4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity
**se
;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq
**cfs_rq
;
259 unsigned long shares
;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity
**rt_se
;
264 struct rt_rq
**rt_rq
;
266 struct rt_bandwidth rt_bandwidth
;
270 struct list_head list
;
272 struct task_group
*parent
;
273 struct list_head siblings
;
274 struct list_head children
;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct
*user
)
282 user
->tg
->uid
= user
->uid
;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group
;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq
, init_tg_cfs_rq
);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq
, init_rt_rq
);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock
);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group
.children
);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group
;
346 /* return group to which a task belongs */
347 static inline struct task_group
*task_group(struct task_struct
*p
)
349 struct task_group
*tg
;
351 #ifdef CONFIG_USER_SCHED
353 tg
= __task_cred(p
)->user
->tg
;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
357 struct task_group
, css
);
359 tg
= &init_task_group
;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
369 p
->se
.parent
= task_group(p
)->se
[cpu
];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
374 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
380 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
381 static inline struct task_group
*task_group(struct task_struct
*p
)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load
;
391 unsigned long nr_running
;
396 struct rb_root tasks_timeline
;
397 struct rb_node
*rb_leftmost
;
399 struct list_head tasks
;
400 struct list_head
*balance_iterator
;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity
*curr
, *next
, *last
;
408 unsigned int nr_spread_over
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list
;
422 struct task_group
*tg
; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight
;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load
;
439 * this cpu's part of tg->shares
441 unsigned long shares
;
444 * load.weight at the time we set shares
446 unsigned long rq_weight
;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active
;
454 unsigned long rt_nr_running
;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr
; /* highest queued rt task prio */
459 int next
; /* next highest */
464 unsigned long rt_nr_migratory
;
465 unsigned long rt_nr_total
;
467 struct plist_head pushable_tasks
;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock
;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted
;
479 struct list_head leaf_rt_rq_list
;
480 struct task_group
*tg
;
481 struct sched_rt_entity
*rt_se
;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online
;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask
;
507 struct cpupri cpupri
;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain
;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running
;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
538 unsigned long last_tick_seen
;
539 unsigned char in_nohz_recently
;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load
;
543 unsigned long nr_load_updates
;
545 u64 nr_migrations_in
;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list
;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list
;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible
;
566 struct task_struct
*curr
, *idle
;
567 unsigned long next_balance
;
568 struct mm_struct
*prev_mm
;
575 struct root_domain
*rd
;
576 struct sched_domain
*sd
;
578 unsigned char idle_at_tick
;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task
;
589 struct task_struct
*migration_thread
;
590 struct list_head migration_queue
;
598 /* calc_load related fields */
599 unsigned long calc_load_update
;
600 long calc_load_active
;
602 #ifdef CONFIG_SCHED_HRTICK
604 int hrtick_csd_pending
;
605 struct call_single_data hrtick_csd
;
607 struct hrtimer hrtick_timer
;
610 #ifdef CONFIG_SCHEDSTATS
612 struct sched_info rq_sched_info
;
613 unsigned long long rq_cpu_time
;
614 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
616 /* sys_sched_yield() stats */
617 unsigned int yld_count
;
619 /* schedule() stats */
620 unsigned int sched_switch
;
621 unsigned int sched_count
;
622 unsigned int sched_goidle
;
624 /* try_to_wake_up() stats */
625 unsigned int ttwu_count
;
626 unsigned int ttwu_local
;
629 unsigned int bkl_count
;
633 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
636 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
638 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
641 static inline int cpu_of(struct rq
*rq
)
651 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
652 * See detach_destroy_domains: synchronize_sched for details.
654 * The domain tree of any CPU may only be accessed from within
655 * preempt-disabled sections.
657 #define for_each_domain(cpu, __sd) \
658 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
660 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
661 #define this_rq() (&__get_cpu_var(runqueues))
662 #define task_rq(p) cpu_rq(task_cpu(p))
663 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
664 #define raw_rq() (&__raw_get_cpu_var(runqueues))
666 inline void update_rq_clock(struct rq
*rq
)
668 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
672 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 #ifdef CONFIG_SCHED_DEBUG
675 # define const_debug __read_mostly
677 # define const_debug static const
682 * @cpu: the processor in question.
684 * Returns true if the current cpu runqueue is locked.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu
)
690 return spin_is_locked(&cpu_rq(cpu
)->lock
);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug
unsigned int sysctl_sched_features
=
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly
char *sched_feat_names
[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file
*m
, void *v
)
730 for (i
= 0; sched_feat_names
[i
]; i
++) {
731 if (!(sysctl_sched_features
& (1UL << i
)))
733 seq_printf(m
, "%s ", sched_feat_names
[i
]);
741 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
742 size_t cnt
, loff_t
*ppos
)
752 if (copy_from_user(&buf
, ubuf
, cnt
))
757 if (strncmp(buf
, "NO_", 3) == 0) {
762 for (i
= 0; sched_feat_names
[i
]; i
++) {
763 int len
= strlen(sched_feat_names
[i
]);
765 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
767 sysctl_sched_features
&= ~(1UL << i
);
769 sysctl_sched_features
|= (1UL << i
);
774 if (!sched_feat_names
[i
])
782 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
784 return single_open(filp
, sched_feat_show
, NULL
);
787 static const struct file_operations sched_feat_fops
= {
788 .open
= sched_feat_open
,
789 .write
= sched_feat_write
,
792 .release
= single_release
,
795 static __init
int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
802 late_initcall(sched_init_debug
);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit
= 250000;
819 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
822 * Inject some fuzzyness into changing the per-cpu group shares
823 * this avoids remote rq-locks at the expense of fairness.
826 unsigned int sysctl_sched_shares_thresh
= 4;
829 * period over which we average the RT time consumption, measured
834 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
837 * period over which we measure -rt task cpu usage in us.
840 unsigned int sysctl_sched_rt_period
= 1000000;
842 static __read_mostly
int scheduler_running
;
845 * part of the period that we allow rt tasks to run in us.
848 int sysctl_sched_rt_runtime
= 950000;
850 static inline u64
global_rt_period(void)
852 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
855 static inline u64
global_rt_runtime(void)
857 if (sysctl_sched_rt_runtime
< 0)
860 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
863 #ifndef prepare_arch_switch
864 # define prepare_arch_switch(next) do { } while (0)
866 #ifndef finish_arch_switch
867 # define finish_arch_switch(prev) do { } while (0)
870 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
872 return rq
->curr
== p
;
875 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
876 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
878 return task_current(rq
, p
);
881 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
885 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
887 #ifdef CONFIG_DEBUG_SPINLOCK
888 /* this is a valid case when another task releases the spinlock */
889 rq
->lock
.owner
= current
;
892 * If we are tracking spinlock dependencies then we have to
893 * fix up the runqueue lock - which gets 'carried over' from
896 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
898 spin_unlock_irq(&rq
->lock
);
901 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
902 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
907 return task_current(rq
, p
);
911 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
915 * We can optimise this out completely for !SMP, because the
916 * SMP rebalancing from interrupt is the only thing that cares
921 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 spin_unlock_irq(&rq
->lock
);
924 spin_unlock(&rq
->lock
);
928 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
932 * After ->oncpu is cleared, the task can be moved to a different CPU.
933 * We must ensure this doesn't happen until the switch is completely
939 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
946 * __task_rq_lock - lock the runqueue a given task resides on.
947 * Must be called interrupts disabled.
949 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
953 struct rq
*rq
= task_rq(p
);
954 spin_lock(&rq
->lock
);
955 if (likely(rq
== task_rq(p
)))
957 spin_unlock(&rq
->lock
);
962 * task_rq_lock - lock the runqueue a given task resides on and disable
963 * interrupts. Note the ordering: we can safely lookup the task_rq without
964 * explicitly disabling preemption.
966 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
972 local_irq_save(*flags
);
974 spin_lock(&rq
->lock
);
975 if (likely(rq
== task_rq(p
)))
977 spin_unlock_irqrestore(&rq
->lock
, *flags
);
981 void task_rq_unlock_wait(struct task_struct
*p
)
983 struct rq
*rq
= task_rq(p
);
985 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
986 spin_unlock_wait(&rq
->lock
);
989 static void __task_rq_unlock(struct rq
*rq
)
992 spin_unlock(&rq
->lock
);
995 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
998 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1002 * this_rq_lock - lock this runqueue and disable interrupts.
1004 static struct rq
*this_rq_lock(void)
1005 __acquires(rq
->lock
)
1009 local_irq_disable();
1011 spin_lock(&rq
->lock
);
1016 #ifdef CONFIG_SCHED_HRTICK
1018 * Use HR-timers to deliver accurate preemption points.
1020 * Its all a bit involved since we cannot program an hrt while holding the
1021 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1024 * When we get rescheduled we reprogram the hrtick_timer outside of the
1030 * - enabled by features
1031 * - hrtimer is actually high res
1033 static inline int hrtick_enabled(struct rq
*rq
)
1035 if (!sched_feat(HRTICK
))
1037 if (!cpu_active(cpu_of(rq
)))
1039 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1042 static void hrtick_clear(struct rq
*rq
)
1044 if (hrtimer_active(&rq
->hrtick_timer
))
1045 hrtimer_cancel(&rq
->hrtick_timer
);
1049 * High-resolution timer tick.
1050 * Runs from hardirq context with interrupts disabled.
1052 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1054 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1056 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1058 spin_lock(&rq
->lock
);
1059 update_rq_clock(rq
);
1060 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1061 spin_unlock(&rq
->lock
);
1063 return HRTIMER_NORESTART
;
1068 * called from hardirq (IPI) context
1070 static void __hrtick_start(void *arg
)
1072 struct rq
*rq
= arg
;
1074 spin_lock(&rq
->lock
);
1075 hrtimer_restart(&rq
->hrtick_timer
);
1076 rq
->hrtick_csd_pending
= 0;
1077 spin_unlock(&rq
->lock
);
1081 * Called to set the hrtick timer state.
1083 * called with rq->lock held and irqs disabled
1085 static void hrtick_start(struct rq
*rq
, u64 delay
)
1087 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1088 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1090 hrtimer_set_expires(timer
, time
);
1092 if (rq
== this_rq()) {
1093 hrtimer_restart(timer
);
1094 } else if (!rq
->hrtick_csd_pending
) {
1095 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1096 rq
->hrtick_csd_pending
= 1;
1101 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1103 int cpu
= (int)(long)hcpu
;
1106 case CPU_UP_CANCELED
:
1107 case CPU_UP_CANCELED_FROZEN
:
1108 case CPU_DOWN_PREPARE
:
1109 case CPU_DOWN_PREPARE_FROZEN
:
1111 case CPU_DEAD_FROZEN
:
1112 hrtick_clear(cpu_rq(cpu
));
1119 static __init
void init_hrtick(void)
1121 hotcpu_notifier(hotplug_hrtick
, 0);
1125 * Called to set the hrtick timer state.
1127 * called with rq->lock held and irqs disabled
1129 static void hrtick_start(struct rq
*rq
, u64 delay
)
1131 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1132 HRTIMER_MODE_REL_PINNED
, 0);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq
*rq
)
1143 rq
->hrtick_csd_pending
= 0;
1145 rq
->hrtick_csd
.flags
= 0;
1146 rq
->hrtick_csd
.func
= __hrtick_start
;
1147 rq
->hrtick_csd
.info
= rq
;
1150 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1151 rq
->hrtick_timer
.function
= hrtick
;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq
*rq
)
1158 static inline void init_rq_hrtick(struct rq
*rq
)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct
*p
)
1184 assert_spin_locked(&task_rq(p
)->lock
);
1186 if (test_tsk_need_resched(p
))
1189 set_tsk_need_resched(p
);
1192 if (cpu
== smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p
))
1198 smp_send_reschedule(cpu
);
1201 static void resched_cpu(int cpu
)
1203 struct rq
*rq
= cpu_rq(cpu
);
1204 unsigned long flags
;
1206 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1208 resched_task(cpu_curr(cpu
));
1209 spin_unlock_irqrestore(&rq
->lock
, flags
);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu
)
1225 struct rq
*rq
= cpu_rq(cpu
);
1227 if (cpu
== smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq
->curr
!= rq
->idle
)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_need_resched(rq
->idle
);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq
->idle
))
1250 smp_send_reschedule(cpu
);
1252 #endif /* CONFIG_NO_HZ */
1254 static u64
sched_avg_period(void)
1256 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1259 static void sched_avg_update(struct rq
*rq
)
1261 s64 period
= sched_avg_period();
1263 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1264 rq
->age_stamp
+= period
;
1269 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1271 rq
->rt_avg
+= rt_delta
;
1272 sched_avg_update(rq
);
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct
*p
)
1278 assert_spin_locked(&task_rq(p
)->lock
);
1279 set_tsk_need_resched(p
);
1282 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1305 struct load_weight
*lw
)
1309 if (!lw
->inv_weight
) {
1310 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1313 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1317 tmp
= (u64
)delta_exec
* weight
;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp
> WMULT_CONST
))
1322 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1325 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1327 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1330 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1336 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight
[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult
[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator
{
1404 struct task_struct
*(*start
)(void *);
1405 struct task_struct
*(*next
)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1411 unsigned long max_load_move
, struct sched_domain
*sd
,
1412 enum cpu_idle_type idle
, int *all_pinned
,
1413 int *this_best_prio
, struct rq_iterator
*iterator
);
1416 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1417 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1418 struct rq_iterator
*iterator
);
1421 /* Time spent by the tasks of the cpu accounting group executing in ... */
1422 enum cpuacct_stat_index
{
1423 CPUACCT_STAT_USER
, /* ... user mode */
1424 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1426 CPUACCT_STAT_NSTATS
,
1429 #ifdef CONFIG_CGROUP_CPUACCT
1430 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1431 static void cpuacct_update_stats(struct task_struct
*tsk
,
1432 enum cpuacct_stat_index idx
, cputime_t val
);
1434 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1435 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1436 enum cpuacct_stat_index idx
, cputime_t val
) {}
1439 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1441 update_load_add(&rq
->load
, load
);
1444 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1446 update_load_sub(&rq
->load
, load
);
1449 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1450 typedef int (*tg_visitor
)(struct task_group
*, void *);
1453 * Iterate the full tree, calling @down when first entering a node and @up when
1454 * leaving it for the final time.
1456 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1458 struct task_group
*parent
, *child
;
1462 parent
= &root_task_group
;
1464 ret
= (*down
)(parent
, data
);
1467 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1474 ret
= (*up
)(parent
, data
);
1479 parent
= parent
->parent
;
1488 static int tg_nop(struct task_group
*tg
, void *data
)
1495 /* Used instead of source_load when we know the type == 0 */
1496 static unsigned long weighted_cpuload(const int cpu
)
1498 return cpu_rq(cpu
)->load
.weight
;
1502 * Return a low guess at the load of a migration-source cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 * We want to under-estimate the load of migration sources, to
1506 * balance conservatively.
1508 static unsigned long source_load(int cpu
, int type
)
1510 struct rq
*rq
= cpu_rq(cpu
);
1511 unsigned long total
= weighted_cpuload(cpu
);
1513 if (type
== 0 || !sched_feat(LB_BIAS
))
1516 return min(rq
->cpu_load
[type
-1], total
);
1520 * Return a high guess at the load of a migration-target cpu weighted
1521 * according to the scheduling class and "nice" value.
1523 static unsigned long target_load(int cpu
, int type
)
1525 struct rq
*rq
= cpu_rq(cpu
);
1526 unsigned long total
= weighted_cpuload(cpu
);
1528 if (type
== 0 || !sched_feat(LB_BIAS
))
1531 return max(rq
->cpu_load
[type
-1], total
);
1534 static struct sched_group
*group_of(int cpu
)
1536 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1544 static unsigned long power_of(int cpu
)
1546 struct sched_group
*group
= group_of(cpu
);
1549 return SCHED_LOAD_SCALE
;
1551 return group
->cpu_power
;
1554 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1556 static unsigned long cpu_avg_load_per_task(int cpu
)
1558 struct rq
*rq
= cpu_rq(cpu
);
1559 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1562 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1564 rq
->avg_load_per_task
= 0;
1566 return rq
->avg_load_per_task
;
1569 #ifdef CONFIG_FAIR_GROUP_SCHED
1571 static __read_mostly
unsigned long *update_shares_data
;
1573 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1579 unsigned long sd_shares
,
1580 unsigned long sd_rq_weight
,
1581 unsigned long *usd_rq_weight
)
1583 unsigned long shares
, rq_weight
;
1586 rq_weight
= usd_rq_weight
[cpu
];
1589 rq_weight
= NICE_0_LOAD
;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1598 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1600 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1601 sysctl_sched_shares_thresh
) {
1602 struct rq
*rq
= cpu_rq(cpu
);
1603 unsigned long flags
;
1605 spin_lock_irqsave(&rq
->lock
, flags
);
1606 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1607 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1608 __set_se_shares(tg
->se
[cpu
], shares
);
1609 spin_unlock_irqrestore(&rq
->lock
, flags
);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group
*tg
, void *data
)
1620 unsigned long weight
, rq_weight
= 0, shares
= 0;
1621 unsigned long *usd_rq_weight
;
1622 struct sched_domain
*sd
= data
;
1623 unsigned long flags
;
1629 local_irq_save(flags
);
1630 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1632 for_each_cpu(i
, sched_domain_span(sd
)) {
1633 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1634 usd_rq_weight
[i
] = weight
;
1637 * If there are currently no tasks on the cpu pretend there
1638 * is one of average load so that when a new task gets to
1639 * run here it will not get delayed by group starvation.
1642 weight
= NICE_0_LOAD
;
1644 rq_weight
+= weight
;
1645 shares
+= tg
->cfs_rq
[i
]->shares
;
1648 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1649 shares
= tg
->shares
;
1651 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1652 shares
= tg
->shares
;
1654 for_each_cpu(i
, sched_domain_span(sd
))
1655 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1657 local_irq_restore(flags
);
1663 * Compute the cpu's hierarchical load factor for each task group.
1664 * This needs to be done in a top-down fashion because the load of a child
1665 * group is a fraction of its parents load.
1667 static int tg_load_down(struct task_group
*tg
, void *data
)
1670 long cpu
= (long)data
;
1673 load
= cpu_rq(cpu
)->load
.weight
;
1675 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1676 load
*= tg
->cfs_rq
[cpu
]->shares
;
1677 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1680 tg
->cfs_rq
[cpu
]->h_load
= load
;
1685 static void update_shares(struct sched_domain
*sd
)
1690 if (root_task_group_empty())
1693 now
= cpu_clock(raw_smp_processor_id());
1694 elapsed
= now
- sd
->last_update
;
1696 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1697 sd
->last_update
= now
;
1698 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1702 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1704 if (root_task_group_empty())
1707 spin_unlock(&rq
->lock
);
1709 spin_lock(&rq
->lock
);
1712 static void update_h_load(long cpu
)
1714 if (root_task_group_empty())
1717 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1722 static inline void update_shares(struct sched_domain
*sd
)
1726 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1732 #ifdef CONFIG_PREEMPT
1734 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1737 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1738 * way at the expense of forcing extra atomic operations in all
1739 * invocations. This assures that the double_lock is acquired using the
1740 * same underlying policy as the spinlock_t on this architecture, which
1741 * reduces latency compared to the unfair variant below. However, it
1742 * also adds more overhead and therefore may reduce throughput.
1744 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1745 __releases(this_rq
->lock
)
1746 __acquires(busiest
->lock
)
1747 __acquires(this_rq
->lock
)
1749 spin_unlock(&this_rq
->lock
);
1750 double_rq_lock(this_rq
, busiest
);
1757 * Unfair double_lock_balance: Optimizes throughput at the expense of
1758 * latency by eliminating extra atomic operations when the locks are
1759 * already in proper order on entry. This favors lower cpu-ids and will
1760 * grant the double lock to lower cpus over higher ids under contention,
1761 * regardless of entry order into the function.
1763 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1764 __releases(this_rq
->lock
)
1765 __acquires(busiest
->lock
)
1766 __acquires(this_rq
->lock
)
1770 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1771 if (busiest
< this_rq
) {
1772 spin_unlock(&this_rq
->lock
);
1773 spin_lock(&busiest
->lock
);
1774 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1777 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1782 #endif /* CONFIG_PREEMPT */
1785 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1787 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1789 if (unlikely(!irqs_disabled())) {
1790 /* printk() doesn't work good under rq->lock */
1791 spin_unlock(&this_rq
->lock
);
1795 return _double_lock_balance(this_rq
, busiest
);
1798 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1799 __releases(busiest
->lock
)
1801 spin_unlock(&busiest
->lock
);
1802 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1810 cfs_rq
->shares
= shares
;
1815 static void calc_load_account_active(struct rq
*this_rq
);
1816 static void update_sysctl(void);
1818 #include "sched_stats.h"
1819 #include "sched_idletask.c"
1820 #include "sched_fair.c"
1821 #include "sched_rt.c"
1822 #ifdef CONFIG_SCHED_DEBUG
1823 # include "sched_debug.c"
1826 #define sched_class_highest (&rt_sched_class)
1827 #define for_each_class(class) \
1828 for (class = sched_class_highest; class; class = class->next)
1830 static void inc_nr_running(struct rq
*rq
)
1835 static void dec_nr_running(struct rq
*rq
)
1840 static void set_load_weight(struct task_struct
*p
)
1842 if (task_has_rt_policy(p
)) {
1843 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1844 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1849 * SCHED_IDLE tasks get minimal weight:
1851 if (p
->policy
== SCHED_IDLE
) {
1852 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1853 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1857 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1858 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1861 static void update_avg(u64
*avg
, u64 sample
)
1863 s64 diff
= sample
- *avg
;
1867 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1870 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1872 sched_info_queued(p
);
1873 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1877 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1880 if (p
->se
.last_wakeup
) {
1881 update_avg(&p
->se
.avg_overlap
,
1882 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1883 p
->se
.last_wakeup
= 0;
1885 update_avg(&p
->se
.avg_wakeup
,
1886 sysctl_sched_wakeup_granularity
);
1890 sched_info_dequeued(p
);
1891 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1896 * __normal_prio - return the priority that is based on the static prio
1898 static inline int __normal_prio(struct task_struct
*p
)
1900 return p
->static_prio
;
1904 * Calculate the expected normal priority: i.e. priority
1905 * without taking RT-inheritance into account. Might be
1906 * boosted by interactivity modifiers. Changes upon fork,
1907 * setprio syscalls, and whenever the interactivity
1908 * estimator recalculates.
1910 static inline int normal_prio(struct task_struct
*p
)
1914 if (task_has_rt_policy(p
))
1915 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1917 prio
= __normal_prio(p
);
1922 * Calculate the current priority, i.e. the priority
1923 * taken into account by the scheduler. This value might
1924 * be boosted by RT tasks, or might be boosted by
1925 * interactivity modifiers. Will be RT if the task got
1926 * RT-boosted. If not then it returns p->normal_prio.
1928 static int effective_prio(struct task_struct
*p
)
1930 p
->normal_prio
= normal_prio(p
);
1932 * If we are RT tasks or we were boosted to RT priority,
1933 * keep the priority unchanged. Otherwise, update priority
1934 * to the normal priority:
1936 if (!rt_prio(p
->prio
))
1937 return p
->normal_prio
;
1942 * activate_task - move a task to the runqueue.
1944 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1946 if (task_contributes_to_load(p
))
1947 rq
->nr_uninterruptible
--;
1949 enqueue_task(rq
, p
, wakeup
);
1954 * deactivate_task - remove a task from the runqueue.
1956 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1958 if (task_contributes_to_load(p
))
1959 rq
->nr_uninterruptible
++;
1961 dequeue_task(rq
, p
, sleep
);
1966 * task_curr - is this task currently executing on a CPU?
1967 * @p: the task in question.
1969 inline int task_curr(const struct task_struct
*p
)
1971 return cpu_curr(task_cpu(p
)) == p
;
1974 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1976 set_task_rq(p
, cpu
);
1979 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1980 * successfuly executed on another CPU. We must ensure that updates of
1981 * per-task data have been completed by this moment.
1984 task_thread_info(p
)->cpu
= cpu
;
1988 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1989 const struct sched_class
*prev_class
,
1990 int oldprio
, int running
)
1992 if (prev_class
!= p
->sched_class
) {
1993 if (prev_class
->switched_from
)
1994 prev_class
->switched_from(rq
, p
, running
);
1995 p
->sched_class
->switched_to(rq
, p
, running
);
1997 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2001 * kthread_bind - bind a just-created kthread to a cpu.
2002 * @p: thread created by kthread_create().
2003 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2005 * Description: This function is equivalent to set_cpus_allowed(),
2006 * except that @cpu doesn't need to be online, and the thread must be
2007 * stopped (i.e., just returned from kthread_create()).
2009 * Function lives here instead of kthread.c because it messes with
2010 * scheduler internals which require locking.
2012 void kthread_bind(struct task_struct
*p
, unsigned int cpu
)
2014 struct rq
*rq
= cpu_rq(cpu
);
2015 unsigned long flags
;
2017 /* Must have done schedule() in kthread() before we set_task_cpu */
2018 if (!wait_task_inactive(p
, TASK_UNINTERRUPTIBLE
)) {
2023 spin_lock_irqsave(&rq
->lock
, flags
);
2024 set_task_cpu(p
, cpu
);
2025 p
->cpus_allowed
= cpumask_of_cpu(cpu
);
2026 p
->rt
.nr_cpus_allowed
= 1;
2027 p
->flags
|= PF_THREAD_BOUND
;
2028 spin_unlock_irqrestore(&rq
->lock
, flags
);
2030 EXPORT_SYMBOL(kthread_bind
);
2034 * Is this task likely cache-hot:
2037 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2041 if (p
->sched_class
!= &fair_sched_class
)
2045 * Buddy candidates are cache hot:
2047 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2048 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2049 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2052 if (sysctl_sched_migration_cost
== -1)
2054 if (sysctl_sched_migration_cost
== 0)
2057 delta
= now
- p
->se
.exec_start
;
2059 return delta
< (s64
)sysctl_sched_migration_cost
;
2063 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2065 int old_cpu
= task_cpu(p
);
2066 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2067 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2068 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2071 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2073 trace_sched_migrate_task(p
, new_cpu
);
2075 #ifdef CONFIG_SCHEDSTATS
2076 if (p
->se
.wait_start
)
2077 p
->se
.wait_start
-= clock_offset
;
2078 if (p
->se
.sleep_start
)
2079 p
->se
.sleep_start
-= clock_offset
;
2080 if (p
->se
.block_start
)
2081 p
->se
.block_start
-= clock_offset
;
2083 if (old_cpu
!= new_cpu
) {
2084 p
->se
.nr_migrations
++;
2085 new_rq
->nr_migrations_in
++;
2086 #ifdef CONFIG_SCHEDSTATS
2087 if (task_hot(p
, old_rq
->clock
, NULL
))
2088 schedstat_inc(p
, se
.nr_forced2_migrations
);
2090 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2093 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2094 new_cfsrq
->min_vruntime
;
2096 __set_task_cpu(p
, new_cpu
);
2099 struct migration_req
{
2100 struct list_head list
;
2102 struct task_struct
*task
;
2105 struct completion done
;
2109 * The task's runqueue lock must be held.
2110 * Returns true if you have to wait for migration thread.
2113 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2115 struct rq
*rq
= task_rq(p
);
2118 * If the task is not on a runqueue (and not running), then
2119 * it is sufficient to simply update the task's cpu field.
2121 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2122 set_task_cpu(p
, dest_cpu
);
2126 init_completion(&req
->done
);
2128 req
->dest_cpu
= dest_cpu
;
2129 list_add(&req
->list
, &rq
->migration_queue
);
2135 * wait_task_context_switch - wait for a thread to complete at least one
2138 * @p must not be current.
2140 void wait_task_context_switch(struct task_struct
*p
)
2142 unsigned long nvcsw
, nivcsw
, flags
;
2150 * The runqueue is assigned before the actual context
2151 * switch. We need to take the runqueue lock.
2153 * We could check initially without the lock but it is
2154 * very likely that we need to take the lock in every
2157 rq
= task_rq_lock(p
, &flags
);
2158 running
= task_running(rq
, p
);
2159 task_rq_unlock(rq
, &flags
);
2161 if (likely(!running
))
2164 * The switch count is incremented before the actual
2165 * context switch. We thus wait for two switches to be
2166 * sure at least one completed.
2168 if ((p
->nvcsw
- nvcsw
) > 1)
2170 if ((p
->nivcsw
- nivcsw
) > 1)
2178 * wait_task_inactive - wait for a thread to unschedule.
2180 * If @match_state is nonzero, it's the @p->state value just checked and
2181 * not expected to change. If it changes, i.e. @p might have woken up,
2182 * then return zero. When we succeed in waiting for @p to be off its CPU,
2183 * we return a positive number (its total switch count). If a second call
2184 * a short while later returns the same number, the caller can be sure that
2185 * @p has remained unscheduled the whole time.
2187 * The caller must ensure that the task *will* unschedule sometime soon,
2188 * else this function might spin for a *long* time. This function can't
2189 * be called with interrupts off, or it may introduce deadlock with
2190 * smp_call_function() if an IPI is sent by the same process we are
2191 * waiting to become inactive.
2193 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2195 unsigned long flags
;
2202 * We do the initial early heuristics without holding
2203 * any task-queue locks at all. We'll only try to get
2204 * the runqueue lock when things look like they will
2210 * If the task is actively running on another CPU
2211 * still, just relax and busy-wait without holding
2214 * NOTE! Since we don't hold any locks, it's not
2215 * even sure that "rq" stays as the right runqueue!
2216 * But we don't care, since "task_running()" will
2217 * return false if the runqueue has changed and p
2218 * is actually now running somewhere else!
2220 while (task_running(rq
, p
)) {
2221 if (match_state
&& unlikely(p
->state
!= match_state
))
2227 * Ok, time to look more closely! We need the rq
2228 * lock now, to be *sure*. If we're wrong, we'll
2229 * just go back and repeat.
2231 rq
= task_rq_lock(p
, &flags
);
2232 trace_sched_wait_task(rq
, p
);
2233 running
= task_running(rq
, p
);
2234 on_rq
= p
->se
.on_rq
;
2236 if (!match_state
|| p
->state
== match_state
)
2237 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2238 task_rq_unlock(rq
, &flags
);
2241 * If it changed from the expected state, bail out now.
2243 if (unlikely(!ncsw
))
2247 * Was it really running after all now that we
2248 * checked with the proper locks actually held?
2250 * Oops. Go back and try again..
2252 if (unlikely(running
)) {
2258 * It's not enough that it's not actively running,
2259 * it must be off the runqueue _entirely_, and not
2262 * So if it was still runnable (but just not actively
2263 * running right now), it's preempted, and we should
2264 * yield - it could be a while.
2266 if (unlikely(on_rq
)) {
2267 schedule_timeout_uninterruptible(1);
2272 * Ahh, all good. It wasn't running, and it wasn't
2273 * runnable, which means that it will never become
2274 * running in the future either. We're all done!
2283 * kick_process - kick a running thread to enter/exit the kernel
2284 * @p: the to-be-kicked thread
2286 * Cause a process which is running on another CPU to enter
2287 * kernel-mode, without any delay. (to get signals handled.)
2289 * NOTE: this function doesnt have to take the runqueue lock,
2290 * because all it wants to ensure is that the remote task enters
2291 * the kernel. If the IPI races and the task has been migrated
2292 * to another CPU then no harm is done and the purpose has been
2295 void kick_process(struct task_struct
*p
)
2301 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2302 smp_send_reschedule(cpu
);
2305 EXPORT_SYMBOL_GPL(kick_process
);
2306 #endif /* CONFIG_SMP */
2309 * task_oncpu_function_call - call a function on the cpu on which a task runs
2310 * @p: the task to evaluate
2311 * @func: the function to be called
2312 * @info: the function call argument
2314 * Calls the function @func when the task is currently running. This might
2315 * be on the current CPU, which just calls the function directly
2317 void task_oncpu_function_call(struct task_struct
*p
,
2318 void (*func
) (void *info
), void *info
)
2325 smp_call_function_single(cpu
, func
, info
, 1);
2330 * try_to_wake_up - wake up a thread
2331 * @p: the to-be-woken-up thread
2332 * @state: the mask of task states that can be woken
2333 * @sync: do a synchronous wakeup?
2335 * Put it on the run-queue if it's not already there. The "current"
2336 * thread is always on the run-queue (except when the actual
2337 * re-schedule is in progress), and as such you're allowed to do
2338 * the simpler "current->state = TASK_RUNNING" to mark yourself
2339 * runnable without the overhead of this.
2341 * returns failure only if the task is already active.
2343 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2346 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2347 unsigned long flags
;
2348 struct rq
*rq
, *orig_rq
;
2350 if (!sched_feat(SYNC_WAKEUPS
))
2351 wake_flags
&= ~WF_SYNC
;
2353 this_cpu
= get_cpu();
2356 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2357 update_rq_clock(rq
);
2358 if (!(p
->state
& state
))
2368 if (unlikely(task_running(rq
, p
)))
2372 * In order to handle concurrent wakeups and release the rq->lock
2373 * we put the task in TASK_WAKING state.
2375 * First fix up the nr_uninterruptible count:
2377 if (task_contributes_to_load(p
))
2378 rq
->nr_uninterruptible
--;
2379 p
->state
= TASK_WAKING
;
2380 task_rq_unlock(rq
, &flags
);
2382 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2383 if (cpu
!= orig_cpu
)
2384 set_task_cpu(p
, cpu
);
2386 rq
= task_rq_lock(p
, &flags
);
2389 update_rq_clock(rq
);
2391 WARN_ON(p
->state
!= TASK_WAKING
);
2394 #ifdef CONFIG_SCHEDSTATS
2395 schedstat_inc(rq
, ttwu_count
);
2396 if (cpu
== this_cpu
)
2397 schedstat_inc(rq
, ttwu_local
);
2399 struct sched_domain
*sd
;
2400 for_each_domain(this_cpu
, sd
) {
2401 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2402 schedstat_inc(sd
, ttwu_wake_remote
);
2407 #endif /* CONFIG_SCHEDSTATS */
2410 #endif /* CONFIG_SMP */
2411 schedstat_inc(p
, se
.nr_wakeups
);
2412 if (wake_flags
& WF_SYNC
)
2413 schedstat_inc(p
, se
.nr_wakeups_sync
);
2414 if (orig_cpu
!= cpu
)
2415 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2416 if (cpu
== this_cpu
)
2417 schedstat_inc(p
, se
.nr_wakeups_local
);
2419 schedstat_inc(p
, se
.nr_wakeups_remote
);
2420 activate_task(rq
, p
, 1);
2424 * Only attribute actual wakeups done by this task.
2426 if (!in_interrupt()) {
2427 struct sched_entity
*se
= ¤t
->se
;
2428 u64 sample
= se
->sum_exec_runtime
;
2430 if (se
->last_wakeup
)
2431 sample
-= se
->last_wakeup
;
2433 sample
-= se
->start_runtime
;
2434 update_avg(&se
->avg_wakeup
, sample
);
2436 se
->last_wakeup
= se
->sum_exec_runtime
;
2440 trace_sched_wakeup(rq
, p
, success
);
2441 check_preempt_curr(rq
, p
, wake_flags
);
2443 p
->state
= TASK_RUNNING
;
2445 if (p
->sched_class
->task_wake_up
)
2446 p
->sched_class
->task_wake_up(rq
, p
);
2448 if (unlikely(rq
->idle_stamp
)) {
2449 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2450 u64 max
= 2*sysctl_sched_migration_cost
;
2455 update_avg(&rq
->avg_idle
, delta
);
2460 task_rq_unlock(rq
, &flags
);
2467 * wake_up_process - Wake up a specific process
2468 * @p: The process to be woken up.
2470 * Attempt to wake up the nominated process and move it to the set of runnable
2471 * processes. Returns 1 if the process was woken up, 0 if it was already
2474 * It may be assumed that this function implies a write memory barrier before
2475 * changing the task state if and only if any tasks are woken up.
2477 int wake_up_process(struct task_struct
*p
)
2479 return try_to_wake_up(p
, TASK_ALL
, 0);
2481 EXPORT_SYMBOL(wake_up_process
);
2483 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2485 return try_to_wake_up(p
, state
, 0);
2489 * Perform scheduler related setup for a newly forked process p.
2490 * p is forked by current.
2492 * __sched_fork() is basic setup used by init_idle() too:
2494 static void __sched_fork(struct task_struct
*p
)
2496 p
->se
.exec_start
= 0;
2497 p
->se
.sum_exec_runtime
= 0;
2498 p
->se
.prev_sum_exec_runtime
= 0;
2499 p
->se
.nr_migrations
= 0;
2500 p
->se
.last_wakeup
= 0;
2501 p
->se
.avg_overlap
= 0;
2502 p
->se
.start_runtime
= 0;
2503 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2504 p
->se
.avg_running
= 0;
2506 #ifdef CONFIG_SCHEDSTATS
2507 p
->se
.wait_start
= 0;
2509 p
->se
.wait_count
= 0;
2512 p
->se
.sleep_start
= 0;
2513 p
->se
.sleep_max
= 0;
2514 p
->se
.sum_sleep_runtime
= 0;
2516 p
->se
.block_start
= 0;
2517 p
->se
.block_max
= 0;
2519 p
->se
.slice_max
= 0;
2521 p
->se
.nr_migrations_cold
= 0;
2522 p
->se
.nr_failed_migrations_affine
= 0;
2523 p
->se
.nr_failed_migrations_running
= 0;
2524 p
->se
.nr_failed_migrations_hot
= 0;
2525 p
->se
.nr_forced_migrations
= 0;
2526 p
->se
.nr_forced2_migrations
= 0;
2528 p
->se
.nr_wakeups
= 0;
2529 p
->se
.nr_wakeups_sync
= 0;
2530 p
->se
.nr_wakeups_migrate
= 0;
2531 p
->se
.nr_wakeups_local
= 0;
2532 p
->se
.nr_wakeups_remote
= 0;
2533 p
->se
.nr_wakeups_affine
= 0;
2534 p
->se
.nr_wakeups_affine_attempts
= 0;
2535 p
->se
.nr_wakeups_passive
= 0;
2536 p
->se
.nr_wakeups_idle
= 0;
2540 INIT_LIST_HEAD(&p
->rt
.run_list
);
2542 INIT_LIST_HEAD(&p
->se
.group_node
);
2544 #ifdef CONFIG_PREEMPT_NOTIFIERS
2545 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2549 * We mark the process as running here, but have not actually
2550 * inserted it onto the runqueue yet. This guarantees that
2551 * nobody will actually run it, and a signal or other external
2552 * event cannot wake it up and insert it on the runqueue either.
2554 p
->state
= TASK_RUNNING
;
2558 * fork()/clone()-time setup:
2560 void sched_fork(struct task_struct
*p
, int clone_flags
)
2562 int cpu
= get_cpu();
2567 * Revert to default priority/policy on fork if requested.
2569 if (unlikely(p
->sched_reset_on_fork
)) {
2570 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2571 p
->policy
= SCHED_NORMAL
;
2572 p
->normal_prio
= p
->static_prio
;
2575 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2576 p
->static_prio
= NICE_TO_PRIO(0);
2577 p
->normal_prio
= p
->static_prio
;
2582 * We don't need the reset flag anymore after the fork. It has
2583 * fulfilled its duty:
2585 p
->sched_reset_on_fork
= 0;
2589 * Make sure we do not leak PI boosting priority to the child.
2591 p
->prio
= current
->normal_prio
;
2593 if (!rt_prio(p
->prio
))
2594 p
->sched_class
= &fair_sched_class
;
2597 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2599 set_task_cpu(p
, cpu
);
2601 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2602 if (likely(sched_info_on()))
2603 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2605 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2608 #ifdef CONFIG_PREEMPT
2609 /* Want to start with kernel preemption disabled. */
2610 task_thread_info(p
)->preempt_count
= 1;
2612 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2618 * wake_up_new_task - wake up a newly created task for the first time.
2620 * This function will do some initial scheduler statistics housekeeping
2621 * that must be done for every newly created context, then puts the task
2622 * on the runqueue and wakes it.
2624 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2626 unsigned long flags
;
2629 rq
= task_rq_lock(p
, &flags
);
2630 BUG_ON(p
->state
!= TASK_RUNNING
);
2631 update_rq_clock(rq
);
2633 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2634 activate_task(rq
, p
, 0);
2637 * Let the scheduling class do new task startup
2638 * management (if any):
2640 p
->sched_class
->task_new(rq
, p
);
2643 trace_sched_wakeup_new(rq
, p
, 1);
2644 check_preempt_curr(rq
, p
, WF_FORK
);
2646 if (p
->sched_class
->task_wake_up
)
2647 p
->sched_class
->task_wake_up(rq
, p
);
2649 task_rq_unlock(rq
, &flags
);
2652 #ifdef CONFIG_PREEMPT_NOTIFIERS
2655 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2656 * @notifier: notifier struct to register
2658 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2660 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2662 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2665 * preempt_notifier_unregister - no longer interested in preemption notifications
2666 * @notifier: notifier struct to unregister
2668 * This is safe to call from within a preemption notifier.
2670 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2672 hlist_del(¬ifier
->link
);
2674 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2676 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2678 struct preempt_notifier
*notifier
;
2679 struct hlist_node
*node
;
2681 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2682 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2686 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2687 struct task_struct
*next
)
2689 struct preempt_notifier
*notifier
;
2690 struct hlist_node
*node
;
2692 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2693 notifier
->ops
->sched_out(notifier
, next
);
2696 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2698 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2703 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2704 struct task_struct
*next
)
2708 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2711 * prepare_task_switch - prepare to switch tasks
2712 * @rq: the runqueue preparing to switch
2713 * @prev: the current task that is being switched out
2714 * @next: the task we are going to switch to.
2716 * This is called with the rq lock held and interrupts off. It must
2717 * be paired with a subsequent finish_task_switch after the context
2720 * prepare_task_switch sets up locking and calls architecture specific
2724 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2725 struct task_struct
*next
)
2727 fire_sched_out_preempt_notifiers(prev
, next
);
2728 prepare_lock_switch(rq
, next
);
2729 prepare_arch_switch(next
);
2733 * finish_task_switch - clean up after a task-switch
2734 * @rq: runqueue associated with task-switch
2735 * @prev: the thread we just switched away from.
2737 * finish_task_switch must be called after the context switch, paired
2738 * with a prepare_task_switch call before the context switch.
2739 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2740 * and do any other architecture-specific cleanup actions.
2742 * Note that we may have delayed dropping an mm in context_switch(). If
2743 * so, we finish that here outside of the runqueue lock. (Doing it
2744 * with the lock held can cause deadlocks; see schedule() for
2747 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2748 __releases(rq
->lock
)
2750 struct mm_struct
*mm
= rq
->prev_mm
;
2756 * A task struct has one reference for the use as "current".
2757 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2758 * schedule one last time. The schedule call will never return, and
2759 * the scheduled task must drop that reference.
2760 * The test for TASK_DEAD must occur while the runqueue locks are
2761 * still held, otherwise prev could be scheduled on another cpu, die
2762 * there before we look at prev->state, and then the reference would
2764 * Manfred Spraul <manfred@colorfullife.com>
2766 prev_state
= prev
->state
;
2767 finish_arch_switch(prev
);
2768 perf_event_task_sched_in(current
, cpu_of(rq
));
2769 finish_lock_switch(rq
, prev
);
2771 fire_sched_in_preempt_notifiers(current
);
2774 if (unlikely(prev_state
== TASK_DEAD
)) {
2776 * Remove function-return probe instances associated with this
2777 * task and put them back on the free list.
2779 kprobe_flush_task(prev
);
2780 put_task_struct(prev
);
2786 /* assumes rq->lock is held */
2787 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2789 if (prev
->sched_class
->pre_schedule
)
2790 prev
->sched_class
->pre_schedule(rq
, prev
);
2793 /* rq->lock is NOT held, but preemption is disabled */
2794 static inline void post_schedule(struct rq
*rq
)
2796 if (rq
->post_schedule
) {
2797 unsigned long flags
;
2799 spin_lock_irqsave(&rq
->lock
, flags
);
2800 if (rq
->curr
->sched_class
->post_schedule
)
2801 rq
->curr
->sched_class
->post_schedule(rq
);
2802 spin_unlock_irqrestore(&rq
->lock
, flags
);
2804 rq
->post_schedule
= 0;
2810 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2814 static inline void post_schedule(struct rq
*rq
)
2821 * schedule_tail - first thing a freshly forked thread must call.
2822 * @prev: the thread we just switched away from.
2824 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2825 __releases(rq
->lock
)
2827 struct rq
*rq
= this_rq();
2829 finish_task_switch(rq
, prev
);
2832 * FIXME: do we need to worry about rq being invalidated by the
2837 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2838 /* In this case, finish_task_switch does not reenable preemption */
2841 if (current
->set_child_tid
)
2842 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2846 * context_switch - switch to the new MM and the new
2847 * thread's register state.
2850 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2851 struct task_struct
*next
)
2853 struct mm_struct
*mm
, *oldmm
;
2855 prepare_task_switch(rq
, prev
, next
);
2856 trace_sched_switch(rq
, prev
, next
);
2858 oldmm
= prev
->active_mm
;
2860 * For paravirt, this is coupled with an exit in switch_to to
2861 * combine the page table reload and the switch backend into
2864 arch_start_context_switch(prev
);
2866 if (unlikely(!mm
)) {
2867 next
->active_mm
= oldmm
;
2868 atomic_inc(&oldmm
->mm_count
);
2869 enter_lazy_tlb(oldmm
, next
);
2871 switch_mm(oldmm
, mm
, next
);
2873 if (unlikely(!prev
->mm
)) {
2874 prev
->active_mm
= NULL
;
2875 rq
->prev_mm
= oldmm
;
2878 * Since the runqueue lock will be released by the next
2879 * task (which is an invalid locking op but in the case
2880 * of the scheduler it's an obvious special-case), so we
2881 * do an early lockdep release here:
2883 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2884 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2887 /* Here we just switch the register state and the stack. */
2888 switch_to(prev
, next
, prev
);
2892 * this_rq must be evaluated again because prev may have moved
2893 * CPUs since it called schedule(), thus the 'rq' on its stack
2894 * frame will be invalid.
2896 finish_task_switch(this_rq(), prev
);
2900 * nr_running, nr_uninterruptible and nr_context_switches:
2902 * externally visible scheduler statistics: current number of runnable
2903 * threads, current number of uninterruptible-sleeping threads, total
2904 * number of context switches performed since bootup.
2906 unsigned long nr_running(void)
2908 unsigned long i
, sum
= 0;
2910 for_each_online_cpu(i
)
2911 sum
+= cpu_rq(i
)->nr_running
;
2916 unsigned long nr_uninterruptible(void)
2918 unsigned long i
, sum
= 0;
2920 for_each_possible_cpu(i
)
2921 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2924 * Since we read the counters lockless, it might be slightly
2925 * inaccurate. Do not allow it to go below zero though:
2927 if (unlikely((long)sum
< 0))
2933 unsigned long long nr_context_switches(void)
2936 unsigned long long sum
= 0;
2938 for_each_possible_cpu(i
)
2939 sum
+= cpu_rq(i
)->nr_switches
;
2944 unsigned long nr_iowait(void)
2946 unsigned long i
, sum
= 0;
2948 for_each_possible_cpu(i
)
2949 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2954 unsigned long nr_iowait_cpu(void)
2956 struct rq
*this = this_rq();
2957 return atomic_read(&this->nr_iowait
);
2960 unsigned long this_cpu_load(void)
2962 struct rq
*this = this_rq();
2963 return this->cpu_load
[0];
2967 /* Variables and functions for calc_load */
2968 static atomic_long_t calc_load_tasks
;
2969 static unsigned long calc_load_update
;
2970 unsigned long avenrun
[3];
2971 EXPORT_SYMBOL(avenrun
);
2974 * get_avenrun - get the load average array
2975 * @loads: pointer to dest load array
2976 * @offset: offset to add
2977 * @shift: shift count to shift the result left
2979 * These values are estimates at best, so no need for locking.
2981 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2983 loads
[0] = (avenrun
[0] + offset
) << shift
;
2984 loads
[1] = (avenrun
[1] + offset
) << shift
;
2985 loads
[2] = (avenrun
[2] + offset
) << shift
;
2988 static unsigned long
2989 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2992 load
+= active
* (FIXED_1
- exp
);
2993 return load
>> FSHIFT
;
2997 * calc_load - update the avenrun load estimates 10 ticks after the
2998 * CPUs have updated calc_load_tasks.
3000 void calc_global_load(void)
3002 unsigned long upd
= calc_load_update
+ 10;
3005 if (time_before(jiffies
, upd
))
3008 active
= atomic_long_read(&calc_load_tasks
);
3009 active
= active
> 0 ? active
* FIXED_1
: 0;
3011 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3012 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3013 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3015 calc_load_update
+= LOAD_FREQ
;
3019 * Either called from update_cpu_load() or from a cpu going idle
3021 static void calc_load_account_active(struct rq
*this_rq
)
3023 long nr_active
, delta
;
3025 nr_active
= this_rq
->nr_running
;
3026 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3028 if (nr_active
!= this_rq
->calc_load_active
) {
3029 delta
= nr_active
- this_rq
->calc_load_active
;
3030 this_rq
->calc_load_active
= nr_active
;
3031 atomic_long_add(delta
, &calc_load_tasks
);
3036 * Externally visible per-cpu scheduler statistics:
3037 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3039 u64
cpu_nr_migrations(int cpu
)
3041 return cpu_rq(cpu
)->nr_migrations_in
;
3045 * Update rq->cpu_load[] statistics. This function is usually called every
3046 * scheduler tick (TICK_NSEC).
3048 static void update_cpu_load(struct rq
*this_rq
)
3050 unsigned long this_load
= this_rq
->load
.weight
;
3053 this_rq
->nr_load_updates
++;
3055 /* Update our load: */
3056 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3057 unsigned long old_load
, new_load
;
3059 /* scale is effectively 1 << i now, and >> i divides by scale */
3061 old_load
= this_rq
->cpu_load
[i
];
3062 new_load
= this_load
;
3064 * Round up the averaging division if load is increasing. This
3065 * prevents us from getting stuck on 9 if the load is 10, for
3068 if (new_load
> old_load
)
3069 new_load
+= scale
-1;
3070 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3073 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3074 this_rq
->calc_load_update
+= LOAD_FREQ
;
3075 calc_load_account_active(this_rq
);
3082 * double_rq_lock - safely lock two runqueues
3084 * Note this does not disable interrupts like task_rq_lock,
3085 * you need to do so manually before calling.
3087 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3088 __acquires(rq1
->lock
)
3089 __acquires(rq2
->lock
)
3091 BUG_ON(!irqs_disabled());
3093 spin_lock(&rq1
->lock
);
3094 __acquire(rq2
->lock
); /* Fake it out ;) */
3097 spin_lock(&rq1
->lock
);
3098 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3100 spin_lock(&rq2
->lock
);
3101 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3104 update_rq_clock(rq1
);
3105 update_rq_clock(rq2
);
3109 * double_rq_unlock - safely unlock two runqueues
3111 * Note this does not restore interrupts like task_rq_unlock,
3112 * you need to do so manually after calling.
3114 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3115 __releases(rq1
->lock
)
3116 __releases(rq2
->lock
)
3118 spin_unlock(&rq1
->lock
);
3120 spin_unlock(&rq2
->lock
);
3122 __release(rq2
->lock
);
3126 * If dest_cpu is allowed for this process, migrate the task to it.
3127 * This is accomplished by forcing the cpu_allowed mask to only
3128 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3129 * the cpu_allowed mask is restored.
3131 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3133 struct migration_req req
;
3134 unsigned long flags
;
3137 rq
= task_rq_lock(p
, &flags
);
3138 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3139 || unlikely(!cpu_active(dest_cpu
)))
3142 /* force the process onto the specified CPU */
3143 if (migrate_task(p
, dest_cpu
, &req
)) {
3144 /* Need to wait for migration thread (might exit: take ref). */
3145 struct task_struct
*mt
= rq
->migration_thread
;
3147 get_task_struct(mt
);
3148 task_rq_unlock(rq
, &flags
);
3149 wake_up_process(mt
);
3150 put_task_struct(mt
);
3151 wait_for_completion(&req
.done
);
3156 task_rq_unlock(rq
, &flags
);
3160 * sched_exec - execve() is a valuable balancing opportunity, because at
3161 * this point the task has the smallest effective memory and cache footprint.
3163 void sched_exec(void)
3165 int new_cpu
, this_cpu
= get_cpu();
3166 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3168 if (new_cpu
!= this_cpu
)
3169 sched_migrate_task(current
, new_cpu
);
3173 * pull_task - move a task from a remote runqueue to the local runqueue.
3174 * Both runqueues must be locked.
3176 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3177 struct rq
*this_rq
, int this_cpu
)
3179 deactivate_task(src_rq
, p
, 0);
3180 set_task_cpu(p
, this_cpu
);
3181 activate_task(this_rq
, p
, 0);
3182 check_preempt_curr(this_rq
, p
, 0);
3186 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3189 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3190 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3193 int tsk_cache_hot
= 0;
3195 * We do not migrate tasks that are:
3196 * 1) running (obviously), or
3197 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3198 * 3) are cache-hot on their current CPU.
3200 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3201 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3206 if (task_running(rq
, p
)) {
3207 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3212 * Aggressive migration if:
3213 * 1) task is cache cold, or
3214 * 2) too many balance attempts have failed.
3217 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3218 if (!tsk_cache_hot
||
3219 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3220 #ifdef CONFIG_SCHEDSTATS
3221 if (tsk_cache_hot
) {
3222 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3223 schedstat_inc(p
, se
.nr_forced_migrations
);
3229 if (tsk_cache_hot
) {
3230 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3236 static unsigned long
3237 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3238 unsigned long max_load_move
, struct sched_domain
*sd
,
3239 enum cpu_idle_type idle
, int *all_pinned
,
3240 int *this_best_prio
, struct rq_iterator
*iterator
)
3242 int loops
= 0, pulled
= 0, pinned
= 0;
3243 struct task_struct
*p
;
3244 long rem_load_move
= max_load_move
;
3246 if (max_load_move
== 0)
3252 * Start the load-balancing iterator:
3254 p
= iterator
->start(iterator
->arg
);
3256 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3259 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3260 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3261 p
= iterator
->next(iterator
->arg
);
3265 pull_task(busiest
, p
, this_rq
, this_cpu
);
3267 rem_load_move
-= p
->se
.load
.weight
;
3269 #ifdef CONFIG_PREEMPT
3271 * NEWIDLE balancing is a source of latency, so preemptible kernels
3272 * will stop after the first task is pulled to minimize the critical
3275 if (idle
== CPU_NEWLY_IDLE
)
3280 * We only want to steal up to the prescribed amount of weighted load.
3282 if (rem_load_move
> 0) {
3283 if (p
->prio
< *this_best_prio
)
3284 *this_best_prio
= p
->prio
;
3285 p
= iterator
->next(iterator
->arg
);
3290 * Right now, this is one of only two places pull_task() is called,
3291 * so we can safely collect pull_task() stats here rather than
3292 * inside pull_task().
3294 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3297 *all_pinned
= pinned
;
3299 return max_load_move
- rem_load_move
;
3303 * move_tasks tries to move up to max_load_move weighted load from busiest to
3304 * this_rq, as part of a balancing operation within domain "sd".
3305 * Returns 1 if successful and 0 otherwise.
3307 * Called with both runqueues locked.
3309 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3310 unsigned long max_load_move
,
3311 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3314 const struct sched_class
*class = sched_class_highest
;
3315 unsigned long total_load_moved
= 0;
3316 int this_best_prio
= this_rq
->curr
->prio
;
3320 class->load_balance(this_rq
, this_cpu
, busiest
,
3321 max_load_move
- total_load_moved
,
3322 sd
, idle
, all_pinned
, &this_best_prio
);
3323 class = class->next
;
3325 #ifdef CONFIG_PREEMPT
3327 * NEWIDLE balancing is a source of latency, so preemptible
3328 * kernels will stop after the first task is pulled to minimize
3329 * the critical section.
3331 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3334 } while (class && max_load_move
> total_load_moved
);
3336 return total_load_moved
> 0;
3340 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3341 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3342 struct rq_iterator
*iterator
)
3344 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3348 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3349 pull_task(busiest
, p
, this_rq
, this_cpu
);
3351 * Right now, this is only the second place pull_task()
3352 * is called, so we can safely collect pull_task()
3353 * stats here rather than inside pull_task().
3355 schedstat_inc(sd
, lb_gained
[idle
]);
3359 p
= iterator
->next(iterator
->arg
);
3366 * move_one_task tries to move exactly one task from busiest to this_rq, as
3367 * part of active balancing operations within "domain".
3368 * Returns 1 if successful and 0 otherwise.
3370 * Called with both runqueues locked.
3372 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3373 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3375 const struct sched_class
*class;
3377 for_each_class(class) {
3378 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3384 /********** Helpers for find_busiest_group ************************/
3386 * sd_lb_stats - Structure to store the statistics of a sched_domain
3387 * during load balancing.
3389 struct sd_lb_stats
{
3390 struct sched_group
*busiest
; /* Busiest group in this sd */
3391 struct sched_group
*this; /* Local group in this sd */
3392 unsigned long total_load
; /* Total load of all groups in sd */
3393 unsigned long total_pwr
; /* Total power of all groups in sd */
3394 unsigned long avg_load
; /* Average load across all groups in sd */
3396 /** Statistics of this group */
3397 unsigned long this_load
;
3398 unsigned long this_load_per_task
;
3399 unsigned long this_nr_running
;
3401 /* Statistics of the busiest group */
3402 unsigned long max_load
;
3403 unsigned long busiest_load_per_task
;
3404 unsigned long busiest_nr_running
;
3406 int group_imb
; /* Is there imbalance in this sd */
3407 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3408 int power_savings_balance
; /* Is powersave balance needed for this sd */
3409 struct sched_group
*group_min
; /* Least loaded group in sd */
3410 struct sched_group
*group_leader
; /* Group which relieves group_min */
3411 unsigned long min_load_per_task
; /* load_per_task in group_min */
3412 unsigned long leader_nr_running
; /* Nr running of group_leader */
3413 unsigned long min_nr_running
; /* Nr running of group_min */
3418 * sg_lb_stats - stats of a sched_group required for load_balancing
3420 struct sg_lb_stats
{
3421 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3422 unsigned long group_load
; /* Total load over the CPUs of the group */
3423 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3424 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3425 unsigned long group_capacity
;
3426 int group_imb
; /* Is there an imbalance in the group ? */
3430 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3431 * @group: The group whose first cpu is to be returned.
3433 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3435 return cpumask_first(sched_group_cpus(group
));
3439 * get_sd_load_idx - Obtain the load index for a given sched domain.
3440 * @sd: The sched_domain whose load_idx is to be obtained.
3441 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3443 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3444 enum cpu_idle_type idle
)
3450 load_idx
= sd
->busy_idx
;
3453 case CPU_NEWLY_IDLE
:
3454 load_idx
= sd
->newidle_idx
;
3457 load_idx
= sd
->idle_idx
;
3465 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3467 * init_sd_power_savings_stats - Initialize power savings statistics for
3468 * the given sched_domain, during load balancing.
3470 * @sd: Sched domain whose power-savings statistics are to be initialized.
3471 * @sds: Variable containing the statistics for sd.
3472 * @idle: Idle status of the CPU at which we're performing load-balancing.
3474 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3475 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3478 * Busy processors will not participate in power savings
3481 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3482 sds
->power_savings_balance
= 0;
3484 sds
->power_savings_balance
= 1;
3485 sds
->min_nr_running
= ULONG_MAX
;
3486 sds
->leader_nr_running
= 0;
3491 * update_sd_power_savings_stats - Update the power saving stats for a
3492 * sched_domain while performing load balancing.
3494 * @group: sched_group belonging to the sched_domain under consideration.
3495 * @sds: Variable containing the statistics of the sched_domain
3496 * @local_group: Does group contain the CPU for which we're performing
3498 * @sgs: Variable containing the statistics of the group.
3500 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3501 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3504 if (!sds
->power_savings_balance
)
3508 * If the local group is idle or completely loaded
3509 * no need to do power savings balance at this domain
3511 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3512 !sds
->this_nr_running
))
3513 sds
->power_savings_balance
= 0;
3516 * If a group is already running at full capacity or idle,
3517 * don't include that group in power savings calculations
3519 if (!sds
->power_savings_balance
||
3520 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3521 !sgs
->sum_nr_running
)
3525 * Calculate the group which has the least non-idle load.
3526 * This is the group from where we need to pick up the load
3529 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3530 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3531 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3532 sds
->group_min
= group
;
3533 sds
->min_nr_running
= sgs
->sum_nr_running
;
3534 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3535 sgs
->sum_nr_running
;
3539 * Calculate the group which is almost near its
3540 * capacity but still has some space to pick up some load
3541 * from other group and save more power
3543 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3546 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3547 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3548 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3549 sds
->group_leader
= group
;
3550 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3555 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3556 * @sds: Variable containing the statistics of the sched_domain
3557 * under consideration.
3558 * @this_cpu: Cpu at which we're currently performing load-balancing.
3559 * @imbalance: Variable to store the imbalance.
3562 * Check if we have potential to perform some power-savings balance.
3563 * If yes, set the busiest group to be the least loaded group in the
3564 * sched_domain, so that it's CPUs can be put to idle.
3566 * Returns 1 if there is potential to perform power-savings balance.
3569 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3570 int this_cpu
, unsigned long *imbalance
)
3572 if (!sds
->power_savings_balance
)
3575 if (sds
->this != sds
->group_leader
||
3576 sds
->group_leader
== sds
->group_min
)
3579 *imbalance
= sds
->min_load_per_task
;
3580 sds
->busiest
= sds
->group_min
;
3585 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3586 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3587 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3592 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3593 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3598 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3599 int this_cpu
, unsigned long *imbalance
)
3603 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3606 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3608 return SCHED_LOAD_SCALE
;
3611 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3613 return default_scale_freq_power(sd
, cpu
);
3616 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3618 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3619 unsigned long smt_gain
= sd
->smt_gain
;
3626 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3628 return default_scale_smt_power(sd
, cpu
);
3631 unsigned long scale_rt_power(int cpu
)
3633 struct rq
*rq
= cpu_rq(cpu
);
3634 u64 total
, available
;
3636 sched_avg_update(rq
);
3638 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3639 available
= total
- rq
->rt_avg
;
3641 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3642 total
= SCHED_LOAD_SCALE
;
3644 total
>>= SCHED_LOAD_SHIFT
;
3646 return div_u64(available
, total
);
3649 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3651 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3652 unsigned long power
= SCHED_LOAD_SCALE
;
3653 struct sched_group
*sdg
= sd
->groups
;
3655 if (sched_feat(ARCH_POWER
))
3656 power
*= arch_scale_freq_power(sd
, cpu
);
3658 power
*= default_scale_freq_power(sd
, cpu
);
3660 power
>>= SCHED_LOAD_SHIFT
;
3662 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3663 if (sched_feat(ARCH_POWER
))
3664 power
*= arch_scale_smt_power(sd
, cpu
);
3666 power
*= default_scale_smt_power(sd
, cpu
);
3668 power
>>= SCHED_LOAD_SHIFT
;
3671 power
*= scale_rt_power(cpu
);
3672 power
>>= SCHED_LOAD_SHIFT
;
3677 sdg
->cpu_power
= power
;
3680 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3682 struct sched_domain
*child
= sd
->child
;
3683 struct sched_group
*group
, *sdg
= sd
->groups
;
3684 unsigned long power
;
3687 update_cpu_power(sd
, cpu
);
3693 group
= child
->groups
;
3695 power
+= group
->cpu_power
;
3696 group
= group
->next
;
3697 } while (group
!= child
->groups
);
3699 sdg
->cpu_power
= power
;
3703 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3704 * @sd: The sched_domain whose statistics are to be updated.
3705 * @group: sched_group whose statistics are to be updated.
3706 * @this_cpu: Cpu for which load balance is currently performed.
3707 * @idle: Idle status of this_cpu
3708 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3709 * @sd_idle: Idle status of the sched_domain containing group.
3710 * @local_group: Does group contain this_cpu.
3711 * @cpus: Set of cpus considered for load balancing.
3712 * @balance: Should we balance.
3713 * @sgs: variable to hold the statistics for this group.
3715 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3716 struct sched_group
*group
, int this_cpu
,
3717 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3718 int local_group
, const struct cpumask
*cpus
,
3719 int *balance
, struct sg_lb_stats
*sgs
)
3721 unsigned long load
, max_cpu_load
, min_cpu_load
;
3723 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3724 unsigned long sum_avg_load_per_task
;
3725 unsigned long avg_load_per_task
;
3728 balance_cpu
= group_first_cpu(group
);
3729 if (balance_cpu
== this_cpu
)
3730 update_group_power(sd
, this_cpu
);
3733 /* Tally up the load of all CPUs in the group */
3734 sum_avg_load_per_task
= avg_load_per_task
= 0;
3736 min_cpu_load
= ~0UL;
3738 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3739 struct rq
*rq
= cpu_rq(i
);
3741 if (*sd_idle
&& rq
->nr_running
)
3744 /* Bias balancing toward cpus of our domain */
3746 if (idle_cpu(i
) && !first_idle_cpu
) {
3751 load
= target_load(i
, load_idx
);
3753 load
= source_load(i
, load_idx
);
3754 if (load
> max_cpu_load
)
3755 max_cpu_load
= load
;
3756 if (min_cpu_load
> load
)
3757 min_cpu_load
= load
;
3760 sgs
->group_load
+= load
;
3761 sgs
->sum_nr_running
+= rq
->nr_running
;
3762 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3764 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3768 * First idle cpu or the first cpu(busiest) in this sched group
3769 * is eligible for doing load balancing at this and above
3770 * domains. In the newly idle case, we will allow all the cpu's
3771 * to do the newly idle load balance.
3773 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3774 balance_cpu
!= this_cpu
&& balance
) {
3779 /* Adjust by relative CPU power of the group */
3780 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3784 * Consider the group unbalanced when the imbalance is larger
3785 * than the average weight of two tasks.
3787 * APZ: with cgroup the avg task weight can vary wildly and
3788 * might not be a suitable number - should we keep a
3789 * normalized nr_running number somewhere that negates
3792 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3795 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3798 sgs
->group_capacity
=
3799 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3803 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3804 * @sd: sched_domain whose statistics are to be updated.
3805 * @this_cpu: Cpu for which load balance is currently performed.
3806 * @idle: Idle status of this_cpu
3807 * @sd_idle: Idle status of the sched_domain containing group.
3808 * @cpus: Set of cpus considered for load balancing.
3809 * @balance: Should we balance.
3810 * @sds: variable to hold the statistics for this sched_domain.
3812 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3813 enum cpu_idle_type idle
, int *sd_idle
,
3814 const struct cpumask
*cpus
, int *balance
,
3815 struct sd_lb_stats
*sds
)
3817 struct sched_domain
*child
= sd
->child
;
3818 struct sched_group
*group
= sd
->groups
;
3819 struct sg_lb_stats sgs
;
3820 int load_idx
, prefer_sibling
= 0;
3822 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3825 init_sd_power_savings_stats(sd
, sds
, idle
);
3826 load_idx
= get_sd_load_idx(sd
, idle
);
3831 local_group
= cpumask_test_cpu(this_cpu
,
3832 sched_group_cpus(group
));
3833 memset(&sgs
, 0, sizeof(sgs
));
3834 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3835 local_group
, cpus
, balance
, &sgs
);
3837 if (local_group
&& balance
&& !(*balance
))
3840 sds
->total_load
+= sgs
.group_load
;
3841 sds
->total_pwr
+= group
->cpu_power
;
3844 * In case the child domain prefers tasks go to siblings
3845 * first, lower the group capacity to one so that we'll try
3846 * and move all the excess tasks away.
3849 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3852 sds
->this_load
= sgs
.avg_load
;
3854 sds
->this_nr_running
= sgs
.sum_nr_running
;
3855 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3856 } else if (sgs
.avg_load
> sds
->max_load
&&
3857 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3859 sds
->max_load
= sgs
.avg_load
;
3860 sds
->busiest
= group
;
3861 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3862 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3863 sds
->group_imb
= sgs
.group_imb
;
3866 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3867 group
= group
->next
;
3868 } while (group
!= sd
->groups
);
3872 * fix_small_imbalance - Calculate the minor imbalance that exists
3873 * amongst the groups of a sched_domain, during
3875 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3876 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3877 * @imbalance: Variable to store the imbalance.
3879 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3880 int this_cpu
, unsigned long *imbalance
)
3882 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3883 unsigned int imbn
= 2;
3885 if (sds
->this_nr_running
) {
3886 sds
->this_load_per_task
/= sds
->this_nr_running
;
3887 if (sds
->busiest_load_per_task
>
3888 sds
->this_load_per_task
)
3891 sds
->this_load_per_task
=
3892 cpu_avg_load_per_task(this_cpu
);
3894 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3895 sds
->busiest_load_per_task
* imbn
) {
3896 *imbalance
= sds
->busiest_load_per_task
;
3901 * OK, we don't have enough imbalance to justify moving tasks,
3902 * however we may be able to increase total CPU power used by
3906 pwr_now
+= sds
->busiest
->cpu_power
*
3907 min(sds
->busiest_load_per_task
, sds
->max_load
);
3908 pwr_now
+= sds
->this->cpu_power
*
3909 min(sds
->this_load_per_task
, sds
->this_load
);
3910 pwr_now
/= SCHED_LOAD_SCALE
;
3912 /* Amount of load we'd subtract */
3913 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3914 sds
->busiest
->cpu_power
;
3915 if (sds
->max_load
> tmp
)
3916 pwr_move
+= sds
->busiest
->cpu_power
*
3917 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3919 /* Amount of load we'd add */
3920 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3921 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3922 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3923 sds
->this->cpu_power
;
3925 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3926 sds
->this->cpu_power
;
3927 pwr_move
+= sds
->this->cpu_power
*
3928 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3929 pwr_move
/= SCHED_LOAD_SCALE
;
3931 /* Move if we gain throughput */
3932 if (pwr_move
> pwr_now
)
3933 *imbalance
= sds
->busiest_load_per_task
;
3937 * calculate_imbalance - Calculate the amount of imbalance present within the
3938 * groups of a given sched_domain during load balance.
3939 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3940 * @this_cpu: Cpu for which currently load balance is being performed.
3941 * @imbalance: The variable to store the imbalance.
3943 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3944 unsigned long *imbalance
)
3946 unsigned long max_pull
;
3948 * In the presence of smp nice balancing, certain scenarios can have
3949 * max load less than avg load(as we skip the groups at or below
3950 * its cpu_power, while calculating max_load..)
3952 if (sds
->max_load
< sds
->avg_load
) {
3954 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3957 /* Don't want to pull so many tasks that a group would go idle */
3958 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3959 sds
->max_load
- sds
->busiest_load_per_task
);
3961 /* How much load to actually move to equalise the imbalance */
3962 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3963 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3967 * if *imbalance is less than the average load per runnable task
3968 * there is no gaurantee that any tasks will be moved so we'll have
3969 * a think about bumping its value to force at least one task to be
3972 if (*imbalance
< sds
->busiest_load_per_task
)
3973 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3976 /******* find_busiest_group() helpers end here *********************/
3979 * find_busiest_group - Returns the busiest group within the sched_domain
3980 * if there is an imbalance. If there isn't an imbalance, and
3981 * the user has opted for power-savings, it returns a group whose
3982 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3983 * such a group exists.
3985 * Also calculates the amount of weighted load which should be moved
3986 * to restore balance.
3988 * @sd: The sched_domain whose busiest group is to be returned.
3989 * @this_cpu: The cpu for which load balancing is currently being performed.
3990 * @imbalance: Variable which stores amount of weighted load which should
3991 * be moved to restore balance/put a group to idle.
3992 * @idle: The idle status of this_cpu.
3993 * @sd_idle: The idleness of sd
3994 * @cpus: The set of CPUs under consideration for load-balancing.
3995 * @balance: Pointer to a variable indicating if this_cpu
3996 * is the appropriate cpu to perform load balancing at this_level.
3998 * Returns: - the busiest group if imbalance exists.
3999 * - If no imbalance and user has opted for power-savings balance,
4000 * return the least loaded group whose CPUs can be
4001 * put to idle by rebalancing its tasks onto our group.
4003 static struct sched_group
*
4004 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4005 unsigned long *imbalance
, enum cpu_idle_type idle
,
4006 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4008 struct sd_lb_stats sds
;
4010 memset(&sds
, 0, sizeof(sds
));
4013 * Compute the various statistics relavent for load balancing at
4016 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4019 /* Cases where imbalance does not exist from POV of this_cpu */
4020 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4022 * 2) There is no busy sibling group to pull from.
4023 * 3) This group is the busiest group.
4024 * 4) This group is more busy than the avg busieness at this
4026 * 5) The imbalance is within the specified limit.
4027 * 6) Any rebalance would lead to ping-pong
4029 if (balance
&& !(*balance
))
4032 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4035 if (sds
.this_load
>= sds
.max_load
)
4038 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4040 if (sds
.this_load
>= sds
.avg_load
)
4043 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4046 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4048 sds
.busiest_load_per_task
=
4049 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4052 * We're trying to get all the cpus to the average_load, so we don't
4053 * want to push ourselves above the average load, nor do we wish to
4054 * reduce the max loaded cpu below the average load, as either of these
4055 * actions would just result in more rebalancing later, and ping-pong
4056 * tasks around. Thus we look for the minimum possible imbalance.
4057 * Negative imbalances (*we* are more loaded than anyone else) will
4058 * be counted as no imbalance for these purposes -- we can't fix that
4059 * by pulling tasks to us. Be careful of negative numbers as they'll
4060 * appear as very large values with unsigned longs.
4062 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4065 /* Looks like there is an imbalance. Compute it */
4066 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4071 * There is no obvious imbalance. But check if we can do some balancing
4074 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4082 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4085 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4086 unsigned long imbalance
, const struct cpumask
*cpus
)
4088 struct rq
*busiest
= NULL
, *rq
;
4089 unsigned long max_load
= 0;
4092 for_each_cpu(i
, sched_group_cpus(group
)) {
4093 unsigned long power
= power_of(i
);
4094 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4097 if (!cpumask_test_cpu(i
, cpus
))
4101 wl
= weighted_cpuload(i
);
4104 * When comparing with imbalance, use weighted_cpuload()
4105 * which is not scaled with the cpu power.
4107 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4111 * For the load comparisons with the other cpu's, consider
4112 * the weighted_cpuload() scaled with the cpu power, so that
4113 * the load can be moved away from the cpu that is potentially
4114 * running at a lower capacity.
4116 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
4118 if (wl
> max_load
) {
4128 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4129 * so long as it is large enough.
4131 #define MAX_PINNED_INTERVAL 512
4133 /* Working cpumask for load_balance and load_balance_newidle. */
4134 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4137 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4138 * tasks if there is an imbalance.
4140 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4141 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4144 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4145 struct sched_group
*group
;
4146 unsigned long imbalance
;
4148 unsigned long flags
;
4149 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4151 cpumask_copy(cpus
, cpu_active_mask
);
4154 * When power savings policy is enabled for the parent domain, idle
4155 * sibling can pick up load irrespective of busy siblings. In this case,
4156 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4157 * portraying it as CPU_NOT_IDLE.
4159 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4160 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4163 schedstat_inc(sd
, lb_count
[idle
]);
4167 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4174 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4178 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4180 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4184 BUG_ON(busiest
== this_rq
);
4186 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4189 if (busiest
->nr_running
> 1) {
4191 * Attempt to move tasks. If find_busiest_group has found
4192 * an imbalance but busiest->nr_running <= 1, the group is
4193 * still unbalanced. ld_moved simply stays zero, so it is
4194 * correctly treated as an imbalance.
4196 local_irq_save(flags
);
4197 double_rq_lock(this_rq
, busiest
);
4198 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4199 imbalance
, sd
, idle
, &all_pinned
);
4200 double_rq_unlock(this_rq
, busiest
);
4201 local_irq_restore(flags
);
4204 * some other cpu did the load balance for us.
4206 if (ld_moved
&& this_cpu
!= smp_processor_id())
4207 resched_cpu(this_cpu
);
4209 /* All tasks on this runqueue were pinned by CPU affinity */
4210 if (unlikely(all_pinned
)) {
4211 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4212 if (!cpumask_empty(cpus
))
4219 schedstat_inc(sd
, lb_failed
[idle
]);
4220 sd
->nr_balance_failed
++;
4222 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4224 spin_lock_irqsave(&busiest
->lock
, flags
);
4226 /* don't kick the migration_thread, if the curr
4227 * task on busiest cpu can't be moved to this_cpu
4229 if (!cpumask_test_cpu(this_cpu
,
4230 &busiest
->curr
->cpus_allowed
)) {
4231 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4233 goto out_one_pinned
;
4236 if (!busiest
->active_balance
) {
4237 busiest
->active_balance
= 1;
4238 busiest
->push_cpu
= this_cpu
;
4241 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4243 wake_up_process(busiest
->migration_thread
);
4246 * We've kicked active balancing, reset the failure
4249 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4252 sd
->nr_balance_failed
= 0;
4254 if (likely(!active_balance
)) {
4255 /* We were unbalanced, so reset the balancing interval */
4256 sd
->balance_interval
= sd
->min_interval
;
4259 * If we've begun active balancing, start to back off. This
4260 * case may not be covered by the all_pinned logic if there
4261 * is only 1 task on the busy runqueue (because we don't call
4264 if (sd
->balance_interval
< sd
->max_interval
)
4265 sd
->balance_interval
*= 2;
4268 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4269 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4275 schedstat_inc(sd
, lb_balanced
[idle
]);
4277 sd
->nr_balance_failed
= 0;
4280 /* tune up the balancing interval */
4281 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4282 (sd
->balance_interval
< sd
->max_interval
))
4283 sd
->balance_interval
*= 2;
4285 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4286 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4297 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4298 * tasks if there is an imbalance.
4300 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4301 * this_rq is locked.
4304 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4306 struct sched_group
*group
;
4307 struct rq
*busiest
= NULL
;
4308 unsigned long imbalance
;
4312 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4314 cpumask_copy(cpus
, cpu_active_mask
);
4317 * When power savings policy is enabled for the parent domain, idle
4318 * sibling can pick up load irrespective of busy siblings. In this case,
4319 * let the state of idle sibling percolate up as IDLE, instead of
4320 * portraying it as CPU_NOT_IDLE.
4322 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4323 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4326 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4328 update_shares_locked(this_rq
, sd
);
4329 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4330 &sd_idle
, cpus
, NULL
);
4332 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4336 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4338 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4342 BUG_ON(busiest
== this_rq
);
4344 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4347 if (busiest
->nr_running
> 1) {
4348 /* Attempt to move tasks */
4349 double_lock_balance(this_rq
, busiest
);
4350 /* this_rq->clock is already updated */
4351 update_rq_clock(busiest
);
4352 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4353 imbalance
, sd
, CPU_NEWLY_IDLE
,
4355 double_unlock_balance(this_rq
, busiest
);
4357 if (unlikely(all_pinned
)) {
4358 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4359 if (!cpumask_empty(cpus
))
4365 int active_balance
= 0;
4367 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4368 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4369 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4372 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4375 if (sd
->nr_balance_failed
++ < 2)
4379 * The only task running in a non-idle cpu can be moved to this
4380 * cpu in an attempt to completely freeup the other CPU
4381 * package. The same method used to move task in load_balance()
4382 * have been extended for load_balance_newidle() to speedup
4383 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4385 * The package power saving logic comes from
4386 * find_busiest_group(). If there are no imbalance, then
4387 * f_b_g() will return NULL. However when sched_mc={1,2} then
4388 * f_b_g() will select a group from which a running task may be
4389 * pulled to this cpu in order to make the other package idle.
4390 * If there is no opportunity to make a package idle and if
4391 * there are no imbalance, then f_b_g() will return NULL and no
4392 * action will be taken in load_balance_newidle().
4394 * Under normal task pull operation due to imbalance, there
4395 * will be more than one task in the source run queue and
4396 * move_tasks() will succeed. ld_moved will be true and this
4397 * active balance code will not be triggered.
4400 /* Lock busiest in correct order while this_rq is held */
4401 double_lock_balance(this_rq
, busiest
);
4404 * don't kick the migration_thread, if the curr
4405 * task on busiest cpu can't be moved to this_cpu
4407 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4408 double_unlock_balance(this_rq
, busiest
);
4413 if (!busiest
->active_balance
) {
4414 busiest
->active_balance
= 1;
4415 busiest
->push_cpu
= this_cpu
;
4419 double_unlock_balance(this_rq
, busiest
);
4421 * Should not call ttwu while holding a rq->lock
4423 spin_unlock(&this_rq
->lock
);
4425 wake_up_process(busiest
->migration_thread
);
4426 spin_lock(&this_rq
->lock
);
4429 sd
->nr_balance_failed
= 0;
4431 update_shares_locked(this_rq
, sd
);
4435 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4436 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4437 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4439 sd
->nr_balance_failed
= 0;
4445 * idle_balance is called by schedule() if this_cpu is about to become
4446 * idle. Attempts to pull tasks from other CPUs.
4448 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4450 struct sched_domain
*sd
;
4451 int pulled_task
= 0;
4452 unsigned long next_balance
= jiffies
+ HZ
;
4454 this_rq
->idle_stamp
= this_rq
->clock
;
4456 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4459 for_each_domain(this_cpu
, sd
) {
4460 unsigned long interval
;
4462 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4465 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4466 /* If we've pulled tasks over stop searching: */
4467 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4470 interval
= msecs_to_jiffies(sd
->balance_interval
);
4471 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4472 next_balance
= sd
->last_balance
+ interval
;
4474 this_rq
->idle_stamp
= 0;
4478 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4480 * We are going idle. next_balance may be set based on
4481 * a busy processor. So reset next_balance.
4483 this_rq
->next_balance
= next_balance
;
4488 * active_load_balance is run by migration threads. It pushes running tasks
4489 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4490 * running on each physical CPU where possible, and avoids physical /
4491 * logical imbalances.
4493 * Called with busiest_rq locked.
4495 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4497 int target_cpu
= busiest_rq
->push_cpu
;
4498 struct sched_domain
*sd
;
4499 struct rq
*target_rq
;
4501 /* Is there any task to move? */
4502 if (busiest_rq
->nr_running
<= 1)
4505 target_rq
= cpu_rq(target_cpu
);
4508 * This condition is "impossible", if it occurs
4509 * we need to fix it. Originally reported by
4510 * Bjorn Helgaas on a 128-cpu setup.
4512 BUG_ON(busiest_rq
== target_rq
);
4514 /* move a task from busiest_rq to target_rq */
4515 double_lock_balance(busiest_rq
, target_rq
);
4516 update_rq_clock(busiest_rq
);
4517 update_rq_clock(target_rq
);
4519 /* Search for an sd spanning us and the target CPU. */
4520 for_each_domain(target_cpu
, sd
) {
4521 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4522 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4527 schedstat_inc(sd
, alb_count
);
4529 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4531 schedstat_inc(sd
, alb_pushed
);
4533 schedstat_inc(sd
, alb_failed
);
4535 double_unlock_balance(busiest_rq
, target_rq
);
4540 atomic_t load_balancer
;
4541 cpumask_var_t cpu_mask
;
4542 cpumask_var_t ilb_grp_nohz_mask
;
4543 } nohz ____cacheline_aligned
= {
4544 .load_balancer
= ATOMIC_INIT(-1),
4547 int get_nohz_load_balancer(void)
4549 return atomic_read(&nohz
.load_balancer
);
4552 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4554 * lowest_flag_domain - Return lowest sched_domain containing flag.
4555 * @cpu: The cpu whose lowest level of sched domain is to
4557 * @flag: The flag to check for the lowest sched_domain
4558 * for the given cpu.
4560 * Returns the lowest sched_domain of a cpu which contains the given flag.
4562 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4564 struct sched_domain
*sd
;
4566 for_each_domain(cpu
, sd
)
4567 if (sd
&& (sd
->flags
& flag
))
4574 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4575 * @cpu: The cpu whose domains we're iterating over.
4576 * @sd: variable holding the value of the power_savings_sd
4578 * @flag: The flag to filter the sched_domains to be iterated.
4580 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4581 * set, starting from the lowest sched_domain to the highest.
4583 #define for_each_flag_domain(cpu, sd, flag) \
4584 for (sd = lowest_flag_domain(cpu, flag); \
4585 (sd && (sd->flags & flag)); sd = sd->parent)
4588 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4589 * @ilb_group: group to be checked for semi-idleness
4591 * Returns: 1 if the group is semi-idle. 0 otherwise.
4593 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4594 * and atleast one non-idle CPU. This helper function checks if the given
4595 * sched_group is semi-idle or not.
4597 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4599 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4600 sched_group_cpus(ilb_group
));
4603 * A sched_group is semi-idle when it has atleast one busy cpu
4604 * and atleast one idle cpu.
4606 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4609 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4615 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4616 * @cpu: The cpu which is nominating a new idle_load_balancer.
4618 * Returns: Returns the id of the idle load balancer if it exists,
4619 * Else, returns >= nr_cpu_ids.
4621 * This algorithm picks the idle load balancer such that it belongs to a
4622 * semi-idle powersavings sched_domain. The idea is to try and avoid
4623 * completely idle packages/cores just for the purpose of idle load balancing
4624 * when there are other idle cpu's which are better suited for that job.
4626 static int find_new_ilb(int cpu
)
4628 struct sched_domain
*sd
;
4629 struct sched_group
*ilb_group
;
4632 * Have idle load balancer selection from semi-idle packages only
4633 * when power-aware load balancing is enabled
4635 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4639 * Optimize for the case when we have no idle CPUs or only one
4640 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4642 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4645 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4646 ilb_group
= sd
->groups
;
4649 if (is_semi_idle_group(ilb_group
))
4650 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4652 ilb_group
= ilb_group
->next
;
4654 } while (ilb_group
!= sd
->groups
);
4658 return cpumask_first(nohz
.cpu_mask
);
4660 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4661 static inline int find_new_ilb(int call_cpu
)
4663 return cpumask_first(nohz
.cpu_mask
);
4668 * This routine will try to nominate the ilb (idle load balancing)
4669 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4670 * load balancing on behalf of all those cpus. If all the cpus in the system
4671 * go into this tickless mode, then there will be no ilb owner (as there is
4672 * no need for one) and all the cpus will sleep till the next wakeup event
4675 * For the ilb owner, tick is not stopped. And this tick will be used
4676 * for idle load balancing. ilb owner will still be part of
4679 * While stopping the tick, this cpu will become the ilb owner if there
4680 * is no other owner. And will be the owner till that cpu becomes busy
4681 * or if all cpus in the system stop their ticks at which point
4682 * there is no need for ilb owner.
4684 * When the ilb owner becomes busy, it nominates another owner, during the
4685 * next busy scheduler_tick()
4687 int select_nohz_load_balancer(int stop_tick
)
4689 int cpu
= smp_processor_id();
4692 cpu_rq(cpu
)->in_nohz_recently
= 1;
4694 if (!cpu_active(cpu
)) {
4695 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4699 * If we are going offline and still the leader,
4702 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4708 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4710 /* time for ilb owner also to sleep */
4711 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
4712 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4713 atomic_set(&nohz
.load_balancer
, -1);
4717 if (atomic_read(&nohz
.load_balancer
) == -1) {
4718 /* make me the ilb owner */
4719 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4721 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4724 if (!(sched_smt_power_savings
||
4725 sched_mc_power_savings
))
4728 * Check to see if there is a more power-efficient
4731 new_ilb
= find_new_ilb(cpu
);
4732 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4733 atomic_set(&nohz
.load_balancer
, -1);
4734 resched_cpu(new_ilb
);
4740 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4743 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4745 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4746 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4753 static DEFINE_SPINLOCK(balancing
);
4756 * It checks each scheduling domain to see if it is due to be balanced,
4757 * and initiates a balancing operation if so.
4759 * Balancing parameters are set up in arch_init_sched_domains.
4761 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4764 struct rq
*rq
= cpu_rq(cpu
);
4765 unsigned long interval
;
4766 struct sched_domain
*sd
;
4767 /* Earliest time when we have to do rebalance again */
4768 unsigned long next_balance
= jiffies
+ 60*HZ
;
4769 int update_next_balance
= 0;
4772 for_each_domain(cpu
, sd
) {
4773 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4776 interval
= sd
->balance_interval
;
4777 if (idle
!= CPU_IDLE
)
4778 interval
*= sd
->busy_factor
;
4780 /* scale ms to jiffies */
4781 interval
= msecs_to_jiffies(interval
);
4782 if (unlikely(!interval
))
4784 if (interval
> HZ
*NR_CPUS
/10)
4785 interval
= HZ
*NR_CPUS
/10;
4787 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4789 if (need_serialize
) {
4790 if (!spin_trylock(&balancing
))
4794 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4795 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4797 * We've pulled tasks over so either we're no
4798 * longer idle, or one of our SMT siblings is
4801 idle
= CPU_NOT_IDLE
;
4803 sd
->last_balance
= jiffies
;
4806 spin_unlock(&balancing
);
4808 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4809 next_balance
= sd
->last_balance
+ interval
;
4810 update_next_balance
= 1;
4814 * Stop the load balance at this level. There is another
4815 * CPU in our sched group which is doing load balancing more
4823 * next_balance will be updated only when there is a need.
4824 * When the cpu is attached to null domain for ex, it will not be
4827 if (likely(update_next_balance
))
4828 rq
->next_balance
= next_balance
;
4832 * run_rebalance_domains is triggered when needed from the scheduler tick.
4833 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4834 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4836 static void run_rebalance_domains(struct softirq_action
*h
)
4838 int this_cpu
= smp_processor_id();
4839 struct rq
*this_rq
= cpu_rq(this_cpu
);
4840 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4841 CPU_IDLE
: CPU_NOT_IDLE
;
4843 rebalance_domains(this_cpu
, idle
);
4847 * If this cpu is the owner for idle load balancing, then do the
4848 * balancing on behalf of the other idle cpus whose ticks are
4851 if (this_rq
->idle_at_tick
&&
4852 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4856 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4857 if (balance_cpu
== this_cpu
)
4861 * If this cpu gets work to do, stop the load balancing
4862 * work being done for other cpus. Next load
4863 * balancing owner will pick it up.
4868 rebalance_domains(balance_cpu
, CPU_IDLE
);
4870 rq
= cpu_rq(balance_cpu
);
4871 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4872 this_rq
->next_balance
= rq
->next_balance
;
4878 static inline int on_null_domain(int cpu
)
4880 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4884 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4886 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4887 * idle load balancing owner or decide to stop the periodic load balancing,
4888 * if the whole system is idle.
4890 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4894 * If we were in the nohz mode recently and busy at the current
4895 * scheduler tick, then check if we need to nominate new idle
4898 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4899 rq
->in_nohz_recently
= 0;
4901 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4902 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4903 atomic_set(&nohz
.load_balancer
, -1);
4906 if (atomic_read(&nohz
.load_balancer
) == -1) {
4907 int ilb
= find_new_ilb(cpu
);
4909 if (ilb
< nr_cpu_ids
)
4915 * If this cpu is idle and doing idle load balancing for all the
4916 * cpus with ticks stopped, is it time for that to stop?
4918 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4919 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4925 * If this cpu is idle and the idle load balancing is done by
4926 * someone else, then no need raise the SCHED_SOFTIRQ
4928 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4929 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4932 /* Don't need to rebalance while attached to NULL domain */
4933 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4934 likely(!on_null_domain(cpu
)))
4935 raise_softirq(SCHED_SOFTIRQ
);
4938 #else /* CONFIG_SMP */
4941 * on UP we do not need to balance between CPUs:
4943 static inline void idle_balance(int cpu
, struct rq
*rq
)
4949 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4951 EXPORT_PER_CPU_SYMBOL(kstat
);
4954 * Return any ns on the sched_clock that have not yet been accounted in
4955 * @p in case that task is currently running.
4957 * Called with task_rq_lock() held on @rq.
4959 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4963 if (task_current(rq
, p
)) {
4964 update_rq_clock(rq
);
4965 ns
= rq
->clock
- p
->se
.exec_start
;
4973 unsigned long long task_delta_exec(struct task_struct
*p
)
4975 unsigned long flags
;
4979 rq
= task_rq_lock(p
, &flags
);
4980 ns
= do_task_delta_exec(p
, rq
);
4981 task_rq_unlock(rq
, &flags
);
4987 * Return accounted runtime for the task.
4988 * In case the task is currently running, return the runtime plus current's
4989 * pending runtime that have not been accounted yet.
4991 unsigned long long task_sched_runtime(struct task_struct
*p
)
4993 unsigned long flags
;
4997 rq
= task_rq_lock(p
, &flags
);
4998 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4999 task_rq_unlock(rq
, &flags
);
5005 * Return sum_exec_runtime for the thread group.
5006 * In case the task is currently running, return the sum plus current's
5007 * pending runtime that have not been accounted yet.
5009 * Note that the thread group might have other running tasks as well,
5010 * so the return value not includes other pending runtime that other
5011 * running tasks might have.
5013 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5015 struct task_cputime totals
;
5016 unsigned long flags
;
5020 rq
= task_rq_lock(p
, &flags
);
5021 thread_group_cputime(p
, &totals
);
5022 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5023 task_rq_unlock(rq
, &flags
);
5029 * Account user cpu time to a process.
5030 * @p: the process that the cpu time gets accounted to
5031 * @cputime: the cpu time spent in user space since the last update
5032 * @cputime_scaled: cputime scaled by cpu frequency
5034 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5035 cputime_t cputime_scaled
)
5037 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5040 /* Add user time to process. */
5041 p
->utime
= cputime_add(p
->utime
, cputime
);
5042 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5043 account_group_user_time(p
, cputime
);
5045 /* Add user time to cpustat. */
5046 tmp
= cputime_to_cputime64(cputime
);
5047 if (TASK_NICE(p
) > 0)
5048 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5050 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5052 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5053 /* Account for user time used */
5054 acct_update_integrals(p
);
5058 * Account guest cpu time to a process.
5059 * @p: the process that the cpu time gets accounted to
5060 * @cputime: the cpu time spent in virtual machine since the last update
5061 * @cputime_scaled: cputime scaled by cpu frequency
5063 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5064 cputime_t cputime_scaled
)
5067 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5069 tmp
= cputime_to_cputime64(cputime
);
5071 /* Add guest time to process. */
5072 p
->utime
= cputime_add(p
->utime
, cputime
);
5073 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5074 account_group_user_time(p
, cputime
);
5075 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5077 /* Add guest time to cpustat. */
5078 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5079 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5083 * Account system cpu time to a process.
5084 * @p: the process that the cpu time gets accounted to
5085 * @hardirq_offset: the offset to subtract from hardirq_count()
5086 * @cputime: the cpu time spent in kernel space since the last update
5087 * @cputime_scaled: cputime scaled by cpu frequency
5089 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5090 cputime_t cputime
, cputime_t cputime_scaled
)
5092 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5095 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5096 account_guest_time(p
, cputime
, cputime_scaled
);
5100 /* Add system time to process. */
5101 p
->stime
= cputime_add(p
->stime
, cputime
);
5102 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5103 account_group_system_time(p
, cputime
);
5105 /* Add system time to cpustat. */
5106 tmp
= cputime_to_cputime64(cputime
);
5107 if (hardirq_count() - hardirq_offset
)
5108 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5109 else if (softirq_count())
5110 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5112 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5114 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5116 /* Account for system time used */
5117 acct_update_integrals(p
);
5121 * Account for involuntary wait time.
5122 * @steal: the cpu time spent in involuntary wait
5124 void account_steal_time(cputime_t cputime
)
5126 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5127 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5129 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5133 * Account for idle time.
5134 * @cputime: the cpu time spent in idle wait
5136 void account_idle_time(cputime_t cputime
)
5138 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5139 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5140 struct rq
*rq
= this_rq();
5142 if (atomic_read(&rq
->nr_iowait
) > 0)
5143 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5145 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5148 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5151 * Account a single tick of cpu time.
5152 * @p: the process that the cpu time gets accounted to
5153 * @user_tick: indicates if the tick is a user or a system tick
5155 void account_process_tick(struct task_struct
*p
, int user_tick
)
5157 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5158 struct rq
*rq
= this_rq();
5161 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5162 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5163 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5166 account_idle_time(cputime_one_jiffy
);
5170 * Account multiple ticks of steal time.
5171 * @p: the process from which the cpu time has been stolen
5172 * @ticks: number of stolen ticks
5174 void account_steal_ticks(unsigned long ticks
)
5176 account_steal_time(jiffies_to_cputime(ticks
));
5180 * Account multiple ticks of idle time.
5181 * @ticks: number of stolen ticks
5183 void account_idle_ticks(unsigned long ticks
)
5185 account_idle_time(jiffies_to_cputime(ticks
));
5191 * Use precise platform statistics if available:
5193 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5194 cputime_t
task_utime(struct task_struct
*p
)
5199 cputime_t
task_stime(struct task_struct
*p
)
5204 cputime_t
task_utime(struct task_struct
*p
)
5206 clock_t utime
= cputime_to_clock_t(p
->utime
),
5207 total
= utime
+ cputime_to_clock_t(p
->stime
);
5211 * Use CFS's precise accounting:
5213 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5217 do_div(temp
, total
);
5219 utime
= (clock_t)temp
;
5221 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5222 return p
->prev_utime
;
5225 cputime_t
task_stime(struct task_struct
*p
)
5230 * Use CFS's precise accounting. (we subtract utime from
5231 * the total, to make sure the total observed by userspace
5232 * grows monotonically - apps rely on that):
5234 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5235 cputime_to_clock_t(task_utime(p
));
5238 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5240 return p
->prev_stime
;
5244 inline cputime_t
task_gtime(struct task_struct
*p
)
5250 * This function gets called by the timer code, with HZ frequency.
5251 * We call it with interrupts disabled.
5253 * It also gets called by the fork code, when changing the parent's
5256 void scheduler_tick(void)
5258 int cpu
= smp_processor_id();
5259 struct rq
*rq
= cpu_rq(cpu
);
5260 struct task_struct
*curr
= rq
->curr
;
5264 spin_lock(&rq
->lock
);
5265 update_rq_clock(rq
);
5266 update_cpu_load(rq
);
5267 curr
->sched_class
->task_tick(rq
, curr
, 0);
5268 spin_unlock(&rq
->lock
);
5270 perf_event_task_tick(curr
, cpu
);
5273 rq
->idle_at_tick
= idle_cpu(cpu
);
5274 trigger_load_balance(rq
, cpu
);
5278 notrace
unsigned long get_parent_ip(unsigned long addr
)
5280 if (in_lock_functions(addr
)) {
5281 addr
= CALLER_ADDR2
;
5282 if (in_lock_functions(addr
))
5283 addr
= CALLER_ADDR3
;
5288 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5289 defined(CONFIG_PREEMPT_TRACER))
5291 void __kprobes
add_preempt_count(int val
)
5293 #ifdef CONFIG_DEBUG_PREEMPT
5297 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5300 preempt_count() += val
;
5301 #ifdef CONFIG_DEBUG_PREEMPT
5303 * Spinlock count overflowing soon?
5305 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5308 if (preempt_count() == val
)
5309 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5311 EXPORT_SYMBOL(add_preempt_count
);
5313 void __kprobes
sub_preempt_count(int val
)
5315 #ifdef CONFIG_DEBUG_PREEMPT
5319 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5322 * Is the spinlock portion underflowing?
5324 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5325 !(preempt_count() & PREEMPT_MASK
)))
5329 if (preempt_count() == val
)
5330 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5331 preempt_count() -= val
;
5333 EXPORT_SYMBOL(sub_preempt_count
);
5338 * Print scheduling while atomic bug:
5340 static noinline
void __schedule_bug(struct task_struct
*prev
)
5342 struct pt_regs
*regs
= get_irq_regs();
5344 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5345 prev
->comm
, prev
->pid
, preempt_count());
5347 debug_show_held_locks(prev
);
5349 if (irqs_disabled())
5350 print_irqtrace_events(prev
);
5359 * Various schedule()-time debugging checks and statistics:
5361 static inline void schedule_debug(struct task_struct
*prev
)
5364 * Test if we are atomic. Since do_exit() needs to call into
5365 * schedule() atomically, we ignore that path for now.
5366 * Otherwise, whine if we are scheduling when we should not be.
5368 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5369 __schedule_bug(prev
);
5371 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5373 schedstat_inc(this_rq(), sched_count
);
5374 #ifdef CONFIG_SCHEDSTATS
5375 if (unlikely(prev
->lock_depth
>= 0)) {
5376 schedstat_inc(this_rq(), bkl_count
);
5377 schedstat_inc(prev
, sched_info
.bkl_count
);
5382 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5384 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5386 update_avg(&p
->se
.avg_running
, runtime
);
5388 if (p
->state
== TASK_RUNNING
) {
5390 * In order to avoid avg_overlap growing stale when we are
5391 * indeed overlapping and hence not getting put to sleep, grow
5392 * the avg_overlap on preemption.
5394 * We use the average preemption runtime because that
5395 * correlates to the amount of cache footprint a task can
5398 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5399 update_avg(&p
->se
.avg_overlap
, runtime
);
5401 update_avg(&p
->se
.avg_running
, 0);
5403 p
->sched_class
->put_prev_task(rq
, p
);
5407 * Pick up the highest-prio task:
5409 static inline struct task_struct
*
5410 pick_next_task(struct rq
*rq
)
5412 const struct sched_class
*class;
5413 struct task_struct
*p
;
5416 * Optimization: we know that if all tasks are in
5417 * the fair class we can call that function directly:
5419 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5420 p
= fair_sched_class
.pick_next_task(rq
);
5425 class = sched_class_highest
;
5427 p
= class->pick_next_task(rq
);
5431 * Will never be NULL as the idle class always
5432 * returns a non-NULL p:
5434 class = class->next
;
5439 * schedule() is the main scheduler function.
5441 asmlinkage
void __sched
schedule(void)
5443 struct task_struct
*prev
, *next
;
5444 unsigned long *switch_count
;
5450 cpu
= smp_processor_id();
5454 switch_count
= &prev
->nivcsw
;
5456 release_kernel_lock(prev
);
5457 need_resched_nonpreemptible
:
5459 schedule_debug(prev
);
5461 if (sched_feat(HRTICK
))
5464 spin_lock_irq(&rq
->lock
);
5465 update_rq_clock(rq
);
5466 clear_tsk_need_resched(prev
);
5468 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5469 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5470 prev
->state
= TASK_RUNNING
;
5472 deactivate_task(rq
, prev
, 1);
5473 switch_count
= &prev
->nvcsw
;
5476 pre_schedule(rq
, prev
);
5478 if (unlikely(!rq
->nr_running
))
5479 idle_balance(cpu
, rq
);
5481 put_prev_task(rq
, prev
);
5482 next
= pick_next_task(rq
);
5484 if (likely(prev
!= next
)) {
5485 sched_info_switch(prev
, next
);
5486 perf_event_task_sched_out(prev
, next
, cpu
);
5492 context_switch(rq
, prev
, next
); /* unlocks the rq */
5494 * the context switch might have flipped the stack from under
5495 * us, hence refresh the local variables.
5497 cpu
= smp_processor_id();
5500 spin_unlock_irq(&rq
->lock
);
5504 if (unlikely(reacquire_kernel_lock(current
) < 0))
5505 goto need_resched_nonpreemptible
;
5507 preempt_enable_no_resched();
5511 EXPORT_SYMBOL(schedule
);
5515 * Look out! "owner" is an entirely speculative pointer
5516 * access and not reliable.
5518 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5523 if (!sched_feat(OWNER_SPIN
))
5526 #ifdef CONFIG_DEBUG_PAGEALLOC
5528 * Need to access the cpu field knowing that
5529 * DEBUG_PAGEALLOC could have unmapped it if
5530 * the mutex owner just released it and exited.
5532 if (probe_kernel_address(&owner
->cpu
, cpu
))
5539 * Even if the access succeeded (likely case),
5540 * the cpu field may no longer be valid.
5542 if (cpu
>= nr_cpumask_bits
)
5546 * We need to validate that we can do a
5547 * get_cpu() and that we have the percpu area.
5549 if (!cpu_online(cpu
))
5556 * Owner changed, break to re-assess state.
5558 if (lock
->owner
!= owner
)
5562 * Is that owner really running on that cpu?
5564 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5574 #ifdef CONFIG_PREEMPT
5576 * this is the entry point to schedule() from in-kernel preemption
5577 * off of preempt_enable. Kernel preemptions off return from interrupt
5578 * occur there and call schedule directly.
5580 asmlinkage
void __sched
preempt_schedule(void)
5582 struct thread_info
*ti
= current_thread_info();
5585 * If there is a non-zero preempt_count or interrupts are disabled,
5586 * we do not want to preempt the current task. Just return..
5588 if (likely(ti
->preempt_count
|| irqs_disabled()))
5592 add_preempt_count(PREEMPT_ACTIVE
);
5594 sub_preempt_count(PREEMPT_ACTIVE
);
5597 * Check again in case we missed a preemption opportunity
5598 * between schedule and now.
5601 } while (need_resched());
5603 EXPORT_SYMBOL(preempt_schedule
);
5606 * this is the entry point to schedule() from kernel preemption
5607 * off of irq context.
5608 * Note, that this is called and return with irqs disabled. This will
5609 * protect us against recursive calling from irq.
5611 asmlinkage
void __sched
preempt_schedule_irq(void)
5613 struct thread_info
*ti
= current_thread_info();
5615 /* Catch callers which need to be fixed */
5616 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5619 add_preempt_count(PREEMPT_ACTIVE
);
5622 local_irq_disable();
5623 sub_preempt_count(PREEMPT_ACTIVE
);
5626 * Check again in case we missed a preemption opportunity
5627 * between schedule and now.
5630 } while (need_resched());
5633 #endif /* CONFIG_PREEMPT */
5635 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5638 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5640 EXPORT_SYMBOL(default_wake_function
);
5643 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5644 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5645 * number) then we wake all the non-exclusive tasks and one exclusive task.
5647 * There are circumstances in which we can try to wake a task which has already
5648 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5649 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5651 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5652 int nr_exclusive
, int wake_flags
, void *key
)
5654 wait_queue_t
*curr
, *next
;
5656 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5657 unsigned flags
= curr
->flags
;
5659 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5660 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5666 * __wake_up - wake up threads blocked on a waitqueue.
5668 * @mode: which threads
5669 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5670 * @key: is directly passed to the wakeup function
5672 * It may be assumed that this function implies a write memory barrier before
5673 * changing the task state if and only if any tasks are woken up.
5675 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5676 int nr_exclusive
, void *key
)
5678 unsigned long flags
;
5680 spin_lock_irqsave(&q
->lock
, flags
);
5681 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5682 spin_unlock_irqrestore(&q
->lock
, flags
);
5684 EXPORT_SYMBOL(__wake_up
);
5687 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5689 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5691 __wake_up_common(q
, mode
, 1, 0, NULL
);
5694 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5696 __wake_up_common(q
, mode
, 1, 0, key
);
5700 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5702 * @mode: which threads
5703 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5704 * @key: opaque value to be passed to wakeup targets
5706 * The sync wakeup differs that the waker knows that it will schedule
5707 * away soon, so while the target thread will be woken up, it will not
5708 * be migrated to another CPU - ie. the two threads are 'synchronized'
5709 * with each other. This can prevent needless bouncing between CPUs.
5711 * On UP it can prevent extra preemption.
5713 * It may be assumed that this function implies a write memory barrier before
5714 * changing the task state if and only if any tasks are woken up.
5716 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5717 int nr_exclusive
, void *key
)
5719 unsigned long flags
;
5720 int wake_flags
= WF_SYNC
;
5725 if (unlikely(!nr_exclusive
))
5728 spin_lock_irqsave(&q
->lock
, flags
);
5729 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5730 spin_unlock_irqrestore(&q
->lock
, flags
);
5732 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5735 * __wake_up_sync - see __wake_up_sync_key()
5737 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5739 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5741 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5744 * complete: - signals a single thread waiting on this completion
5745 * @x: holds the state of this particular completion
5747 * This will wake up a single thread waiting on this completion. Threads will be
5748 * awakened in the same order in which they were queued.
5750 * See also complete_all(), wait_for_completion() and related routines.
5752 * It may be assumed that this function implies a write memory barrier before
5753 * changing the task state if and only if any tasks are woken up.
5755 void complete(struct completion
*x
)
5757 unsigned long flags
;
5759 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5761 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5762 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5764 EXPORT_SYMBOL(complete
);
5767 * complete_all: - signals all threads waiting on this completion
5768 * @x: holds the state of this particular completion
5770 * This will wake up all threads waiting on this particular completion event.
5772 * It may be assumed that this function implies a write memory barrier before
5773 * changing the task state if and only if any tasks are woken up.
5775 void complete_all(struct completion
*x
)
5777 unsigned long flags
;
5779 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5780 x
->done
+= UINT_MAX
/2;
5781 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5782 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5784 EXPORT_SYMBOL(complete_all
);
5786 static inline long __sched
5787 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5790 DECLARE_WAITQUEUE(wait
, current
);
5792 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5793 __add_wait_queue_tail(&x
->wait
, &wait
);
5795 if (signal_pending_state(state
, current
)) {
5796 timeout
= -ERESTARTSYS
;
5799 __set_current_state(state
);
5800 spin_unlock_irq(&x
->wait
.lock
);
5801 timeout
= schedule_timeout(timeout
);
5802 spin_lock_irq(&x
->wait
.lock
);
5803 } while (!x
->done
&& timeout
);
5804 __remove_wait_queue(&x
->wait
, &wait
);
5809 return timeout
?: 1;
5813 wait_for_common(struct completion
*x
, long timeout
, int state
)
5817 spin_lock_irq(&x
->wait
.lock
);
5818 timeout
= do_wait_for_common(x
, timeout
, state
);
5819 spin_unlock_irq(&x
->wait
.lock
);
5824 * wait_for_completion: - waits for completion of a task
5825 * @x: holds the state of this particular completion
5827 * This waits to be signaled for completion of a specific task. It is NOT
5828 * interruptible and there is no timeout.
5830 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5831 * and interrupt capability. Also see complete().
5833 void __sched
wait_for_completion(struct completion
*x
)
5835 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5837 EXPORT_SYMBOL(wait_for_completion
);
5840 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5841 * @x: holds the state of this particular completion
5842 * @timeout: timeout value in jiffies
5844 * This waits for either a completion of a specific task to be signaled or for a
5845 * specified timeout to expire. The timeout is in jiffies. It is not
5848 unsigned long __sched
5849 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5851 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5853 EXPORT_SYMBOL(wait_for_completion_timeout
);
5856 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5857 * @x: holds the state of this particular completion
5859 * This waits for completion of a specific task to be signaled. It is
5862 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5864 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5865 if (t
== -ERESTARTSYS
)
5869 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5872 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5873 * @x: holds the state of this particular completion
5874 * @timeout: timeout value in jiffies
5876 * This waits for either a completion of a specific task to be signaled or for a
5877 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5879 unsigned long __sched
5880 wait_for_completion_interruptible_timeout(struct completion
*x
,
5881 unsigned long timeout
)
5883 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5885 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5888 * wait_for_completion_killable: - waits for completion of a task (killable)
5889 * @x: holds the state of this particular completion
5891 * This waits to be signaled for completion of a specific task. It can be
5892 * interrupted by a kill signal.
5894 int __sched
wait_for_completion_killable(struct completion
*x
)
5896 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5897 if (t
== -ERESTARTSYS
)
5901 EXPORT_SYMBOL(wait_for_completion_killable
);
5904 * try_wait_for_completion - try to decrement a completion without blocking
5905 * @x: completion structure
5907 * Returns: 0 if a decrement cannot be done without blocking
5908 * 1 if a decrement succeeded.
5910 * If a completion is being used as a counting completion,
5911 * attempt to decrement the counter without blocking. This
5912 * enables us to avoid waiting if the resource the completion
5913 * is protecting is not available.
5915 bool try_wait_for_completion(struct completion
*x
)
5919 spin_lock_irq(&x
->wait
.lock
);
5924 spin_unlock_irq(&x
->wait
.lock
);
5927 EXPORT_SYMBOL(try_wait_for_completion
);
5930 * completion_done - Test to see if a completion has any waiters
5931 * @x: completion structure
5933 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5934 * 1 if there are no waiters.
5937 bool completion_done(struct completion
*x
)
5941 spin_lock_irq(&x
->wait
.lock
);
5944 spin_unlock_irq(&x
->wait
.lock
);
5947 EXPORT_SYMBOL(completion_done
);
5950 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5952 unsigned long flags
;
5955 init_waitqueue_entry(&wait
, current
);
5957 __set_current_state(state
);
5959 spin_lock_irqsave(&q
->lock
, flags
);
5960 __add_wait_queue(q
, &wait
);
5961 spin_unlock(&q
->lock
);
5962 timeout
= schedule_timeout(timeout
);
5963 spin_lock_irq(&q
->lock
);
5964 __remove_wait_queue(q
, &wait
);
5965 spin_unlock_irqrestore(&q
->lock
, flags
);
5970 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5972 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5974 EXPORT_SYMBOL(interruptible_sleep_on
);
5977 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5979 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5981 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5983 void __sched
sleep_on(wait_queue_head_t
*q
)
5985 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5987 EXPORT_SYMBOL(sleep_on
);
5989 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5991 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5993 EXPORT_SYMBOL(sleep_on_timeout
);
5995 #ifdef CONFIG_RT_MUTEXES
5998 * rt_mutex_setprio - set the current priority of a task
6000 * @prio: prio value (kernel-internal form)
6002 * This function changes the 'effective' priority of a task. It does
6003 * not touch ->normal_prio like __setscheduler().
6005 * Used by the rt_mutex code to implement priority inheritance logic.
6007 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6009 unsigned long flags
;
6010 int oldprio
, on_rq
, running
;
6012 const struct sched_class
*prev_class
;
6014 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6016 rq
= task_rq_lock(p
, &flags
);
6017 update_rq_clock(rq
);
6020 prev_class
= p
->sched_class
;
6021 on_rq
= p
->se
.on_rq
;
6022 running
= task_current(rq
, p
);
6024 dequeue_task(rq
, p
, 0);
6026 p
->sched_class
->put_prev_task(rq
, p
);
6029 p
->sched_class
= &rt_sched_class
;
6031 p
->sched_class
= &fair_sched_class
;
6036 p
->sched_class
->set_curr_task(rq
);
6038 enqueue_task(rq
, p
, 0);
6040 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6042 task_rq_unlock(rq
, &flags
);
6047 void set_user_nice(struct task_struct
*p
, long nice
)
6049 int old_prio
, delta
, on_rq
;
6050 unsigned long flags
;
6053 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6056 * We have to be careful, if called from sys_setpriority(),
6057 * the task might be in the middle of scheduling on another CPU.
6059 rq
= task_rq_lock(p
, &flags
);
6060 update_rq_clock(rq
);
6062 * The RT priorities are set via sched_setscheduler(), but we still
6063 * allow the 'normal' nice value to be set - but as expected
6064 * it wont have any effect on scheduling until the task is
6065 * SCHED_FIFO/SCHED_RR:
6067 if (task_has_rt_policy(p
)) {
6068 p
->static_prio
= NICE_TO_PRIO(nice
);
6071 on_rq
= p
->se
.on_rq
;
6073 dequeue_task(rq
, p
, 0);
6075 p
->static_prio
= NICE_TO_PRIO(nice
);
6078 p
->prio
= effective_prio(p
);
6079 delta
= p
->prio
- old_prio
;
6082 enqueue_task(rq
, p
, 0);
6084 * If the task increased its priority or is running and
6085 * lowered its priority, then reschedule its CPU:
6087 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6088 resched_task(rq
->curr
);
6091 task_rq_unlock(rq
, &flags
);
6093 EXPORT_SYMBOL(set_user_nice
);
6096 * can_nice - check if a task can reduce its nice value
6100 int can_nice(const struct task_struct
*p
, const int nice
)
6102 /* convert nice value [19,-20] to rlimit style value [1,40] */
6103 int nice_rlim
= 20 - nice
;
6105 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6106 capable(CAP_SYS_NICE
));
6109 #ifdef __ARCH_WANT_SYS_NICE
6112 * sys_nice - change the priority of the current process.
6113 * @increment: priority increment
6115 * sys_setpriority is a more generic, but much slower function that
6116 * does similar things.
6118 SYSCALL_DEFINE1(nice
, int, increment
)
6123 * Setpriority might change our priority at the same moment.
6124 * We don't have to worry. Conceptually one call occurs first
6125 * and we have a single winner.
6127 if (increment
< -40)
6132 nice
= TASK_NICE(current
) + increment
;
6138 if (increment
< 0 && !can_nice(current
, nice
))
6141 retval
= security_task_setnice(current
, nice
);
6145 set_user_nice(current
, nice
);
6152 * task_prio - return the priority value of a given task.
6153 * @p: the task in question.
6155 * This is the priority value as seen by users in /proc.
6156 * RT tasks are offset by -200. Normal tasks are centered
6157 * around 0, value goes from -16 to +15.
6159 int task_prio(const struct task_struct
*p
)
6161 return p
->prio
- MAX_RT_PRIO
;
6165 * task_nice - return the nice value of a given task.
6166 * @p: the task in question.
6168 int task_nice(const struct task_struct
*p
)
6170 return TASK_NICE(p
);
6172 EXPORT_SYMBOL(task_nice
);
6175 * idle_cpu - is a given cpu idle currently?
6176 * @cpu: the processor in question.
6178 int idle_cpu(int cpu
)
6180 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6184 * idle_task - return the idle task for a given cpu.
6185 * @cpu: the processor in question.
6187 struct task_struct
*idle_task(int cpu
)
6189 return cpu_rq(cpu
)->idle
;
6193 * find_process_by_pid - find a process with a matching PID value.
6194 * @pid: the pid in question.
6196 static struct task_struct
*find_process_by_pid(pid_t pid
)
6198 return pid
? find_task_by_vpid(pid
) : current
;
6201 /* Actually do priority change: must hold rq lock. */
6203 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6205 BUG_ON(p
->se
.on_rq
);
6208 switch (p
->policy
) {
6212 p
->sched_class
= &fair_sched_class
;
6216 p
->sched_class
= &rt_sched_class
;
6220 p
->rt_priority
= prio
;
6221 p
->normal_prio
= normal_prio(p
);
6222 /* we are holding p->pi_lock already */
6223 p
->prio
= rt_mutex_getprio(p
);
6228 * check the target process has a UID that matches the current process's
6230 static bool check_same_owner(struct task_struct
*p
)
6232 const struct cred
*cred
= current_cred(), *pcred
;
6236 pcred
= __task_cred(p
);
6237 match
= (cred
->euid
== pcred
->euid
||
6238 cred
->euid
== pcred
->uid
);
6243 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6244 struct sched_param
*param
, bool user
)
6246 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6247 unsigned long flags
;
6248 const struct sched_class
*prev_class
;
6252 /* may grab non-irq protected spin_locks */
6253 BUG_ON(in_interrupt());
6255 /* double check policy once rq lock held */
6257 reset_on_fork
= p
->sched_reset_on_fork
;
6258 policy
= oldpolicy
= p
->policy
;
6260 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6261 policy
&= ~SCHED_RESET_ON_FORK
;
6263 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6264 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6265 policy
!= SCHED_IDLE
)
6270 * Valid priorities for SCHED_FIFO and SCHED_RR are
6271 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6272 * SCHED_BATCH and SCHED_IDLE is 0.
6274 if (param
->sched_priority
< 0 ||
6275 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6276 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6278 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6282 * Allow unprivileged RT tasks to decrease priority:
6284 if (user
&& !capable(CAP_SYS_NICE
)) {
6285 if (rt_policy(policy
)) {
6286 unsigned long rlim_rtprio
;
6288 if (!lock_task_sighand(p
, &flags
))
6290 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6291 unlock_task_sighand(p
, &flags
);
6293 /* can't set/change the rt policy */
6294 if (policy
!= p
->policy
&& !rlim_rtprio
)
6297 /* can't increase priority */
6298 if (param
->sched_priority
> p
->rt_priority
&&
6299 param
->sched_priority
> rlim_rtprio
)
6303 * Like positive nice levels, dont allow tasks to
6304 * move out of SCHED_IDLE either:
6306 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6309 /* can't change other user's priorities */
6310 if (!check_same_owner(p
))
6313 /* Normal users shall not reset the sched_reset_on_fork flag */
6314 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6319 #ifdef CONFIG_RT_GROUP_SCHED
6321 * Do not allow realtime tasks into groups that have no runtime
6324 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6325 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6329 retval
= security_task_setscheduler(p
, policy
, param
);
6335 * make sure no PI-waiters arrive (or leave) while we are
6336 * changing the priority of the task:
6338 spin_lock_irqsave(&p
->pi_lock
, flags
);
6340 * To be able to change p->policy safely, the apropriate
6341 * runqueue lock must be held.
6343 rq
= __task_rq_lock(p
);
6344 /* recheck policy now with rq lock held */
6345 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6346 policy
= oldpolicy
= -1;
6347 __task_rq_unlock(rq
);
6348 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6351 update_rq_clock(rq
);
6352 on_rq
= p
->se
.on_rq
;
6353 running
= task_current(rq
, p
);
6355 deactivate_task(rq
, p
, 0);
6357 p
->sched_class
->put_prev_task(rq
, p
);
6359 p
->sched_reset_on_fork
= reset_on_fork
;
6362 prev_class
= p
->sched_class
;
6363 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6366 p
->sched_class
->set_curr_task(rq
);
6368 activate_task(rq
, p
, 0);
6370 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6372 __task_rq_unlock(rq
);
6373 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6375 rt_mutex_adjust_pi(p
);
6381 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6382 * @p: the task in question.
6383 * @policy: new policy.
6384 * @param: structure containing the new RT priority.
6386 * NOTE that the task may be already dead.
6388 int sched_setscheduler(struct task_struct
*p
, int policy
,
6389 struct sched_param
*param
)
6391 return __sched_setscheduler(p
, policy
, param
, true);
6393 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6396 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6397 * @p: the task in question.
6398 * @policy: new policy.
6399 * @param: structure containing the new RT priority.
6401 * Just like sched_setscheduler, only don't bother checking if the
6402 * current context has permission. For example, this is needed in
6403 * stop_machine(): we create temporary high priority worker threads,
6404 * but our caller might not have that capability.
6406 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6407 struct sched_param
*param
)
6409 return __sched_setscheduler(p
, policy
, param
, false);
6413 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6415 struct sched_param lparam
;
6416 struct task_struct
*p
;
6419 if (!param
|| pid
< 0)
6421 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6426 p
= find_process_by_pid(pid
);
6428 retval
= sched_setscheduler(p
, policy
, &lparam
);
6435 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6436 * @pid: the pid in question.
6437 * @policy: new policy.
6438 * @param: structure containing the new RT priority.
6440 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6441 struct sched_param __user
*, param
)
6443 /* negative values for policy are not valid */
6447 return do_sched_setscheduler(pid
, policy
, param
);
6451 * sys_sched_setparam - set/change the RT priority of a thread
6452 * @pid: the pid in question.
6453 * @param: structure containing the new RT priority.
6455 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6457 return do_sched_setscheduler(pid
, -1, param
);
6461 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6462 * @pid: the pid in question.
6464 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6466 struct task_struct
*p
;
6473 read_lock(&tasklist_lock
);
6474 p
= find_process_by_pid(pid
);
6476 retval
= security_task_getscheduler(p
);
6479 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6481 read_unlock(&tasklist_lock
);
6486 * sys_sched_getparam - get the RT priority of a thread
6487 * @pid: the pid in question.
6488 * @param: structure containing the RT priority.
6490 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6492 struct sched_param lp
;
6493 struct task_struct
*p
;
6496 if (!param
|| pid
< 0)
6499 read_lock(&tasklist_lock
);
6500 p
= find_process_by_pid(pid
);
6505 retval
= security_task_getscheduler(p
);
6509 lp
.sched_priority
= p
->rt_priority
;
6510 read_unlock(&tasklist_lock
);
6513 * This one might sleep, we cannot do it with a spinlock held ...
6515 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6520 read_unlock(&tasklist_lock
);
6524 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6526 cpumask_var_t cpus_allowed
, new_mask
;
6527 struct task_struct
*p
;
6531 read_lock(&tasklist_lock
);
6533 p
= find_process_by_pid(pid
);
6535 read_unlock(&tasklist_lock
);
6541 * It is not safe to call set_cpus_allowed with the
6542 * tasklist_lock held. We will bump the task_struct's
6543 * usage count and then drop tasklist_lock.
6546 read_unlock(&tasklist_lock
);
6548 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6552 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6554 goto out_free_cpus_allowed
;
6557 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6560 retval
= security_task_setscheduler(p
, 0, NULL
);
6564 cpuset_cpus_allowed(p
, cpus_allowed
);
6565 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6567 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6570 cpuset_cpus_allowed(p
, cpus_allowed
);
6571 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6573 * We must have raced with a concurrent cpuset
6574 * update. Just reset the cpus_allowed to the
6575 * cpuset's cpus_allowed
6577 cpumask_copy(new_mask
, cpus_allowed
);
6582 free_cpumask_var(new_mask
);
6583 out_free_cpus_allowed
:
6584 free_cpumask_var(cpus_allowed
);
6591 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6592 struct cpumask
*new_mask
)
6594 if (len
< cpumask_size())
6595 cpumask_clear(new_mask
);
6596 else if (len
> cpumask_size())
6597 len
= cpumask_size();
6599 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6603 * sys_sched_setaffinity - set the cpu affinity of a process
6604 * @pid: pid of the process
6605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6606 * @user_mask_ptr: user-space pointer to the new cpu mask
6608 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6609 unsigned long __user
*, user_mask_ptr
)
6611 cpumask_var_t new_mask
;
6614 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6617 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6619 retval
= sched_setaffinity(pid
, new_mask
);
6620 free_cpumask_var(new_mask
);
6624 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6626 struct task_struct
*p
;
6630 read_lock(&tasklist_lock
);
6633 p
= find_process_by_pid(pid
);
6637 retval
= security_task_getscheduler(p
);
6641 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6644 read_unlock(&tasklist_lock
);
6651 * sys_sched_getaffinity - get the cpu affinity of a process
6652 * @pid: pid of the process
6653 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6654 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6656 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6657 unsigned long __user
*, user_mask_ptr
)
6662 if (len
< cpumask_size())
6665 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6668 ret
= sched_getaffinity(pid
, mask
);
6670 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6673 ret
= cpumask_size();
6675 free_cpumask_var(mask
);
6681 * sys_sched_yield - yield the current processor to other threads.
6683 * This function yields the current CPU to other tasks. If there are no
6684 * other threads running on this CPU then this function will return.
6686 SYSCALL_DEFINE0(sched_yield
)
6688 struct rq
*rq
= this_rq_lock();
6690 schedstat_inc(rq
, yld_count
);
6691 current
->sched_class
->yield_task(rq
);
6694 * Since we are going to call schedule() anyway, there's
6695 * no need to preempt or enable interrupts:
6697 __release(rq
->lock
);
6698 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6699 _raw_spin_unlock(&rq
->lock
);
6700 preempt_enable_no_resched();
6707 static inline int should_resched(void)
6709 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6712 static void __cond_resched(void)
6714 add_preempt_count(PREEMPT_ACTIVE
);
6716 sub_preempt_count(PREEMPT_ACTIVE
);
6719 int __sched
_cond_resched(void)
6721 if (should_resched()) {
6727 EXPORT_SYMBOL(_cond_resched
);
6730 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6731 * call schedule, and on return reacquire the lock.
6733 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6734 * operations here to prevent schedule() from being called twice (once via
6735 * spin_unlock(), once by hand).
6737 int __cond_resched_lock(spinlock_t
*lock
)
6739 int resched
= should_resched();
6742 lockdep_assert_held(lock
);
6744 if (spin_needbreak(lock
) || resched
) {
6755 EXPORT_SYMBOL(__cond_resched_lock
);
6757 int __sched
__cond_resched_softirq(void)
6759 BUG_ON(!in_softirq());
6761 if (should_resched()) {
6769 EXPORT_SYMBOL(__cond_resched_softirq
);
6772 * yield - yield the current processor to other threads.
6774 * This is a shortcut for kernel-space yielding - it marks the
6775 * thread runnable and calls sys_sched_yield().
6777 void __sched
yield(void)
6779 set_current_state(TASK_RUNNING
);
6782 EXPORT_SYMBOL(yield
);
6785 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6786 * that process accounting knows that this is a task in IO wait state.
6788 void __sched
io_schedule(void)
6790 struct rq
*rq
= raw_rq();
6792 delayacct_blkio_start();
6793 atomic_inc(&rq
->nr_iowait
);
6794 current
->in_iowait
= 1;
6796 current
->in_iowait
= 0;
6797 atomic_dec(&rq
->nr_iowait
);
6798 delayacct_blkio_end();
6800 EXPORT_SYMBOL(io_schedule
);
6802 long __sched
io_schedule_timeout(long timeout
)
6804 struct rq
*rq
= raw_rq();
6807 delayacct_blkio_start();
6808 atomic_inc(&rq
->nr_iowait
);
6809 current
->in_iowait
= 1;
6810 ret
= schedule_timeout(timeout
);
6811 current
->in_iowait
= 0;
6812 atomic_dec(&rq
->nr_iowait
);
6813 delayacct_blkio_end();
6818 * sys_sched_get_priority_max - return maximum RT priority.
6819 * @policy: scheduling class.
6821 * this syscall returns the maximum rt_priority that can be used
6822 * by a given scheduling class.
6824 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6831 ret
= MAX_USER_RT_PRIO
-1;
6843 * sys_sched_get_priority_min - return minimum RT priority.
6844 * @policy: scheduling class.
6846 * this syscall returns the minimum rt_priority that can be used
6847 * by a given scheduling class.
6849 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6867 * sys_sched_rr_get_interval - return the default timeslice of a process.
6868 * @pid: pid of the process.
6869 * @interval: userspace pointer to the timeslice value.
6871 * this syscall writes the default timeslice value of a given process
6872 * into the user-space timespec buffer. A value of '0' means infinity.
6874 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6875 struct timespec __user
*, interval
)
6877 struct task_struct
*p
;
6878 unsigned int time_slice
;
6886 read_lock(&tasklist_lock
);
6887 p
= find_process_by_pid(pid
);
6891 retval
= security_task_getscheduler(p
);
6895 time_slice
= p
->sched_class
->get_rr_interval(p
);
6897 read_unlock(&tasklist_lock
);
6898 jiffies_to_timespec(time_slice
, &t
);
6899 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6903 read_unlock(&tasklist_lock
);
6907 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6909 void sched_show_task(struct task_struct
*p
)
6911 unsigned long free
= 0;
6914 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6915 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6916 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6917 #if BITS_PER_LONG == 32
6918 if (state
== TASK_RUNNING
)
6919 printk(KERN_CONT
" running ");
6921 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6923 if (state
== TASK_RUNNING
)
6924 printk(KERN_CONT
" running task ");
6926 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6928 #ifdef CONFIG_DEBUG_STACK_USAGE
6929 free
= stack_not_used(p
);
6931 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6932 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6933 (unsigned long)task_thread_info(p
)->flags
);
6935 show_stack(p
, NULL
);
6938 void show_state_filter(unsigned long state_filter
)
6940 struct task_struct
*g
, *p
;
6942 #if BITS_PER_LONG == 32
6944 " task PC stack pid father\n");
6947 " task PC stack pid father\n");
6949 read_lock(&tasklist_lock
);
6950 do_each_thread(g
, p
) {
6952 * reset the NMI-timeout, listing all files on a slow
6953 * console might take alot of time:
6955 touch_nmi_watchdog();
6956 if (!state_filter
|| (p
->state
& state_filter
))
6958 } while_each_thread(g
, p
);
6960 touch_all_softlockup_watchdogs();
6962 #ifdef CONFIG_SCHED_DEBUG
6963 sysrq_sched_debug_show();
6965 read_unlock(&tasklist_lock
);
6967 * Only show locks if all tasks are dumped:
6969 if (state_filter
== -1)
6970 debug_show_all_locks();
6973 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6975 idle
->sched_class
= &idle_sched_class
;
6979 * init_idle - set up an idle thread for a given CPU
6980 * @idle: task in question
6981 * @cpu: cpu the idle task belongs to
6983 * NOTE: this function does not set the idle thread's NEED_RESCHED
6984 * flag, to make booting more robust.
6986 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6988 struct rq
*rq
= cpu_rq(cpu
);
6989 unsigned long flags
;
6991 spin_lock_irqsave(&rq
->lock
, flags
);
6994 idle
->se
.exec_start
= sched_clock();
6996 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6997 __set_task_cpu(idle
, cpu
);
6999 rq
->curr
= rq
->idle
= idle
;
7000 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7003 spin_unlock_irqrestore(&rq
->lock
, flags
);
7005 /* Set the preempt count _outside_ the spinlocks! */
7006 #if defined(CONFIG_PREEMPT)
7007 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7009 task_thread_info(idle
)->preempt_count
= 0;
7012 * The idle tasks have their own, simple scheduling class:
7014 idle
->sched_class
= &idle_sched_class
;
7015 ftrace_graph_init_task(idle
);
7019 * In a system that switches off the HZ timer nohz_cpu_mask
7020 * indicates which cpus entered this state. This is used
7021 * in the rcu update to wait only for active cpus. For system
7022 * which do not switch off the HZ timer nohz_cpu_mask should
7023 * always be CPU_BITS_NONE.
7025 cpumask_var_t nohz_cpu_mask
;
7028 * Increase the granularity value when there are more CPUs,
7029 * because with more CPUs the 'effective latency' as visible
7030 * to users decreases. But the relationship is not linear,
7031 * so pick a second-best guess by going with the log2 of the
7034 * This idea comes from the SD scheduler of Con Kolivas:
7036 static void update_sysctl(void)
7038 unsigned int cpus
= min(num_online_cpus(), 8U);
7039 unsigned int factor
= 1 + ilog2(cpus
);
7041 #define SET_SYSCTL(name) \
7042 (sysctl_##name = (factor) * normalized_sysctl_##name)
7043 SET_SYSCTL(sched_min_granularity
);
7044 SET_SYSCTL(sched_latency
);
7045 SET_SYSCTL(sched_wakeup_granularity
);
7046 SET_SYSCTL(sched_shares_ratelimit
);
7050 static inline void sched_init_granularity(void)
7057 * This is how migration works:
7059 * 1) we queue a struct migration_req structure in the source CPU's
7060 * runqueue and wake up that CPU's migration thread.
7061 * 2) we down() the locked semaphore => thread blocks.
7062 * 3) migration thread wakes up (implicitly it forces the migrated
7063 * thread off the CPU)
7064 * 4) it gets the migration request and checks whether the migrated
7065 * task is still in the wrong runqueue.
7066 * 5) if it's in the wrong runqueue then the migration thread removes
7067 * it and puts it into the right queue.
7068 * 6) migration thread up()s the semaphore.
7069 * 7) we wake up and the migration is done.
7073 * Change a given task's CPU affinity. Migrate the thread to a
7074 * proper CPU and schedule it away if the CPU it's executing on
7075 * is removed from the allowed bitmask.
7077 * NOTE: the caller must have a valid reference to the task, the
7078 * task must not exit() & deallocate itself prematurely. The
7079 * call is not atomic; no spinlocks may be held.
7081 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7083 struct migration_req req
;
7084 unsigned long flags
;
7088 rq
= task_rq_lock(p
, &flags
);
7089 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
7094 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7095 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7100 if (p
->sched_class
->set_cpus_allowed
)
7101 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7103 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7104 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7107 /* Can the task run on the task's current CPU? If so, we're done */
7108 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7111 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
7112 /* Need help from migration thread: drop lock and wait. */
7113 struct task_struct
*mt
= rq
->migration_thread
;
7115 get_task_struct(mt
);
7116 task_rq_unlock(rq
, &flags
);
7117 wake_up_process(rq
->migration_thread
);
7118 put_task_struct(mt
);
7119 wait_for_completion(&req
.done
);
7120 tlb_migrate_finish(p
->mm
);
7124 task_rq_unlock(rq
, &flags
);
7128 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7131 * Move (not current) task off this cpu, onto dest cpu. We're doing
7132 * this because either it can't run here any more (set_cpus_allowed()
7133 * away from this CPU, or CPU going down), or because we're
7134 * attempting to rebalance this task on exec (sched_exec).
7136 * So we race with normal scheduler movements, but that's OK, as long
7137 * as the task is no longer on this CPU.
7139 * Returns non-zero if task was successfully migrated.
7141 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7143 struct rq
*rq_dest
, *rq_src
;
7146 if (unlikely(!cpu_active(dest_cpu
)))
7149 rq_src
= cpu_rq(src_cpu
);
7150 rq_dest
= cpu_rq(dest_cpu
);
7152 double_rq_lock(rq_src
, rq_dest
);
7153 /* Already moved. */
7154 if (task_cpu(p
) != src_cpu
)
7156 /* Affinity changed (again). */
7157 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7160 on_rq
= p
->se
.on_rq
;
7162 deactivate_task(rq_src
, p
, 0);
7164 set_task_cpu(p
, dest_cpu
);
7166 activate_task(rq_dest
, p
, 0);
7167 check_preempt_curr(rq_dest
, p
, 0);
7172 double_rq_unlock(rq_src
, rq_dest
);
7176 #define RCU_MIGRATION_IDLE 0
7177 #define RCU_MIGRATION_NEED_QS 1
7178 #define RCU_MIGRATION_GOT_QS 2
7179 #define RCU_MIGRATION_MUST_SYNC 3
7182 * migration_thread - this is a highprio system thread that performs
7183 * thread migration by bumping thread off CPU then 'pushing' onto
7186 static int migration_thread(void *data
)
7189 int cpu
= (long)data
;
7193 BUG_ON(rq
->migration_thread
!= current
);
7195 set_current_state(TASK_INTERRUPTIBLE
);
7196 while (!kthread_should_stop()) {
7197 struct migration_req
*req
;
7198 struct list_head
*head
;
7200 spin_lock_irq(&rq
->lock
);
7202 if (cpu_is_offline(cpu
)) {
7203 spin_unlock_irq(&rq
->lock
);
7207 if (rq
->active_balance
) {
7208 active_load_balance(rq
, cpu
);
7209 rq
->active_balance
= 0;
7212 head
= &rq
->migration_queue
;
7214 if (list_empty(head
)) {
7215 spin_unlock_irq(&rq
->lock
);
7217 set_current_state(TASK_INTERRUPTIBLE
);
7220 req
= list_entry(head
->next
, struct migration_req
, list
);
7221 list_del_init(head
->next
);
7223 if (req
->task
!= NULL
) {
7224 spin_unlock(&rq
->lock
);
7225 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7226 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7227 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7228 spin_unlock(&rq
->lock
);
7230 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7231 spin_unlock(&rq
->lock
);
7232 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7236 complete(&req
->done
);
7238 __set_current_state(TASK_RUNNING
);
7243 #ifdef CONFIG_HOTPLUG_CPU
7245 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7249 local_irq_disable();
7250 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7256 * Figure out where task on dead CPU should go, use force if necessary.
7258 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7261 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7264 /* Look for allowed, online CPU in same node. */
7265 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
7266 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7269 /* Any allowed, online CPU? */
7270 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
7271 if (dest_cpu
< nr_cpu_ids
)
7274 /* No more Mr. Nice Guy. */
7275 if (dest_cpu
>= nr_cpu_ids
) {
7276 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7277 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
7280 * Don't tell them about moving exiting tasks or
7281 * kernel threads (both mm NULL), since they never
7284 if (p
->mm
&& printk_ratelimit()) {
7285 printk(KERN_INFO
"process %d (%s) no "
7286 "longer affine to cpu%d\n",
7287 task_pid_nr(p
), p
->comm
, dead_cpu
);
7292 /* It can have affinity changed while we were choosing. */
7293 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7298 * While a dead CPU has no uninterruptible tasks queued at this point,
7299 * it might still have a nonzero ->nr_uninterruptible counter, because
7300 * for performance reasons the counter is not stricly tracking tasks to
7301 * their home CPUs. So we just add the counter to another CPU's counter,
7302 * to keep the global sum constant after CPU-down:
7304 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7306 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
7307 unsigned long flags
;
7309 local_irq_save(flags
);
7310 double_rq_lock(rq_src
, rq_dest
);
7311 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7312 rq_src
->nr_uninterruptible
= 0;
7313 double_rq_unlock(rq_src
, rq_dest
);
7314 local_irq_restore(flags
);
7317 /* Run through task list and migrate tasks from the dead cpu. */
7318 static void migrate_live_tasks(int src_cpu
)
7320 struct task_struct
*p
, *t
;
7322 read_lock(&tasklist_lock
);
7324 do_each_thread(t
, p
) {
7328 if (task_cpu(p
) == src_cpu
)
7329 move_task_off_dead_cpu(src_cpu
, p
);
7330 } while_each_thread(t
, p
);
7332 read_unlock(&tasklist_lock
);
7336 * Schedules idle task to be the next runnable task on current CPU.
7337 * It does so by boosting its priority to highest possible.
7338 * Used by CPU offline code.
7340 void sched_idle_next(void)
7342 int this_cpu
= smp_processor_id();
7343 struct rq
*rq
= cpu_rq(this_cpu
);
7344 struct task_struct
*p
= rq
->idle
;
7345 unsigned long flags
;
7347 /* cpu has to be offline */
7348 BUG_ON(cpu_online(this_cpu
));
7351 * Strictly not necessary since rest of the CPUs are stopped by now
7352 * and interrupts disabled on the current cpu.
7354 spin_lock_irqsave(&rq
->lock
, flags
);
7356 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7358 update_rq_clock(rq
);
7359 activate_task(rq
, p
, 0);
7361 spin_unlock_irqrestore(&rq
->lock
, flags
);
7365 * Ensures that the idle task is using init_mm right before its cpu goes
7368 void idle_task_exit(void)
7370 struct mm_struct
*mm
= current
->active_mm
;
7372 BUG_ON(cpu_online(smp_processor_id()));
7375 switch_mm(mm
, &init_mm
, current
);
7379 /* called under rq->lock with disabled interrupts */
7380 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7382 struct rq
*rq
= cpu_rq(dead_cpu
);
7384 /* Must be exiting, otherwise would be on tasklist. */
7385 BUG_ON(!p
->exit_state
);
7387 /* Cannot have done final schedule yet: would have vanished. */
7388 BUG_ON(p
->state
== TASK_DEAD
);
7393 * Drop lock around migration; if someone else moves it,
7394 * that's OK. No task can be added to this CPU, so iteration is
7397 spin_unlock_irq(&rq
->lock
);
7398 move_task_off_dead_cpu(dead_cpu
, p
);
7399 spin_lock_irq(&rq
->lock
);
7404 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7405 static void migrate_dead_tasks(unsigned int dead_cpu
)
7407 struct rq
*rq
= cpu_rq(dead_cpu
);
7408 struct task_struct
*next
;
7411 if (!rq
->nr_running
)
7413 update_rq_clock(rq
);
7414 next
= pick_next_task(rq
);
7417 next
->sched_class
->put_prev_task(rq
, next
);
7418 migrate_dead(dead_cpu
, next
);
7424 * remove the tasks which were accounted by rq from calc_load_tasks.
7426 static void calc_global_load_remove(struct rq
*rq
)
7428 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7429 rq
->calc_load_active
= 0;
7431 #endif /* CONFIG_HOTPLUG_CPU */
7433 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7435 static struct ctl_table sd_ctl_dir
[] = {
7437 .procname
= "sched_domain",
7443 static struct ctl_table sd_ctl_root
[] = {
7445 .ctl_name
= CTL_KERN
,
7446 .procname
= "kernel",
7448 .child
= sd_ctl_dir
,
7453 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7455 struct ctl_table
*entry
=
7456 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7461 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7463 struct ctl_table
*entry
;
7466 * In the intermediate directories, both the child directory and
7467 * procname are dynamically allocated and could fail but the mode
7468 * will always be set. In the lowest directory the names are
7469 * static strings and all have proc handlers.
7471 for (entry
= *tablep
; entry
->mode
; entry
++) {
7473 sd_free_ctl_entry(&entry
->child
);
7474 if (entry
->proc_handler
== NULL
)
7475 kfree(entry
->procname
);
7483 set_table_entry(struct ctl_table
*entry
,
7484 const char *procname
, void *data
, int maxlen
,
7485 mode_t mode
, proc_handler
*proc_handler
)
7487 entry
->procname
= procname
;
7489 entry
->maxlen
= maxlen
;
7491 entry
->proc_handler
= proc_handler
;
7494 static struct ctl_table
*
7495 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7497 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7502 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7503 sizeof(long), 0644, proc_doulongvec_minmax
);
7504 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7505 sizeof(long), 0644, proc_doulongvec_minmax
);
7506 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7507 sizeof(int), 0644, proc_dointvec_minmax
);
7508 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7509 sizeof(int), 0644, proc_dointvec_minmax
);
7510 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7511 sizeof(int), 0644, proc_dointvec_minmax
);
7512 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7513 sizeof(int), 0644, proc_dointvec_minmax
);
7514 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7515 sizeof(int), 0644, proc_dointvec_minmax
);
7516 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7517 sizeof(int), 0644, proc_dointvec_minmax
);
7518 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7519 sizeof(int), 0644, proc_dointvec_minmax
);
7520 set_table_entry(&table
[9], "cache_nice_tries",
7521 &sd
->cache_nice_tries
,
7522 sizeof(int), 0644, proc_dointvec_minmax
);
7523 set_table_entry(&table
[10], "flags", &sd
->flags
,
7524 sizeof(int), 0644, proc_dointvec_minmax
);
7525 set_table_entry(&table
[11], "name", sd
->name
,
7526 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7527 /* &table[12] is terminator */
7532 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7534 struct ctl_table
*entry
, *table
;
7535 struct sched_domain
*sd
;
7536 int domain_num
= 0, i
;
7539 for_each_domain(cpu
, sd
)
7541 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7546 for_each_domain(cpu
, sd
) {
7547 snprintf(buf
, 32, "domain%d", i
);
7548 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7550 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7557 static struct ctl_table_header
*sd_sysctl_header
;
7558 static void register_sched_domain_sysctl(void)
7560 int i
, cpu_num
= num_possible_cpus();
7561 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7564 WARN_ON(sd_ctl_dir
[0].child
);
7565 sd_ctl_dir
[0].child
= entry
;
7570 for_each_possible_cpu(i
) {
7571 snprintf(buf
, 32, "cpu%d", i
);
7572 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7574 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7578 WARN_ON(sd_sysctl_header
);
7579 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7582 /* may be called multiple times per register */
7583 static void unregister_sched_domain_sysctl(void)
7585 if (sd_sysctl_header
)
7586 unregister_sysctl_table(sd_sysctl_header
);
7587 sd_sysctl_header
= NULL
;
7588 if (sd_ctl_dir
[0].child
)
7589 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7592 static void register_sched_domain_sysctl(void)
7595 static void unregister_sched_domain_sysctl(void)
7600 static void set_rq_online(struct rq
*rq
)
7603 const struct sched_class
*class;
7605 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7608 for_each_class(class) {
7609 if (class->rq_online
)
7610 class->rq_online(rq
);
7615 static void set_rq_offline(struct rq
*rq
)
7618 const struct sched_class
*class;
7620 for_each_class(class) {
7621 if (class->rq_offline
)
7622 class->rq_offline(rq
);
7625 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7631 * migration_call - callback that gets triggered when a CPU is added.
7632 * Here we can start up the necessary migration thread for the new CPU.
7634 static int __cpuinit
7635 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7637 struct task_struct
*p
;
7638 int cpu
= (long)hcpu
;
7639 unsigned long flags
;
7644 case CPU_UP_PREPARE
:
7645 case CPU_UP_PREPARE_FROZEN
:
7646 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7649 kthread_bind(p
, cpu
);
7650 /* Must be high prio: stop_machine expects to yield to it. */
7651 rq
= task_rq_lock(p
, &flags
);
7652 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7653 task_rq_unlock(rq
, &flags
);
7655 cpu_rq(cpu
)->migration_thread
= p
;
7656 rq
->calc_load_update
= calc_load_update
;
7660 case CPU_ONLINE_FROZEN
:
7661 /* Strictly unnecessary, as first user will wake it. */
7662 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7664 /* Update our root-domain */
7666 spin_lock_irqsave(&rq
->lock
, flags
);
7668 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7672 spin_unlock_irqrestore(&rq
->lock
, flags
);
7675 #ifdef CONFIG_HOTPLUG_CPU
7676 case CPU_UP_CANCELED
:
7677 case CPU_UP_CANCELED_FROZEN
:
7678 if (!cpu_rq(cpu
)->migration_thread
)
7680 /* Unbind it from offline cpu so it can run. Fall thru. */
7681 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7682 cpumask_any(cpu_online_mask
));
7683 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7684 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7685 cpu_rq(cpu
)->migration_thread
= NULL
;
7689 case CPU_DEAD_FROZEN
:
7690 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7691 migrate_live_tasks(cpu
);
7693 kthread_stop(rq
->migration_thread
);
7694 put_task_struct(rq
->migration_thread
);
7695 rq
->migration_thread
= NULL
;
7696 /* Idle task back to normal (off runqueue, low prio) */
7697 spin_lock_irq(&rq
->lock
);
7698 update_rq_clock(rq
);
7699 deactivate_task(rq
, rq
->idle
, 0);
7700 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7701 rq
->idle
->sched_class
= &idle_sched_class
;
7702 migrate_dead_tasks(cpu
);
7703 spin_unlock_irq(&rq
->lock
);
7705 migrate_nr_uninterruptible(rq
);
7706 BUG_ON(rq
->nr_running
!= 0);
7707 calc_global_load_remove(rq
);
7709 * No need to migrate the tasks: it was best-effort if
7710 * they didn't take sched_hotcpu_mutex. Just wake up
7713 spin_lock_irq(&rq
->lock
);
7714 while (!list_empty(&rq
->migration_queue
)) {
7715 struct migration_req
*req
;
7717 req
= list_entry(rq
->migration_queue
.next
,
7718 struct migration_req
, list
);
7719 list_del_init(&req
->list
);
7720 spin_unlock_irq(&rq
->lock
);
7721 complete(&req
->done
);
7722 spin_lock_irq(&rq
->lock
);
7724 spin_unlock_irq(&rq
->lock
);
7728 case CPU_DYING_FROZEN
:
7729 /* Update our root-domain */
7731 spin_lock_irqsave(&rq
->lock
, flags
);
7733 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7736 spin_unlock_irqrestore(&rq
->lock
, flags
);
7744 * Register at high priority so that task migration (migrate_all_tasks)
7745 * happens before everything else. This has to be lower priority than
7746 * the notifier in the perf_event subsystem, though.
7748 static struct notifier_block __cpuinitdata migration_notifier
= {
7749 .notifier_call
= migration_call
,
7753 static int __init
migration_init(void)
7755 void *cpu
= (void *)(long)smp_processor_id();
7758 /* Start one for the boot CPU: */
7759 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7760 BUG_ON(err
== NOTIFY_BAD
);
7761 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7762 register_cpu_notifier(&migration_notifier
);
7766 early_initcall(migration_init
);
7771 #ifdef CONFIG_SCHED_DEBUG
7773 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7774 struct cpumask
*groupmask
)
7776 struct sched_group
*group
= sd
->groups
;
7779 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7780 cpumask_clear(groupmask
);
7782 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7784 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7785 printk("does not load-balance\n");
7787 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7792 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7794 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7795 printk(KERN_ERR
"ERROR: domain->span does not contain "
7798 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7799 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7803 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7807 printk(KERN_ERR
"ERROR: group is NULL\n");
7811 if (!group
->cpu_power
) {
7812 printk(KERN_CONT
"\n");
7813 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7818 if (!cpumask_weight(sched_group_cpus(group
))) {
7819 printk(KERN_CONT
"\n");
7820 printk(KERN_ERR
"ERROR: empty group\n");
7824 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7825 printk(KERN_CONT
"\n");
7826 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7830 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7832 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7834 printk(KERN_CONT
" %s", str
);
7835 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7836 printk(KERN_CONT
" (cpu_power = %d)",
7840 group
= group
->next
;
7841 } while (group
!= sd
->groups
);
7842 printk(KERN_CONT
"\n");
7844 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7845 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7848 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7849 printk(KERN_ERR
"ERROR: parent span is not a superset "
7850 "of domain->span\n");
7854 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7856 cpumask_var_t groupmask
;
7860 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7864 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7866 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7867 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7872 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7879 free_cpumask_var(groupmask
);
7881 #else /* !CONFIG_SCHED_DEBUG */
7882 # define sched_domain_debug(sd, cpu) do { } while (0)
7883 #endif /* CONFIG_SCHED_DEBUG */
7885 static int sd_degenerate(struct sched_domain
*sd
)
7887 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7890 /* Following flags need at least 2 groups */
7891 if (sd
->flags
& (SD_LOAD_BALANCE
|
7892 SD_BALANCE_NEWIDLE
|
7896 SD_SHARE_PKG_RESOURCES
)) {
7897 if (sd
->groups
!= sd
->groups
->next
)
7901 /* Following flags don't use groups */
7902 if (sd
->flags
& (SD_WAKE_AFFINE
))
7909 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7911 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7913 if (sd_degenerate(parent
))
7916 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7919 /* Flags needing groups don't count if only 1 group in parent */
7920 if (parent
->groups
== parent
->groups
->next
) {
7921 pflags
&= ~(SD_LOAD_BALANCE
|
7922 SD_BALANCE_NEWIDLE
|
7926 SD_SHARE_PKG_RESOURCES
);
7927 if (nr_node_ids
== 1)
7928 pflags
&= ~SD_SERIALIZE
;
7930 if (~cflags
& pflags
)
7936 static void free_rootdomain(struct root_domain
*rd
)
7938 synchronize_sched();
7940 cpupri_cleanup(&rd
->cpupri
);
7942 free_cpumask_var(rd
->rto_mask
);
7943 free_cpumask_var(rd
->online
);
7944 free_cpumask_var(rd
->span
);
7948 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7950 struct root_domain
*old_rd
= NULL
;
7951 unsigned long flags
;
7953 spin_lock_irqsave(&rq
->lock
, flags
);
7958 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7961 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7964 * If we dont want to free the old_rt yet then
7965 * set old_rd to NULL to skip the freeing later
7968 if (!atomic_dec_and_test(&old_rd
->refcount
))
7972 atomic_inc(&rd
->refcount
);
7975 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7976 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
7979 spin_unlock_irqrestore(&rq
->lock
, flags
);
7982 free_rootdomain(old_rd
);
7985 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7987 gfp_t gfp
= GFP_KERNEL
;
7989 memset(rd
, 0, sizeof(*rd
));
7994 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7996 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7998 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8001 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8006 free_cpumask_var(rd
->rto_mask
);
8008 free_cpumask_var(rd
->online
);
8010 free_cpumask_var(rd
->span
);
8015 static void init_defrootdomain(void)
8017 init_rootdomain(&def_root_domain
, true);
8019 atomic_set(&def_root_domain
.refcount
, 1);
8022 static struct root_domain
*alloc_rootdomain(void)
8024 struct root_domain
*rd
;
8026 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8030 if (init_rootdomain(rd
, false) != 0) {
8039 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8040 * hold the hotplug lock.
8043 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8045 struct rq
*rq
= cpu_rq(cpu
);
8046 struct sched_domain
*tmp
;
8048 /* Remove the sched domains which do not contribute to scheduling. */
8049 for (tmp
= sd
; tmp
; ) {
8050 struct sched_domain
*parent
= tmp
->parent
;
8054 if (sd_parent_degenerate(tmp
, parent
)) {
8055 tmp
->parent
= parent
->parent
;
8057 parent
->parent
->child
= tmp
;
8062 if (sd
&& sd_degenerate(sd
)) {
8068 sched_domain_debug(sd
, cpu
);
8070 rq_attach_root(rq
, rd
);
8071 rcu_assign_pointer(rq
->sd
, sd
);
8074 /* cpus with isolated domains */
8075 static cpumask_var_t cpu_isolated_map
;
8077 /* Setup the mask of cpus configured for isolated domains */
8078 static int __init
isolated_cpu_setup(char *str
)
8080 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8081 cpulist_parse(str
, cpu_isolated_map
);
8085 __setup("isolcpus=", isolated_cpu_setup
);
8088 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8089 * to a function which identifies what group(along with sched group) a CPU
8090 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8091 * (due to the fact that we keep track of groups covered with a struct cpumask).
8093 * init_sched_build_groups will build a circular linked list of the groups
8094 * covered by the given span, and will set each group's ->cpumask correctly,
8095 * and ->cpu_power to 0.
8098 init_sched_build_groups(const struct cpumask
*span
,
8099 const struct cpumask
*cpu_map
,
8100 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8101 struct sched_group
**sg
,
8102 struct cpumask
*tmpmask
),
8103 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8105 struct sched_group
*first
= NULL
, *last
= NULL
;
8108 cpumask_clear(covered
);
8110 for_each_cpu(i
, span
) {
8111 struct sched_group
*sg
;
8112 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8115 if (cpumask_test_cpu(i
, covered
))
8118 cpumask_clear(sched_group_cpus(sg
));
8121 for_each_cpu(j
, span
) {
8122 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8125 cpumask_set_cpu(j
, covered
);
8126 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8137 #define SD_NODES_PER_DOMAIN 16
8142 * find_next_best_node - find the next node to include in a sched_domain
8143 * @node: node whose sched_domain we're building
8144 * @used_nodes: nodes already in the sched_domain
8146 * Find the next node to include in a given scheduling domain. Simply
8147 * finds the closest node not already in the @used_nodes map.
8149 * Should use nodemask_t.
8151 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8153 int i
, n
, val
, min_val
, best_node
= 0;
8157 for (i
= 0; i
< nr_node_ids
; i
++) {
8158 /* Start at @node */
8159 n
= (node
+ i
) % nr_node_ids
;
8161 if (!nr_cpus_node(n
))
8164 /* Skip already used nodes */
8165 if (node_isset(n
, *used_nodes
))
8168 /* Simple min distance search */
8169 val
= node_distance(node
, n
);
8171 if (val
< min_val
) {
8177 node_set(best_node
, *used_nodes
);
8182 * sched_domain_node_span - get a cpumask for a node's sched_domain
8183 * @node: node whose cpumask we're constructing
8184 * @span: resulting cpumask
8186 * Given a node, construct a good cpumask for its sched_domain to span. It
8187 * should be one that prevents unnecessary balancing, but also spreads tasks
8190 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8192 nodemask_t used_nodes
;
8195 cpumask_clear(span
);
8196 nodes_clear(used_nodes
);
8198 cpumask_or(span
, span
, cpumask_of_node(node
));
8199 node_set(node
, used_nodes
);
8201 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8202 int next_node
= find_next_best_node(node
, &used_nodes
);
8204 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8207 #endif /* CONFIG_NUMA */
8209 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8212 * The cpus mask in sched_group and sched_domain hangs off the end.
8214 * ( See the the comments in include/linux/sched.h:struct sched_group
8215 * and struct sched_domain. )
8217 struct static_sched_group
{
8218 struct sched_group sg
;
8219 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8222 struct static_sched_domain
{
8223 struct sched_domain sd
;
8224 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8230 cpumask_var_t domainspan
;
8231 cpumask_var_t covered
;
8232 cpumask_var_t notcovered
;
8234 cpumask_var_t nodemask
;
8235 cpumask_var_t this_sibling_map
;
8236 cpumask_var_t this_core_map
;
8237 cpumask_var_t send_covered
;
8238 cpumask_var_t tmpmask
;
8239 struct sched_group
**sched_group_nodes
;
8240 struct root_domain
*rd
;
8244 sa_sched_groups
= 0,
8249 sa_this_sibling_map
,
8251 sa_sched_group_nodes
,
8261 * SMT sched-domains:
8263 #ifdef CONFIG_SCHED_SMT
8264 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8265 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8268 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8269 struct sched_group
**sg
, struct cpumask
*unused
)
8272 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8275 #endif /* CONFIG_SCHED_SMT */
8278 * multi-core sched-domains:
8280 #ifdef CONFIG_SCHED_MC
8281 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8282 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8283 #endif /* CONFIG_SCHED_MC */
8285 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8287 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8288 struct sched_group
**sg
, struct cpumask
*mask
)
8292 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8293 group
= cpumask_first(mask
);
8295 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8298 #elif defined(CONFIG_SCHED_MC)
8300 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8301 struct sched_group
**sg
, struct cpumask
*unused
)
8304 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8309 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8310 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8313 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8314 struct sched_group
**sg
, struct cpumask
*mask
)
8317 #ifdef CONFIG_SCHED_MC
8318 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8319 group
= cpumask_first(mask
);
8320 #elif defined(CONFIG_SCHED_SMT)
8321 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8322 group
= cpumask_first(mask
);
8327 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8333 * The init_sched_build_groups can't handle what we want to do with node
8334 * groups, so roll our own. Now each node has its own list of groups which
8335 * gets dynamically allocated.
8337 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8338 static struct sched_group
***sched_group_nodes_bycpu
;
8340 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8341 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8343 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8344 struct sched_group
**sg
,
8345 struct cpumask
*nodemask
)
8349 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8350 group
= cpumask_first(nodemask
);
8353 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8357 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8359 struct sched_group
*sg
= group_head
;
8365 for_each_cpu(j
, sched_group_cpus(sg
)) {
8366 struct sched_domain
*sd
;
8368 sd
= &per_cpu(phys_domains
, j
).sd
;
8369 if (j
!= group_first_cpu(sd
->groups
)) {
8371 * Only add "power" once for each
8377 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8380 } while (sg
!= group_head
);
8383 static int build_numa_sched_groups(struct s_data
*d
,
8384 const struct cpumask
*cpu_map
, int num
)
8386 struct sched_domain
*sd
;
8387 struct sched_group
*sg
, *prev
;
8390 cpumask_clear(d
->covered
);
8391 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8392 if (cpumask_empty(d
->nodemask
)) {
8393 d
->sched_group_nodes
[num
] = NULL
;
8397 sched_domain_node_span(num
, d
->domainspan
);
8398 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8400 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8403 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8407 d
->sched_group_nodes
[num
] = sg
;
8409 for_each_cpu(j
, d
->nodemask
) {
8410 sd
= &per_cpu(node_domains
, j
).sd
;
8415 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8417 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8420 for (j
= 0; j
< nr_node_ids
; j
++) {
8421 n
= (num
+ j
) % nr_node_ids
;
8422 cpumask_complement(d
->notcovered
, d
->covered
);
8423 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8424 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8425 if (cpumask_empty(d
->tmpmask
))
8427 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8428 if (cpumask_empty(d
->tmpmask
))
8430 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8434 "Can not alloc domain group for node %d\n", j
);
8438 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8439 sg
->next
= prev
->next
;
8440 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8447 #endif /* CONFIG_NUMA */
8450 /* Free memory allocated for various sched_group structures */
8451 static void free_sched_groups(const struct cpumask
*cpu_map
,
8452 struct cpumask
*nodemask
)
8456 for_each_cpu(cpu
, cpu_map
) {
8457 struct sched_group
**sched_group_nodes
8458 = sched_group_nodes_bycpu
[cpu
];
8460 if (!sched_group_nodes
)
8463 for (i
= 0; i
< nr_node_ids
; i
++) {
8464 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8466 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8467 if (cpumask_empty(nodemask
))
8477 if (oldsg
!= sched_group_nodes
[i
])
8480 kfree(sched_group_nodes
);
8481 sched_group_nodes_bycpu
[cpu
] = NULL
;
8484 #else /* !CONFIG_NUMA */
8485 static void free_sched_groups(const struct cpumask
*cpu_map
,
8486 struct cpumask
*nodemask
)
8489 #endif /* CONFIG_NUMA */
8492 * Initialize sched groups cpu_power.
8494 * cpu_power indicates the capacity of sched group, which is used while
8495 * distributing the load between different sched groups in a sched domain.
8496 * Typically cpu_power for all the groups in a sched domain will be same unless
8497 * there are asymmetries in the topology. If there are asymmetries, group
8498 * having more cpu_power will pickup more load compared to the group having
8501 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8503 struct sched_domain
*child
;
8504 struct sched_group
*group
;
8508 WARN_ON(!sd
|| !sd
->groups
);
8510 if (cpu
!= group_first_cpu(sd
->groups
))
8515 sd
->groups
->cpu_power
= 0;
8518 power
= SCHED_LOAD_SCALE
;
8519 weight
= cpumask_weight(sched_domain_span(sd
));
8521 * SMT siblings share the power of a single core.
8522 * Usually multiple threads get a better yield out of
8523 * that one core than a single thread would have,
8524 * reflect that in sd->smt_gain.
8526 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8527 power
*= sd
->smt_gain
;
8529 power
>>= SCHED_LOAD_SHIFT
;
8531 sd
->groups
->cpu_power
+= power
;
8536 * Add cpu_power of each child group to this groups cpu_power.
8538 group
= child
->groups
;
8540 sd
->groups
->cpu_power
+= group
->cpu_power
;
8541 group
= group
->next
;
8542 } while (group
!= child
->groups
);
8546 * Initializers for schedule domains
8547 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8550 #ifdef CONFIG_SCHED_DEBUG
8551 # define SD_INIT_NAME(sd, type) sd->name = #type
8553 # define SD_INIT_NAME(sd, type) do { } while (0)
8556 #define SD_INIT(sd, type) sd_init_##type(sd)
8558 #define SD_INIT_FUNC(type) \
8559 static noinline void sd_init_##type(struct sched_domain *sd) \
8561 memset(sd, 0, sizeof(*sd)); \
8562 *sd = SD_##type##_INIT; \
8563 sd->level = SD_LV_##type; \
8564 SD_INIT_NAME(sd, type); \
8569 SD_INIT_FUNC(ALLNODES
)
8572 #ifdef CONFIG_SCHED_SMT
8573 SD_INIT_FUNC(SIBLING
)
8575 #ifdef CONFIG_SCHED_MC
8579 static int default_relax_domain_level
= -1;
8581 static int __init
setup_relax_domain_level(char *str
)
8585 val
= simple_strtoul(str
, NULL
, 0);
8586 if (val
< SD_LV_MAX
)
8587 default_relax_domain_level
= val
;
8591 __setup("relax_domain_level=", setup_relax_domain_level
);
8593 static void set_domain_attribute(struct sched_domain
*sd
,
8594 struct sched_domain_attr
*attr
)
8598 if (!attr
|| attr
->relax_domain_level
< 0) {
8599 if (default_relax_domain_level
< 0)
8602 request
= default_relax_domain_level
;
8604 request
= attr
->relax_domain_level
;
8605 if (request
< sd
->level
) {
8606 /* turn off idle balance on this domain */
8607 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8609 /* turn on idle balance on this domain */
8610 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8614 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8615 const struct cpumask
*cpu_map
)
8618 case sa_sched_groups
:
8619 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8620 d
->sched_group_nodes
= NULL
;
8622 free_rootdomain(d
->rd
); /* fall through */
8624 free_cpumask_var(d
->tmpmask
); /* fall through */
8625 case sa_send_covered
:
8626 free_cpumask_var(d
->send_covered
); /* fall through */
8627 case sa_this_core_map
:
8628 free_cpumask_var(d
->this_core_map
); /* fall through */
8629 case sa_this_sibling_map
:
8630 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8632 free_cpumask_var(d
->nodemask
); /* fall through */
8633 case sa_sched_group_nodes
:
8635 kfree(d
->sched_group_nodes
); /* fall through */
8637 free_cpumask_var(d
->notcovered
); /* fall through */
8639 free_cpumask_var(d
->covered
); /* fall through */
8641 free_cpumask_var(d
->domainspan
); /* fall through */
8648 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8649 const struct cpumask
*cpu_map
)
8652 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8654 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8655 return sa_domainspan
;
8656 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8658 /* Allocate the per-node list of sched groups */
8659 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8660 sizeof(struct sched_group
*), GFP_KERNEL
);
8661 if (!d
->sched_group_nodes
) {
8662 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8663 return sa_notcovered
;
8665 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8667 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8668 return sa_sched_group_nodes
;
8669 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8671 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8672 return sa_this_sibling_map
;
8673 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8674 return sa_this_core_map
;
8675 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8676 return sa_send_covered
;
8677 d
->rd
= alloc_rootdomain();
8679 printk(KERN_WARNING
"Cannot alloc root domain\n");
8682 return sa_rootdomain
;
8685 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8686 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8688 struct sched_domain
*sd
= NULL
;
8690 struct sched_domain
*parent
;
8693 if (cpumask_weight(cpu_map
) >
8694 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8695 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8696 SD_INIT(sd
, ALLNODES
);
8697 set_domain_attribute(sd
, attr
);
8698 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8699 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8704 sd
= &per_cpu(node_domains
, i
).sd
;
8706 set_domain_attribute(sd
, attr
);
8707 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8708 sd
->parent
= parent
;
8711 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8716 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8717 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8718 struct sched_domain
*parent
, int i
)
8720 struct sched_domain
*sd
;
8721 sd
= &per_cpu(phys_domains
, i
).sd
;
8723 set_domain_attribute(sd
, attr
);
8724 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8725 sd
->parent
= parent
;
8728 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8732 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8733 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8734 struct sched_domain
*parent
, int i
)
8736 struct sched_domain
*sd
= parent
;
8737 #ifdef CONFIG_SCHED_MC
8738 sd
= &per_cpu(core_domains
, i
).sd
;
8740 set_domain_attribute(sd
, attr
);
8741 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8742 sd
->parent
= parent
;
8744 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8749 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8750 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8751 struct sched_domain
*parent
, int i
)
8753 struct sched_domain
*sd
= parent
;
8754 #ifdef CONFIG_SCHED_SMT
8755 sd
= &per_cpu(cpu_domains
, i
).sd
;
8756 SD_INIT(sd
, SIBLING
);
8757 set_domain_attribute(sd
, attr
);
8758 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8759 sd
->parent
= parent
;
8761 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8766 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8767 const struct cpumask
*cpu_map
, int cpu
)
8770 #ifdef CONFIG_SCHED_SMT
8771 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8772 cpumask_and(d
->this_sibling_map
, cpu_map
,
8773 topology_thread_cpumask(cpu
));
8774 if (cpu
== cpumask_first(d
->this_sibling_map
))
8775 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8777 d
->send_covered
, d
->tmpmask
);
8780 #ifdef CONFIG_SCHED_MC
8781 case SD_LV_MC
: /* set up multi-core groups */
8782 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8783 if (cpu
== cpumask_first(d
->this_core_map
))
8784 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8786 d
->send_covered
, d
->tmpmask
);
8789 case SD_LV_CPU
: /* set up physical groups */
8790 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8791 if (!cpumask_empty(d
->nodemask
))
8792 init_sched_build_groups(d
->nodemask
, cpu_map
,
8794 d
->send_covered
, d
->tmpmask
);
8797 case SD_LV_ALLNODES
:
8798 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8799 d
->send_covered
, d
->tmpmask
);
8808 * Build sched domains for a given set of cpus and attach the sched domains
8809 * to the individual cpus
8811 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8812 struct sched_domain_attr
*attr
)
8814 enum s_alloc alloc_state
= sa_none
;
8816 struct sched_domain
*sd
;
8822 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8823 if (alloc_state
!= sa_rootdomain
)
8825 alloc_state
= sa_sched_groups
;
8828 * Set up domains for cpus specified by the cpu_map.
8830 for_each_cpu(i
, cpu_map
) {
8831 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8834 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8835 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8836 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8837 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8840 for_each_cpu(i
, cpu_map
) {
8841 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8842 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8845 /* Set up physical groups */
8846 for (i
= 0; i
< nr_node_ids
; i
++)
8847 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8850 /* Set up node groups */
8852 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8854 for (i
= 0; i
< nr_node_ids
; i
++)
8855 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8859 /* Calculate CPU power for physical packages and nodes */
8860 #ifdef CONFIG_SCHED_SMT
8861 for_each_cpu(i
, cpu_map
) {
8862 sd
= &per_cpu(cpu_domains
, i
).sd
;
8863 init_sched_groups_power(i
, sd
);
8866 #ifdef CONFIG_SCHED_MC
8867 for_each_cpu(i
, cpu_map
) {
8868 sd
= &per_cpu(core_domains
, i
).sd
;
8869 init_sched_groups_power(i
, sd
);
8873 for_each_cpu(i
, cpu_map
) {
8874 sd
= &per_cpu(phys_domains
, i
).sd
;
8875 init_sched_groups_power(i
, sd
);
8879 for (i
= 0; i
< nr_node_ids
; i
++)
8880 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8882 if (d
.sd_allnodes
) {
8883 struct sched_group
*sg
;
8885 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8887 init_numa_sched_groups_power(sg
);
8891 /* Attach the domains */
8892 for_each_cpu(i
, cpu_map
) {
8893 #ifdef CONFIG_SCHED_SMT
8894 sd
= &per_cpu(cpu_domains
, i
).sd
;
8895 #elif defined(CONFIG_SCHED_MC)
8896 sd
= &per_cpu(core_domains
, i
).sd
;
8898 sd
= &per_cpu(phys_domains
, i
).sd
;
8900 cpu_attach_domain(sd
, d
.rd
, i
);
8903 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8904 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8908 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8912 static int build_sched_domains(const struct cpumask
*cpu_map
)
8914 return __build_sched_domains(cpu_map
, NULL
);
8917 static struct cpumask
*doms_cur
; /* current sched domains */
8918 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8919 static struct sched_domain_attr
*dattr_cur
;
8920 /* attribues of custom domains in 'doms_cur' */
8923 * Special case: If a kmalloc of a doms_cur partition (array of
8924 * cpumask) fails, then fallback to a single sched domain,
8925 * as determined by the single cpumask fallback_doms.
8927 static cpumask_var_t fallback_doms
;
8930 * arch_update_cpu_topology lets virtualized architectures update the
8931 * cpu core maps. It is supposed to return 1 if the topology changed
8932 * or 0 if it stayed the same.
8934 int __attribute__((weak
)) arch_update_cpu_topology(void)
8940 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8941 * For now this just excludes isolated cpus, but could be used to
8942 * exclude other special cases in the future.
8944 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8948 arch_update_cpu_topology();
8950 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8952 doms_cur
= fallback_doms
;
8953 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8955 err
= build_sched_domains(doms_cur
);
8956 register_sched_domain_sysctl();
8961 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8962 struct cpumask
*tmpmask
)
8964 free_sched_groups(cpu_map
, tmpmask
);
8968 * Detach sched domains from a group of cpus specified in cpu_map
8969 * These cpus will now be attached to the NULL domain
8971 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8973 /* Save because hotplug lock held. */
8974 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8977 for_each_cpu(i
, cpu_map
)
8978 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8979 synchronize_sched();
8980 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8983 /* handle null as "default" */
8984 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8985 struct sched_domain_attr
*new, int idx_new
)
8987 struct sched_domain_attr tmp
;
8994 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8995 new ? (new + idx_new
) : &tmp
,
8996 sizeof(struct sched_domain_attr
));
9000 * Partition sched domains as specified by the 'ndoms_new'
9001 * cpumasks in the array doms_new[] of cpumasks. This compares
9002 * doms_new[] to the current sched domain partitioning, doms_cur[].
9003 * It destroys each deleted domain and builds each new domain.
9005 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9006 * The masks don't intersect (don't overlap.) We should setup one
9007 * sched domain for each mask. CPUs not in any of the cpumasks will
9008 * not be load balanced. If the same cpumask appears both in the
9009 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9012 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9013 * ownership of it and will kfree it when done with it. If the caller
9014 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9015 * ndoms_new == 1, and partition_sched_domains() will fallback to
9016 * the single partition 'fallback_doms', it also forces the domains
9019 * If doms_new == NULL it will be replaced with cpu_online_mask.
9020 * ndoms_new == 0 is a special case for destroying existing domains,
9021 * and it will not create the default domain.
9023 * Call with hotplug lock held
9025 /* FIXME: Change to struct cpumask *doms_new[] */
9026 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9027 struct sched_domain_attr
*dattr_new
)
9032 mutex_lock(&sched_domains_mutex
);
9034 /* always unregister in case we don't destroy any domains */
9035 unregister_sched_domain_sysctl();
9037 /* Let architecture update cpu core mappings. */
9038 new_topology
= arch_update_cpu_topology();
9040 n
= doms_new
? ndoms_new
: 0;
9042 /* Destroy deleted domains */
9043 for (i
= 0; i
< ndoms_cur
; i
++) {
9044 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9045 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9046 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9049 /* no match - a current sched domain not in new doms_new[] */
9050 detach_destroy_domains(doms_cur
+ i
);
9055 if (doms_new
== NULL
) {
9057 doms_new
= fallback_doms
;
9058 cpumask_andnot(&doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
9059 WARN_ON_ONCE(dattr_new
);
9062 /* Build new domains */
9063 for (i
= 0; i
< ndoms_new
; i
++) {
9064 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9065 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9066 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9069 /* no match - add a new doms_new */
9070 __build_sched_domains(doms_new
+ i
,
9071 dattr_new
? dattr_new
+ i
: NULL
);
9076 /* Remember the new sched domains */
9077 if (doms_cur
!= fallback_doms
)
9079 kfree(dattr_cur
); /* kfree(NULL) is safe */
9080 doms_cur
= doms_new
;
9081 dattr_cur
= dattr_new
;
9082 ndoms_cur
= ndoms_new
;
9084 register_sched_domain_sysctl();
9086 mutex_unlock(&sched_domains_mutex
);
9089 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9090 static void arch_reinit_sched_domains(void)
9094 /* Destroy domains first to force the rebuild */
9095 partition_sched_domains(0, NULL
, NULL
);
9097 rebuild_sched_domains();
9101 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9103 unsigned int level
= 0;
9105 if (sscanf(buf
, "%u", &level
) != 1)
9109 * level is always be positive so don't check for
9110 * level < POWERSAVINGS_BALANCE_NONE which is 0
9111 * What happens on 0 or 1 byte write,
9112 * need to check for count as well?
9115 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9119 sched_smt_power_savings
= level
;
9121 sched_mc_power_savings
= level
;
9123 arch_reinit_sched_domains();
9128 #ifdef CONFIG_SCHED_MC
9129 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9132 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9134 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9135 const char *buf
, size_t count
)
9137 return sched_power_savings_store(buf
, count
, 0);
9139 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9140 sched_mc_power_savings_show
,
9141 sched_mc_power_savings_store
);
9144 #ifdef CONFIG_SCHED_SMT
9145 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9148 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9150 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9151 const char *buf
, size_t count
)
9153 return sched_power_savings_store(buf
, count
, 1);
9155 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9156 sched_smt_power_savings_show
,
9157 sched_smt_power_savings_store
);
9160 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9164 #ifdef CONFIG_SCHED_SMT
9166 err
= sysfs_create_file(&cls
->kset
.kobj
,
9167 &attr_sched_smt_power_savings
.attr
);
9169 #ifdef CONFIG_SCHED_MC
9170 if (!err
&& mc_capable())
9171 err
= sysfs_create_file(&cls
->kset
.kobj
,
9172 &attr_sched_mc_power_savings
.attr
);
9176 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9178 #ifndef CONFIG_CPUSETS
9180 * Add online and remove offline CPUs from the scheduler domains.
9181 * When cpusets are enabled they take over this function.
9183 static int update_sched_domains(struct notifier_block
*nfb
,
9184 unsigned long action
, void *hcpu
)
9188 case CPU_ONLINE_FROZEN
:
9189 case CPU_DOWN_PREPARE
:
9190 case CPU_DOWN_PREPARE_FROZEN
:
9191 case CPU_DOWN_FAILED
:
9192 case CPU_DOWN_FAILED_FROZEN
:
9193 partition_sched_domains(1, NULL
, NULL
);
9202 static int update_runtime(struct notifier_block
*nfb
,
9203 unsigned long action
, void *hcpu
)
9205 int cpu
= (int)(long)hcpu
;
9208 case CPU_DOWN_PREPARE
:
9209 case CPU_DOWN_PREPARE_FROZEN
:
9210 disable_runtime(cpu_rq(cpu
));
9213 case CPU_DOWN_FAILED
:
9214 case CPU_DOWN_FAILED_FROZEN
:
9216 case CPU_ONLINE_FROZEN
:
9217 enable_runtime(cpu_rq(cpu
));
9225 void __init
sched_init_smp(void)
9227 cpumask_var_t non_isolated_cpus
;
9229 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9230 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9232 #if defined(CONFIG_NUMA)
9233 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9235 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9238 mutex_lock(&sched_domains_mutex
);
9239 arch_init_sched_domains(cpu_active_mask
);
9240 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9241 if (cpumask_empty(non_isolated_cpus
))
9242 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9243 mutex_unlock(&sched_domains_mutex
);
9246 #ifndef CONFIG_CPUSETS
9247 /* XXX: Theoretical race here - CPU may be hotplugged now */
9248 hotcpu_notifier(update_sched_domains
, 0);
9251 /* RT runtime code needs to handle some hotplug events */
9252 hotcpu_notifier(update_runtime
, 0);
9256 /* Move init over to a non-isolated CPU */
9257 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9259 sched_init_granularity();
9260 free_cpumask_var(non_isolated_cpus
);
9262 init_sched_rt_class();
9265 void __init
sched_init_smp(void)
9267 sched_init_granularity();
9269 #endif /* CONFIG_SMP */
9271 const_debug
unsigned int sysctl_timer_migration
= 1;
9273 int in_sched_functions(unsigned long addr
)
9275 return in_lock_functions(addr
) ||
9276 (addr
>= (unsigned long)__sched_text_start
9277 && addr
< (unsigned long)__sched_text_end
);
9280 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9282 cfs_rq
->tasks_timeline
= RB_ROOT
;
9283 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9284 #ifdef CONFIG_FAIR_GROUP_SCHED
9287 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9290 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9292 struct rt_prio_array
*array
;
9295 array
= &rt_rq
->active
;
9296 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9297 INIT_LIST_HEAD(array
->queue
+ i
);
9298 __clear_bit(i
, array
->bitmap
);
9300 /* delimiter for bitsearch: */
9301 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9303 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9304 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9306 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9310 rt_rq
->rt_nr_migratory
= 0;
9311 rt_rq
->overloaded
= 0;
9312 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9316 rt_rq
->rt_throttled
= 0;
9317 rt_rq
->rt_runtime
= 0;
9318 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9320 #ifdef CONFIG_RT_GROUP_SCHED
9321 rt_rq
->rt_nr_boosted
= 0;
9326 #ifdef CONFIG_FAIR_GROUP_SCHED
9327 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9328 struct sched_entity
*se
, int cpu
, int add
,
9329 struct sched_entity
*parent
)
9331 struct rq
*rq
= cpu_rq(cpu
);
9332 tg
->cfs_rq
[cpu
] = cfs_rq
;
9333 init_cfs_rq(cfs_rq
, rq
);
9336 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9339 /* se could be NULL for init_task_group */
9344 se
->cfs_rq
= &rq
->cfs
;
9346 se
->cfs_rq
= parent
->my_q
;
9349 se
->load
.weight
= tg
->shares
;
9350 se
->load
.inv_weight
= 0;
9351 se
->parent
= parent
;
9355 #ifdef CONFIG_RT_GROUP_SCHED
9356 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9357 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9358 struct sched_rt_entity
*parent
)
9360 struct rq
*rq
= cpu_rq(cpu
);
9362 tg
->rt_rq
[cpu
] = rt_rq
;
9363 init_rt_rq(rt_rq
, rq
);
9365 rt_rq
->rt_se
= rt_se
;
9366 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9368 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9370 tg
->rt_se
[cpu
] = rt_se
;
9375 rt_se
->rt_rq
= &rq
->rt
;
9377 rt_se
->rt_rq
= parent
->my_q
;
9379 rt_se
->my_q
= rt_rq
;
9380 rt_se
->parent
= parent
;
9381 INIT_LIST_HEAD(&rt_se
->run_list
);
9385 void __init
sched_init(void)
9388 unsigned long alloc_size
= 0, ptr
;
9390 #ifdef CONFIG_FAIR_GROUP_SCHED
9391 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9393 #ifdef CONFIG_RT_GROUP_SCHED
9394 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9396 #ifdef CONFIG_USER_SCHED
9399 #ifdef CONFIG_CPUMASK_OFFSTACK
9400 alloc_size
+= num_possible_cpus() * cpumask_size();
9403 * As sched_init() is called before page_alloc is setup,
9404 * we use alloc_bootmem().
9407 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9409 #ifdef CONFIG_FAIR_GROUP_SCHED
9410 init_task_group
.se
= (struct sched_entity
**)ptr
;
9411 ptr
+= nr_cpu_ids
* sizeof(void **);
9413 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9414 ptr
+= nr_cpu_ids
* sizeof(void **);
9416 #ifdef CONFIG_USER_SCHED
9417 root_task_group
.se
= (struct sched_entity
**)ptr
;
9418 ptr
+= nr_cpu_ids
* sizeof(void **);
9420 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9421 ptr
+= nr_cpu_ids
* sizeof(void **);
9422 #endif /* CONFIG_USER_SCHED */
9423 #endif /* CONFIG_FAIR_GROUP_SCHED */
9424 #ifdef CONFIG_RT_GROUP_SCHED
9425 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9426 ptr
+= nr_cpu_ids
* sizeof(void **);
9428 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9429 ptr
+= nr_cpu_ids
* sizeof(void **);
9431 #ifdef CONFIG_USER_SCHED
9432 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9433 ptr
+= nr_cpu_ids
* sizeof(void **);
9435 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9436 ptr
+= nr_cpu_ids
* sizeof(void **);
9437 #endif /* CONFIG_USER_SCHED */
9438 #endif /* CONFIG_RT_GROUP_SCHED */
9439 #ifdef CONFIG_CPUMASK_OFFSTACK
9440 for_each_possible_cpu(i
) {
9441 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9442 ptr
+= cpumask_size();
9444 #endif /* CONFIG_CPUMASK_OFFSTACK */
9448 init_defrootdomain();
9451 init_rt_bandwidth(&def_rt_bandwidth
,
9452 global_rt_period(), global_rt_runtime());
9454 #ifdef CONFIG_RT_GROUP_SCHED
9455 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9456 global_rt_period(), global_rt_runtime());
9457 #ifdef CONFIG_USER_SCHED
9458 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9459 global_rt_period(), RUNTIME_INF
);
9460 #endif /* CONFIG_USER_SCHED */
9461 #endif /* CONFIG_RT_GROUP_SCHED */
9463 #ifdef CONFIG_GROUP_SCHED
9464 list_add(&init_task_group
.list
, &task_groups
);
9465 INIT_LIST_HEAD(&init_task_group
.children
);
9467 #ifdef CONFIG_USER_SCHED
9468 INIT_LIST_HEAD(&root_task_group
.children
);
9469 init_task_group
.parent
= &root_task_group
;
9470 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9471 #endif /* CONFIG_USER_SCHED */
9472 #endif /* CONFIG_GROUP_SCHED */
9474 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9475 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9476 __alignof__(unsigned long));
9478 for_each_possible_cpu(i
) {
9482 spin_lock_init(&rq
->lock
);
9484 rq
->calc_load_active
= 0;
9485 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9486 init_cfs_rq(&rq
->cfs
, rq
);
9487 init_rt_rq(&rq
->rt
, rq
);
9488 #ifdef CONFIG_FAIR_GROUP_SCHED
9489 init_task_group
.shares
= init_task_group_load
;
9490 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9491 #ifdef CONFIG_CGROUP_SCHED
9493 * How much cpu bandwidth does init_task_group get?
9495 * In case of task-groups formed thr' the cgroup filesystem, it
9496 * gets 100% of the cpu resources in the system. This overall
9497 * system cpu resource is divided among the tasks of
9498 * init_task_group and its child task-groups in a fair manner,
9499 * based on each entity's (task or task-group's) weight
9500 * (se->load.weight).
9502 * In other words, if init_task_group has 10 tasks of weight
9503 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9504 * then A0's share of the cpu resource is:
9506 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9508 * We achieve this by letting init_task_group's tasks sit
9509 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9511 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9512 #elif defined CONFIG_USER_SCHED
9513 root_task_group
.shares
= NICE_0_LOAD
;
9514 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9516 * In case of task-groups formed thr' the user id of tasks,
9517 * init_task_group represents tasks belonging to root user.
9518 * Hence it forms a sibling of all subsequent groups formed.
9519 * In this case, init_task_group gets only a fraction of overall
9520 * system cpu resource, based on the weight assigned to root
9521 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9522 * by letting tasks of init_task_group sit in a separate cfs_rq
9523 * (init_tg_cfs_rq) and having one entity represent this group of
9524 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9526 init_tg_cfs_entry(&init_task_group
,
9527 &per_cpu(init_tg_cfs_rq
, i
),
9528 &per_cpu(init_sched_entity
, i
), i
, 1,
9529 root_task_group
.se
[i
]);
9532 #endif /* CONFIG_FAIR_GROUP_SCHED */
9534 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9535 #ifdef CONFIG_RT_GROUP_SCHED
9536 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9537 #ifdef CONFIG_CGROUP_SCHED
9538 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9539 #elif defined CONFIG_USER_SCHED
9540 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9541 init_tg_rt_entry(&init_task_group
,
9542 &per_cpu(init_rt_rq
, i
),
9543 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9544 root_task_group
.rt_se
[i
]);
9548 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9549 rq
->cpu_load
[j
] = 0;
9553 rq
->post_schedule
= 0;
9554 rq
->active_balance
= 0;
9555 rq
->next_balance
= jiffies
;
9559 rq
->migration_thread
= NULL
;
9561 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9562 INIT_LIST_HEAD(&rq
->migration_queue
);
9563 rq_attach_root(rq
, &def_root_domain
);
9566 atomic_set(&rq
->nr_iowait
, 0);
9569 set_load_weight(&init_task
);
9571 #ifdef CONFIG_PREEMPT_NOTIFIERS
9572 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9576 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9579 #ifdef CONFIG_RT_MUTEXES
9580 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9584 * The boot idle thread does lazy MMU switching as well:
9586 atomic_inc(&init_mm
.mm_count
);
9587 enter_lazy_tlb(&init_mm
, current
);
9590 * Make us the idle thread. Technically, schedule() should not be
9591 * called from this thread, however somewhere below it might be,
9592 * but because we are the idle thread, we just pick up running again
9593 * when this runqueue becomes "idle".
9595 init_idle(current
, smp_processor_id());
9597 calc_load_update
= jiffies
+ LOAD_FREQ
;
9600 * During early bootup we pretend to be a normal task:
9602 current
->sched_class
= &fair_sched_class
;
9604 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9605 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9608 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9609 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9611 /* May be allocated at isolcpus cmdline parse time */
9612 if (cpu_isolated_map
== NULL
)
9613 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9618 scheduler_running
= 1;
9621 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9622 static inline int preempt_count_equals(int preempt_offset
)
9624 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9626 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9629 void __might_sleep(char *file
, int line
, int preempt_offset
)
9632 static unsigned long prev_jiffy
; /* ratelimiting */
9634 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9635 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9637 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9639 prev_jiffy
= jiffies
;
9642 "BUG: sleeping function called from invalid context at %s:%d\n",
9645 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9646 in_atomic(), irqs_disabled(),
9647 current
->pid
, current
->comm
);
9649 debug_show_held_locks(current
);
9650 if (irqs_disabled())
9651 print_irqtrace_events(current
);
9655 EXPORT_SYMBOL(__might_sleep
);
9658 #ifdef CONFIG_MAGIC_SYSRQ
9659 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9663 update_rq_clock(rq
);
9664 on_rq
= p
->se
.on_rq
;
9666 deactivate_task(rq
, p
, 0);
9667 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9669 activate_task(rq
, p
, 0);
9670 resched_task(rq
->curr
);
9674 void normalize_rt_tasks(void)
9676 struct task_struct
*g
, *p
;
9677 unsigned long flags
;
9680 read_lock_irqsave(&tasklist_lock
, flags
);
9681 do_each_thread(g
, p
) {
9683 * Only normalize user tasks:
9688 p
->se
.exec_start
= 0;
9689 #ifdef CONFIG_SCHEDSTATS
9690 p
->se
.wait_start
= 0;
9691 p
->se
.sleep_start
= 0;
9692 p
->se
.block_start
= 0;
9697 * Renice negative nice level userspace
9700 if (TASK_NICE(p
) < 0 && p
->mm
)
9701 set_user_nice(p
, 0);
9705 spin_lock(&p
->pi_lock
);
9706 rq
= __task_rq_lock(p
);
9708 normalize_task(rq
, p
);
9710 __task_rq_unlock(rq
);
9711 spin_unlock(&p
->pi_lock
);
9712 } while_each_thread(g
, p
);
9714 read_unlock_irqrestore(&tasklist_lock
, flags
);
9717 #endif /* CONFIG_MAGIC_SYSRQ */
9721 * These functions are only useful for the IA64 MCA handling.
9723 * They can only be called when the whole system has been
9724 * stopped - every CPU needs to be quiescent, and no scheduling
9725 * activity can take place. Using them for anything else would
9726 * be a serious bug, and as a result, they aren't even visible
9727 * under any other configuration.
9731 * curr_task - return the current task for a given cpu.
9732 * @cpu: the processor in question.
9734 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9736 struct task_struct
*curr_task(int cpu
)
9738 return cpu_curr(cpu
);
9742 * set_curr_task - set the current task for a given cpu.
9743 * @cpu: the processor in question.
9744 * @p: the task pointer to set.
9746 * Description: This function must only be used when non-maskable interrupts
9747 * are serviced on a separate stack. It allows the architecture to switch the
9748 * notion of the current task on a cpu in a non-blocking manner. This function
9749 * must be called with all CPU's synchronized, and interrupts disabled, the
9750 * and caller must save the original value of the current task (see
9751 * curr_task() above) and restore that value before reenabling interrupts and
9752 * re-starting the system.
9754 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9756 void set_curr_task(int cpu
, struct task_struct
*p
)
9763 #ifdef CONFIG_FAIR_GROUP_SCHED
9764 static void free_fair_sched_group(struct task_group
*tg
)
9768 for_each_possible_cpu(i
) {
9770 kfree(tg
->cfs_rq
[i
]);
9780 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9782 struct cfs_rq
*cfs_rq
;
9783 struct sched_entity
*se
;
9787 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9790 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9794 tg
->shares
= NICE_0_LOAD
;
9796 for_each_possible_cpu(i
) {
9799 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9800 GFP_KERNEL
, cpu_to_node(i
));
9804 se
= kzalloc_node(sizeof(struct sched_entity
),
9805 GFP_KERNEL
, cpu_to_node(i
));
9809 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9818 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9820 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9821 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9824 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9826 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9828 #else /* !CONFG_FAIR_GROUP_SCHED */
9829 static inline void free_fair_sched_group(struct task_group
*tg
)
9834 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9839 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9843 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9846 #endif /* CONFIG_FAIR_GROUP_SCHED */
9848 #ifdef CONFIG_RT_GROUP_SCHED
9849 static void free_rt_sched_group(struct task_group
*tg
)
9853 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9855 for_each_possible_cpu(i
) {
9857 kfree(tg
->rt_rq
[i
]);
9859 kfree(tg
->rt_se
[i
]);
9867 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9869 struct rt_rq
*rt_rq
;
9870 struct sched_rt_entity
*rt_se
;
9874 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9877 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9881 init_rt_bandwidth(&tg
->rt_bandwidth
,
9882 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9884 for_each_possible_cpu(i
) {
9887 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9888 GFP_KERNEL
, cpu_to_node(i
));
9892 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9893 GFP_KERNEL
, cpu_to_node(i
));
9897 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9906 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9908 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9909 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9912 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9914 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9916 #else /* !CONFIG_RT_GROUP_SCHED */
9917 static inline void free_rt_sched_group(struct task_group
*tg
)
9922 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9927 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9931 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9934 #endif /* CONFIG_RT_GROUP_SCHED */
9936 #ifdef CONFIG_GROUP_SCHED
9937 static void free_sched_group(struct task_group
*tg
)
9939 free_fair_sched_group(tg
);
9940 free_rt_sched_group(tg
);
9944 /* allocate runqueue etc for a new task group */
9945 struct task_group
*sched_create_group(struct task_group
*parent
)
9947 struct task_group
*tg
;
9948 unsigned long flags
;
9951 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9953 return ERR_PTR(-ENOMEM
);
9955 if (!alloc_fair_sched_group(tg
, parent
))
9958 if (!alloc_rt_sched_group(tg
, parent
))
9961 spin_lock_irqsave(&task_group_lock
, flags
);
9962 for_each_possible_cpu(i
) {
9963 register_fair_sched_group(tg
, i
);
9964 register_rt_sched_group(tg
, i
);
9966 list_add_rcu(&tg
->list
, &task_groups
);
9968 WARN_ON(!parent
); /* root should already exist */
9970 tg
->parent
= parent
;
9971 INIT_LIST_HEAD(&tg
->children
);
9972 list_add_rcu(&tg
->siblings
, &parent
->children
);
9973 spin_unlock_irqrestore(&task_group_lock
, flags
);
9978 free_sched_group(tg
);
9979 return ERR_PTR(-ENOMEM
);
9982 /* rcu callback to free various structures associated with a task group */
9983 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9985 /* now it should be safe to free those cfs_rqs */
9986 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9989 /* Destroy runqueue etc associated with a task group */
9990 void sched_destroy_group(struct task_group
*tg
)
9992 unsigned long flags
;
9995 spin_lock_irqsave(&task_group_lock
, flags
);
9996 for_each_possible_cpu(i
) {
9997 unregister_fair_sched_group(tg
, i
);
9998 unregister_rt_sched_group(tg
, i
);
10000 list_del_rcu(&tg
->list
);
10001 list_del_rcu(&tg
->siblings
);
10002 spin_unlock_irqrestore(&task_group_lock
, flags
);
10004 /* wait for possible concurrent references to cfs_rqs complete */
10005 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10008 /* change task's runqueue when it moves between groups.
10009 * The caller of this function should have put the task in its new group
10010 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10011 * reflect its new group.
10013 void sched_move_task(struct task_struct
*tsk
)
10015 int on_rq
, running
;
10016 unsigned long flags
;
10019 rq
= task_rq_lock(tsk
, &flags
);
10021 update_rq_clock(rq
);
10023 running
= task_current(rq
, tsk
);
10024 on_rq
= tsk
->se
.on_rq
;
10027 dequeue_task(rq
, tsk
, 0);
10028 if (unlikely(running
))
10029 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10031 set_task_rq(tsk
, task_cpu(tsk
));
10033 #ifdef CONFIG_FAIR_GROUP_SCHED
10034 if (tsk
->sched_class
->moved_group
)
10035 tsk
->sched_class
->moved_group(tsk
);
10038 if (unlikely(running
))
10039 tsk
->sched_class
->set_curr_task(rq
);
10041 enqueue_task(rq
, tsk
, 0);
10043 task_rq_unlock(rq
, &flags
);
10045 #endif /* CONFIG_GROUP_SCHED */
10047 #ifdef CONFIG_FAIR_GROUP_SCHED
10048 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10050 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10055 dequeue_entity(cfs_rq
, se
, 0);
10057 se
->load
.weight
= shares
;
10058 se
->load
.inv_weight
= 0;
10061 enqueue_entity(cfs_rq
, se
, 0);
10064 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10066 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10067 struct rq
*rq
= cfs_rq
->rq
;
10068 unsigned long flags
;
10070 spin_lock_irqsave(&rq
->lock
, flags
);
10071 __set_se_shares(se
, shares
);
10072 spin_unlock_irqrestore(&rq
->lock
, flags
);
10075 static DEFINE_MUTEX(shares_mutex
);
10077 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10080 unsigned long flags
;
10083 * We can't change the weight of the root cgroup.
10088 if (shares
< MIN_SHARES
)
10089 shares
= MIN_SHARES
;
10090 else if (shares
> MAX_SHARES
)
10091 shares
= MAX_SHARES
;
10093 mutex_lock(&shares_mutex
);
10094 if (tg
->shares
== shares
)
10097 spin_lock_irqsave(&task_group_lock
, flags
);
10098 for_each_possible_cpu(i
)
10099 unregister_fair_sched_group(tg
, i
);
10100 list_del_rcu(&tg
->siblings
);
10101 spin_unlock_irqrestore(&task_group_lock
, flags
);
10103 /* wait for any ongoing reference to this group to finish */
10104 synchronize_sched();
10107 * Now we are free to modify the group's share on each cpu
10108 * w/o tripping rebalance_share or load_balance_fair.
10110 tg
->shares
= shares
;
10111 for_each_possible_cpu(i
) {
10113 * force a rebalance
10115 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10116 set_se_shares(tg
->se
[i
], shares
);
10120 * Enable load balance activity on this group, by inserting it back on
10121 * each cpu's rq->leaf_cfs_rq_list.
10123 spin_lock_irqsave(&task_group_lock
, flags
);
10124 for_each_possible_cpu(i
)
10125 register_fair_sched_group(tg
, i
);
10126 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10127 spin_unlock_irqrestore(&task_group_lock
, flags
);
10129 mutex_unlock(&shares_mutex
);
10133 unsigned long sched_group_shares(struct task_group
*tg
)
10139 #ifdef CONFIG_RT_GROUP_SCHED
10141 * Ensure that the real time constraints are schedulable.
10143 static DEFINE_MUTEX(rt_constraints_mutex
);
10145 static unsigned long to_ratio(u64 period
, u64 runtime
)
10147 if (runtime
== RUNTIME_INF
)
10150 return div64_u64(runtime
<< 20, period
);
10153 /* Must be called with tasklist_lock held */
10154 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10156 struct task_struct
*g
, *p
;
10158 do_each_thread(g
, p
) {
10159 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10161 } while_each_thread(g
, p
);
10166 struct rt_schedulable_data
{
10167 struct task_group
*tg
;
10172 static int tg_schedulable(struct task_group
*tg
, void *data
)
10174 struct rt_schedulable_data
*d
= data
;
10175 struct task_group
*child
;
10176 unsigned long total
, sum
= 0;
10177 u64 period
, runtime
;
10179 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10180 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10183 period
= d
->rt_period
;
10184 runtime
= d
->rt_runtime
;
10187 #ifdef CONFIG_USER_SCHED
10188 if (tg
== &root_task_group
) {
10189 period
= global_rt_period();
10190 runtime
= global_rt_runtime();
10195 * Cannot have more runtime than the period.
10197 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10201 * Ensure we don't starve existing RT tasks.
10203 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10206 total
= to_ratio(period
, runtime
);
10209 * Nobody can have more than the global setting allows.
10211 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10215 * The sum of our children's runtime should not exceed our own.
10217 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10218 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10219 runtime
= child
->rt_bandwidth
.rt_runtime
;
10221 if (child
== d
->tg
) {
10222 period
= d
->rt_period
;
10223 runtime
= d
->rt_runtime
;
10226 sum
+= to_ratio(period
, runtime
);
10235 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10237 struct rt_schedulable_data data
= {
10239 .rt_period
= period
,
10240 .rt_runtime
= runtime
,
10243 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10246 static int tg_set_bandwidth(struct task_group
*tg
,
10247 u64 rt_period
, u64 rt_runtime
)
10251 mutex_lock(&rt_constraints_mutex
);
10252 read_lock(&tasklist_lock
);
10253 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10257 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10258 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10259 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10261 for_each_possible_cpu(i
) {
10262 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10264 spin_lock(&rt_rq
->rt_runtime_lock
);
10265 rt_rq
->rt_runtime
= rt_runtime
;
10266 spin_unlock(&rt_rq
->rt_runtime_lock
);
10268 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10270 read_unlock(&tasklist_lock
);
10271 mutex_unlock(&rt_constraints_mutex
);
10276 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10278 u64 rt_runtime
, rt_period
;
10280 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10281 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10282 if (rt_runtime_us
< 0)
10283 rt_runtime
= RUNTIME_INF
;
10285 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10288 long sched_group_rt_runtime(struct task_group
*tg
)
10292 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10295 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10296 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10297 return rt_runtime_us
;
10300 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10302 u64 rt_runtime
, rt_period
;
10304 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10305 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10307 if (rt_period
== 0)
10310 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10313 long sched_group_rt_period(struct task_group
*tg
)
10317 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10318 do_div(rt_period_us
, NSEC_PER_USEC
);
10319 return rt_period_us
;
10322 static int sched_rt_global_constraints(void)
10324 u64 runtime
, period
;
10327 if (sysctl_sched_rt_period
<= 0)
10330 runtime
= global_rt_runtime();
10331 period
= global_rt_period();
10334 * Sanity check on the sysctl variables.
10336 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10339 mutex_lock(&rt_constraints_mutex
);
10340 read_lock(&tasklist_lock
);
10341 ret
= __rt_schedulable(NULL
, 0, 0);
10342 read_unlock(&tasklist_lock
);
10343 mutex_unlock(&rt_constraints_mutex
);
10348 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10350 /* Don't accept realtime tasks when there is no way for them to run */
10351 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10357 #else /* !CONFIG_RT_GROUP_SCHED */
10358 static int sched_rt_global_constraints(void)
10360 unsigned long flags
;
10363 if (sysctl_sched_rt_period
<= 0)
10367 * There's always some RT tasks in the root group
10368 * -- migration, kstopmachine etc..
10370 if (sysctl_sched_rt_runtime
== 0)
10373 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10374 for_each_possible_cpu(i
) {
10375 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10377 spin_lock(&rt_rq
->rt_runtime_lock
);
10378 rt_rq
->rt_runtime
= global_rt_runtime();
10379 spin_unlock(&rt_rq
->rt_runtime_lock
);
10381 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10385 #endif /* CONFIG_RT_GROUP_SCHED */
10387 int sched_rt_handler(struct ctl_table
*table
, int write
,
10388 void __user
*buffer
, size_t *lenp
,
10392 int old_period
, old_runtime
;
10393 static DEFINE_MUTEX(mutex
);
10395 mutex_lock(&mutex
);
10396 old_period
= sysctl_sched_rt_period
;
10397 old_runtime
= sysctl_sched_rt_runtime
;
10399 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10401 if (!ret
&& write
) {
10402 ret
= sched_rt_global_constraints();
10404 sysctl_sched_rt_period
= old_period
;
10405 sysctl_sched_rt_runtime
= old_runtime
;
10407 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10408 def_rt_bandwidth
.rt_period
=
10409 ns_to_ktime(global_rt_period());
10412 mutex_unlock(&mutex
);
10417 #ifdef CONFIG_CGROUP_SCHED
10419 /* return corresponding task_group object of a cgroup */
10420 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10422 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10423 struct task_group
, css
);
10426 static struct cgroup_subsys_state
*
10427 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10429 struct task_group
*tg
, *parent
;
10431 if (!cgrp
->parent
) {
10432 /* This is early initialization for the top cgroup */
10433 return &init_task_group
.css
;
10436 parent
= cgroup_tg(cgrp
->parent
);
10437 tg
= sched_create_group(parent
);
10439 return ERR_PTR(-ENOMEM
);
10445 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10447 struct task_group
*tg
= cgroup_tg(cgrp
);
10449 sched_destroy_group(tg
);
10453 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10455 #ifdef CONFIG_RT_GROUP_SCHED
10456 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10459 /* We don't support RT-tasks being in separate groups */
10460 if (tsk
->sched_class
!= &fair_sched_class
)
10467 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10468 struct task_struct
*tsk
, bool threadgroup
)
10470 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10474 struct task_struct
*c
;
10476 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10477 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10489 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10490 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10493 sched_move_task(tsk
);
10495 struct task_struct
*c
;
10497 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10498 sched_move_task(c
);
10504 #ifdef CONFIG_FAIR_GROUP_SCHED
10505 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10508 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10511 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10513 struct task_group
*tg
= cgroup_tg(cgrp
);
10515 return (u64
) tg
->shares
;
10517 #endif /* CONFIG_FAIR_GROUP_SCHED */
10519 #ifdef CONFIG_RT_GROUP_SCHED
10520 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10523 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10526 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10528 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10531 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10534 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10537 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10539 return sched_group_rt_period(cgroup_tg(cgrp
));
10541 #endif /* CONFIG_RT_GROUP_SCHED */
10543 static struct cftype cpu_files
[] = {
10544 #ifdef CONFIG_FAIR_GROUP_SCHED
10547 .read_u64
= cpu_shares_read_u64
,
10548 .write_u64
= cpu_shares_write_u64
,
10551 #ifdef CONFIG_RT_GROUP_SCHED
10553 .name
= "rt_runtime_us",
10554 .read_s64
= cpu_rt_runtime_read
,
10555 .write_s64
= cpu_rt_runtime_write
,
10558 .name
= "rt_period_us",
10559 .read_u64
= cpu_rt_period_read_uint
,
10560 .write_u64
= cpu_rt_period_write_uint
,
10565 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10567 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10570 struct cgroup_subsys cpu_cgroup_subsys
= {
10572 .create
= cpu_cgroup_create
,
10573 .destroy
= cpu_cgroup_destroy
,
10574 .can_attach
= cpu_cgroup_can_attach
,
10575 .attach
= cpu_cgroup_attach
,
10576 .populate
= cpu_cgroup_populate
,
10577 .subsys_id
= cpu_cgroup_subsys_id
,
10581 #endif /* CONFIG_CGROUP_SCHED */
10583 #ifdef CONFIG_CGROUP_CPUACCT
10586 * CPU accounting code for task groups.
10588 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10589 * (balbir@in.ibm.com).
10592 /* track cpu usage of a group of tasks and its child groups */
10594 struct cgroup_subsys_state css
;
10595 /* cpuusage holds pointer to a u64-type object on every cpu */
10597 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10598 struct cpuacct
*parent
;
10601 struct cgroup_subsys cpuacct_subsys
;
10603 /* return cpu accounting group corresponding to this container */
10604 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10606 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10607 struct cpuacct
, css
);
10610 /* return cpu accounting group to which this task belongs */
10611 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10613 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10614 struct cpuacct
, css
);
10617 /* create a new cpu accounting group */
10618 static struct cgroup_subsys_state
*cpuacct_create(
10619 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10621 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10627 ca
->cpuusage
= alloc_percpu(u64
);
10631 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10632 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10633 goto out_free_counters
;
10636 ca
->parent
= cgroup_ca(cgrp
->parent
);
10642 percpu_counter_destroy(&ca
->cpustat
[i
]);
10643 free_percpu(ca
->cpuusage
);
10647 return ERR_PTR(-ENOMEM
);
10650 /* destroy an existing cpu accounting group */
10652 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10654 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10657 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10658 percpu_counter_destroy(&ca
->cpustat
[i
]);
10659 free_percpu(ca
->cpuusage
);
10663 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10665 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10668 #ifndef CONFIG_64BIT
10670 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10672 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10674 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10682 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10684 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10686 #ifndef CONFIG_64BIT
10688 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10690 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10692 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10698 /* return total cpu usage (in nanoseconds) of a group */
10699 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10701 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10702 u64 totalcpuusage
= 0;
10705 for_each_present_cpu(i
)
10706 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10708 return totalcpuusage
;
10711 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10714 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10723 for_each_present_cpu(i
)
10724 cpuacct_cpuusage_write(ca
, i
, 0);
10730 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10731 struct seq_file
*m
)
10733 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10737 for_each_present_cpu(i
) {
10738 percpu
= cpuacct_cpuusage_read(ca
, i
);
10739 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10741 seq_printf(m
, "\n");
10745 static const char *cpuacct_stat_desc
[] = {
10746 [CPUACCT_STAT_USER
] = "user",
10747 [CPUACCT_STAT_SYSTEM
] = "system",
10750 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10751 struct cgroup_map_cb
*cb
)
10753 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10756 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10757 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10758 val
= cputime64_to_clock_t(val
);
10759 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10764 static struct cftype files
[] = {
10767 .read_u64
= cpuusage_read
,
10768 .write_u64
= cpuusage_write
,
10771 .name
= "usage_percpu",
10772 .read_seq_string
= cpuacct_percpu_seq_read
,
10776 .read_map
= cpuacct_stats_show
,
10780 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10782 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10786 * charge this task's execution time to its accounting group.
10788 * called with rq->lock held.
10790 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10792 struct cpuacct
*ca
;
10795 if (unlikely(!cpuacct_subsys
.active
))
10798 cpu
= task_cpu(tsk
);
10804 for (; ca
; ca
= ca
->parent
) {
10805 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10806 *cpuusage
+= cputime
;
10813 * Charge the system/user time to the task's accounting group.
10815 static void cpuacct_update_stats(struct task_struct
*tsk
,
10816 enum cpuacct_stat_index idx
, cputime_t val
)
10818 struct cpuacct
*ca
;
10820 if (unlikely(!cpuacct_subsys
.active
))
10827 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10833 struct cgroup_subsys cpuacct_subsys
= {
10835 .create
= cpuacct_create
,
10836 .destroy
= cpuacct_destroy
,
10837 .populate
= cpuacct_populate
,
10838 .subsys_id
= cpuacct_subsys_id
,
10840 #endif /* CONFIG_CGROUP_CPUACCT */
10844 int rcu_expedited_torture_stats(char *page
)
10848 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10850 void synchronize_sched_expedited(void)
10853 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10855 #else /* #ifndef CONFIG_SMP */
10857 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10858 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10860 #define RCU_EXPEDITED_STATE_POST -2
10861 #define RCU_EXPEDITED_STATE_IDLE -1
10863 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10865 int rcu_expedited_torture_stats(char *page
)
10870 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10871 for_each_online_cpu(cpu
) {
10872 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10873 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10875 cnt
+= sprintf(&page
[cnt
], "\n");
10878 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10880 static long synchronize_sched_expedited_count
;
10883 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10884 * approach to force grace period to end quickly. This consumes
10885 * significant time on all CPUs, and is thus not recommended for
10886 * any sort of common-case code.
10888 * Note that it is illegal to call this function while holding any
10889 * lock that is acquired by a CPU-hotplug notifier. Failing to
10890 * observe this restriction will result in deadlock.
10892 void synchronize_sched_expedited(void)
10895 unsigned long flags
;
10896 bool need_full_sync
= 0;
10898 struct migration_req
*req
;
10902 smp_mb(); /* ensure prior mod happens before capturing snap. */
10903 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10905 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10907 if (trycount
++ < 10)
10908 udelay(trycount
* num_online_cpus());
10910 synchronize_sched();
10913 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10914 smp_mb(); /* ensure test happens before caller kfree */
10919 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10920 for_each_online_cpu(cpu
) {
10922 req
= &per_cpu(rcu_migration_req
, cpu
);
10923 init_completion(&req
->done
);
10925 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10926 spin_lock_irqsave(&rq
->lock
, flags
);
10927 list_add(&req
->list
, &rq
->migration_queue
);
10928 spin_unlock_irqrestore(&rq
->lock
, flags
);
10929 wake_up_process(rq
->migration_thread
);
10931 for_each_online_cpu(cpu
) {
10932 rcu_expedited_state
= cpu
;
10933 req
= &per_cpu(rcu_migration_req
, cpu
);
10935 wait_for_completion(&req
->done
);
10936 spin_lock_irqsave(&rq
->lock
, flags
);
10937 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10938 need_full_sync
= 1;
10939 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10940 spin_unlock_irqrestore(&rq
->lock
, flags
);
10942 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10943 mutex_unlock(&rcu_sched_expedited_mutex
);
10945 if (need_full_sync
)
10946 synchronize_sched();
10948 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10950 #endif /* #else #ifndef CONFIG_SMP */