Bjorn Helgaas on a 128-cpu setup.
4498 BUG_ON(busiest_rq
== target_rq
);
4500 /* move a task from busiest_rq to target_rq */
4501 double_lock_balance(busiest_rq
, target_rq
);
4502 update_rq_clock(busiest_rq
);
4503 update_rq_clock(target_rq
);
4505 /* Search for an sd spanning us and the target CPU. */
4506 for_each_domain(target_cpu
, sd
) {
4507 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4508 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4513 schedstat_inc(sd
, alb_count
);
4515 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4517 schedstat_inc(sd
, alb_pushed
);
4519 schedstat_inc(sd
, alb_failed
);
4521 double_unlock_balance(busiest_rq
, target_rq
);
4526 atomic_t load_balancer
;
4527 cpumask_var_t cpu_mask
;
4528 cpumask_var_t ilb_grp_nohz_mask
;
4529 } nohz ____cacheline_aligned
= {
4530 .load_balancer
= ATOMIC_INIT(-1),
4533 int get_nohz_load_balancer(void)
4535 return atomic_read(&nohz
.load_balancer
);
4538 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4540 * lowest_flag_domain - Return lowest sched_domain containing flag.
4541 * @cpu: The cpu whose lowest level of sched domain is to
4543 * @flag: The flag to check for the lowest sched_domain
4544 * for the given cpu.
4546 * Returns the lowest sched_domain of a cpu which contains the given flag.
4548 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4550 struct sched_domain
*sd
;
4552 for_each_domain(cpu
, sd
)
4553 if (sd
&& (sd
->flags
& flag
))
4560 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4561 * @cpu: The cpu whose domains we're iterating over.
4562 * @sd: variable holding the value of the power_savings_sd
4564 * @flag: The flag to filter the sched_domains to be iterated.
4566 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4567 * set, starting from the lowest sched_domain to the highest.
4569 #define for_each_flag_domain(cpu, sd, flag) \
4570 for (sd = lowest_flag_domain(cpu, flag); \
4571 (sd && (sd->flags & flag)); sd = sd->parent)
4574 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4575 * @ilb_group: group to be checked for semi-idleness
4577 * Returns: 1 if the group is semi-idle. 0 otherwise.
4579 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4580 * and atleast one non-idle CPU. This helper function checks if the given
4581 * sched_group is semi-idle or not.
4583 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4585 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4586 sched_group_cpus(ilb_group
));
4589 * A sched_group is semi-idle when it has atleast one busy cpu
4590 * and atleast one idle cpu.
4592 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4595 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4601 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4602 * @cpu: The cpu which is nominating a new idle_load_balancer.
4604 * Returns: Returns the id of the idle load balancer if it exists,
4605 * Else, returns >= nr_cpu_ids.
4607 * This algorithm picks the idle load balancer such that it belongs to a
4608 * semi-idle powersavings sched_domain. The idea is to try and avoid
4609 * completely idle packages/cores just for the purpose of idle load balancing
4610 * when there are other idle cpu's which are better suited for that job.
4612 static int find_new_ilb(int cpu
)
4614 struct sched_domain
*sd
;
4615 struct sched_group
*ilb_group
;
4618 * Have idle load balancer selection from semi-idle packages only
4619 * when power-aware load balancing is enabled
4621 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4625 * Optimize for the case when we have no idle CPUs or only one
4626 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4628 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4631 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4632 ilb_group
= sd
->groups
;
4635 if (is_semi_idle_group(ilb_group
))
4636 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4638 ilb_group
= ilb_group
->next
;
4640 } while (ilb_group
!= sd
->groups
);
4644 return cpumask_first(nohz
.cpu_mask
);
4646 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4647 static inline int find_new_ilb(int call_cpu
)
4649 return cpumask_first(nohz
.cpu_mask
);
4654 * This routine will try to nominate the ilb (idle load balancing)
4655 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4656 * load balancing on behalf of all those cpus. If all the cpus in the system
4657 * go into this tickless mode, then there will be no ilb owner (as there is
4658 * no need for one) and all the cpus will sleep till the next wakeup event
4661 * For the ilb owner, tick is not stopped. And this tick will be used
4662 * for idle load balancing. ilb owner will still be part of
4665 * While stopping the tick, this cpu will become the ilb owner if there
4666 * is no other owner. And will be the owner till that cpu becomes busy
4667 * or if all cpus in the system stop their ticks at which point
4668 * there is no need for ilb owner.
4670 * When the ilb owner becomes busy, it nominates another owner, during the
4671 * next busy scheduler_tick()
4673 int select_nohz_load_balancer(int stop_tick
)
4675 int cpu
= smp_processor_id();
4678 cpu_rq(cpu
)->in_nohz_recently
= 1;
4680 if (!cpu_active(cpu
)) {
4681 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4685 * If we are going offline and still the leader,
4688 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4694 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4696 /* time for ilb owner also to sleep */
4697 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
4698 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4699 atomic_set(&nohz
.load_balancer
, -1);
4703 if (atomic_read(&nohz
.load_balancer
) == -1) {
4704 /* make me the ilb owner */
4705 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4707 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4710 if (!(sched_smt_power_savings
||
4711 sched_mc_power_savings
))
4714 * Check to see if there is a more power-efficient
4717 new_ilb
= find_new_ilb(cpu
);
4718 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4719 atomic_set(&nohz
.load_balancer
, -1);
4720 resched_cpu(new_ilb
);
4726 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4729 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4731 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4732 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4739 static DEFINE_SPINLOCK(balancing
);
4742 * It checks each scheduling domain to see if it is due to be balanced,
4743 * and initiates a balancing operation if so.
4745 * Balancing parameters are set up in arch_init_sched_domains.
4747 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4750 struct rq
*rq
= cpu_rq(cpu
);
4751 unsigned long interval
;
4752 struct sched_domain
*sd
;
4753 /* Earliest time when we have to do rebalance again */
4754 unsigned long next_balance
= jiffies
+ 60*HZ
;
4755 int update_next_balance
= 0;
4758 for_each_domain(cpu
, sd
) {
4759 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4762 interval
= sd
->balance_interval
;
4763 if (idle
!= CPU_IDLE
)
4764 interval
*= sd
->busy_factor
;
4766 /* scale ms to jiffies */
4767 interval
= msecs_to_jiffies(interval
);
4768 if (unlikely(!interval
))
4770 if (interval
> HZ
*NR_CPUS
/10)
4771 interval
= HZ
*NR_CPUS
/10;
4773 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4775 if (need_serialize
) {
4776 if (!spin_trylock(&balancing
))
4780 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4781 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4783 * We've pulled tasks over so either we're no
4784 * longer idle, or one of our SMT siblings is
4787 idle
= CPU_NOT_IDLE
;
4789 sd
->last_balance
= jiffies
;
4792 spin_unlock(&balancing
);
4794 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4795 next_balance
= sd
->last_balance
+ interval
;
4796 update_next_balance
= 1;
4800 * Stop the load balance at this level. There is another
4801 * CPU in our sched group which is doing load balancing more
4809 * next_balance will be updated only when there is a need.
4810 * When the cpu is attached to null domain for ex, it will not be
4813 if (likely(update_next_balance
))
4814 rq
->next_balance
= next_balance
;
4818 * run_rebalance_domains is triggered when needed from the scheduler tick.
4819 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4820 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4822 static void run_rebalance_domains(struct softirq_action
*h
)
4824 int this_cpu
= smp_processor_id();
4825 struct rq
*this_rq
= cpu_rq(this_cpu
);
4826 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4827 CPU_IDLE
: CPU_NOT_IDLE
;
4829 rebalance_domains(this_cpu
, idle
);
4833 * If this cpu is the owner for idle load balancing, then do the
4834 * balancing on behalf of the other idle cpus whose ticks are
4837 if (this_rq
->idle_at_tick
&&
4838 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4842 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4843 if (balance_cpu
== this_cpu
)
4847 * If this cpu gets work to do, stop the load balancing
4848 * work being done for other cpus. Next load
4849 * balancing owner will pick it up.
4854 rebalance_domains(balance_cpu
, CPU_IDLE
);
4856 rq
= cpu_rq(balance_cpu
);
4857 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4858 this_rq
->next_balance
= rq
->next_balance
;
4864 static inline int on_null_domain(int cpu
)
4866 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4870 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4872 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4873 * idle load balancing owner or decide to stop the periodic load balancing,
4874 * if the whole system is idle.
4876 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4880 * If we were in the nohz mode recently and busy at the current
4881 * scheduler tick, then check if we need to nominate new idle
4884 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4885 rq
->in_nohz_recently
= 0;
4887 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4888 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4889 atomic_set(&nohz
.load_balancer
, -1);
4892 if (atomic_read(&nohz
.load_balancer
) == -1) {
4893 int ilb
= find_new_ilb(cpu
);
4895 if (ilb
< nr_cpu_ids
)
4901 * If this cpu is idle and doing idle load balancing for all the
4902 * cpus with ticks stopped, is it time for that to stop?
4904 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4905 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4911 * If this cpu is idle and the idle load balancing is done by
4912 * someone else, then no need raise the SCHED_SOFTIRQ
4914 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4915 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4918 /* Don't need to rebalance while attached to NULL domain */
4919 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4920 likely(!on_null_domain(cpu
)))
4921 raise_softirq(SCHED_SOFTIRQ
);
4924 #else /* CONFIG_SMP */
4927 * on UP we do not need to balance between CPUs:
4929 static inline void idle_balance(int cpu
, struct rq
*rq
)
4935 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4937 EXPORT_PER_CPU_SYMBOL(kstat
);
4940 * Return any ns on the sched_clock that have not yet been accounted in
4941 * @p in case that task is currently running.
4943 * Called with task_rq_lock() held on @rq.
4945 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4949 if (task_current(rq
, p
)) {
4950 update_rq_clock(rq
);
4951 ns
= rq
->clock
- p
->se
.exec_start
;
4959 unsigned long long task_delta_exec(struct task_struct
*p
)
4961 unsigned long flags
;
4965 rq
= task_rq_lock(p
, &flags
);
4966 ns
= do_task_delta_exec(p
, rq
);
4967 task_rq_unlock(rq
, &flags
);
4973 * Return accounted runtime for the task.
4974 * In case the task is currently running, return the runtime plus current's
4975 * pending runtime that have not been accounted yet.
4977 unsigned long long task_sched_runtime(struct task_struct
*p
)
4979 unsigned long flags
;
4983 rq
= task_rq_lock(p
, &flags
);
4984 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4985 task_rq_unlock(rq
, &flags
);
4991 * Return sum_exec_runtime for the thread group.
4992 * In case the task is currently running, return the sum plus current's
4993 * pending runtime that have not been accounted yet.
4995 * Note that the thread group might have other running tasks as well,
4996 * so the return value not includes other pending runtime that other
4997 * running tasks might have.
4999 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5001 struct task_cputime totals
;
5002 unsigned long flags
;
5006 rq
= task_rq_lock(p
, &flags
);
5007 thread_group_cputime(p
, &totals
);
5008 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5009 task_rq_unlock(rq
, &flags
);
5015 * Account user cpu time to a process.
5016 * @p: the process that the cpu time gets accounted to
5017 * @cputime: the cpu time spent in user space since the last update
5018 * @cputime_scaled: cputime scaled by cpu frequency
5020 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5021 cputime_t cputime_scaled
)
5023 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5026 /* Add user time to process. */
5027 p
->utime
= cputime_add(p
->utime
, cputime
);
5028 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5029 account_group_user_time(p
, cputime
);
5031 /* Add user time to cpustat. */
5032 tmp
= cputime_to_cputime64(cputime
);
5033 if (TASK_NICE(p
) > 0)
5034 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5036 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5038 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5039 /* Account for user time used */
5040 acct_update_integrals(p
);
5044 * Account guest cpu time to a process.
5045 * @p: the process that the cpu time gets accounted to
5046 * @cputime: the cpu time spent in virtual machine since the last update
5047 * @cputime_scaled: cputime scaled by cpu frequency
5049 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5050 cputime_t cputime_scaled
)
5053 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5055 tmp
= cputime_to_cputime64(cputime
);
5057 /* Add guest time to process. */
5058 p
->utime
= cputime_add(p
->utime
, cputime
);
5059 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5060 account_group_user_time(p
, cputime
);
5061 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5063 /* Add guest time to cpustat. */
5064 if (TASK_NICE(p
) > 0) {
5065 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5066 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
5068 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5069 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 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5191 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5193 struct task_cputime cputime
;
5195 thread_group_cputime(p
, &cputime
);
5197 *ut
= cputime
.utime
;
5198 *st
= cputime
.stime
;
5202 #ifndef nsecs_to_cputime
5203 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5206 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5208 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
5211 * Use CFS's precise accounting:
5213 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
5218 temp
= (u64
)(rtime
* utime
);
5219 do_div(temp
, total
);
5220 utime
= (cputime_t
)temp
;
5225 * Compare with previous values, to keep monotonicity:
5227 p
->prev_utime
= max(p
->prev_utime
, utime
);
5228 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
5230 *ut
= p
->prev_utime
;
5231 *st
= p
->prev_stime
;
5235 * Must be called with siglock held.
5237 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5239 struct signal_struct
*sig
= p
->signal
;
5240 struct task_cputime cputime
;
5241 cputime_t rtime
, utime
, total
;
5243 thread_group_cputime(p
, &cputime
);
5245 total
= cputime_add(cputime
.utime
, cputime
.stime
);
5246 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
5251 temp
= (u64
)(rtime
* cputime
.utime
);
5252 do_div(temp
, total
);
5253 utime
= (cputime_t
)temp
;
5257 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
5258 sig
->prev_stime
= max(sig
->prev_stime
,
5259 cputime_sub(rtime
, sig
->prev_utime
));
5261 *ut
= sig
->prev_utime
;
5262 *st
= sig
->prev_stime
;
5267 * This function gets called by the timer code, with HZ frequency.
5268 * We call it with interrupts disabled.
5270 * It also gets called by the fork code, when changing the parent's
5273 void scheduler_tick(void)
5275 int cpu
= smp_processor_id();
5276 struct rq
*rq
= cpu_rq(cpu
);
5277 struct task_struct
*curr
= rq
->curr
;
5281 spin_lock(&rq
->lock
);
5282 update_rq_clock(rq
);
5283 update_cpu_load(rq
);
5284 curr
->sched_class
->task_tick(rq
, curr
, 0);
5285 spin_unlock(&rq
->lock
);
5287 perf_event_task_tick(curr
, cpu
);
5290 rq
->idle_at_tick
= idle_cpu(cpu
);
5291 trigger_load_balance(rq
, cpu
);
5295 notrace
unsigned long get_parent_ip(unsigned long addr
)
5297 if (in_lock_functions(addr
)) {
5298 addr
= CALLER_ADDR2
;
5299 if (in_lock_functions(addr
))
5300 addr
= CALLER_ADDR3
;
5305 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5306 defined(CONFIG_PREEMPT_TRACER))
5308 void __kprobes
add_preempt_count(int val
)
5310 #ifdef CONFIG_DEBUG_PREEMPT
5314 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5317 preempt_count() += val
;
5318 #ifdef CONFIG_DEBUG_PREEMPT
5320 * Spinlock count overflowing soon?
5322 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5325 if (preempt_count() == val
)
5326 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5328 EXPORT_SYMBOL(add_preempt_count
);
5330 void __kprobes
sub_preempt_count(int val
)
5332 #ifdef CONFIG_DEBUG_PREEMPT
5336 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5339 * Is the spinlock portion underflowing?
5341 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5342 !(preempt_count() & PREEMPT_MASK
)))
5346 if (preempt_count() == val
)
5347 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5348 preempt_count() -= val
;
5350 EXPORT_SYMBOL(sub_preempt_count
);
5355 * Print scheduling while atomic bug:
5357 static noinline
void __schedule_bug(struct task_struct
*prev
)
5359 struct pt_regs
*regs
= get_irq_regs();
5361 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5362 prev
->comm
, prev
->pid
, preempt_count());
5364 debug_show_held_locks(prev
);
5366 if (irqs_disabled())
5367 print_irqtrace_events(prev
);
5376 * Various schedule()-time debugging checks and statistics:
5378 static inline void schedule_debug(struct task_struct
*prev
)
5381 * Test if we are atomic. Since do_exit() needs to call into
5382 * schedule() atomically, we ignore that path for now.
5383 * Otherwise, whine if we are scheduling when we should not be.
5385 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5386 __schedule_bug(prev
);
5388 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5390 schedstat_inc(this_rq(), sched_count
);
5391 #ifdef CONFIG_SCHEDSTATS
5392 if (unlikely(prev
->lock_depth
>= 0)) {
5393 schedstat_inc(this_rq(), bkl_count
);
5394 schedstat_inc(prev
, sched_info
.bkl_count
);
5399 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5401 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5403 update_avg(&p
->se
.avg_running
, runtime
);
5405 if (p
->state
== TASK_RUNNING
) {
5407 * In order to avoid avg_overlap growing stale when we are
5408 * indeed overlapping and hence not getting put to sleep, grow
5409 * the avg_overlap on preemption.
5411 * We use the average preemption runtime because that
5412 * correlates to the amount of cache footprint a task can
5415 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5416 update_avg(&p
->se
.avg_overlap
, runtime
);
5418 update_avg(&p
->se
.avg_running
, 0);
5420 p
->sched_class
->put_prev_task(rq
, p
);
5424 * Pick up the highest-prio task:
5426 static inline struct task_struct
*
5427 pick_next_task(struct rq
*rq
)
5429 const struct sched_class
*class;
5430 struct task_struct
*p
;
5433 * Optimization: we know that if all tasks are in
5434 * the fair class we can call that function directly:
5436 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5437 p
= fair_sched_class
.pick_next_task(rq
);
5442 class = sched_class_highest
;
5444 p
= class->pick_next_task(rq
);
5448 * Will never be NULL as the idle class always
5449 * returns a non-NULL p:
5451 class = class->next
;
5456 * schedule() is the main scheduler function.
5458 asmlinkage
void __sched
schedule(void)
5460 struct task_struct
*prev
, *next
;
5461 unsigned long *switch_count
;
5467 cpu
= smp_processor_id();
5471 switch_count
= &prev
->nivcsw
;
5473 release_kernel_lock(prev
);
5474 need_resched_nonpreemptible
:
5476 schedule_debug(prev
);
5478 if (sched_feat(HRTICK
))
5481 spin_lock_irq(&rq
->lock
);
5482 update_rq_clock(rq
);
5483 clear_tsk_need_resched(prev
);
5485 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5486 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5487 prev
->state
= TASK_RUNNING
;
5489 deactivate_task(rq
, prev
, 1);
5490 switch_count
= &prev
->nvcsw
;
5493 pre_schedule(rq
, prev
);
5495 if (unlikely(!rq
->nr_running
))
5496 idle_balance(cpu
, rq
);
5498 put_prev_task(rq
, prev
);
5499 next
= pick_next_task(rq
);
5501 if (likely(prev
!= next
)) {
5502 sched_info_switch(prev
, next
);
5503 perf_event_task_sched_out(prev
, next
, cpu
);
5509 context_switch(rq
, prev
, next
); /* unlocks the rq */
5511 * the context switch might have flipped the stack from under
5512 * us, hence refresh the local variables.
5514 cpu
= smp_processor_id();
5517 spin_unlock_irq(&rq
->lock
);
5521 if (unlikely(reacquire_kernel_lock(current
) < 0))
5522 goto need_resched_nonpreemptible
;
5524 preempt_enable_no_resched();
5528 EXPORT_SYMBOL(schedule
);
5530 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5532 * Look out! "owner" is an entirely speculative pointer
5533 * access and not reliable.
5535 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5540 if (!sched_feat(OWNER_SPIN
))
5543 #ifdef CONFIG_DEBUG_PAGEALLOC
5545 * Need to access the cpu field knowing that
5546 * DEBUG_PAGEALLOC could have unmapped it if
5547 * the mutex owner just released it and exited.
5549 if (probe_kernel_address(&owner
->cpu
, cpu
))
5556 * Even if the access succeeded (likely case),
5557 * the cpu field may no longer be valid.
5559 if (cpu
>= nr_cpumask_bits
)
5563 * We need to validate that we can do a
5564 * get_cpu() and that we have the percpu area.
5566 if (!cpu_online(cpu
))
5573 * Owner changed, break to re-assess state.
5575 if (lock
->owner
!= owner
)
5579 * Is that owner really running on that cpu?
5581 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5591 #ifdef CONFIG_PREEMPT
5593 * this is the entry point to schedule() from in-kernel preemption
5594 * off of preempt_enable. Kernel preemptions off return from interrupt
5595 * occur there and call schedule directly.
5597 asmlinkage
void __sched
preempt_schedule(void)
5599 struct thread_info
*ti
= current_thread_info();
5602 * If there is a non-zero preempt_count or interrupts are disabled,
5603 * we do not want to preempt the current task. Just return..
5605 if (likely(ti
->preempt_count
|| irqs_disabled()))
5609 add_preempt_count(PREEMPT_ACTIVE
);
5611 sub_preempt_count(PREEMPT_ACTIVE
);
5614 * Check again in case we missed a preemption opportunity
5615 * between schedule and now.
5618 } while (need_resched());
5620 EXPORT_SYMBOL(preempt_schedule
);
5623 * this is the entry point to schedule() from kernel preemption
5624 * off of irq context.
5625 * Note, that this is called and return with irqs disabled. This will
5626 * protect us against recursive calling from irq.
5628 asmlinkage
void __sched
preempt_schedule_irq(void)
5630 struct thread_info
*ti
= current_thread_info();
5632 /* Catch callers which need to be fixed */
5633 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5636 add_preempt_count(PREEMPT_ACTIVE
);
5639 local_irq_disable();
5640 sub_preempt_count(PREEMPT_ACTIVE
);
5643 * Check again in case we missed a preemption opportunity
5644 * between schedule and now.
5647 } while (need_resched());
5650 #endif /* CONFIG_PREEMPT */
5652 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5655 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5657 EXPORT_SYMBOL(default_wake_function
);
5660 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5661 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5662 * number) then we wake all the non-exclusive tasks and one exclusive task.
5664 * There are circumstances in which we can try to wake a task which has already
5665 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5666 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5668 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5669 int nr_exclusive
, int wake_flags
, void *key
)
5671 wait_queue_t
*curr
, *next
;
5673 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5674 unsigned flags
= curr
->flags
;
5676 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5677 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5683 * __wake_up - wake up threads blocked on a waitqueue.
5685 * @mode: which threads
5686 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5687 * @key: is directly passed to the wakeup function
5689 * It may be assumed that this function implies a write memory barrier before
5690 * changing the task state if and only if any tasks are woken up.
5692 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5693 int nr_exclusive
, void *key
)
5695 unsigned long flags
;
5697 spin_lock_irqsave(&q
->lock
, flags
);
5698 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5699 spin_unlock_irqrestore(&q
->lock
, flags
);
5701 EXPORT_SYMBOL(__wake_up
);
5704 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5706 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5708 __wake_up_common(q
, mode
, 1, 0, NULL
);
5711 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5713 __wake_up_common(q
, mode
, 1, 0, key
);
5717 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5719 * @mode: which threads
5720 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5721 * @key: opaque value to be passed to wakeup targets
5723 * The sync wakeup differs that the waker knows that it will schedule
5724 * away soon, so while the target thread will be woken up, it will not
5725 * be migrated to another CPU - ie. the two threads are 'synchronized'
5726 * with each other. This can prevent needless bouncing between CPUs.
5728 * On UP it can prevent extra preemption.
5730 * It may be assumed that this function implies a write memory barrier before
5731 * changing the task state if and only if any tasks are woken up.
5733 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5734 int nr_exclusive
, void *key
)
5736 unsigned long flags
;
5737 int wake_flags
= WF_SYNC
;
5742 if (unlikely(!nr_exclusive
))
5745 spin_lock_irqsave(&q
->lock
, flags
);
5746 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5747 spin_unlock_irqrestore(&q
->lock
, flags
);
5749 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5752 * __wake_up_sync - see __wake_up_sync_key()
5754 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5756 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5758 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5761 * complete: - signals a single thread waiting on this completion
5762 * @x: holds the state of this particular completion
5764 * This will wake up a single thread waiting on this completion. Threads will be
5765 * awakened in the same order in which they were queued.
5767 * See also complete_all(), wait_for_completion() and related routines.
5769 * It may be assumed that this function implies a write memory barrier before
5770 * changing the task state if and only if any tasks are woken up.
5772 void complete(struct completion
*x
)
5774 unsigned long flags
;
5776 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5778 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5779 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5781 EXPORT_SYMBOL(complete
);
5784 * complete_all: - signals all threads waiting on this completion
5785 * @x: holds the state of this particular completion
5787 * This will wake up all threads waiting on this particular completion event.
5789 * It may be assumed that this function implies a write memory barrier before
5790 * changing the task state if and only if any tasks are woken up.
5792 void complete_all(struct completion
*x
)
5794 unsigned long flags
;
5796 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5797 x
->done
+= UINT_MAX
/2;
5798 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5799 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5801 EXPORT_SYMBOL(complete_all
);
5803 static inline long __sched
5804 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5807 DECLARE_WAITQUEUE(wait
, current
);
5809 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5810 __add_wait_queue_tail(&x
->wait
, &wait
);
5812 if (signal_pending_state(state
, current
)) {
5813 timeout
= -ERESTARTSYS
;
5816 __set_current_state(state
);
5817 spin_unlock_irq(&x
->wait
.lock
);
5818 timeout
= schedule_timeout(timeout
);
5819 spin_lock_irq(&x
->wait
.lock
);
5820 } while (!x
->done
&& timeout
);
5821 __remove_wait_queue(&x
->wait
, &wait
);
5826 return timeout
?: 1;
5830 wait_for_common(struct completion
*x
, long timeout
, int state
)
5834 spin_lock_irq(&x
->wait
.lock
);
5835 timeout
= do_wait_for_common(x
, timeout
, state
);
5836 spin_unlock_irq(&x
->wait
.lock
);
5841 * wait_for_completion: - waits for completion of a task
5842 * @x: holds the state of this particular completion
5844 * This waits to be signaled for completion of a specific task. It is NOT
5845 * interruptible and there is no timeout.
5847 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5848 * and interrupt capability. Also see complete().
5850 void __sched
wait_for_completion(struct completion
*x
)
5852 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5854 EXPORT_SYMBOL(wait_for_completion
);
5857 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5858 * @x: holds the state of this particular completion
5859 * @timeout: timeout value in jiffies
5861 * This waits for either a completion of a specific task to be signaled or for a
5862 * specified timeout to expire. The timeout is in jiffies. It is not
5865 unsigned long __sched
5866 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5868 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5870 EXPORT_SYMBOL(wait_for_completion_timeout
);
5873 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5874 * @x: holds the state of this particular completion
5876 * This waits for completion of a specific task to be signaled. It is
5879 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5881 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5882 if (t
== -ERESTARTSYS
)
5886 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5889 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5890 * @x: holds the state of this particular completion
5891 * @timeout: timeout value in jiffies
5893 * This waits for either a completion of a specific task to be signaled or for a
5894 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5896 unsigned long __sched
5897 wait_for_completion_interruptible_timeout(struct completion
*x
,
5898 unsigned long timeout
)
5900 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5902 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5905 * wait_for_completion_killable: - waits for completion of a task (killable)
5906 * @x: holds the state of this particular completion
5908 * This waits to be signaled for completion of a specific task. It can be
5909 * interrupted by a kill signal.
5911 int __sched
wait_for_completion_killable(struct completion
*x
)
5913 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5914 if (t
== -ERESTARTSYS
)
5918 EXPORT_SYMBOL(wait_for_completion_killable
);
5921 * try_wait_for_completion - try to decrement a completion without blocking
5922 * @x: completion structure
5924 * Returns: 0 if a decrement cannot be done without blocking
5925 * 1 if a decrement succeeded.
5927 * If a completion is being used as a counting completion,
5928 * attempt to decrement the counter without blocking. This
5929 * enables us to avoid waiting if the resource the completion
5930 * is protecting is not available.
5932 bool try_wait_for_completion(struct completion
*x
)
5936 spin_lock_irq(&x
->wait
.lock
);
5941 spin_unlock_irq(&x
->wait
.lock
);
5944 EXPORT_SYMBOL(try_wait_for_completion
);
5947 * completion_done - Test to see if a completion has any waiters
5948 * @x: completion structure
5950 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5951 * 1 if there are no waiters.
5954 bool completion_done(struct completion
*x
)
5958 spin_lock_irq(&x
->wait
.lock
);
5961 spin_unlock_irq(&x
->wait
.lock
);
5964 EXPORT_SYMBOL(completion_done
);
5967 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5969 unsigned long flags
;
5972 init_waitqueue_entry(&wait
, current
);
5974 __set_current_state(state
);
5976 spin_lock_irqsave(&q
->lock
, flags
);
5977 __add_wait_queue(q
, &wait
);
5978 spin_unlock(&q
->lock
);
5979 timeout
= schedule_timeout(timeout
);
5980 spin_lock_irq(&q
->lock
);
5981 __remove_wait_queue(q
, &wait
);
5982 spin_unlock_irqrestore(&q
->lock
, flags
);
5987 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5989 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5991 EXPORT_SYMBOL(interruptible_sleep_on
);
5994 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5996 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5998 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6000 void __sched
sleep_on(wait_queue_head_t
*q
)
6002 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6004 EXPORT_SYMBOL(sleep_on
);
6006 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6008 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6010 EXPORT_SYMBOL(sleep_on_timeout
);
6012 #ifdef CONFIG_RT_MUTEXES
6015 * rt_mutex_setprio - set the current priority of a task
6017 * @prio: prio value (kernel-internal form)
6019 * This function changes the 'effective' priority of a task. It does
6020 * not touch ->normal_prio like __setscheduler().
6022 * Used by the rt_mutex code to implement priority inheritance logic.
6024 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6026 unsigned long flags
;
6027 int oldprio
, on_rq
, running
;
6029 const struct sched_class
*prev_class
= p
->sched_class
;
6031 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6033 rq
= task_rq_lock(p
, &flags
);
6034 update_rq_clock(rq
);
6037 on_rq
= p
->se
.on_rq
;
6038 running
= task_current(rq
, p
);
6040 dequeue_task(rq
, p
, 0);
6042 p
->sched_class
->put_prev_task(rq
, p
);
6045 p
->sched_class
= &rt_sched_class
;
6047 p
->sched_class
= &fair_sched_class
;
6052 p
->sched_class
->set_curr_task(rq
);
6054 enqueue_task(rq
, p
, 0);
6056 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6058 task_rq_unlock(rq
, &flags
);
6063 void set_user_nice(struct task_struct
*p
, long nice
)
6065 int old_prio
, delta
, on_rq
;
6066 unsigned long flags
;
6069 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6072 * We have to be careful, if called from sys_setpriority(),
6073 * the task might be in the middle of scheduling on another CPU.
6075 rq
= task_rq_lock(p
, &flags
);
6076 update_rq_clock(rq
);
6078 * The RT priorities are set via sched_setscheduler(), but we still
6079 * allow the 'normal' nice value to be set - but as expected
6080 * it wont have any effect on scheduling until the task is
6081 * SCHED_FIFO/SCHED_RR:
6083 if (task_has_rt_policy(p
)) {
6084 p
->static_prio
= NICE_TO_PRIO(nice
);
6087 on_rq
= p
->se
.on_rq
;
6089 dequeue_task(rq
, p
, 0);
6091 p
->static_prio
= NICE_TO_PRIO(nice
);
6094 p
->prio
= effective_prio(p
);
6095 delta
= p
->prio
- old_prio
;
6098 enqueue_task(rq
, p
, 0);
6100 * If the task increased its priority or is running and
6101 * lowered its priority, then reschedule its CPU:
6103 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6104 resched_task(rq
->curr
);
6107 task_rq_unlock(rq
, &flags
);
6109 EXPORT_SYMBOL(set_user_nice
);
6112 * can_nice - check if a task can reduce its nice value
6116 int can_nice(const struct task_struct
*p
, const int nice
)
6118 /* convert nice value [19,-20] to rlimit style value [1,40] */
6119 int nice_rlim
= 20 - nice
;
6121 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6122 capable(CAP_SYS_NICE
));
6125 #ifdef __ARCH_WANT_SYS_NICE
6128 * sys_nice - change the priority of the current process.
6129 * @increment: priority increment
6131 * sys_setpriority is a more generic, but much slower function that
6132 * does similar things.
6134 SYSCALL_DEFINE1(nice
, int, increment
)
6139 * Setpriority might change our priority at the same moment.
6140 * We don't have to worry. Conceptually one call occurs first
6141 * and we have a single winner.
6143 if (increment
< -40)
6148 nice
= TASK_NICE(current
) + increment
;
6154 if (increment
< 0 && !can_nice(current
, nice
))
6157 retval
= security_task_setnice(current
, nice
);
6161 set_user_nice(current
, nice
);
6168 * task_prio - return the priority value of a given task.
6169 * @p: the task in question.
6171 * This is the priority value as seen by users in /proc.
6172 * RT tasks are offset by -200. Normal tasks are centered
6173 * around 0, value goes from -16 to +15.
6175 int task_prio(const struct task_struct
*p
)
6177 return p
->prio
- MAX_RT_PRIO
;
6181 * task_nice - return the nice value of a given task.
6182 * @p: the task in question.
6184 int task_nice(const struct task_struct
*p
)
6186 return TASK_NICE(p
);
6188 EXPORT_SYMBOL(task_nice
);
6191 * idle_cpu - is a given cpu idle currently?
6192 * @cpu: the processor in question.
6194 int idle_cpu(int cpu
)
6196 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6200 * idle_task - return the idle task for a given cpu.
6201 * @cpu: the processor in question.
6203 struct task_struct
*idle_task(int cpu
)
6205 return cpu_rq(cpu
)->idle
;
6209 * find_process_by_pid - find a process with a matching PID value.
6210 * @pid: the pid in question.
6212 static struct task_struct
*find_process_by_pid(pid_t pid
)
6214 return pid
? find_task_by_vpid(pid
) : current
;
6217 /* Actually do priority change: must hold rq lock. */
6219 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6221 BUG_ON(p
->se
.on_rq
);
6224 p
->rt_priority
= prio
;
6225 p
->normal_prio
= normal_prio(p
);
6226 /* we are holding p->pi_lock already */
6227 p
->prio
= rt_mutex_getprio(p
);
6228 if (rt_prio(p
->prio
))
6229 p
->sched_class
= &rt_sched_class
;
6231 p
->sched_class
= &fair_sched_class
;
6236 * check the target process has a UID that matches the current process's
6238 static bool check_same_owner(struct task_struct
*p
)
6240 const struct cred
*cred
= current_cred(), *pcred
;
6244 pcred
= __task_cred(p
);
6245 match
= (cred
->euid
== pcred
->euid
||
6246 cred
->euid
== pcred
->uid
);
6251 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6252 struct sched_param
*param
, bool user
)
6254 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6255 unsigned long flags
;
6256 const struct sched_class
*prev_class
= p
->sched_class
;
6260 /* may grab non-irq protected spin_locks */
6261 BUG_ON(in_interrupt());
6263 /* double check policy once rq lock held */
6265 reset_on_fork
= p
->sched_reset_on_fork
;
6266 policy
= oldpolicy
= p
->policy
;
6268 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6269 policy
&= ~SCHED_RESET_ON_FORK
;
6271 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6272 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6273 policy
!= SCHED_IDLE
)
6278 * Valid priorities for SCHED_FIFO and SCHED_RR are
6279 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6280 * SCHED_BATCH and SCHED_IDLE is 0.
6282 if (param
->sched_priority
< 0 ||
6283 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6284 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6286 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6290 * Allow unprivileged RT tasks to decrease priority:
6292 if (user
&& !capable(CAP_SYS_NICE
)) {
6293 if (rt_policy(policy
)) {
6294 unsigned long rlim_rtprio
;
6296 if (!lock_task_sighand(p
, &flags
))
6298 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6299 unlock_task_sighand(p
, &flags
);
6301 /* can't set/change the rt policy */
6302 if (policy
!= p
->policy
&& !rlim_rtprio
)
6305 /* can't increase priority */
6306 if (param
->sched_priority
> p
->rt_priority
&&
6307 param
->sched_priority
> rlim_rtprio
)
6311 * Like positive nice levels, dont allow tasks to
6312 * move out of SCHED_IDLE either:
6314 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6317 /* can't change other user's priorities */
6318 if (!check_same_owner(p
))
6321 /* Normal users shall not reset the sched_reset_on_fork flag */
6322 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6327 #ifdef CONFIG_RT_GROUP_SCHED
6329 * Do not allow realtime tasks into groups that have no runtime
6332 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6333 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6337 retval
= security_task_setscheduler(p
, policy
, param
);
6343 * make sure no PI-waiters arrive (or leave) while we are
6344 * changing the priority of the task:
6346 spin_lock_irqsave(&p
->pi_lock
, flags
);
6348 * To be able to change p->policy safely, the apropriate
6349 * runqueue lock must be held.
6351 rq
= __task_rq_lock(p
);
6352 /* recheck policy now with rq lock held */
6353 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6354 policy
= oldpolicy
= -1;
6355 __task_rq_unlock(rq
);
6356 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6359 update_rq_clock(rq
);
6360 on_rq
= p
->se
.on_rq
;
6361 running
= task_current(rq
, p
);
6363 deactivate_task(rq
, p
, 0);
6365 p
->sched_class
->put_prev_task(rq
, p
);
6367 p
->sched_reset_on_fork
= reset_on_fork
;
6370 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6373 p
->sched_class
->set_curr_task(rq
);
6375 activate_task(rq
, p
, 0);
6377 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6379 __task_rq_unlock(rq
);
6380 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6382 rt_mutex_adjust_pi(p
);
6388 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6389 * @p: the task in question.
6390 * @policy: new policy.
6391 * @param: structure containing the new RT priority.
6393 * NOTE that the task may be already dead.
6395 int sched_setscheduler(struct task_struct
*p
, int policy
,
6396 struct sched_param
*param
)
6398 return __sched_setscheduler(p
, policy
, param
, true);
6400 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6403 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6404 * @p: the task in question.
6405 * @policy: new policy.
6406 * @param: structure containing the new RT priority.
6408 * Just like sched_setscheduler, only don't bother checking if the
6409 * current context has permission. For example, this is needed in
6410 * stop_machine(): we create temporary high priority worker threads,
6411 * but our caller might not have that capability.
6413 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6414 struct sched_param
*param
)
6416 return __sched_setscheduler(p
, policy
, param
, false);
6420 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6422 struct sched_param lparam
;
6423 struct task_struct
*p
;
6426 if (!param
|| pid
< 0)
6428 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6433 p
= find_process_by_pid(pid
);
6435 retval
= sched_setscheduler(p
, policy
, &lparam
);
6442 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6443 * @pid: the pid in question.
6444 * @policy: new policy.
6445 * @param: structure containing the new RT priority.
6447 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6448 struct sched_param __user
*, param
)
6450 /* negative values for policy are not valid */
6454 return do_sched_setscheduler(pid
, policy
, param
);
6458 * sys_sched_setparam - set/change the RT priority of a thread
6459 * @pid: the pid in question.
6460 * @param: structure containing the new RT priority.
6462 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6464 return do_sched_setscheduler(pid
, -1, param
);
6468 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6469 * @pid: the pid in question.
6471 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6473 struct task_struct
*p
;
6480 read_lock(&tasklist_lock
);
6481 p
= find_process_by_pid(pid
);
6483 retval
= security_task_getscheduler(p
);
6486 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6488 read_unlock(&tasklist_lock
);
6493 * sys_sched_getparam - get the RT priority of a thread
6494 * @pid: the pid in question.
6495 * @param: structure containing the RT priority.
6497 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6499 struct sched_param lp
;
6500 struct task_struct
*p
;
6503 if (!param
|| pid
< 0)
6506 read_lock(&tasklist_lock
);
6507 p
= find_process_by_pid(pid
);
6512 retval
= security_task_getscheduler(p
);
6516 lp
.sched_priority
= p
->rt_priority
;
6517 read_unlock(&tasklist_lock
);
6520 * This one might sleep, we cannot do it with a spinlock held ...
6522 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6527 read_unlock(&tasklist_lock
);
6531 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6533 cpumask_var_t cpus_allowed
, new_mask
;
6534 struct task_struct
*p
;
6538 read_lock(&tasklist_lock
);
6540 p
= find_process_by_pid(pid
);
6542 read_unlock(&tasklist_lock
);
6548 * It is not safe to call set_cpus_allowed with the
6549 * tasklist_lock held. We will bump the task_struct's
6550 * usage count and then drop tasklist_lock.
6553 read_unlock(&tasklist_lock
);
6555 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6559 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6561 goto out_free_cpus_allowed
;
6564 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6567 retval
= security_task_setscheduler(p
, 0, NULL
);
6571 cpuset_cpus_allowed(p
, cpus_allowed
);
6572 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6574 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6577 cpuset_cpus_allowed(p
, cpus_allowed
);
6578 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6580 * We must have raced with a concurrent cpuset
6581 * update. Just reset the cpus_allowed to the
6582 * cpuset's cpus_allowed
6584 cpumask_copy(new_mask
, cpus_allowed
);
6589 free_cpumask_var(new_mask
);
6590 out_free_cpus_allowed
:
6591 free_cpumask_var(cpus_allowed
);
6598 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6599 struct cpumask
*new_mask
)
6601 if (len
< cpumask_size())
6602 cpumask_clear(new_mask
);
6603 else if (len
> cpumask_size())
6604 len
= cpumask_size();
6606 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6610 * sys_sched_setaffinity - set the cpu affinity of a process
6611 * @pid: pid of the process
6612 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6613 * @user_mask_ptr: user-space pointer to the new cpu mask
6615 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6616 unsigned long __user
*, user_mask_ptr
)
6618 cpumask_var_t new_mask
;
6621 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6624 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6626 retval
= sched_setaffinity(pid
, new_mask
);
6627 free_cpumask_var(new_mask
);
6631 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6633 struct task_struct
*p
;
6634 unsigned long flags
;
6639 read_lock(&tasklist_lock
);
6642 p
= find_process_by_pid(pid
);
6646 retval
= security_task_getscheduler(p
);
6650 rq
= task_rq_lock(p
, &flags
);
6651 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6652 task_rq_unlock(rq
, &flags
);
6655 read_unlock(&tasklist_lock
);
6662 * sys_sched_getaffinity - get the cpu affinity of a process
6663 * @pid: pid of the process
6664 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6665 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6667 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6668 unsigned long __user
*, user_mask_ptr
)
6673 if (len
< cpumask_size())
6676 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6679 ret
= sched_getaffinity(pid
, mask
);
6681 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6684 ret
= cpumask_size();
6686 free_cpumask_var(mask
);
6692 * sys_sched_yield - yield the current processor to other threads.
6694 * This function yields the current CPU to other tasks. If there are no
6695 * other threads running on this CPU then this function will return.
6697 SYSCALL_DEFINE0(sched_yield
)
6699 struct rq
*rq
= this_rq_lock();
6701 schedstat_inc(rq
, yld_count
);
6702 current
->sched_class
->yield_task(rq
);
6705 * Since we are going to call schedule() anyway, there's
6706 * no need to preempt or enable interrupts:
6708 __release(rq
->lock
);
6709 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6710 _raw_spin_unlock(&rq
->lock
);
6711 preempt_enable_no_resched();
6718 static inline int should_resched(void)
6720 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6723 static void __cond_resched(void)
6725 add_preempt_count(PREEMPT_ACTIVE
);
6727 sub_preempt_count(PREEMPT_ACTIVE
);
6730 int __sched
_cond_resched(void)
6732 if (should_resched()) {
6738 EXPORT_SYMBOL(_cond_resched
);
6741 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6742 * call schedule, and on return reacquire the lock.
6744 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6745 * operations here to prevent schedule() from being called twice (once via
6746 * spin_unlock(), once by hand).
6748 int __cond_resched_lock(spinlock_t
*lock
)
6750 int resched
= should_resched();
6753 lockdep_assert_held(lock
);
6755 if (spin_needbreak(lock
) || resched
) {
6766 EXPORT_SYMBOL(__cond_resched_lock
);
6768 int __sched
__cond_resched_softirq(void)
6770 BUG_ON(!in_softirq());
6772 if (should_resched()) {
6780 EXPORT_SYMBOL(__cond_resched_softirq
);
6783 * yield - yield the current processor to other threads.
6785 * This is a shortcut for kernel-space yielding - it marks the
6786 * thread runnable and calls sys_sched_yield().
6788 void __sched
yield(void)
6790 set_current_state(TASK_RUNNING
);
6793 EXPORT_SYMBOL(yield
);
6796 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6797 * that process accounting knows that this is a task in IO wait state.
6799 void __sched
io_schedule(void)
6801 struct rq
*rq
= raw_rq();
6803 delayacct_blkio_start();
6804 atomic_inc(&rq
->nr_iowait
);
6805 current
->in_iowait
= 1;
6807 current
->in_iowait
= 0;
6808 atomic_dec(&rq
->nr_iowait
);
6809 delayacct_blkio_end();
6811 EXPORT_SYMBOL(io_schedule
);
6813 long __sched
io_schedule_timeout(long timeout
)
6815 struct rq
*rq
= raw_rq();
6818 delayacct_blkio_start();
6819 atomic_inc(&rq
->nr_iowait
);
6820 current
->in_iowait
= 1;
6821 ret
= schedule_timeout(timeout
);
6822 current
->in_iowait
= 0;
6823 atomic_dec(&rq
->nr_iowait
);
6824 delayacct_blkio_end();
6829 * sys_sched_get_priority_max - return maximum RT priority.
6830 * @policy: scheduling class.
6832 * this syscall returns the maximum rt_priority that can be used
6833 * by a given scheduling class.
6835 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6842 ret
= MAX_USER_RT_PRIO
-1;
6854 * sys_sched_get_priority_min - return minimum RT priority.
6855 * @policy: scheduling class.
6857 * this syscall returns the minimum rt_priority that can be used
6858 * by a given scheduling class.
6860 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6878 * sys_sched_rr_get_interval - return the default timeslice of a process.
6879 * @pid: pid of the process.
6880 * @interval: userspace pointer to the timeslice value.
6882 * this syscall writes the default timeslice value of a given process
6883 * into the user-space timespec buffer. A value of '0' means infinity.
6885 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6886 struct timespec __user
*, interval
)
6888 struct task_struct
*p
;
6889 unsigned int time_slice
;
6890 unsigned long flags
;
6899 read_lock(&tasklist_lock
);
6900 p
= find_process_by_pid(pid
);
6904 retval
= security_task_getscheduler(p
);
6908 rq
= task_rq_lock(p
, &flags
);
6909 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
6910 task_rq_unlock(rq
, &flags
);
6912 read_unlock(&tasklist_lock
);
6913 jiffies_to_timespec(time_slice
, &t
);
6914 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6918 read_unlock(&tasklist_lock
);
6922 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6924 void sched_show_task(struct task_struct
*p
)
6926 unsigned long free
= 0;
6929 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6930 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6931 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6932 #if BITS_PER_LONG == 32
6933 if (state
== TASK_RUNNING
)
6934 printk(KERN_CONT
" running ");
6936 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6938 if (state
== TASK_RUNNING
)
6939 printk(KERN_CONT
" running task ");
6941 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6943 #ifdef CONFIG_DEBUG_STACK_USAGE
6944 free
= stack_not_used(p
);
6946 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6947 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6948 (unsigned long)task_thread_info(p
)->flags
);
6950 show_stack(p
, NULL
);
6953 void show_state_filter(unsigned long state_filter
)
6955 struct task_struct
*g
, *p
;
6957 #if BITS_PER_LONG == 32
6959 " task PC stack pid father\n");
6962 " task PC stack pid father\n");
6964 read_lock(&tasklist_lock
);
6965 do_each_thread(g
, p
) {
6967 * reset the NMI-timeout, listing all files on a slow
6968 * console might take alot of time:
6970 touch_nmi_watchdog();
6971 if (!state_filter
|| (p
->state
& state_filter
))
6973 } while_each_thread(g
, p
);
6975 touch_all_softlockup_watchdogs();
6977 #ifdef CONFIG_SCHED_DEBUG
6978 sysrq_sched_debug_show();
6980 read_unlock(&tasklist_lock
);
6982 * Only show locks if all tasks are dumped:
6985 debug_show_all_locks();
6988 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6990 idle
->sched_class
= &idle_sched_class
;
6994 * init_idle - set up an idle thread for a given CPU
6995 * @idle: task in question
6996 * @cpu: cpu the idle task belongs to
6998 * NOTE: this function does not set the idle thread's NEED_RESCHED
6999 * flag, to make booting more robust.
7001 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7003 struct rq
*rq
= cpu_rq(cpu
);
7004 unsigned long flags
;
7006 spin_lock_irqsave(&rq
->lock
, flags
);
7009 idle
->se
.exec_start
= sched_clock();
7011 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
7012 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7013 __set_task_cpu(idle
, cpu
);
7015 rq
->curr
= rq
->idle
= idle
;
7016 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7019 spin_unlock_irqrestore(&rq
->lock
, flags
);
7021 /* Set the preempt count _outside_ the spinlocks! */
7022 #if defined(CONFIG_PREEMPT)
7023 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7025 task_thread_info(idle
)->preempt_count
= 0;
7028 * The idle tasks have their own, simple scheduling class:
7030 idle
->sched_class
= &idle_sched_class
;
7031 ftrace_graph_init_task(idle
);
7035 * In a system that switches off the HZ timer nohz_cpu_mask
7036 * indicates which cpus entered this state. This is used
7037 * in the rcu update to wait only for active cpus. For system
7038 * which do not switch off the HZ timer nohz_cpu_mask should
7039 * always be CPU_BITS_NONE.
7041 cpumask_var_t nohz_cpu_mask
;
7044 * Increase the granularity value when there are more CPUs,
7045 * because with more CPUs the 'effective latency' as visible
7046 * to users decreases. But the relationship is not linear,
7047 * so pick a second-best guess by going with the log2 of the
7050 * This idea comes from the SD scheduler of Con Kolivas:
7052 static inline void sched_init_granularity(void)
7054 unsigned int factor
= 1 + ilog2(num_online_cpus());
7055 const unsigned long limit
= 200000000;
7057 sysctl_sched_min_granularity
*= factor
;
7058 if (sysctl_sched_min_granularity
> limit
)
7059 sysctl_sched_min_granularity
= limit
;
7061 sysctl_sched_latency
*= factor
;
7062 if (sysctl_sched_latency
> limit
)
7063 sysctl_sched_latency
= limit
;
7065 sysctl_sched_wakeup_granularity
*= factor
;
7067 sysctl_sched_shares_ratelimit
*= factor
;
7072 * This is how migration works:
7074 * 1) we queue a struct migration_req structure in the source CPU's
7075 * runqueue and wake up that CPU's migration thread.
7076 * 2) we down() the locked semaphore => thread blocks.
7077 * 3) migration thread wakes up (implicitly it forces the migrated
7078 * thread off the CPU)
7079 * 4) it gets the migration request and checks whether the migrated
7080 * task is still in the wrong runqueue.
7081 * 5) if it's in the wrong runqueue then the migration thread removes
7082 * it and puts it into the right queue.
7083 * 6) migration thread up()s the semaphore.
7084 * 7) we wake up and the migration is done.
7088 * Change a given task's CPU affinity. Migrate the thread to a
7089 * proper CPU and schedule it away if the CPU it's executing on
7090 * is removed from the allowed bitmask.
7092 * NOTE: the caller must have a valid reference to the task, the
7093 * task must not exit() & deallocate itself prematurely. The
7094 * call is not atomic; no spinlocks may be held.
7096 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7098 struct migration_req req
;
7099 unsigned long flags
;
7103 rq
= task_rq_lock(p
, &flags
);
7104 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
7109 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7110 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7115 if (p
->sched_class
->set_cpus_allowed
)
7116 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7118 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7119 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7122 /* Can the task run on the task's current CPU? If so, we're done */
7123 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7126 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
7127 /* Need help from migration thread: drop lock and wait. */
7128 struct task_struct
*mt
= rq
->migration_thread
;
7130 get_task_struct(mt
);
7131 task_rq_unlock(rq
, &flags
);
7132 wake_up_process(rq
->migration_thread
);
7133 put_task_struct(mt
);
7134 wait_for_completion(&req
.done
);
7135 tlb_migrate_finish(p
->mm
);
7139 task_rq_unlock(rq
, &flags
);
7143 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7146 * Move (not current) task off this cpu, onto dest cpu. We're doing
7147 * this because either it can't run here any more (set_cpus_allowed()
7148 * away from this CPU, or CPU going down), or because we're
7149 * attempting to rebalance this task on exec (sched_exec).
7151 * So we race with normal scheduler movements, but that's OK, as long
7152 * as the task is no longer on this CPU.
7154 * Returns non-zero if task was successfully migrated.
7156 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7158 struct rq
*rq_dest
, *rq_src
;
7161 if (unlikely(!cpu_active(dest_cpu
)))
7164 rq_src
= cpu_rq(src_cpu
);
7165 rq_dest
= cpu_rq(dest_cpu
);
7167 double_rq_lock(rq_src
, rq_dest
);
7168 /* Already moved. */
7169 if (task_cpu(p
) != src_cpu
)
7171 /* Affinity changed (again). */
7172 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7175 on_rq
= p
->se
.on_rq
;
7177 deactivate_task(rq_src
, p
, 0);
7179 set_task_cpu(p
, dest_cpu
);
7181 activate_task(rq_dest
, p
, 0);
7182 check_preempt_curr(rq_dest
, p
, 0);
7187 double_rq_unlock(rq_src
, rq_dest
);
7191 #define RCU_MIGRATION_IDLE 0
7192 #define RCU_MIGRATION_NEED_QS 1
7193 #define RCU_MIGRATION_GOT_QS 2
7194 #define RCU_MIGRATION_MUST_SYNC 3
7197 * migration_thread - this is a highprio system thread that performs
7198 * thread migration by bumping thread off CPU then 'pushing' onto
7201 static int migration_thread(void *data
)
7204 int cpu
= (long)data
;
7208 BUG_ON(rq
->migration_thread
!= current
);
7210 set_current_state(TASK_INTERRUPTIBLE
);
7211 while (!kthread_should_stop()) {
7212 struct migration_req
*req
;
7213 struct list_head
*head
;
7215 spin_lock_irq(&rq
->lock
);
7217 if (cpu_is_offline(cpu
)) {
7218 spin_unlock_irq(&rq
->lock
);
7222 if (rq
->active_balance
) {
7223 active_load_balance(rq
, cpu
);
7224 rq
->active_balance
= 0;
7227 head
= &rq
->migration_queue
;
7229 if (list_empty(head
)) {
7230 spin_unlock_irq(&rq
->lock
);
7232 set_current_state(TASK_INTERRUPTIBLE
);
7235 req
= list_entry(head
->next
, struct migration_req
, list
);
7236 list_del_init(head
->next
);
7238 if (req
->task
!= NULL
) {
7239 spin_unlock(&rq
->lock
);
7240 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7241 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7242 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7243 spin_unlock(&rq
->lock
);
7245 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7246 spin_unlock(&rq
->lock
);
7247 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7251 complete(&req
->done
);
7253 __set_current_state(TASK_RUNNING
);
7258 #ifdef CONFIG_HOTPLUG_CPU
7260 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7264 local_irq_disable();
7265 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7271 * Figure out where task on dead CPU should go, use force if necessary.
7273 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7276 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7279 /* Look for allowed, online CPU in same node. */
7280 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
7281 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7284 /* Any allowed, online CPU? */
7285 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
7286 if (dest_cpu
< nr_cpu_ids
)
7289 /* No more Mr. Nice Guy. */
7290 if (dest_cpu
>= nr_cpu_ids
) {
7291 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7292 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
7295 * Don't tell them about moving exiting tasks or
7296 * kernel threads (both mm NULL), since they never
7299 if (p
->mm
&& printk_ratelimit()) {
7300 printk(KERN_INFO
"process %d (%s) no "
7301 "longer affine to cpu%d\n",
7302 task_pid_nr(p
), p
->comm
, dead_cpu
);
7307 /* It can have affinity changed while we were choosing. */
7308 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7313 * While a dead CPU has no uninterruptible tasks queued at this point,
7314 * it might still have a nonzero ->nr_uninterruptible counter, because
7315 * for performance reasons the counter is not stricly tracking tasks to
7316 * their home CPUs. So we just add the counter to another CPU's counter,
7317 * to keep the global sum constant after CPU-down:
7319 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7321 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
7322 unsigned long flags
;
7324 local_irq_save(flags
);
7325 double_rq_lock(rq_src
, rq_dest
);
7326 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7327 rq_src
->nr_uninterruptible
= 0;
7328 double_rq_unlock(rq_src
, rq_dest
);
7329 local_irq_restore(flags
);
7332 /* Run through task list and migrate tasks from the dead cpu. */
7333 static void migrate_live_tasks(int src_cpu
)
7335 struct task_struct
*p
, *t
;
7337 read_lock(&tasklist_lock
);
7339 do_each_thread(t
, p
) {
7343 if (task_cpu(p
) == src_cpu
)
7344 move_task_off_dead_cpu(src_cpu
, p
);
7345 } while_each_thread(t
, p
);
7347 read_unlock(&tasklist_lock
);
7351 * Schedules idle task to be the next runnable task on current CPU.
7352 * It does so by boosting its priority to highest possible.
7353 * Used by CPU offline code.
7355 void sched_idle_next(void)
7357 int this_cpu
= smp_processor_id();
7358 struct rq
*rq
= cpu_rq(this_cpu
);
7359 struct task_struct
*p
= rq
->idle
;
7360 unsigned long flags
;
7362 /* cpu has to be offline */
7363 BUG_ON(cpu_online(this_cpu
));
7366 * Strictly not necessary since rest of the CPUs are stopped by now
7367 * and interrupts disabled on the current cpu.
7369 spin_lock_irqsave(&rq
->lock
, flags
);
7371 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7373 update_rq_clock(rq
);
7374 activate_task(rq
, p
, 0);
7376 spin_unlock_irqrestore(&rq
->lock
, flags
);
7380 * Ensures that the idle task is using init_mm right before its cpu goes
7383 void idle_task_exit(void)
7385 struct mm_struct
*mm
= current
->active_mm
;
7387 BUG_ON(cpu_online(smp_processor_id()));
7390 switch_mm(mm
, &init_mm
, current
);
7394 /* called under rq->lock with disabled interrupts */
7395 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7397 struct rq
*rq
= cpu_rq(dead_cpu
);
7399 /* Must be exiting, otherwise would be on tasklist. */
7400 BUG_ON(!p
->exit_state
);
7402 /* Cannot have done final schedule yet: would have vanished. */
7403 BUG_ON(p
->state
== TASK_DEAD
);
7408 * Drop lock around migration; if someone else moves it,
7409 * that's OK. No task can be added to this CPU, so iteration is
7412 spin_unlock_irq(&rq
->lock
);
7413 move_task_off_dead_cpu(dead_cpu
, p
);
7414 spin_lock_irq(&rq
->lock
);
7419 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7420 static void migrate_dead_tasks(unsigned int dead_cpu
)
7422 struct rq
*rq
= cpu_rq(dead_cpu
);
7423 struct task_struct
*next
;
7426 if (!rq
->nr_running
)
7428 update_rq_clock(rq
);
7429 next
= pick_next_task(rq
);
7432 next
->sched_class
->put_prev_task(rq
, next
);
7433 migrate_dead(dead_cpu
, next
);
7439 * remove the tasks which were accounted by rq from calc_load_tasks.
7441 static void calc_global_load_remove(struct rq
*rq
)
7443 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7444 rq
->calc_load_active
= 0;
7446 #endif /* CONFIG_HOTPLUG_CPU */
7448 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7450 static struct ctl_table sd_ctl_dir
[] = {
7452 .procname
= "sched_domain",
7458 static struct ctl_table sd_ctl_root
[] = {
7460 .ctl_name
= CTL_KERN
,
7461 .procname
= "kernel",
7463 .child
= sd_ctl_dir
,
7468 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7470 struct ctl_table
*entry
=
7471 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7476 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7478 struct ctl_table
*entry
;
7481 * In the intermediate directories, both the child directory and
7482 * procname are dynamically allocated and could fail but the mode
7483 * will always be set. In the lowest directory the names are
7484 * static strings and all have proc handlers.
7486 for (entry
= *tablep
; entry
->mode
; entry
++) {
7488 sd_free_ctl_entry(&entry
->child
);
7489 if (entry
->proc_handler
== NULL
)
7490 kfree(entry
->procname
);
7498 set_table_entry(struct ctl_table
*entry
,
7499 const char *procname
, void *data
, int maxlen
,
7500 mode_t mode
, proc_handler
*proc_handler
)
7502 entry
->procname
= procname
;
7504 entry
->maxlen
= maxlen
;
7506 entry
->proc_handler
= proc_handler
;
7509 static struct ctl_table
*
7510 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7512 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7517 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7518 sizeof(long), 0644, proc_doulongvec_minmax
);
7519 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7520 sizeof(long), 0644, proc_doulongvec_minmax
);
7521 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7522 sizeof(int), 0644, proc_dointvec_minmax
);
7523 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7524 sizeof(int), 0644, proc_dointvec_minmax
);
7525 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7526 sizeof(int), 0644, proc_dointvec_minmax
);
7527 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7528 sizeof(int), 0644, proc_dointvec_minmax
);
7529 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7530 sizeof(int), 0644, proc_dointvec_minmax
);
7531 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7532 sizeof(int), 0644, proc_dointvec_minmax
);
7533 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7534 sizeof(int), 0644, proc_dointvec_minmax
);
7535 set_table_entry(&table
[9], "cache_nice_tries",
7536 &sd
->cache_nice_tries
,
7537 sizeof(int), 0644, proc_dointvec_minmax
);
7538 set_table_entry(&table
[10], "flags", &sd
->flags
,
7539 sizeof(int), 0644, proc_dointvec_minmax
);
7540 set_table_entry(&table
[11], "name", sd
->name
,
7541 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7542 /* &table[12] is terminator */
7547 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7549 struct ctl_table
*entry
, *table
;
7550 struct sched_domain
*sd
;
7551 int domain_num
= 0, i
;
7554 for_each_domain(cpu
, sd
)
7556 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7561 for_each_domain(cpu
, sd
) {
7562 snprintf(buf
, 32, "domain%d", i
);
7563 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7565 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7572 static struct ctl_table_header
*sd_sysctl_header
;
7573 static void register_sched_domain_sysctl(void)
7575 int i
, cpu_num
= num_possible_cpus();
7576 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7579 WARN_ON(sd_ctl_dir
[0].child
);
7580 sd_ctl_dir
[0].child
= entry
;
7585 for_each_possible_cpu(i
) {
7586 snprintf(buf
, 32, "cpu%d", i
);
7587 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7589 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7593 WARN_ON(sd_sysctl_header
);
7594 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7597 /* may be called multiple times per register */
7598 static void unregister_sched_domain_sysctl(void)
7600 if (sd_sysctl_header
)
7601 unregister_sysctl_table(sd_sysctl_header
);
7602 sd_sysctl_header
= NULL
;
7603 if (sd_ctl_dir
[0].child
)
7604 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7607 static void register_sched_domain_sysctl(void)
7610 static void unregister_sched_domain_sysctl(void)
7615 static void set_rq_online(struct rq
*rq
)
7618 const struct sched_class
*class;
7620 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7623 for_each_class(class) {
7624 if (class->rq_online
)
7625 class->rq_online(rq
);
7630 static void set_rq_offline(struct rq
*rq
)
7633 const struct sched_class
*class;
7635 for_each_class(class) {
7636 if (class->rq_offline
)
7637 class->rq_offline(rq
);
7640 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7646 * migration_call - callback that gets triggered when a CPU is added.
7647 * Here we can start up the necessary migration thread for the new CPU.
7649 static int __cpuinit
7650 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7652 struct task_struct
*p
;
7653 int cpu
= (long)hcpu
;
7654 unsigned long flags
;
7659 case CPU_UP_PREPARE
:
7660 case CPU_UP_PREPARE_FROZEN
:
7661 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7664 kthread_bind(p
, cpu
);
7665 /* Must be high prio: stop_machine expects to yield to it. */
7666 rq
= task_rq_lock(p
, &flags
);
7667 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7668 task_rq_unlock(rq
, &flags
);
7670 cpu_rq(cpu
)->migration_thread
= p
;
7671 rq
->calc_load_update
= calc_load_update
;
7675 case CPU_ONLINE_FROZEN
:
7676 /* Strictly unnecessary, as first user will wake it. */
7677 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7679 /* Update our root-domain */
7681 spin_lock_irqsave(&rq
->lock
, flags
);
7683 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7687 spin_unlock_irqrestore(&rq
->lock
, flags
);
7690 #ifdef CONFIG_HOTPLUG_CPU
7691 case CPU_UP_CANCELED
:
7692 case CPU_UP_CANCELED_FROZEN
:
7693 if (!cpu_rq(cpu
)->migration_thread
)
7695 /* Unbind it from offline cpu so it can run. Fall thru. */
7696 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7697 cpumask_any(cpu_online_mask
));
7698 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7699 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7700 cpu_rq(cpu
)->migration_thread
= NULL
;
7704 case CPU_DEAD_FROZEN
:
7705 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7706 migrate_live_tasks(cpu
);
7708 kthread_stop(rq
->migration_thread
);
7709 put_task_struct(rq
->migration_thread
);
7710 rq
->migration_thread
= NULL
;
7711 /* Idle task back to normal (off runqueue, low prio) */
7712 spin_lock_irq(&rq
->lock
);
7713 update_rq_clock(rq
);
7714 deactivate_task(rq
, rq
->idle
, 0);
7715 rq
->idle
->static_prio
= MAX_PRIO
;
7716 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7717 rq
->idle
->sched_class
= &idle_sched_class
;
7718 migrate_dead_tasks(cpu
);
7719 spin_unlock_irq(&rq
->lock
);
7721 migrate_nr_uninterruptible(rq
);
7722 BUG_ON(rq
->nr_running
!= 0);
7723 calc_global_load_remove(rq
);
7725 * No need to migrate the tasks: it was best-effort if
7726 * they didn't take sched_hotcpu_mutex. Just wake up
7729 spin_lock_irq(&rq
->lock
);
7730 while (!list_empty(&rq
->migration_queue
)) {
7731 struct migration_req
*req
;
7733 req
= list_entry(rq
->migration_queue
.next
,
7734 struct migration_req
, list
);
7735 list_del_init(&req
->list
);
7736 spin_unlock_irq(&rq
->lock
);
7737 complete(&req
->done
);
7738 spin_lock_irq(&rq
->lock
);
7740 spin_unlock_irq(&rq
->lock
);
7744 case CPU_DYING_FROZEN
:
7745 /* Update our root-domain */
7747 spin_lock_irqsave(&rq
->lock
, flags
);
7749 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7752 spin_unlock_irqrestore(&rq
->lock
, flags
);
7760 * Register at high priority so that task migration (migrate_all_tasks)
7761 * happens before everything else. This has to be lower priority than
7762 * the notifier in the perf_event subsystem, though.
7764 static struct notifier_block __cpuinitdata migration_notifier
= {
7765 .notifier_call
= migration_call
,
7769 static int __init
migration_init(void)
7771 void *cpu
= (void *)(long)smp_processor_id();
7774 /* Start one for the boot CPU: */
7775 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7776 BUG_ON(err
== NOTIFY_BAD
);
7777 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7778 register_cpu_notifier(&migration_notifier
);
7782 early_initcall(migration_init
);
7787 #ifdef CONFIG_SCHED_DEBUG
7789 static __read_mostly
int sched_domain_debug_enabled
;
7791 static int __init
sched_domain_debug_setup(char *str
)
7793 sched_domain_debug_enabled
= 1;
7797 early_param("sched_debug", sched_domain_debug_setup
);
7799 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7800 struct cpumask
*groupmask
)
7802 struct sched_group
*group
= sd
->groups
;
7805 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7806 cpumask_clear(groupmask
);
7808 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7810 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7811 printk("does not load-balance\n");
7813 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7818 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7820 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7821 printk(KERN_ERR
"ERROR: domain->span does not contain "
7824 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7825 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7829 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7833 printk(KERN_ERR
"ERROR: group is NULL\n");
7837 if (!group
->cpu_power
) {
7838 printk(KERN_CONT
"\n");
7839 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7844 if (!cpumask_weight(sched_group_cpus(group
))) {
7845 printk(KERN_CONT
"\n");
7846 printk(KERN_ERR
"ERROR: empty group\n");
7850 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7851 printk(KERN_CONT
"\n");
7852 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7856 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7858 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7860 printk(KERN_CONT
" %s", str
);
7861 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7862 printk(KERN_CONT
" (cpu_power = %d)",
7866 group
= group
->next
;
7867 } while (group
!= sd
->groups
);
7868 printk(KERN_CONT
"\n");
7870 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7871 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7874 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7875 printk(KERN_ERR
"ERROR: parent span is not a superset "
7876 "of domain->span\n");
7880 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7882 cpumask_var_t groupmask
;
7885 if (!sched_domain_debug_enabled
)
7889 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7893 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7895 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7896 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7901 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7908 free_cpumask_var(groupmask
);
7910 #else /* !CONFIG_SCHED_DEBUG */
7911 # define sched_domain_debug(sd, cpu) do { } while (0)
7912 #endif /* CONFIG_SCHED_DEBUG */
7914 static int sd_degenerate(struct sched_domain
*sd
)
7916 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7919 /* Following flags need at least 2 groups */
7920 if (sd
->flags
& (SD_LOAD_BALANCE
|
7921 SD_BALANCE_NEWIDLE
|
7925 SD_SHARE_PKG_RESOURCES
)) {
7926 if (sd
->groups
!= sd
->groups
->next
)
7930 /* Following flags don't use groups */
7931 if (sd
->flags
& (SD_WAKE_AFFINE
))
7938 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7940 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7942 if (sd_degenerate(parent
))
7945 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7948 /* Flags needing groups don't count if only 1 group in parent */
7949 if (parent
->groups
== parent
->groups
->next
) {
7950 pflags
&= ~(SD_LOAD_BALANCE
|
7951 SD_BALANCE_NEWIDLE
|
7955 SD_SHARE_PKG_RESOURCES
);
7956 if (nr_node_ids
== 1)
7957 pflags
&= ~SD_SERIALIZE
;
7959 if (~cflags
& pflags
)
7965 static void free_rootdomain(struct root_domain
*rd
)
7967 synchronize_sched();
7969 cpupri_cleanup(&rd
->cpupri
);
7971 free_cpumask_var(rd
->rto_mask
);
7972 free_cpumask_var(rd
->online
);
7973 free_cpumask_var(rd
->span
);
7977 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7979 struct root_domain
*old_rd
= NULL
;
7980 unsigned long flags
;
7982 spin_lock_irqsave(&rq
->lock
, flags
);
7987 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7990 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7993 * If we dont want to free the old_rt yet then
7994 * set old_rd to NULL to skip the freeing later
7997 if (!atomic_dec_and_test(&old_rd
->refcount
))
8001 atomic_inc(&rd
->refcount
);
8004 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8005 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8008 spin_unlock_irqrestore(&rq
->lock
, flags
);
8011 free_rootdomain(old_rd
);
8014 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8016 gfp_t gfp
= GFP_KERNEL
;
8018 memset(rd
, 0, sizeof(*rd
));
8023 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8025 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8027 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8030 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8035 free_cpumask_var(rd
->rto_mask
);
8037 free_cpumask_var(rd
->online
);
8039 free_cpumask_var(rd
->span
);
8044 static void init_defrootdomain(void)
8046 init_rootdomain(&def_root_domain
, true);
8048 atomic_set(&def_root_domain
.refcount
, 1);
8051 static struct root_domain
*alloc_rootdomain(void)
8053 struct root_domain
*rd
;
8055 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8059 if (init_rootdomain(rd
, false) != 0) {
8068 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8069 * hold the hotplug lock.
8072 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8074 struct rq
*rq
= cpu_rq(cpu
);
8075 struct sched_domain
*tmp
;
8077 /* Remove the sched domains which do not contribute to scheduling. */
8078 for (tmp
= sd
; tmp
; ) {
8079 struct sched_domain
*parent
= tmp
->parent
;
8083 if (sd_parent_degenerate(tmp
, parent
)) {
8084 tmp
->parent
= parent
->parent
;
8086 parent
->parent
->child
= tmp
;
8091 if (sd
&& sd_degenerate(sd
)) {
8097 sched_domain_debug(sd
, cpu
);
8099 rq_attach_root(rq
, rd
);
8100 rcu_assign_pointer(rq
->sd
, sd
);
8103 /* cpus with isolated domains */
8104 static cpumask_var_t cpu_isolated_map
;
8106 /* Setup the mask of cpus configured for isolated domains */
8107 static int __init
isolated_cpu_setup(char *str
)
8109 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8110 cpulist_parse(str
, cpu_isolated_map
);
8114 __setup("isolcpus=", isolated_cpu_setup
);
8117 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8118 * to a function which identifies what group(along with sched group) a CPU
8119 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8120 * (due to the fact that we keep track of groups covered with a struct cpumask).
8122 * init_sched_build_groups will build a circular linked list of the groups
8123 * covered by the given span, and will set each group's ->cpumask correctly,
8124 * and ->cpu_power to 0.
8127 init_sched_build_groups(const struct cpumask
*span
,
8128 const struct cpumask
*cpu_map
,
8129 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8130 struct sched_group
**sg
,
8131 struct cpumask
*tmpmask
),
8132 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8134 struct sched_group
*first
= NULL
, *last
= NULL
;
8137 cpumask_clear(covered
);
8139 for_each_cpu(i
, span
) {
8140 struct sched_group
*sg
;
8141 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8144 if (cpumask_test_cpu(i
, covered
))
8147 cpumask_clear(sched_group_cpus(sg
));
8150 for_each_cpu(j
, span
) {
8151 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8154 cpumask_set_cpu(j
, covered
);
8155 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8166 #define SD_NODES_PER_DOMAIN 16
8171 * find_next_best_node - find the next node to include in a sched_domain
8172 * @node: node whose sched_domain we're building
8173 * @used_nodes: nodes already in the sched_domain
8175 * Find the next node to include in a given scheduling domain. Simply
8176 * finds the closest node not already in the @used_nodes map.
8178 * Should use nodemask_t.
8180 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8182 int i
, n
, val
, min_val
, best_node
= 0;
8186 for (i
= 0; i
< nr_node_ids
; i
++) {
8187 /* Start at @node */
8188 n
= (node
+ i
) % nr_node_ids
;
8190 if (!nr_cpus_node(n
))
8193 /* Skip already used nodes */
8194 if (node_isset(n
, *used_nodes
))
8197 /* Simple min distance search */
8198 val
= node_distance(node
, n
);
8200 if (val
< min_val
) {
8206 node_set(best_node
, *used_nodes
);
8211 * sched_domain_node_span - get a cpumask for a node's sched_domain
8212 * @node: node whose cpumask we're constructing
8213 * @span: resulting cpumask
8215 * Given a node, construct a good cpumask for its sched_domain to span. It
8216 * should be one that prevents unnecessary balancing, but also spreads tasks
8219 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8221 nodemask_t used_nodes
;
8224 cpumask_clear(span
);
8225 nodes_clear(used_nodes
);
8227 cpumask_or(span
, span
, cpumask_of_node(node
));
8228 node_set(node
, used_nodes
);
8230 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8231 int next_node
= find_next_best_node(node
, &used_nodes
);
8233 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8236 #endif /* CONFIG_NUMA */
8238 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8241 * The cpus mask in sched_group and sched_domain hangs off the end.
8243 * ( See the the comments in include/linux/sched.h:struct sched_group
8244 * and struct sched_domain. )
8246 struct static_sched_group
{
8247 struct sched_group sg
;
8248 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8251 struct static_sched_domain
{
8252 struct sched_domain sd
;
8253 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8259 cpumask_var_t domainspan
;
8260 cpumask_var_t covered
;
8261 cpumask_var_t notcovered
;
8263 cpumask_var_t nodemask
;
8264 cpumask_var_t this_sibling_map
;
8265 cpumask_var_t this_core_map
;
8266 cpumask_var_t send_covered
;
8267 cpumask_var_t tmpmask
;
8268 struct sched_group
**sched_group_nodes
;
8269 struct root_domain
*rd
;
8273 sa_sched_groups
= 0,
8278 sa_this_sibling_map
,
8280 sa_sched_group_nodes
,
8290 * SMT sched-domains:
8292 #ifdef CONFIG_SCHED_SMT
8293 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8294 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8297 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8298 struct sched_group
**sg
, struct cpumask
*unused
)
8301 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8304 #endif /* CONFIG_SCHED_SMT */
8307 * multi-core sched-domains:
8309 #ifdef CONFIG_SCHED_MC
8310 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8311 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8312 #endif /* CONFIG_SCHED_MC */
8314 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8316 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8317 struct sched_group
**sg
, struct cpumask
*mask
)
8321 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8322 group
= cpumask_first(mask
);
8324 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8327 #elif defined(CONFIG_SCHED_MC)
8329 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8330 struct sched_group
**sg
, struct cpumask
*unused
)
8333 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8338 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8339 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8342 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8343 struct sched_group
**sg
, struct cpumask
*mask
)
8346 #ifdef CONFIG_SCHED_MC
8347 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8348 group
= cpumask_first(mask
);
8349 #elif defined(CONFIG_SCHED_SMT)
8350 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8351 group
= cpumask_first(mask
);
8356 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8362 * The init_sched_build_groups can't handle what we want to do with node
8363 * groups, so roll our own. Now each node has its own list of groups which
8364 * gets dynamically allocated.
8366 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8367 static struct sched_group
***sched_group_nodes_bycpu
;
8369 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8370 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8372 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8373 struct sched_group
**sg
,
8374 struct cpumask
*nodemask
)
8378 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8379 group
= cpumask_first(nodemask
);
8382 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8386 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8388 struct sched_group
*sg
= group_head
;
8394 for_each_cpu(j
, sched_group_cpus(sg
)) {
8395 struct sched_domain
*sd
;
8397 sd
= &per_cpu(phys_domains
, j
).sd
;
8398 if (j
!= group_first_cpu(sd
->groups
)) {
8400 * Only add "power" once for each
8406 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8409 } while (sg
!= group_head
);
8412 static int build_numa_sched_groups(struct s_data
*d
,
8413 const struct cpumask
*cpu_map
, int num
)
8415 struct sched_domain
*sd
;
8416 struct sched_group
*sg
, *prev
;
8419 cpumask_clear(d
->covered
);
8420 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8421 if (cpumask_empty(d
->nodemask
)) {
8422 d
->sched_group_nodes
[num
] = NULL
;
8426 sched_domain_node_span(num
, d
->domainspan
);
8427 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8429 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8432 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8436 d
->sched_group_nodes
[num
] = sg
;
8438 for_each_cpu(j
, d
->nodemask
) {
8439 sd
= &per_cpu(node_domains
, j
).sd
;
8444 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8446 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8449 for (j
= 0; j
< nr_node_ids
; j
++) {
8450 n
= (num
+ j
) % nr_node_ids
;
8451 cpumask_complement(d
->notcovered
, d
->covered
);
8452 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8453 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8454 if (cpumask_empty(d
->tmpmask
))
8456 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8457 if (cpumask_empty(d
->tmpmask
))
8459 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8463 "Can not alloc domain group for node %d\n", j
);
8467 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8468 sg
->next
= prev
->next
;
8469 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8476 #endif /* CONFIG_NUMA */
8479 /* Free memory allocated for various sched_group structures */
8480 static void free_sched_groups(const struct cpumask
*cpu_map
,
8481 struct cpumask
*nodemask
)
8485 for_each_cpu(cpu
, cpu_map
) {
8486 struct sched_group
**sched_group_nodes
8487 = sched_group_nodes_bycpu
[cpu
];
8489 if (!sched_group_nodes
)
8492 for (i
= 0; i
< nr_node_ids
; i
++) {
8493 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8495 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8496 if (cpumask_empty(nodemask
))
8506 if (oldsg
!= sched_group_nodes
[i
])
8509 kfree(sched_group_nodes
);
8510 sched_group_nodes_bycpu
[cpu
] = NULL
;
8513 #else /* !CONFIG_NUMA */
8514 static void free_sched_groups(const struct cpumask
*cpu_map
,
8515 struct cpumask
*nodemask
)
8518 #endif /* CONFIG_NUMA */
8521 * Initialize sched groups cpu_power.
8523 * cpu_power indicates the capacity of sched group, which is used while
8524 * distributing the load between different sched groups in a sched domain.
8525 * Typically cpu_power for all the groups in a sched domain will be same unless
8526 * there are asymmetries in the topology. If there are asymmetries, group
8527 * having more cpu_power will pickup more load compared to the group having
8530 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8532 struct sched_domain
*child
;
8533 struct sched_group
*group
;
8537 WARN_ON(!sd
|| !sd
->groups
);
8539 if (cpu
!= group_first_cpu(sd
->groups
))
8544 sd
->groups
->cpu_power
= 0;
8547 power
= SCHED_LOAD_SCALE
;
8548 weight
= cpumask_weight(sched_domain_span(sd
));
8550 * SMT siblings share the power of a single core.
8551 * Usually multiple threads get a better yield out of
8552 * that one core than a single thread would have,
8553 * reflect that in sd->smt_gain.
8555 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8556 power
*= sd
->smt_gain
;
8558 power
>>= SCHED_LOAD_SHIFT
;
8560 sd
->groups
->cpu_power
+= power
;
8565 * Add cpu_power of each child group to this groups cpu_power.
8567 group
= child
->groups
;
8569 sd
->groups
->cpu_power
+= group
->cpu_power
;
8570 group
= group
->next
;
8571 } while (group
!= child
->groups
);
8575 * Initializers for schedule domains
8576 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8579 #ifdef CONFIG_SCHED_DEBUG
8580 # define SD_INIT_NAME(sd, type) sd->name = #type
8582 # define SD_INIT_NAME(sd, type) do { } while (0)
8585 #define SD_INIT(sd, type) sd_init_##type(sd)
8587 #define SD_INIT_FUNC(type) \
8588 static noinline void sd_init_##type(struct sched_domain *sd) \
8590 memset(sd, 0, sizeof(*sd)); \
8591 *sd = SD_##type##_INIT; \
8592 sd->level = SD_LV_##type; \
8593 SD_INIT_NAME(sd, type); \
8598 SD_INIT_FUNC(ALLNODES
)
8601 #ifdef CONFIG_SCHED_SMT
8602 SD_INIT_FUNC(SIBLING
)
8604 #ifdef CONFIG_SCHED_MC
8608 static int default_relax_domain_level
= -1;
8610 static int __init
setup_relax_domain_level(char *str
)
8614 val
= simple_strtoul(str
, NULL
, 0);
8615 if (val
< SD_LV_MAX
)
8616 default_relax_domain_level
= val
;
8620 __setup("relax_domain_level=", setup_relax_domain_level
);
8622 static void set_domain_attribute(struct sched_domain
*sd
,
8623 struct sched_domain_attr
*attr
)
8627 if (!attr
|| attr
->relax_domain_level
< 0) {
8628 if (default_relax_domain_level
< 0)
8631 request
= default_relax_domain_level
;
8633 request
= attr
->relax_domain_level
;
8634 if (request
< sd
->level
) {
8635 /* turn off idle balance on this domain */
8636 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8638 /* turn on idle balance on this domain */
8639 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8643 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8644 const struct cpumask
*cpu_map
)
8647 case sa_sched_groups
:
8648 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8649 d
->sched_group_nodes
= NULL
;
8651 free_rootdomain(d
->rd
); /* fall through */
8653 free_cpumask_var(d
->tmpmask
); /* fall through */
8654 case sa_send_covered
:
8655 free_cpumask_var(d
->send_covered
); /* fall through */
8656 case sa_this_core_map
:
8657 free_cpumask_var(d
->this_core_map
); /* fall through */
8658 case sa_this_sibling_map
:
8659 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8661 free_cpumask_var(d
->nodemask
); /* fall through */
8662 case sa_sched_group_nodes
:
8664 kfree(d
->sched_group_nodes
); /* fall through */
8666 free_cpumask_var(d
->notcovered
); /* fall through */
8668 free_cpumask_var(d
->covered
); /* fall through */
8670 free_cpumask_var(d
->domainspan
); /* fall through */
8677 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8678 const struct cpumask
*cpu_map
)
8681 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8683 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8684 return sa_domainspan
;
8685 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8687 /* Allocate the per-node list of sched groups */
8688 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8689 sizeof(struct sched_group
*), GFP_KERNEL
);
8690 if (!d
->sched_group_nodes
) {
8691 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8692 return sa_notcovered
;
8694 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8696 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8697 return sa_sched_group_nodes
;
8698 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8700 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8701 return sa_this_sibling_map
;
8702 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8703 return sa_this_core_map
;
8704 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8705 return sa_send_covered
;
8706 d
->rd
= alloc_rootdomain();
8708 printk(KERN_WARNING
"Cannot alloc root domain\n");
8711 return sa_rootdomain
;
8714 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8715 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8717 struct sched_domain
*sd
= NULL
;
8719 struct sched_domain
*parent
;
8722 if (cpumask_weight(cpu_map
) >
8723 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8724 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8725 SD_INIT(sd
, ALLNODES
);
8726 set_domain_attribute(sd
, attr
);
8727 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8728 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8733 sd
= &per_cpu(node_domains
, i
).sd
;
8735 set_domain_attribute(sd
, attr
);
8736 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8737 sd
->parent
= parent
;
8740 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8745 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8746 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8747 struct sched_domain
*parent
, int i
)
8749 struct sched_domain
*sd
;
8750 sd
= &per_cpu(phys_domains
, i
).sd
;
8752 set_domain_attribute(sd
, attr
);
8753 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8754 sd
->parent
= parent
;
8757 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8761 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8762 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8763 struct sched_domain
*parent
, int i
)
8765 struct sched_domain
*sd
= parent
;
8766 #ifdef CONFIG_SCHED_MC
8767 sd
= &per_cpu(core_domains
, i
).sd
;
8769 set_domain_attribute(sd
, attr
);
8770 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8771 sd
->parent
= parent
;
8773 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8778 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8779 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8780 struct sched_domain
*parent
, int i
)
8782 struct sched_domain
*sd
= parent
;
8783 #ifdef CONFIG_SCHED_SMT
8784 sd
= &per_cpu(cpu_domains
, i
).sd
;
8785 SD_INIT(sd
, SIBLING
);
8786 set_domain_attribute(sd
, attr
);
8787 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8788 sd
->parent
= parent
;
8790 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8795 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8796 const struct cpumask
*cpu_map
, int cpu
)
8799 #ifdef CONFIG_SCHED_SMT
8800 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8801 cpumask_and(d
->this_sibling_map
, cpu_map
,
8802 topology_thread_cpumask(cpu
));
8803 if (cpu
== cpumask_first(d
->this_sibling_map
))
8804 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8806 d
->send_covered
, d
->tmpmask
);
8809 #ifdef CONFIG_SCHED_MC
8810 case SD_LV_MC
: /* set up multi-core groups */
8811 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8812 if (cpu
== cpumask_first(d
->this_core_map
))
8813 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8815 d
->send_covered
, d
->tmpmask
);
8818 case SD_LV_CPU
: /* set up physical groups */
8819 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8820 if (!cpumask_empty(d
->nodemask
))
8821 init_sched_build_groups(d
->nodemask
, cpu_map
,
8823 d
->send_covered
, d
->tmpmask
);
8826 case SD_LV_ALLNODES
:
8827 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8828 d
->send_covered
, d
->tmpmask
);
8837 * Build sched domains for a given set of cpus and attach the sched domains
8838 * to the individual cpus
8840 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8841 struct sched_domain_attr
*attr
)
8843 enum s_alloc alloc_state
= sa_none
;
8845 struct sched_domain
*sd
;
8851 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8852 if (alloc_state
!= sa_rootdomain
)
8854 alloc_state
= sa_sched_groups
;
8857 * Set up domains for cpus specified by the cpu_map.
8859 for_each_cpu(i
, cpu_map
) {
8860 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8863 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8864 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8865 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8866 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8869 for_each_cpu(i
, cpu_map
) {
8870 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8871 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8874 /* Set up physical groups */
8875 for (i
= 0; i
< nr_node_ids
; i
++)
8876 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8879 /* Set up node groups */
8881 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8883 for (i
= 0; i
< nr_node_ids
; i
++)
8884 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8888 /* Calculate CPU power for physical packages and nodes */
8889 #ifdef CONFIG_SCHED_SMT
8890 for_each_cpu(i
, cpu_map
) {
8891 sd
= &per_cpu(cpu_domains
, i
).sd
;
8892 init_sched_groups_power(i
, sd
);
8895 #ifdef CONFIG_SCHED_MC
8896 for_each_cpu(i
, cpu_map
) {
8897 sd
= &per_cpu(core_domains
, i
).sd
;
8898 init_sched_groups_power(i
, sd
);
8902 for_each_cpu(i
, cpu_map
) {
8903 sd
= &per_cpu(phys_domains
, i
).sd
;
8904 init_sched_groups_power(i
, sd
);
8908 for (i
= 0; i
< nr_node_ids
; i
++)
8909 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8911 if (d
.sd_allnodes
) {
8912 struct sched_group
*sg
;
8914 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8916 init_numa_sched_groups_power(sg
);
8920 /* Attach the domains */
8921 for_each_cpu(i
, cpu_map
) {
8922 #ifdef CONFIG_SCHED_SMT
8923 sd
= &per_cpu(cpu_domains
, i
).sd
;
8924 #elif defined(CONFIG_SCHED_MC)
8925 sd
= &per_cpu(core_domains
, i
).sd
;
8927 sd
= &per_cpu(phys_domains
, i
).sd
;
8929 cpu_attach_domain(sd
, d
.rd
, i
);
8932 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8933 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8937 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8941 static int build_sched_domains(const struct cpumask
*cpu_map
)
8943 return __build_sched_domains(cpu_map
, NULL
);
8946 static cpumask_var_t
*doms_cur
; /* current sched domains */
8947 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8948 static struct sched_domain_attr
*dattr_cur
;
8949 /* attribues of custom domains in 'doms_cur' */
8952 * Special case: If a kmalloc of a doms_cur partition (array of
8953 * cpumask) fails, then fallback to a single sched domain,
8954 * as determined by the single cpumask fallback_doms.
8956 static cpumask_var_t fallback_doms
;
8959 * arch_update_cpu_topology lets virtualized architectures update the
8960 * cpu core maps. It is supposed to return 1 if the topology changed
8961 * or 0 if it stayed the same.
8963 int __attribute__((weak
)) arch_update_cpu_topology(void)
8968 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
8971 cpumask_var_t
*doms
;
8973 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
8976 for (i
= 0; i
< ndoms
; i
++) {
8977 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
8978 free_sched_domains(doms
, i
);
8985 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
8988 for (i
= 0; i
< ndoms
; i
++)
8989 free_cpumask_var(doms
[i
]);
8994 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8995 * For now this just excludes isolated cpus, but could be used to
8996 * exclude other special cases in the future.
8998 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9002 arch_update_cpu_topology();
9004 doms_cur
= alloc_sched_domains(ndoms_cur
);
9006 doms_cur
= &fallback_doms
;
9007 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
9009 err
= build_sched_domains(doms_cur
[0]);
9010 register_sched_domain_sysctl();
9015 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9016 struct cpumask
*tmpmask
)
9018 free_sched_groups(cpu_map
, tmpmask
);
9022 * Detach sched domains from a group of cpus specified in cpu_map
9023 * These cpus will now be attached to the NULL domain
9025 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9027 /* Save because hotplug lock held. */
9028 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9031 for_each_cpu(i
, cpu_map
)
9032 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9033 synchronize_sched();
9034 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9037 /* handle null as "default" */
9038 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9039 struct sched_domain_attr
*new, int idx_new
)
9041 struct sched_domain_attr tmp
;
9048 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9049 new ? (new + idx_new
) : &tmp
,
9050 sizeof(struct sched_domain_attr
));
9054 * Partition sched domains as specified by the 'ndoms_new'
9055 * cpumasks in the array doms_new[] of cpumasks. This compares
9056 * doms_new[] to the current sched domain partitioning, doms_cur[].
9057 * It destroys each deleted domain and builds each new domain.
9059 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9060 * The masks don't intersect (don't overlap.) We should setup one
9061 * sched domain for each mask. CPUs not in any of the cpumasks will
9062 * not be load balanced. If the same cpumask appears both in the
9063 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9066 * The passed in 'doms_new' should be allocated using
9067 * alloc_sched_domains. This routine takes ownership of it and will
9068 * free_sched_domains it when done with it. If the caller failed the
9069 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9070 * and partition_sched_domains() will fallback to the single partition
9071 * 'fallback_doms', it also forces the domains to be rebuilt.
9073 * If doms_new == NULL it will be replaced with cpu_online_mask.
9074 * ndoms_new == 0 is a special case for destroying existing domains,
9075 * and it will not create the default domain.
9077 * Call with hotplug lock held
9079 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
9080 struct sched_domain_attr
*dattr_new
)
9085 mutex_lock(&sched_domains_mutex
);
9087 /* always unregister in case we don't destroy any domains */
9088 unregister_sched_domain_sysctl();
9090 /* Let architecture update cpu core mappings. */
9091 new_topology
= arch_update_cpu_topology();
9093 n
= doms_new
? ndoms_new
: 0;
9095 /* Destroy deleted domains */
9096 for (i
= 0; i
< ndoms_cur
; i
++) {
9097 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9098 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
9099 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9102 /* no match - a current sched domain not in new doms_new[] */
9103 detach_destroy_domains(doms_cur
[i
]);
9108 if (doms_new
== NULL
) {
9110 doms_new
= &fallback_doms
;
9111 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
9112 WARN_ON_ONCE(dattr_new
);
9115 /* Build new domains */
9116 for (i
= 0; i
< ndoms_new
; i
++) {
9117 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9118 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
9119 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9122 /* no match - add a new doms_new */
9123 __build_sched_domains(doms_new
[i
],
9124 dattr_new
? dattr_new
+ i
: NULL
);
9129 /* Remember the new sched domains */
9130 if (doms_cur
!= &fallback_doms
)
9131 free_sched_domains(doms_cur
, ndoms_cur
);
9132 kfree(dattr_cur
); /* kfree(NULL) is safe */
9133 doms_cur
= doms_new
;
9134 dattr_cur
= dattr_new
;
9135 ndoms_cur
= ndoms_new
;
9137 register_sched_domain_sysctl();
9139 mutex_unlock(&sched_domains_mutex
);
9142 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9143 static void arch_reinit_sched_domains(void)
9147 /* Destroy domains first to force the rebuild */
9148 partition_sched_domains(0, NULL
, NULL
);
9150 rebuild_sched_domains();
9154 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9156 unsigned int level
= 0;
9158 if (sscanf(buf
, "%u", &level
) != 1)
9162 * level is always be positive so don't check for
9163 * level < POWERSAVINGS_BALANCE_NONE which is 0
9164 * What happens on 0 or 1 byte write,
9165 * need to check for count as well?
9168 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9172 sched_smt_power_savings
= level
;
9174 sched_mc_power_savings
= level
;
9176 arch_reinit_sched_domains();
9181 #ifdef CONFIG_SCHED_MC
9182 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9185 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9187 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9188 const char *buf
, size_t count
)
9190 return sched_power_savings_store(buf
, count
, 0);
9192 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9193 sched_mc_power_savings_show
,
9194 sched_mc_power_savings_store
);
9197 #ifdef CONFIG_SCHED_SMT
9198 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9201 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9203 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9204 const char *buf
, size_t count
)
9206 return sched_power_savings_store(buf
, count
, 1);
9208 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9209 sched_smt_power_savings_show
,
9210 sched_smt_power_savings_store
);
9213 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9217 #ifdef CONFIG_SCHED_SMT
9219 err
= sysfs_create_file(&cls
->kset
.kobj
,
9220 &attr_sched_smt_power_savings
.attr
);
9222 #ifdef CONFIG_SCHED_MC
9223 if (!err
&& mc_capable())
9224 err
= sysfs_create_file(&cls
->kset
.kobj
,
9225 &attr_sched_mc_power_savings
.attr
);
9229 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9231 #ifndef CONFIG_CPUSETS
9233 * Add online and remove offline CPUs from the scheduler domains.
9234 * When cpusets are enabled they take over this function.
9236 static int update_sched_domains(struct notifier_block
*nfb
,
9237 unsigned long action
, void *hcpu
)
9241 case CPU_ONLINE_FROZEN
:
9242 case CPU_DOWN_PREPARE
:
9243 case CPU_DOWN_PREPARE_FROZEN
:
9244 case CPU_DOWN_FAILED
:
9245 case CPU_DOWN_FAILED_FROZEN
:
9246 partition_sched_domains(1, NULL
, NULL
);
9255 static int update_runtime(struct notifier_block
*nfb
,
9256 unsigned long action
, void *hcpu
)
9258 int cpu
= (int)(long)hcpu
;
9261 case CPU_DOWN_PREPARE
:
9262 case CPU_DOWN_PREPARE_FROZEN
:
9263 disable_runtime(cpu_rq(cpu
));
9266 case CPU_DOWN_FAILED
:
9267 case CPU_DOWN_FAILED_FROZEN
:
9269 case CPU_ONLINE_FROZEN
:
9270 enable_runtime(cpu_rq(cpu
));
9278 void __init
sched_init_smp(void)
9280 cpumask_var_t non_isolated_cpus
;
9282 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9283 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9285 #if defined(CONFIG_NUMA)
9286 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9288 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9291 mutex_lock(&sched_domains_mutex
);
9292 arch_init_sched_domains(cpu_active_mask
);
9293 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9294 if (cpumask_empty(non_isolated_cpus
))
9295 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9296 mutex_unlock(&sched_domains_mutex
);
9299 #ifndef CONFIG_CPUSETS
9300 /* XXX: Theoretical race here - CPU may be hotplugged now */
9301 hotcpu_notifier(update_sched_domains
, 0);
9304 /* RT runtime code needs to handle some hotplug events */
9305 hotcpu_notifier(update_runtime
, 0);
9309 /* Move init over to a non-isolated CPU */
9310 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9312 sched_init_granularity();
9313 free_cpumask_var(non_isolated_cpus
);
9315 init_sched_rt_class();
9318 void __init
sched_init_smp(void)
9320 sched_init_granularity();
9322 #endif /* CONFIG_SMP */
9324 const_debug
unsigned int sysctl_timer_migration
= 1;
9326 int in_sched_functions(unsigned long addr
)
9328 return in_lock_functions(addr
) ||
9329 (addr
>= (unsigned long)__sched_text_start
9330 && addr
< (unsigned long)__sched_text_end
);
9333 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9335 cfs_rq
->tasks_timeline
= RB_ROOT
;
9336 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9337 #ifdef CONFIG_FAIR_GROUP_SCHED
9340 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9343 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9345 struct rt_prio_array
*array
;
9348 array
= &rt_rq
->active
;
9349 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9350 INIT_LIST_HEAD(array
->queue
+ i
);
9351 __clear_bit(i
, array
->bitmap
);
9353 /* delimiter for bitsearch: */
9354 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9356 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9357 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9359 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9363 rt_rq
->rt_nr_migratory
= 0;
9364 rt_rq
->overloaded
= 0;
9365 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9369 rt_rq
->rt_throttled
= 0;
9370 rt_rq
->rt_runtime
= 0;
9371 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9373 #ifdef CONFIG_RT_GROUP_SCHED
9374 rt_rq
->rt_nr_boosted
= 0;
9379 #ifdef CONFIG_FAIR_GROUP_SCHED
9380 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9381 struct sched_entity
*se
, int cpu
, int add
,
9382 struct sched_entity
*parent
)
9384 struct rq
*rq
= cpu_rq(cpu
);
9385 tg
->cfs_rq
[cpu
] = cfs_rq
;
9386 init_cfs_rq(cfs_rq
, rq
);
9389 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9392 /* se could be NULL for init_task_group */
9397 se
->cfs_rq
= &rq
->cfs
;
9399 se
->cfs_rq
= parent
->my_q
;
9402 se
->load
.weight
= tg
->shares
;
9403 se
->load
.inv_weight
= 0;
9404 se
->parent
= parent
;
9408 #ifdef CONFIG_RT_GROUP_SCHED
9409 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9410 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9411 struct sched_rt_entity
*parent
)
9413 struct rq
*rq
= cpu_rq(cpu
);
9415 tg
->rt_rq
[cpu
] = rt_rq
;
9416 init_rt_rq(rt_rq
, rq
);
9418 rt_rq
->rt_se
= rt_se
;
9419 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9421 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9423 tg
->rt_se
[cpu
] = rt_se
;
9428 rt_se
->rt_rq
= &rq
->rt
;
9430 rt_se
->rt_rq
= parent
->my_q
;
9432 rt_se
->my_q
= rt_rq
;
9433 rt_se
->parent
= parent
;
9434 INIT_LIST_HEAD(&rt_se
->run_list
);
9438 void __init
sched_init(void)
9441 unsigned long alloc_size
= 0, ptr
;
9443 #ifdef CONFIG_FAIR_GROUP_SCHED
9444 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9446 #ifdef CONFIG_RT_GROUP_SCHED
9447 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9449 #ifdef CONFIG_USER_SCHED
9452 #ifdef CONFIG_CPUMASK_OFFSTACK
9453 alloc_size
+= num_possible_cpus() * cpumask_size();
9456 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9458 #ifdef CONFIG_FAIR_GROUP_SCHED
9459 init_task_group
.se
= (struct sched_entity
**)ptr
;
9460 ptr
+= nr_cpu_ids
* sizeof(void **);
9462 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9463 ptr
+= nr_cpu_ids
* sizeof(void **);
9465 #ifdef CONFIG_USER_SCHED
9466 root_task_group
.se
= (struct sched_entity
**)ptr
;
9467 ptr
+= nr_cpu_ids
* sizeof(void **);
9469 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9470 ptr
+= nr_cpu_ids
* sizeof(void **);
9471 #endif /* CONFIG_USER_SCHED */
9472 #endif /* CONFIG_FAIR_GROUP_SCHED */
9473 #ifdef CONFIG_RT_GROUP_SCHED
9474 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9475 ptr
+= nr_cpu_ids
* sizeof(void **);
9477 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9478 ptr
+= nr_cpu_ids
* sizeof(void **);
9480 #ifdef CONFIG_USER_SCHED
9481 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9482 ptr
+= nr_cpu_ids
* sizeof(void **);
9484 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9485 ptr
+= nr_cpu_ids
* sizeof(void **);
9486 #endif /* CONFIG_USER_SCHED */
9487 #endif /* CONFIG_RT_GROUP_SCHED */
9488 #ifdef CONFIG_CPUMASK_OFFSTACK
9489 for_each_possible_cpu(i
) {
9490 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9491 ptr
+= cpumask_size();
9493 #endif /* CONFIG_CPUMASK_OFFSTACK */
9497 init_defrootdomain();
9500 init_rt_bandwidth(&def_rt_bandwidth
,
9501 global_rt_period(), global_rt_runtime());
9503 #ifdef CONFIG_RT_GROUP_SCHED
9504 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9505 global_rt_period(), global_rt_runtime());
9506 #ifdef CONFIG_USER_SCHED
9507 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9508 global_rt_period(), RUNTIME_INF
);
9509 #endif /* CONFIG_USER_SCHED */
9510 #endif /* CONFIG_RT_GROUP_SCHED */
9512 #ifdef CONFIG_GROUP_SCHED
9513 list_add(&init_task_group
.list
, &task_groups
);
9514 INIT_LIST_HEAD(&init_task_group
.children
);
9516 #ifdef CONFIG_USER_SCHED
9517 INIT_LIST_HEAD(&root_task_group
.children
);
9518 init_task_group
.parent
= &root_task_group
;
9519 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9520 #endif /* CONFIG_USER_SCHED */
9521 #endif /* CONFIG_GROUP_SCHED */
9523 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9524 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9525 __alignof__(unsigned long));
9527 for_each_possible_cpu(i
) {
9531 spin_lock_init(&rq
->lock
);
9533 rq
->calc_load_active
= 0;
9534 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9535 init_cfs_rq(&rq
->cfs
, rq
);
9536 init_rt_rq(&rq
->rt
, rq
);
9537 #ifdef CONFIG_FAIR_GROUP_SCHED
9538 init_task_group
.shares
= init_task_group_load
;
9539 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9540 #ifdef CONFIG_CGROUP_SCHED
9542 * How much cpu bandwidth does init_task_group get?
9544 * In case of task-groups formed thr' the cgroup filesystem, it
9545 * gets 100% of the cpu resources in the system. This overall
9546 * system cpu resource is divided among the tasks of
9547 * init_task_group and its child task-groups in a fair manner,
9548 * based on each entity's (task or task-group's) weight
9549 * (se->load.weight).
9551 * In other words, if init_task_group has 10 tasks of weight
9552 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9553 * then A0's share of the cpu resource is:
9555 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9557 * We achieve this by letting init_task_group's tasks sit
9558 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9560 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9561 #elif defined CONFIG_USER_SCHED
9562 root_task_group
.shares
= NICE_0_LOAD
;
9563 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9565 * In case of task-groups formed thr' the user id of tasks,
9566 * init_task_group represents tasks belonging to root user.
9567 * Hence it forms a sibling of all subsequent groups formed.
9568 * In this case, init_task_group gets only a fraction of overall
9569 * system cpu resource, based on the weight assigned to root
9570 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9571 * by letting tasks of init_task_group sit in a separate cfs_rq
9572 * (init_tg_cfs_rq) and having one entity represent this group of
9573 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9575 init_tg_cfs_entry(&init_task_group
,
9576 &per_cpu(init_tg_cfs_rq
, i
),
9577 &per_cpu(init_sched_entity
, i
), i
, 1,
9578 root_task_group
.se
[i
]);
9581 #endif /* CONFIG_FAIR_GROUP_SCHED */
9583 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9584 #ifdef CONFIG_RT_GROUP_SCHED
9585 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9586 #ifdef CONFIG_CGROUP_SCHED
9587 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9588 #elif defined CONFIG_USER_SCHED
9589 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9590 init_tg_rt_entry(&init_task_group
,
9591 &per_cpu(init_rt_rq
, i
),
9592 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9593 root_task_group
.rt_se
[i
]);
9597 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9598 rq
->cpu_load
[j
] = 0;
9602 rq
->post_schedule
= 0;
9603 rq
->active_balance
= 0;
9604 rq
->next_balance
= jiffies
;
9608 rq
->migration_thread
= NULL
;
9610 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9611 INIT_LIST_HEAD(&rq
->migration_queue
);
9612 rq_attach_root(rq
, &def_root_domain
);
9615 atomic_set(&rq
->nr_iowait
, 0);
9618 set_load_weight(&init_task
);
9620 #ifdef CONFIG_PREEMPT_NOTIFIERS
9621 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9625 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9628 #ifdef CONFIG_RT_MUTEXES
9629 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9633 * The boot idle thread does lazy MMU switching as well:
9635 atomic_inc(&init_mm
.mm_count
);
9636 enter_lazy_tlb(&init_mm
, current
);
9639 * Make us the idle thread. Technically, schedule() should not be
9640 * called from this thread, however somewhere below it might be,
9641 * but because we are the idle thread, we just pick up running again
9642 * when this runqueue becomes "idle".
9644 init_idle(current
, smp_processor_id());
9646 calc_load_update
= jiffies
+ LOAD_FREQ
;
9649 * During early bootup we pretend to be a normal task:
9651 current
->sched_class
= &fair_sched_class
;
9653 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9654 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9657 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9658 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9660 /* May be allocated at isolcpus cmdline parse time */
9661 if (cpu_isolated_map
== NULL
)
9662 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9667 scheduler_running
= 1;
9670 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9671 static inline int preempt_count_equals(int preempt_offset
)
9673 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9675 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9678 void __might_sleep(char *file
, int line
, int preempt_offset
)
9681 static unsigned long prev_jiffy
; /* ratelimiting */
9683 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9684 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9686 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9688 prev_jiffy
= jiffies
;
9691 "BUG: sleeping function called from invalid context at %s:%d\n",
9694 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9695 in_atomic(), irqs_disabled(),
9696 current
->pid
, current
->comm
);
9698 debug_show_held_locks(current
);
9699 if (irqs_disabled())
9700 print_irqtrace_events(current
);
9704 EXPORT_SYMBOL(__might_sleep
);
9707 #ifdef CONFIG_MAGIC_SYSRQ
9708 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9712 update_rq_clock(rq
);
9713 on_rq
= p
->se
.on_rq
;
9715 deactivate_task(rq
, p
, 0);
9716 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9718 activate_task(rq
, p
, 0);
9719 resched_task(rq
->curr
);
9723 void normalize_rt_tasks(void)
9725 struct task_struct
*g
, *p
;
9726 unsigned long flags
;
9729 read_lock_irqsave(&tasklist_lock
, flags
);
9730 do_each_thread(g
, p
) {
9732 * Only normalize user tasks:
9737 p
->se
.exec_start
= 0;
9738 #ifdef CONFIG_SCHEDSTATS
9739 p
->se
.wait_start
= 0;
9740 p
->se
.sleep_start
= 0;
9741 p
->se
.block_start
= 0;
9746 * Renice negative nice level userspace
9749 if (TASK_NICE(p
) < 0 && p
->mm
)
9750 set_user_nice(p
, 0);
9754 spin_lock(&p
->pi_lock
);
9755 rq
= __task_rq_lock(p
);
9757 normalize_task(rq
, p
);
9759 __task_rq_unlock(rq
);
9760 spin_unlock(&p
->pi_lock
);
9761 } while_each_thread(g
, p
);
9763 read_unlock_irqrestore(&tasklist_lock
, flags
);
9766 #endif /* CONFIG_MAGIC_SYSRQ */
9770 * These functions are only useful for the IA64 MCA handling.
9772 * They can only be called when the whole system has been
9773 * stopped - every CPU needs to be quiescent, and no scheduling
9774 * activity can take place. Using them for anything else would
9775 * be a serious bug, and as a result, they aren't even visible
9776 * under any other configuration.
9780 * curr_task - return the current task for a given cpu.
9781 * @cpu: the processor in question.
9783 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9785 struct task_struct
*curr_task(int cpu
)
9787 return cpu_curr(cpu
);
9791 * set_curr_task - set the current task for a given cpu.
9792 * @cpu: the processor in question.
9793 * @p: the task pointer to set.
9795 * Description: This function must only be used when non-maskable interrupts
9796 * are serviced on a separate stack. It allows the architecture to switch the
9797 * notion of the current task on a cpu in a non-blocking manner. This function
9798 * must be called with all CPU's synchronized, and interrupts disabled, the
9799 * and caller must save the original value of the current task (see
9800 * curr_task() above) and restore that value before reenabling interrupts and
9801 * re-starting the system.
9803 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9805 void set_curr_task(int cpu
, struct task_struct
*p
)
9812 #ifdef CONFIG_FAIR_GROUP_SCHED
9813 static void free_fair_sched_group(struct task_group
*tg
)
9817 for_each_possible_cpu(i
) {
9819 kfree(tg
->cfs_rq
[i
]);
9829 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9831 struct cfs_rq
*cfs_rq
;
9832 struct sched_entity
*se
;
9836 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9839 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9843 tg
->shares
= NICE_0_LOAD
;
9845 for_each_possible_cpu(i
) {
9848 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9849 GFP_KERNEL
, cpu_to_node(i
));
9853 se
= kzalloc_node(sizeof(struct sched_entity
),
9854 GFP_KERNEL
, cpu_to_node(i
));
9858 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9867 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9869 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9870 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9873 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9875 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9877 #else /* !CONFG_FAIR_GROUP_SCHED */
9878 static inline void free_fair_sched_group(struct task_group
*tg
)
9883 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9888 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9892 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9895 #endif /* CONFIG_FAIR_GROUP_SCHED */
9897 #ifdef CONFIG_RT_GROUP_SCHED
9898 static void free_rt_sched_group(struct task_group
*tg
)
9902 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9904 for_each_possible_cpu(i
) {
9906 kfree(tg
->rt_rq
[i
]);
9908 kfree(tg
->rt_se
[i
]);
9916 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9918 struct rt_rq
*rt_rq
;
9919 struct sched_rt_entity
*rt_se
;
9923 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9926 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9930 init_rt_bandwidth(&tg
->rt_bandwidth
,
9931 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9933 for_each_possible_cpu(i
) {
9936 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9937 GFP_KERNEL
, cpu_to_node(i
));
9941 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9942 GFP_KERNEL
, cpu_to_node(i
));
9946 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9955 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9957 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9958 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9961 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9963 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9965 #else /* !CONFIG_RT_GROUP_SCHED */
9966 static inline void free_rt_sched_group(struct task_group
*tg
)
9971 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9976 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9980 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9983 #endif /* CONFIG_RT_GROUP_SCHED */
9985 #ifdef CONFIG_GROUP_SCHED
9986 static void free_sched_group(struct task_group
*tg
)
9988 free_fair_sched_group(tg
);
9989 free_rt_sched_group(tg
);
9993 /* allocate runqueue etc for a new task group */
9994 struct task_group
*sched_create_group(struct task_group
*parent
)
9996 struct task_group
*tg
;
9997 unsigned long flags
;
10000 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10002 return ERR_PTR(-ENOMEM
);
10004 if (!alloc_fair_sched_group(tg
, parent
))
10007 if (!alloc_rt_sched_group(tg
, parent
))
10010 spin_lock_irqsave(&task_group_lock
, flags
);
10011 for_each_possible_cpu(i
) {
10012 register_fair_sched_group(tg
, i
);
10013 register_rt_sched_group(tg
, i
);
10015 list_add_rcu(&tg
->list
, &task_groups
);
10017 WARN_ON(!parent
); /* root should already exist */
10019 tg
->parent
= parent
;
10020 INIT_LIST_HEAD(&tg
->children
);
10021 list_add_rcu(&tg
->siblings
, &parent
->children
);
10022 spin_unlock_irqrestore(&task_group_lock
, flags
);
10027 free_sched_group(tg
);
10028 return ERR_PTR(-ENOMEM
);
10031 /* rcu callback to free various structures associated with a task group */
10032 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10034 /* now it should be safe to free those cfs_rqs */
10035 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10038 /* Destroy runqueue etc associated with a task group */
10039 void sched_destroy_group(struct task_group
*tg
)
10041 unsigned long flags
;
10044 spin_lock_irqsave(&task_group_lock
, flags
);
10045 for_each_possible_cpu(i
) {
10046 unregister_fair_sched_group(tg
, i
);
10047 unregister_rt_sched_group(tg
, i
);
10049 list_del_rcu(&tg
->list
);
10050 list_del_rcu(&tg
->siblings
);
10051 spin_unlock_irqrestore(&task_group_lock
, flags
);
10053 /* wait for possible concurrent references to cfs_rqs complete */
10054 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10057 /* change task's runqueue when it moves between groups.
10058 * The caller of this function should have put the task in its new group
10059 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10060 * reflect its new group.
10062 void sched_move_task(struct task_struct
*tsk
)
10064 int on_rq
, running
;
10065 unsigned long flags
;
10068 rq
= task_rq_lock(tsk
, &flags
);
10070 update_rq_clock(rq
);
10072 running
= task_current(rq
, tsk
);
10073 on_rq
= tsk
->se
.on_rq
;
10076 dequeue_task(rq
, tsk
, 0);
10077 if (unlikely(running
))
10078 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10080 set_task_rq(tsk
, task_cpu(tsk
));
10082 #ifdef CONFIG_FAIR_GROUP_SCHED
10083 if (tsk
->sched_class
->moved_group
)
10084 tsk
->sched_class
->moved_group(tsk
);
10087 if (unlikely(running
))
10088 tsk
->sched_class
->set_curr_task(rq
);
10090 enqueue_task(rq
, tsk
, 0);
10092 task_rq_unlock(rq
, &flags
);
10094 #endif /* CONFIG_GROUP_SCHED */
10096 #ifdef CONFIG_FAIR_GROUP_SCHED
10097 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10099 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10104 dequeue_entity(cfs_rq
, se
, 0);
10106 se
->load
.weight
= shares
;
10107 se
->load
.inv_weight
= 0;
10110 enqueue_entity(cfs_rq
, se
, 0);
10113 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10115 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10116 struct rq
*rq
= cfs_rq
->rq
;
10117 unsigned long flags
;
10119 spin_lock_irqsave(&rq
->lock
, flags
);
10120 __set_se_shares(se
, shares
);
10121 spin_unlock_irqrestore(&rq
->lock
, flags
);
10124 static DEFINE_MUTEX(shares_mutex
);
10126 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10129 unsigned long flags
;
10132 * We can't change the weight of the root cgroup.
10137 if (shares
< MIN_SHARES
)
10138 shares
= MIN_SHARES
;
10139 else if (shares
> MAX_SHARES
)
10140 shares
= MAX_SHARES
;
10142 mutex_lock(&shares_mutex
);
10143 if (tg
->shares
== shares
)
10146 spin_lock_irqsave(&task_group_lock
, flags
);
10147 for_each_possible_cpu(i
)
10148 unregister_fair_sched_group(tg
, i
);
10149 list_del_rcu(&tg
->siblings
);
10150 spin_unlock_irqrestore(&task_group_lock
, flags
);
10152 /* wait for any ongoing reference to this group to finish */
10153 synchronize_sched();
10156 * Now we are free to modify the group's share on each cpu
10157 * w/o tripping rebalance_share or load_balance_fair.
10159 tg
->shares
= shares
;
10160 for_each_possible_cpu(i
) {
10162 * force a rebalance
10164 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10165 set_se_shares(tg
->se
[i
], shares
);
10169 * Enable load balance activity on this group, by inserting it back on
10170 * each cpu's rq->leaf_cfs_rq_list.
10172 spin_lock_irqsave(&task_group_lock
, flags
);
10173 for_each_possible_cpu(i
)
10174 register_fair_sched_group(tg
, i
);
10175 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10176 spin_unlock_irqrestore(&task_group_lock
, flags
);
10178 mutex_unlock(&shares_mutex
);
10182 unsigned long sched_group_shares(struct task_group
*tg
)
10188 #ifdef CONFIG_RT_GROUP_SCHED
10190 * Ensure that the real time constraints are schedulable.
10192 static DEFINE_MUTEX(rt_constraints_mutex
);
10194 static unsigned long to_ratio(u64 period
, u64 runtime
)
10196 if (runtime
== RUNTIME_INF
)
10199 return div64_u64(runtime
<< 20, period
);
10202 /* Must be called with tasklist_lock held */
10203 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10205 struct task_struct
*g
, *p
;
10207 do_each_thread(g
, p
) {
10208 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10210 } while_each_thread(g
, p
);
10215 struct rt_schedulable_data
{
10216 struct task_group
*tg
;
10221 static int tg_schedulable(struct task_group
*tg
, void *data
)
10223 struct rt_schedulable_data
*d
= data
;
10224 struct task_group
*child
;
10225 unsigned long total
, sum
= 0;
10226 u64 period
, runtime
;
10228 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10229 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10232 period
= d
->rt_period
;
10233 runtime
= d
->rt_runtime
;
10236 #ifdef CONFIG_USER_SCHED
10237 if (tg
== &root_task_group
) {
10238 period
= global_rt_period();
10239 runtime
= global_rt_runtime();
10244 * Cannot have more runtime than the period.
10246 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10250 * Ensure we don't starve existing RT tasks.
10252 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10255 total
= to_ratio(period
, runtime
);
10258 * Nobody can have more than the global setting allows.
10260 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10264 * The sum of our children's runtime should not exceed our own.
10266 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10267 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10268 runtime
= child
->rt_bandwidth
.rt_runtime
;
10270 if (child
== d
->tg
) {
10271 period
= d
->rt_period
;
10272 runtime
= d
->rt_runtime
;
10275 sum
+= to_ratio(period
, runtime
);
10284 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10286 struct rt_schedulable_data data
= {
10288 .rt_period
= period
,
10289 .rt_runtime
= runtime
,
10292 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10295 static int tg_set_bandwidth(struct task_group
*tg
,
10296 u64 rt_period
, u64 rt_runtime
)
10300 mutex_lock(&rt_constraints_mutex
);
10301 read_lock(&tasklist_lock
);
10302 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10306 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10307 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10308 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10310 for_each_possible_cpu(i
) {
10311 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10313 spin_lock(&rt_rq
->rt_runtime_lock
);
10314 rt_rq
->rt_runtime
= rt_runtime
;
10315 spin_unlock(&rt_rq
->rt_runtime_lock
);
10317 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10319 read_unlock(&tasklist_lock
);
10320 mutex_unlock(&rt_constraints_mutex
);
10325 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10327 u64 rt_runtime
, rt_period
;
10329 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10330 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10331 if (rt_runtime_us
< 0)
10332 rt_runtime
= RUNTIME_INF
;
10334 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10337 long sched_group_rt_runtime(struct task_group
*tg
)
10341 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10344 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10345 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10346 return rt_runtime_us
;
10349 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10351 u64 rt_runtime
, rt_period
;
10353 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10354 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10356 if (rt_period
== 0)
10359 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10362 long sched_group_rt_period(struct task_group
*tg
)
10366 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10367 do_div(rt_period_us
, NSEC_PER_USEC
);
10368 return rt_period_us
;
10371 static int sched_rt_global_constraints(void)
10373 u64 runtime
, period
;
10376 if (sysctl_sched_rt_period
<= 0)
10379 runtime
= global_rt_runtime();
10380 period
= global_rt_period();
10383 * Sanity check on the sysctl variables.
10385 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10388 mutex_lock(&rt_constraints_mutex
);
10389 read_lock(&tasklist_lock
);
10390 ret
= __rt_schedulable(NULL
, 0, 0);
10391 read_unlock(&tasklist_lock
);
10392 mutex_unlock(&rt_constraints_mutex
);
10397 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10399 /* Don't accept realtime tasks when there is no way for them to run */
10400 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10406 #else /* !CONFIG_RT_GROUP_SCHED */
10407 static int sched_rt_global_constraints(void)
10409 unsigned long flags
;
10412 if (sysctl_sched_rt_period
<= 0)
10416 * There's always some RT tasks in the root group
10417 * -- migration, kstopmachine etc..
10419 if (sysctl_sched_rt_runtime
== 0)
10422 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10423 for_each_possible_cpu(i
) {
10424 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10426 spin_lock(&rt_rq
->rt_runtime_lock
);
10427 rt_rq
->rt_runtime
= global_rt_runtime();
10428 spin_unlock(&rt_rq
->rt_runtime_lock
);
10430 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10434 #endif /* CONFIG_RT_GROUP_SCHED */
10436 int sched_rt_handler(struct ctl_table
*table
, int write
,
10437 void __user
*buffer
, size_t *lenp
,
10441 int old_period
, old_runtime
;
10442 static DEFINE_MUTEX(mutex
);
10444 mutex_lock(&mutex
);
10445 old_period
= sysctl_sched_rt_period
;
10446 old_runtime
= sysctl_sched_rt_runtime
;
10448 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10450 if (!ret
&& write
) {
10451 ret
= sched_rt_global_constraints();
10453 sysctl_sched_rt_period
= old_period
;
10454 sysctl_sched_rt_runtime
= old_runtime
;
10456 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10457 def_rt_bandwidth
.rt_period
=
10458 ns_to_ktime(global_rt_period());
10461 mutex_unlock(&mutex
);
10466 #ifdef CONFIG_CGROUP_SCHED
10468 /* return corresponding task_group object of a cgroup */
10469 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10471 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10472 struct task_group
, css
);
10475 static struct cgroup_subsys_state
*
10476 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10478 struct task_group
*tg
, *parent
;
10480 if (!cgrp
->parent
) {
10481 /* This is early initialization for the top cgroup */
10482 return &init_task_group
.css
;
10485 parent
= cgroup_tg(cgrp
->parent
);
10486 tg
= sched_create_group(parent
);
10488 return ERR_PTR(-ENOMEM
);
10494 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10496 struct task_group
*tg
= cgroup_tg(cgrp
);
10498 sched_destroy_group(tg
);
10502 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10504 #ifdef CONFIG_RT_GROUP_SCHED
10505 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10508 /* We don't support RT-tasks being in separate groups */
10509 if (tsk
->sched_class
!= &fair_sched_class
)
10516 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10517 struct task_struct
*tsk
, bool threadgroup
)
10519 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10523 struct task_struct
*c
;
10525 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10526 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10538 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10539 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10542 sched_move_task(tsk
);
10544 struct task_struct
*c
;
10546 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10547 sched_move_task(c
);
10553 #ifdef CONFIG_FAIR_GROUP_SCHED
10554 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10557 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10560 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10562 struct task_group
*tg
= cgroup_tg(cgrp
);
10564 return (u64
) tg
->shares
;
10566 #endif /* CONFIG_FAIR_GROUP_SCHED */
10568 #ifdef CONFIG_RT_GROUP_SCHED
10569 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10572 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10575 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10577 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10580 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10583 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10586 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10588 return sched_group_rt_period(cgroup_tg(cgrp
));
10590 #endif /* CONFIG_RT_GROUP_SCHED */
10592 static struct cftype cpu_files
[] = {
10593 #ifdef CONFIG_FAIR_GROUP_SCHED
10596 .read_u64
= cpu_shares_read_u64
,
10597 .write_u64
= cpu_shares_write_u64
,
10600 #ifdef CONFIG_RT_GROUP_SCHED
10602 .name
= "rt_runtime_us",
10603 .read_s64
= cpu_rt_runtime_read
,
10604 .write_s64
= cpu_rt_runtime_write
,
10607 .name
= "rt_period_us",
10608 .read_u64
= cpu_rt_period_read_uint
,
10609 .write_u64
= cpu_rt_period_write_uint
,
10614 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10616 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10619 struct cgroup_subsys cpu_cgroup_subsys
= {
10621 .create
= cpu_cgroup_create
,
10622 .destroy
= cpu_cgroup_destroy
,
10623 .can_attach
= cpu_cgroup_can_attach
,
10624 .attach
= cpu_cgroup_attach
,
10625 .populate
= cpu_cgroup_populate
,
10626 .subsys_id
= cpu_cgroup_subsys_id
,
10630 #endif /* CONFIG_CGROUP_SCHED */
10632 #ifdef CONFIG_CGROUP_CPUACCT
10635 * CPU accounting code for task groups.
10637 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10638 * (balbir@in.ibm.com).
10641 /* track cpu usage of a group of tasks and its child groups */
10643 struct cgroup_subsys_state css
;
10644 /* cpuusage holds pointer to a u64-type object on every cpu */
10646 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10647 struct cpuacct
*parent
;
10650 struct cgroup_subsys cpuacct_subsys
;
10652 /* return cpu accounting group corresponding to this container */
10653 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10655 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10656 struct cpuacct
, css
);
10659 /* return cpu accounting group to which this task belongs */
10660 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10662 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10663 struct cpuacct
, css
);
10666 /* create a new cpu accounting group */
10667 static struct cgroup_subsys_state
*cpuacct_create(
10668 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10670 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10676 ca
->cpuusage
= alloc_percpu(u64
);
10680 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10681 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10682 goto out_free_counters
;
10685 ca
->parent
= cgroup_ca(cgrp
->parent
);
10691 percpu_counter_destroy(&ca
->cpustat
[i
]);
10692 free_percpu(ca
->cpuusage
);
10696 return ERR_PTR(-ENOMEM
);
10699 /* destroy an existing cpu accounting group */
10701 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10703 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10706 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10707 percpu_counter_destroy(&ca
->cpustat
[i
]);
10708 free_percpu(ca
->cpuusage
);
10712 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10714 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10717 #ifndef CONFIG_64BIT
10719 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10721 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10723 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10731 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10733 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10735 #ifndef CONFIG_64BIT
10737 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10739 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10741 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10747 /* return total cpu usage (in nanoseconds) of a group */
10748 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10750 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10751 u64 totalcpuusage
= 0;
10754 for_each_present_cpu(i
)
10755 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10757 return totalcpuusage
;
10760 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10763 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10772 for_each_present_cpu(i
)
10773 cpuacct_cpuusage_write(ca
, i
, 0);
10779 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10780 struct seq_file
*m
)
10782 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10786 for_each_present_cpu(i
) {
10787 percpu
= cpuacct_cpuusage_read(ca
, i
);
10788 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10790 seq_printf(m
, "\n");
10794 static const char *cpuacct_stat_desc
[] = {
10795 [CPUACCT_STAT_USER
] = "user",
10796 [CPUACCT_STAT_SYSTEM
] = "system",
10799 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10800 struct cgroup_map_cb
*cb
)
10802 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10805 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10806 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10807 val
= cputime64_to_clock_t(val
);
10808 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10813 static struct cftype files
[] = {
10816 .read_u64
= cpuusage_read
,
10817 .write_u64
= cpuusage_write
,
10820 .name
= "usage_percpu",
10821 .read_seq_string
= cpuacct_percpu_seq_read
,
10825 .read_map
= cpuacct_stats_show
,
10829 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10831 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10835 * charge this task's execution time to its accounting group.
10837 * called with rq->lock held.
10839 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10841 struct cpuacct
*ca
;
10844 if (unlikely(!cpuacct_subsys
.active
))
10847 cpu
= task_cpu(tsk
);
10853 for (; ca
; ca
= ca
->parent
) {
10854 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10855 *cpuusage
+= cputime
;
10862 * Charge the system/user time to the task's accounting group.
10864 static void cpuacct_update_stats(struct task_struct
*tsk
,
10865 enum cpuacct_stat_index idx
, cputime_t val
)
10867 struct cpuacct
*ca
;
10869 if (unlikely(!cpuacct_subsys
.active
))
10876 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10882 struct cgroup_subsys cpuacct_subsys
= {
10884 .create
= cpuacct_create
,
10885 .destroy
= cpuacct_destroy
,
10886 .populate
= cpuacct_populate
,
10887 .subsys_id
= cpuacct_subsys_id
,
10889 #endif /* CONFIG_CGROUP_CPUACCT */
10893 int rcu_expedited_torture_stats(char *page
)
10897 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10899 void synchronize_sched_expedited(void)
10902 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10904 #else /* #ifndef CONFIG_SMP */
10906 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10907 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10909 #define RCU_EXPEDITED_STATE_POST -2
10910 #define RCU_EXPEDITED_STATE_IDLE -1
10912 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10914 int rcu_expedited_torture_stats(char *page
)
10919 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10920 for_each_online_cpu(cpu
) {
10921 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10922 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10924 cnt
+= sprintf(&page
[cnt
], "\n");
10927 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10929 static long synchronize_sched_expedited_count
;
10932 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10933 * approach to force grace period to end quickly. This consumes
10934 * significant time on all CPUs, and is thus not recommended for
10935 * any sort of common-case code.
10937 * Note that it is illegal to call this function while holding any
10938 * lock that is acquired by a CPU-hotplug notifier. Failing to
10939 * observe this restriction will result in deadlock.
10941 void synchronize_sched_expedited(void)
10944 unsigned long flags
;
10945 bool need_full_sync
= 0;
10947 struct migration_req
*req
;
10951 smp_mb(); /* ensure prior mod happens before capturing snap. */
10952 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10954 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10956 if (trycount
++ < 10)
10957 udelay(trycount
* num_online_cpus());
10959 synchronize_sched();
10962 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10963 smp_mb(); /* ensure test happens before caller kfree */
10968 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10969 for_each_online_cpu(cpu
) {
10971 req
= &per_cpu(rcu_migration_req
, cpu
);
10972 init_completion(&req
->done
);
10974 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10975 spin_lock_irqsave(&rq
->lock
, flags
);
10976 list_add(&req
->list
, &rq
->migration_queue
);
10977 spin_unlock_irqrestore(&rq
->lock
, flags
);
10978 wake_up_process(rq
->migration_thread
);
10980 for_each_online_cpu(cpu
) {
10981 rcu_expedited_state
= cpu
;
10982 req
= &per_cpu(rcu_migration_req
, cpu
);
10984 wait_for_completion(&req
->done
);
10985 spin_lock_irqsave(&rq
->lock
, flags
);
10986 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10987 need_full_sync
= 1;
10988 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10989 spin_unlock_irqrestore(&rq
->lock
, flags
);
10991 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10992 synchronize_sched_expedited_count
++;
10993 mutex_unlock(&rcu_sched_expedited_mutex
);
10995 if (need_full_sync
)
10996 synchronize_sched();
10998 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
11000 #endif /* #else #ifndef CONFIG_SMP */