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
) {
1265 * Inline assembly required to prevent the compiler
1266 * optimising this loop into a divmod call.
1267 * See __iter_div_u64_rem() for another example of this.
1269 asm("" : "+rm" (rq
->age_stamp
));
1270 rq
->age_stamp
+= period
;
1275 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1277 rq
->rt_avg
+= rt_delta
;
1278 sched_avg_update(rq
);
1281 #else /* !CONFIG_SMP */
1282 static void resched_task(struct task_struct
*p
)
1284 assert_spin_locked(&task_rq(p
)->lock
);
1285 set_tsk_need_resched(p
);
1288 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1291 #endif /* CONFIG_SMP */
1293 #if BITS_PER_LONG == 32
1294 # define WMULT_CONST (~0UL)
1296 # define WMULT_CONST (1UL << 32)
1299 #define WMULT_SHIFT 32
1302 * Shift right and round:
1304 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1307 * delta *= weight / lw
1309 static unsigned long
1310 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1311 struct load_weight
*lw
)
1315 if (!lw
->inv_weight
) {
1316 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1319 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1323 tmp
= (u64
)delta_exec
* weight
;
1325 * Check whether we'd overflow the 64-bit multiplication:
1327 if (unlikely(tmp
> WMULT_CONST
))
1328 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1331 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1333 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1336 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1342 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1349 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1350 * of tasks with abnormal "nice" values across CPUs the contribution that
1351 * each task makes to its run queue's load is weighted according to its
1352 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1353 * scaled version of the new time slice allocation that they receive on time
1357 #define WEIGHT_IDLEPRIO 3
1358 #define WMULT_IDLEPRIO 1431655765
1361 * Nice levels are multiplicative, with a gentle 10% change for every
1362 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1363 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1364 * that remained on nice 0.
1366 * The "10% effect" is relative and cumulative: from _any_ nice level,
1367 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1368 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1369 * If a task goes up by ~10% and another task goes down by ~10% then
1370 * the relative distance between them is ~25%.)
1372 static const int prio_to_weight
[40] = {
1373 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1374 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1375 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1376 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1377 /* 0 */ 1024, 820, 655, 526, 423,
1378 /* 5 */ 335, 272, 215, 172, 137,
1379 /* 10 */ 110, 87, 70, 56, 45,
1380 /* 15 */ 36, 29, 23, 18, 15,
1384 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1386 * In cases where the weight does not change often, we can use the
1387 * precalculated inverse to speed up arithmetics by turning divisions
1388 * into multiplications:
1390 static const u32 prio_to_wmult
[40] = {
1391 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1392 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1393 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1394 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1395 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1396 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1397 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1398 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1401 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1404 * runqueue iterator, to support SMP load-balancing between different
1405 * scheduling classes, without having to expose their internal data
1406 * structures to the load-balancing proper:
1408 struct rq_iterator
{
1410 struct task_struct
*(*start
)(void *);
1411 struct task_struct
*(*next
)(void *);
1415 static unsigned long
1416 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1417 unsigned long max_load_move
, struct sched_domain
*sd
,
1418 enum cpu_idle_type idle
, int *all_pinned
,
1419 int *this_best_prio
, struct rq_iterator
*iterator
);
1422 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1423 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1424 struct rq_iterator
*iterator
);
1427 /* Time spent by the tasks of the cpu accounting group executing in ... */
1428 enum cpuacct_stat_index
{
1429 CPUACCT_STAT_USER
, /* ... user mode */
1430 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1432 CPUACCT_STAT_NSTATS
,
1435 #ifdef CONFIG_CGROUP_CPUACCT
1436 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1437 static void cpuacct_update_stats(struct task_struct
*tsk
,
1438 enum cpuacct_stat_index idx
, cputime_t val
);
1440 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1441 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1442 enum cpuacct_stat_index idx
, cputime_t val
) {}
1445 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1447 update_load_add(&rq
->load
, load
);
1450 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1452 update_load_sub(&rq
->load
, load
);
1455 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1456 typedef int (*tg_visitor
)(struct task_group
*, void *);
1459 * Iterate the full tree, calling @down when first entering a node and @up when
1460 * leaving it for the final time.
1462 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1464 struct task_group
*parent
, *child
;
1468 parent
= &root_task_group
;
1470 ret
= (*down
)(parent
, data
);
1473 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1480 ret
= (*up
)(parent
, data
);
1485 parent
= parent
->parent
;
1494 static int tg_nop(struct task_group
*tg
, void *data
)
1501 /* Used instead of source_load when we know the type == 0 */
1502 static unsigned long weighted_cpuload(const int cpu
)
1504 return cpu_rq(cpu
)->load
.weight
;
1508 * Return a low guess at the load of a migration-source cpu weighted
1509 * according to the scheduling class and "nice" value.
1511 * We want to under-estimate the load of migration sources, to
1512 * balance conservatively.
1514 static unsigned long source_load(int cpu
, int type
)
1516 struct rq
*rq
= cpu_rq(cpu
);
1517 unsigned long total
= weighted_cpuload(cpu
);
1519 if (type
== 0 || !sched_feat(LB_BIAS
))
1522 return min(rq
->cpu_load
[type
-1], total
);
1526 * Return a high guess at the load of a migration-target cpu weighted
1527 * according to the scheduling class and "nice" value.
1529 static unsigned long target_load(int cpu
, int type
)
1531 struct rq
*rq
= cpu_rq(cpu
);
1532 unsigned long total
= weighted_cpuload(cpu
);
1534 if (type
== 0 || !sched_feat(LB_BIAS
))
1537 return max(rq
->cpu_load
[type
-1], total
);
1540 static struct sched_group
*group_of(int cpu
)
1542 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1550 static unsigned long power_of(int cpu
)
1552 struct sched_group
*group
= group_of(cpu
);
1555 return SCHED_LOAD_SCALE
;
1557 return group
->cpu_power
;
1560 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1562 static unsigned long cpu_avg_load_per_task(int cpu
)
1564 struct rq
*rq
= cpu_rq(cpu
);
1565 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1568 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1570 rq
->avg_load_per_task
= 0;
1572 return rq
->avg_load_per_task
;
1575 #ifdef CONFIG_FAIR_GROUP_SCHED
1577 static __read_mostly
unsigned long *update_shares_data
;
1579 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1582 * Calculate and set the cpu's group shares.
1584 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1585 unsigned long sd_shares
,
1586 unsigned long sd_rq_weight
,
1587 unsigned long *usd_rq_weight
)
1589 unsigned long shares
, rq_weight
;
1592 rq_weight
= usd_rq_weight
[cpu
];
1595 rq_weight
= NICE_0_LOAD
;
1599 * \Sum_j shares_j * rq_weight_i
1600 * shares_i = -----------------------------
1601 * \Sum_j rq_weight_j
1603 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1604 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1606 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1607 sysctl_sched_shares_thresh
) {
1608 struct rq
*rq
= cpu_rq(cpu
);
1609 unsigned long flags
;
1611 spin_lock_irqsave(&rq
->lock
, flags
);
1612 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1613 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1614 __set_se_shares(tg
->se
[cpu
], shares
);
1615 spin_unlock_irqrestore(&rq
->lock
, flags
);
1620 * Re-compute the task group their per cpu shares over the given domain.
1621 * This needs to be done in a bottom-up fashion because the rq weight of a
1622 * parent group depends on the shares of its child groups.
1624 static int tg_shares_up(struct task_group
*tg
, void *data
)
1626 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1627 unsigned long *usd_rq_weight
;
1628 struct sched_domain
*sd
= data
;
1629 unsigned long flags
;
1635 local_irq_save(flags
);
1636 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1638 for_each_cpu(i
, sched_domain_span(sd
)) {
1639 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1640 usd_rq_weight
[i
] = weight
;
1642 rq_weight
+= weight
;
1644 * If there are currently no tasks on the cpu pretend there
1645 * is one of average load so that when a new task gets to
1646 * run here it will not get delayed by group starvation.
1649 weight
= NICE_0_LOAD
;
1651 sum_weight
+= weight
;
1652 shares
+= tg
->cfs_rq
[i
]->shares
;
1656 rq_weight
= sum_weight
;
1658 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1659 shares
= tg
->shares
;
1661 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1662 shares
= tg
->shares
;
1664 for_each_cpu(i
, sched_domain_span(sd
))
1665 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1667 local_irq_restore(flags
);
1673 * Compute the cpu's hierarchical load factor for each task group.
1674 * This needs to be done in a top-down fashion because the load of a child
1675 * group is a fraction of its parents load.
1677 static int tg_load_down(struct task_group
*tg
, void *data
)
1680 long cpu
= (long)data
;
1683 load
= cpu_rq(cpu
)->load
.weight
;
1685 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1686 load
*= tg
->cfs_rq
[cpu
]->shares
;
1687 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1690 tg
->cfs_rq
[cpu
]->h_load
= load
;
1695 static void update_shares(struct sched_domain
*sd
)
1700 if (root_task_group_empty())
1703 now
= cpu_clock(raw_smp_processor_id());
1704 elapsed
= now
- sd
->last_update
;
1706 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1707 sd
->last_update
= now
;
1708 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1712 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1714 if (root_task_group_empty())
1717 spin_unlock(&rq
->lock
);
1719 spin_lock(&rq
->lock
);
1722 static void update_h_load(long cpu
)
1724 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1729 static inline void update_shares(struct sched_domain
*sd
)
1733 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1739 #ifdef CONFIG_PREEMPT
1741 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1744 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1745 * way at the expense of forcing extra atomic operations in all
1746 * invocations. This assures that the double_lock is acquired using the
1747 * same underlying policy as the spinlock_t on this architecture, which
1748 * reduces latency compared to the unfair variant below. However, it
1749 * also adds more overhead and therefore may reduce throughput.
1751 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1752 __releases(this_rq
->lock
)
1753 __acquires(busiest
->lock
)
1754 __acquires(this_rq
->lock
)
1756 spin_unlock(&this_rq
->lock
);
1757 double_rq_lock(this_rq
, busiest
);
1764 * Unfair double_lock_balance: Optimizes throughput at the expense of
1765 * latency by eliminating extra atomic operations when the locks are
1766 * already in proper order on entry. This favors lower cpu-ids and will
1767 * grant the double lock to lower cpus over higher ids under contention,
1768 * regardless of entry order into the function.
1770 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1771 __releases(this_rq
->lock
)
1772 __acquires(busiest
->lock
)
1773 __acquires(this_rq
->lock
)
1777 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1778 if (busiest
< this_rq
) {
1779 spin_unlock(&this_rq
->lock
);
1780 spin_lock(&busiest
->lock
);
1781 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1784 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1789 #endif /* CONFIG_PREEMPT */
1792 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1794 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1796 if (unlikely(!irqs_disabled())) {
1797 /* printk() doesn't work good under rq->lock */
1798 spin_unlock(&this_rq
->lock
);
1802 return _double_lock_balance(this_rq
, busiest
);
1805 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1806 __releases(busiest
->lock
)
1808 spin_unlock(&busiest
->lock
);
1809 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1813 #ifdef CONFIG_FAIR_GROUP_SCHED
1814 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1817 cfs_rq
->shares
= shares
;
1822 static void calc_load_account_active(struct rq
*this_rq
);
1823 static void update_sysctl(void);
1825 #include "sched_stats.h"
1826 #include "sched_idletask.c"
1827 #include "sched_fair.c"
1828 #include "sched_rt.c"
1829 #ifdef CONFIG_SCHED_DEBUG
1830 # include "sched_debug.c"
1833 #define sched_class_highest (&rt_sched_class)
1834 #define for_each_class(class) \
1835 for (class = sched_class_highest; class; class = class->next)
1837 static void inc_nr_running(struct rq
*rq
)
1842 static void dec_nr_running(struct rq
*rq
)
1847 static void set_load_weight(struct task_struct
*p
)
1849 if (task_has_rt_policy(p
)) {
1850 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1851 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1856 * SCHED_IDLE tasks get minimal weight:
1858 if (p
->policy
== SCHED_IDLE
) {
1859 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1860 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1864 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1865 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1868 static void update_avg(u64
*avg
, u64 sample
)
1870 s64 diff
= sample
- *avg
;
1874 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1877 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1879 sched_info_queued(p
);
1880 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1884 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1887 if (p
->se
.last_wakeup
) {
1888 update_avg(&p
->se
.avg_overlap
,
1889 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1890 p
->se
.last_wakeup
= 0;
1892 update_avg(&p
->se
.avg_wakeup
,
1893 sysctl_sched_wakeup_granularity
);
1897 sched_info_dequeued(p
);
1898 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1903 * __normal_prio - return the priority that is based on the static prio
1905 static inline int __normal_prio(struct task_struct
*p
)
1907 return p
->static_prio
;
1911 * Calculate the expected normal priority: i.e. priority
1912 * without taking RT-inheritance into account. Might be
1913 * boosted by interactivity modifiers. Changes upon fork,
1914 * setprio syscalls, and whenever the interactivity
1915 * estimator recalculates.
1917 static inline int normal_prio(struct task_struct
*p
)
1921 if (task_has_rt_policy(p
))
1922 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1924 prio
= __normal_prio(p
);
1929 * Calculate the current priority, i.e. the priority
1930 * taken into account by the scheduler. This value might
1931 * be boosted by RT tasks, or might be boosted by
1932 * interactivity modifiers. Will be RT if the task got
1933 * RT-boosted. If not then it returns p->normal_prio.
1935 static int effective_prio(struct task_struct
*p
)
1937 p
->normal_prio
= normal_prio(p
);
1939 * If we are RT tasks or we were boosted to RT priority,
1940 * keep the priority unchanged. Otherwise, update priority
1941 * to the normal priority:
1943 if (!rt_prio(p
->prio
))
1944 return p
->normal_prio
;
1949 * activate_task - move a task to the runqueue.
1951 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1953 if (task_contributes_to_load(p
))
1954 rq
->nr_uninterruptible
--;
1956 enqueue_task(rq
, p
, wakeup
);
1961 * deactivate_task - remove a task from the runqueue.
1963 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1965 if (task_contributes_to_load(p
))
1966 rq
->nr_uninterruptible
++;
1968 dequeue_task(rq
, p
, sleep
);
1973 * task_curr - is this task currently executing on a CPU?
1974 * @p: the task in question.
1976 inline int task_curr(const struct task_struct
*p
)
1978 return cpu_curr(task_cpu(p
)) == p
;
1981 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1983 set_task_rq(p
, cpu
);
1986 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1987 * successfuly executed on another CPU. We must ensure that updates of
1988 * per-task data have been completed by this moment.
1991 task_thread_info(p
)->cpu
= cpu
;
1995 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1996 const struct sched_class
*prev_class
,
1997 int oldprio
, int running
)
1999 if (prev_class
!= p
->sched_class
) {
2000 if (prev_class
->switched_from
)
2001 prev_class
->switched_from(rq
, p
, running
);
2002 p
->sched_class
->switched_to(rq
, p
, running
);
2004 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2008 * kthread_bind - bind a just-created kthread to a cpu.
2009 * @p: thread created by kthread_create().
2010 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2012 * Description: This function is equivalent to set_cpus_allowed(),
2013 * except that @cpu doesn't need to be online, and the thread must be
2014 * stopped (i.e., just returned from kthread_create()).
2016 * Function lives here instead of kthread.c because it messes with
2017 * scheduler internals which require locking.
2019 void kthread_bind(struct task_struct
*p
, unsigned int cpu
)
2021 struct rq
*rq
= cpu_rq(cpu
);
2022 unsigned long flags
;
2024 /* Must have done schedule() in kthread() before we set_task_cpu */
2025 if (!wait_task_inactive(p
, TASK_UNINTERRUPTIBLE
)) {
2030 spin_lock_irqsave(&rq
->lock
, flags
);
2031 set_task_cpu(p
, cpu
);
2032 p
->cpus_allowed
= cpumask_of_cpu(cpu
);
2033 p
->rt
.nr_cpus_allowed
= 1;
2034 p
->flags
|= PF_THREAD_BOUND
;
2035 spin_unlock_irqrestore(&rq
->lock
, flags
);
2037 EXPORT_SYMBOL(kthread_bind
);
2041 * Is this task likely cache-hot:
2044 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2048 if (p
->sched_class
!= &fair_sched_class
)
2052 * Buddy candidates are cache hot:
2054 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2055 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2056 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2059 if (sysctl_sched_migration_cost
== -1)
2061 if (sysctl_sched_migration_cost
== 0)
2064 delta
= now
- p
->se
.exec_start
;
2066 return delta
< (s64
)sysctl_sched_migration_cost
;
2070 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2072 int old_cpu
= task_cpu(p
);
2073 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2074 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2075 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2078 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2080 trace_sched_migrate_task(p
, new_cpu
);
2082 #ifdef CONFIG_SCHEDSTATS
2083 if (p
->se
.wait_start
)
2084 p
->se
.wait_start
-= clock_offset
;
2085 if (p
->se
.sleep_start
)
2086 p
->se
.sleep_start
-= clock_offset
;
2087 if (p
->se
.block_start
)
2088 p
->se
.block_start
-= clock_offset
;
2090 if (old_cpu
!= new_cpu
) {
2091 p
->se
.nr_migrations
++;
2092 new_rq
->nr_migrations_in
++;
2093 #ifdef CONFIG_SCHEDSTATS
2094 if (task_hot(p
, old_rq
->clock
, NULL
))
2095 schedstat_inc(p
, se
.nr_forced2_migrations
);
2097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2100 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2101 new_cfsrq
->min_vruntime
;
2103 __set_task_cpu(p
, new_cpu
);
2106 struct migration_req
{
2107 struct list_head list
;
2109 struct task_struct
*task
;
2112 struct completion done
;
2116 * The task's runqueue lock must be held.
2117 * Returns true if you have to wait for migration thread.
2120 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2122 struct rq
*rq
= task_rq(p
);
2125 * If the task is not on a runqueue (and not running), then
2126 * the next wake-up will properly place the task.
2128 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2131 init_completion(&req
->done
);
2133 req
->dest_cpu
= dest_cpu
;
2134 list_add(&req
->list
, &rq
->migration_queue
);
2140 * wait_task_context_switch - wait for a thread to complete at least one
2143 * @p must not be current.
2145 void wait_task_context_switch(struct task_struct
*p
)
2147 unsigned long nvcsw
, nivcsw
, flags
;
2155 * The runqueue is assigned before the actual context
2156 * switch. We need to take the runqueue lock.
2158 * We could check initially without the lock but it is
2159 * very likely that we need to take the lock in every
2162 rq
= task_rq_lock(p
, &flags
);
2163 running
= task_running(rq
, p
);
2164 task_rq_unlock(rq
, &flags
);
2166 if (likely(!running
))
2169 * The switch count is incremented before the actual
2170 * context switch. We thus wait for two switches to be
2171 * sure at least one completed.
2173 if ((p
->nvcsw
- nvcsw
) > 1)
2175 if ((p
->nivcsw
- nivcsw
) > 1)
2183 * wait_task_inactive - wait for a thread to unschedule.
2185 * If @match_state is nonzero, it's the @p->state value just checked and
2186 * not expected to change. If it changes, i.e. @p might have woken up,
2187 * then return zero. When we succeed in waiting for @p to be off its CPU,
2188 * we return a positive number (its total switch count). If a second call
2189 * a short while later returns the same number, the caller can be sure that
2190 * @p has remained unscheduled the whole time.
2192 * The caller must ensure that the task *will* unschedule sometime soon,
2193 * else this function might spin for a *long* time. This function can't
2194 * be called with interrupts off, or it may introduce deadlock with
2195 * smp_call_function() if an IPI is sent by the same process we are
2196 * waiting to become inactive.
2198 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2200 unsigned long flags
;
2207 * We do the initial early heuristics without holding
2208 * any task-queue locks at all. We'll only try to get
2209 * the runqueue lock when things look like they will
2215 * If the task is actively running on another CPU
2216 * still, just relax and busy-wait without holding
2219 * NOTE! Since we don't hold any locks, it's not
2220 * even sure that "rq" stays as the right runqueue!
2221 * But we don't care, since "task_running()" will
2222 * return false if the runqueue has changed and p
2223 * is actually now running somewhere else!
2225 while (task_running(rq
, p
)) {
2226 if (match_state
&& unlikely(p
->state
!= match_state
))
2232 * Ok, time to look more closely! We need the rq
2233 * lock now, to be *sure*. If we're wrong, we'll
2234 * just go back and repeat.
2236 rq
= task_rq_lock(p
, &flags
);
2237 trace_sched_wait_task(rq
, p
);
2238 running
= task_running(rq
, p
);
2239 on_rq
= p
->se
.on_rq
;
2241 if (!match_state
|| p
->state
== match_state
)
2242 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2243 task_rq_unlock(rq
, &flags
);
2246 * If it changed from the expected state, bail out now.
2248 if (unlikely(!ncsw
))
2252 * Was it really running after all now that we
2253 * checked with the proper locks actually held?
2255 * Oops. Go back and try again..
2257 if (unlikely(running
)) {
2263 * It's not enough that it's not actively running,
2264 * it must be off the runqueue _entirely_, and not
2267 * So if it was still runnable (but just not actively
2268 * running right now), it's preempted, and we should
2269 * yield - it could be a while.
2271 if (unlikely(on_rq
)) {
2272 schedule_timeout_uninterruptible(1);
2277 * Ahh, all good. It wasn't running, and it wasn't
2278 * runnable, which means that it will never become
2279 * running in the future either. We're all done!
2288 * kick_process - kick a running thread to enter/exit the kernel
2289 * @p: the to-be-kicked thread
2291 * Cause a process which is running on another CPU to enter
2292 * kernel-mode, without any delay. (to get signals handled.)
2294 * NOTE: this function doesnt have to take the runqueue lock,
2295 * because all it wants to ensure is that the remote task enters
2296 * the kernel. If the IPI races and the task has been migrated
2297 * to another CPU then no harm is done and the purpose has been
2300 void kick_process(struct task_struct
*p
)
2306 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2307 smp_send_reschedule(cpu
);
2310 EXPORT_SYMBOL_GPL(kick_process
);
2311 #endif /* CONFIG_SMP */
2314 * task_oncpu_function_call - call a function on the cpu on which a task runs
2315 * @p: the task to evaluate
2316 * @func: the function to be called
2317 * @info: the function call argument
2319 * Calls the function @func when the task is currently running. This might
2320 * be on the current CPU, which just calls the function directly
2322 void task_oncpu_function_call(struct task_struct
*p
,
2323 void (*func
) (void *info
), void *info
)
2330 smp_call_function_single(cpu
, func
, info
, 1);
2335 * try_to_wake_up - wake up a thread
2336 * @p: the to-be-woken-up thread
2337 * @state: the mask of task states that can be woken
2338 * @sync: do a synchronous wakeup?
2340 * Put it on the run-queue if it's not already there. The "current"
2341 * thread is always on the run-queue (except when the actual
2342 * re-schedule is in progress), and as such you're allowed to do
2343 * the simpler "current->state = TASK_RUNNING" to mark yourself
2344 * runnable without the overhead of this.
2346 * returns failure only if the task is already active.
2348 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2351 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2352 unsigned long flags
;
2353 struct rq
*rq
, *orig_rq
;
2355 if (!sched_feat(SYNC_WAKEUPS
))
2356 wake_flags
&= ~WF_SYNC
;
2358 this_cpu
= get_cpu();
2361 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2362 update_rq_clock(rq
);
2363 if (!(p
->state
& state
))
2373 if (unlikely(task_running(rq
, p
)))
2377 * In order to handle concurrent wakeups and release the rq->lock
2378 * we put the task in TASK_WAKING state.
2380 * First fix up the nr_uninterruptible count:
2382 if (task_contributes_to_load(p
))
2383 rq
->nr_uninterruptible
--;
2384 p
->state
= TASK_WAKING
;
2385 task_rq_unlock(rq
, &flags
);
2387 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2388 if (cpu
!= orig_cpu
)
2389 set_task_cpu(p
, cpu
);
2391 rq
= task_rq_lock(p
, &flags
);
2394 update_rq_clock(rq
);
2396 WARN_ON(p
->state
!= TASK_WAKING
);
2399 #ifdef CONFIG_SCHEDSTATS
2400 schedstat_inc(rq
, ttwu_count
);
2401 if (cpu
== this_cpu
)
2402 schedstat_inc(rq
, ttwu_local
);
2404 struct sched_domain
*sd
;
2405 for_each_domain(this_cpu
, sd
) {
2406 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2407 schedstat_inc(sd
, ttwu_wake_remote
);
2412 #endif /* CONFIG_SCHEDSTATS */
2415 #endif /* CONFIG_SMP */
2416 schedstat_inc(p
, se
.nr_wakeups
);
2417 if (wake_flags
& WF_SYNC
)
2418 schedstat_inc(p
, se
.nr_wakeups_sync
);
2419 if (orig_cpu
!= cpu
)
2420 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2421 if (cpu
== this_cpu
)
2422 schedstat_inc(p
, se
.nr_wakeups_local
);
2424 schedstat_inc(p
, se
.nr_wakeups_remote
);
2425 activate_task(rq
, p
, 1);
2429 * Only attribute actual wakeups done by this task.
2431 if (!in_interrupt()) {
2432 struct sched_entity
*se
= ¤t
->se
;
2433 u64 sample
= se
->sum_exec_runtime
;
2435 if (se
->last_wakeup
)
2436 sample
-= se
->last_wakeup
;
2438 sample
-= se
->start_runtime
;
2439 update_avg(&se
->avg_wakeup
, sample
);
2441 se
->last_wakeup
= se
->sum_exec_runtime
;
2445 trace_sched_wakeup(rq
, p
, success
);
2446 check_preempt_curr(rq
, p
, wake_flags
);
2448 p
->state
= TASK_RUNNING
;
2450 if (p
->sched_class
->task_wake_up
)
2451 p
->sched_class
->task_wake_up(rq
, p
);
2453 if (unlikely(rq
->idle_stamp
)) {
2454 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2455 u64 max
= 2*sysctl_sched_migration_cost
;
2460 update_avg(&rq
->avg_idle
, delta
);
2465 task_rq_unlock(rq
, &flags
);
2472 * wake_up_process - Wake up a specific process
2473 * @p: The process to be woken up.
2475 * Attempt to wake up the nominated process and move it to the set of runnable
2476 * processes. Returns 1 if the process was woken up, 0 if it was already
2479 * It may be assumed that this function implies a write memory barrier before
2480 * changing the task state if and only if any tasks are woken up.
2482 int wake_up_process(struct task_struct
*p
)
2484 return try_to_wake_up(p
, TASK_ALL
, 0);
2486 EXPORT_SYMBOL(wake_up_process
);
2488 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2490 return try_to_wake_up(p
, state
, 0);
2494 * Perform scheduler related setup for a newly forked process p.
2495 * p is forked by current.
2497 * __sched_fork() is basic setup used by init_idle() too:
2499 static void __sched_fork(struct task_struct
*p
)
2501 p
->se
.exec_start
= 0;
2502 p
->se
.sum_exec_runtime
= 0;
2503 p
->se
.prev_sum_exec_runtime
= 0;
2504 p
->se
.nr_migrations
= 0;
2505 p
->se
.last_wakeup
= 0;
2506 p
->se
.avg_overlap
= 0;
2507 p
->se
.start_runtime
= 0;
2508 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2509 p
->se
.avg_running
= 0;
2511 #ifdef CONFIG_SCHEDSTATS
2512 p
->se
.wait_start
= 0;
2514 p
->se
.wait_count
= 0;
2517 p
->se
.sleep_start
= 0;
2518 p
->se
.sleep_max
= 0;
2519 p
->se
.sum_sleep_runtime
= 0;
2521 p
->se
.block_start
= 0;
2522 p
->se
.block_max
= 0;
2524 p
->se
.slice_max
= 0;
2526 p
->se
.nr_migrations_cold
= 0;
2527 p
->se
.nr_failed_migrations_affine
= 0;
2528 p
->se
.nr_failed_migrations_running
= 0;
2529 p
->se
.nr_failed_migrations_hot
= 0;
2530 p
->se
.nr_forced_migrations
= 0;
2531 p
->se
.nr_forced2_migrations
= 0;
2533 p
->se
.nr_wakeups
= 0;
2534 p
->se
.nr_wakeups_sync
= 0;
2535 p
->se
.nr_wakeups_migrate
= 0;
2536 p
->se
.nr_wakeups_local
= 0;
2537 p
->se
.nr_wakeups_remote
= 0;
2538 p
->se
.nr_wakeups_affine
= 0;
2539 p
->se
.nr_wakeups_affine_attempts
= 0;
2540 p
->se
.nr_wakeups_passive
= 0;
2541 p
->se
.nr_wakeups_idle
= 0;
2545 INIT_LIST_HEAD(&p
->rt
.run_list
);
2547 INIT_LIST_HEAD(&p
->se
.group_node
);
2549 #ifdef CONFIG_PREEMPT_NOTIFIERS
2550 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2554 * We mark the process as running here, but have not actually
2555 * inserted it onto the runqueue yet. This guarantees that
2556 * nobody will actually run it, and a signal or other external
2557 * event cannot wake it up and insert it on the runqueue either.
2559 p
->state
= TASK_RUNNING
;
2563 * fork()/clone()-time setup:
2565 void sched_fork(struct task_struct
*p
, int clone_flags
)
2567 int cpu
= get_cpu();
2572 * Revert to default priority/policy on fork if requested.
2574 if (unlikely(p
->sched_reset_on_fork
)) {
2575 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2576 p
->policy
= SCHED_NORMAL
;
2577 p
->normal_prio
= p
->static_prio
;
2580 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2581 p
->static_prio
= NICE_TO_PRIO(0);
2582 p
->normal_prio
= p
->static_prio
;
2587 * We don't need the reset flag anymore after the fork. It has
2588 * fulfilled its duty:
2590 p
->sched_reset_on_fork
= 0;
2594 * Make sure we do not leak PI boosting priority to the child.
2596 p
->prio
= current
->normal_prio
;
2598 if (!rt_prio(p
->prio
))
2599 p
->sched_class
= &fair_sched_class
;
2602 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2604 set_task_cpu(p
, cpu
);
2606 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2607 if (likely(sched_info_on()))
2608 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2610 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2613 #ifdef CONFIG_PREEMPT
2614 /* Want to start with kernel preemption disabled. */
2615 task_thread_info(p
)->preempt_count
= 1;
2617 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2623 * wake_up_new_task - wake up a newly created task for the first time.
2625 * This function will do some initial scheduler statistics housekeeping
2626 * that must be done for every newly created context, then puts the task
2627 * on the runqueue and wakes it.
2629 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2631 unsigned long flags
;
2634 rq
= task_rq_lock(p
, &flags
);
2635 BUG_ON(p
->state
!= TASK_RUNNING
);
2636 update_rq_clock(rq
);
2638 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2639 activate_task(rq
, p
, 0);
2642 * Let the scheduling class do new task startup
2643 * management (if any):
2645 p
->sched_class
->task_new(rq
, p
);
2648 trace_sched_wakeup_new(rq
, p
, 1);
2649 check_preempt_curr(rq
, p
, WF_FORK
);
2651 if (p
->sched_class
->task_wake_up
)
2652 p
->sched_class
->task_wake_up(rq
, p
);
2654 task_rq_unlock(rq
, &flags
);
2657 #ifdef CONFIG_PREEMPT_NOTIFIERS
2660 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2661 * @notifier: notifier struct to register
2663 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2665 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2667 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2670 * preempt_notifier_unregister - no longer interested in preemption notifications
2671 * @notifier: notifier struct to unregister
2673 * This is safe to call from within a preemption notifier.
2675 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2677 hlist_del(¬ifier
->link
);
2679 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2681 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2683 struct preempt_notifier
*notifier
;
2684 struct hlist_node
*node
;
2686 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2687 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2691 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2692 struct task_struct
*next
)
2694 struct preempt_notifier
*notifier
;
2695 struct hlist_node
*node
;
2697 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2698 notifier
->ops
->sched_out(notifier
, next
);
2701 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2703 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2708 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2709 struct task_struct
*next
)
2713 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2716 * prepare_task_switch - prepare to switch tasks
2717 * @rq: the runqueue preparing to switch
2718 * @prev: the current task that is being switched out
2719 * @next: the task we are going to switch to.
2721 * This is called with the rq lock held and interrupts off. It must
2722 * be paired with a subsequent finish_task_switch after the context
2725 * prepare_task_switch sets up locking and calls architecture specific
2729 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2730 struct task_struct
*next
)
2732 fire_sched_out_preempt_notifiers(prev
, next
);
2733 prepare_lock_switch(rq
, next
);
2734 prepare_arch_switch(next
);
2738 * finish_task_switch - clean up after a task-switch
2739 * @rq: runqueue associated with task-switch
2740 * @prev: the thread we just switched away from.
2742 * finish_task_switch must be called after the context switch, paired
2743 * with a prepare_task_switch call before the context switch.
2744 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2745 * and do any other architecture-specific cleanup actions.
2747 * Note that we may have delayed dropping an mm in context_switch(). If
2748 * so, we finish that here outside of the runqueue lock. (Doing it
2749 * with the lock held can cause deadlocks; see schedule() for
2752 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2753 __releases(rq
->lock
)
2755 struct mm_struct
*mm
= rq
->prev_mm
;
2761 * A task struct has one reference for the use as "current".
2762 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2763 * schedule one last time. The schedule call will never return, and
2764 * the scheduled task must drop that reference.
2765 * The test for TASK_DEAD must occur while the runqueue locks are
2766 * still held, otherwise prev could be scheduled on another cpu, die
2767 * there before we look at prev->state, and then the reference would
2769 * Manfred Spraul <manfred@colorfullife.com>
2771 prev_state
= prev
->state
;
2772 finish_arch_switch(prev
);
2773 perf_event_task_sched_in(current
, cpu_of(rq
));
2774 finish_lock_switch(rq
, prev
);
2776 fire_sched_in_preempt_notifiers(current
);
2779 if (unlikely(prev_state
== TASK_DEAD
)) {
2781 * Remove function-return probe instances associated with this
2782 * task and put them back on the free list.
2784 kprobe_flush_task(prev
);
2785 put_task_struct(prev
);
2791 /* assumes rq->lock is held */
2792 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2794 if (prev
->sched_class
->pre_schedule
)
2795 prev
->sched_class
->pre_schedule(rq
, prev
);
2798 /* rq->lock is NOT held, but preemption is disabled */
2799 static inline void post_schedule(struct rq
*rq
)
2801 if (rq
->post_schedule
) {
2802 unsigned long flags
;
2804 spin_lock_irqsave(&rq
->lock
, flags
);
2805 if (rq
->curr
->sched_class
->post_schedule
)
2806 rq
->curr
->sched_class
->post_schedule(rq
);
2807 spin_unlock_irqrestore(&rq
->lock
, flags
);
2809 rq
->post_schedule
= 0;
2815 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2819 static inline void post_schedule(struct rq
*rq
)
2826 * schedule_tail - first thing a freshly forked thread must call.
2827 * @prev: the thread we just switched away from.
2829 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2830 __releases(rq
->lock
)
2832 struct rq
*rq
= this_rq();
2834 finish_task_switch(rq
, prev
);
2837 * FIXME: do we need to worry about rq being invalidated by the
2842 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2843 /* In this case, finish_task_switch does not reenable preemption */
2846 if (current
->set_child_tid
)
2847 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2851 * context_switch - switch to the new MM and the new
2852 * thread's register state.
2855 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2856 struct task_struct
*next
)
2858 struct mm_struct
*mm
, *oldmm
;
2860 prepare_task_switch(rq
, prev
, next
);
2861 trace_sched_switch(rq
, prev
, next
);
2863 oldmm
= prev
->active_mm
;
2865 * For paravirt, this is coupled with an exit in switch_to to
2866 * combine the page table reload and the switch backend into
2869 arch_start_context_switch(prev
);
2871 if (unlikely(!mm
)) {
2872 next
->active_mm
= oldmm
;
2873 atomic_inc(&oldmm
->mm_count
);
2874 enter_lazy_tlb(oldmm
, next
);
2876 switch_mm(oldmm
, mm
, next
);
2878 if (unlikely(!prev
->mm
)) {
2879 prev
->active_mm
= NULL
;
2880 rq
->prev_mm
= oldmm
;
2883 * Since the runqueue lock will be released by the next
2884 * task (which is an invalid locking op but in the case
2885 * of the scheduler it's an obvious special-case), so we
2886 * do an early lockdep release here:
2888 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2889 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2892 /* Here we just switch the register state and the stack. */
2893 switch_to(prev
, next
, prev
);
2897 * this_rq must be evaluated again because prev may have moved
2898 * CPUs since it called schedule(), thus the 'rq' on its stack
2899 * frame will be invalid.
2901 finish_task_switch(this_rq(), prev
);
2905 * nr_running, nr_uninterruptible and nr_context_switches:
2907 * externally visible scheduler statistics: current number of runnable
2908 * threads, current number of uninterruptible-sleeping threads, total
2909 * number of context switches performed since bootup.
2911 unsigned long nr_running(void)
2913 unsigned long i
, sum
= 0;
2915 for_each_online_cpu(i
)
2916 sum
+= cpu_rq(i
)->nr_running
;
2921 unsigned long nr_uninterruptible(void)
2923 unsigned long i
, sum
= 0;
2925 for_each_possible_cpu(i
)
2926 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2929 * Since we read the counters lockless, it might be slightly
2930 * inaccurate. Do not allow it to go below zero though:
2932 if (unlikely((long)sum
< 0))
2938 unsigned long long nr_context_switches(void)
2941 unsigned long long sum
= 0;
2943 for_each_possible_cpu(i
)
2944 sum
+= cpu_rq(i
)->nr_switches
;
2949 unsigned long nr_iowait(void)
2951 unsigned long i
, sum
= 0;
2953 for_each_possible_cpu(i
)
2954 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2959 unsigned long nr_iowait_cpu(void)
2961 struct rq
*this = this_rq();
2962 return atomic_read(&this->nr_iowait
);
2965 unsigned long this_cpu_load(void)
2967 struct rq
*this = this_rq();
2968 return this->cpu_load
[0];
2972 /* Variables and functions for calc_load */
2973 static atomic_long_t calc_load_tasks
;
2974 static unsigned long calc_load_update
;
2975 unsigned long avenrun
[3];
2976 EXPORT_SYMBOL(avenrun
);
2979 * get_avenrun - get the load average array
2980 * @loads: pointer to dest load array
2981 * @offset: offset to add
2982 * @shift: shift count to shift the result left
2984 * These values are estimates at best, so no need for locking.
2986 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2988 loads
[0] = (avenrun
[0] + offset
) << shift
;
2989 loads
[1] = (avenrun
[1] + offset
) << shift
;
2990 loads
[2] = (avenrun
[2] + offset
) << shift
;
2993 static unsigned long
2994 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2997 load
+= active
* (FIXED_1
- exp
);
2998 return load
>> FSHIFT
;
3002 * calc_load - update the avenrun load estimates 10 ticks after the
3003 * CPUs have updated calc_load_tasks.
3005 void calc_global_load(void)
3007 unsigned long upd
= calc_load_update
+ 10;
3010 if (time_before(jiffies
, upd
))
3013 active
= atomic_long_read(&calc_load_tasks
);
3014 active
= active
> 0 ? active
* FIXED_1
: 0;
3016 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3017 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3018 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3020 calc_load_update
+= LOAD_FREQ
;
3024 * Either called from update_cpu_load() or from a cpu going idle
3026 static void calc_load_account_active(struct rq
*this_rq
)
3028 long nr_active
, delta
;
3030 nr_active
= this_rq
->nr_running
;
3031 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3033 if (nr_active
!= this_rq
->calc_load_active
) {
3034 delta
= nr_active
- this_rq
->calc_load_active
;
3035 this_rq
->calc_load_active
= nr_active
;
3036 atomic_long_add(delta
, &calc_load_tasks
);
3041 * Externally visible per-cpu scheduler statistics:
3042 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3044 u64
cpu_nr_migrations(int cpu
)
3046 return cpu_rq(cpu
)->nr_migrations_in
;
3050 * Update rq->cpu_load[] statistics. This function is usually called every
3051 * scheduler tick (TICK_NSEC).
3053 static void update_cpu_load(struct rq
*this_rq
)
3055 unsigned long this_load
= this_rq
->load
.weight
;
3058 this_rq
->nr_load_updates
++;
3060 /* Update our load: */
3061 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3062 unsigned long old_load
, new_load
;
3064 /* scale is effectively 1 << i now, and >> i divides by scale */
3066 old_load
= this_rq
->cpu_load
[i
];
3067 new_load
= this_load
;
3069 * Round up the averaging division if load is increasing. This
3070 * prevents us from getting stuck on 9 if the load is 10, for
3073 if (new_load
> old_load
)
3074 new_load
+= scale
-1;
3075 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3078 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3079 this_rq
->calc_load_update
+= LOAD_FREQ
;
3080 calc_load_account_active(this_rq
);
3087 * double_rq_lock - safely lock two runqueues
3089 * Note this does not disable interrupts like task_rq_lock,
3090 * you need to do so manually before calling.
3092 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3093 __acquires(rq1
->lock
)
3094 __acquires(rq2
->lock
)
3096 BUG_ON(!irqs_disabled());
3098 spin_lock(&rq1
->lock
);
3099 __acquire(rq2
->lock
); /* Fake it out ;) */
3102 spin_lock(&rq1
->lock
);
3103 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3105 spin_lock(&rq2
->lock
);
3106 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3109 update_rq_clock(rq1
);
3110 update_rq_clock(rq2
);
3114 * double_rq_unlock - safely unlock two runqueues
3116 * Note this does not restore interrupts like task_rq_unlock,
3117 * you need to do so manually after calling.
3119 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3120 __releases(rq1
->lock
)
3121 __releases(rq2
->lock
)
3123 spin_unlock(&rq1
->lock
);
3125 spin_unlock(&rq2
->lock
);
3127 __release(rq2
->lock
);
3131 * If dest_cpu is allowed for this process, migrate the task to it.
3132 * This is accomplished by forcing the cpu_allowed mask to only
3133 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3134 * the cpu_allowed mask is restored.
3136 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3138 struct migration_req req
;
3139 unsigned long flags
;
3142 rq
= task_rq_lock(p
, &flags
);
3143 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3144 || unlikely(!cpu_active(dest_cpu
)))
3147 /* force the process onto the specified CPU */
3148 if (migrate_task(p
, dest_cpu
, &req
)) {
3149 /* Need to wait for migration thread (might exit: take ref). */
3150 struct task_struct
*mt
= rq
->migration_thread
;
3152 get_task_struct(mt
);
3153 task_rq_unlock(rq
, &flags
);
3154 wake_up_process(mt
);
3155 put_task_struct(mt
);
3156 wait_for_completion(&req
.done
);
3161 task_rq_unlock(rq
, &flags
);
3165 * sched_exec - execve() is a valuable balancing opportunity, because at
3166 * this point the task has the smallest effective memory and cache footprint.
3168 void sched_exec(void)
3170 int new_cpu
, this_cpu
= get_cpu();
3171 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3173 if (new_cpu
!= this_cpu
)
3174 sched_migrate_task(current
, new_cpu
);
3178 * pull_task - move a task from a remote runqueue to the local runqueue.
3179 * Both runqueues must be locked.
3181 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3182 struct rq
*this_rq
, int this_cpu
)
3184 deactivate_task(src_rq
, p
, 0);
3185 set_task_cpu(p
, this_cpu
);
3186 activate_task(this_rq
, p
, 0);
3187 check_preempt_curr(this_rq
, p
, 0);
3191 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3194 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3195 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3198 int tsk_cache_hot
= 0;
3200 * We do not migrate tasks that are:
3201 * 1) running (obviously), or
3202 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3203 * 3) are cache-hot on their current CPU.
3205 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3206 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3211 if (task_running(rq
, p
)) {
3212 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3217 * Aggressive migration if:
3218 * 1) task is cache cold, or
3219 * 2) too many balance attempts have failed.
3222 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3223 if (!tsk_cache_hot
||
3224 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3225 #ifdef CONFIG_SCHEDSTATS
3226 if (tsk_cache_hot
) {
3227 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3228 schedstat_inc(p
, se
.nr_forced_migrations
);
3234 if (tsk_cache_hot
) {
3235 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3241 static unsigned long
3242 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3243 unsigned long max_load_move
, struct sched_domain
*sd
,
3244 enum cpu_idle_type idle
, int *all_pinned
,
3245 int *this_best_prio
, struct rq_iterator
*iterator
)
3247 int loops
= 0, pulled
= 0, pinned
= 0;
3248 struct task_struct
*p
;
3249 long rem_load_move
= max_load_move
;
3251 if (max_load_move
== 0)
3257 * Start the load-balancing iterator:
3259 p
= iterator
->start(iterator
->arg
);
3261 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3264 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3265 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3266 p
= iterator
->next(iterator
->arg
);
3270 pull_task(busiest
, p
, this_rq
, this_cpu
);
3272 rem_load_move
-= p
->se
.load
.weight
;
3274 #ifdef CONFIG_PREEMPT
3276 * NEWIDLE balancing is a source of latency, so preemptible kernels
3277 * will stop after the first task is pulled to minimize the critical
3280 if (idle
== CPU_NEWLY_IDLE
)
3285 * We only want to steal up to the prescribed amount of weighted load.
3287 if (rem_load_move
> 0) {
3288 if (p
->prio
< *this_best_prio
)
3289 *this_best_prio
= p
->prio
;
3290 p
= iterator
->next(iterator
->arg
);
3295 * Right now, this is one of only two places pull_task() is called,
3296 * so we can safely collect pull_task() stats here rather than
3297 * inside pull_task().
3299 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3302 *all_pinned
= pinned
;
3304 return max_load_move
- rem_load_move
;
3308 * move_tasks tries to move up to max_load_move weighted load from busiest to
3309 * this_rq, as part of a balancing operation within domain "sd".
3310 * Returns 1 if successful and 0 otherwise.
3312 * Called with both runqueues locked.
3314 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3315 unsigned long max_load_move
,
3316 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3319 const struct sched_class
*class = sched_class_highest
;
3320 unsigned long total_load_moved
= 0;
3321 int this_best_prio
= this_rq
->curr
->prio
;
3325 class->load_balance(this_rq
, this_cpu
, busiest
,
3326 max_load_move
- total_load_moved
,
3327 sd
, idle
, all_pinned
, &this_best_prio
);
3328 class = class->next
;
3330 #ifdef CONFIG_PREEMPT
3332 * NEWIDLE balancing is a source of latency, so preemptible
3333 * kernels will stop after the first task is pulled to minimize
3334 * the critical section.
3336 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3339 } while (class && max_load_move
> total_load_moved
);
3341 return total_load_moved
> 0;
3345 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3346 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3347 struct rq_iterator
*iterator
)
3349 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3353 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3354 pull_task(busiest
, p
, this_rq
, this_cpu
);
3356 * Right now, this is only the second place pull_task()
3357 * is called, so we can safely collect pull_task()
3358 * stats here rather than inside pull_task().
3360 schedstat_inc(sd
, lb_gained
[idle
]);
3364 p
= iterator
->next(iterator
->arg
);
3371 * move_one_task tries to move exactly one task from busiest to this_rq, as
3372 * part of active balancing operations within "domain".
3373 * Returns 1 if successful and 0 otherwise.
3375 * Called with both runqueues locked.
3377 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3378 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3380 const struct sched_class
*class;
3382 for_each_class(class) {
3383 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3389 /********** Helpers for find_busiest_group ************************/
3391 * sd_lb_stats - Structure to store the statistics of a sched_domain
3392 * during load balancing.
3394 struct sd_lb_stats
{
3395 struct sched_group
*busiest
; /* Busiest group in this sd */
3396 struct sched_group
*this; /* Local group in this sd */
3397 unsigned long total_load
; /* Total load of all groups in sd */
3398 unsigned long total_pwr
; /* Total power of all groups in sd */
3399 unsigned long avg_load
; /* Average load across all groups in sd */
3401 /** Statistics of this group */
3402 unsigned long this_load
;
3403 unsigned long this_load_per_task
;
3404 unsigned long this_nr_running
;
3406 /* Statistics of the busiest group */
3407 unsigned long max_load
;
3408 unsigned long busiest_load_per_task
;
3409 unsigned long busiest_nr_running
;
3410 unsigned long busiest_group_capacity
;
3412 int group_imb
; /* Is there imbalance in this sd */
3413 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3414 int power_savings_balance
; /* Is powersave balance needed for this sd */
3415 struct sched_group
*group_min
; /* Least loaded group in sd */
3416 struct sched_group
*group_leader
; /* Group which relieves group_min */
3417 unsigned long min_load_per_task
; /* load_per_task in group_min */
3418 unsigned long leader_nr_running
; /* Nr running of group_leader */
3419 unsigned long min_nr_running
; /* Nr running of group_min */
3424 * sg_lb_stats - stats of a sched_group required for load_balancing
3426 struct sg_lb_stats
{
3427 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3428 unsigned long group_load
; /* Total load over the CPUs of the group */
3429 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3430 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3431 unsigned long group_capacity
;
3432 int group_imb
; /* Is there an imbalance in the group ? */
3436 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3437 * @group: The group whose first cpu is to be returned.
3439 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3441 return cpumask_first(sched_group_cpus(group
));
3445 * get_sd_load_idx - Obtain the load index for a given sched domain.
3446 * @sd: The sched_domain whose load_idx is to be obtained.
3447 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3449 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3450 enum cpu_idle_type idle
)
3456 load_idx
= sd
->busy_idx
;
3459 case CPU_NEWLY_IDLE
:
3460 load_idx
= sd
->newidle_idx
;
3463 load_idx
= sd
->idle_idx
;
3471 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3473 * init_sd_power_savings_stats - Initialize power savings statistics for
3474 * the given sched_domain, during load balancing.
3476 * @sd: Sched domain whose power-savings statistics are to be initialized.
3477 * @sds: Variable containing the statistics for sd.
3478 * @idle: Idle status of the CPU at which we're performing load-balancing.
3480 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3481 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3484 * Busy processors will not participate in power savings
3487 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3488 sds
->power_savings_balance
= 0;
3490 sds
->power_savings_balance
= 1;
3491 sds
->min_nr_running
= ULONG_MAX
;
3492 sds
->leader_nr_running
= 0;
3497 * update_sd_power_savings_stats - Update the power saving stats for a
3498 * sched_domain while performing load balancing.
3500 * @group: sched_group belonging to the sched_domain under consideration.
3501 * @sds: Variable containing the statistics of the sched_domain
3502 * @local_group: Does group contain the CPU for which we're performing
3504 * @sgs: Variable containing the statistics of the group.
3506 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3507 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3510 if (!sds
->power_savings_balance
)
3514 * If the local group is idle or completely loaded
3515 * no need to do power savings balance at this domain
3517 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3518 !sds
->this_nr_running
))
3519 sds
->power_savings_balance
= 0;
3522 * If a group is already running at full capacity or idle,
3523 * don't include that group in power savings calculations
3525 if (!sds
->power_savings_balance
||
3526 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3527 !sgs
->sum_nr_running
)
3531 * Calculate the group which has the least non-idle load.
3532 * This is the group from where we need to pick up the load
3535 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3536 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3537 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3538 sds
->group_min
= group
;
3539 sds
->min_nr_running
= sgs
->sum_nr_running
;
3540 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3541 sgs
->sum_nr_running
;
3545 * Calculate the group which is almost near its
3546 * capacity but still has some space to pick up some load
3547 * from other group and save more power
3549 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3552 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3553 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3554 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3555 sds
->group_leader
= group
;
3556 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3561 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3562 * @sds: Variable containing the statistics of the sched_domain
3563 * under consideration.
3564 * @this_cpu: Cpu at which we're currently performing load-balancing.
3565 * @imbalance: Variable to store the imbalance.
3568 * Check if we have potential to perform some power-savings balance.
3569 * If yes, set the busiest group to be the least loaded group in the
3570 * sched_domain, so that it's CPUs can be put to idle.
3572 * Returns 1 if there is potential to perform power-savings balance.
3575 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3576 int this_cpu
, unsigned long *imbalance
)
3578 if (!sds
->power_savings_balance
)
3581 if (sds
->this != sds
->group_leader
||
3582 sds
->group_leader
== sds
->group_min
)
3585 *imbalance
= sds
->min_load_per_task
;
3586 sds
->busiest
= sds
->group_min
;
3591 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3592 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3593 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3598 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3599 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3604 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3605 int this_cpu
, unsigned long *imbalance
)
3609 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3612 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3614 return SCHED_LOAD_SCALE
;
3617 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3619 return default_scale_freq_power(sd
, cpu
);
3622 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3624 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3625 unsigned long smt_gain
= sd
->smt_gain
;
3632 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3634 return default_scale_smt_power(sd
, cpu
);
3637 unsigned long scale_rt_power(int cpu
)
3639 struct rq
*rq
= cpu_rq(cpu
);
3640 u64 total
, available
;
3642 sched_avg_update(rq
);
3644 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3645 available
= total
- rq
->rt_avg
;
3647 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3648 total
= SCHED_LOAD_SCALE
;
3650 total
>>= SCHED_LOAD_SHIFT
;
3652 return div_u64(available
, total
);
3655 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3657 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3658 unsigned long power
= SCHED_LOAD_SCALE
;
3659 struct sched_group
*sdg
= sd
->groups
;
3661 if (sched_feat(ARCH_POWER
))
3662 power
*= arch_scale_freq_power(sd
, cpu
);
3664 power
*= default_scale_freq_power(sd
, cpu
);
3666 power
>>= SCHED_LOAD_SHIFT
;
3668 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3669 if (sched_feat(ARCH_POWER
))
3670 power
*= arch_scale_smt_power(sd
, cpu
);
3672 power
*= default_scale_smt_power(sd
, cpu
);
3674 power
>>= SCHED_LOAD_SHIFT
;
3677 power
*= scale_rt_power(cpu
);
3678 power
>>= SCHED_LOAD_SHIFT
;
3683 sdg
->cpu_power
= power
;
3686 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3688 struct sched_domain
*child
= sd
->child
;
3689 struct sched_group
*group
, *sdg
= sd
->groups
;
3690 unsigned long power
;
3693 update_cpu_power(sd
, cpu
);
3699 group
= child
->groups
;
3701 power
+= group
->cpu_power
;
3702 group
= group
->next
;
3703 } while (group
!= child
->groups
);
3705 sdg
->cpu_power
= power
;
3709 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3710 * @sd: The sched_domain whose statistics are to be updated.
3711 * @group: sched_group whose statistics are to be updated.
3712 * @this_cpu: Cpu for which load balance is currently performed.
3713 * @idle: Idle status of this_cpu
3714 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3715 * @sd_idle: Idle status of the sched_domain containing group.
3716 * @local_group: Does group contain this_cpu.
3717 * @cpus: Set of cpus considered for load balancing.
3718 * @balance: Should we balance.
3719 * @sgs: variable to hold the statistics for this group.
3721 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3722 struct sched_group
*group
, int this_cpu
,
3723 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3724 int local_group
, const struct cpumask
*cpus
,
3725 int *balance
, struct sg_lb_stats
*sgs
)
3727 unsigned long load
, max_cpu_load
, min_cpu_load
;
3729 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3730 unsigned long avg_load_per_task
= 0;
3733 balance_cpu
= group_first_cpu(group
);
3734 if (balance_cpu
== this_cpu
)
3735 update_group_power(sd
, this_cpu
);
3738 /* Tally up the load of all CPUs in the group */
3740 min_cpu_load
= ~0UL;
3742 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3743 struct rq
*rq
= cpu_rq(i
);
3745 if (*sd_idle
&& rq
->nr_running
)
3748 /* Bias balancing toward cpus of our domain */
3750 if (idle_cpu(i
) && !first_idle_cpu
) {
3755 load
= target_load(i
, load_idx
);
3757 load
= source_load(i
, load_idx
);
3758 if (load
> max_cpu_load
)
3759 max_cpu_load
= load
;
3760 if (min_cpu_load
> load
)
3761 min_cpu_load
= load
;
3764 sgs
->group_load
+= load
;
3765 sgs
->sum_nr_running
+= rq
->nr_running
;
3766 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3771 * First idle cpu or the first cpu(busiest) in this sched group
3772 * is eligible for doing load balancing at this and above
3773 * domains. In the newly idle case, we will allow all the cpu's
3774 * to do the newly idle load balance.
3776 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3777 balance_cpu
!= this_cpu
&& balance
) {
3782 /* Adjust by relative CPU power of the group */
3783 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3786 * Consider the group unbalanced when the imbalance is larger
3787 * than the average weight of two tasks.
3789 * APZ: with cgroup the avg task weight can vary wildly and
3790 * might not be a suitable number - should we keep a
3791 * normalized nr_running number somewhere that negates
3794 if (sgs
->sum_nr_running
)
3795 avg_load_per_task
= sgs
->sum_weighted_load
/ sgs
->sum_nr_running
;
3797 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3800 sgs
->group_capacity
=
3801 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3805 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3806 * @sd: sched_domain whose statistics are to be updated.
3807 * @this_cpu: Cpu for which load balance is currently performed.
3808 * @idle: Idle status of this_cpu
3809 * @sd_idle: Idle status of the sched_domain containing group.
3810 * @cpus: Set of cpus considered for load balancing.
3811 * @balance: Should we balance.
3812 * @sds: variable to hold the statistics for this sched_domain.
3814 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3815 enum cpu_idle_type idle
, int *sd_idle
,
3816 const struct cpumask
*cpus
, int *balance
,
3817 struct sd_lb_stats
*sds
)
3819 struct sched_domain
*child
= sd
->child
;
3820 struct sched_group
*group
= sd
->groups
;
3821 struct sg_lb_stats sgs
;
3822 int load_idx
, prefer_sibling
= 0;
3824 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3827 init_sd_power_savings_stats(sd
, sds
, idle
);
3828 load_idx
= get_sd_load_idx(sd
, idle
);
3833 local_group
= cpumask_test_cpu(this_cpu
,
3834 sched_group_cpus(group
));
3835 memset(&sgs
, 0, sizeof(sgs
));
3836 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3837 local_group
, cpus
, balance
, &sgs
);
3839 if (local_group
&& balance
&& !(*balance
))
3842 sds
->total_load
+= sgs
.group_load
;
3843 sds
->total_pwr
+= group
->cpu_power
;
3846 * In case the child domain prefers tasks go to siblings
3847 * first, lower the group capacity to one so that we'll try
3848 * and move all the excess tasks away.
3851 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3854 sds
->this_load
= sgs
.avg_load
;
3856 sds
->this_nr_running
= sgs
.sum_nr_running
;
3857 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3858 } else if (sgs
.avg_load
> sds
->max_load
&&
3859 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3861 sds
->max_load
= sgs
.avg_load
;
3862 sds
->busiest
= group
;
3863 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3864 sds
->busiest_group_capacity
= sgs
.group_capacity
;
3865 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3866 sds
->group_imb
= sgs
.group_imb
;
3869 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3870 group
= group
->next
;
3871 } while (group
!= sd
->groups
);
3875 * fix_small_imbalance - Calculate the minor imbalance that exists
3876 * amongst the groups of a sched_domain, during
3878 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3879 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3880 * @imbalance: Variable to store the imbalance.
3882 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3883 int this_cpu
, unsigned long *imbalance
)
3885 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3886 unsigned int imbn
= 2;
3887 unsigned long scaled_busy_load_per_task
;
3889 if (sds
->this_nr_running
) {
3890 sds
->this_load_per_task
/= sds
->this_nr_running
;
3891 if (sds
->busiest_load_per_task
>
3892 sds
->this_load_per_task
)
3895 sds
->this_load_per_task
=
3896 cpu_avg_load_per_task(this_cpu
);
3898 scaled_busy_load_per_task
= sds
->busiest_load_per_task
3900 scaled_busy_load_per_task
/= sds
->busiest
->cpu_power
;
3902 if (sds
->max_load
- sds
->this_load
+ scaled_busy_load_per_task
>=
3903 (scaled_busy_load_per_task
* imbn
)) {
3904 *imbalance
= sds
->busiest_load_per_task
;
3909 * OK, we don't have enough imbalance to justify moving tasks,
3910 * however we may be able to increase total CPU power used by
3914 pwr_now
+= sds
->busiest
->cpu_power
*
3915 min(sds
->busiest_load_per_task
, sds
->max_load
);
3916 pwr_now
+= sds
->this->cpu_power
*
3917 min(sds
->this_load_per_task
, sds
->this_load
);
3918 pwr_now
/= SCHED_LOAD_SCALE
;
3920 /* Amount of load we'd subtract */
3921 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3922 sds
->busiest
->cpu_power
;
3923 if (sds
->max_load
> tmp
)
3924 pwr_move
+= sds
->busiest
->cpu_power
*
3925 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3927 /* Amount of load we'd add */
3928 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3929 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3930 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3931 sds
->this->cpu_power
;
3933 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3934 sds
->this->cpu_power
;
3935 pwr_move
+= sds
->this->cpu_power
*
3936 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3937 pwr_move
/= SCHED_LOAD_SCALE
;
3939 /* Move if we gain throughput */
3940 if (pwr_move
> pwr_now
)
3941 *imbalance
= sds
->busiest_load_per_task
;
3945 * calculate_imbalance - Calculate the amount of imbalance present within the
3946 * groups of a given sched_domain during load balance.
3947 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3948 * @this_cpu: Cpu for which currently load balance is being performed.
3949 * @imbalance: The variable to store the imbalance.
3951 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3952 unsigned long *imbalance
)
3954 unsigned long max_pull
, load_above_capacity
= ~0UL;
3956 sds
->busiest_load_per_task
/= sds
->busiest_nr_running
;
3957 if (sds
->group_imb
) {
3958 sds
->busiest_load_per_task
=
3959 min(sds
->busiest_load_per_task
, sds
->avg_load
);
3963 * In the presence of smp nice balancing, certain scenarios can have
3964 * max load less than avg load(as we skip the groups at or below
3965 * its cpu_power, while calculating max_load..)
3967 if (sds
->max_load
< sds
->avg_load
) {
3969 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3972 if (!sds
->group_imb
) {
3974 * Don't want to pull so many tasks that a group would go idle.
3976 load_above_capacity
= (sds
->busiest_nr_running
-
3977 sds
->busiest_group_capacity
);
3979 load_above_capacity
*= (SCHED_LOAD_SCALE
* SCHED_LOAD_SCALE
);
3981 load_above_capacity
/= sds
->busiest
->cpu_power
;
3985 * We're trying to get all the cpus to the average_load, so we don't
3986 * want to push ourselves above the average load, nor do we wish to
3987 * reduce the max loaded cpu below the average load. At the same time,
3988 * we also don't want to reduce the group load below the group capacity
3989 * (so that we can implement power-savings policies etc). Thus we look
3990 * for the minimum possible imbalance.
3991 * Be careful of negative numbers as they'll appear as very large values
3992 * with unsigned longs.
3994 max_pull
= min(sds
->max_load
- sds
->avg_load
, load_above_capacity
);
3996 /* How much load to actually move to equalise the imbalance */
3997 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3998 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
4002 * if *imbalance is less than the average load per runnable task
4003 * there is no gaurantee that any tasks will be moved so we'll have
4004 * a think about bumping its value to force at least one task to be
4007 if (*imbalance
< sds
->busiest_load_per_task
)
4008 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4011 /******* find_busiest_group() helpers end here *********************/
4014 * find_busiest_group - Returns the busiest group within the sched_domain
4015 * if there is an imbalance. If there isn't an imbalance, and
4016 * the user has opted for power-savings, it returns a group whose
4017 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4018 * such a group exists.
4020 * Also calculates the amount of weighted load which should be moved
4021 * to restore balance.
4023 * @sd: The sched_domain whose busiest group is to be returned.
4024 * @this_cpu: The cpu for which load balancing is currently being performed.
4025 * @imbalance: Variable which stores amount of weighted load which should
4026 * be moved to restore balance/put a group to idle.
4027 * @idle: The idle status of this_cpu.
4028 * @sd_idle: The idleness of sd
4029 * @cpus: The set of CPUs under consideration for load-balancing.
4030 * @balance: Pointer to a variable indicating if this_cpu
4031 * is the appropriate cpu to perform load balancing at this_level.
4033 * Returns: - the busiest group if imbalance exists.
4034 * - If no imbalance and user has opted for power-savings balance,
4035 * return the least loaded group whose CPUs can be
4036 * put to idle by rebalancing its tasks onto our group.
4038 static struct sched_group
*
4039 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4040 unsigned long *imbalance
, enum cpu_idle_type idle
,
4041 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4043 struct sd_lb_stats sds
;
4045 memset(&sds
, 0, sizeof(sds
));
4048 * Compute the various statistics relavent for load balancing at
4051 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4054 /* Cases where imbalance does not exist from POV of this_cpu */
4055 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4057 * 2) There is no busy sibling group to pull from.
4058 * 3) This group is the busiest group.
4059 * 4) This group is more busy than the avg busieness at this
4061 * 5) The imbalance is within the specified limit.
4063 if (balance
&& !(*balance
))
4066 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4069 if (sds
.this_load
>= sds
.max_load
)
4072 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4074 if (sds
.this_load
>= sds
.avg_load
)
4077 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4080 /* Looks like there is an imbalance. Compute it */
4081 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4086 * There is no obvious imbalance. But check if we can do some balancing
4089 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4097 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4100 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4101 unsigned long imbalance
, const struct cpumask
*cpus
)
4103 struct rq
*busiest
= NULL
, *rq
;
4104 unsigned long max_load
= 0;
4107 for_each_cpu(i
, sched_group_cpus(group
)) {
4108 unsigned long power
= power_of(i
);
4109 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4112 if (!cpumask_test_cpu(i
, cpus
))
4116 wl
= weighted_cpuload(i
);
4119 * When comparing with imbalance, use weighted_cpuload()
4120 * which is not scaled with the cpu power.
4122 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4126 * For the load comparisons with the other cpu's, consider
4127 * the weighted_cpuload() scaled with the cpu power, so that
4128 * the load can be moved away from the cpu that is potentially
4129 * running at a lower capacity.
4131 wl
= (wl
* SCHED_LOAD_SCALE
) / power
;
4133 if (wl
> max_load
) {
4143 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4144 * so long as it is large enough.
4146 #define MAX_PINNED_INTERVAL 512
4148 /* Working cpumask for load_balance and load_balance_newidle. */
4149 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4152 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4153 * tasks if there is an imbalance.
4155 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4156 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4159 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4160 struct sched_group
*group
;
4161 unsigned long imbalance
;
4163 unsigned long flags
;
4164 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4166 cpumask_copy(cpus
, cpu_active_mask
);
4169 * When power savings policy is enabled for the parent domain, idle
4170 * sibling can pick up load irrespective of busy siblings. In this case,
4171 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4172 * portraying it as CPU_NOT_IDLE.
4174 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4175 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4178 schedstat_inc(sd
, lb_count
[idle
]);
4182 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4189 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4193 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4195 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4199 BUG_ON(busiest
== this_rq
);
4201 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4204 if (busiest
->nr_running
> 1) {
4206 * Attempt to move tasks. If find_busiest_group has found
4207 * an imbalance but busiest->nr_running <= 1, the group is
4208 * still unbalanced. ld_moved simply stays zero, so it is
4209 * correctly treated as an imbalance.
4211 local_irq_save(flags
);
4212 double_rq_lock(this_rq
, busiest
);
4213 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4214 imbalance
, sd
, idle
, &all_pinned
);
4215 double_rq_unlock(this_rq
, busiest
);
4216 local_irq_restore(flags
);
4219 * some other cpu did the load balance for us.
4221 if (ld_moved
&& this_cpu
!= smp_processor_id())
4222 resched_cpu(this_cpu
);
4224 /* All tasks on this runqueue were pinned by CPU affinity */
4225 if (unlikely(all_pinned
)) {
4226 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4227 if (!cpumask_empty(cpus
))
4234 schedstat_inc(sd
, lb_failed
[idle
]);
4235 sd
->nr_balance_failed
++;
4237 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4239 spin_lock_irqsave(&busiest
->lock
, flags
);
4241 /* don't kick the migration_thread, if the curr
4242 * task on busiest cpu can't be moved to this_cpu
4244 if (!cpumask_test_cpu(this_cpu
,
4245 &busiest
->curr
->cpus_allowed
)) {
4246 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4248 goto out_one_pinned
;
4251 if (!busiest
->active_balance
) {
4252 busiest
->active_balance
= 1;
4253 busiest
->push_cpu
= this_cpu
;
4256 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4258 wake_up_process(busiest
->migration_thread
);
4261 * We've kicked active balancing, reset the failure
4264 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4267 sd
->nr_balance_failed
= 0;
4269 if (likely(!active_balance
)) {
4270 /* We were unbalanced, so reset the balancing interval */
4271 sd
->balance_interval
= sd
->min_interval
;
4274 * If we've begun active balancing, start to back off. This
4275 * case may not be covered by the all_pinned logic if there
4276 * is only 1 task on the busy runqueue (because we don't call
4279 if (sd
->balance_interval
< sd
->max_interval
)
4280 sd
->balance_interval
*= 2;
4283 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4284 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4290 schedstat_inc(sd
, lb_balanced
[idle
]);
4292 sd
->nr_balance_failed
= 0;
4295 /* tune up the balancing interval */
4296 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4297 (sd
->balance_interval
< sd
->max_interval
))
4298 sd
->balance_interval
*= 2;
4300 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4301 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4312 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4313 * tasks if there is an imbalance.
4315 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4316 * this_rq is locked.
4319 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4321 struct sched_group
*group
;
4322 struct rq
*busiest
= NULL
;
4323 unsigned long imbalance
;
4327 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4329 cpumask_copy(cpus
, cpu_active_mask
);
4332 * When power savings policy is enabled for the parent domain, idle
4333 * sibling can pick up load irrespective of busy siblings. In this case,
4334 * let the state of idle sibling percolate up as IDLE, instead of
4335 * portraying it as CPU_NOT_IDLE.
4337 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4338 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4341 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4343 update_shares_locked(this_rq
, sd
);
4344 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4345 &sd_idle
, cpus
, NULL
);
4347 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4351 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4353 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4357 BUG_ON(busiest
== this_rq
);
4359 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4362 if (busiest
->nr_running
> 1) {
4363 /* Attempt to move tasks */
4364 double_lock_balance(this_rq
, busiest
);
4365 /* this_rq->clock is already updated */
4366 update_rq_clock(busiest
);
4367 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4368 imbalance
, sd
, CPU_NEWLY_IDLE
,
4370 double_unlock_balance(this_rq
, busiest
);
4372 if (unlikely(all_pinned
)) {
4373 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4374 if (!cpumask_empty(cpus
))
4380 int active_balance
= 0;
4382 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4383 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4384 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4387 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4390 if (sd
->nr_balance_failed
++ < 2)
4394 * The only task running in a non-idle cpu can be moved to this
4395 * cpu in an attempt to completely freeup the other CPU
4396 * package. The same method used to move task in load_balance()
4397 * have been extended for load_balance_newidle() to speedup
4398 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4400 * The package power saving logic comes from
4401 * find_busiest_group(). If there are no imbalance, then
4402 * f_b_g() will return NULL. However when sched_mc={1,2} then
4403 * f_b_g() will select a group from which a running task may be
4404 * pulled to this cpu in order to make the other package idle.
4405 * If there is no opportunity to make a package idle and if
4406 * there are no imbalance, then f_b_g() will return NULL and no
4407 * action will be taken in load_balance_newidle().
4409 * Under normal task pull operation due to imbalance, there
4410 * will be more than one task in the source run queue and
4411 * move_tasks() will succeed. ld_moved will be true and this
4412 * active balance code will not be triggered.
4415 /* Lock busiest in correct order while this_rq is held */
4416 double_lock_balance(this_rq
, busiest
);
4419 * don't kick the migration_thread, if the curr
4420 * task on busiest cpu can't be moved to this_cpu
4422 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4423 double_unlock_balance(this_rq
, busiest
);
4428 if (!busiest
->active_balance
) {
4429 busiest
->active_balance
= 1;
4430 busiest
->push_cpu
= this_cpu
;
4434 double_unlock_balance(this_rq
, busiest
);
4436 * Should not call ttwu while holding a rq->lock
4438 spin_unlock(&this_rq
->lock
);
4440 wake_up_process(busiest
->migration_thread
);
4441 spin_lock(&this_rq
->lock
);
4444 sd
->nr_balance_failed
= 0;
4446 update_shares_locked(this_rq
, sd
);
4450 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4451 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4452 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4454 sd
->nr_balance_failed
= 0;
4460 * idle_balance is called by schedule() if this_cpu is about to become
4461 * idle. Attempts to pull tasks from other CPUs.
4463 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4465 struct sched_domain
*sd
;
4466 int pulled_task
= 0;
4467 unsigned long next_balance
= jiffies
+ HZ
;
4469 this_rq
->idle_stamp
= this_rq
->clock
;
4471 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4474 for_each_domain(this_cpu
, sd
) {
4475 unsigned long interval
;
4477 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4480 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4481 /* If we've pulled tasks over stop searching: */
4482 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4485 interval
= msecs_to_jiffies(sd
->balance_interval
);
4486 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4487 next_balance
= sd
->last_balance
+ interval
;
4489 this_rq
->idle_stamp
= 0;
4493 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4495 * We are going idle. next_balance may be set based on
4496 * a busy processor. So reset next_balance.
4498 this_rq
->next_balance
= next_balance
;
4503 * active_load_balance is run by migration threads. It pushes running tasks
4504 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4505 * running on each physical CPU where possible, and avoids physical /
4506 * logical imbalances.
4508 * Called with busiest_rq locked.
4510 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4512 int target_cpu
= busiest_rq
->push_cpu
;
4513 struct sched_domain
*sd
;
4514 struct rq
*target_rq
;
4516 /* Is there any task to move? */
4517 if (busiest_rq
->nr_running
<= 1)
4520 target_rq
= cpu_rq(target_cpu
);
4523 * This condition is "impossible", if it occurs
4524 * we need to fix it. Originally reported by
4525 * Bjorn Helgaas on a 128-cpu setup.
4527 BUG_ON(busiest_rq
== target_rq
);
4529 /* move a task from busiest_rq to target_rq */
4530 double_lock_balance(busiest_rq
, target_rq
);
4531 update_rq_clock(busiest_rq
);
4532 update_rq_clock(target_rq
);
4534 /* Search for an sd spanning us and the target CPU. */
4535 for_each_domain(target_cpu
, sd
) {
4536 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4537 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4542 schedstat_inc(sd
, alb_count
);
4544 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4546 schedstat_inc(sd
, alb_pushed
);
4548 schedstat_inc(sd
, alb_failed
);
4550 double_unlock_balance(busiest_rq
, target_rq
);
4555 atomic_t load_balancer
;
4556 cpumask_var_t cpu_mask
;
4557 cpumask_var_t ilb_grp_nohz_mask
;
4558 } nohz ____cacheline_aligned
= {
4559 .load_balancer
= ATOMIC_INIT(-1),
4562 int get_nohz_load_balancer(void)
4564 return atomic_read(&nohz
.load_balancer
);
4567 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4569 * lowest_flag_domain - Return lowest sched_domain containing flag.
4570 * @cpu: The cpu whose lowest level of sched domain is to
4572 * @flag: The flag to check for the lowest sched_domain
4573 * for the given cpu.
4575 * Returns the lowest sched_domain of a cpu which contains the given flag.
4577 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4579 struct sched_domain
*sd
;
4581 for_each_domain(cpu
, sd
)
4582 if (sd
&& (sd
->flags
& flag
))
4589 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4590 * @cpu: The cpu whose domains we're iterating over.
4591 * @sd: variable holding the value of the power_savings_sd
4593 * @flag: The flag to filter the sched_domains to be iterated.
4595 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4596 * set, starting from the lowest sched_domain to the highest.
4598 #define for_each_flag_domain(cpu, sd, flag) \
4599 for (sd = lowest_flag_domain(cpu, flag); \
4600 (sd && (sd->flags & flag)); sd = sd->parent)
4603 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4604 * @ilb_group: group to be checked for semi-idleness
4606 * Returns: 1 if the group is semi-idle. 0 otherwise.
4608 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4609 * and atleast one non-idle CPU. This helper function checks if the given
4610 * sched_group is semi-idle or not.
4612 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4614 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4615 sched_group_cpus(ilb_group
));
4618 * A sched_group is semi-idle when it has atleast one busy cpu
4619 * and atleast one idle cpu.
4621 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4624 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4630 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4631 * @cpu: The cpu which is nominating a new idle_load_balancer.
4633 * Returns: Returns the id of the idle load balancer if it exists,
4634 * Else, returns >= nr_cpu_ids.
4636 * This algorithm picks the idle load balancer such that it belongs to a
4637 * semi-idle powersavings sched_domain. The idea is to try and avoid
4638 * completely idle packages/cores just for the purpose of idle load balancing
4639 * when there are other idle cpu's which are better suited for that job.
4641 static int find_new_ilb(int cpu
)
4643 struct sched_domain
*sd
;
4644 struct sched_group
*ilb_group
;
4647 * Have idle load balancer selection from semi-idle packages only
4648 * when power-aware load balancing is enabled
4650 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4654 * Optimize for the case when we have no idle CPUs or only one
4655 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4657 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4660 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4661 ilb_group
= sd
->groups
;
4664 if (is_semi_idle_group(ilb_group
))
4665 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4667 ilb_group
= ilb_group
->next
;
4669 } while (ilb_group
!= sd
->groups
);
4673 return cpumask_first(nohz
.cpu_mask
);
4675 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4676 static inline int find_new_ilb(int call_cpu
)
4678 return cpumask_first(nohz
.cpu_mask
);
4683 * This routine will try to nominate the ilb (idle load balancing)
4684 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4685 * load balancing on behalf of all those cpus. If all the cpus in the system
4686 * go into this tickless mode, then there will be no ilb owner (as there is
4687 * no need for one) and all the cpus will sleep till the next wakeup event
4690 * For the ilb owner, tick is not stopped. And this tick will be used
4691 * for idle load balancing. ilb owner will still be part of
4694 * While stopping the tick, this cpu will become the ilb owner if there
4695 * is no other owner. And will be the owner till that cpu becomes busy
4696 * or if all cpus in the system stop their ticks at which point
4697 * there is no need for ilb owner.
4699 * When the ilb owner becomes busy, it nominates another owner, during the
4700 * next busy scheduler_tick()
4702 int select_nohz_load_balancer(int stop_tick
)
4704 int cpu
= smp_processor_id();
4707 cpu_rq(cpu
)->in_nohz_recently
= 1;
4709 if (!cpu_active(cpu
)) {
4710 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4714 * If we are going offline and still the leader,
4717 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4723 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4725 /* time for ilb owner also to sleep */
4726 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
4727 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4728 atomic_set(&nohz
.load_balancer
, -1);
4732 if (atomic_read(&nohz
.load_balancer
) == -1) {
4733 /* make me the ilb owner */
4734 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4736 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4739 if (!(sched_smt_power_savings
||
4740 sched_mc_power_savings
))
4743 * Check to see if there is a more power-efficient
4746 new_ilb
= find_new_ilb(cpu
);
4747 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4748 atomic_set(&nohz
.load_balancer
, -1);
4749 resched_cpu(new_ilb
);
4755 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4758 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4760 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4761 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4768 static DEFINE_SPINLOCK(balancing
);
4771 * It checks each scheduling domain to see if it is due to be balanced,
4772 * and initiates a balancing operation if so.
4774 * Balancing parameters are set up in arch_init_sched_domains.
4776 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4779 struct rq
*rq
= cpu_rq(cpu
);
4780 unsigned long interval
;
4781 struct sched_domain
*sd
;
4782 /* Earliest time when we have to do rebalance again */
4783 unsigned long next_balance
= jiffies
+ 60*HZ
;
4784 int update_next_balance
= 0;
4787 for_each_domain(cpu
, sd
) {
4788 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4791 interval
= sd
->balance_interval
;
4792 if (idle
!= CPU_IDLE
)
4793 interval
*= sd
->busy_factor
;
4795 /* scale ms to jiffies */
4796 interval
= msecs_to_jiffies(interval
);
4797 if (unlikely(!interval
))
4799 if (interval
> HZ
*NR_CPUS
/10)
4800 interval
= HZ
*NR_CPUS
/10;
4802 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4804 if (need_serialize
) {
4805 if (!spin_trylock(&balancing
))
4809 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4810 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4812 * We've pulled tasks over so either we're no
4813 * longer idle, or one of our SMT siblings is
4816 idle
= CPU_NOT_IDLE
;
4818 sd
->last_balance
= jiffies
;
4821 spin_unlock(&balancing
);
4823 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4824 next_balance
= sd
->last_balance
+ interval
;
4825 update_next_balance
= 1;
4829 * Stop the load balance at this level. There is another
4830 * CPU in our sched group which is doing load balancing more
4838 * next_balance will be updated only when there is a need.
4839 * When the cpu is attached to null domain for ex, it will not be
4842 if (likely(update_next_balance
))
4843 rq
->next_balance
= next_balance
;
4847 * run_rebalance_domains is triggered when needed from the scheduler tick.
4848 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4849 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4851 static void run_rebalance_domains(struct softirq_action
*h
)
4853 int this_cpu
= smp_processor_id();
4854 struct rq
*this_rq
= cpu_rq(this_cpu
);
4855 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4856 CPU_IDLE
: CPU_NOT_IDLE
;
4858 rebalance_domains(this_cpu
, idle
);
4862 * If this cpu is the owner for idle load balancing, then do the
4863 * balancing on behalf of the other idle cpus whose ticks are
4866 if (this_rq
->idle_at_tick
&&
4867 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4871 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4872 if (balance_cpu
== this_cpu
)
4876 * If this cpu gets work to do, stop the load balancing
4877 * work being done for other cpus. Next load
4878 * balancing owner will pick it up.
4883 rebalance_domains(balance_cpu
, CPU_IDLE
);
4885 rq
= cpu_rq(balance_cpu
);
4886 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4887 this_rq
->next_balance
= rq
->next_balance
;
4893 static inline int on_null_domain(int cpu
)
4895 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4899 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4901 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4902 * idle load balancing owner or decide to stop the periodic load balancing,
4903 * if the whole system is idle.
4905 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4909 * If we were in the nohz mode recently and busy at the current
4910 * scheduler tick, then check if we need to nominate new idle
4913 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4914 rq
->in_nohz_recently
= 0;
4916 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4917 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4918 atomic_set(&nohz
.load_balancer
, -1);
4921 if (atomic_read(&nohz
.load_balancer
) == -1) {
4922 int ilb
= find_new_ilb(cpu
);
4924 if (ilb
< nr_cpu_ids
)
4930 * If this cpu is idle and doing idle load balancing for all the
4931 * cpus with ticks stopped, is it time for that to stop?
4933 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4934 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4940 * If this cpu is idle and the idle load balancing is done by
4941 * someone else, then no need raise the SCHED_SOFTIRQ
4943 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4944 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4947 /* Don't need to rebalance while attached to NULL domain */
4948 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4949 likely(!on_null_domain(cpu
)))
4950 raise_softirq(SCHED_SOFTIRQ
);
4953 #else /* CONFIG_SMP */
4956 * on UP we do not need to balance between CPUs:
4958 static inline void idle_balance(int cpu
, struct rq
*rq
)
4964 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4966 EXPORT_PER_CPU_SYMBOL(kstat
);
4969 * Return any ns on the sched_clock that have not yet been accounted in
4970 * @p in case that task is currently running.
4972 * Called with task_rq_lock() held on @rq.
4974 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4978 if (task_current(rq
, p
)) {
4979 update_rq_clock(rq
);
4980 ns
= rq
->clock
- p
->se
.exec_start
;
4988 unsigned long long task_delta_exec(struct task_struct
*p
)
4990 unsigned long flags
;
4994 rq
= task_rq_lock(p
, &flags
);
4995 ns
= do_task_delta_exec(p
, rq
);
4996 task_rq_unlock(rq
, &flags
);
5002 * Return accounted runtime for the task.
5003 * In case the task is currently running, return the runtime plus current's
5004 * pending runtime that have not been accounted yet.
5006 unsigned long long task_sched_runtime(struct task_struct
*p
)
5008 unsigned long flags
;
5012 rq
= task_rq_lock(p
, &flags
);
5013 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5014 task_rq_unlock(rq
, &flags
);
5020 * Return sum_exec_runtime for the thread group.
5021 * In case the task is currently running, return the sum plus current's
5022 * pending runtime that have not been accounted yet.
5024 * Note that the thread group might have other running tasks as well,
5025 * so the return value not includes other pending runtime that other
5026 * running tasks might have.
5028 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5030 struct task_cputime totals
;
5031 unsigned long flags
;
5035 rq
= task_rq_lock(p
, &flags
);
5036 thread_group_cputime(p
, &totals
);
5037 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5038 task_rq_unlock(rq
, &flags
);
5044 * Account user cpu time to a process.
5045 * @p: the process that the cpu time gets accounted to
5046 * @cputime: the cpu time spent in user space since the last update
5047 * @cputime_scaled: cputime scaled by cpu frequency
5049 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5050 cputime_t cputime_scaled
)
5052 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5055 /* Add user time to process. */
5056 p
->utime
= cputime_add(p
->utime
, cputime
);
5057 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5058 account_group_user_time(p
, cputime
);
5060 /* Add user time to cpustat. */
5061 tmp
= cputime_to_cputime64(cputime
);
5062 if (TASK_NICE(p
) > 0)
5063 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5065 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5067 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5068 /* Account for user time used */
5069 acct_update_integrals(p
);
5073 * Account guest cpu time to a process.
5074 * @p: the process that the cpu time gets accounted to
5075 * @cputime: the cpu time spent in virtual machine since the last update
5076 * @cputime_scaled: cputime scaled by cpu frequency
5078 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5079 cputime_t cputime_scaled
)
5082 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5084 tmp
= cputime_to_cputime64(cputime
);
5086 /* Add guest time to process. */
5087 p
->utime
= cputime_add(p
->utime
, cputime
);
5088 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5089 account_group_user_time(p
, cputime
);
5090 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5092 /* Add guest time to cpustat. */
5093 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5094 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5098 * Account system cpu time to a process.
5099 * @p: the process that the cpu time gets accounted to
5100 * @hardirq_offset: the offset to subtract from hardirq_count()
5101 * @cputime: the cpu time spent in kernel space since the last update
5102 * @cputime_scaled: cputime scaled by cpu frequency
5104 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5105 cputime_t cputime
, cputime_t cputime_scaled
)
5107 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5110 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5111 account_guest_time(p
, cputime
, cputime_scaled
);
5115 /* Add system time to process. */
5116 p
->stime
= cputime_add(p
->stime
, cputime
);
5117 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5118 account_group_system_time(p
, cputime
);
5120 /* Add system time to cpustat. */
5121 tmp
= cputime_to_cputime64(cputime
);
5122 if (hardirq_count() - hardirq_offset
)
5123 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5124 else if (softirq_count())
5125 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5127 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5129 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5131 /* Account for system time used */
5132 acct_update_integrals(p
);
5136 * Account for involuntary wait time.
5137 * @steal: the cpu time spent in involuntary wait
5139 void account_steal_time(cputime_t cputime
)
5141 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5142 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5144 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5148 * Account for idle time.
5149 * @cputime: the cpu time spent in idle wait
5151 void account_idle_time(cputime_t cputime
)
5153 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5154 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5155 struct rq
*rq
= this_rq();
5157 if (atomic_read(&rq
->nr_iowait
) > 0)
5158 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5160 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5163 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5166 * Account a single tick of cpu time.
5167 * @p: the process that the cpu time gets accounted to
5168 * @user_tick: indicates if the tick is a user or a system tick
5170 void account_process_tick(struct task_struct
*p
, int user_tick
)
5172 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5173 struct rq
*rq
= this_rq();
5176 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5177 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5178 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5181 account_idle_time(cputime_one_jiffy
);
5185 * Account multiple ticks of steal time.
5186 * @p: the process from which the cpu time has been stolen
5187 * @ticks: number of stolen ticks
5189 void account_steal_ticks(unsigned long ticks
)
5191 account_steal_time(jiffies_to_cputime(ticks
));
5195 * Account multiple ticks of idle time.
5196 * @ticks: number of stolen ticks
5198 void account_idle_ticks(unsigned long ticks
)
5200 account_idle_time(jiffies_to_cputime(ticks
));
5206 * Use precise platform statistics if available:
5208 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5209 cputime_t
task_utime(struct task_struct
*p
)
5214 cputime_t
task_stime(struct task_struct
*p
)
5219 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5221 struct task_cputime cputime
;
5223 thread_group_cputime(p
, &cputime
);
5225 *ut
= cputime
.utime
;
5226 *st
= cputime
.stime
;
5230 #ifndef nsecs_to_cputime
5231 # define nsecs_to_cputime(__nsecs) \
5232 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5235 cputime_t
task_utime(struct task_struct
*p
)
5237 cputime_t utime
= p
->utime
, total
= utime
+ p
->stime
;
5241 * Use CFS's precise accounting:
5243 temp
= (u64
)nsecs_to_cputime(p
->se
.sum_exec_runtime
);
5247 do_div(temp
, total
);
5249 utime
= (cputime_t
)temp
;
5251 p
->prev_utime
= max(p
->prev_utime
, utime
);
5252 return p
->prev_utime
;
5255 cputime_t
task_stime(struct task_struct
*p
)
5260 * Use CFS's precise accounting. (we subtract utime from
5261 * the total, to make sure the total observed by userspace
5262 * grows monotonically - apps rely on that):
5264 stime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
) - task_utime(p
);
5267 p
->prev_stime
= max(p
->prev_stime
, stime
);
5269 return p
->prev_stime
;
5273 * Must be called with siglock held.
5275 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5277 struct signal_struct
*sig
= p
->signal
;
5278 struct task_cputime cputime
;
5279 cputime_t rtime
, utime
, total
;
5281 thread_group_cputime(p
, &cputime
);
5283 total
= cputime_add(cputime
.utime
, cputime
.stime
);
5284 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
5289 temp
= (u64
)(rtime
* cputime
.utime
);
5290 do_div(temp
, total
);
5291 utime
= (cputime_t
)temp
;
5295 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
5296 sig
->prev_stime
= max(sig
->prev_stime
,
5297 cputime_sub(rtime
, sig
->prev_utime
));
5299 *ut
= sig
->prev_utime
;
5300 *st
= sig
->prev_stime
;
5304 inline cputime_t
task_gtime(struct task_struct
*p
)
5310 * This function gets called by the timer code, with HZ frequency.
5311 * We call it with interrupts disabled.
5313 * It also gets called by the fork code, when changing the parent's
5316 void scheduler_tick(void)
5318 int cpu
= smp_processor_id();
5319 struct rq
*rq
= cpu_rq(cpu
);
5320 struct task_struct
*curr
= rq
->curr
;
5324 spin_lock(&rq
->lock
);
5325 update_rq_clock(rq
);
5326 update_cpu_load(rq
);
5327 curr
->sched_class
->task_tick(rq
, curr
, 0);
5328 spin_unlock(&rq
->lock
);
5330 perf_event_task_tick(curr
, cpu
);
5333 rq
->idle_at_tick
= idle_cpu(cpu
);
5334 trigger_load_balance(rq
, cpu
);
5338 notrace
unsigned long get_parent_ip(unsigned long addr
)
5340 if (in_lock_functions(addr
)) {
5341 addr
= CALLER_ADDR2
;
5342 if (in_lock_functions(addr
))
5343 addr
= CALLER_ADDR3
;
5348 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5349 defined(CONFIG_PREEMPT_TRACER))
5351 void __kprobes
add_preempt_count(int val
)
5353 #ifdef CONFIG_DEBUG_PREEMPT
5357 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5360 preempt_count() += val
;
5361 #ifdef CONFIG_DEBUG_PREEMPT
5363 * Spinlock count overflowing soon?
5365 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5368 if (preempt_count() == val
)
5369 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5371 EXPORT_SYMBOL(add_preempt_count
);
5373 void __kprobes
sub_preempt_count(int val
)
5375 #ifdef CONFIG_DEBUG_PREEMPT
5379 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5382 * Is the spinlock portion underflowing?
5384 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5385 !(preempt_count() & PREEMPT_MASK
)))
5389 if (preempt_count() == val
)
5390 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5391 preempt_count() -= val
;
5393 EXPORT_SYMBOL(sub_preempt_count
);
5398 * Print scheduling while atomic bug:
5400 static noinline
void __schedule_bug(struct task_struct
*prev
)
5402 struct pt_regs
*regs
= get_irq_regs();
5404 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5405 prev
->comm
, prev
->pid
, preempt_count());
5407 debug_show_held_locks(prev
);
5409 if (irqs_disabled())
5410 print_irqtrace_events(prev
);
5419 * Various schedule()-time debugging checks and statistics:
5421 static inline void schedule_debug(struct task_struct
*prev
)
5424 * Test if we are atomic. Since do_exit() needs to call into
5425 * schedule() atomically, we ignore that path for now.
5426 * Otherwise, whine if we are scheduling when we should not be.
5428 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5429 __schedule_bug(prev
);
5431 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5433 schedstat_inc(this_rq(), sched_count
);
5434 #ifdef CONFIG_SCHEDSTATS
5435 if (unlikely(prev
->lock_depth
>= 0)) {
5436 schedstat_inc(this_rq(), bkl_count
);
5437 schedstat_inc(prev
, sched_info
.bkl_count
);
5442 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5444 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5446 update_avg(&p
->se
.avg_running
, runtime
);
5448 if (p
->state
== TASK_RUNNING
) {
5450 * In order to avoid avg_overlap growing stale when we are
5451 * indeed overlapping and hence not getting put to sleep, grow
5452 * the avg_overlap on preemption.
5454 * We use the average preemption runtime because that
5455 * correlates to the amount of cache footprint a task can
5458 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5459 update_avg(&p
->se
.avg_overlap
, runtime
);
5461 update_avg(&p
->se
.avg_running
, 0);
5463 p
->sched_class
->put_prev_task(rq
, p
);
5467 * Pick up the highest-prio task:
5469 static inline struct task_struct
*
5470 pick_next_task(struct rq
*rq
)
5472 const struct sched_class
*class;
5473 struct task_struct
*p
;
5476 * Optimization: we know that if all tasks are in
5477 * the fair class we can call that function directly:
5479 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5480 p
= fair_sched_class
.pick_next_task(rq
);
5485 class = sched_class_highest
;
5487 p
= class->pick_next_task(rq
);
5491 * Will never be NULL as the idle class always
5492 * returns a non-NULL p:
5494 class = class->next
;
5499 * schedule() is the main scheduler function.
5501 asmlinkage
void __sched
schedule(void)
5503 struct task_struct
*prev
, *next
;
5504 unsigned long *switch_count
;
5510 cpu
= smp_processor_id();
5514 switch_count
= &prev
->nivcsw
;
5516 release_kernel_lock(prev
);
5517 need_resched_nonpreemptible
:
5519 schedule_debug(prev
);
5521 if (sched_feat(HRTICK
))
5524 spin_lock_irq(&rq
->lock
);
5525 update_rq_clock(rq
);
5526 clear_tsk_need_resched(prev
);
5528 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5529 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5530 prev
->state
= TASK_RUNNING
;
5532 deactivate_task(rq
, prev
, 1);
5533 switch_count
= &prev
->nvcsw
;
5536 pre_schedule(rq
, prev
);
5538 if (unlikely(!rq
->nr_running
))
5539 idle_balance(cpu
, rq
);
5541 put_prev_task(rq
, prev
);
5542 next
= pick_next_task(rq
);
5544 if (likely(prev
!= next
)) {
5545 sched_info_switch(prev
, next
);
5546 perf_event_task_sched_out(prev
, next
, cpu
);
5552 context_switch(rq
, prev
, next
); /* unlocks the rq */
5554 * the context switch might have flipped the stack from under
5555 * us, hence refresh the local variables.
5557 cpu
= smp_processor_id();
5560 spin_unlock_irq(&rq
->lock
);
5564 if (unlikely(reacquire_kernel_lock(current
) < 0))
5565 goto need_resched_nonpreemptible
;
5567 preempt_enable_no_resched();
5571 EXPORT_SYMBOL(schedule
);
5575 * Look out! "owner" is an entirely speculative pointer
5576 * access and not reliable.
5578 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5583 if (!sched_feat(OWNER_SPIN
))
5586 #ifdef CONFIG_DEBUG_PAGEALLOC
5588 * Need to access the cpu field knowing that
5589 * DEBUG_PAGEALLOC could have unmapped it if
5590 * the mutex owner just released it and exited.
5592 if (probe_kernel_address(&owner
->cpu
, cpu
))
5599 * Even if the access succeeded (likely case),
5600 * the cpu field may no longer be valid.
5602 if (cpu
>= nr_cpumask_bits
)
5606 * We need to validate that we can do a
5607 * get_cpu() and that we have the percpu area.
5609 if (!cpu_online(cpu
))
5616 * Owner changed, break to re-assess state.
5618 if (lock
->owner
!= owner
)
5622 * Is that owner really running on that cpu?
5624 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5634 #ifdef CONFIG_PREEMPT
5636 * this is the entry point to schedule() from in-kernel preemption
5637 * off of preempt_enable. Kernel preemptions off return from interrupt
5638 * occur there and call schedule directly.
5640 asmlinkage
void __sched
preempt_schedule(void)
5642 struct thread_info
*ti
= current_thread_info();
5645 * If there is a non-zero preempt_count or interrupts are disabled,
5646 * we do not want to preempt the current task. Just return..
5648 if (likely(ti
->preempt_count
|| irqs_disabled()))
5652 add_preempt_count(PREEMPT_ACTIVE
);
5654 sub_preempt_count(PREEMPT_ACTIVE
);
5657 * Check again in case we missed a preemption opportunity
5658 * between schedule and now.
5661 } while (need_resched());
5663 EXPORT_SYMBOL(preempt_schedule
);
5666 * this is the entry point to schedule() from kernel preemption
5667 * off of irq context.
5668 * Note, that this is called and return with irqs disabled. This will
5669 * protect us against recursive calling from irq.
5671 asmlinkage
void __sched
preempt_schedule_irq(void)
5673 struct thread_info
*ti
= current_thread_info();
5675 /* Catch callers which need to be fixed */
5676 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5679 add_preempt_count(PREEMPT_ACTIVE
);
5682 local_irq_disable();
5683 sub_preempt_count(PREEMPT_ACTIVE
);
5686 * Check again in case we missed a preemption opportunity
5687 * between schedule and now.
5690 } while (need_resched());
5693 #endif /* CONFIG_PREEMPT */
5695 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5698 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5700 EXPORT_SYMBOL(default_wake_function
);
5703 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5704 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5705 * number) then we wake all the non-exclusive tasks and one exclusive task.
5707 * There are circumstances in which we can try to wake a task which has already
5708 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5709 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5711 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5712 int nr_exclusive
, int wake_flags
, void *key
)
5714 wait_queue_t
*curr
, *next
;
5716 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5717 unsigned flags
= curr
->flags
;
5719 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5720 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5726 * __wake_up - wake up threads blocked on a waitqueue.
5728 * @mode: which threads
5729 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5730 * @key: is directly passed to the wakeup function
5732 * It may be assumed that this function implies a write memory barrier before
5733 * changing the task state if and only if any tasks are woken up.
5735 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5736 int nr_exclusive
, void *key
)
5738 unsigned long flags
;
5740 spin_lock_irqsave(&q
->lock
, flags
);
5741 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5742 spin_unlock_irqrestore(&q
->lock
, flags
);
5744 EXPORT_SYMBOL(__wake_up
);
5747 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5749 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5751 __wake_up_common(q
, mode
, 1, 0, NULL
);
5754 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5756 __wake_up_common(q
, mode
, 1, 0, key
);
5760 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5762 * @mode: which threads
5763 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5764 * @key: opaque value to be passed to wakeup targets
5766 * The sync wakeup differs that the waker knows that it will schedule
5767 * away soon, so while the target thread will be woken up, it will not
5768 * be migrated to another CPU - ie. the two threads are 'synchronized'
5769 * with each other. This can prevent needless bouncing between CPUs.
5771 * On UP it can prevent extra preemption.
5773 * It may be assumed that this function implies a write memory barrier before
5774 * changing the task state if and only if any tasks are woken up.
5776 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5777 int nr_exclusive
, void *key
)
5779 unsigned long flags
;
5780 int wake_flags
= WF_SYNC
;
5785 if (unlikely(!nr_exclusive
))
5788 spin_lock_irqsave(&q
->lock
, flags
);
5789 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5790 spin_unlock_irqrestore(&q
->lock
, flags
);
5792 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5795 * __wake_up_sync - see __wake_up_sync_key()
5797 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5799 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5801 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5804 * complete: - signals a single thread waiting on this completion
5805 * @x: holds the state of this particular completion
5807 * This will wake up a single thread waiting on this completion. Threads will be
5808 * awakened in the same order in which they were queued.
5810 * See also complete_all(), wait_for_completion() and related routines.
5812 * It may be assumed that this function implies a write memory barrier before
5813 * changing the task state if and only if any tasks are woken up.
5815 void complete(struct completion
*x
)
5817 unsigned long flags
;
5819 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5821 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5822 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5824 EXPORT_SYMBOL(complete
);
5827 * complete_all: - signals all threads waiting on this completion
5828 * @x: holds the state of this particular completion
5830 * This will wake up all threads waiting on this particular completion event.
5832 * It may be assumed that this function implies a write memory barrier before
5833 * changing the task state if and only if any tasks are woken up.
5835 void complete_all(struct completion
*x
)
5837 unsigned long flags
;
5839 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5840 x
->done
+= UINT_MAX
/2;
5841 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5842 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5844 EXPORT_SYMBOL(complete_all
);
5846 static inline long __sched
5847 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5850 DECLARE_WAITQUEUE(wait
, current
);
5852 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5853 __add_wait_queue_tail(&x
->wait
, &wait
);
5855 if (signal_pending_state(state
, current
)) {
5856 timeout
= -ERESTARTSYS
;
5859 __set_current_state(state
);
5860 spin_unlock_irq(&x
->wait
.lock
);
5861 timeout
= schedule_timeout(timeout
);
5862 spin_lock_irq(&x
->wait
.lock
);
5863 } while (!x
->done
&& timeout
);
5864 __remove_wait_queue(&x
->wait
, &wait
);
5869 return timeout
?: 1;
5873 wait_for_common(struct completion
*x
, long timeout
, int state
)
5877 spin_lock_irq(&x
->wait
.lock
);
5878 timeout
= do_wait_for_common(x
, timeout
, state
);
5879 spin_unlock_irq(&x
->wait
.lock
);
5884 * wait_for_completion: - waits for completion of a task
5885 * @x: holds the state of this particular completion
5887 * This waits to be signaled for completion of a specific task. It is NOT
5888 * interruptible and there is no timeout.
5890 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5891 * and interrupt capability. Also see complete().
5893 void __sched
wait_for_completion(struct completion
*x
)
5895 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5897 EXPORT_SYMBOL(wait_for_completion
);
5900 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5901 * @x: holds the state of this particular completion
5902 * @timeout: timeout value in jiffies
5904 * This waits for either a completion of a specific task to be signaled or for a
5905 * specified timeout to expire. The timeout is in jiffies. It is not
5908 unsigned long __sched
5909 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5911 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5913 EXPORT_SYMBOL(wait_for_completion_timeout
);
5916 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5917 * @x: holds the state of this particular completion
5919 * This waits for completion of a specific task to be signaled. It is
5922 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5924 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5925 if (t
== -ERESTARTSYS
)
5929 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5932 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5933 * @x: holds the state of this particular completion
5934 * @timeout: timeout value in jiffies
5936 * This waits for either a completion of a specific task to be signaled or for a
5937 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5939 unsigned long __sched
5940 wait_for_completion_interruptible_timeout(struct completion
*x
,
5941 unsigned long timeout
)
5943 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5945 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5948 * wait_for_completion_killable: - waits for completion of a task (killable)
5949 * @x: holds the state of this particular completion
5951 * This waits to be signaled for completion of a specific task. It can be
5952 * interrupted by a kill signal.
5954 int __sched
wait_for_completion_killable(struct completion
*x
)
5956 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5957 if (t
== -ERESTARTSYS
)
5961 EXPORT_SYMBOL(wait_for_completion_killable
);
5964 * try_wait_for_completion - try to decrement a completion without blocking
5965 * @x: completion structure
5967 * Returns: 0 if a decrement cannot be done without blocking
5968 * 1 if a decrement succeeded.
5970 * If a completion is being used as a counting completion,
5971 * attempt to decrement the counter without blocking. This
5972 * enables us to avoid waiting if the resource the completion
5973 * is protecting is not available.
5975 bool try_wait_for_completion(struct completion
*x
)
5979 spin_lock_irq(&x
->wait
.lock
);
5984 spin_unlock_irq(&x
->wait
.lock
);
5987 EXPORT_SYMBOL(try_wait_for_completion
);
5990 * completion_done - Test to see if a completion has any waiters
5991 * @x: completion structure
5993 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5994 * 1 if there are no waiters.
5997 bool completion_done(struct completion
*x
)
6001 spin_lock_irq(&x
->wait
.lock
);
6004 spin_unlock_irq(&x
->wait
.lock
);
6007 EXPORT_SYMBOL(completion_done
);
6010 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
6012 unsigned long flags
;
6015 init_waitqueue_entry(&wait
, current
);
6017 __set_current_state(state
);
6019 spin_lock_irqsave(&q
->lock
, flags
);
6020 __add_wait_queue(q
, &wait
);
6021 spin_unlock(&q
->lock
);
6022 timeout
= schedule_timeout(timeout
);
6023 spin_lock_irq(&q
->lock
);
6024 __remove_wait_queue(q
, &wait
);
6025 spin_unlock_irqrestore(&q
->lock
, flags
);
6030 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
6032 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6034 EXPORT_SYMBOL(interruptible_sleep_on
);
6037 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6039 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
6041 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6043 void __sched
sleep_on(wait_queue_head_t
*q
)
6045 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6047 EXPORT_SYMBOL(sleep_on
);
6049 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6051 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6053 EXPORT_SYMBOL(sleep_on_timeout
);
6055 #ifdef CONFIG_RT_MUTEXES
6058 * rt_mutex_setprio - set the current priority of a task
6060 * @prio: prio value (kernel-internal form)
6062 * This function changes the 'effective' priority of a task. It does
6063 * not touch ->normal_prio like __setscheduler().
6065 * Used by the rt_mutex code to implement priority inheritance logic.
6067 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6069 unsigned long flags
;
6070 int oldprio
, on_rq
, running
;
6072 const struct sched_class
*prev_class
;
6074 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6076 rq
= task_rq_lock(p
, &flags
);
6077 update_rq_clock(rq
);
6080 prev_class
= p
->sched_class
;
6081 on_rq
= p
->se
.on_rq
;
6082 running
= task_current(rq
, p
);
6084 dequeue_task(rq
, p
, 0);
6086 p
->sched_class
->put_prev_task(rq
, p
);
6089 p
->sched_class
= &rt_sched_class
;
6091 p
->sched_class
= &fair_sched_class
;
6096 p
->sched_class
->set_curr_task(rq
);
6098 enqueue_task(rq
, p
, 0);
6100 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6102 task_rq_unlock(rq
, &flags
);
6107 void set_user_nice(struct task_struct
*p
, long nice
)
6109 int old_prio
, delta
, on_rq
;
6110 unsigned long flags
;
6113 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6116 * We have to be careful, if called from sys_setpriority(),
6117 * the task might be in the middle of scheduling on another CPU.
6119 rq
= task_rq_lock(p
, &flags
);
6120 update_rq_clock(rq
);
6122 * The RT priorities are set via sched_setscheduler(), but we still
6123 * allow the 'normal' nice value to be set - but as expected
6124 * it wont have any effect on scheduling until the task is
6125 * SCHED_FIFO/SCHED_RR:
6127 if (task_has_rt_policy(p
)) {
6128 p
->static_prio
= NICE_TO_PRIO(nice
);
6131 on_rq
= p
->se
.on_rq
;
6133 dequeue_task(rq
, p
, 0);
6135 p
->static_prio
= NICE_TO_PRIO(nice
);
6138 p
->prio
= effective_prio(p
);
6139 delta
= p
->prio
- old_prio
;
6142 enqueue_task(rq
, p
, 0);
6144 * If the task increased its priority or is running and
6145 * lowered its priority, then reschedule its CPU:
6147 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6148 resched_task(rq
->curr
);
6151 task_rq_unlock(rq
, &flags
);
6153 EXPORT_SYMBOL(set_user_nice
);
6156 * can_nice - check if a task can reduce its nice value
6160 int can_nice(const struct task_struct
*p
, const int nice
)
6162 /* convert nice value [19,-20] to rlimit style value [1,40] */
6163 int nice_rlim
= 20 - nice
;
6165 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6166 capable(CAP_SYS_NICE
));
6169 #ifdef __ARCH_WANT_SYS_NICE
6172 * sys_nice - change the priority of the current process.
6173 * @increment: priority increment
6175 * sys_setpriority is a more generic, but much slower function that
6176 * does similar things.
6178 SYSCALL_DEFINE1(nice
, int, increment
)
6183 * Setpriority might change our priority at the same moment.
6184 * We don't have to worry. Conceptually one call occurs first
6185 * and we have a single winner.
6187 if (increment
< -40)
6192 nice
= TASK_NICE(current
) + increment
;
6198 if (increment
< 0 && !can_nice(current
, nice
))
6201 retval
= security_task_setnice(current
, nice
);
6205 set_user_nice(current
, nice
);
6212 * task_prio - return the priority value of a given task.
6213 * @p: the task in question.
6215 * This is the priority value as seen by users in /proc.
6216 * RT tasks are offset by -200. Normal tasks are centered
6217 * around 0, value goes from -16 to +15.
6219 int task_prio(const struct task_struct
*p
)
6221 return p
->prio
- MAX_RT_PRIO
;
6225 * task_nice - return the nice value of a given task.
6226 * @p: the task in question.
6228 int task_nice(const struct task_struct
*p
)
6230 return TASK_NICE(p
);
6232 EXPORT_SYMBOL(task_nice
);
6235 * idle_cpu - is a given cpu idle currently?
6236 * @cpu: the processor in question.
6238 int idle_cpu(int cpu
)
6240 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6244 * idle_task - return the idle task for a given cpu.
6245 * @cpu: the processor in question.
6247 struct task_struct
*idle_task(int cpu
)
6249 return cpu_rq(cpu
)->idle
;
6253 * find_process_by_pid - find a process with a matching PID value.
6254 * @pid: the pid in question.
6256 static struct task_struct
*find_process_by_pid(pid_t pid
)
6258 return pid
? find_task_by_vpid(pid
) : current
;
6261 /* Actually do priority change: must hold rq lock. */
6263 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6265 BUG_ON(p
->se
.on_rq
);
6268 switch (p
->policy
) {
6272 p
->sched_class
= &fair_sched_class
;
6276 p
->sched_class
= &rt_sched_class
;
6280 p
->rt_priority
= prio
;
6281 p
->normal_prio
= normal_prio(p
);
6282 /* we are holding p->pi_lock already */
6283 p
->prio
= rt_mutex_getprio(p
);
6288 * check the target process has a UID that matches the current process's
6290 static bool check_same_owner(struct task_struct
*p
)
6292 const struct cred
*cred
= current_cred(), *pcred
;
6296 pcred
= __task_cred(p
);
6297 match
= (cred
->euid
== pcred
->euid
||
6298 cred
->euid
== pcred
->uid
);
6303 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6304 struct sched_param
*param
, bool user
)
6306 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6307 unsigned long flags
;
6308 const struct sched_class
*prev_class
;
6312 /* may grab non-irq protected spin_locks */
6313 BUG_ON(in_interrupt());
6315 /* double check policy once rq lock held */
6317 reset_on_fork
= p
->sched_reset_on_fork
;
6318 policy
= oldpolicy
= p
->policy
;
6320 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6321 policy
&= ~SCHED_RESET_ON_FORK
;
6323 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6324 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6325 policy
!= SCHED_IDLE
)
6330 * Valid priorities for SCHED_FIFO and SCHED_RR are
6331 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6332 * SCHED_BATCH and SCHED_IDLE is 0.
6334 if (param
->sched_priority
< 0 ||
6335 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6336 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6338 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6342 * Allow unprivileged RT tasks to decrease priority:
6344 if (user
&& !capable(CAP_SYS_NICE
)) {
6345 if (rt_policy(policy
)) {
6346 unsigned long rlim_rtprio
;
6348 if (!lock_task_sighand(p
, &flags
))
6350 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6351 unlock_task_sighand(p
, &flags
);
6353 /* can't set/change the rt policy */
6354 if (policy
!= p
->policy
&& !rlim_rtprio
)
6357 /* can't increase priority */
6358 if (param
->sched_priority
> p
->rt_priority
&&
6359 param
->sched_priority
> rlim_rtprio
)
6363 * Like positive nice levels, dont allow tasks to
6364 * move out of SCHED_IDLE either:
6366 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6369 /* can't change other user's priorities */
6370 if (!check_same_owner(p
))
6373 /* Normal users shall not reset the sched_reset_on_fork flag */
6374 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6379 #ifdef CONFIG_RT_GROUP_SCHED
6381 * Do not allow realtime tasks into groups that have no runtime
6384 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6385 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6389 retval
= security_task_setscheduler(p
, policy
, param
);
6395 * make sure no PI-waiters arrive (or leave) while we are
6396 * changing the priority of the task:
6398 spin_lock_irqsave(&p
->pi_lock
, flags
);
6400 * To be able to change p->policy safely, the apropriate
6401 * runqueue lock must be held.
6403 rq
= __task_rq_lock(p
);
6404 /* recheck policy now with rq lock held */
6405 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6406 policy
= oldpolicy
= -1;
6407 __task_rq_unlock(rq
);
6408 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6411 update_rq_clock(rq
);
6412 on_rq
= p
->se
.on_rq
;
6413 running
= task_current(rq
, p
);
6415 deactivate_task(rq
, p
, 0);
6417 p
->sched_class
->put_prev_task(rq
, p
);
6419 p
->sched_reset_on_fork
= reset_on_fork
;
6422 prev_class
= p
->sched_class
;
6423 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6426 p
->sched_class
->set_curr_task(rq
);
6428 activate_task(rq
, p
, 0);
6430 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6432 __task_rq_unlock(rq
);
6433 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6435 rt_mutex_adjust_pi(p
);
6441 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6442 * @p: the task in question.
6443 * @policy: new policy.
6444 * @param: structure containing the new RT priority.
6446 * NOTE that the task may be already dead.
6448 int sched_setscheduler(struct task_struct
*p
, int policy
,
6449 struct sched_param
*param
)
6451 return __sched_setscheduler(p
, policy
, param
, true);
6453 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6456 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6457 * @p: the task in question.
6458 * @policy: new policy.
6459 * @param: structure containing the new RT priority.
6461 * Just like sched_setscheduler, only don't bother checking if the
6462 * current context has permission. For example, this is needed in
6463 * stop_machine(): we create temporary high priority worker threads,
6464 * but our caller might not have that capability.
6466 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6467 struct sched_param
*param
)
6469 return __sched_setscheduler(p
, policy
, param
, false);
6473 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6475 struct sched_param lparam
;
6476 struct task_struct
*p
;
6479 if (!param
|| pid
< 0)
6481 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6486 p
= find_process_by_pid(pid
);
6488 retval
= sched_setscheduler(p
, policy
, &lparam
);
6495 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6496 * @pid: the pid in question.
6497 * @policy: new policy.
6498 * @param: structure containing the new RT priority.
6500 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6501 struct sched_param __user
*, param
)
6503 /* negative values for policy are not valid */
6507 return do_sched_setscheduler(pid
, policy
, param
);
6511 * sys_sched_setparam - set/change the RT priority of a thread
6512 * @pid: the pid in question.
6513 * @param: structure containing the new RT priority.
6515 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6517 return do_sched_setscheduler(pid
, -1, param
);
6521 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6522 * @pid: the pid in question.
6524 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6526 struct task_struct
*p
;
6533 read_lock(&tasklist_lock
);
6534 p
= find_process_by_pid(pid
);
6536 retval
= security_task_getscheduler(p
);
6539 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6541 read_unlock(&tasklist_lock
);
6546 * sys_sched_getparam - get the RT priority of a thread
6547 * @pid: the pid in question.
6548 * @param: structure containing the RT priority.
6550 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6552 struct sched_param lp
;
6553 struct task_struct
*p
;
6556 if (!param
|| pid
< 0)
6559 read_lock(&tasklist_lock
);
6560 p
= find_process_by_pid(pid
);
6565 retval
= security_task_getscheduler(p
);
6569 lp
.sched_priority
= p
->rt_priority
;
6570 read_unlock(&tasklist_lock
);
6573 * This one might sleep, we cannot do it with a spinlock held ...
6575 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6580 read_unlock(&tasklist_lock
);
6584 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6586 cpumask_var_t cpus_allowed
, new_mask
;
6587 struct task_struct
*p
;
6591 read_lock(&tasklist_lock
);
6593 p
= find_process_by_pid(pid
);
6595 read_unlock(&tasklist_lock
);
6601 * It is not safe to call set_cpus_allowed with the
6602 * tasklist_lock held. We will bump the task_struct's
6603 * usage count and then drop tasklist_lock.
6606 read_unlock(&tasklist_lock
);
6608 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6612 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6614 goto out_free_cpus_allowed
;
6617 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6620 retval
= security_task_setscheduler(p
, 0, NULL
);
6624 cpuset_cpus_allowed(p
, cpus_allowed
);
6625 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6627 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6630 cpuset_cpus_allowed(p
, cpus_allowed
);
6631 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6633 * We must have raced with a concurrent cpuset
6634 * update. Just reset the cpus_allowed to the
6635 * cpuset's cpus_allowed
6637 cpumask_copy(new_mask
, cpus_allowed
);
6642 free_cpumask_var(new_mask
);
6643 out_free_cpus_allowed
:
6644 free_cpumask_var(cpus_allowed
);
6651 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6652 struct cpumask
*new_mask
)
6654 if (len
< cpumask_size())
6655 cpumask_clear(new_mask
);
6656 else if (len
> cpumask_size())
6657 len
= cpumask_size();
6659 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6663 * sys_sched_setaffinity - set the cpu affinity of a process
6664 * @pid: pid of the process
6665 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6666 * @user_mask_ptr: user-space pointer to the new cpu mask
6668 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6669 unsigned long __user
*, user_mask_ptr
)
6671 cpumask_var_t new_mask
;
6674 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6677 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6679 retval
= sched_setaffinity(pid
, new_mask
);
6680 free_cpumask_var(new_mask
);
6684 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6686 struct task_struct
*p
;
6690 read_lock(&tasklist_lock
);
6693 p
= find_process_by_pid(pid
);
6697 retval
= security_task_getscheduler(p
);
6701 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6704 read_unlock(&tasklist_lock
);
6711 * sys_sched_getaffinity - get the cpu affinity of a process
6712 * @pid: pid of the process
6713 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6714 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6716 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6717 unsigned long __user
*, user_mask_ptr
)
6722 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
6724 if (len
& (sizeof(unsigned long)-1))
6727 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6730 ret
= sched_getaffinity(pid
, mask
);
6732 size_t retlen
= min_t(size_t, len
, cpumask_size());
6734 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
6739 free_cpumask_var(mask
);
6745 * sys_sched_yield - yield the current processor to other threads.
6747 * This function yields the current CPU to other tasks. If there are no
6748 * other threads running on this CPU then this function will return.
6750 SYSCALL_DEFINE0(sched_yield
)
6752 struct rq
*rq
= this_rq_lock();
6754 schedstat_inc(rq
, yld_count
);
6755 current
->sched_class
->yield_task(rq
);
6758 * Since we are going to call schedule() anyway, there's
6759 * no need to preempt or enable interrupts:
6761 __release(rq
->lock
);
6762 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6763 _raw_spin_unlock(&rq
->lock
);
6764 preempt_enable_no_resched();
6771 static inline int should_resched(void)
6773 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6776 static void __cond_resched(void)
6778 add_preempt_count(PREEMPT_ACTIVE
);
6780 sub_preempt_count(PREEMPT_ACTIVE
);
6783 int __sched
_cond_resched(void)
6785 if (should_resched()) {
6791 EXPORT_SYMBOL(_cond_resched
);
6794 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6795 * call schedule, and on return reacquire the lock.
6797 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6798 * operations here to prevent schedule() from being called twice (once via
6799 * spin_unlock(), once by hand).
6801 int __cond_resched_lock(spinlock_t
*lock
)
6803 int resched
= should_resched();
6806 lockdep_assert_held(lock
);
6808 if (spin_needbreak(lock
) || resched
) {
6819 EXPORT_SYMBOL(__cond_resched_lock
);
6821 int __sched
__cond_resched_softirq(void)
6823 BUG_ON(!in_softirq());
6825 if (should_resched()) {
6833 EXPORT_SYMBOL(__cond_resched_softirq
);
6836 * yield - yield the current processor to other threads.
6838 * This is a shortcut for kernel-space yielding - it marks the
6839 * thread runnable and calls sys_sched_yield().
6841 void __sched
yield(void)
6843 set_current_state(TASK_RUNNING
);
6846 EXPORT_SYMBOL(yield
);
6849 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6850 * that process accounting knows that this is a task in IO wait state.
6852 void __sched
io_schedule(void)
6854 struct rq
*rq
= raw_rq();
6856 delayacct_blkio_start();
6857 atomic_inc(&rq
->nr_iowait
);
6858 current
->in_iowait
= 1;
6860 current
->in_iowait
= 0;
6861 atomic_dec(&rq
->nr_iowait
);
6862 delayacct_blkio_end();
6864 EXPORT_SYMBOL(io_schedule
);
6866 long __sched
io_schedule_timeout(long timeout
)
6868 struct rq
*rq
= raw_rq();
6871 delayacct_blkio_start();
6872 atomic_inc(&rq
->nr_iowait
);
6873 current
->in_iowait
= 1;
6874 ret
= schedule_timeout(timeout
);
6875 current
->in_iowait
= 0;
6876 atomic_dec(&rq
->nr_iowait
);
6877 delayacct_blkio_end();
6882 * sys_sched_get_priority_max - return maximum RT priority.
6883 * @policy: scheduling class.
6885 * this syscall returns the maximum rt_priority that can be used
6886 * by a given scheduling class.
6888 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6895 ret
= MAX_USER_RT_PRIO
-1;
6907 * sys_sched_get_priority_min - return minimum RT priority.
6908 * @policy: scheduling class.
6910 * this syscall returns the minimum rt_priority that can be used
6911 * by a given scheduling class.
6913 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6931 * sys_sched_rr_get_interval - return the default timeslice of a process.
6932 * @pid: pid of the process.
6933 * @interval: userspace pointer to the timeslice value.
6935 * this syscall writes the default timeslice value of a given process
6936 * into the user-space timespec buffer. A value of '0' means infinity.
6938 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6939 struct timespec __user
*, interval
)
6941 struct task_struct
*p
;
6942 unsigned int time_slice
;
6950 read_lock(&tasklist_lock
);
6951 p
= find_process_by_pid(pid
);
6955 retval
= security_task_getscheduler(p
);
6959 time_slice
= p
->sched_class
->get_rr_interval(p
);
6961 read_unlock(&tasklist_lock
);
6962 jiffies_to_timespec(time_slice
, &t
);
6963 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6967 read_unlock(&tasklist_lock
);
6971 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6973 void sched_show_task(struct task_struct
*p
)
6975 unsigned long free
= 0;
6978 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6979 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6980 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6981 #if BITS_PER_LONG == 32
6982 if (state
== TASK_RUNNING
)
6983 printk(KERN_CONT
" running ");
6985 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6987 if (state
== TASK_RUNNING
)
6988 printk(KERN_CONT
" running task ");
6990 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6992 #ifdef CONFIG_DEBUG_STACK_USAGE
6993 free
= stack_not_used(p
);
6995 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6996 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6997 (unsigned long)task_thread_info(p
)->flags
);
6999 show_stack(p
, NULL
);
7002 void show_state_filter(unsigned long state_filter
)
7004 struct task_struct
*g
, *p
;
7006 #if BITS_PER_LONG == 32
7008 " task PC stack pid father\n");
7011 " task PC stack pid father\n");
7013 read_lock(&tasklist_lock
);
7014 do_each_thread(g
, p
) {
7016 * reset the NMI-timeout, listing all files on a slow
7017 * console might take alot of time:
7019 touch_nmi_watchdog();
7020 if (!state_filter
|| (p
->state
& state_filter
))
7022 } while_each_thread(g
, p
);
7024 touch_all_softlockup_watchdogs();
7026 #ifdef CONFIG_SCHED_DEBUG
7027 sysrq_sched_debug_show();
7029 read_unlock(&tasklist_lock
);
7031 * Only show locks if all tasks are dumped:
7033 if (state_filter
== -1)
7034 debug_show_all_locks();
7037 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
7039 idle
->sched_class
= &idle_sched_class
;
7043 * init_idle - set up an idle thread for a given CPU
7044 * @idle: task in question
7045 * @cpu: cpu the idle task belongs to
7047 * NOTE: this function does not set the idle thread's NEED_RESCHED
7048 * flag, to make booting more robust.
7050 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7052 struct rq
*rq
= cpu_rq(cpu
);
7053 unsigned long flags
;
7055 spin_lock_irqsave(&rq
->lock
, flags
);
7058 idle
->se
.exec_start
= sched_clock();
7060 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7061 __set_task_cpu(idle
, cpu
);
7063 rq
->curr
= rq
->idle
= idle
;
7064 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7067 spin_unlock_irqrestore(&rq
->lock
, flags
);
7069 /* Set the preempt count _outside_ the spinlocks! */
7070 #if defined(CONFIG_PREEMPT)
7071 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7073 task_thread_info(idle
)->preempt_count
= 0;
7076 * The idle tasks have their own, simple scheduling class:
7078 idle
->sched_class
= &idle_sched_class
;
7079 ftrace_graph_init_task(idle
);
7083 * In a system that switches off the HZ timer nohz_cpu_mask
7084 * indicates which cpus entered this state. This is used
7085 * in the rcu update to wait only for active cpus. For system
7086 * which do not switch off the HZ timer nohz_cpu_mask should
7087 * always be CPU_BITS_NONE.
7089 cpumask_var_t nohz_cpu_mask
;
7092 * Increase the granularity value when there are more CPUs,
7093 * because with more CPUs the 'effective latency' as visible
7094 * to users decreases. But the relationship is not linear,
7095 * so pick a second-best guess by going with the log2 of the
7098 * This idea comes from the SD scheduler of Con Kolivas:
7100 static void update_sysctl(void)
7102 unsigned int cpus
= min(num_online_cpus(), 8U);
7103 unsigned int factor
= 1 + ilog2(cpus
);
7105 #define SET_SYSCTL(name) \
7106 (sysctl_##name = (factor) * normalized_sysctl_##name)
7107 SET_SYSCTL(sched_min_granularity
);
7108 SET_SYSCTL(sched_latency
);
7109 SET_SYSCTL(sched_wakeup_granularity
);
7110 SET_SYSCTL(sched_shares_ratelimit
);
7114 static inline void sched_init_granularity(void)
7121 * This is how migration works:
7123 * 1) we queue a struct migration_req structure in the source CPU's
7124 * runqueue and wake up that CPU's migration thread.
7125 * 2) we down() the locked semaphore => thread blocks.
7126 * 3) migration thread wakes up (implicitly it forces the migrated
7127 * thread off the CPU)
7128 * 4) it gets the migration request and checks whether the migrated
7129 * task is still in the wrong runqueue.
7130 * 5) if it's in the wrong runqueue then the migration thread removes
7131 * it and puts it into the right queue.
7132 * 6) migration thread up()s the semaphore.
7133 * 7) we wake up and the migration is done.
7137 * Change a given task's CPU affinity. Migrate the thread to a
7138 * proper CPU and schedule it away if the CPU it's executing on
7139 * is removed from the allowed bitmask.
7141 * NOTE: the caller must have a valid reference to the task, the
7142 * task must not exit() & deallocate itself prematurely. The
7143 * call is not atomic; no spinlocks may be held.
7145 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7147 struct migration_req req
;
7148 unsigned long flags
;
7152 rq
= task_rq_lock(p
, &flags
);
7153 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
7158 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7159 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7164 if (p
->sched_class
->set_cpus_allowed
)
7165 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7167 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7168 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7171 /* Can the task run on the task's current CPU? If so, we're done */
7172 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7175 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
7176 /* Need help from migration thread: drop lock and wait. */
7177 struct task_struct
*mt
= rq
->migration_thread
;
7179 get_task_struct(mt
);
7180 task_rq_unlock(rq
, &flags
);
7181 wake_up_process(rq
->migration_thread
);
7182 put_task_struct(mt
);
7183 wait_for_completion(&req
.done
);
7184 tlb_migrate_finish(p
->mm
);
7188 task_rq_unlock(rq
, &flags
);
7192 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7195 * Move (not current) task off this cpu, onto dest cpu. We're doing
7196 * this because either it can't run here any more (set_cpus_allowed()
7197 * away from this CPU, or CPU going down), or because we're
7198 * attempting to rebalance this task on exec (sched_exec).
7200 * So we race with normal scheduler movements, but that's OK, as long
7201 * as the task is no longer on this CPU.
7203 * Returns non-zero if task was successfully migrated.
7205 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7207 struct rq
*rq_dest
, *rq_src
;
7210 if (unlikely(!cpu_active(dest_cpu
)))
7213 rq_src
= cpu_rq(src_cpu
);
7214 rq_dest
= cpu_rq(dest_cpu
);
7216 double_rq_lock(rq_src
, rq_dest
);
7217 /* Already moved. */
7218 if (task_cpu(p
) != src_cpu
)
7220 /* Waking up, don't get in the way of try_to_wake_up(). */
7221 if (p
->state
== TASK_WAKING
)
7223 /* Affinity changed (again). */
7224 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7227 on_rq
= p
->se
.on_rq
;
7229 deactivate_task(rq_src
, p
, 0);
7231 set_task_cpu(p
, dest_cpu
);
7233 activate_task(rq_dest
, p
, 0);
7234 check_preempt_curr(rq_dest
, p
, 0);
7239 double_rq_unlock(rq_src
, rq_dest
);
7243 #define RCU_MIGRATION_IDLE 0
7244 #define RCU_MIGRATION_NEED_QS 1
7245 #define RCU_MIGRATION_GOT_QS 2
7246 #define RCU_MIGRATION_MUST_SYNC 3
7249 * migration_thread - this is a highprio system thread that performs
7250 * thread migration by bumping thread off CPU then 'pushing' onto
7253 static int migration_thread(void *data
)
7256 int cpu
= (long)data
;
7260 BUG_ON(rq
->migration_thread
!= current
);
7262 set_current_state(TASK_INTERRUPTIBLE
);
7263 while (!kthread_should_stop()) {
7264 struct migration_req
*req
;
7265 struct list_head
*head
;
7267 spin_lock_irq(&rq
->lock
);
7269 if (cpu_is_offline(cpu
)) {
7270 spin_unlock_irq(&rq
->lock
);
7274 if (rq
->active_balance
) {
7275 active_load_balance(rq
, cpu
);
7276 rq
->active_balance
= 0;
7279 head
= &rq
->migration_queue
;
7281 if (list_empty(head
)) {
7282 spin_unlock_irq(&rq
->lock
);
7284 set_current_state(TASK_INTERRUPTIBLE
);
7287 req
= list_entry(head
->next
, struct migration_req
, list
);
7288 list_del_init(head
->next
);
7290 if (req
->task
!= NULL
) {
7291 spin_unlock(&rq
->lock
);
7292 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7293 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7294 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7295 spin_unlock(&rq
->lock
);
7297 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7298 spin_unlock(&rq
->lock
);
7299 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7303 complete(&req
->done
);
7305 __set_current_state(TASK_RUNNING
);
7310 #ifdef CONFIG_HOTPLUG_CPU
7312 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7316 local_irq_disable();
7317 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7323 * Figure out where task on dead CPU should go, use force if necessary.
7325 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7328 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7331 /* Look for allowed, online CPU in same node. */
7332 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
7333 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7336 /* Any allowed, online CPU? */
7337 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
7338 if (dest_cpu
< nr_cpu_ids
)
7341 /* No more Mr. Nice Guy. */
7342 if (dest_cpu
>= nr_cpu_ids
) {
7343 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7344 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
7347 * Don't tell them about moving exiting tasks or
7348 * kernel threads (both mm NULL), since they never
7351 if (p
->mm
&& printk_ratelimit()) {
7352 printk(KERN_INFO
"process %d (%s) no "
7353 "longer affine to cpu%d\n",
7354 task_pid_nr(p
), p
->comm
, dead_cpu
);
7359 /* It can have affinity changed while we were choosing. */
7360 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7365 * While a dead CPU has no uninterruptible tasks queued at this point,
7366 * it might still have a nonzero ->nr_uninterruptible counter, because
7367 * for performance reasons the counter is not stricly tracking tasks to
7368 * their home CPUs. So we just add the counter to another CPU's counter,
7369 * to keep the global sum constant after CPU-down:
7371 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7373 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
7374 unsigned long flags
;
7376 local_irq_save(flags
);
7377 double_rq_lock(rq_src
, rq_dest
);
7378 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7379 rq_src
->nr_uninterruptible
= 0;
7380 double_rq_unlock(rq_src
, rq_dest
);
7381 local_irq_restore(flags
);
7384 /* Run through task list and migrate tasks from the dead cpu. */
7385 static void migrate_live_tasks(int src_cpu
)
7387 struct task_struct
*p
, *t
;
7389 read_lock(&tasklist_lock
);
7391 do_each_thread(t
, p
) {
7395 if (task_cpu(p
) == src_cpu
)
7396 move_task_off_dead_cpu(src_cpu
, p
);
7397 } while_each_thread(t
, p
);
7399 read_unlock(&tasklist_lock
);
7403 * Schedules idle task to be the next runnable task on current CPU.
7404 * It does so by boosting its priority to highest possible.
7405 * Used by CPU offline code.
7407 void sched_idle_next(void)
7409 int this_cpu
= smp_processor_id();
7410 struct rq
*rq
= cpu_rq(this_cpu
);
7411 struct task_struct
*p
= rq
->idle
;
7412 unsigned long flags
;
7414 /* cpu has to be offline */
7415 BUG_ON(cpu_online(this_cpu
));
7418 * Strictly not necessary since rest of the CPUs are stopped by now
7419 * and interrupts disabled on the current cpu.
7421 spin_lock_irqsave(&rq
->lock
, flags
);
7423 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7425 update_rq_clock(rq
);
7426 activate_task(rq
, p
, 0);
7428 spin_unlock_irqrestore(&rq
->lock
, flags
);
7432 * Ensures that the idle task is using init_mm right before its cpu goes
7435 void idle_task_exit(void)
7437 struct mm_struct
*mm
= current
->active_mm
;
7439 BUG_ON(cpu_online(smp_processor_id()));
7442 switch_mm(mm
, &init_mm
, current
);
7446 /* called under rq->lock with disabled interrupts */
7447 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7449 struct rq
*rq
= cpu_rq(dead_cpu
);
7451 /* Must be exiting, otherwise would be on tasklist. */
7452 BUG_ON(!p
->exit_state
);
7454 /* Cannot have done final schedule yet: would have vanished. */
7455 BUG_ON(p
->state
== TASK_DEAD
);
7460 * Drop lock around migration; if someone else moves it,
7461 * that's OK. No task can be added to this CPU, so iteration is
7464 spin_unlock_irq(&rq
->lock
);
7465 move_task_off_dead_cpu(dead_cpu
, p
);
7466 spin_lock_irq(&rq
->lock
);
7471 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7472 static void migrate_dead_tasks(unsigned int dead_cpu
)
7474 struct rq
*rq
= cpu_rq(dead_cpu
);
7475 struct task_struct
*next
;
7478 if (!rq
->nr_running
)
7480 update_rq_clock(rq
);
7481 next
= pick_next_task(rq
);
7484 next
->sched_class
->put_prev_task(rq
, next
);
7485 migrate_dead(dead_cpu
, next
);
7491 * remove the tasks which were accounted by rq from calc_load_tasks.
7493 static void calc_global_load_remove(struct rq
*rq
)
7495 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7496 rq
->calc_load_active
= 0;
7498 #endif /* CONFIG_HOTPLUG_CPU */
7500 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7502 static struct ctl_table sd_ctl_dir
[] = {
7504 .procname
= "sched_domain",
7510 static struct ctl_table sd_ctl_root
[] = {
7512 .ctl_name
= CTL_KERN
,
7513 .procname
= "kernel",
7515 .child
= sd_ctl_dir
,
7520 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7522 struct ctl_table
*entry
=
7523 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7528 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7530 struct ctl_table
*entry
;
7533 * In the intermediate directories, both the child directory and
7534 * procname are dynamically allocated and could fail but the mode
7535 * will always be set. In the lowest directory the names are
7536 * static strings and all have proc handlers.
7538 for (entry
= *tablep
; entry
->mode
; entry
++) {
7540 sd_free_ctl_entry(&entry
->child
);
7541 if (entry
->proc_handler
== NULL
)
7542 kfree(entry
->procname
);
7550 set_table_entry(struct ctl_table
*entry
,
7551 const char *procname
, void *data
, int maxlen
,
7552 mode_t mode
, proc_handler
*proc_handler
)
7554 entry
->procname
= procname
;
7556 entry
->maxlen
= maxlen
;
7558 entry
->proc_handler
= proc_handler
;
7561 static struct ctl_table
*
7562 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7564 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7569 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7570 sizeof(long), 0644, proc_doulongvec_minmax
);
7571 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7572 sizeof(long), 0644, proc_doulongvec_minmax
);
7573 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7574 sizeof(int), 0644, proc_dointvec_minmax
);
7575 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7576 sizeof(int), 0644, proc_dointvec_minmax
);
7577 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7578 sizeof(int), 0644, proc_dointvec_minmax
);
7579 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7580 sizeof(int), 0644, proc_dointvec_minmax
);
7581 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7582 sizeof(int), 0644, proc_dointvec_minmax
);
7583 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7584 sizeof(int), 0644, proc_dointvec_minmax
);
7585 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7586 sizeof(int), 0644, proc_dointvec_minmax
);
7587 set_table_entry(&table
[9], "cache_nice_tries",
7588 &sd
->cache_nice_tries
,
7589 sizeof(int), 0644, proc_dointvec_minmax
);
7590 set_table_entry(&table
[10], "flags", &sd
->flags
,
7591 sizeof(int), 0644, proc_dointvec_minmax
);
7592 set_table_entry(&table
[11], "name", sd
->name
,
7593 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7594 /* &table[12] is terminator */
7599 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7601 struct ctl_table
*entry
, *table
;
7602 struct sched_domain
*sd
;
7603 int domain_num
= 0, i
;
7606 for_each_domain(cpu
, sd
)
7608 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7613 for_each_domain(cpu
, sd
) {
7614 snprintf(buf
, 32, "domain%d", i
);
7615 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7617 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7624 static struct ctl_table_header
*sd_sysctl_header
;
7625 static void register_sched_domain_sysctl(void)
7627 int i
, cpu_num
= num_possible_cpus();
7628 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7631 WARN_ON(sd_ctl_dir
[0].child
);
7632 sd_ctl_dir
[0].child
= entry
;
7637 for_each_possible_cpu(i
) {
7638 snprintf(buf
, 32, "cpu%d", i
);
7639 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7641 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7645 WARN_ON(sd_sysctl_header
);
7646 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7649 /* may be called multiple times per register */
7650 static void unregister_sched_domain_sysctl(void)
7652 if (sd_sysctl_header
)
7653 unregister_sysctl_table(sd_sysctl_header
);
7654 sd_sysctl_header
= NULL
;
7655 if (sd_ctl_dir
[0].child
)
7656 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7659 static void register_sched_domain_sysctl(void)
7662 static void unregister_sched_domain_sysctl(void)
7667 static void set_rq_online(struct rq
*rq
)
7670 const struct sched_class
*class;
7672 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7675 for_each_class(class) {
7676 if (class->rq_online
)
7677 class->rq_online(rq
);
7682 static void set_rq_offline(struct rq
*rq
)
7685 const struct sched_class
*class;
7687 for_each_class(class) {
7688 if (class->rq_offline
)
7689 class->rq_offline(rq
);
7692 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7698 * migration_call - callback that gets triggered when a CPU is added.
7699 * Here we can start up the necessary migration thread for the new CPU.
7701 static int __cpuinit
7702 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7704 struct task_struct
*p
;
7705 int cpu
= (long)hcpu
;
7706 unsigned long flags
;
7711 case CPU_UP_PREPARE
:
7712 case CPU_UP_PREPARE_FROZEN
:
7713 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7716 kthread_bind(p
, cpu
);
7717 /* Must be high prio: stop_machine expects to yield to it. */
7718 rq
= task_rq_lock(p
, &flags
);
7719 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7720 task_rq_unlock(rq
, &flags
);
7722 cpu_rq(cpu
)->migration_thread
= p
;
7723 rq
->calc_load_update
= calc_load_update
;
7727 case CPU_ONLINE_FROZEN
:
7728 /* Strictly unnecessary, as first user will wake it. */
7729 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7731 /* Update our root-domain */
7733 spin_lock_irqsave(&rq
->lock
, flags
);
7735 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7739 spin_unlock_irqrestore(&rq
->lock
, flags
);
7742 #ifdef CONFIG_HOTPLUG_CPU
7743 case CPU_UP_CANCELED
:
7744 case CPU_UP_CANCELED_FROZEN
:
7745 if (!cpu_rq(cpu
)->migration_thread
)
7747 /* Unbind it from offline cpu so it can run. Fall thru. */
7748 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7749 cpumask_any(cpu_online_mask
));
7750 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7751 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7752 cpu_rq(cpu
)->migration_thread
= NULL
;
7756 case CPU_DEAD_FROZEN
:
7757 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7758 migrate_live_tasks(cpu
);
7760 kthread_stop(rq
->migration_thread
);
7761 put_task_struct(rq
->migration_thread
);
7762 rq
->migration_thread
= NULL
;
7763 /* Idle task back to normal (off runqueue, low prio) */
7764 spin_lock_irq(&rq
->lock
);
7765 update_rq_clock(rq
);
7766 deactivate_task(rq
, rq
->idle
, 0);
7767 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7768 rq
->idle
->sched_class
= &idle_sched_class
;
7769 migrate_dead_tasks(cpu
);
7770 spin_unlock_irq(&rq
->lock
);
7772 migrate_nr_uninterruptible(rq
);
7773 BUG_ON(rq
->nr_running
!= 0);
7774 calc_global_load_remove(rq
);
7776 * No need to migrate the tasks: it was best-effort if
7777 * they didn't take sched_hotcpu_mutex. Just wake up
7780 spin_lock_irq(&rq
->lock
);
7781 while (!list_empty(&rq
->migration_queue
)) {
7782 struct migration_req
*req
;
7784 req
= list_entry(rq
->migration_queue
.next
,
7785 struct migration_req
, list
);
7786 list_del_init(&req
->list
);
7787 spin_unlock_irq(&rq
->lock
);
7788 complete(&req
->done
);
7789 spin_lock_irq(&rq
->lock
);
7791 spin_unlock_irq(&rq
->lock
);
7795 case CPU_DYING_FROZEN
:
7796 /* Update our root-domain */
7798 spin_lock_irqsave(&rq
->lock
, flags
);
7800 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7803 spin_unlock_irqrestore(&rq
->lock
, flags
);
7811 * Register at high priority so that task migration (migrate_all_tasks)
7812 * happens before everything else. This has to be lower priority than
7813 * the notifier in the perf_event subsystem, though.
7815 static struct notifier_block __cpuinitdata migration_notifier
= {
7816 .notifier_call
= migration_call
,
7820 static int __init
migration_init(void)
7822 void *cpu
= (void *)(long)smp_processor_id();
7825 /* Start one for the boot CPU: */
7826 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7827 BUG_ON(err
== NOTIFY_BAD
);
7828 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7829 register_cpu_notifier(&migration_notifier
);
7833 early_initcall(migration_init
);
7838 #ifdef CONFIG_SCHED_DEBUG
7840 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7841 struct cpumask
*groupmask
)
7843 struct sched_group
*group
= sd
->groups
;
7846 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7847 cpumask_clear(groupmask
);
7849 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7851 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7852 printk("does not load-balance\n");
7854 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7859 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7861 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7862 printk(KERN_ERR
"ERROR: domain->span does not contain "
7865 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7866 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7870 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7874 printk(KERN_ERR
"ERROR: group is NULL\n");
7878 if (!group
->cpu_power
) {
7879 printk(KERN_CONT
"\n");
7880 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7885 if (!cpumask_weight(sched_group_cpus(group
))) {
7886 printk(KERN_CONT
"\n");
7887 printk(KERN_ERR
"ERROR: empty group\n");
7891 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7892 printk(KERN_CONT
"\n");
7893 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7897 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7899 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7901 printk(KERN_CONT
" %s", str
);
7902 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7903 printk(KERN_CONT
" (cpu_power = %d)",
7907 group
= group
->next
;
7908 } while (group
!= sd
->groups
);
7909 printk(KERN_CONT
"\n");
7911 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7912 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7915 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7916 printk(KERN_ERR
"ERROR: parent span is not a superset "
7917 "of domain->span\n");
7921 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7923 cpumask_var_t groupmask
;
7927 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7931 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7933 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7934 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7939 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7946 free_cpumask_var(groupmask
);
7948 #else /* !CONFIG_SCHED_DEBUG */
7949 # define sched_domain_debug(sd, cpu) do { } while (0)
7950 #endif /* CONFIG_SCHED_DEBUG */
7952 static int sd_degenerate(struct sched_domain
*sd
)
7954 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7957 /* Following flags need at least 2 groups */
7958 if (sd
->flags
& (SD_LOAD_BALANCE
|
7959 SD_BALANCE_NEWIDLE
|
7963 SD_SHARE_PKG_RESOURCES
)) {
7964 if (sd
->groups
!= sd
->groups
->next
)
7968 /* Following flags don't use groups */
7969 if (sd
->flags
& (SD_WAKE_AFFINE
))
7976 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7978 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7980 if (sd_degenerate(parent
))
7983 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7986 /* Flags needing groups don't count if only 1 group in parent */
7987 if (parent
->groups
== parent
->groups
->next
) {
7988 pflags
&= ~(SD_LOAD_BALANCE
|
7989 SD_BALANCE_NEWIDLE
|
7993 SD_SHARE_PKG_RESOURCES
);
7994 if (nr_node_ids
== 1)
7995 pflags
&= ~SD_SERIALIZE
;
7997 if (~cflags
& pflags
)
8003 static void free_rootdomain(struct root_domain
*rd
)
8005 synchronize_sched();
8007 cpupri_cleanup(&rd
->cpupri
);
8009 free_cpumask_var(rd
->rto_mask
);
8010 free_cpumask_var(rd
->online
);
8011 free_cpumask_var(rd
->span
);
8015 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8017 struct root_domain
*old_rd
= NULL
;
8018 unsigned long flags
;
8020 spin_lock_irqsave(&rq
->lock
, flags
);
8025 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8028 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8031 * If we dont want to free the old_rt yet then
8032 * set old_rd to NULL to skip the freeing later
8035 if (!atomic_dec_and_test(&old_rd
->refcount
))
8039 atomic_inc(&rd
->refcount
);
8042 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8043 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8046 spin_unlock_irqrestore(&rq
->lock
, flags
);
8049 free_rootdomain(old_rd
);
8052 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8054 gfp_t gfp
= GFP_KERNEL
;
8056 memset(rd
, 0, sizeof(*rd
));
8061 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8063 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8065 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8068 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8073 free_cpumask_var(rd
->rto_mask
);
8075 free_cpumask_var(rd
->online
);
8077 free_cpumask_var(rd
->span
);
8082 static void init_defrootdomain(void)
8084 init_rootdomain(&def_root_domain
, true);
8086 atomic_set(&def_root_domain
.refcount
, 1);
8089 static struct root_domain
*alloc_rootdomain(void)
8091 struct root_domain
*rd
;
8093 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8097 if (init_rootdomain(rd
, false) != 0) {
8106 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8107 * hold the hotplug lock.
8110 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8112 struct rq
*rq
= cpu_rq(cpu
);
8113 struct sched_domain
*tmp
;
8115 /* Remove the sched domains which do not contribute to scheduling. */
8116 for (tmp
= sd
; tmp
; ) {
8117 struct sched_domain
*parent
= tmp
->parent
;
8121 if (sd_parent_degenerate(tmp
, parent
)) {
8122 tmp
->parent
= parent
->parent
;
8124 parent
->parent
->child
= tmp
;
8129 if (sd
&& sd_degenerate(sd
)) {
8135 sched_domain_debug(sd
, cpu
);
8137 rq_attach_root(rq
, rd
);
8138 rcu_assign_pointer(rq
->sd
, sd
);
8141 /* cpus with isolated domains */
8142 static cpumask_var_t cpu_isolated_map
;
8144 /* Setup the mask of cpus configured for isolated domains */
8145 static int __init
isolated_cpu_setup(char *str
)
8147 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8148 cpulist_parse(str
, cpu_isolated_map
);
8152 __setup("isolcpus=", isolated_cpu_setup
);
8155 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8156 * to a function which identifies what group(along with sched group) a CPU
8157 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8158 * (due to the fact that we keep track of groups covered with a struct cpumask).
8160 * init_sched_build_groups will build a circular linked list of the groups
8161 * covered by the given span, and will set each group's ->cpumask correctly,
8162 * and ->cpu_power to 0.
8165 init_sched_build_groups(const struct cpumask
*span
,
8166 const struct cpumask
*cpu_map
,
8167 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8168 struct sched_group
**sg
,
8169 struct cpumask
*tmpmask
),
8170 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8172 struct sched_group
*first
= NULL
, *last
= NULL
;
8175 cpumask_clear(covered
);
8177 for_each_cpu(i
, span
) {
8178 struct sched_group
*sg
;
8179 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8182 if (cpumask_test_cpu(i
, covered
))
8185 cpumask_clear(sched_group_cpus(sg
));
8188 for_each_cpu(j
, span
) {
8189 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8192 cpumask_set_cpu(j
, covered
);
8193 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8204 #define SD_NODES_PER_DOMAIN 16
8209 * find_next_best_node - find the next node to include in a sched_domain
8210 * @node: node whose sched_domain we're building
8211 * @used_nodes: nodes already in the sched_domain
8213 * Find the next node to include in a given scheduling domain. Simply
8214 * finds the closest node not already in the @used_nodes map.
8216 * Should use nodemask_t.
8218 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8220 int i
, n
, val
, min_val
, best_node
= 0;
8224 for (i
= 0; i
< nr_node_ids
; i
++) {
8225 /* Start at @node */
8226 n
= (node
+ i
) % nr_node_ids
;
8228 if (!nr_cpus_node(n
))
8231 /* Skip already used nodes */
8232 if (node_isset(n
, *used_nodes
))
8235 /* Simple min distance search */
8236 val
= node_distance(node
, n
);
8238 if (val
< min_val
) {
8244 node_set(best_node
, *used_nodes
);
8249 * sched_domain_node_span - get a cpumask for a node's sched_domain
8250 * @node: node whose cpumask we're constructing
8251 * @span: resulting cpumask
8253 * Given a node, construct a good cpumask for its sched_domain to span. It
8254 * should be one that prevents unnecessary balancing, but also spreads tasks
8257 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8259 nodemask_t used_nodes
;
8262 cpumask_clear(span
);
8263 nodes_clear(used_nodes
);
8265 cpumask_or(span
, span
, cpumask_of_node(node
));
8266 node_set(node
, used_nodes
);
8268 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8269 int next_node
= find_next_best_node(node
, &used_nodes
);
8271 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8274 #endif /* CONFIG_NUMA */
8276 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8279 * The cpus mask in sched_group and sched_domain hangs off the end.
8281 * ( See the the comments in include/linux/sched.h:struct sched_group
8282 * and struct sched_domain. )
8284 struct static_sched_group
{
8285 struct sched_group sg
;
8286 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8289 struct static_sched_domain
{
8290 struct sched_domain sd
;
8291 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8297 cpumask_var_t domainspan
;
8298 cpumask_var_t covered
;
8299 cpumask_var_t notcovered
;
8301 cpumask_var_t nodemask
;
8302 cpumask_var_t this_sibling_map
;
8303 cpumask_var_t this_core_map
;
8304 cpumask_var_t send_covered
;
8305 cpumask_var_t tmpmask
;
8306 struct sched_group
**sched_group_nodes
;
8307 struct root_domain
*rd
;
8311 sa_sched_groups
= 0,
8316 sa_this_sibling_map
,
8318 sa_sched_group_nodes
,
8328 * SMT sched-domains:
8330 #ifdef CONFIG_SCHED_SMT
8331 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8332 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8335 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8336 struct sched_group
**sg
, struct cpumask
*unused
)
8339 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8342 #endif /* CONFIG_SCHED_SMT */
8345 * multi-core sched-domains:
8347 #ifdef CONFIG_SCHED_MC
8348 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8349 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8350 #endif /* CONFIG_SCHED_MC */
8352 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8354 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8355 struct sched_group
**sg
, struct cpumask
*mask
)
8359 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8360 group
= cpumask_first(mask
);
8362 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8365 #elif defined(CONFIG_SCHED_MC)
8367 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8368 struct sched_group
**sg
, struct cpumask
*unused
)
8371 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8376 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8377 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8380 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8381 struct sched_group
**sg
, struct cpumask
*mask
)
8384 #ifdef CONFIG_SCHED_MC
8385 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8386 group
= cpumask_first(mask
);
8387 #elif defined(CONFIG_SCHED_SMT)
8388 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8389 group
= cpumask_first(mask
);
8394 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8400 * The init_sched_build_groups can't handle what we want to do with node
8401 * groups, so roll our own. Now each node has its own list of groups which
8402 * gets dynamically allocated.
8404 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8405 static struct sched_group
***sched_group_nodes_bycpu
;
8407 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8408 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8410 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8411 struct sched_group
**sg
,
8412 struct cpumask
*nodemask
)
8416 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8417 group
= cpumask_first(nodemask
);
8420 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8424 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8426 struct sched_group
*sg
= group_head
;
8432 for_each_cpu(j
, sched_group_cpus(sg
)) {
8433 struct sched_domain
*sd
;
8435 sd
= &per_cpu(phys_domains
, j
).sd
;
8436 if (j
!= group_first_cpu(sd
->groups
)) {
8438 * Only add "power" once for each
8444 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8447 } while (sg
!= group_head
);
8450 static int build_numa_sched_groups(struct s_data
*d
,
8451 const struct cpumask
*cpu_map
, int num
)
8453 struct sched_domain
*sd
;
8454 struct sched_group
*sg
, *prev
;
8457 cpumask_clear(d
->covered
);
8458 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8459 if (cpumask_empty(d
->nodemask
)) {
8460 d
->sched_group_nodes
[num
] = NULL
;
8464 sched_domain_node_span(num
, d
->domainspan
);
8465 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8467 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8470 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8474 d
->sched_group_nodes
[num
] = sg
;
8476 for_each_cpu(j
, d
->nodemask
) {
8477 sd
= &per_cpu(node_domains
, j
).sd
;
8482 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8484 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8487 for (j
= 0; j
< nr_node_ids
; j
++) {
8488 n
= (num
+ j
) % nr_node_ids
;
8489 cpumask_complement(d
->notcovered
, d
->covered
);
8490 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8491 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8492 if (cpumask_empty(d
->tmpmask
))
8494 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8495 if (cpumask_empty(d
->tmpmask
))
8497 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8501 "Can not alloc domain group for node %d\n", j
);
8505 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8506 sg
->next
= prev
->next
;
8507 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8514 #endif /* CONFIG_NUMA */
8517 /* Free memory allocated for various sched_group structures */
8518 static void free_sched_groups(const struct cpumask
*cpu_map
,
8519 struct cpumask
*nodemask
)
8523 for_each_cpu(cpu
, cpu_map
) {
8524 struct sched_group
**sched_group_nodes
8525 = sched_group_nodes_bycpu
[cpu
];
8527 if (!sched_group_nodes
)
8530 for (i
= 0; i
< nr_node_ids
; i
++) {
8531 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8533 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8534 if (cpumask_empty(nodemask
))
8544 if (oldsg
!= sched_group_nodes
[i
])
8547 kfree(sched_group_nodes
);
8548 sched_group_nodes_bycpu
[cpu
] = NULL
;
8551 #else /* !CONFIG_NUMA */
8552 static void free_sched_groups(const struct cpumask
*cpu_map
,
8553 struct cpumask
*nodemask
)
8556 #endif /* CONFIG_NUMA */
8559 * Initialize sched groups cpu_power.
8561 * cpu_power indicates the capacity of sched group, which is used while
8562 * distributing the load between different sched groups in a sched domain.
8563 * Typically cpu_power for all the groups in a sched domain will be same unless
8564 * there are asymmetries in the topology. If there are asymmetries, group
8565 * having more cpu_power will pickup more load compared to the group having
8568 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8570 struct sched_domain
*child
;
8571 struct sched_group
*group
;
8575 WARN_ON(!sd
|| !sd
->groups
);
8577 if (cpu
!= group_first_cpu(sd
->groups
))
8582 sd
->groups
->cpu_power
= 0;
8585 power
= SCHED_LOAD_SCALE
;
8586 weight
= cpumask_weight(sched_domain_span(sd
));
8588 * SMT siblings share the power of a single core.
8589 * Usually multiple threads get a better yield out of
8590 * that one core than a single thread would have,
8591 * reflect that in sd->smt_gain.
8593 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8594 power
*= sd
->smt_gain
;
8596 power
>>= SCHED_LOAD_SHIFT
;
8598 sd
->groups
->cpu_power
+= power
;
8603 * Add cpu_power of each child group to this groups cpu_power.
8605 group
= child
->groups
;
8607 sd
->groups
->cpu_power
+= group
->cpu_power
;
8608 group
= group
->next
;
8609 } while (group
!= child
->groups
);
8613 * Initializers for schedule domains
8614 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8617 #ifdef CONFIG_SCHED_DEBUG
8618 # define SD_INIT_NAME(sd, type) sd->name = #type
8620 # define SD_INIT_NAME(sd, type) do { } while (0)
8623 #define SD_INIT(sd, type) sd_init_##type(sd)
8625 #define SD_INIT_FUNC(type) \
8626 static noinline void sd_init_##type(struct sched_domain *sd) \
8628 memset(sd, 0, sizeof(*sd)); \
8629 *sd = SD_##type##_INIT; \
8630 sd->level = SD_LV_##type; \
8631 SD_INIT_NAME(sd, type); \
8636 SD_INIT_FUNC(ALLNODES
)
8639 #ifdef CONFIG_SCHED_SMT
8640 SD_INIT_FUNC(SIBLING
)
8642 #ifdef CONFIG_SCHED_MC
8646 static int default_relax_domain_level
= -1;
8648 static int __init
setup_relax_domain_level(char *str
)
8652 val
= simple_strtoul(str
, NULL
, 0);
8653 if (val
< SD_LV_MAX
)
8654 default_relax_domain_level
= val
;
8658 __setup("relax_domain_level=", setup_relax_domain_level
);
8660 static void set_domain_attribute(struct sched_domain
*sd
,
8661 struct sched_domain_attr
*attr
)
8665 if (!attr
|| attr
->relax_domain_level
< 0) {
8666 if (default_relax_domain_level
< 0)
8669 request
= default_relax_domain_level
;
8671 request
= attr
->relax_domain_level
;
8672 if (request
< sd
->level
) {
8673 /* turn off idle balance on this domain */
8674 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8676 /* turn on idle balance on this domain */
8677 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8681 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8682 const struct cpumask
*cpu_map
)
8685 case sa_sched_groups
:
8686 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8687 d
->sched_group_nodes
= NULL
;
8689 free_rootdomain(d
->rd
); /* fall through */
8691 free_cpumask_var(d
->tmpmask
); /* fall through */
8692 case sa_send_covered
:
8693 free_cpumask_var(d
->send_covered
); /* fall through */
8694 case sa_this_core_map
:
8695 free_cpumask_var(d
->this_core_map
); /* fall through */
8696 case sa_this_sibling_map
:
8697 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8699 free_cpumask_var(d
->nodemask
); /* fall through */
8700 case sa_sched_group_nodes
:
8702 kfree(d
->sched_group_nodes
); /* fall through */
8704 free_cpumask_var(d
->notcovered
); /* fall through */
8706 free_cpumask_var(d
->covered
); /* fall through */
8708 free_cpumask_var(d
->domainspan
); /* fall through */
8715 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8716 const struct cpumask
*cpu_map
)
8719 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8721 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8722 return sa_domainspan
;
8723 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8725 /* Allocate the per-node list of sched groups */
8726 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8727 sizeof(struct sched_group
*), GFP_KERNEL
);
8728 if (!d
->sched_group_nodes
) {
8729 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8730 return sa_notcovered
;
8732 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8734 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8735 return sa_sched_group_nodes
;
8736 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8738 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8739 return sa_this_sibling_map
;
8740 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8741 return sa_this_core_map
;
8742 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8743 return sa_send_covered
;
8744 d
->rd
= alloc_rootdomain();
8746 printk(KERN_WARNING
"Cannot alloc root domain\n");
8749 return sa_rootdomain
;
8752 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8753 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8755 struct sched_domain
*sd
= NULL
;
8757 struct sched_domain
*parent
;
8760 if (cpumask_weight(cpu_map
) >
8761 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8762 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8763 SD_INIT(sd
, ALLNODES
);
8764 set_domain_attribute(sd
, attr
);
8765 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8766 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8771 sd
= &per_cpu(node_domains
, i
).sd
;
8773 set_domain_attribute(sd
, attr
);
8774 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8775 sd
->parent
= parent
;
8778 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8783 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8784 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8785 struct sched_domain
*parent
, int i
)
8787 struct sched_domain
*sd
;
8788 sd
= &per_cpu(phys_domains
, i
).sd
;
8790 set_domain_attribute(sd
, attr
);
8791 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8792 sd
->parent
= parent
;
8795 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8799 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8800 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8801 struct sched_domain
*parent
, int i
)
8803 struct sched_domain
*sd
= parent
;
8804 #ifdef CONFIG_SCHED_MC
8805 sd
= &per_cpu(core_domains
, i
).sd
;
8807 set_domain_attribute(sd
, attr
);
8808 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8809 sd
->parent
= parent
;
8811 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8816 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8817 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8818 struct sched_domain
*parent
, int i
)
8820 struct sched_domain
*sd
= parent
;
8821 #ifdef CONFIG_SCHED_SMT
8822 sd
= &per_cpu(cpu_domains
, i
).sd
;
8823 SD_INIT(sd
, SIBLING
);
8824 set_domain_attribute(sd
, attr
);
8825 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8826 sd
->parent
= parent
;
8828 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8833 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8834 const struct cpumask
*cpu_map
, int cpu
)
8837 #ifdef CONFIG_SCHED_SMT
8838 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8839 cpumask_and(d
->this_sibling_map
, cpu_map
,
8840 topology_thread_cpumask(cpu
));
8841 if (cpu
== cpumask_first(d
->this_sibling_map
))
8842 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8844 d
->send_covered
, d
->tmpmask
);
8847 #ifdef CONFIG_SCHED_MC
8848 case SD_LV_MC
: /* set up multi-core groups */
8849 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8850 if (cpu
== cpumask_first(d
->this_core_map
))
8851 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8853 d
->send_covered
, d
->tmpmask
);
8856 case SD_LV_CPU
: /* set up physical groups */
8857 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8858 if (!cpumask_empty(d
->nodemask
))
8859 init_sched_build_groups(d
->nodemask
, cpu_map
,
8861 d
->send_covered
, d
->tmpmask
);
8864 case SD_LV_ALLNODES
:
8865 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8866 d
->send_covered
, d
->tmpmask
);
8875 * Build sched domains for a given set of cpus and attach the sched domains
8876 * to the individual cpus
8878 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8879 struct sched_domain_attr
*attr
)
8881 enum s_alloc alloc_state
= sa_none
;
8883 struct sched_domain
*sd
;
8889 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8890 if (alloc_state
!= sa_rootdomain
)
8892 alloc_state
= sa_sched_groups
;
8895 * Set up domains for cpus specified by the cpu_map.
8897 for_each_cpu(i
, cpu_map
) {
8898 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8901 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8902 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8903 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8904 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8907 for_each_cpu(i
, cpu_map
) {
8908 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8909 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8912 /* Set up physical groups */
8913 for (i
= 0; i
< nr_node_ids
; i
++)
8914 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8917 /* Set up node groups */
8919 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8921 for (i
= 0; i
< nr_node_ids
; i
++)
8922 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8926 /* Calculate CPU power for physical packages and nodes */
8927 #ifdef CONFIG_SCHED_SMT
8928 for_each_cpu(i
, cpu_map
) {
8929 sd
= &per_cpu(cpu_domains
, i
).sd
;
8930 init_sched_groups_power(i
, sd
);
8933 #ifdef CONFIG_SCHED_MC
8934 for_each_cpu(i
, cpu_map
) {
8935 sd
= &per_cpu(core_domains
, i
).sd
;
8936 init_sched_groups_power(i
, sd
);
8940 for_each_cpu(i
, cpu_map
) {
8941 sd
= &per_cpu(phys_domains
, i
).sd
;
8942 init_sched_groups_power(i
, sd
);
8946 for (i
= 0; i
< nr_node_ids
; i
++)
8947 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8949 if (d
.sd_allnodes
) {
8950 struct sched_group
*sg
;
8952 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8954 init_numa_sched_groups_power(sg
);
8958 /* Attach the domains */
8959 for_each_cpu(i
, cpu_map
) {
8960 #ifdef CONFIG_SCHED_SMT
8961 sd
= &per_cpu(cpu_domains
, i
).sd
;
8962 #elif defined(CONFIG_SCHED_MC)
8963 sd
= &per_cpu(core_domains
, i
).sd
;
8965 sd
= &per_cpu(phys_domains
, i
).sd
;
8967 cpu_attach_domain(sd
, d
.rd
, i
);
8970 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8971 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8975 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8979 static int build_sched_domains(const struct cpumask
*cpu_map
)
8981 return __build_sched_domains(cpu_map
, NULL
);
8984 static struct cpumask
*doms_cur
; /* current sched domains */
8985 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8986 static struct sched_domain_attr
*dattr_cur
;
8987 /* attribues of custom domains in 'doms_cur' */
8990 * Special case: If a kmalloc of a doms_cur partition (array of
8991 * cpumask) fails, then fallback to a single sched domain,
8992 * as determined by the single cpumask fallback_doms.
8994 static cpumask_var_t fallback_doms
;
8997 * arch_update_cpu_topology lets virtualized architectures update the
8998 * cpu core maps. It is supposed to return 1 if the topology changed
8999 * or 0 if it stayed the same.
9001 int __attribute__((weak
)) arch_update_cpu_topology(void)
9007 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9008 * For now this just excludes isolated cpus, but could be used to
9009 * exclude other special cases in the future.
9011 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9015 arch_update_cpu_topology();
9017 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
9019 doms_cur
= fallback_doms
;
9020 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
9022 err
= build_sched_domains(doms_cur
);
9023 register_sched_domain_sysctl();
9028 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9029 struct cpumask
*tmpmask
)
9031 free_sched_groups(cpu_map
, tmpmask
);
9035 * Detach sched domains from a group of cpus specified in cpu_map
9036 * These cpus will now be attached to the NULL domain
9038 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9040 /* Save because hotplug lock held. */
9041 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9044 for_each_cpu(i
, cpu_map
)
9045 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9046 synchronize_sched();
9047 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9050 /* handle null as "default" */
9051 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9052 struct sched_domain_attr
*new, int idx_new
)
9054 struct sched_domain_attr tmp
;
9061 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9062 new ? (new + idx_new
) : &tmp
,
9063 sizeof(struct sched_domain_attr
));
9067 * Partition sched domains as specified by the 'ndoms_new'
9068 * cpumasks in the array doms_new[] of cpumasks. This compares
9069 * doms_new[] to the current sched domain partitioning, doms_cur[].
9070 * It destroys each deleted domain and builds each new domain.
9072 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9073 * The masks don't intersect (don't overlap.) We should setup one
9074 * sched domain for each mask. CPUs not in any of the cpumasks will
9075 * not be load balanced. If the same cpumask appears both in the
9076 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9079 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9080 * ownership of it and will kfree it when done with it. If the caller
9081 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9082 * ndoms_new == 1, and partition_sched_domains() will fallback to
9083 * the single partition 'fallback_doms', it also forces the domains
9086 * If doms_new == NULL it will be replaced with cpu_online_mask.
9087 * ndoms_new == 0 is a special case for destroying existing domains,
9088 * and it will not create the default domain.
9090 * Call with hotplug lock held
9092 /* FIXME: Change to struct cpumask *doms_new[] */
9093 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9094 struct sched_domain_attr
*dattr_new
)
9099 mutex_lock(&sched_domains_mutex
);
9101 /* always unregister in case we don't destroy any domains */
9102 unregister_sched_domain_sysctl();
9104 /* Let architecture update cpu core mappings. */
9105 new_topology
= arch_update_cpu_topology();
9107 n
= doms_new
? ndoms_new
: 0;
9109 /* Destroy deleted domains */
9110 for (i
= 0; i
< ndoms_cur
; i
++) {
9111 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9112 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9113 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9116 /* no match - a current sched domain not in new doms_new[] */
9117 detach_destroy_domains(doms_cur
+ i
);
9122 if (doms_new
== NULL
) {
9124 doms_new
= fallback_doms
;
9125 cpumask_andnot(&doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
9126 WARN_ON_ONCE(dattr_new
);
9129 /* Build new domains */
9130 for (i
= 0; i
< ndoms_new
; i
++) {
9131 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9132 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9133 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9136 /* no match - add a new doms_new */
9137 __build_sched_domains(doms_new
+ i
,
9138 dattr_new
? dattr_new
+ i
: NULL
);
9143 /* Remember the new sched domains */
9144 if (doms_cur
!= fallback_doms
)
9146 kfree(dattr_cur
); /* kfree(NULL) is safe */
9147 doms_cur
= doms_new
;
9148 dattr_cur
= dattr_new
;
9149 ndoms_cur
= ndoms_new
;
9151 register_sched_domain_sysctl();
9153 mutex_unlock(&sched_domains_mutex
);
9156 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9157 static void arch_reinit_sched_domains(void)
9161 /* Destroy domains first to force the rebuild */
9162 partition_sched_domains(0, NULL
, NULL
);
9164 rebuild_sched_domains();
9168 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9170 unsigned int level
= 0;
9172 if (sscanf(buf
, "%u", &level
) != 1)
9176 * level is always be positive so don't check for
9177 * level < POWERSAVINGS_BALANCE_NONE which is 0
9178 * What happens on 0 or 1 byte write,
9179 * need to check for count as well?
9182 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9186 sched_smt_power_savings
= level
;
9188 sched_mc_power_savings
= level
;
9190 arch_reinit_sched_domains();
9195 #ifdef CONFIG_SCHED_MC
9196 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9199 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9201 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9202 const char *buf
, size_t count
)
9204 return sched_power_savings_store(buf
, count
, 0);
9206 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9207 sched_mc_power_savings_show
,
9208 sched_mc_power_savings_store
);
9211 #ifdef CONFIG_SCHED_SMT
9212 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9215 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9217 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9218 const char *buf
, size_t count
)
9220 return sched_power_savings_store(buf
, count
, 1);
9222 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9223 sched_smt_power_savings_show
,
9224 sched_smt_power_savings_store
);
9227 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9231 #ifdef CONFIG_SCHED_SMT
9233 err
= sysfs_create_file(&cls
->kset
.kobj
,
9234 &attr_sched_smt_power_savings
.attr
);
9236 #ifdef CONFIG_SCHED_MC
9237 if (!err
&& mc_capable())
9238 err
= sysfs_create_file(&cls
->kset
.kobj
,
9239 &attr_sched_mc_power_savings
.attr
);
9243 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9245 #ifndef CONFIG_CPUSETS
9247 * Add online and remove offline CPUs from the scheduler domains.
9248 * When cpusets are enabled they take over this function.
9250 static int update_sched_domains(struct notifier_block
*nfb
,
9251 unsigned long action
, void *hcpu
)
9255 case CPU_ONLINE_FROZEN
:
9256 case CPU_DOWN_PREPARE
:
9257 case CPU_DOWN_PREPARE_FROZEN
:
9258 case CPU_DOWN_FAILED
:
9259 case CPU_DOWN_FAILED_FROZEN
:
9260 partition_sched_domains(1, NULL
, NULL
);
9269 static int update_runtime(struct notifier_block
*nfb
,
9270 unsigned long action
, void *hcpu
)
9272 int cpu
= (int)(long)hcpu
;
9275 case CPU_DOWN_PREPARE
:
9276 case CPU_DOWN_PREPARE_FROZEN
:
9277 disable_runtime(cpu_rq(cpu
));
9280 case CPU_DOWN_FAILED
:
9281 case CPU_DOWN_FAILED_FROZEN
:
9283 case CPU_ONLINE_FROZEN
:
9284 enable_runtime(cpu_rq(cpu
));
9292 void __init
sched_init_smp(void)
9294 cpumask_var_t non_isolated_cpus
;
9296 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9297 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9299 #if defined(CONFIG_NUMA)
9300 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9302 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9305 mutex_lock(&sched_domains_mutex
);
9306 arch_init_sched_domains(cpu_active_mask
);
9307 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9308 if (cpumask_empty(non_isolated_cpus
))
9309 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9310 mutex_unlock(&sched_domains_mutex
);
9313 #ifndef CONFIG_CPUSETS
9314 /* XXX: Theoretical race here - CPU may be hotplugged now */
9315 hotcpu_notifier(update_sched_domains
, 0);
9318 /* RT runtime code needs to handle some hotplug events */
9319 hotcpu_notifier(update_runtime
, 0);
9323 /* Move init over to a non-isolated CPU */
9324 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9326 sched_init_granularity();
9327 free_cpumask_var(non_isolated_cpus
);
9329 init_sched_rt_class();
9332 void __init
sched_init_smp(void)
9334 sched_init_granularity();
9336 #endif /* CONFIG_SMP */
9338 const_debug
unsigned int sysctl_timer_migration
= 1;
9340 int in_sched_functions(unsigned long addr
)
9342 return in_lock_functions(addr
) ||
9343 (addr
>= (unsigned long)__sched_text_start
9344 && addr
< (unsigned long)__sched_text_end
);
9347 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9349 cfs_rq
->tasks_timeline
= RB_ROOT
;
9350 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9351 #ifdef CONFIG_FAIR_GROUP_SCHED
9354 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9357 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9359 struct rt_prio_array
*array
;
9362 array
= &rt_rq
->active
;
9363 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9364 INIT_LIST_HEAD(array
->queue
+ i
);
9365 __clear_bit(i
, array
->bitmap
);
9367 /* delimiter for bitsearch: */
9368 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9370 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9371 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9373 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9377 rt_rq
->rt_nr_migratory
= 0;
9378 rt_rq
->overloaded
= 0;
9379 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9383 rt_rq
->rt_throttled
= 0;
9384 rt_rq
->rt_runtime
= 0;
9385 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9387 #ifdef CONFIG_RT_GROUP_SCHED
9388 rt_rq
->rt_nr_boosted
= 0;
9393 #ifdef CONFIG_FAIR_GROUP_SCHED
9394 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9395 struct sched_entity
*se
, int cpu
, int add
,
9396 struct sched_entity
*parent
)
9398 struct rq
*rq
= cpu_rq(cpu
);
9399 tg
->cfs_rq
[cpu
] = cfs_rq
;
9400 init_cfs_rq(cfs_rq
, rq
);
9403 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9406 /* se could be NULL for init_task_group */
9411 se
->cfs_rq
= &rq
->cfs
;
9413 se
->cfs_rq
= parent
->my_q
;
9416 se
->load
.weight
= tg
->shares
;
9417 se
->load
.inv_weight
= 0;
9418 se
->parent
= parent
;
9422 #ifdef CONFIG_RT_GROUP_SCHED
9423 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9424 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9425 struct sched_rt_entity
*parent
)
9427 struct rq
*rq
= cpu_rq(cpu
);
9429 tg
->rt_rq
[cpu
] = rt_rq
;
9430 init_rt_rq(rt_rq
, rq
);
9432 rt_rq
->rt_se
= rt_se
;
9433 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9435 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9437 tg
->rt_se
[cpu
] = rt_se
;
9442 rt_se
->rt_rq
= &rq
->rt
;
9444 rt_se
->rt_rq
= parent
->my_q
;
9446 rt_se
->my_q
= rt_rq
;
9447 rt_se
->parent
= parent
;
9448 INIT_LIST_HEAD(&rt_se
->run_list
);
9452 void __init
sched_init(void)
9455 unsigned long alloc_size
= 0, ptr
;
9457 #ifdef CONFIG_FAIR_GROUP_SCHED
9458 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9460 #ifdef CONFIG_RT_GROUP_SCHED
9461 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9463 #ifdef CONFIG_USER_SCHED
9466 #ifdef CONFIG_CPUMASK_OFFSTACK
9467 alloc_size
+= num_possible_cpus() * cpumask_size();
9470 * As sched_init() is called before page_alloc is setup,
9471 * we use alloc_bootmem().
9474 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9476 #ifdef CONFIG_FAIR_GROUP_SCHED
9477 init_task_group
.se
= (struct sched_entity
**)ptr
;
9478 ptr
+= nr_cpu_ids
* sizeof(void **);
9480 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9481 ptr
+= nr_cpu_ids
* sizeof(void **);
9483 #ifdef CONFIG_USER_SCHED
9484 root_task_group
.se
= (struct sched_entity
**)ptr
;
9485 ptr
+= nr_cpu_ids
* sizeof(void **);
9487 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9488 ptr
+= nr_cpu_ids
* sizeof(void **);
9489 #endif /* CONFIG_USER_SCHED */
9490 #endif /* CONFIG_FAIR_GROUP_SCHED */
9491 #ifdef CONFIG_RT_GROUP_SCHED
9492 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9493 ptr
+= nr_cpu_ids
* sizeof(void **);
9495 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9496 ptr
+= nr_cpu_ids
* sizeof(void **);
9498 #ifdef CONFIG_USER_SCHED
9499 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9500 ptr
+= nr_cpu_ids
* sizeof(void **);
9502 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9503 ptr
+= nr_cpu_ids
* sizeof(void **);
9504 #endif /* CONFIG_USER_SCHED */
9505 #endif /* CONFIG_RT_GROUP_SCHED */
9506 #ifdef CONFIG_CPUMASK_OFFSTACK
9507 for_each_possible_cpu(i
) {
9508 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9509 ptr
+= cpumask_size();
9511 #endif /* CONFIG_CPUMASK_OFFSTACK */
9515 init_defrootdomain();
9518 init_rt_bandwidth(&def_rt_bandwidth
,
9519 global_rt_period(), global_rt_runtime());
9521 #ifdef CONFIG_RT_GROUP_SCHED
9522 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9523 global_rt_period(), global_rt_runtime());
9524 #ifdef CONFIG_USER_SCHED
9525 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9526 global_rt_period(), RUNTIME_INF
);
9527 #endif /* CONFIG_USER_SCHED */
9528 #endif /* CONFIG_RT_GROUP_SCHED */
9530 #ifdef CONFIG_GROUP_SCHED
9531 list_add(&init_task_group
.list
, &task_groups
);
9532 INIT_LIST_HEAD(&init_task_group
.children
);
9534 #ifdef CONFIG_USER_SCHED
9535 INIT_LIST_HEAD(&root_task_group
.children
);
9536 init_task_group
.parent
= &root_task_group
;
9537 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9538 #endif /* CONFIG_USER_SCHED */
9539 #endif /* CONFIG_GROUP_SCHED */
9541 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9542 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9543 __alignof__(unsigned long));
9545 for_each_possible_cpu(i
) {
9549 spin_lock_init(&rq
->lock
);
9551 rq
->calc_load_active
= 0;
9552 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9553 init_cfs_rq(&rq
->cfs
, rq
);
9554 init_rt_rq(&rq
->rt
, rq
);
9555 #ifdef CONFIG_FAIR_GROUP_SCHED
9556 init_task_group
.shares
= init_task_group_load
;
9557 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9558 #ifdef CONFIG_CGROUP_SCHED
9560 * How much cpu bandwidth does init_task_group get?
9562 * In case of task-groups formed thr' the cgroup filesystem, it
9563 * gets 100% of the cpu resources in the system. This overall
9564 * system cpu resource is divided among the tasks of
9565 * init_task_group and its child task-groups in a fair manner,
9566 * based on each entity's (task or task-group's) weight
9567 * (se->load.weight).
9569 * In other words, if init_task_group has 10 tasks of weight
9570 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9571 * then A0's share of the cpu resource is:
9573 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9575 * We achieve this by letting init_task_group's tasks sit
9576 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9578 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9579 #elif defined CONFIG_USER_SCHED
9580 root_task_group
.shares
= NICE_0_LOAD
;
9581 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9583 * In case of task-groups formed thr' the user id of tasks,
9584 * init_task_group represents tasks belonging to root user.
9585 * Hence it forms a sibling of all subsequent groups formed.
9586 * In this case, init_task_group gets only a fraction of overall
9587 * system cpu resource, based on the weight assigned to root
9588 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9589 * by letting tasks of init_task_group sit in a separate cfs_rq
9590 * (init_tg_cfs_rq) and having one entity represent this group of
9591 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9593 init_tg_cfs_entry(&init_task_group
,
9594 &per_cpu(init_tg_cfs_rq
, i
),
9595 &per_cpu(init_sched_entity
, i
), i
, 1,
9596 root_task_group
.se
[i
]);
9599 #endif /* CONFIG_FAIR_GROUP_SCHED */
9601 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9602 #ifdef CONFIG_RT_GROUP_SCHED
9603 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9604 #ifdef CONFIG_CGROUP_SCHED
9605 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9606 #elif defined CONFIG_USER_SCHED
9607 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9608 init_tg_rt_entry(&init_task_group
,
9609 &per_cpu(init_rt_rq
, i
),
9610 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9611 root_task_group
.rt_se
[i
]);
9615 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9616 rq
->cpu_load
[j
] = 0;
9620 rq
->post_schedule
= 0;
9621 rq
->active_balance
= 0;
9622 rq
->next_balance
= jiffies
;
9626 rq
->migration_thread
= NULL
;
9628 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9629 INIT_LIST_HEAD(&rq
->migration_queue
);
9630 rq_attach_root(rq
, &def_root_domain
);
9633 atomic_set(&rq
->nr_iowait
, 0);
9636 set_load_weight(&init_task
);
9638 #ifdef CONFIG_PREEMPT_NOTIFIERS
9639 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9643 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9646 #ifdef CONFIG_RT_MUTEXES
9647 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9651 * The boot idle thread does lazy MMU switching as well:
9653 atomic_inc(&init_mm
.mm_count
);
9654 enter_lazy_tlb(&init_mm
, current
);
9657 * Make us the idle thread. Technically, schedule() should not be
9658 * called from this thread, however somewhere below it might be,
9659 * but because we are the idle thread, we just pick up running again
9660 * when this runqueue becomes "idle".
9662 init_idle(current
, smp_processor_id());
9664 calc_load_update
= jiffies
+ LOAD_FREQ
;
9667 * During early bootup we pretend to be a normal task:
9669 current
->sched_class
= &fair_sched_class
;
9671 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9672 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9675 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9676 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9678 /* May be allocated at isolcpus cmdline parse time */
9679 if (cpu_isolated_map
== NULL
)
9680 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9685 scheduler_running
= 1;
9688 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9689 static inline int preempt_count_equals(int preempt_offset
)
9691 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9693 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9696 void __might_sleep(char *file
, int line
, int preempt_offset
)
9699 static unsigned long prev_jiffy
; /* ratelimiting */
9701 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9702 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9704 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9706 prev_jiffy
= jiffies
;
9709 "BUG: sleeping function called from invalid context at %s:%d\n",
9712 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9713 in_atomic(), irqs_disabled(),
9714 current
->pid
, current
->comm
);
9716 debug_show_held_locks(current
);
9717 if (irqs_disabled())
9718 print_irqtrace_events(current
);
9722 EXPORT_SYMBOL(__might_sleep
);
9725 #ifdef CONFIG_MAGIC_SYSRQ
9726 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9730 update_rq_clock(rq
);
9731 on_rq
= p
->se
.on_rq
;
9733 deactivate_task(rq
, p
, 0);
9734 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9736 activate_task(rq
, p
, 0);
9737 resched_task(rq
->curr
);
9741 void normalize_rt_tasks(void)
9743 struct task_struct
*g
, *p
;
9744 unsigned long flags
;
9747 read_lock_irqsave(&tasklist_lock
, flags
);
9748 do_each_thread(g
, p
) {
9750 * Only normalize user tasks:
9755 p
->se
.exec_start
= 0;
9756 #ifdef CONFIG_SCHEDSTATS
9757 p
->se
.wait_start
= 0;
9758 p
->se
.sleep_start
= 0;
9759 p
->se
.block_start
= 0;
9764 * Renice negative nice level userspace
9767 if (TASK_NICE(p
) < 0 && p
->mm
)
9768 set_user_nice(p
, 0);
9772 spin_lock(&p
->pi_lock
);
9773 rq
= __task_rq_lock(p
);
9775 normalize_task(rq
, p
);
9777 __task_rq_unlock(rq
);
9778 spin_unlock(&p
->pi_lock
);
9779 } while_each_thread(g
, p
);
9781 read_unlock_irqrestore(&tasklist_lock
, flags
);
9784 #endif /* CONFIG_MAGIC_SYSRQ */
9788 * These functions are only useful for the IA64 MCA handling.
9790 * They can only be called when the whole system has been
9791 * stopped - every CPU needs to be quiescent, and no scheduling
9792 * activity can take place. Using them for anything else would
9793 * be a serious bug, and as a result, they aren't even visible
9794 * under any other configuration.
9798 * curr_task - return the current task for a given cpu.
9799 * @cpu: the processor in question.
9801 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9803 struct task_struct
*curr_task(int cpu
)
9805 return cpu_curr(cpu
);
9809 * set_curr_task - set the current task for a given cpu.
9810 * @cpu: the processor in question.
9811 * @p: the task pointer to set.
9813 * Description: This function must only be used when non-maskable interrupts
9814 * are serviced on a separate stack. It allows the architecture to switch the
9815 * notion of the current task on a cpu in a non-blocking manner. This function
9816 * must be called with all CPU's synchronized, and interrupts disabled, the
9817 * and caller must save the original value of the current task (see
9818 * curr_task() above) and restore that value before reenabling interrupts and
9819 * re-starting the system.
9821 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9823 void set_curr_task(int cpu
, struct task_struct
*p
)
9830 #ifdef CONFIG_FAIR_GROUP_SCHED
9831 static void free_fair_sched_group(struct task_group
*tg
)
9835 for_each_possible_cpu(i
) {
9837 kfree(tg
->cfs_rq
[i
]);
9847 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9849 struct cfs_rq
*cfs_rq
;
9850 struct sched_entity
*se
;
9854 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9857 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9861 tg
->shares
= NICE_0_LOAD
;
9863 for_each_possible_cpu(i
) {
9866 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9867 GFP_KERNEL
, cpu_to_node(i
));
9871 se
= kzalloc_node(sizeof(struct sched_entity
),
9872 GFP_KERNEL
, cpu_to_node(i
));
9876 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9885 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9887 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9888 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9891 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9893 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9895 #else /* !CONFG_FAIR_GROUP_SCHED */
9896 static inline void free_fair_sched_group(struct task_group
*tg
)
9901 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9906 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9910 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9913 #endif /* CONFIG_FAIR_GROUP_SCHED */
9915 #ifdef CONFIG_RT_GROUP_SCHED
9916 static void free_rt_sched_group(struct task_group
*tg
)
9920 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9922 for_each_possible_cpu(i
) {
9924 kfree(tg
->rt_rq
[i
]);
9926 kfree(tg
->rt_se
[i
]);
9934 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9936 struct rt_rq
*rt_rq
;
9937 struct sched_rt_entity
*rt_se
;
9941 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9944 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9948 init_rt_bandwidth(&tg
->rt_bandwidth
,
9949 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9951 for_each_possible_cpu(i
) {
9954 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9955 GFP_KERNEL
, cpu_to_node(i
));
9959 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9960 GFP_KERNEL
, cpu_to_node(i
));
9964 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9973 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9975 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9976 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9979 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9981 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9983 #else /* !CONFIG_RT_GROUP_SCHED */
9984 static inline void free_rt_sched_group(struct task_group
*tg
)
9989 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9994 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9998 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10001 #endif /* CONFIG_RT_GROUP_SCHED */
10003 #ifdef CONFIG_GROUP_SCHED
10004 static void free_sched_group(struct task_group
*tg
)
10006 free_fair_sched_group(tg
);
10007 free_rt_sched_group(tg
);
10011 /* allocate runqueue etc for a new task group */
10012 struct task_group
*sched_create_group(struct task_group
*parent
)
10014 struct task_group
*tg
;
10015 unsigned long flags
;
10018 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10020 return ERR_PTR(-ENOMEM
);
10022 if (!alloc_fair_sched_group(tg
, parent
))
10025 if (!alloc_rt_sched_group(tg
, parent
))
10028 spin_lock_irqsave(&task_group_lock
, flags
);
10029 for_each_possible_cpu(i
) {
10030 register_fair_sched_group(tg
, i
);
10031 register_rt_sched_group(tg
, i
);
10033 list_add_rcu(&tg
->list
, &task_groups
);
10035 WARN_ON(!parent
); /* root should already exist */
10037 tg
->parent
= parent
;
10038 INIT_LIST_HEAD(&tg
->children
);
10039 list_add_rcu(&tg
->siblings
, &parent
->children
);
10040 spin_unlock_irqrestore(&task_group_lock
, flags
);
10045 free_sched_group(tg
);
10046 return ERR_PTR(-ENOMEM
);
10049 /* rcu callback to free various structures associated with a task group */
10050 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10052 /* now it should be safe to free those cfs_rqs */
10053 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10056 /* Destroy runqueue etc associated with a task group */
10057 void sched_destroy_group(struct task_group
*tg
)
10059 unsigned long flags
;
10062 spin_lock_irqsave(&task_group_lock
, flags
);
10063 for_each_possible_cpu(i
) {
10064 unregister_fair_sched_group(tg
, i
);
10065 unregister_rt_sched_group(tg
, i
);
10067 list_del_rcu(&tg
->list
);
10068 list_del_rcu(&tg
->siblings
);
10069 spin_unlock_irqrestore(&task_group_lock
, flags
);
10071 /* wait for possible concurrent references to cfs_rqs complete */
10072 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10075 /* change task's runqueue when it moves between groups.
10076 * The caller of this function should have put the task in its new group
10077 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10078 * reflect its new group.
10080 void sched_move_task(struct task_struct
*tsk
)
10082 int on_rq
, running
;
10083 unsigned long flags
;
10086 rq
= task_rq_lock(tsk
, &flags
);
10088 update_rq_clock(rq
);
10090 running
= task_current(rq
, tsk
);
10091 on_rq
= tsk
->se
.on_rq
;
10094 dequeue_task(rq
, tsk
, 0);
10095 if (unlikely(running
))
10096 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10098 set_task_rq(tsk
, task_cpu(tsk
));
10100 #ifdef CONFIG_FAIR_GROUP_SCHED
10101 if (tsk
->sched_class
->moved_group
)
10102 tsk
->sched_class
->moved_group(tsk
);
10105 if (unlikely(running
))
10106 tsk
->sched_class
->set_curr_task(rq
);
10108 enqueue_task(rq
, tsk
, 0);
10110 task_rq_unlock(rq
, &flags
);
10112 #endif /* CONFIG_GROUP_SCHED */
10114 #ifdef CONFIG_FAIR_GROUP_SCHED
10115 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10117 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10122 dequeue_entity(cfs_rq
, se
, 0);
10124 se
->load
.weight
= shares
;
10125 se
->load
.inv_weight
= 0;
10128 enqueue_entity(cfs_rq
, se
, 0);
10131 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10133 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10134 struct rq
*rq
= cfs_rq
->rq
;
10135 unsigned long flags
;
10137 spin_lock_irqsave(&rq
->lock
, flags
);
10138 __set_se_shares(se
, shares
);
10139 spin_unlock_irqrestore(&rq
->lock
, flags
);
10142 static DEFINE_MUTEX(shares_mutex
);
10144 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10147 unsigned long flags
;
10150 * We can't change the weight of the root cgroup.
10155 if (shares
< MIN_SHARES
)
10156 shares
= MIN_SHARES
;
10157 else if (shares
> MAX_SHARES
)
10158 shares
= MAX_SHARES
;
10160 mutex_lock(&shares_mutex
);
10161 if (tg
->shares
== shares
)
10164 spin_lock_irqsave(&task_group_lock
, flags
);
10165 for_each_possible_cpu(i
)
10166 unregister_fair_sched_group(tg
, i
);
10167 list_del_rcu(&tg
->siblings
);
10168 spin_unlock_irqrestore(&task_group_lock
, flags
);
10170 /* wait for any ongoing reference to this group to finish */
10171 synchronize_sched();
10174 * Now we are free to modify the group's share on each cpu
10175 * w/o tripping rebalance_share or load_balance_fair.
10177 tg
->shares
= shares
;
10178 for_each_possible_cpu(i
) {
10180 * force a rebalance
10182 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10183 set_se_shares(tg
->se
[i
], shares
);
10187 * Enable load balance activity on this group, by inserting it back on
10188 * each cpu's rq->leaf_cfs_rq_list.
10190 spin_lock_irqsave(&task_group_lock
, flags
);
10191 for_each_possible_cpu(i
)
10192 register_fair_sched_group(tg
, i
);
10193 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10194 spin_unlock_irqrestore(&task_group_lock
, flags
);
10196 mutex_unlock(&shares_mutex
);
10200 unsigned long sched_group_shares(struct task_group
*tg
)
10206 #ifdef CONFIG_RT_GROUP_SCHED
10208 * Ensure that the real time constraints are schedulable.
10210 static DEFINE_MUTEX(rt_constraints_mutex
);
10212 static unsigned long to_ratio(u64 period
, u64 runtime
)
10214 if (runtime
== RUNTIME_INF
)
10217 return div64_u64(runtime
<< 20, period
);
10220 /* Must be called with tasklist_lock held */
10221 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10223 struct task_struct
*g
, *p
;
10225 do_each_thread(g
, p
) {
10226 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10228 } while_each_thread(g
, p
);
10233 struct rt_schedulable_data
{
10234 struct task_group
*tg
;
10239 static int tg_schedulable(struct task_group
*tg
, void *data
)
10241 struct rt_schedulable_data
*d
= data
;
10242 struct task_group
*child
;
10243 unsigned long total
, sum
= 0;
10244 u64 period
, runtime
;
10246 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10247 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10250 period
= d
->rt_period
;
10251 runtime
= d
->rt_runtime
;
10254 #ifdef CONFIG_USER_SCHED
10255 if (tg
== &root_task_group
) {
10256 period
= global_rt_period();
10257 runtime
= global_rt_runtime();
10262 * Cannot have more runtime than the period.
10264 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10268 * Ensure we don't starve existing RT tasks.
10270 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10273 total
= to_ratio(period
, runtime
);
10276 * Nobody can have more than the global setting allows.
10278 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10282 * The sum of our children's runtime should not exceed our own.
10284 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10285 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10286 runtime
= child
->rt_bandwidth
.rt_runtime
;
10288 if (child
== d
->tg
) {
10289 period
= d
->rt_period
;
10290 runtime
= d
->rt_runtime
;
10293 sum
+= to_ratio(period
, runtime
);
10302 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10304 struct rt_schedulable_data data
= {
10306 .rt_period
= period
,
10307 .rt_runtime
= runtime
,
10310 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10313 static int tg_set_bandwidth(struct task_group
*tg
,
10314 u64 rt_period
, u64 rt_runtime
)
10318 mutex_lock(&rt_constraints_mutex
);
10319 read_lock(&tasklist_lock
);
10320 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10324 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10325 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10326 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10328 for_each_possible_cpu(i
) {
10329 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10331 spin_lock(&rt_rq
->rt_runtime_lock
);
10332 rt_rq
->rt_runtime
= rt_runtime
;
10333 spin_unlock(&rt_rq
->rt_runtime_lock
);
10335 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10337 read_unlock(&tasklist_lock
);
10338 mutex_unlock(&rt_constraints_mutex
);
10343 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10345 u64 rt_runtime
, rt_period
;
10347 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10348 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10349 if (rt_runtime_us
< 0)
10350 rt_runtime
= RUNTIME_INF
;
10352 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10355 long sched_group_rt_runtime(struct task_group
*tg
)
10359 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10362 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10363 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10364 return rt_runtime_us
;
10367 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10369 u64 rt_runtime
, rt_period
;
10371 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10372 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10374 if (rt_period
== 0)
10377 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10380 long sched_group_rt_period(struct task_group
*tg
)
10384 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10385 do_div(rt_period_us
, NSEC_PER_USEC
);
10386 return rt_period_us
;
10389 static int sched_rt_global_constraints(void)
10391 u64 runtime
, period
;
10394 if (sysctl_sched_rt_period
<= 0)
10397 runtime
= global_rt_runtime();
10398 period
= global_rt_period();
10401 * Sanity check on the sysctl variables.
10403 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10406 mutex_lock(&rt_constraints_mutex
);
10407 read_lock(&tasklist_lock
);
10408 ret
= __rt_schedulable(NULL
, 0, 0);
10409 read_unlock(&tasklist_lock
);
10410 mutex_unlock(&rt_constraints_mutex
);
10415 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10417 /* Don't accept realtime tasks when there is no way for them to run */
10418 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10424 #else /* !CONFIG_RT_GROUP_SCHED */
10425 static int sched_rt_global_constraints(void)
10427 unsigned long flags
;
10430 if (sysctl_sched_rt_period
<= 0)
10434 * There's always some RT tasks in the root group
10435 * -- migration, kstopmachine etc..
10437 if (sysctl_sched_rt_runtime
== 0)
10440 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10441 for_each_possible_cpu(i
) {
10442 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10444 spin_lock(&rt_rq
->rt_runtime_lock
);
10445 rt_rq
->rt_runtime
= global_rt_runtime();
10446 spin_unlock(&rt_rq
->rt_runtime_lock
);
10448 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10452 #endif /* CONFIG_RT_GROUP_SCHED */
10454 int sched_rt_handler(struct ctl_table
*table
, int write
,
10455 void __user
*buffer
, size_t *lenp
,
10459 int old_period
, old_runtime
;
10460 static DEFINE_MUTEX(mutex
);
10462 mutex_lock(&mutex
);
10463 old_period
= sysctl_sched_rt_period
;
10464 old_runtime
= sysctl_sched_rt_runtime
;
10466 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10468 if (!ret
&& write
) {
10469 ret
= sched_rt_global_constraints();
10471 sysctl_sched_rt_period
= old_period
;
10472 sysctl_sched_rt_runtime
= old_runtime
;
10474 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10475 def_rt_bandwidth
.rt_period
=
10476 ns_to_ktime(global_rt_period());
10479 mutex_unlock(&mutex
);
10484 #ifdef CONFIG_CGROUP_SCHED
10486 /* return corresponding task_group object of a cgroup */
10487 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10489 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10490 struct task_group
, css
);
10493 static struct cgroup_subsys_state
*
10494 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10496 struct task_group
*tg
, *parent
;
10498 if (!cgrp
->parent
) {
10499 /* This is early initialization for the top cgroup */
10500 return &init_task_group
.css
;
10503 parent
= cgroup_tg(cgrp
->parent
);
10504 tg
= sched_create_group(parent
);
10506 return ERR_PTR(-ENOMEM
);
10512 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10514 struct task_group
*tg
= cgroup_tg(cgrp
);
10516 sched_destroy_group(tg
);
10520 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10522 #ifdef CONFIG_RT_GROUP_SCHED
10523 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10526 /* We don't support RT-tasks being in separate groups */
10527 if (tsk
->sched_class
!= &fair_sched_class
)
10534 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10535 struct task_struct
*tsk
, bool threadgroup
)
10537 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10541 struct task_struct
*c
;
10543 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10544 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10556 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10557 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10560 sched_move_task(tsk
);
10562 struct task_struct
*c
;
10564 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10565 sched_move_task(c
);
10571 #ifdef CONFIG_FAIR_GROUP_SCHED
10572 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10575 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10578 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10580 struct task_group
*tg
= cgroup_tg(cgrp
);
10582 return (u64
) tg
->shares
;
10584 #endif /* CONFIG_FAIR_GROUP_SCHED */
10586 #ifdef CONFIG_RT_GROUP_SCHED
10587 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10590 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10593 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10595 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10598 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10601 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10604 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10606 return sched_group_rt_period(cgroup_tg(cgrp
));
10608 #endif /* CONFIG_RT_GROUP_SCHED */
10610 static struct cftype cpu_files
[] = {
10611 #ifdef CONFIG_FAIR_GROUP_SCHED
10614 .read_u64
= cpu_shares_read_u64
,
10615 .write_u64
= cpu_shares_write_u64
,
10618 #ifdef CONFIG_RT_GROUP_SCHED
10620 .name
= "rt_runtime_us",
10621 .read_s64
= cpu_rt_runtime_read
,
10622 .write_s64
= cpu_rt_runtime_write
,
10625 .name
= "rt_period_us",
10626 .read_u64
= cpu_rt_period_read_uint
,
10627 .write_u64
= cpu_rt_period_write_uint
,
10632 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10634 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10637 struct cgroup_subsys cpu_cgroup_subsys
= {
10639 .create
= cpu_cgroup_create
,
10640 .destroy
= cpu_cgroup_destroy
,
10641 .can_attach
= cpu_cgroup_can_attach
,
10642 .attach
= cpu_cgroup_attach
,
10643 .populate
= cpu_cgroup_populate
,
10644 .subsys_id
= cpu_cgroup_subsys_id
,
10648 #endif /* CONFIG_CGROUP_SCHED */
10650 #ifdef CONFIG_CGROUP_CPUACCT
10653 * CPU accounting code for task groups.
10655 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10656 * (balbir@in.ibm.com).
10659 /* track cpu usage of a group of tasks and its child groups */
10661 struct cgroup_subsys_state css
;
10662 /* cpuusage holds pointer to a u64-type object on every cpu */
10664 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10665 struct cpuacct
*parent
;
10668 struct cgroup_subsys cpuacct_subsys
;
10670 /* return cpu accounting group corresponding to this container */
10671 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10673 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10674 struct cpuacct
, css
);
10677 /* return cpu accounting group to which this task belongs */
10678 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10680 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10681 struct cpuacct
, css
);
10684 /* create a new cpu accounting group */
10685 static struct cgroup_subsys_state
*cpuacct_create(
10686 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10688 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10694 ca
->cpuusage
= alloc_percpu(u64
);
10698 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10699 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10700 goto out_free_counters
;
10703 ca
->parent
= cgroup_ca(cgrp
->parent
);
10709 percpu_counter_destroy(&ca
->cpustat
[i
]);
10710 free_percpu(ca
->cpuusage
);
10714 return ERR_PTR(-ENOMEM
);
10717 /* destroy an existing cpu accounting group */
10719 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10721 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10724 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10725 percpu_counter_destroy(&ca
->cpustat
[i
]);
10726 free_percpu(ca
->cpuusage
);
10730 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10732 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10735 #ifndef CONFIG_64BIT
10737 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10739 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10741 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10749 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10751 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10753 #ifndef CONFIG_64BIT
10755 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10757 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10759 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10765 /* return total cpu usage (in nanoseconds) of a group */
10766 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10768 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10769 u64 totalcpuusage
= 0;
10772 for_each_present_cpu(i
)
10773 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10775 return totalcpuusage
;
10778 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10781 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10790 for_each_present_cpu(i
)
10791 cpuacct_cpuusage_write(ca
, i
, 0);
10797 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10798 struct seq_file
*m
)
10800 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10804 for_each_present_cpu(i
) {
10805 percpu
= cpuacct_cpuusage_read(ca
, i
);
10806 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10808 seq_printf(m
, "\n");
10812 static const char *cpuacct_stat_desc
[] = {
10813 [CPUACCT_STAT_USER
] = "user",
10814 [CPUACCT_STAT_SYSTEM
] = "system",
10817 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10818 struct cgroup_map_cb
*cb
)
10820 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10823 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10824 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10825 val
= cputime64_to_clock_t(val
);
10826 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10831 static struct cftype files
[] = {
10834 .read_u64
= cpuusage_read
,
10835 .write_u64
= cpuusage_write
,
10838 .name
= "usage_percpu",
10839 .read_seq_string
= cpuacct_percpu_seq_read
,
10843 .read_map
= cpuacct_stats_show
,
10847 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10849 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10853 * charge this task's execution time to its accounting group.
10855 * called with rq->lock held.
10857 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10859 struct cpuacct
*ca
;
10862 if (unlikely(!cpuacct_subsys
.active
))
10865 cpu
= task_cpu(tsk
);
10871 for (; ca
; ca
= ca
->parent
) {
10872 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10873 *cpuusage
+= cputime
;
10880 * Charge the system/user time to the task's accounting group.
10882 static void cpuacct_update_stats(struct task_struct
*tsk
,
10883 enum cpuacct_stat_index idx
, cputime_t val
)
10885 struct cpuacct
*ca
;
10887 if (unlikely(!cpuacct_subsys
.active
))
10894 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10900 struct cgroup_subsys cpuacct_subsys
= {
10902 .create
= cpuacct_create
,
10903 .destroy
= cpuacct_destroy
,
10904 .populate
= cpuacct_populate
,
10905 .subsys_id
= cpuacct_subsys_id
,
10907 #endif /* CONFIG_CGROUP_CPUACCT */
10911 int rcu_expedited_torture_stats(char *page
)
10915 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10917 void synchronize_sched_expedited(void)
10920 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10922 #else /* #ifndef CONFIG_SMP */
10924 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10925 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10927 #define RCU_EXPEDITED_STATE_POST -2
10928 #define RCU_EXPEDITED_STATE_IDLE -1
10930 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10932 int rcu_expedited_torture_stats(char *page
)
10937 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10938 for_each_online_cpu(cpu
) {
10939 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10940 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10942 cnt
+= sprintf(&page
[cnt
], "\n");
10945 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10947 static long synchronize_sched_expedited_count
;
10950 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10951 * approach to force grace period to end quickly. This consumes
10952 * significant time on all CPUs, and is thus not recommended for
10953 * any sort of common-case code.
10955 * Note that it is illegal to call this function while holding any
10956 * lock that is acquired by a CPU-hotplug notifier. Failing to
10957 * observe this restriction will result in deadlock.
10959 void synchronize_sched_expedited(void)
10962 unsigned long flags
;
10963 bool need_full_sync
= 0;
10965 struct migration_req
*req
;
10969 smp_mb(); /* ensure prior mod happens before capturing snap. */
10970 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10972 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10974 if (trycount
++ < 10)
10975 udelay(trycount
* num_online_cpus());
10977 synchronize_sched();
10980 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10981 smp_mb(); /* ensure test happens before caller kfree */
10986 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10987 for_each_online_cpu(cpu
) {
10989 req
= &per_cpu(rcu_migration_req
, cpu
);
10990 init_completion(&req
->done
);
10992 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10993 spin_lock_irqsave(&rq
->lock
, flags
);
10994 list_add(&req
->list
, &rq
->migration_queue
);
10995 spin_unlock_irqrestore(&rq
->lock
, flags
);
10996 wake_up_process(rq
->migration_thread
);
10998 for_each_online_cpu(cpu
) {
10999 rcu_expedited_state
= cpu
;
11000 req
= &per_cpu(rcu_migration_req
, cpu
);
11002 wait_for_completion(&req
->done
);
11003 spin_lock_irqsave(&rq
->lock
, flags
);
11004 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
11005 need_full_sync
= 1;
11006 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
11007 spin_unlock_irqrestore(&rq
->lock
, flags
);
11009 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
11010 mutex_unlock(&rcu_sched_expedited_mutex
);
11012 if (need_full_sync
)
11013 synchronize_sched();
11015 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
11017 #endif /* #else #ifndef CONFIG_SMP */