4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css
;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity
**se
;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq
**cfs_rq
;
259 unsigned long shares
;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity
**rt_se
;
264 struct rt_rq
**rt_rq
;
266 struct rt_bandwidth rt_bandwidth
;
270 struct list_head list
;
272 struct task_group
*parent
;
273 struct list_head siblings
;
274 struct list_head children
;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct
*user
)
282 user
->tg
->uid
= user
->uid
;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group
;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq
, init_tg_cfs_rq
);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq
, init_rt_rq
);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock
);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group
.children
);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group
;
345 /* return group to which a task belongs */
346 static inline struct task_group
*task_group(struct task_struct
*p
)
348 struct task_group
*tg
;
350 #ifdef CONFIG_USER_SCHED
352 tg
= __task_cred(p
)->user
->tg
;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
356 struct task_group
, css
);
358 tg
= &init_task_group
;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
368 p
->se
.parent
= task_group(p
)->se
[cpu
];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
373 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
379 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
380 static inline struct task_group
*task_group(struct task_struct
*p
)
385 #endif /* CONFIG_GROUP_SCHED */
387 /* CFS-related fields in a runqueue */
389 struct load_weight load
;
390 unsigned long nr_running
;
395 struct rb_root tasks_timeline
;
396 struct rb_node
*rb_leftmost
;
398 struct list_head tasks
;
399 struct list_head
*balance_iterator
;
402 * 'curr' points to currently running entity on this cfs_rq.
403 * It is set to NULL otherwise (i.e when none are currently running).
405 struct sched_entity
*curr
, *next
, *last
;
407 unsigned int nr_spread_over
;
409 #ifdef CONFIG_FAIR_GROUP_SCHED
410 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
420 struct list_head leaf_cfs_rq_list
;
421 struct task_group
*tg
; /* group that "owns" this runqueue */
425 * the part of load.weight contributed by tasks
427 unsigned long task_weight
;
430 * h_load = weight * f(tg)
432 * Where f(tg) is the recursive weight fraction assigned to
435 unsigned long h_load
;
438 * this cpu's part of tg->shares
440 unsigned long shares
;
443 * load.weight at the time we set shares
445 unsigned long rq_weight
;
450 /* Real-Time classes' related field in a runqueue: */
452 struct rt_prio_array active
;
453 unsigned long rt_nr_running
;
454 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int curr
; /* highest queued rt task prio */
458 int next
; /* next highest */
463 unsigned long rt_nr_migratory
;
464 unsigned long rt_nr_total
;
466 struct plist_head pushable_tasks
;
471 /* Nests inside the rq lock: */
472 spinlock_t rt_runtime_lock
;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 unsigned long rt_nr_boosted
;
478 struct list_head leaf_rt_rq_list
;
479 struct task_group
*tg
;
480 struct sched_rt_entity
*rt_se
;
487 * We add the notion of a root-domain which will be used to define per-domain
488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
490 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online
;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask
;
506 struct cpupri cpupri
;
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
514 static struct root_domain def_root_domain
;
519 * This is the main, per-CPU runqueue data structure.
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
533 unsigned long nr_running
;
534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
537 unsigned long last_tick_seen
;
538 unsigned char in_nohz_recently
;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load
;
542 unsigned long nr_load_updates
;
544 u64 nr_migrations_in
;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list
;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list
;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible
;
565 struct task_struct
*curr
, *idle
;
566 unsigned long next_balance
;
567 struct mm_struct
*prev_mm
;
574 struct root_domain
*rd
;
575 struct sched_domain
*sd
;
577 unsigned char idle_at_tick
;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task
;
588 struct task_struct
*migration_thread
;
589 struct list_head migration_queue
;
595 /* calc_load related fields */
596 unsigned long calc_load_update
;
597 long calc_load_active
;
599 #ifdef CONFIG_SCHED_HRTICK
601 int hrtick_csd_pending
;
602 struct call_single_data hrtick_csd
;
604 struct hrtimer hrtick_timer
;
607 #ifdef CONFIG_SCHEDSTATS
609 struct sched_info rq_sched_info
;
610 unsigned long long rq_cpu_time
;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
613 /* sys_sched_yield() stats */
614 unsigned int yld_count
;
616 /* schedule() stats */
617 unsigned int sched_switch
;
618 unsigned int sched_count
;
619 unsigned int sched_goidle
;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count
;
623 unsigned int ttwu_local
;
626 unsigned int bkl_count
;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
633 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
635 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
638 static inline int cpu_of(struct rq
*rq
)
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 #define raw_rq() (&__raw_get_cpu_var(runqueues))
663 inline void update_rq_clock(struct rq
*rq
)
665 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 #ifdef CONFIG_SCHED_DEBUG
672 # define const_debug __read_mostly
674 # define const_debug static const
679 * @cpu: the processor in question.
681 * Returns true if the current cpu runqueue is locked.
682 * This interface allows printk to be called with the runqueue lock
683 * held and know whether or not it is OK to wake up the klogd.
685 int runqueue_is_locked(int cpu
)
687 return spin_is_locked(&cpu_rq(cpu
)->lock
);
691 * Debugging: various feature bits
694 #define SCHED_FEAT(name, enabled) \
695 __SCHED_FEAT_##name ,
698 #include "sched_features.h"
703 #define SCHED_FEAT(name, enabled) \
704 (1UL << __SCHED_FEAT_##name) * enabled |
706 const_debug
unsigned int sysctl_sched_features
=
707 #include "sched_features.h"
712 #ifdef CONFIG_SCHED_DEBUG
713 #define SCHED_FEAT(name, enabled) \
716 static __read_mostly
char *sched_feat_names
[] = {
717 #include "sched_features.h"
723 static int sched_feat_show(struct seq_file
*m
, void *v
)
727 for (i
= 0; sched_feat_names
[i
]; i
++) {
728 if (!(sysctl_sched_features
& (1UL << i
)))
730 seq_printf(m
, "%s ", sched_feat_names
[i
]);
738 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
739 size_t cnt
, loff_t
*ppos
)
749 if (copy_from_user(&buf
, ubuf
, cnt
))
754 if (strncmp(buf
, "NO_", 3) == 0) {
759 for (i
= 0; sched_feat_names
[i
]; i
++) {
760 int len
= strlen(sched_feat_names
[i
]);
762 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
764 sysctl_sched_features
&= ~(1UL << i
);
766 sysctl_sched_features
|= (1UL << i
);
771 if (!sched_feat_names
[i
])
779 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
781 return single_open(filp
, sched_feat_show
, NULL
);
784 static const struct file_operations sched_feat_fops
= {
785 .open
= sched_feat_open
,
786 .write
= sched_feat_write
,
789 .release
= single_release
,
792 static __init
int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
799 late_initcall(sched_init_debug
);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
812 * ratelimit for updating the group shares.
815 unsigned int sysctl_sched_shares_ratelimit
= 250000;
818 * Inject some fuzzyness into changing the per-cpu group shares
819 * this avoids remote rq-locks at the expense of fairness.
822 unsigned int sysctl_sched_shares_thresh
= 4;
825 * period over which we average the RT time consumption, measured
830 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
833 * period over which we measure -rt task cpu usage in us.
836 unsigned int sysctl_sched_rt_period
= 1000000;
838 static __read_mostly
int scheduler_running
;
841 * part of the period that we allow rt tasks to run in us.
844 int sysctl_sched_rt_runtime
= 950000;
846 static inline u64
global_rt_period(void)
848 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
851 static inline u64
global_rt_runtime(void)
853 if (sysctl_sched_rt_runtime
< 0)
856 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
859 #ifndef prepare_arch_switch
860 # define prepare_arch_switch(next) do { } while (0)
862 #ifndef finish_arch_switch
863 # define finish_arch_switch(prev) do { } while (0)
866 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
868 return rq
->curr
== p
;
871 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
872 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
874 return task_current(rq
, p
);
877 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
881 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
883 #ifdef CONFIG_DEBUG_SPINLOCK
884 /* this is a valid case when another task releases the spinlock */
885 rq
->lock
.owner
= current
;
888 * If we are tracking spinlock dependencies then we have to
889 * fix up the runqueue lock - which gets 'carried over' from
892 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
894 spin_unlock_irq(&rq
->lock
);
897 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
898 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
903 return task_current(rq
, p
);
907 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
911 * We can optimise this out completely for !SMP, because the
912 * SMP rebalancing from interrupt is the only thing that cares
917 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
918 spin_unlock_irq(&rq
->lock
);
920 spin_unlock(&rq
->lock
);
924 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
928 * After ->oncpu is cleared, the task can be moved to a different CPU.
929 * We must ensure this doesn't happen until the switch is completely
935 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
939 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
942 * __task_rq_lock - lock the runqueue a given task resides on.
943 * Must be called interrupts disabled.
945 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
949 struct rq
*rq
= task_rq(p
);
950 spin_lock(&rq
->lock
);
951 if (likely(rq
== task_rq(p
)))
953 spin_unlock(&rq
->lock
);
958 * task_rq_lock - lock the runqueue a given task resides on and disable
959 * interrupts. Note the ordering: we can safely lookup the task_rq without
960 * explicitly disabling preemption.
962 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
968 local_irq_save(*flags
);
970 spin_lock(&rq
->lock
);
971 if (likely(rq
== task_rq(p
)))
973 spin_unlock_irqrestore(&rq
->lock
, *flags
);
977 void task_rq_unlock_wait(struct task_struct
*p
)
979 struct rq
*rq
= task_rq(p
);
981 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
982 spin_unlock_wait(&rq
->lock
);
985 static void __task_rq_unlock(struct rq
*rq
)
988 spin_unlock(&rq
->lock
);
991 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
994 spin_unlock_irqrestore(&rq
->lock
, *flags
);
998 * this_rq_lock - lock this runqueue and disable interrupts.
1000 static struct rq
*this_rq_lock(void)
1001 __acquires(rq
->lock
)
1005 local_irq_disable();
1007 spin_lock(&rq
->lock
);
1012 #ifdef CONFIG_SCHED_HRTICK
1014 * Use HR-timers to deliver accurate preemption points.
1016 * Its all a bit involved since we cannot program an hrt while holding the
1017 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * When we get rescheduled we reprogram the hrtick_timer outside of the
1026 * - enabled by features
1027 * - hrtimer is actually high res
1029 static inline int hrtick_enabled(struct rq
*rq
)
1031 if (!sched_feat(HRTICK
))
1033 if (!cpu_active(cpu_of(rq
)))
1035 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1038 static void hrtick_clear(struct rq
*rq
)
1040 if (hrtimer_active(&rq
->hrtick_timer
))
1041 hrtimer_cancel(&rq
->hrtick_timer
);
1045 * High-resolution timer tick.
1046 * Runs from hardirq context with interrupts disabled.
1048 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1050 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1052 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1054 spin_lock(&rq
->lock
);
1055 update_rq_clock(rq
);
1056 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1057 spin_unlock(&rq
->lock
);
1059 return HRTIMER_NORESTART
;
1064 * called from hardirq (IPI) context
1066 static void __hrtick_start(void *arg
)
1068 struct rq
*rq
= arg
;
1070 spin_lock(&rq
->lock
);
1071 hrtimer_restart(&rq
->hrtick_timer
);
1072 rq
->hrtick_csd_pending
= 0;
1073 spin_unlock(&rq
->lock
);
1077 * Called to set the hrtick timer state.
1079 * called with rq->lock held and irqs disabled
1081 static void hrtick_start(struct rq
*rq
, u64 delay
)
1083 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1084 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1086 hrtimer_set_expires(timer
, time
);
1088 if (rq
== this_rq()) {
1089 hrtimer_restart(timer
);
1090 } else if (!rq
->hrtick_csd_pending
) {
1091 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1092 rq
->hrtick_csd_pending
= 1;
1097 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1099 int cpu
= (int)(long)hcpu
;
1102 case CPU_UP_CANCELED
:
1103 case CPU_UP_CANCELED_FROZEN
:
1104 case CPU_DOWN_PREPARE
:
1105 case CPU_DOWN_PREPARE_FROZEN
:
1107 case CPU_DEAD_FROZEN
:
1108 hrtick_clear(cpu_rq(cpu
));
1115 static __init
void init_hrtick(void)
1117 hotcpu_notifier(hotplug_hrtick
, 0);
1121 * Called to set the hrtick timer state.
1123 * called with rq->lock held and irqs disabled
1125 static void hrtick_start(struct rq
*rq
, u64 delay
)
1127 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1128 HRTIMER_MODE_REL_PINNED
, 0);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq
*rq
)
1139 rq
->hrtick_csd_pending
= 0;
1141 rq
->hrtick_csd
.flags
= 0;
1142 rq
->hrtick_csd
.func
= __hrtick_start
;
1143 rq
->hrtick_csd
.info
= rq
;
1146 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1147 rq
->hrtick_timer
.function
= hrtick
;
1149 #else /* CONFIG_SCHED_HRTICK */
1150 static inline void hrtick_clear(struct rq
*rq
)
1154 static inline void init_rq_hrtick(struct rq
*rq
)
1158 static inline void init_hrtick(void)
1161 #endif /* CONFIG_SCHED_HRTICK */
1164 * resched_task - mark a task 'to be rescheduled now'.
1166 * On UP this means the setting of the need_resched flag, on SMP it
1167 * might also involve a cross-CPU call to trigger the scheduler on
1172 #ifndef tsk_is_polling
1173 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176 static void resched_task(struct task_struct
*p
)
1180 assert_spin_locked(&task_rq(p
)->lock
);
1182 if (test_tsk_need_resched(p
))
1185 set_tsk_need_resched(p
);
1188 if (cpu
== smp_processor_id())
1191 /* NEED_RESCHED must be visible before we test polling */
1193 if (!tsk_is_polling(p
))
1194 smp_send_reschedule(cpu
);
1197 static void resched_cpu(int cpu
)
1199 struct rq
*rq
= cpu_rq(cpu
);
1200 unsigned long flags
;
1202 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1204 resched_task(cpu_curr(cpu
));
1205 spin_unlock_irqrestore(&rq
->lock
, flags
);
1210 * When add_timer_on() enqueues a timer into the timer wheel of an
1211 * idle CPU then this timer might expire before the next timer event
1212 * which is scheduled to wake up that CPU. In case of a completely
1213 * idle system the next event might even be infinite time into the
1214 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1215 * leaves the inner idle loop so the newly added timer is taken into
1216 * account when the CPU goes back to idle and evaluates the timer
1217 * wheel for the next timer event.
1219 void wake_up_idle_cpu(int cpu
)
1221 struct rq
*rq
= cpu_rq(cpu
);
1223 if (cpu
== smp_processor_id())
1227 * This is safe, as this function is called with the timer
1228 * wheel base lock of (cpu) held. When the CPU is on the way
1229 * to idle and has not yet set rq->curr to idle then it will
1230 * be serialized on the timer wheel base lock and take the new
1231 * timer into account automatically.
1233 if (rq
->curr
!= rq
->idle
)
1237 * We can set TIF_RESCHED on the idle task of the other CPU
1238 * lockless. The worst case is that the other CPU runs the
1239 * idle task through an additional NOOP schedule()
1241 set_tsk_need_resched(rq
->idle
);
1243 /* NEED_RESCHED must be visible before we test polling */
1245 if (!tsk_is_polling(rq
->idle
))
1246 smp_send_reschedule(cpu
);
1248 #endif /* CONFIG_NO_HZ */
1250 static u64
sched_avg_period(void)
1252 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1255 static void sched_avg_update(struct rq
*rq
)
1257 s64 period
= sched_avg_period();
1259 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1260 rq
->age_stamp
+= period
;
1265 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1267 rq
->rt_avg
+= rt_delta
;
1268 sched_avg_update(rq
);
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct
*p
)
1274 assert_spin_locked(&task_rq(p
)->lock
);
1275 set_tsk_need_resched(p
);
1278 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1301 struct load_weight
*lw
)
1305 if (!lw
->inv_weight
) {
1306 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1309 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1313 tmp
= (u64
)delta_exec
* weight
;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp
> WMULT_CONST
))
1318 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1321 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1323 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1326 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1332 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight
[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult
[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator
{
1400 struct task_struct
*(*start
)(void *);
1401 struct task_struct
*(*next
)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1407 unsigned long max_load_move
, struct sched_domain
*sd
,
1408 enum cpu_idle_type idle
, int *all_pinned
,
1409 int *this_best_prio
, struct rq_iterator
*iterator
);
1412 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1413 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1414 struct rq_iterator
*iterator
);
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index
{
1419 CPUACCT_STAT_USER
, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS
,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1427 static void cpuacct_update_stats(struct task_struct
*tsk
,
1428 enum cpuacct_stat_index idx
, cputime_t val
);
1430 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1431 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1432 enum cpuacct_stat_index idx
, cputime_t val
) {}
1435 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1437 update_load_add(&rq
->load
, load
);
1440 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1442 update_load_sub(&rq
->load
, load
);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor
)(struct task_group
*, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1454 struct task_group
*parent
, *child
;
1458 parent
= &root_task_group
;
1460 ret
= (*down
)(parent
, data
);
1463 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1470 ret
= (*up
)(parent
, data
);
1475 parent
= parent
->parent
;
1484 static int tg_nop(struct task_group
*tg
, void *data
)
1491 /* Used instead of source_load when we know the type == 0 */
1492 static unsigned long weighted_cpuload(const int cpu
)
1494 return cpu_rq(cpu
)->load
.weight
;
1498 * Return a low guess at the load of a migration-source cpu weighted
1499 * according to the scheduling class and "nice" value.
1501 * We want to under-estimate the load of migration sources, to
1502 * balance conservatively.
1504 static unsigned long source_load(int cpu
, int type
)
1506 struct rq
*rq
= cpu_rq(cpu
);
1507 unsigned long total
= weighted_cpuload(cpu
);
1509 if (type
== 0 || !sched_feat(LB_BIAS
))
1512 return min(rq
->cpu_load
[type
-1], total
);
1516 * Return a high guess at the load of a migration-target cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 static unsigned long target_load(int cpu
, int type
)
1521 struct rq
*rq
= cpu_rq(cpu
);
1522 unsigned long total
= weighted_cpuload(cpu
);
1524 if (type
== 0 || !sched_feat(LB_BIAS
))
1527 return max(rq
->cpu_load
[type
-1], total
);
1530 static struct sched_group
*group_of(int cpu
)
1532 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1540 static unsigned long power_of(int cpu
)
1542 struct sched_group
*group
= group_of(cpu
);
1545 return SCHED_LOAD_SCALE
;
1547 return group
->cpu_power
;
1550 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1552 static unsigned long cpu_avg_load_per_task(int cpu
)
1554 struct rq
*rq
= cpu_rq(cpu
);
1555 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1558 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1560 rq
->avg_load_per_task
= 0;
1562 return rq
->avg_load_per_task
;
1565 #ifdef CONFIG_FAIR_GROUP_SCHED
1567 struct update_shares_data
{
1568 unsigned long rq_weight
[NR_CPUS
];
1571 static DEFINE_PER_CPU(struct update_shares_data
, update_shares_data
);
1573 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1579 unsigned long sd_shares
,
1580 unsigned long sd_rq_weight
,
1581 struct update_shares_data
*usd
)
1583 unsigned long shares
, rq_weight
;
1586 rq_weight
= usd
->rq_weight
[cpu
];
1589 rq_weight
= NICE_0_LOAD
;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1598 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1600 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1601 sysctl_sched_shares_thresh
) {
1602 struct rq
*rq
= cpu_rq(cpu
);
1603 unsigned long flags
;
1605 spin_lock_irqsave(&rq
->lock
, flags
);
1606 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1607 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1608 __set_se_shares(tg
->se
[cpu
], shares
);
1609 spin_unlock_irqrestore(&rq
->lock
, flags
);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group
*tg
, void *data
)
1620 unsigned long weight
, rq_weight
= 0, shares
= 0;
1621 struct update_shares_data
*usd
;
1622 struct sched_domain
*sd
= data
;
1623 unsigned long flags
;
1629 local_irq_save(flags
);
1630 usd
= &__get_cpu_var(update_shares_data
);
1632 for_each_cpu(i
, sched_domain_span(sd
)) {
1633 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1634 usd
->rq_weight
[i
] = weight
;
1637 * If there are currently no tasks on the cpu pretend there
1638 * is one of average load so that when a new task gets to
1639 * run here it will not get delayed by group starvation.
1642 weight
= NICE_0_LOAD
;
1644 rq_weight
+= weight
;
1645 shares
+= tg
->cfs_rq
[i
]->shares
;
1648 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1649 shares
= tg
->shares
;
1651 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1652 shares
= tg
->shares
;
1654 for_each_cpu(i
, sched_domain_span(sd
))
1655 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd
);
1657 local_irq_restore(flags
);
1663 * Compute the cpu's hierarchical load factor for each task group.
1664 * This needs to be done in a top-down fashion because the load of a child
1665 * group is a fraction of its parents load.
1667 static int tg_load_down(struct task_group
*tg
, void *data
)
1670 long cpu
= (long)data
;
1673 load
= cpu_rq(cpu
)->load
.weight
;
1675 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1676 load
*= tg
->cfs_rq
[cpu
]->shares
;
1677 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1680 tg
->cfs_rq
[cpu
]->h_load
= load
;
1685 static void update_shares(struct sched_domain
*sd
)
1690 if (root_task_group_empty())
1693 now
= cpu_clock(raw_smp_processor_id());
1694 elapsed
= now
- sd
->last_update
;
1696 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1697 sd
->last_update
= now
;
1698 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1702 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1704 if (root_task_group_empty())
1707 spin_unlock(&rq
->lock
);
1709 spin_lock(&rq
->lock
);
1712 static void update_h_load(long cpu
)
1714 if (root_task_group_empty())
1717 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1722 static inline void update_shares(struct sched_domain
*sd
)
1726 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1732 #ifdef CONFIG_PREEMPT
1734 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1737 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1738 * way at the expense of forcing extra atomic operations in all
1739 * invocations. This assures that the double_lock is acquired using the
1740 * same underlying policy as the spinlock_t on this architecture, which
1741 * reduces latency compared to the unfair variant below. However, it
1742 * also adds more overhead and therefore may reduce throughput.
1744 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1745 __releases(this_rq
->lock
)
1746 __acquires(busiest
->lock
)
1747 __acquires(this_rq
->lock
)
1749 spin_unlock(&this_rq
->lock
);
1750 double_rq_lock(this_rq
, busiest
);
1757 * Unfair double_lock_balance: Optimizes throughput at the expense of
1758 * latency by eliminating extra atomic operations when the locks are
1759 * already in proper order on entry. This favors lower cpu-ids and will
1760 * grant the double lock to lower cpus over higher ids under contention,
1761 * regardless of entry order into the function.
1763 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1764 __releases(this_rq
->lock
)
1765 __acquires(busiest
->lock
)
1766 __acquires(this_rq
->lock
)
1770 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1771 if (busiest
< this_rq
) {
1772 spin_unlock(&this_rq
->lock
);
1773 spin_lock(&busiest
->lock
);
1774 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1777 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1782 #endif /* CONFIG_PREEMPT */
1785 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1787 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1789 if (unlikely(!irqs_disabled())) {
1790 /* printk() doesn't work good under rq->lock */
1791 spin_unlock(&this_rq
->lock
);
1795 return _double_lock_balance(this_rq
, busiest
);
1798 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1799 __releases(busiest
->lock
)
1801 spin_unlock(&busiest
->lock
);
1802 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1810 cfs_rq
->shares
= shares
;
1815 static void calc_load_account_active(struct rq
*this_rq
);
1817 #include "sched_stats.h"
1818 #include "sched_idletask.c"
1819 #include "sched_fair.c"
1820 #include "sched_rt.c"
1821 #ifdef CONFIG_SCHED_DEBUG
1822 # include "sched_debug.c"
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 static void inc_nr_running(struct rq
*rq
)
1834 static void dec_nr_running(struct rq
*rq
)
1839 static void set_load_weight(struct task_struct
*p
)
1841 if (task_has_rt_policy(p
)) {
1842 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1843 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1848 * SCHED_IDLE tasks get minimal weight:
1850 if (p
->policy
== SCHED_IDLE
) {
1851 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1852 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1856 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1857 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1860 static void update_avg(u64
*avg
, u64 sample
)
1862 s64 diff
= sample
- *avg
;
1866 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1869 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1871 sched_info_queued(p
);
1872 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1876 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1879 if (p
->se
.last_wakeup
) {
1880 update_avg(&p
->se
.avg_overlap
,
1881 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1882 p
->se
.last_wakeup
= 0;
1884 update_avg(&p
->se
.avg_wakeup
,
1885 sysctl_sched_wakeup_granularity
);
1889 sched_info_dequeued(p
);
1890 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1895 * __normal_prio - return the priority that is based on the static prio
1897 static inline int __normal_prio(struct task_struct
*p
)
1899 return p
->static_prio
;
1903 * Calculate the expected normal priority: i.e. priority
1904 * without taking RT-inheritance into account. Might be
1905 * boosted by interactivity modifiers. Changes upon fork,
1906 * setprio syscalls, and whenever the interactivity
1907 * estimator recalculates.
1909 static inline int normal_prio(struct task_struct
*p
)
1913 if (task_has_rt_policy(p
))
1914 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1916 prio
= __normal_prio(p
);
1921 * Calculate the current priority, i.e. the priority
1922 * taken into account by the scheduler. This value might
1923 * be boosted by RT tasks, or might be boosted by
1924 * interactivity modifiers. Will be RT if the task got
1925 * RT-boosted. If not then it returns p->normal_prio.
1927 static int effective_prio(struct task_struct
*p
)
1929 p
->normal_prio
= normal_prio(p
);
1931 * If we are RT tasks or we were boosted to RT priority,
1932 * keep the priority unchanged. Otherwise, update priority
1933 * to the normal priority:
1935 if (!rt_prio(p
->prio
))
1936 return p
->normal_prio
;
1941 * activate_task - move a task to the runqueue.
1943 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1945 if (task_contributes_to_load(p
))
1946 rq
->nr_uninterruptible
--;
1948 enqueue_task(rq
, p
, wakeup
);
1953 * deactivate_task - remove a task from the runqueue.
1955 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1957 if (task_contributes_to_load(p
))
1958 rq
->nr_uninterruptible
++;
1960 dequeue_task(rq
, p
, sleep
);
1965 * task_curr - is this task currently executing on a CPU?
1966 * @p: the task in question.
1968 inline int task_curr(const struct task_struct
*p
)
1970 return cpu_curr(task_cpu(p
)) == p
;
1973 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1975 set_task_rq(p
, cpu
);
1978 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1979 * successfuly executed on another CPU. We must ensure that updates of
1980 * per-task data have been completed by this moment.
1983 task_thread_info(p
)->cpu
= cpu
;
1987 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1988 const struct sched_class
*prev_class
,
1989 int oldprio
, int running
)
1991 if (prev_class
!= p
->sched_class
) {
1992 if (prev_class
->switched_from
)
1993 prev_class
->switched_from(rq
, p
, running
);
1994 p
->sched_class
->switched_to(rq
, p
, running
);
1996 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2001 * Is this task likely cache-hot:
2004 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2009 * Buddy candidates are cache hot:
2011 if (sched_feat(CACHE_HOT_BUDDY
) &&
2012 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2013 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2016 if (p
->sched_class
!= &fair_sched_class
)
2019 if (sysctl_sched_migration_cost
== -1)
2021 if (sysctl_sched_migration_cost
== 0)
2024 delta
= now
- p
->se
.exec_start
;
2026 return delta
< (s64
)sysctl_sched_migration_cost
;
2030 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2032 int old_cpu
= task_cpu(p
);
2033 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2034 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2035 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2038 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2040 trace_sched_migrate_task(p
, new_cpu
);
2042 #ifdef CONFIG_SCHEDSTATS
2043 if (p
->se
.wait_start
)
2044 p
->se
.wait_start
-= clock_offset
;
2045 if (p
->se
.sleep_start
)
2046 p
->se
.sleep_start
-= clock_offset
;
2047 if (p
->se
.block_start
)
2048 p
->se
.block_start
-= clock_offset
;
2050 if (old_cpu
!= new_cpu
) {
2051 p
->se
.nr_migrations
++;
2052 new_rq
->nr_migrations_in
++;
2053 #ifdef CONFIG_SCHEDSTATS
2054 if (task_hot(p
, old_rq
->clock
, NULL
))
2055 schedstat_inc(p
, se
.nr_forced2_migrations
);
2057 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2060 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2061 new_cfsrq
->min_vruntime
;
2063 __set_task_cpu(p
, new_cpu
);
2066 struct migration_req
{
2067 struct list_head list
;
2069 struct task_struct
*task
;
2072 struct completion done
;
2076 * The task's runqueue lock must be held.
2077 * Returns true if you have to wait for migration thread.
2080 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2082 struct rq
*rq
= task_rq(p
);
2085 * If the task is not on a runqueue (and not running), then
2086 * it is sufficient to simply update the task's cpu field.
2088 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2089 set_task_cpu(p
, dest_cpu
);
2093 init_completion(&req
->done
);
2095 req
->dest_cpu
= dest_cpu
;
2096 list_add(&req
->list
, &rq
->migration_queue
);
2102 * wait_task_context_switch - wait for a thread to complete at least one
2105 * @p must not be current.
2107 void wait_task_context_switch(struct task_struct
*p
)
2109 unsigned long nvcsw
, nivcsw
, flags
;
2117 * The runqueue is assigned before the actual context
2118 * switch. We need to take the runqueue lock.
2120 * We could check initially without the lock but it is
2121 * very likely that we need to take the lock in every
2124 rq
= task_rq_lock(p
, &flags
);
2125 running
= task_running(rq
, p
);
2126 task_rq_unlock(rq
, &flags
);
2128 if (likely(!running
))
2131 * The switch count is incremented before the actual
2132 * context switch. We thus wait for two switches to be
2133 * sure at least one completed.
2135 if ((p
->nvcsw
- nvcsw
) > 1)
2137 if ((p
->nivcsw
- nivcsw
) > 1)
2145 * wait_task_inactive - wait for a thread to unschedule.
2147 * If @match_state is nonzero, it's the @p->state value just checked and
2148 * not expected to change. If it changes, i.e. @p might have woken up,
2149 * then return zero. When we succeed in waiting for @p to be off its CPU,
2150 * we return a positive number (its total switch count). If a second call
2151 * a short while later returns the same number, the caller can be sure that
2152 * @p has remained unscheduled the whole time.
2154 * The caller must ensure that the task *will* unschedule sometime soon,
2155 * else this function might spin for a *long* time. This function can't
2156 * be called with interrupts off, or it may introduce deadlock with
2157 * smp_call_function() if an IPI is sent by the same process we are
2158 * waiting to become inactive.
2160 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2162 unsigned long flags
;
2169 * We do the initial early heuristics without holding
2170 * any task-queue locks at all. We'll only try to get
2171 * the runqueue lock when things look like they will
2177 * If the task is actively running on another CPU
2178 * still, just relax and busy-wait without holding
2181 * NOTE! Since we don't hold any locks, it's not
2182 * even sure that "rq" stays as the right runqueue!
2183 * But we don't care, since "task_running()" will
2184 * return false if the runqueue has changed and p
2185 * is actually now running somewhere else!
2187 while (task_running(rq
, p
)) {
2188 if (match_state
&& unlikely(p
->state
!= match_state
))
2194 * Ok, time to look more closely! We need the rq
2195 * lock now, to be *sure*. If we're wrong, we'll
2196 * just go back and repeat.
2198 rq
= task_rq_lock(p
, &flags
);
2199 trace_sched_wait_task(rq
, p
);
2200 running
= task_running(rq
, p
);
2201 on_rq
= p
->se
.on_rq
;
2203 if (!match_state
|| p
->state
== match_state
)
2204 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2205 task_rq_unlock(rq
, &flags
);
2208 * If it changed from the expected state, bail out now.
2210 if (unlikely(!ncsw
))
2214 * Was it really running after all now that we
2215 * checked with the proper locks actually held?
2217 * Oops. Go back and try again..
2219 if (unlikely(running
)) {
2225 * It's not enough that it's not actively running,
2226 * it must be off the runqueue _entirely_, and not
2229 * So if it was still runnable (but just not actively
2230 * running right now), it's preempted, and we should
2231 * yield - it could be a while.
2233 if (unlikely(on_rq
)) {
2234 schedule_timeout_uninterruptible(1);
2239 * Ahh, all good. It wasn't running, and it wasn't
2240 * runnable, which means that it will never become
2241 * running in the future either. We're all done!
2250 * kick_process - kick a running thread to enter/exit the kernel
2251 * @p: the to-be-kicked thread
2253 * Cause a process which is running on another CPU to enter
2254 * kernel-mode, without any delay. (to get signals handled.)
2256 * NOTE: this function doesnt have to take the runqueue lock,
2257 * because all it wants to ensure is that the remote task enters
2258 * the kernel. If the IPI races and the task has been migrated
2259 * to another CPU then no harm is done and the purpose has been
2262 void kick_process(struct task_struct
*p
)
2268 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2269 smp_send_reschedule(cpu
);
2272 EXPORT_SYMBOL_GPL(kick_process
);
2273 #endif /* CONFIG_SMP */
2276 * task_oncpu_function_call - call a function on the cpu on which a task runs
2277 * @p: the task to evaluate
2278 * @func: the function to be called
2279 * @info: the function call argument
2281 * Calls the function @func when the task is currently running. This might
2282 * be on the current CPU, which just calls the function directly
2284 void task_oncpu_function_call(struct task_struct
*p
,
2285 void (*func
) (void *info
), void *info
)
2292 smp_call_function_single(cpu
, func
, info
, 1);
2297 * try_to_wake_up - wake up a thread
2298 * @p: the to-be-woken-up thread
2299 * @state: the mask of task states that can be woken
2300 * @sync: do a synchronous wakeup?
2302 * Put it on the run-queue if it's not already there. The "current"
2303 * thread is always on the run-queue (except when the actual
2304 * re-schedule is in progress), and as such you're allowed to do
2305 * the simpler "current->state = TASK_RUNNING" to mark yourself
2306 * runnable without the overhead of this.
2308 * returns failure only if the task is already active.
2310 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2313 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2314 unsigned long flags
;
2315 struct rq
*rq
, *orig_rq
;
2317 if (!sched_feat(SYNC_WAKEUPS
))
2318 wake_flags
&= ~WF_SYNC
;
2320 this_cpu
= get_cpu();
2323 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2324 update_rq_clock(rq
);
2325 if (!(p
->state
& state
))
2335 if (unlikely(task_running(rq
, p
)))
2339 * In order to handle concurrent wakeups and release the rq->lock
2340 * we put the task in TASK_WAKING state.
2342 * First fix up the nr_uninterruptible count:
2344 if (task_contributes_to_load(p
))
2345 rq
->nr_uninterruptible
--;
2346 p
->state
= TASK_WAKING
;
2347 task_rq_unlock(rq
, &flags
);
2349 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2350 if (cpu
!= orig_cpu
)
2351 set_task_cpu(p
, cpu
);
2353 rq
= task_rq_lock(p
, &flags
);
2356 update_rq_clock(rq
);
2358 WARN_ON(p
->state
!= TASK_WAKING
);
2361 #ifdef CONFIG_SCHEDSTATS
2362 schedstat_inc(rq
, ttwu_count
);
2363 if (cpu
== this_cpu
)
2364 schedstat_inc(rq
, ttwu_local
);
2366 struct sched_domain
*sd
;
2367 for_each_domain(this_cpu
, sd
) {
2368 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2369 schedstat_inc(sd
, ttwu_wake_remote
);
2374 #endif /* CONFIG_SCHEDSTATS */
2377 #endif /* CONFIG_SMP */
2378 schedstat_inc(p
, se
.nr_wakeups
);
2379 if (wake_flags
& WF_SYNC
)
2380 schedstat_inc(p
, se
.nr_wakeups_sync
);
2381 if (orig_cpu
!= cpu
)
2382 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2383 if (cpu
== this_cpu
)
2384 schedstat_inc(p
, se
.nr_wakeups_local
);
2386 schedstat_inc(p
, se
.nr_wakeups_remote
);
2387 activate_task(rq
, p
, 1);
2391 * Only attribute actual wakeups done by this task.
2393 if (!in_interrupt()) {
2394 struct sched_entity
*se
= ¤t
->se
;
2395 u64 sample
= se
->sum_exec_runtime
;
2397 if (se
->last_wakeup
)
2398 sample
-= se
->last_wakeup
;
2400 sample
-= se
->start_runtime
;
2401 update_avg(&se
->avg_wakeup
, sample
);
2403 se
->last_wakeup
= se
->sum_exec_runtime
;
2407 trace_sched_wakeup(rq
, p
, success
);
2408 check_preempt_curr(rq
, p
, wake_flags
);
2410 p
->state
= TASK_RUNNING
;
2412 if (p
->sched_class
->task_wake_up
)
2413 p
->sched_class
->task_wake_up(rq
, p
);
2416 task_rq_unlock(rq
, &flags
);
2423 * wake_up_process - Wake up a specific process
2424 * @p: The process to be woken up.
2426 * Attempt to wake up the nominated process and move it to the set of runnable
2427 * processes. Returns 1 if the process was woken up, 0 if it was already
2430 * It may be assumed that this function implies a write memory barrier before
2431 * changing the task state if and only if any tasks are woken up.
2433 int wake_up_process(struct task_struct
*p
)
2435 return try_to_wake_up(p
, TASK_ALL
, 0);
2437 EXPORT_SYMBOL(wake_up_process
);
2439 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2441 return try_to_wake_up(p
, state
, 0);
2445 * Perform scheduler related setup for a newly forked process p.
2446 * p is forked by current.
2448 * __sched_fork() is basic setup used by init_idle() too:
2450 static void __sched_fork(struct task_struct
*p
)
2452 p
->se
.exec_start
= 0;
2453 p
->se
.sum_exec_runtime
= 0;
2454 p
->se
.prev_sum_exec_runtime
= 0;
2455 p
->se
.nr_migrations
= 0;
2456 p
->se
.last_wakeup
= 0;
2457 p
->se
.avg_overlap
= 0;
2458 p
->se
.start_runtime
= 0;
2459 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2460 p
->se
.avg_running
= 0;
2462 #ifdef CONFIG_SCHEDSTATS
2463 p
->se
.wait_start
= 0;
2465 p
->se
.wait_count
= 0;
2468 p
->se
.sleep_start
= 0;
2469 p
->se
.sleep_max
= 0;
2470 p
->se
.sum_sleep_runtime
= 0;
2472 p
->se
.block_start
= 0;
2473 p
->se
.block_max
= 0;
2475 p
->se
.slice_max
= 0;
2477 p
->se
.nr_migrations_cold
= 0;
2478 p
->se
.nr_failed_migrations_affine
= 0;
2479 p
->se
.nr_failed_migrations_running
= 0;
2480 p
->se
.nr_failed_migrations_hot
= 0;
2481 p
->se
.nr_forced_migrations
= 0;
2482 p
->se
.nr_forced2_migrations
= 0;
2484 p
->se
.nr_wakeups
= 0;
2485 p
->se
.nr_wakeups_sync
= 0;
2486 p
->se
.nr_wakeups_migrate
= 0;
2487 p
->se
.nr_wakeups_local
= 0;
2488 p
->se
.nr_wakeups_remote
= 0;
2489 p
->se
.nr_wakeups_affine
= 0;
2490 p
->se
.nr_wakeups_affine_attempts
= 0;
2491 p
->se
.nr_wakeups_passive
= 0;
2492 p
->se
.nr_wakeups_idle
= 0;
2496 INIT_LIST_HEAD(&p
->rt
.run_list
);
2498 INIT_LIST_HEAD(&p
->se
.group_node
);
2500 #ifdef CONFIG_PREEMPT_NOTIFIERS
2501 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2505 * We mark the process as running here, but have not actually
2506 * inserted it onto the runqueue yet. This guarantees that
2507 * nobody will actually run it, and a signal or other external
2508 * event cannot wake it up and insert it on the runqueue either.
2510 p
->state
= TASK_RUNNING
;
2514 * fork()/clone()-time setup:
2516 void sched_fork(struct task_struct
*p
, int clone_flags
)
2518 int cpu
= get_cpu();
2523 * Revert to default priority/policy on fork if requested.
2525 if (unlikely(p
->sched_reset_on_fork
)) {
2526 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2527 p
->policy
= SCHED_NORMAL
;
2528 p
->normal_prio
= p
->static_prio
;
2531 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2532 p
->static_prio
= NICE_TO_PRIO(0);
2533 p
->normal_prio
= p
->static_prio
;
2538 * We don't need the reset flag anymore after the fork. It has
2539 * fulfilled its duty:
2541 p
->sched_reset_on_fork
= 0;
2545 * Make sure we do not leak PI boosting priority to the child.
2547 p
->prio
= current
->normal_prio
;
2549 if (!rt_prio(p
->prio
))
2550 p
->sched_class
= &fair_sched_class
;
2553 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_FORK
, 0);
2555 set_task_cpu(p
, cpu
);
2557 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2558 if (likely(sched_info_on()))
2559 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2561 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2564 #ifdef CONFIG_PREEMPT
2565 /* Want to start with kernel preemption disabled. */
2566 task_thread_info(p
)->preempt_count
= 1;
2568 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2574 * wake_up_new_task - wake up a newly created task for the first time.
2576 * This function will do some initial scheduler statistics housekeeping
2577 * that must be done for every newly created context, then puts the task
2578 * on the runqueue and wakes it.
2580 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2582 unsigned long flags
;
2585 rq
= task_rq_lock(p
, &flags
);
2586 BUG_ON(p
->state
!= TASK_RUNNING
);
2587 update_rq_clock(rq
);
2589 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2590 activate_task(rq
, p
, 0);
2593 * Let the scheduling class do new task startup
2594 * management (if any):
2596 p
->sched_class
->task_new(rq
, p
);
2599 trace_sched_wakeup_new(rq
, p
, 1);
2600 check_preempt_curr(rq
, p
, WF_FORK
);
2602 if (p
->sched_class
->task_wake_up
)
2603 p
->sched_class
->task_wake_up(rq
, p
);
2605 task_rq_unlock(rq
, &flags
);
2608 #ifdef CONFIG_PREEMPT_NOTIFIERS
2611 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2612 * @notifier: notifier struct to register
2614 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2616 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2618 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2621 * preempt_notifier_unregister - no longer interested in preemption notifications
2622 * @notifier: notifier struct to unregister
2624 * This is safe to call from within a preemption notifier.
2626 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2628 hlist_del(¬ifier
->link
);
2630 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2632 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2634 struct preempt_notifier
*notifier
;
2635 struct hlist_node
*node
;
2637 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2638 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2642 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2643 struct task_struct
*next
)
2645 struct preempt_notifier
*notifier
;
2646 struct hlist_node
*node
;
2648 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2649 notifier
->ops
->sched_out(notifier
, next
);
2652 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2654 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2659 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2660 struct task_struct
*next
)
2664 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2667 * prepare_task_switch - prepare to switch tasks
2668 * @rq: the runqueue preparing to switch
2669 * @prev: the current task that is being switched out
2670 * @next: the task we are going to switch to.
2672 * This is called with the rq lock held and interrupts off. It must
2673 * be paired with a subsequent finish_task_switch after the context
2676 * prepare_task_switch sets up locking and calls architecture specific
2680 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2681 struct task_struct
*next
)
2683 fire_sched_out_preempt_notifiers(prev
, next
);
2684 prepare_lock_switch(rq
, next
);
2685 prepare_arch_switch(next
);
2689 * finish_task_switch - clean up after a task-switch
2690 * @rq: runqueue associated with task-switch
2691 * @prev: the thread we just switched away from.
2693 * finish_task_switch must be called after the context switch, paired
2694 * with a prepare_task_switch call before the context switch.
2695 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2696 * and do any other architecture-specific cleanup actions.
2698 * Note that we may have delayed dropping an mm in context_switch(). If
2699 * so, we finish that here outside of the runqueue lock. (Doing it
2700 * with the lock held can cause deadlocks; see schedule() for
2703 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2704 __releases(rq
->lock
)
2706 struct mm_struct
*mm
= rq
->prev_mm
;
2712 * A task struct has one reference for the use as "current".
2713 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2714 * schedule one last time. The schedule call will never return, and
2715 * the scheduled task must drop that reference.
2716 * The test for TASK_DEAD must occur while the runqueue locks are
2717 * still held, otherwise prev could be scheduled on another cpu, die
2718 * there before we look at prev->state, and then the reference would
2720 * Manfred Spraul <manfred@colorfullife.com>
2722 prev_state
= prev
->state
;
2723 finish_arch_switch(prev
);
2724 perf_event_task_sched_in(current
, cpu_of(rq
));
2725 finish_lock_switch(rq
, prev
);
2727 fire_sched_in_preempt_notifiers(current
);
2730 if (unlikely(prev_state
== TASK_DEAD
)) {
2732 * Remove function-return probe instances associated with this
2733 * task and put them back on the free list.
2735 kprobe_flush_task(prev
);
2736 put_task_struct(prev
);
2742 /* assumes rq->lock is held */
2743 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2745 if (prev
->sched_class
->pre_schedule
)
2746 prev
->sched_class
->pre_schedule(rq
, prev
);
2749 /* rq->lock is NOT held, but preemption is disabled */
2750 static inline void post_schedule(struct rq
*rq
)
2752 if (rq
->post_schedule
) {
2753 unsigned long flags
;
2755 spin_lock_irqsave(&rq
->lock
, flags
);
2756 if (rq
->curr
->sched_class
->post_schedule
)
2757 rq
->curr
->sched_class
->post_schedule(rq
);
2758 spin_unlock_irqrestore(&rq
->lock
, flags
);
2760 rq
->post_schedule
= 0;
2766 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2770 static inline void post_schedule(struct rq
*rq
)
2777 * schedule_tail - first thing a freshly forked thread must call.
2778 * @prev: the thread we just switched away from.
2780 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2781 __releases(rq
->lock
)
2783 struct rq
*rq
= this_rq();
2785 finish_task_switch(rq
, prev
);
2788 * FIXME: do we need to worry about rq being invalidated by the
2793 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2794 /* In this case, finish_task_switch does not reenable preemption */
2797 if (current
->set_child_tid
)
2798 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2802 * context_switch - switch to the new MM and the new
2803 * thread's register state.
2806 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2807 struct task_struct
*next
)
2809 struct mm_struct
*mm
, *oldmm
;
2811 prepare_task_switch(rq
, prev
, next
);
2812 trace_sched_switch(rq
, prev
, next
);
2814 oldmm
= prev
->active_mm
;
2816 * For paravirt, this is coupled with an exit in switch_to to
2817 * combine the page table reload and the switch backend into
2820 arch_start_context_switch(prev
);
2822 if (unlikely(!mm
)) {
2823 next
->active_mm
= oldmm
;
2824 atomic_inc(&oldmm
->mm_count
);
2825 enter_lazy_tlb(oldmm
, next
);
2827 switch_mm(oldmm
, mm
, next
);
2829 if (unlikely(!prev
->mm
)) {
2830 prev
->active_mm
= NULL
;
2831 rq
->prev_mm
= oldmm
;
2834 * Since the runqueue lock will be released by the next
2835 * task (which is an invalid locking op but in the case
2836 * of the scheduler it's an obvious special-case), so we
2837 * do an early lockdep release here:
2839 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2840 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2843 /* Here we just switch the register state and the stack. */
2844 switch_to(prev
, next
, prev
);
2848 * this_rq must be evaluated again because prev may have moved
2849 * CPUs since it called schedule(), thus the 'rq' on its stack
2850 * frame will be invalid.
2852 finish_task_switch(this_rq(), prev
);
2856 * nr_running, nr_uninterruptible and nr_context_switches:
2858 * externally visible scheduler statistics: current number of runnable
2859 * threads, current number of uninterruptible-sleeping threads, total
2860 * number of context switches performed since bootup.
2862 unsigned long nr_running(void)
2864 unsigned long i
, sum
= 0;
2866 for_each_online_cpu(i
)
2867 sum
+= cpu_rq(i
)->nr_running
;
2872 unsigned long nr_uninterruptible(void)
2874 unsigned long i
, sum
= 0;
2876 for_each_possible_cpu(i
)
2877 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2880 * Since we read the counters lockless, it might be slightly
2881 * inaccurate. Do not allow it to go below zero though:
2883 if (unlikely((long)sum
< 0))
2889 unsigned long long nr_context_switches(void)
2892 unsigned long long sum
= 0;
2894 for_each_possible_cpu(i
)
2895 sum
+= cpu_rq(i
)->nr_switches
;
2900 unsigned long nr_iowait(void)
2902 unsigned long i
, sum
= 0;
2904 for_each_possible_cpu(i
)
2905 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2910 unsigned long nr_iowait_cpu(void)
2912 struct rq
*this = this_rq();
2913 return atomic_read(&this->nr_iowait
);
2916 unsigned long this_cpu_load(void)
2918 struct rq
*this = this_rq();
2919 return this->cpu_load
[0];
2923 /* Variables and functions for calc_load */
2924 static atomic_long_t calc_load_tasks
;
2925 static unsigned long calc_load_update
;
2926 unsigned long avenrun
[3];
2927 EXPORT_SYMBOL(avenrun
);
2930 * get_avenrun - get the load average array
2931 * @loads: pointer to dest load array
2932 * @offset: offset to add
2933 * @shift: shift count to shift the result left
2935 * These values are estimates at best, so no need for locking.
2937 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2939 loads
[0] = (avenrun
[0] + offset
) << shift
;
2940 loads
[1] = (avenrun
[1] + offset
) << shift
;
2941 loads
[2] = (avenrun
[2] + offset
) << shift
;
2944 static unsigned long
2945 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2948 load
+= active
* (FIXED_1
- exp
);
2949 return load
>> FSHIFT
;
2953 * calc_load - update the avenrun load estimates 10 ticks after the
2954 * CPUs have updated calc_load_tasks.
2956 void calc_global_load(void)
2958 unsigned long upd
= calc_load_update
+ 10;
2961 if (time_before(jiffies
, upd
))
2964 active
= atomic_long_read(&calc_load_tasks
);
2965 active
= active
> 0 ? active
* FIXED_1
: 0;
2967 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2968 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2969 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2971 calc_load_update
+= LOAD_FREQ
;
2975 * Either called from update_cpu_load() or from a cpu going idle
2977 static void calc_load_account_active(struct rq
*this_rq
)
2979 long nr_active
, delta
;
2981 nr_active
= this_rq
->nr_running
;
2982 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2984 if (nr_active
!= this_rq
->calc_load_active
) {
2985 delta
= nr_active
- this_rq
->calc_load_active
;
2986 this_rq
->calc_load_active
= nr_active
;
2987 atomic_long_add(delta
, &calc_load_tasks
);
2992 * Externally visible per-cpu scheduler statistics:
2993 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2995 u64
cpu_nr_migrations(int cpu
)
2997 return cpu_rq(cpu
)->nr_migrations_in
;
3001 * Update rq->cpu_load[] statistics. This function is usually called every
3002 * scheduler tick (TICK_NSEC).
3004 static void update_cpu_load(struct rq
*this_rq
)
3006 unsigned long this_load
= this_rq
->load
.weight
;
3009 this_rq
->nr_load_updates
++;
3011 /* Update our load: */
3012 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3013 unsigned long old_load
, new_load
;
3015 /* scale is effectively 1 << i now, and >> i divides by scale */
3017 old_load
= this_rq
->cpu_load
[i
];
3018 new_load
= this_load
;
3020 * Round up the averaging division if load is increasing. This
3021 * prevents us from getting stuck on 9 if the load is 10, for
3024 if (new_load
> old_load
)
3025 new_load
+= scale
-1;
3026 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3029 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3030 this_rq
->calc_load_update
+= LOAD_FREQ
;
3031 calc_load_account_active(this_rq
);
3038 * double_rq_lock - safely lock two runqueues
3040 * Note this does not disable interrupts like task_rq_lock,
3041 * you need to do so manually before calling.
3043 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3044 __acquires(rq1
->lock
)
3045 __acquires(rq2
->lock
)
3047 BUG_ON(!irqs_disabled());
3049 spin_lock(&rq1
->lock
);
3050 __acquire(rq2
->lock
); /* Fake it out ;) */
3053 spin_lock(&rq1
->lock
);
3054 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3056 spin_lock(&rq2
->lock
);
3057 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3060 update_rq_clock(rq1
);
3061 update_rq_clock(rq2
);
3065 * double_rq_unlock - safely unlock two runqueues
3067 * Note this does not restore interrupts like task_rq_unlock,
3068 * you need to do so manually after calling.
3070 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3071 __releases(rq1
->lock
)
3072 __releases(rq2
->lock
)
3074 spin_unlock(&rq1
->lock
);
3076 spin_unlock(&rq2
->lock
);
3078 __release(rq2
->lock
);
3082 * If dest_cpu is allowed for this process, migrate the task to it.
3083 * This is accomplished by forcing the cpu_allowed mask to only
3084 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3085 * the cpu_allowed mask is restored.
3087 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3089 struct migration_req req
;
3090 unsigned long flags
;
3093 rq
= task_rq_lock(p
, &flags
);
3094 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3095 || unlikely(!cpu_active(dest_cpu
)))
3098 /* force the process onto the specified CPU */
3099 if (migrate_task(p
, dest_cpu
, &req
)) {
3100 /* Need to wait for migration thread (might exit: take ref). */
3101 struct task_struct
*mt
= rq
->migration_thread
;
3103 get_task_struct(mt
);
3104 task_rq_unlock(rq
, &flags
);
3105 wake_up_process(mt
);
3106 put_task_struct(mt
);
3107 wait_for_completion(&req
.done
);
3112 task_rq_unlock(rq
, &flags
);
3116 * sched_exec - execve() is a valuable balancing opportunity, because at
3117 * this point the task has the smallest effective memory and cache footprint.
3119 void sched_exec(void)
3121 int new_cpu
, this_cpu
= get_cpu();
3122 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3124 if (new_cpu
!= this_cpu
)
3125 sched_migrate_task(current
, new_cpu
);
3129 * pull_task - move a task from a remote runqueue to the local runqueue.
3130 * Both runqueues must be locked.
3132 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3133 struct rq
*this_rq
, int this_cpu
)
3135 deactivate_task(src_rq
, p
, 0);
3136 set_task_cpu(p
, this_cpu
);
3137 activate_task(this_rq
, p
, 0);
3139 * Note that idle threads have a prio of MAX_PRIO, for this test
3140 * to be always true for them.
3142 check_preempt_curr(this_rq
, p
, 0);
3146 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3149 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3150 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3153 int tsk_cache_hot
= 0;
3155 * We do not migrate tasks that are:
3156 * 1) running (obviously), or
3157 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3158 * 3) are cache-hot on their current CPU.
3160 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3161 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3166 if (task_running(rq
, p
)) {
3167 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3172 * Aggressive migration if:
3173 * 1) task is cache cold, or
3174 * 2) too many balance attempts have failed.
3177 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3178 if (!tsk_cache_hot
||
3179 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3180 #ifdef CONFIG_SCHEDSTATS
3181 if (tsk_cache_hot
) {
3182 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3183 schedstat_inc(p
, se
.nr_forced_migrations
);
3189 if (tsk_cache_hot
) {
3190 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3196 static unsigned long
3197 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3198 unsigned long max_load_move
, struct sched_domain
*sd
,
3199 enum cpu_idle_type idle
, int *all_pinned
,
3200 int *this_best_prio
, struct rq_iterator
*iterator
)
3202 int loops
= 0, pulled
= 0, pinned
= 0;
3203 struct task_struct
*p
;
3204 long rem_load_move
= max_load_move
;
3206 if (max_load_move
== 0)
3212 * Start the load-balancing iterator:
3214 p
= iterator
->start(iterator
->arg
);
3216 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3219 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3220 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3221 p
= iterator
->next(iterator
->arg
);
3225 pull_task(busiest
, p
, this_rq
, this_cpu
);
3227 rem_load_move
-= p
->se
.load
.weight
;
3229 #ifdef CONFIG_PREEMPT
3231 * NEWIDLE balancing is a source of latency, so preemptible kernels
3232 * will stop after the first task is pulled to minimize the critical
3235 if (idle
== CPU_NEWLY_IDLE
)
3240 * We only want to steal up to the prescribed amount of weighted load.
3242 if (rem_load_move
> 0) {
3243 if (p
->prio
< *this_best_prio
)
3244 *this_best_prio
= p
->prio
;
3245 p
= iterator
->next(iterator
->arg
);
3250 * Right now, this is one of only two places pull_task() is called,
3251 * so we can safely collect pull_task() stats here rather than
3252 * inside pull_task().
3254 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3257 *all_pinned
= pinned
;
3259 return max_load_move
- rem_load_move
;
3263 * move_tasks tries to move up to max_load_move weighted load from busiest to
3264 * this_rq, as part of a balancing operation within domain "sd".
3265 * Returns 1 if successful and 0 otherwise.
3267 * Called with both runqueues locked.
3269 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3270 unsigned long max_load_move
,
3271 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3274 const struct sched_class
*class = sched_class_highest
;
3275 unsigned long total_load_moved
= 0;
3276 int this_best_prio
= this_rq
->curr
->prio
;
3280 class->load_balance(this_rq
, this_cpu
, busiest
,
3281 max_load_move
- total_load_moved
,
3282 sd
, idle
, all_pinned
, &this_best_prio
);
3283 class = class->next
;
3285 #ifdef CONFIG_PREEMPT
3287 * NEWIDLE balancing is a source of latency, so preemptible
3288 * kernels will stop after the first task is pulled to minimize
3289 * the critical section.
3291 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3294 } while (class && max_load_move
> total_load_moved
);
3296 return total_load_moved
> 0;
3300 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3301 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3302 struct rq_iterator
*iterator
)
3304 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3308 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3309 pull_task(busiest
, p
, this_rq
, this_cpu
);
3311 * Right now, this is only the second place pull_task()
3312 * is called, so we can safely collect pull_task()
3313 * stats here rather than inside pull_task().
3315 schedstat_inc(sd
, lb_gained
[idle
]);
3319 p
= iterator
->next(iterator
->arg
);
3326 * move_one_task tries to move exactly one task from busiest to this_rq, as
3327 * part of active balancing operations within "domain".
3328 * Returns 1 if successful and 0 otherwise.
3330 * Called with both runqueues locked.
3332 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3333 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3335 const struct sched_class
*class;
3337 for_each_class(class) {
3338 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3344 /********** Helpers for find_busiest_group ************************/
3346 * sd_lb_stats - Structure to store the statistics of a sched_domain
3347 * during load balancing.
3349 struct sd_lb_stats
{
3350 struct sched_group
*busiest
; /* Busiest group in this sd */
3351 struct sched_group
*this; /* Local group in this sd */
3352 unsigned long total_load
; /* Total load of all groups in sd */
3353 unsigned long total_pwr
; /* Total power of all groups in sd */
3354 unsigned long avg_load
; /* Average load across all groups in sd */
3356 /** Statistics of this group */
3357 unsigned long this_load
;
3358 unsigned long this_load_per_task
;
3359 unsigned long this_nr_running
;
3361 /* Statistics of the busiest group */
3362 unsigned long max_load
;
3363 unsigned long busiest_load_per_task
;
3364 unsigned long busiest_nr_running
;
3366 int group_imb
; /* Is there imbalance in this sd */
3367 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3368 int power_savings_balance
; /* Is powersave balance needed for this sd */
3369 struct sched_group
*group_min
; /* Least loaded group in sd */
3370 struct sched_group
*group_leader
; /* Group which relieves group_min */
3371 unsigned long min_load_per_task
; /* load_per_task in group_min */
3372 unsigned long leader_nr_running
; /* Nr running of group_leader */
3373 unsigned long min_nr_running
; /* Nr running of group_min */
3378 * sg_lb_stats - stats of a sched_group required for load_balancing
3380 struct sg_lb_stats
{
3381 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3382 unsigned long group_load
; /* Total load over the CPUs of the group */
3383 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3384 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3385 unsigned long group_capacity
;
3386 int group_imb
; /* Is there an imbalance in the group ? */
3390 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3391 * @group: The group whose first cpu is to be returned.
3393 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3395 return cpumask_first(sched_group_cpus(group
));
3399 * get_sd_load_idx - Obtain the load index for a given sched domain.
3400 * @sd: The sched_domain whose load_idx is to be obtained.
3401 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3403 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3404 enum cpu_idle_type idle
)
3410 load_idx
= sd
->busy_idx
;
3413 case CPU_NEWLY_IDLE
:
3414 load_idx
= sd
->newidle_idx
;
3417 load_idx
= sd
->idle_idx
;
3425 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3427 * init_sd_power_savings_stats - Initialize power savings statistics for
3428 * the given sched_domain, during load balancing.
3430 * @sd: Sched domain whose power-savings statistics are to be initialized.
3431 * @sds: Variable containing the statistics for sd.
3432 * @idle: Idle status of the CPU at which we're performing load-balancing.
3434 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3435 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3438 * Busy processors will not participate in power savings
3441 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3442 sds
->power_savings_balance
= 0;
3444 sds
->power_savings_balance
= 1;
3445 sds
->min_nr_running
= ULONG_MAX
;
3446 sds
->leader_nr_running
= 0;
3451 * update_sd_power_savings_stats - Update the power saving stats for a
3452 * sched_domain while performing load balancing.
3454 * @group: sched_group belonging to the sched_domain under consideration.
3455 * @sds: Variable containing the statistics of the sched_domain
3456 * @local_group: Does group contain the CPU for which we're performing
3458 * @sgs: Variable containing the statistics of the group.
3460 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3461 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3464 if (!sds
->power_savings_balance
)
3468 * If the local group is idle or completely loaded
3469 * no need to do power savings balance at this domain
3471 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3472 !sds
->this_nr_running
))
3473 sds
->power_savings_balance
= 0;
3476 * If a group is already running at full capacity or idle,
3477 * don't include that group in power savings calculations
3479 if (!sds
->power_savings_balance
||
3480 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3481 !sgs
->sum_nr_running
)
3485 * Calculate the group which has the least non-idle load.
3486 * This is the group from where we need to pick up the load
3489 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3490 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3491 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3492 sds
->group_min
= group
;
3493 sds
->min_nr_running
= sgs
->sum_nr_running
;
3494 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3495 sgs
->sum_nr_running
;
3499 * Calculate the group which is almost near its
3500 * capacity but still has some space to pick up some load
3501 * from other group and save more power
3503 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3506 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3507 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3508 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3509 sds
->group_leader
= group
;
3510 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3515 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3516 * @sds: Variable containing the statistics of the sched_domain
3517 * under consideration.
3518 * @this_cpu: Cpu at which we're currently performing load-balancing.
3519 * @imbalance: Variable to store the imbalance.
3522 * Check if we have potential to perform some power-savings balance.
3523 * If yes, set the busiest group to be the least loaded group in the
3524 * sched_domain, so that it's CPUs can be put to idle.
3526 * Returns 1 if there is potential to perform power-savings balance.
3529 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3530 int this_cpu
, unsigned long *imbalance
)
3532 if (!sds
->power_savings_balance
)
3535 if (sds
->this != sds
->group_leader
||
3536 sds
->group_leader
== sds
->group_min
)
3539 *imbalance
= sds
->min_load_per_task
;
3540 sds
->busiest
= sds
->group_min
;
3545 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3546 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3547 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3552 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3553 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3558 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3559 int this_cpu
, unsigned long *imbalance
)
3563 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3566 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3568 return SCHED_LOAD_SCALE
;
3571 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3573 return default_scale_freq_power(sd
, cpu
);
3576 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3578 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3579 unsigned long smt_gain
= sd
->smt_gain
;
3586 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3588 return default_scale_smt_power(sd
, cpu
);
3591 unsigned long scale_rt_power(int cpu
)
3593 struct rq
*rq
= cpu_rq(cpu
);
3594 u64 total
, available
;
3596 sched_avg_update(rq
);
3598 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3599 available
= total
- rq
->rt_avg
;
3601 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3602 total
= SCHED_LOAD_SCALE
;
3604 total
>>= SCHED_LOAD_SHIFT
;
3606 return div_u64(available
, total
);
3609 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3611 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3612 unsigned long power
= SCHED_LOAD_SCALE
;
3613 struct sched_group
*sdg
= sd
->groups
;
3615 if (sched_feat(ARCH_POWER
))
3616 power
*= arch_scale_freq_power(sd
, cpu
);
3618 power
*= default_scale_freq_power(sd
, cpu
);
3620 power
>>= SCHED_LOAD_SHIFT
;
3622 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3623 if (sched_feat(ARCH_POWER
))
3624 power
*= arch_scale_smt_power(sd
, cpu
);
3626 power
*= default_scale_smt_power(sd
, cpu
);
3628 power
>>= SCHED_LOAD_SHIFT
;
3631 power
*= scale_rt_power(cpu
);
3632 power
>>= SCHED_LOAD_SHIFT
;
3637 sdg
->cpu_power
= power
;
3640 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3642 struct sched_domain
*child
= sd
->child
;
3643 struct sched_group
*group
, *sdg
= sd
->groups
;
3644 unsigned long power
;
3647 update_cpu_power(sd
, cpu
);
3653 group
= child
->groups
;
3655 power
+= group
->cpu_power
;
3656 group
= group
->next
;
3657 } while (group
!= child
->groups
);
3659 sdg
->cpu_power
= power
;
3663 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3664 * @sd: The sched_domain whose statistics are to be updated.
3665 * @group: sched_group whose statistics are to be updated.
3666 * @this_cpu: Cpu for which load balance is currently performed.
3667 * @idle: Idle status of this_cpu
3668 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3669 * @sd_idle: Idle status of the sched_domain containing group.
3670 * @local_group: Does group contain this_cpu.
3671 * @cpus: Set of cpus considered for load balancing.
3672 * @balance: Should we balance.
3673 * @sgs: variable to hold the statistics for this group.
3675 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3676 struct sched_group
*group
, int this_cpu
,
3677 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3678 int local_group
, const struct cpumask
*cpus
,
3679 int *balance
, struct sg_lb_stats
*sgs
)
3681 unsigned long load
, max_cpu_load
, min_cpu_load
;
3683 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3684 unsigned long sum_avg_load_per_task
;
3685 unsigned long avg_load_per_task
;
3688 balance_cpu
= group_first_cpu(group
);
3689 if (balance_cpu
== this_cpu
)
3690 update_group_power(sd
, this_cpu
);
3693 /* Tally up the load of all CPUs in the group */
3694 sum_avg_load_per_task
= avg_load_per_task
= 0;
3696 min_cpu_load
= ~0UL;
3698 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3699 struct rq
*rq
= cpu_rq(i
);
3701 if (*sd_idle
&& rq
->nr_running
)
3704 /* Bias balancing toward cpus of our domain */
3706 if (idle_cpu(i
) && !first_idle_cpu
) {
3711 load
= target_load(i
, load_idx
);
3713 load
= source_load(i
, load_idx
);
3714 if (load
> max_cpu_load
)
3715 max_cpu_load
= load
;
3716 if (min_cpu_load
> load
)
3717 min_cpu_load
= load
;
3720 sgs
->group_load
+= load
;
3721 sgs
->sum_nr_running
+= rq
->nr_running
;
3722 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3724 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3728 * First idle cpu or the first cpu(busiest) in this sched group
3729 * is eligible for doing load balancing at this and above
3730 * domains. In the newly idle case, we will allow all the cpu's
3731 * to do the newly idle load balance.
3733 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3734 balance_cpu
!= this_cpu
&& balance
) {
3739 /* Adjust by relative CPU power of the group */
3740 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3744 * Consider the group unbalanced when the imbalance is larger
3745 * than the average weight of two tasks.
3747 * APZ: with cgroup the avg task weight can vary wildly and
3748 * might not be a suitable number - should we keep a
3749 * normalized nr_running number somewhere that negates
3752 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3755 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3758 sgs
->group_capacity
=
3759 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3763 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3764 * @sd: sched_domain whose statistics are to be updated.
3765 * @this_cpu: Cpu for which load balance is currently performed.
3766 * @idle: Idle status of this_cpu
3767 * @sd_idle: Idle status of the sched_domain containing group.
3768 * @cpus: Set of cpus considered for load balancing.
3769 * @balance: Should we balance.
3770 * @sds: variable to hold the statistics for this sched_domain.
3772 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3773 enum cpu_idle_type idle
, int *sd_idle
,
3774 const struct cpumask
*cpus
, int *balance
,
3775 struct sd_lb_stats
*sds
)
3777 struct sched_domain
*child
= sd
->child
;
3778 struct sched_group
*group
= sd
->groups
;
3779 struct sg_lb_stats sgs
;
3780 int load_idx
, prefer_sibling
= 0;
3782 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3785 init_sd_power_savings_stats(sd
, sds
, idle
);
3786 load_idx
= get_sd_load_idx(sd
, idle
);
3791 local_group
= cpumask_test_cpu(this_cpu
,
3792 sched_group_cpus(group
));
3793 memset(&sgs
, 0, sizeof(sgs
));
3794 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3795 local_group
, cpus
, balance
, &sgs
);
3797 if (local_group
&& balance
&& !(*balance
))
3800 sds
->total_load
+= sgs
.group_load
;
3801 sds
->total_pwr
+= group
->cpu_power
;
3804 * In case the child domain prefers tasks go to siblings
3805 * first, lower the group capacity to one so that we'll try
3806 * and move all the excess tasks away.
3809 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3812 sds
->this_load
= sgs
.avg_load
;
3814 sds
->this_nr_running
= sgs
.sum_nr_running
;
3815 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3816 } else if (sgs
.avg_load
> sds
->max_load
&&
3817 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3819 sds
->max_load
= sgs
.avg_load
;
3820 sds
->busiest
= group
;
3821 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3822 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3823 sds
->group_imb
= sgs
.group_imb
;
3826 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3827 group
= group
->next
;
3828 } while (group
!= sd
->groups
);
3832 * fix_small_imbalance - Calculate the minor imbalance that exists
3833 * amongst the groups of a sched_domain, during
3835 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3836 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3837 * @imbalance: Variable to store the imbalance.
3839 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3840 int this_cpu
, unsigned long *imbalance
)
3842 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3843 unsigned int imbn
= 2;
3845 if (sds
->this_nr_running
) {
3846 sds
->this_load_per_task
/= sds
->this_nr_running
;
3847 if (sds
->busiest_load_per_task
>
3848 sds
->this_load_per_task
)
3851 sds
->this_load_per_task
=
3852 cpu_avg_load_per_task(this_cpu
);
3854 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3855 sds
->busiest_load_per_task
* imbn
) {
3856 *imbalance
= sds
->busiest_load_per_task
;
3861 * OK, we don't have enough imbalance to justify moving tasks,
3862 * however we may be able to increase total CPU power used by
3866 pwr_now
+= sds
->busiest
->cpu_power
*
3867 min(sds
->busiest_load_per_task
, sds
->max_load
);
3868 pwr_now
+= sds
->this->cpu_power
*
3869 min(sds
->this_load_per_task
, sds
->this_load
);
3870 pwr_now
/= SCHED_LOAD_SCALE
;
3872 /* Amount of load we'd subtract */
3873 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3874 sds
->busiest
->cpu_power
;
3875 if (sds
->max_load
> tmp
)
3876 pwr_move
+= sds
->busiest
->cpu_power
*
3877 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3879 /* Amount of load we'd add */
3880 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3881 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3882 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3883 sds
->this->cpu_power
;
3885 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3886 sds
->this->cpu_power
;
3887 pwr_move
+= sds
->this->cpu_power
*
3888 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3889 pwr_move
/= SCHED_LOAD_SCALE
;
3891 /* Move if we gain throughput */
3892 if (pwr_move
> pwr_now
)
3893 *imbalance
= sds
->busiest_load_per_task
;
3897 * calculate_imbalance - Calculate the amount of imbalance present within the
3898 * groups of a given sched_domain during load balance.
3899 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3900 * @this_cpu: Cpu for which currently load balance is being performed.
3901 * @imbalance: The variable to store the imbalance.
3903 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3904 unsigned long *imbalance
)
3906 unsigned long max_pull
;
3908 * In the presence of smp nice balancing, certain scenarios can have
3909 * max load less than avg load(as we skip the groups at or below
3910 * its cpu_power, while calculating max_load..)
3912 if (sds
->max_load
< sds
->avg_load
) {
3914 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3917 /* Don't want to pull so many tasks that a group would go idle */
3918 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3919 sds
->max_load
- sds
->busiest_load_per_task
);
3921 /* How much load to actually move to equalise the imbalance */
3922 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3923 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3927 * if *imbalance is less than the average load per runnable task
3928 * there is no gaurantee that any tasks will be moved so we'll have
3929 * a think about bumping its value to force at least one task to be
3932 if (*imbalance
< sds
->busiest_load_per_task
)
3933 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3936 /******* find_busiest_group() helpers end here *********************/
3939 * find_busiest_group - Returns the busiest group within the sched_domain
3940 * if there is an imbalance. If there isn't an imbalance, and
3941 * the user has opted for power-savings, it returns a group whose
3942 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3943 * such a group exists.
3945 * Also calculates the amount of weighted load which should be moved
3946 * to restore balance.
3948 * @sd: The sched_domain whose busiest group is to be returned.
3949 * @this_cpu: The cpu for which load balancing is currently being performed.
3950 * @imbalance: Variable which stores amount of weighted load which should
3951 * be moved to restore balance/put a group to idle.
3952 * @idle: The idle status of this_cpu.
3953 * @sd_idle: The idleness of sd
3954 * @cpus: The set of CPUs under consideration for load-balancing.
3955 * @balance: Pointer to a variable indicating if this_cpu
3956 * is the appropriate cpu to perform load balancing at this_level.
3958 * Returns: - the busiest group if imbalance exists.
3959 * - If no imbalance and user has opted for power-savings balance,
3960 * return the least loaded group whose CPUs can be
3961 * put to idle by rebalancing its tasks onto our group.
3963 static struct sched_group
*
3964 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3965 unsigned long *imbalance
, enum cpu_idle_type idle
,
3966 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3968 struct sd_lb_stats sds
;
3970 memset(&sds
, 0, sizeof(sds
));
3973 * Compute the various statistics relavent for load balancing at
3976 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3979 /* Cases where imbalance does not exist from POV of this_cpu */
3980 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3982 * 2) There is no busy sibling group to pull from.
3983 * 3) This group is the busiest group.
3984 * 4) This group is more busy than the avg busieness at this
3986 * 5) The imbalance is within the specified limit.
3987 * 6) Any rebalance would lead to ping-pong
3989 if (balance
&& !(*balance
))
3992 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3995 if (sds
.this_load
>= sds
.max_load
)
3998 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4000 if (sds
.this_load
>= sds
.avg_load
)
4003 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4006 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4008 sds
.busiest_load_per_task
=
4009 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4012 * We're trying to get all the cpus to the average_load, so we don't
4013 * want to push ourselves above the average load, nor do we wish to
4014 * reduce the max loaded cpu below the average load, as either of these
4015 * actions would just result in more rebalancing later, and ping-pong
4016 * tasks around. Thus we look for the minimum possible imbalance.
4017 * Negative imbalances (*we* are more loaded than anyone else) will
4018 * be counted as no imbalance for these purposes -- we can't fix that
4019 * by pulling tasks to us. Be careful of negative numbers as they'll
4020 * appear as very large values with unsigned longs.
4022 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4025 /* Looks like there is an imbalance. Compute it */
4026 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4031 * There is no obvious imbalance. But check if we can do some balancing
4034 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4042 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4045 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4046 unsigned long imbalance
, const struct cpumask
*cpus
)
4048 struct rq
*busiest
= NULL
, *rq
;
4049 unsigned long max_load
= 0;
4052 for_each_cpu(i
, sched_group_cpus(group
)) {
4053 unsigned long power
= power_of(i
);
4054 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4057 if (!cpumask_test_cpu(i
, cpus
))
4061 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4064 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4067 if (wl
> max_load
) {
4077 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4078 * so long as it is large enough.
4080 #define MAX_PINNED_INTERVAL 512
4082 /* Working cpumask for load_balance and load_balance_newidle. */
4083 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4086 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4087 * tasks if there is an imbalance.
4089 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4090 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4093 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4094 struct sched_group
*group
;
4095 unsigned long imbalance
;
4097 unsigned long flags
;
4098 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4100 cpumask_setall(cpus
);
4103 * When power savings policy is enabled for the parent domain, idle
4104 * sibling can pick up load irrespective of busy siblings. In this case,
4105 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4106 * portraying it as CPU_NOT_IDLE.
4108 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4109 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4112 schedstat_inc(sd
, lb_count
[idle
]);
4116 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4123 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4127 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4129 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4133 BUG_ON(busiest
== this_rq
);
4135 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4138 if (busiest
->nr_running
> 1) {
4140 * Attempt to move tasks. If find_busiest_group has found
4141 * an imbalance but busiest->nr_running <= 1, the group is
4142 * still unbalanced. ld_moved simply stays zero, so it is
4143 * correctly treated as an imbalance.
4145 local_irq_save(flags
);
4146 double_rq_lock(this_rq
, busiest
);
4147 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4148 imbalance
, sd
, idle
, &all_pinned
);
4149 double_rq_unlock(this_rq
, busiest
);
4150 local_irq_restore(flags
);
4153 * some other cpu did the load balance for us.
4155 if (ld_moved
&& this_cpu
!= smp_processor_id())
4156 resched_cpu(this_cpu
);
4158 /* All tasks on this runqueue were pinned by CPU affinity */
4159 if (unlikely(all_pinned
)) {
4160 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4161 if (!cpumask_empty(cpus
))
4168 schedstat_inc(sd
, lb_failed
[idle
]);
4169 sd
->nr_balance_failed
++;
4171 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4173 spin_lock_irqsave(&busiest
->lock
, flags
);
4175 /* don't kick the migration_thread, if the curr
4176 * task on busiest cpu can't be moved to this_cpu
4178 if (!cpumask_test_cpu(this_cpu
,
4179 &busiest
->curr
->cpus_allowed
)) {
4180 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4182 goto out_one_pinned
;
4185 if (!busiest
->active_balance
) {
4186 busiest
->active_balance
= 1;
4187 busiest
->push_cpu
= this_cpu
;
4190 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4192 wake_up_process(busiest
->migration_thread
);
4195 * We've kicked active balancing, reset the failure
4198 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4201 sd
->nr_balance_failed
= 0;
4203 if (likely(!active_balance
)) {
4204 /* We were unbalanced, so reset the balancing interval */
4205 sd
->balance_interval
= sd
->min_interval
;
4208 * If we've begun active balancing, start to back off. This
4209 * case may not be covered by the all_pinned logic if there
4210 * is only 1 task on the busy runqueue (because we don't call
4213 if (sd
->balance_interval
< sd
->max_interval
)
4214 sd
->balance_interval
*= 2;
4217 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4218 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4224 schedstat_inc(sd
, lb_balanced
[idle
]);
4226 sd
->nr_balance_failed
= 0;
4229 /* tune up the balancing interval */
4230 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4231 (sd
->balance_interval
< sd
->max_interval
))
4232 sd
->balance_interval
*= 2;
4234 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4235 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4246 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4247 * tasks if there is an imbalance.
4249 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4250 * this_rq is locked.
4253 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4255 struct sched_group
*group
;
4256 struct rq
*busiest
= NULL
;
4257 unsigned long imbalance
;
4261 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4263 cpumask_setall(cpus
);
4266 * When power savings policy is enabled for the parent domain, idle
4267 * sibling can pick up load irrespective of busy siblings. In this case,
4268 * let the state of idle sibling percolate up as IDLE, instead of
4269 * portraying it as CPU_NOT_IDLE.
4271 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4272 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4275 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4277 update_shares_locked(this_rq
, sd
);
4278 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4279 &sd_idle
, cpus
, NULL
);
4281 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4285 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4287 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4291 BUG_ON(busiest
== this_rq
);
4293 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4296 if (busiest
->nr_running
> 1) {
4297 /* Attempt to move tasks */
4298 double_lock_balance(this_rq
, busiest
);
4299 /* this_rq->clock is already updated */
4300 update_rq_clock(busiest
);
4301 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4302 imbalance
, sd
, CPU_NEWLY_IDLE
,
4304 double_unlock_balance(this_rq
, busiest
);
4306 if (unlikely(all_pinned
)) {
4307 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4308 if (!cpumask_empty(cpus
))
4314 int active_balance
= 0;
4316 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4317 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4318 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4321 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4324 if (sd
->nr_balance_failed
++ < 2)
4328 * The only task running in a non-idle cpu can be moved to this
4329 * cpu in an attempt to completely freeup the other CPU
4330 * package. The same method used to move task in load_balance()
4331 * have been extended for load_balance_newidle() to speedup
4332 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4334 * The package power saving logic comes from
4335 * find_busiest_group(). If there are no imbalance, then
4336 * f_b_g() will return NULL. However when sched_mc={1,2} then
4337 * f_b_g() will select a group from which a running task may be
4338 * pulled to this cpu in order to make the other package idle.
4339 * If there is no opportunity to make a package idle and if
4340 * there are no imbalance, then f_b_g() will return NULL and no
4341 * action will be taken in load_balance_newidle().
4343 * Under normal task pull operation due to imbalance, there
4344 * will be more than one task in the source run queue and
4345 * move_tasks() will succeed. ld_moved will be true and this
4346 * active balance code will not be triggered.
4349 /* Lock busiest in correct order while this_rq is held */
4350 double_lock_balance(this_rq
, busiest
);
4353 * don't kick the migration_thread, if the curr
4354 * task on busiest cpu can't be moved to this_cpu
4356 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4357 double_unlock_balance(this_rq
, busiest
);
4362 if (!busiest
->active_balance
) {
4363 busiest
->active_balance
= 1;
4364 busiest
->push_cpu
= this_cpu
;
4368 double_unlock_balance(this_rq
, busiest
);
4370 * Should not call ttwu while holding a rq->lock
4372 spin_unlock(&this_rq
->lock
);
4374 wake_up_process(busiest
->migration_thread
);
4375 spin_lock(&this_rq
->lock
);
4378 sd
->nr_balance_failed
= 0;
4380 update_shares_locked(this_rq
, sd
);
4384 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4385 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4386 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4388 sd
->nr_balance_failed
= 0;
4394 * idle_balance is called by schedule() if this_cpu is about to become
4395 * idle. Attempts to pull tasks from other CPUs.
4397 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4399 struct sched_domain
*sd
;
4400 int pulled_task
= 0;
4401 unsigned long next_balance
= jiffies
+ HZ
;
4403 for_each_domain(this_cpu
, sd
) {
4404 unsigned long interval
;
4406 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4409 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4410 /* If we've pulled tasks over stop searching: */
4411 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4414 interval
= msecs_to_jiffies(sd
->balance_interval
);
4415 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4416 next_balance
= sd
->last_balance
+ interval
;
4420 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4422 * We are going idle. next_balance may be set based on
4423 * a busy processor. So reset next_balance.
4425 this_rq
->next_balance
= next_balance
;
4430 * active_load_balance is run by migration threads. It pushes running tasks
4431 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4432 * running on each physical CPU where possible, and avoids physical /
4433 * logical imbalances.
4435 * Called with busiest_rq locked.
4437 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4439 int target_cpu
= busiest_rq
->push_cpu
;
4440 struct sched_domain
*sd
;
4441 struct rq
*target_rq
;
4443 /* Is there any task to move? */
4444 if (busiest_rq
->nr_running
<= 1)
4447 target_rq
= cpu_rq(target_cpu
);
4450 * This condition is "impossible", if it occurs
4451 * we need to fix it. Originally reported by
4452 * Bjorn Helgaas on a 128-cpu setup.
4454 BUG_ON(busiest_rq
== target_rq
);
4456 /* move a task from busiest_rq to target_rq */
4457 double_lock_balance(busiest_rq
, target_rq
);
4458 update_rq_clock(busiest_rq
);
4459 update_rq_clock(target_rq
);
4461 /* Search for an sd spanning us and the target CPU. */
4462 for_each_domain(target_cpu
, sd
) {
4463 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4464 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4469 schedstat_inc(sd
, alb_count
);
4471 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4473 schedstat_inc(sd
, alb_pushed
);
4475 schedstat_inc(sd
, alb_failed
);
4477 double_unlock_balance(busiest_rq
, target_rq
);
4482 atomic_t load_balancer
;
4483 cpumask_var_t cpu_mask
;
4484 cpumask_var_t ilb_grp_nohz_mask
;
4485 } nohz ____cacheline_aligned
= {
4486 .load_balancer
= ATOMIC_INIT(-1),
4489 int get_nohz_load_balancer(void)
4491 return atomic_read(&nohz
.load_balancer
);
4494 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4496 * lowest_flag_domain - Return lowest sched_domain containing flag.
4497 * @cpu: The cpu whose lowest level of sched domain is to
4499 * @flag: The flag to check for the lowest sched_domain
4500 * for the given cpu.
4502 * Returns the lowest sched_domain of a cpu which contains the given flag.
4504 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4506 struct sched_domain
*sd
;
4508 for_each_domain(cpu
, sd
)
4509 if (sd
&& (sd
->flags
& flag
))
4516 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4517 * @cpu: The cpu whose domains we're iterating over.
4518 * @sd: variable holding the value of the power_savings_sd
4520 * @flag: The flag to filter the sched_domains to be iterated.
4522 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4523 * set, starting from the lowest sched_domain to the highest.
4525 #define for_each_flag_domain(cpu, sd, flag) \
4526 for (sd = lowest_flag_domain(cpu, flag); \
4527 (sd && (sd->flags & flag)); sd = sd->parent)
4530 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4531 * @ilb_group: group to be checked for semi-idleness
4533 * Returns: 1 if the group is semi-idle. 0 otherwise.
4535 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4536 * and atleast one non-idle CPU. This helper function checks if the given
4537 * sched_group is semi-idle or not.
4539 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4541 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4542 sched_group_cpus(ilb_group
));
4545 * A sched_group is semi-idle when it has atleast one busy cpu
4546 * and atleast one idle cpu.
4548 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4551 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4557 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4558 * @cpu: The cpu which is nominating a new idle_load_balancer.
4560 * Returns: Returns the id of the idle load balancer if it exists,
4561 * Else, returns >= nr_cpu_ids.
4563 * This algorithm picks the idle load balancer such that it belongs to a
4564 * semi-idle powersavings sched_domain. The idea is to try and avoid
4565 * completely idle packages/cores just for the purpose of idle load balancing
4566 * when there are other idle cpu's which are better suited for that job.
4568 static int find_new_ilb(int cpu
)
4570 struct sched_domain
*sd
;
4571 struct sched_group
*ilb_group
;
4574 * Have idle load balancer selection from semi-idle packages only
4575 * when power-aware load balancing is enabled
4577 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4581 * Optimize for the case when we have no idle CPUs or only one
4582 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4584 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4587 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4588 ilb_group
= sd
->groups
;
4591 if (is_semi_idle_group(ilb_group
))
4592 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4594 ilb_group
= ilb_group
->next
;
4596 } while (ilb_group
!= sd
->groups
);
4600 return cpumask_first(nohz
.cpu_mask
);
4602 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4603 static inline int find_new_ilb(int call_cpu
)
4605 return cpumask_first(nohz
.cpu_mask
);
4610 * This routine will try to nominate the ilb (idle load balancing)
4611 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4612 * load balancing on behalf of all those cpus. If all the cpus in the system
4613 * go into this tickless mode, then there will be no ilb owner (as there is
4614 * no need for one) and all the cpus will sleep till the next wakeup event
4617 * For the ilb owner, tick is not stopped. And this tick will be used
4618 * for idle load balancing. ilb owner will still be part of
4621 * While stopping the tick, this cpu will become the ilb owner if there
4622 * is no other owner. And will be the owner till that cpu becomes busy
4623 * or if all cpus in the system stop their ticks at which point
4624 * there is no need for ilb owner.
4626 * When the ilb owner becomes busy, it nominates another owner, during the
4627 * next busy scheduler_tick()
4629 int select_nohz_load_balancer(int stop_tick
)
4631 int cpu
= smp_processor_id();
4634 cpu_rq(cpu
)->in_nohz_recently
= 1;
4636 if (!cpu_active(cpu
)) {
4637 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4641 * If we are going offline and still the leader,
4644 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4650 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4652 /* time for ilb owner also to sleep */
4653 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4654 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4655 atomic_set(&nohz
.load_balancer
, -1);
4659 if (atomic_read(&nohz
.load_balancer
) == -1) {
4660 /* make me the ilb owner */
4661 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4663 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4666 if (!(sched_smt_power_savings
||
4667 sched_mc_power_savings
))
4670 * Check to see if there is a more power-efficient
4673 new_ilb
= find_new_ilb(cpu
);
4674 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4675 atomic_set(&nohz
.load_balancer
, -1);
4676 resched_cpu(new_ilb
);
4682 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4685 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4687 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4688 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4695 static DEFINE_SPINLOCK(balancing
);
4698 * It checks each scheduling domain to see if it is due to be balanced,
4699 * and initiates a balancing operation if so.
4701 * Balancing parameters are set up in arch_init_sched_domains.
4703 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4706 struct rq
*rq
= cpu_rq(cpu
);
4707 unsigned long interval
;
4708 struct sched_domain
*sd
;
4709 /* Earliest time when we have to do rebalance again */
4710 unsigned long next_balance
= jiffies
+ 60*HZ
;
4711 int update_next_balance
= 0;
4714 for_each_domain(cpu
, sd
) {
4715 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4718 interval
= sd
->balance_interval
;
4719 if (idle
!= CPU_IDLE
)
4720 interval
*= sd
->busy_factor
;
4722 /* scale ms to jiffies */
4723 interval
= msecs_to_jiffies(interval
);
4724 if (unlikely(!interval
))
4726 if (interval
> HZ
*NR_CPUS
/10)
4727 interval
= HZ
*NR_CPUS
/10;
4729 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4731 if (need_serialize
) {
4732 if (!spin_trylock(&balancing
))
4736 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4737 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4739 * We've pulled tasks over so either we're no
4740 * longer idle, or one of our SMT siblings is
4743 idle
= CPU_NOT_IDLE
;
4745 sd
->last_balance
= jiffies
;
4748 spin_unlock(&balancing
);
4750 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4751 next_balance
= sd
->last_balance
+ interval
;
4752 update_next_balance
= 1;
4756 * Stop the load balance at this level. There is another
4757 * CPU in our sched group which is doing load balancing more
4765 * next_balance will be updated only when there is a need.
4766 * When the cpu is attached to null domain for ex, it will not be
4769 if (likely(update_next_balance
))
4770 rq
->next_balance
= next_balance
;
4774 * run_rebalance_domains is triggered when needed from the scheduler tick.
4775 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4776 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4778 static void run_rebalance_domains(struct softirq_action
*h
)
4780 int this_cpu
= smp_processor_id();
4781 struct rq
*this_rq
= cpu_rq(this_cpu
);
4782 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4783 CPU_IDLE
: CPU_NOT_IDLE
;
4785 rebalance_domains(this_cpu
, idle
);
4789 * If this cpu is the owner for idle load balancing, then do the
4790 * balancing on behalf of the other idle cpus whose ticks are
4793 if (this_rq
->idle_at_tick
&&
4794 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4798 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4799 if (balance_cpu
== this_cpu
)
4803 * If this cpu gets work to do, stop the load balancing
4804 * work being done for other cpus. Next load
4805 * balancing owner will pick it up.
4810 rebalance_domains(balance_cpu
, CPU_IDLE
);
4812 rq
= cpu_rq(balance_cpu
);
4813 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4814 this_rq
->next_balance
= rq
->next_balance
;
4820 static inline int on_null_domain(int cpu
)
4822 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4826 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4828 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4829 * idle load balancing owner or decide to stop the periodic load balancing,
4830 * if the whole system is idle.
4832 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4836 * If we were in the nohz mode recently and busy at the current
4837 * scheduler tick, then check if we need to nominate new idle
4840 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4841 rq
->in_nohz_recently
= 0;
4843 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4844 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4845 atomic_set(&nohz
.load_balancer
, -1);
4848 if (atomic_read(&nohz
.load_balancer
) == -1) {
4849 int ilb
= find_new_ilb(cpu
);
4851 if (ilb
< nr_cpu_ids
)
4857 * If this cpu is idle and doing idle load balancing for all the
4858 * cpus with ticks stopped, is it time for that to stop?
4860 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4861 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4867 * If this cpu is idle and the idle load balancing is done by
4868 * someone else, then no need raise the SCHED_SOFTIRQ
4870 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4871 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4874 /* Don't need to rebalance while attached to NULL domain */
4875 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4876 likely(!on_null_domain(cpu
)))
4877 raise_softirq(SCHED_SOFTIRQ
);
4880 #else /* CONFIG_SMP */
4883 * on UP we do not need to balance between CPUs:
4885 static inline void idle_balance(int cpu
, struct rq
*rq
)
4891 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4893 EXPORT_PER_CPU_SYMBOL(kstat
);
4896 * Return any ns on the sched_clock that have not yet been accounted in
4897 * @p in case that task is currently running.
4899 * Called with task_rq_lock() held on @rq.
4901 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4905 if (task_current(rq
, p
)) {
4906 update_rq_clock(rq
);
4907 ns
= rq
->clock
- p
->se
.exec_start
;
4915 unsigned long long task_delta_exec(struct task_struct
*p
)
4917 unsigned long flags
;
4921 rq
= task_rq_lock(p
, &flags
);
4922 ns
= do_task_delta_exec(p
, rq
);
4923 task_rq_unlock(rq
, &flags
);
4929 * Return accounted runtime for the task.
4930 * In case the task is currently running, return the runtime plus current's
4931 * pending runtime that have not been accounted yet.
4933 unsigned long long task_sched_runtime(struct task_struct
*p
)
4935 unsigned long flags
;
4939 rq
= task_rq_lock(p
, &flags
);
4940 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4941 task_rq_unlock(rq
, &flags
);
4947 * Return sum_exec_runtime for the thread group.
4948 * In case the task is currently running, return the sum plus current's
4949 * pending runtime that have not been accounted yet.
4951 * Note that the thread group might have other running tasks as well,
4952 * so the return value not includes other pending runtime that other
4953 * running tasks might have.
4955 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4957 struct task_cputime totals
;
4958 unsigned long flags
;
4962 rq
= task_rq_lock(p
, &flags
);
4963 thread_group_cputime(p
, &totals
);
4964 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4965 task_rq_unlock(rq
, &flags
);
4971 * Account user cpu time to a process.
4972 * @p: the process that the cpu time gets accounted to
4973 * @cputime: the cpu time spent in user space since the last update
4974 * @cputime_scaled: cputime scaled by cpu frequency
4976 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4977 cputime_t cputime_scaled
)
4979 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4982 /* Add user time to process. */
4983 p
->utime
= cputime_add(p
->utime
, cputime
);
4984 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4985 account_group_user_time(p
, cputime
);
4987 /* Add user time to cpustat. */
4988 tmp
= cputime_to_cputime64(cputime
);
4989 if (TASK_NICE(p
) > 0)
4990 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4992 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4994 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4995 /* Account for user time used */
4996 acct_update_integrals(p
);
5000 * Account guest cpu time to a process.
5001 * @p: the process that the cpu time gets accounted to
5002 * @cputime: the cpu time spent in virtual machine since the last update
5003 * @cputime_scaled: cputime scaled by cpu frequency
5005 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5006 cputime_t cputime_scaled
)
5009 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5011 tmp
= cputime_to_cputime64(cputime
);
5013 /* Add guest time to process. */
5014 p
->utime
= cputime_add(p
->utime
, cputime
);
5015 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5016 account_group_user_time(p
, cputime
);
5017 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5019 /* Add guest time to cpustat. */
5020 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5021 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5025 * Account system cpu time to a process.
5026 * @p: the process that the cpu time gets accounted to
5027 * @hardirq_offset: the offset to subtract from hardirq_count()
5028 * @cputime: the cpu time spent in kernel space since the last update
5029 * @cputime_scaled: cputime scaled by cpu frequency
5031 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5032 cputime_t cputime
, cputime_t cputime_scaled
)
5034 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5037 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5038 account_guest_time(p
, cputime
, cputime_scaled
);
5042 /* Add system time to process. */
5043 p
->stime
= cputime_add(p
->stime
, cputime
);
5044 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5045 account_group_system_time(p
, cputime
);
5047 /* Add system time to cpustat. */
5048 tmp
= cputime_to_cputime64(cputime
);
5049 if (hardirq_count() - hardirq_offset
)
5050 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5051 else if (softirq_count())
5052 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5054 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5056 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5058 /* Account for system time used */
5059 acct_update_integrals(p
);
5063 * Account for involuntary wait time.
5064 * @steal: the cpu time spent in involuntary wait
5066 void account_steal_time(cputime_t cputime
)
5068 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5069 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5071 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5075 * Account for idle time.
5076 * @cputime: the cpu time spent in idle wait
5078 void account_idle_time(cputime_t cputime
)
5080 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5081 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5082 struct rq
*rq
= this_rq();
5084 if (atomic_read(&rq
->nr_iowait
) > 0)
5085 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5087 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5090 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5093 * Account a single tick of cpu time.
5094 * @p: the process that the cpu time gets accounted to
5095 * @user_tick: indicates if the tick is a user or a system tick
5097 void account_process_tick(struct task_struct
*p
, int user_tick
)
5099 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5100 struct rq
*rq
= this_rq();
5103 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5104 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5105 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5108 account_idle_time(cputime_one_jiffy
);
5112 * Account multiple ticks of steal time.
5113 * @p: the process from which the cpu time has been stolen
5114 * @ticks: number of stolen ticks
5116 void account_steal_ticks(unsigned long ticks
)
5118 account_steal_time(jiffies_to_cputime(ticks
));
5122 * Account multiple ticks of idle time.
5123 * @ticks: number of stolen ticks
5125 void account_idle_ticks(unsigned long ticks
)
5127 account_idle_time(jiffies_to_cputime(ticks
));
5133 * Use precise platform statistics if available:
5135 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5136 cputime_t
task_utime(struct task_struct
*p
)
5141 cputime_t
task_stime(struct task_struct
*p
)
5146 cputime_t
task_utime(struct task_struct
*p
)
5148 clock_t utime
= cputime_to_clock_t(p
->utime
),
5149 total
= utime
+ cputime_to_clock_t(p
->stime
);
5153 * Use CFS's precise accounting:
5155 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5159 do_div(temp
, total
);
5161 utime
= (clock_t)temp
;
5163 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5164 return p
->prev_utime
;
5167 cputime_t
task_stime(struct task_struct
*p
)
5172 * Use CFS's precise accounting. (we subtract utime from
5173 * the total, to make sure the total observed by userspace
5174 * grows monotonically - apps rely on that):
5176 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5177 cputime_to_clock_t(task_utime(p
));
5180 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5182 return p
->prev_stime
;
5186 inline cputime_t
task_gtime(struct task_struct
*p
)
5192 * This function gets called by the timer code, with HZ frequency.
5193 * We call it with interrupts disabled.
5195 * It also gets called by the fork code, when changing the parent's
5198 void scheduler_tick(void)
5200 int cpu
= smp_processor_id();
5201 struct rq
*rq
= cpu_rq(cpu
);
5202 struct task_struct
*curr
= rq
->curr
;
5206 spin_lock(&rq
->lock
);
5207 update_rq_clock(rq
);
5208 update_cpu_load(rq
);
5209 curr
->sched_class
->task_tick(rq
, curr
, 0);
5210 spin_unlock(&rq
->lock
);
5212 perf_event_task_tick(curr
, cpu
);
5215 rq
->idle_at_tick
= idle_cpu(cpu
);
5216 trigger_load_balance(rq
, cpu
);
5220 notrace
unsigned long get_parent_ip(unsigned long addr
)
5222 if (in_lock_functions(addr
)) {
5223 addr
= CALLER_ADDR2
;
5224 if (in_lock_functions(addr
))
5225 addr
= CALLER_ADDR3
;
5230 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5231 defined(CONFIG_PREEMPT_TRACER))
5233 void __kprobes
add_preempt_count(int val
)
5235 #ifdef CONFIG_DEBUG_PREEMPT
5239 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5242 preempt_count() += val
;
5243 #ifdef CONFIG_DEBUG_PREEMPT
5245 * Spinlock count overflowing soon?
5247 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5250 if (preempt_count() == val
)
5251 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5253 EXPORT_SYMBOL(add_preempt_count
);
5255 void __kprobes
sub_preempt_count(int val
)
5257 #ifdef CONFIG_DEBUG_PREEMPT
5261 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5264 * Is the spinlock portion underflowing?
5266 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5267 !(preempt_count() & PREEMPT_MASK
)))
5271 if (preempt_count() == val
)
5272 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5273 preempt_count() -= val
;
5275 EXPORT_SYMBOL(sub_preempt_count
);
5280 * Print scheduling while atomic bug:
5282 static noinline
void __schedule_bug(struct task_struct
*prev
)
5284 struct pt_regs
*regs
= get_irq_regs();
5286 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5287 prev
->comm
, prev
->pid
, preempt_count());
5289 debug_show_held_locks(prev
);
5291 if (irqs_disabled())
5292 print_irqtrace_events(prev
);
5301 * Various schedule()-time debugging checks and statistics:
5303 static inline void schedule_debug(struct task_struct
*prev
)
5306 * Test if we are atomic. Since do_exit() needs to call into
5307 * schedule() atomically, we ignore that path for now.
5308 * Otherwise, whine if we are scheduling when we should not be.
5310 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5311 __schedule_bug(prev
);
5313 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5315 schedstat_inc(this_rq(), sched_count
);
5316 #ifdef CONFIG_SCHEDSTATS
5317 if (unlikely(prev
->lock_depth
>= 0)) {
5318 schedstat_inc(this_rq(), bkl_count
);
5319 schedstat_inc(prev
, sched_info
.bkl_count
);
5324 static void put_prev_task(struct rq
*rq
, struct task_struct
*p
)
5326 u64 runtime
= p
->se
.sum_exec_runtime
- p
->se
.prev_sum_exec_runtime
;
5328 update_avg(&p
->se
.avg_running
, runtime
);
5330 if (p
->state
== TASK_RUNNING
) {
5332 * In order to avoid avg_overlap growing stale when we are
5333 * indeed overlapping and hence not getting put to sleep, grow
5334 * the avg_overlap on preemption.
5336 * We use the average preemption runtime because that
5337 * correlates to the amount of cache footprint a task can
5340 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5341 update_avg(&p
->se
.avg_overlap
, runtime
);
5343 update_avg(&p
->se
.avg_running
, 0);
5345 p
->sched_class
->put_prev_task(rq
, p
);
5349 * Pick up the highest-prio task:
5351 static inline struct task_struct
*
5352 pick_next_task(struct rq
*rq
)
5354 const struct sched_class
*class;
5355 struct task_struct
*p
;
5358 * Optimization: we know that if all tasks are in
5359 * the fair class we can call that function directly:
5361 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5362 p
= fair_sched_class
.pick_next_task(rq
);
5367 class = sched_class_highest
;
5369 p
= class->pick_next_task(rq
);
5373 * Will never be NULL as the idle class always
5374 * returns a non-NULL p:
5376 class = class->next
;
5381 * schedule() is the main scheduler function.
5383 asmlinkage
void __sched
schedule(void)
5385 struct task_struct
*prev
, *next
;
5386 unsigned long *switch_count
;
5392 cpu
= smp_processor_id();
5396 switch_count
= &prev
->nivcsw
;
5398 release_kernel_lock(prev
);
5399 need_resched_nonpreemptible
:
5401 schedule_debug(prev
);
5403 if (sched_feat(HRTICK
))
5406 spin_lock_irq(&rq
->lock
);
5407 update_rq_clock(rq
);
5408 clear_tsk_need_resched(prev
);
5410 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5411 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5412 prev
->state
= TASK_RUNNING
;
5414 deactivate_task(rq
, prev
, 1);
5415 switch_count
= &prev
->nvcsw
;
5418 pre_schedule(rq
, prev
);
5420 if (unlikely(!rq
->nr_running
))
5421 idle_balance(cpu
, rq
);
5423 put_prev_task(rq
, prev
);
5424 next
= pick_next_task(rq
);
5426 if (likely(prev
!= next
)) {
5427 sched_info_switch(prev
, next
);
5428 perf_event_task_sched_out(prev
, next
, cpu
);
5434 context_switch(rq
, prev
, next
); /* unlocks the rq */
5436 * the context switch might have flipped the stack from under
5437 * us, hence refresh the local variables.
5439 cpu
= smp_processor_id();
5442 spin_unlock_irq(&rq
->lock
);
5446 if (unlikely(reacquire_kernel_lock(current
) < 0))
5447 goto need_resched_nonpreemptible
;
5449 preempt_enable_no_resched();
5453 EXPORT_SYMBOL(schedule
);
5457 * Look out! "owner" is an entirely speculative pointer
5458 * access and not reliable.
5460 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5465 if (!sched_feat(OWNER_SPIN
))
5468 #ifdef CONFIG_DEBUG_PAGEALLOC
5470 * Need to access the cpu field knowing that
5471 * DEBUG_PAGEALLOC could have unmapped it if
5472 * the mutex owner just released it and exited.
5474 if (probe_kernel_address(&owner
->cpu
, cpu
))
5481 * Even if the access succeeded (likely case),
5482 * the cpu field may no longer be valid.
5484 if (cpu
>= nr_cpumask_bits
)
5488 * We need to validate that we can do a
5489 * get_cpu() and that we have the percpu area.
5491 if (!cpu_online(cpu
))
5498 * Owner changed, break to re-assess state.
5500 if (lock
->owner
!= owner
)
5504 * Is that owner really running on that cpu?
5506 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5516 #ifdef CONFIG_PREEMPT
5518 * this is the entry point to schedule() from in-kernel preemption
5519 * off of preempt_enable. Kernel preemptions off return from interrupt
5520 * occur there and call schedule directly.
5522 asmlinkage
void __sched
preempt_schedule(void)
5524 struct thread_info
*ti
= current_thread_info();
5527 * If there is a non-zero preempt_count or interrupts are disabled,
5528 * we do not want to preempt the current task. Just return..
5530 if (likely(ti
->preempt_count
|| irqs_disabled()))
5534 add_preempt_count(PREEMPT_ACTIVE
);
5536 sub_preempt_count(PREEMPT_ACTIVE
);
5539 * Check again in case we missed a preemption opportunity
5540 * between schedule and now.
5543 } while (need_resched());
5545 EXPORT_SYMBOL(preempt_schedule
);
5548 * this is the entry point to schedule() from kernel preemption
5549 * off of irq context.
5550 * Note, that this is called and return with irqs disabled. This will
5551 * protect us against recursive calling from irq.
5553 asmlinkage
void __sched
preempt_schedule_irq(void)
5555 struct thread_info
*ti
= current_thread_info();
5557 /* Catch callers which need to be fixed */
5558 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5561 add_preempt_count(PREEMPT_ACTIVE
);
5564 local_irq_disable();
5565 sub_preempt_count(PREEMPT_ACTIVE
);
5568 * Check again in case we missed a preemption opportunity
5569 * between schedule and now.
5572 } while (need_resched());
5575 #endif /* CONFIG_PREEMPT */
5577 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5580 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5582 EXPORT_SYMBOL(default_wake_function
);
5585 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5586 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5587 * number) then we wake all the non-exclusive tasks and one exclusive task.
5589 * There are circumstances in which we can try to wake a task which has already
5590 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5591 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5593 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5594 int nr_exclusive
, int wake_flags
, void *key
)
5596 wait_queue_t
*curr
, *next
;
5598 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5599 unsigned flags
= curr
->flags
;
5601 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5602 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5608 * __wake_up - wake up threads blocked on a waitqueue.
5610 * @mode: which threads
5611 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5612 * @key: is directly passed to the wakeup function
5614 * It may be assumed that this function implies a write memory barrier before
5615 * changing the task state if and only if any tasks are woken up.
5617 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5618 int nr_exclusive
, void *key
)
5620 unsigned long flags
;
5622 spin_lock_irqsave(&q
->lock
, flags
);
5623 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5624 spin_unlock_irqrestore(&q
->lock
, flags
);
5626 EXPORT_SYMBOL(__wake_up
);
5629 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5631 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5633 __wake_up_common(q
, mode
, 1, 0, NULL
);
5636 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5638 __wake_up_common(q
, mode
, 1, 0, key
);
5642 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5644 * @mode: which threads
5645 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5646 * @key: opaque value to be passed to wakeup targets
5648 * The sync wakeup differs that the waker knows that it will schedule
5649 * away soon, so while the target thread will be woken up, it will not
5650 * be migrated to another CPU - ie. the two threads are 'synchronized'
5651 * with each other. This can prevent needless bouncing between CPUs.
5653 * On UP it can prevent extra preemption.
5655 * It may be assumed that this function implies a write memory barrier before
5656 * changing the task state if and only if any tasks are woken up.
5658 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5659 int nr_exclusive
, void *key
)
5661 unsigned long flags
;
5662 int wake_flags
= WF_SYNC
;
5667 if (unlikely(!nr_exclusive
))
5670 spin_lock_irqsave(&q
->lock
, flags
);
5671 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5672 spin_unlock_irqrestore(&q
->lock
, flags
);
5674 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5677 * __wake_up_sync - see __wake_up_sync_key()
5679 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5681 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5683 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5686 * complete: - signals a single thread waiting on this completion
5687 * @x: holds the state of this particular completion
5689 * This will wake up a single thread waiting on this completion. Threads will be
5690 * awakened in the same order in which they were queued.
5692 * See also complete_all(), wait_for_completion() and related routines.
5694 * It may be assumed that this function implies a write memory barrier before
5695 * changing the task state if and only if any tasks are woken up.
5697 void complete(struct completion
*x
)
5699 unsigned long flags
;
5701 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5703 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5704 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5706 EXPORT_SYMBOL(complete
);
5709 * complete_all: - signals all threads waiting on this completion
5710 * @x: holds the state of this particular completion
5712 * This will wake up all threads waiting on this particular completion event.
5714 * It may be assumed that this function implies a write memory barrier before
5715 * changing the task state if and only if any tasks are woken up.
5717 void complete_all(struct completion
*x
)
5719 unsigned long flags
;
5721 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5722 x
->done
+= UINT_MAX
/2;
5723 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5724 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5726 EXPORT_SYMBOL(complete_all
);
5728 static inline long __sched
5729 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5732 DECLARE_WAITQUEUE(wait
, current
);
5734 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5735 __add_wait_queue_tail(&x
->wait
, &wait
);
5737 if (signal_pending_state(state
, current
)) {
5738 timeout
= -ERESTARTSYS
;
5741 __set_current_state(state
);
5742 spin_unlock_irq(&x
->wait
.lock
);
5743 timeout
= schedule_timeout(timeout
);
5744 spin_lock_irq(&x
->wait
.lock
);
5745 } while (!x
->done
&& timeout
);
5746 __remove_wait_queue(&x
->wait
, &wait
);
5751 return timeout
?: 1;
5755 wait_for_common(struct completion
*x
, long timeout
, int state
)
5759 spin_lock_irq(&x
->wait
.lock
);
5760 timeout
= do_wait_for_common(x
, timeout
, state
);
5761 spin_unlock_irq(&x
->wait
.lock
);
5766 * wait_for_completion: - waits for completion of a task
5767 * @x: holds the state of this particular completion
5769 * This waits to be signaled for completion of a specific task. It is NOT
5770 * interruptible and there is no timeout.
5772 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5773 * and interrupt capability. Also see complete().
5775 void __sched
wait_for_completion(struct completion
*x
)
5777 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5779 EXPORT_SYMBOL(wait_for_completion
);
5782 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5783 * @x: holds the state of this particular completion
5784 * @timeout: timeout value in jiffies
5786 * This waits for either a completion of a specific task to be signaled or for a
5787 * specified timeout to expire. The timeout is in jiffies. It is not
5790 unsigned long __sched
5791 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5793 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5795 EXPORT_SYMBOL(wait_for_completion_timeout
);
5798 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5799 * @x: holds the state of this particular completion
5801 * This waits for completion of a specific task to be signaled. It is
5804 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5806 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5807 if (t
== -ERESTARTSYS
)
5811 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5814 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5815 * @x: holds the state of this particular completion
5816 * @timeout: timeout value in jiffies
5818 * This waits for either a completion of a specific task to be signaled or for a
5819 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5821 unsigned long __sched
5822 wait_for_completion_interruptible_timeout(struct completion
*x
,
5823 unsigned long timeout
)
5825 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5827 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5830 * wait_for_completion_killable: - waits for completion of a task (killable)
5831 * @x: holds the state of this particular completion
5833 * This waits to be signaled for completion of a specific task. It can be
5834 * interrupted by a kill signal.
5836 int __sched
wait_for_completion_killable(struct completion
*x
)
5838 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5839 if (t
== -ERESTARTSYS
)
5843 EXPORT_SYMBOL(wait_for_completion_killable
);
5846 * try_wait_for_completion - try to decrement a completion without blocking
5847 * @x: completion structure
5849 * Returns: 0 if a decrement cannot be done without blocking
5850 * 1 if a decrement succeeded.
5852 * If a completion is being used as a counting completion,
5853 * attempt to decrement the counter without blocking. This
5854 * enables us to avoid waiting if the resource the completion
5855 * is protecting is not available.
5857 bool try_wait_for_completion(struct completion
*x
)
5861 spin_lock_irq(&x
->wait
.lock
);
5866 spin_unlock_irq(&x
->wait
.lock
);
5869 EXPORT_SYMBOL(try_wait_for_completion
);
5872 * completion_done - Test to see if a completion has any waiters
5873 * @x: completion structure
5875 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5876 * 1 if there are no waiters.
5879 bool completion_done(struct completion
*x
)
5883 spin_lock_irq(&x
->wait
.lock
);
5886 spin_unlock_irq(&x
->wait
.lock
);
5889 EXPORT_SYMBOL(completion_done
);
5892 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5894 unsigned long flags
;
5897 init_waitqueue_entry(&wait
, current
);
5899 __set_current_state(state
);
5901 spin_lock_irqsave(&q
->lock
, flags
);
5902 __add_wait_queue(q
, &wait
);
5903 spin_unlock(&q
->lock
);
5904 timeout
= schedule_timeout(timeout
);
5905 spin_lock_irq(&q
->lock
);
5906 __remove_wait_queue(q
, &wait
);
5907 spin_unlock_irqrestore(&q
->lock
, flags
);
5912 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5914 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5916 EXPORT_SYMBOL(interruptible_sleep_on
);
5919 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5921 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5923 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5925 void __sched
sleep_on(wait_queue_head_t
*q
)
5927 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5929 EXPORT_SYMBOL(sleep_on
);
5931 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5933 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5935 EXPORT_SYMBOL(sleep_on_timeout
);
5937 #ifdef CONFIG_RT_MUTEXES
5940 * rt_mutex_setprio - set the current priority of a task
5942 * @prio: prio value (kernel-internal form)
5944 * This function changes the 'effective' priority of a task. It does
5945 * not touch ->normal_prio like __setscheduler().
5947 * Used by the rt_mutex code to implement priority inheritance logic.
5949 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5951 unsigned long flags
;
5952 int oldprio
, on_rq
, running
;
5954 const struct sched_class
*prev_class
= p
->sched_class
;
5956 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5958 rq
= task_rq_lock(p
, &flags
);
5959 update_rq_clock(rq
);
5962 on_rq
= p
->se
.on_rq
;
5963 running
= task_current(rq
, p
);
5965 dequeue_task(rq
, p
, 0);
5967 p
->sched_class
->put_prev_task(rq
, p
);
5970 p
->sched_class
= &rt_sched_class
;
5972 p
->sched_class
= &fair_sched_class
;
5977 p
->sched_class
->set_curr_task(rq
);
5979 enqueue_task(rq
, p
, 0);
5981 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5983 task_rq_unlock(rq
, &flags
);
5988 void set_user_nice(struct task_struct
*p
, long nice
)
5990 int old_prio
, delta
, on_rq
;
5991 unsigned long flags
;
5994 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5997 * We have to be careful, if called from sys_setpriority(),
5998 * the task might be in the middle of scheduling on another CPU.
6000 rq
= task_rq_lock(p
, &flags
);
6001 update_rq_clock(rq
);
6003 * The RT priorities are set via sched_setscheduler(), but we still
6004 * allow the 'normal' nice value to be set - but as expected
6005 * it wont have any effect on scheduling until the task is
6006 * SCHED_FIFO/SCHED_RR:
6008 if (task_has_rt_policy(p
)) {
6009 p
->static_prio
= NICE_TO_PRIO(nice
);
6012 on_rq
= p
->se
.on_rq
;
6014 dequeue_task(rq
, p
, 0);
6016 p
->static_prio
= NICE_TO_PRIO(nice
);
6019 p
->prio
= effective_prio(p
);
6020 delta
= p
->prio
- old_prio
;
6023 enqueue_task(rq
, p
, 0);
6025 * If the task increased its priority or is running and
6026 * lowered its priority, then reschedule its CPU:
6028 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6029 resched_task(rq
->curr
);
6032 task_rq_unlock(rq
, &flags
);
6034 EXPORT_SYMBOL(set_user_nice
);
6037 * can_nice - check if a task can reduce its nice value
6041 int can_nice(const struct task_struct
*p
, const int nice
)
6043 /* convert nice value [19,-20] to rlimit style value [1,40] */
6044 int nice_rlim
= 20 - nice
;
6046 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6047 capable(CAP_SYS_NICE
));
6050 #ifdef __ARCH_WANT_SYS_NICE
6053 * sys_nice - change the priority of the current process.
6054 * @increment: priority increment
6056 * sys_setpriority is a more generic, but much slower function that
6057 * does similar things.
6059 SYSCALL_DEFINE1(nice
, int, increment
)
6064 * Setpriority might change our priority at the same moment.
6065 * We don't have to worry. Conceptually one call occurs first
6066 * and we have a single winner.
6068 if (increment
< -40)
6073 nice
= TASK_NICE(current
) + increment
;
6079 if (increment
< 0 && !can_nice(current
, nice
))
6082 retval
= security_task_setnice(current
, nice
);
6086 set_user_nice(current
, nice
);
6093 * task_prio - return the priority value of a given task.
6094 * @p: the task in question.
6096 * This is the priority value as seen by users in /proc.
6097 * RT tasks are offset by -200. Normal tasks are centered
6098 * around 0, value goes from -16 to +15.
6100 int task_prio(const struct task_struct
*p
)
6102 return p
->prio
- MAX_RT_PRIO
;
6106 * task_nice - return the nice value of a given task.
6107 * @p: the task in question.
6109 int task_nice(const struct task_struct
*p
)
6111 return TASK_NICE(p
);
6113 EXPORT_SYMBOL(task_nice
);
6116 * idle_cpu - is a given cpu idle currently?
6117 * @cpu: the processor in question.
6119 int idle_cpu(int cpu
)
6121 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6125 * idle_task - return the idle task for a given cpu.
6126 * @cpu: the processor in question.
6128 struct task_struct
*idle_task(int cpu
)
6130 return cpu_rq(cpu
)->idle
;
6134 * find_process_by_pid - find a process with a matching PID value.
6135 * @pid: the pid in question.
6137 static struct task_struct
*find_process_by_pid(pid_t pid
)
6139 return pid
? find_task_by_vpid(pid
) : current
;
6142 /* Actually do priority change: must hold rq lock. */
6144 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6146 BUG_ON(p
->se
.on_rq
);
6149 switch (p
->policy
) {
6153 p
->sched_class
= &fair_sched_class
;
6157 p
->sched_class
= &rt_sched_class
;
6161 p
->rt_priority
= prio
;
6162 p
->normal_prio
= normal_prio(p
);
6163 /* we are holding p->pi_lock already */
6164 p
->prio
= rt_mutex_getprio(p
);
6169 * check the target process has a UID that matches the current process's
6171 static bool check_same_owner(struct task_struct
*p
)
6173 const struct cred
*cred
= current_cred(), *pcred
;
6177 pcred
= __task_cred(p
);
6178 match
= (cred
->euid
== pcred
->euid
||
6179 cred
->euid
== pcred
->uid
);
6184 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6185 struct sched_param
*param
, bool user
)
6187 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6188 unsigned long flags
;
6189 const struct sched_class
*prev_class
= p
->sched_class
;
6193 /* may grab non-irq protected spin_locks */
6194 BUG_ON(in_interrupt());
6196 /* double check policy once rq lock held */
6198 reset_on_fork
= p
->sched_reset_on_fork
;
6199 policy
= oldpolicy
= p
->policy
;
6201 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6202 policy
&= ~SCHED_RESET_ON_FORK
;
6204 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6205 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6206 policy
!= SCHED_IDLE
)
6211 * Valid priorities for SCHED_FIFO and SCHED_RR are
6212 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6213 * SCHED_BATCH and SCHED_IDLE is 0.
6215 if (param
->sched_priority
< 0 ||
6216 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6217 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6219 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6223 * Allow unprivileged RT tasks to decrease priority:
6225 if (user
&& !capable(CAP_SYS_NICE
)) {
6226 if (rt_policy(policy
)) {
6227 unsigned long rlim_rtprio
;
6229 if (!lock_task_sighand(p
, &flags
))
6231 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6232 unlock_task_sighand(p
, &flags
);
6234 /* can't set/change the rt policy */
6235 if (policy
!= p
->policy
&& !rlim_rtprio
)
6238 /* can't increase priority */
6239 if (param
->sched_priority
> p
->rt_priority
&&
6240 param
->sched_priority
> rlim_rtprio
)
6244 * Like positive nice levels, dont allow tasks to
6245 * move out of SCHED_IDLE either:
6247 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6250 /* can't change other user's priorities */
6251 if (!check_same_owner(p
))
6254 /* Normal users shall not reset the sched_reset_on_fork flag */
6255 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6260 #ifdef CONFIG_RT_GROUP_SCHED
6262 * Do not allow realtime tasks into groups that have no runtime
6265 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6266 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6270 retval
= security_task_setscheduler(p
, policy
, param
);
6276 * make sure no PI-waiters arrive (or leave) while we are
6277 * changing the priority of the task:
6279 spin_lock_irqsave(&p
->pi_lock
, flags
);
6281 * To be able to change p->policy safely, the apropriate
6282 * runqueue lock must be held.
6284 rq
= __task_rq_lock(p
);
6285 /* recheck policy now with rq lock held */
6286 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6287 policy
= oldpolicy
= -1;
6288 __task_rq_unlock(rq
);
6289 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6292 update_rq_clock(rq
);
6293 on_rq
= p
->se
.on_rq
;
6294 running
= task_current(rq
, p
);
6296 deactivate_task(rq
, p
, 0);
6298 p
->sched_class
->put_prev_task(rq
, p
);
6300 p
->sched_reset_on_fork
= reset_on_fork
;
6303 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6306 p
->sched_class
->set_curr_task(rq
);
6308 activate_task(rq
, p
, 0);
6310 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6312 __task_rq_unlock(rq
);
6313 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6315 rt_mutex_adjust_pi(p
);
6321 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6322 * @p: the task in question.
6323 * @policy: new policy.
6324 * @param: structure containing the new RT priority.
6326 * NOTE that the task may be already dead.
6328 int sched_setscheduler(struct task_struct
*p
, int policy
,
6329 struct sched_param
*param
)
6331 return __sched_setscheduler(p
, policy
, param
, true);
6333 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6336 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6337 * @p: the task in question.
6338 * @policy: new policy.
6339 * @param: structure containing the new RT priority.
6341 * Just like sched_setscheduler, only don't bother checking if the
6342 * current context has permission. For example, this is needed in
6343 * stop_machine(): we create temporary high priority worker threads,
6344 * but our caller might not have that capability.
6346 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6347 struct sched_param
*param
)
6349 return __sched_setscheduler(p
, policy
, param
, false);
6353 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6355 struct sched_param lparam
;
6356 struct task_struct
*p
;
6359 if (!param
|| pid
< 0)
6361 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6366 p
= find_process_by_pid(pid
);
6368 retval
= sched_setscheduler(p
, policy
, &lparam
);
6375 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6376 * @pid: the pid in question.
6377 * @policy: new policy.
6378 * @param: structure containing the new RT priority.
6380 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6381 struct sched_param __user
*, param
)
6383 /* negative values for policy are not valid */
6387 return do_sched_setscheduler(pid
, policy
, param
);
6391 * sys_sched_setparam - set/change the RT priority of a thread
6392 * @pid: the pid in question.
6393 * @param: structure containing the new RT priority.
6395 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6397 return do_sched_setscheduler(pid
, -1, param
);
6401 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6402 * @pid: the pid in question.
6404 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6406 struct task_struct
*p
;
6413 read_lock(&tasklist_lock
);
6414 p
= find_process_by_pid(pid
);
6416 retval
= security_task_getscheduler(p
);
6419 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6421 read_unlock(&tasklist_lock
);
6426 * sys_sched_getparam - get the RT priority of a thread
6427 * @pid: the pid in question.
6428 * @param: structure containing the RT priority.
6430 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6432 struct sched_param lp
;
6433 struct task_struct
*p
;
6436 if (!param
|| pid
< 0)
6439 read_lock(&tasklist_lock
);
6440 p
= find_process_by_pid(pid
);
6445 retval
= security_task_getscheduler(p
);
6449 lp
.sched_priority
= p
->rt_priority
;
6450 read_unlock(&tasklist_lock
);
6453 * This one might sleep, we cannot do it with a spinlock held ...
6455 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6460 read_unlock(&tasklist_lock
);
6464 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6466 cpumask_var_t cpus_allowed
, new_mask
;
6467 struct task_struct
*p
;
6471 read_lock(&tasklist_lock
);
6473 p
= find_process_by_pid(pid
);
6475 read_unlock(&tasklist_lock
);
6481 * It is not safe to call set_cpus_allowed with the
6482 * tasklist_lock held. We will bump the task_struct's
6483 * usage count and then drop tasklist_lock.
6486 read_unlock(&tasklist_lock
);
6488 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6492 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6494 goto out_free_cpus_allowed
;
6497 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6500 retval
= security_task_setscheduler(p
, 0, NULL
);
6504 cpuset_cpus_allowed(p
, cpus_allowed
);
6505 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6507 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6510 cpuset_cpus_allowed(p
, cpus_allowed
);
6511 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6513 * We must have raced with a concurrent cpuset
6514 * update. Just reset the cpus_allowed to the
6515 * cpuset's cpus_allowed
6517 cpumask_copy(new_mask
, cpus_allowed
);
6522 free_cpumask_var(new_mask
);
6523 out_free_cpus_allowed
:
6524 free_cpumask_var(cpus_allowed
);
6531 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6532 struct cpumask
*new_mask
)
6534 if (len
< cpumask_size())
6535 cpumask_clear(new_mask
);
6536 else if (len
> cpumask_size())
6537 len
= cpumask_size();
6539 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6543 * sys_sched_setaffinity - set the cpu affinity of a process
6544 * @pid: pid of the process
6545 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6546 * @user_mask_ptr: user-space pointer to the new cpu mask
6548 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6549 unsigned long __user
*, user_mask_ptr
)
6551 cpumask_var_t new_mask
;
6554 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6557 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6559 retval
= sched_setaffinity(pid
, new_mask
);
6560 free_cpumask_var(new_mask
);
6564 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6566 struct task_struct
*p
;
6570 read_lock(&tasklist_lock
);
6573 p
= find_process_by_pid(pid
);
6577 retval
= security_task_getscheduler(p
);
6581 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6584 read_unlock(&tasklist_lock
);
6591 * sys_sched_getaffinity - get the cpu affinity of a process
6592 * @pid: pid of the process
6593 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6594 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6596 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6597 unsigned long __user
*, user_mask_ptr
)
6602 if (len
< cpumask_size())
6605 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6608 ret
= sched_getaffinity(pid
, mask
);
6610 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6613 ret
= cpumask_size();
6615 free_cpumask_var(mask
);
6621 * sys_sched_yield - yield the current processor to other threads.
6623 * This function yields the current CPU to other tasks. If there are no
6624 * other threads running on this CPU then this function will return.
6626 SYSCALL_DEFINE0(sched_yield
)
6628 struct rq
*rq
= this_rq_lock();
6630 schedstat_inc(rq
, yld_count
);
6631 current
->sched_class
->yield_task(rq
);
6634 * Since we are going to call schedule() anyway, there's
6635 * no need to preempt or enable interrupts:
6637 __release(rq
->lock
);
6638 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6639 _raw_spin_unlock(&rq
->lock
);
6640 preempt_enable_no_resched();
6647 static inline int should_resched(void)
6649 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6652 static void __cond_resched(void)
6654 add_preempt_count(PREEMPT_ACTIVE
);
6656 sub_preempt_count(PREEMPT_ACTIVE
);
6659 int __sched
_cond_resched(void)
6661 if (should_resched()) {
6667 EXPORT_SYMBOL(_cond_resched
);
6670 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6671 * call schedule, and on return reacquire the lock.
6673 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6674 * operations here to prevent schedule() from being called twice (once via
6675 * spin_unlock(), once by hand).
6677 int __cond_resched_lock(spinlock_t
*lock
)
6679 int resched
= should_resched();
6682 lockdep_assert_held(lock
);
6684 if (spin_needbreak(lock
) || resched
) {
6695 EXPORT_SYMBOL(__cond_resched_lock
);
6697 int __sched
__cond_resched_softirq(void)
6699 BUG_ON(!in_softirq());
6701 if (should_resched()) {
6709 EXPORT_SYMBOL(__cond_resched_softirq
);
6712 * yield - yield the current processor to other threads.
6714 * This is a shortcut for kernel-space yielding - it marks the
6715 * thread runnable and calls sys_sched_yield().
6717 void __sched
yield(void)
6719 set_current_state(TASK_RUNNING
);
6722 EXPORT_SYMBOL(yield
);
6725 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6726 * that process accounting knows that this is a task in IO wait state.
6728 void __sched
io_schedule(void)
6730 struct rq
*rq
= raw_rq();
6732 delayacct_blkio_start();
6733 atomic_inc(&rq
->nr_iowait
);
6734 current
->in_iowait
= 1;
6736 current
->in_iowait
= 0;
6737 atomic_dec(&rq
->nr_iowait
);
6738 delayacct_blkio_end();
6740 EXPORT_SYMBOL(io_schedule
);
6742 long __sched
io_schedule_timeout(long timeout
)
6744 struct rq
*rq
= raw_rq();
6747 delayacct_blkio_start();
6748 atomic_inc(&rq
->nr_iowait
);
6749 current
->in_iowait
= 1;
6750 ret
= schedule_timeout(timeout
);
6751 current
->in_iowait
= 0;
6752 atomic_dec(&rq
->nr_iowait
);
6753 delayacct_blkio_end();
6758 * sys_sched_get_priority_max - return maximum RT priority.
6759 * @policy: scheduling class.
6761 * this syscall returns the maximum rt_priority that can be used
6762 * by a given scheduling class.
6764 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6771 ret
= MAX_USER_RT_PRIO
-1;
6783 * sys_sched_get_priority_min - return minimum RT priority.
6784 * @policy: scheduling class.
6786 * this syscall returns the minimum rt_priority that can be used
6787 * by a given scheduling class.
6789 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6807 * sys_sched_rr_get_interval - return the default timeslice of a process.
6808 * @pid: pid of the process.
6809 * @interval: userspace pointer to the timeslice value.
6811 * this syscall writes the default timeslice value of a given process
6812 * into the user-space timespec buffer. A value of '0' means infinity.
6814 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6815 struct timespec __user
*, interval
)
6817 struct task_struct
*p
;
6818 unsigned int time_slice
;
6826 read_lock(&tasklist_lock
);
6827 p
= find_process_by_pid(pid
);
6831 retval
= security_task_getscheduler(p
);
6835 time_slice
= p
->sched_class
->get_rr_interval(p
);
6837 read_unlock(&tasklist_lock
);
6838 jiffies_to_timespec(time_slice
, &t
);
6839 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6843 read_unlock(&tasklist_lock
);
6847 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6849 void sched_show_task(struct task_struct
*p
)
6851 unsigned long free
= 0;
6854 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6855 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6856 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6857 #if BITS_PER_LONG == 32
6858 if (state
== TASK_RUNNING
)
6859 printk(KERN_CONT
" running ");
6861 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6863 if (state
== TASK_RUNNING
)
6864 printk(KERN_CONT
" running task ");
6866 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6868 #ifdef CONFIG_DEBUG_STACK_USAGE
6869 free
= stack_not_used(p
);
6871 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6872 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6873 (unsigned long)task_thread_info(p
)->flags
);
6875 show_stack(p
, NULL
);
6878 void show_state_filter(unsigned long state_filter
)
6880 struct task_struct
*g
, *p
;
6882 #if BITS_PER_LONG == 32
6884 " task PC stack pid father\n");
6887 " task PC stack pid father\n");
6889 read_lock(&tasklist_lock
);
6890 do_each_thread(g
, p
) {
6892 * reset the NMI-timeout, listing all files on a slow
6893 * console might take alot of time:
6895 touch_nmi_watchdog();
6896 if (!state_filter
|| (p
->state
& state_filter
))
6898 } while_each_thread(g
, p
);
6900 touch_all_softlockup_watchdogs();
6902 #ifdef CONFIG_SCHED_DEBUG
6903 sysrq_sched_debug_show();
6905 read_unlock(&tasklist_lock
);
6907 * Only show locks if all tasks are dumped:
6909 if (state_filter
== -1)
6910 debug_show_all_locks();
6913 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6915 idle
->sched_class
= &idle_sched_class
;
6919 * init_idle - set up an idle thread for a given CPU
6920 * @idle: task in question
6921 * @cpu: cpu the idle task belongs to
6923 * NOTE: this function does not set the idle thread's NEED_RESCHED
6924 * flag, to make booting more robust.
6926 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6928 struct rq
*rq
= cpu_rq(cpu
);
6929 unsigned long flags
;
6931 spin_lock_irqsave(&rq
->lock
, flags
);
6934 idle
->se
.exec_start
= sched_clock();
6936 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6937 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6938 __set_task_cpu(idle
, cpu
);
6940 rq
->curr
= rq
->idle
= idle
;
6941 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6944 spin_unlock_irqrestore(&rq
->lock
, flags
);
6946 /* Set the preempt count _outside_ the spinlocks! */
6947 #if defined(CONFIG_PREEMPT)
6948 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6950 task_thread_info(idle
)->preempt_count
= 0;
6953 * The idle tasks have their own, simple scheduling class:
6955 idle
->sched_class
= &idle_sched_class
;
6956 ftrace_graph_init_task(idle
);
6960 * In a system that switches off the HZ timer nohz_cpu_mask
6961 * indicates which cpus entered this state. This is used
6962 * in the rcu update to wait only for active cpus. For system
6963 * which do not switch off the HZ timer nohz_cpu_mask should
6964 * always be CPU_BITS_NONE.
6966 cpumask_var_t nohz_cpu_mask
;
6969 * Increase the granularity value when there are more CPUs,
6970 * because with more CPUs the 'effective latency' as visible
6971 * to users decreases. But the relationship is not linear,
6972 * so pick a second-best guess by going with the log2 of the
6975 * This idea comes from the SD scheduler of Con Kolivas:
6977 static inline void sched_init_granularity(void)
6979 unsigned int factor
= 1 + ilog2(num_online_cpus());
6980 const unsigned long limit
= 200000000;
6982 sysctl_sched_min_granularity
*= factor
;
6983 if (sysctl_sched_min_granularity
> limit
)
6984 sysctl_sched_min_granularity
= limit
;
6986 sysctl_sched_latency
*= factor
;
6987 if (sysctl_sched_latency
> limit
)
6988 sysctl_sched_latency
= limit
;
6990 sysctl_sched_wakeup_granularity
*= factor
;
6992 sysctl_sched_shares_ratelimit
*= factor
;
6997 * This is how migration works:
6999 * 1) we queue a struct migration_req structure in the source CPU's
7000 * runqueue and wake up that CPU's migration thread.
7001 * 2) we down() the locked semaphore => thread blocks.
7002 * 3) migration thread wakes up (implicitly it forces the migrated
7003 * thread off the CPU)
7004 * 4) it gets the migration request and checks whether the migrated
7005 * task is still in the wrong runqueue.
7006 * 5) if it's in the wrong runqueue then the migration thread removes
7007 * it and puts it into the right queue.
7008 * 6) migration thread up()s the semaphore.
7009 * 7) we wake up and the migration is done.
7013 * Change a given task's CPU affinity. Migrate the thread to a
7014 * proper CPU and schedule it away if the CPU it's executing on
7015 * is removed from the allowed bitmask.
7017 * NOTE: the caller must have a valid reference to the task, the
7018 * task must not exit() & deallocate itself prematurely. The
7019 * call is not atomic; no spinlocks may be held.
7021 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7023 struct migration_req req
;
7024 unsigned long flags
;
7028 rq
= task_rq_lock(p
, &flags
);
7029 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7034 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7035 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7040 if (p
->sched_class
->set_cpus_allowed
)
7041 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7043 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7044 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7047 /* Can the task run on the task's current CPU? If so, we're done */
7048 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7051 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7052 /* Need help from migration thread: drop lock and wait. */
7053 struct task_struct
*mt
= rq
->migration_thread
;
7055 get_task_struct(mt
);
7056 task_rq_unlock(rq
, &flags
);
7057 wake_up_process(rq
->migration_thread
);
7058 put_task_struct(mt
);
7059 wait_for_completion(&req
.done
);
7060 tlb_migrate_finish(p
->mm
);
7064 task_rq_unlock(rq
, &flags
);
7068 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7071 * Move (not current) task off this cpu, onto dest cpu. We're doing
7072 * this because either it can't run here any more (set_cpus_allowed()
7073 * away from this CPU, or CPU going down), or because we're
7074 * attempting to rebalance this task on exec (sched_exec).
7076 * So we race with normal scheduler movements, but that's OK, as long
7077 * as the task is no longer on this CPU.
7079 * Returns non-zero if task was successfully migrated.
7081 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7083 struct rq
*rq_dest
, *rq_src
;
7086 if (unlikely(!cpu_active(dest_cpu
)))
7089 rq_src
= cpu_rq(src_cpu
);
7090 rq_dest
= cpu_rq(dest_cpu
);
7092 double_rq_lock(rq_src
, rq_dest
);
7093 /* Already moved. */
7094 if (task_cpu(p
) != src_cpu
)
7096 /* Affinity changed (again). */
7097 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7100 on_rq
= p
->se
.on_rq
;
7102 deactivate_task(rq_src
, p
, 0);
7104 set_task_cpu(p
, dest_cpu
);
7106 activate_task(rq_dest
, p
, 0);
7107 check_preempt_curr(rq_dest
, p
, 0);
7112 double_rq_unlock(rq_src
, rq_dest
);
7116 #define RCU_MIGRATION_IDLE 0
7117 #define RCU_MIGRATION_NEED_QS 1
7118 #define RCU_MIGRATION_GOT_QS 2
7119 #define RCU_MIGRATION_MUST_SYNC 3
7122 * migration_thread - this is a highprio system thread that performs
7123 * thread migration by bumping thread off CPU then 'pushing' onto
7126 static int migration_thread(void *data
)
7129 int cpu
= (long)data
;
7133 BUG_ON(rq
->migration_thread
!= current
);
7135 set_current_state(TASK_INTERRUPTIBLE
);
7136 while (!kthread_should_stop()) {
7137 struct migration_req
*req
;
7138 struct list_head
*head
;
7140 spin_lock_irq(&rq
->lock
);
7142 if (cpu_is_offline(cpu
)) {
7143 spin_unlock_irq(&rq
->lock
);
7147 if (rq
->active_balance
) {
7148 active_load_balance(rq
, cpu
);
7149 rq
->active_balance
= 0;
7152 head
= &rq
->migration_queue
;
7154 if (list_empty(head
)) {
7155 spin_unlock_irq(&rq
->lock
);
7157 set_current_state(TASK_INTERRUPTIBLE
);
7160 req
= list_entry(head
->next
, struct migration_req
, list
);
7161 list_del_init(head
->next
);
7163 if (req
->task
!= NULL
) {
7164 spin_unlock(&rq
->lock
);
7165 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7166 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7167 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7168 spin_unlock(&rq
->lock
);
7170 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7171 spin_unlock(&rq
->lock
);
7172 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7176 complete(&req
->done
);
7178 __set_current_state(TASK_RUNNING
);
7183 #ifdef CONFIG_HOTPLUG_CPU
7185 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7189 local_irq_disable();
7190 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7196 * Figure out where task on dead CPU should go, use force if necessary.
7198 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7201 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7204 /* Look for allowed, online CPU in same node. */
7205 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7206 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7209 /* Any allowed, online CPU? */
7210 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7211 if (dest_cpu
< nr_cpu_ids
)
7214 /* No more Mr. Nice Guy. */
7215 if (dest_cpu
>= nr_cpu_ids
) {
7216 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7217 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7220 * Don't tell them about moving exiting tasks or
7221 * kernel threads (both mm NULL), since they never
7224 if (p
->mm
&& printk_ratelimit()) {
7225 printk(KERN_INFO
"process %d (%s) no "
7226 "longer affine to cpu%d\n",
7227 task_pid_nr(p
), p
->comm
, dead_cpu
);
7232 /* It can have affinity changed while we were choosing. */
7233 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7238 * While a dead CPU has no uninterruptible tasks queued at this point,
7239 * it might still have a nonzero ->nr_uninterruptible counter, because
7240 * for performance reasons the counter is not stricly tracking tasks to
7241 * their home CPUs. So we just add the counter to another CPU's counter,
7242 * to keep the global sum constant after CPU-down:
7244 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7246 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7247 unsigned long flags
;
7249 local_irq_save(flags
);
7250 double_rq_lock(rq_src
, rq_dest
);
7251 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7252 rq_src
->nr_uninterruptible
= 0;
7253 double_rq_unlock(rq_src
, rq_dest
);
7254 local_irq_restore(flags
);
7257 /* Run through task list and migrate tasks from the dead cpu. */
7258 static void migrate_live_tasks(int src_cpu
)
7260 struct task_struct
*p
, *t
;
7262 read_lock(&tasklist_lock
);
7264 do_each_thread(t
, p
) {
7268 if (task_cpu(p
) == src_cpu
)
7269 move_task_off_dead_cpu(src_cpu
, p
);
7270 } while_each_thread(t
, p
);
7272 read_unlock(&tasklist_lock
);
7276 * Schedules idle task to be the next runnable task on current CPU.
7277 * It does so by boosting its priority to highest possible.
7278 * Used by CPU offline code.
7280 void sched_idle_next(void)
7282 int this_cpu
= smp_processor_id();
7283 struct rq
*rq
= cpu_rq(this_cpu
);
7284 struct task_struct
*p
= rq
->idle
;
7285 unsigned long flags
;
7287 /* cpu has to be offline */
7288 BUG_ON(cpu_online(this_cpu
));
7291 * Strictly not necessary since rest of the CPUs are stopped by now
7292 * and interrupts disabled on the current cpu.
7294 spin_lock_irqsave(&rq
->lock
, flags
);
7296 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7298 update_rq_clock(rq
);
7299 activate_task(rq
, p
, 0);
7301 spin_unlock_irqrestore(&rq
->lock
, flags
);
7305 * Ensures that the idle task is using init_mm right before its cpu goes
7308 void idle_task_exit(void)
7310 struct mm_struct
*mm
= current
->active_mm
;
7312 BUG_ON(cpu_online(smp_processor_id()));
7315 switch_mm(mm
, &init_mm
, current
);
7319 /* called under rq->lock with disabled interrupts */
7320 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7322 struct rq
*rq
= cpu_rq(dead_cpu
);
7324 /* Must be exiting, otherwise would be on tasklist. */
7325 BUG_ON(!p
->exit_state
);
7327 /* Cannot have done final schedule yet: would have vanished. */
7328 BUG_ON(p
->state
== TASK_DEAD
);
7333 * Drop lock around migration; if someone else moves it,
7334 * that's OK. No task can be added to this CPU, so iteration is
7337 spin_unlock_irq(&rq
->lock
);
7338 move_task_off_dead_cpu(dead_cpu
, p
);
7339 spin_lock_irq(&rq
->lock
);
7344 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7345 static void migrate_dead_tasks(unsigned int dead_cpu
)
7347 struct rq
*rq
= cpu_rq(dead_cpu
);
7348 struct task_struct
*next
;
7351 if (!rq
->nr_running
)
7353 update_rq_clock(rq
);
7354 next
= pick_next_task(rq
);
7357 next
->sched_class
->put_prev_task(rq
, next
);
7358 migrate_dead(dead_cpu
, next
);
7364 * remove the tasks which were accounted by rq from calc_load_tasks.
7366 static void calc_global_load_remove(struct rq
*rq
)
7368 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7369 rq
->calc_load_active
= 0;
7371 #endif /* CONFIG_HOTPLUG_CPU */
7373 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7375 static struct ctl_table sd_ctl_dir
[] = {
7377 .procname
= "sched_domain",
7383 static struct ctl_table sd_ctl_root
[] = {
7385 .ctl_name
= CTL_KERN
,
7386 .procname
= "kernel",
7388 .child
= sd_ctl_dir
,
7393 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7395 struct ctl_table
*entry
=
7396 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7401 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7403 struct ctl_table
*entry
;
7406 * In the intermediate directories, both the child directory and
7407 * procname are dynamically allocated and could fail but the mode
7408 * will always be set. In the lowest directory the names are
7409 * static strings and all have proc handlers.
7411 for (entry
= *tablep
; entry
->mode
; entry
++) {
7413 sd_free_ctl_entry(&entry
->child
);
7414 if (entry
->proc_handler
== NULL
)
7415 kfree(entry
->procname
);
7423 set_table_entry(struct ctl_table
*entry
,
7424 const char *procname
, void *data
, int maxlen
,
7425 mode_t mode
, proc_handler
*proc_handler
)
7427 entry
->procname
= procname
;
7429 entry
->maxlen
= maxlen
;
7431 entry
->proc_handler
= proc_handler
;
7434 static struct ctl_table
*
7435 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7437 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7442 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7443 sizeof(long), 0644, proc_doulongvec_minmax
);
7444 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7445 sizeof(long), 0644, proc_doulongvec_minmax
);
7446 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7447 sizeof(int), 0644, proc_dointvec_minmax
);
7448 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7449 sizeof(int), 0644, proc_dointvec_minmax
);
7450 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7451 sizeof(int), 0644, proc_dointvec_minmax
);
7452 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7453 sizeof(int), 0644, proc_dointvec_minmax
);
7454 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7455 sizeof(int), 0644, proc_dointvec_minmax
);
7456 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7457 sizeof(int), 0644, proc_dointvec_minmax
);
7458 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7459 sizeof(int), 0644, proc_dointvec_minmax
);
7460 set_table_entry(&table
[9], "cache_nice_tries",
7461 &sd
->cache_nice_tries
,
7462 sizeof(int), 0644, proc_dointvec_minmax
);
7463 set_table_entry(&table
[10], "flags", &sd
->flags
,
7464 sizeof(int), 0644, proc_dointvec_minmax
);
7465 set_table_entry(&table
[11], "name", sd
->name
,
7466 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7467 /* &table[12] is terminator */
7472 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7474 struct ctl_table
*entry
, *table
;
7475 struct sched_domain
*sd
;
7476 int domain_num
= 0, i
;
7479 for_each_domain(cpu
, sd
)
7481 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7486 for_each_domain(cpu
, sd
) {
7487 snprintf(buf
, 32, "domain%d", i
);
7488 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7490 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7497 static struct ctl_table_header
*sd_sysctl_header
;
7498 static void register_sched_domain_sysctl(void)
7500 int i
, cpu_num
= num_online_cpus();
7501 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7504 WARN_ON(sd_ctl_dir
[0].child
);
7505 sd_ctl_dir
[0].child
= entry
;
7510 for_each_online_cpu(i
) {
7511 snprintf(buf
, 32, "cpu%d", i
);
7512 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7514 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7518 WARN_ON(sd_sysctl_header
);
7519 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7522 /* may be called multiple times per register */
7523 static void unregister_sched_domain_sysctl(void)
7525 if (sd_sysctl_header
)
7526 unregister_sysctl_table(sd_sysctl_header
);
7527 sd_sysctl_header
= NULL
;
7528 if (sd_ctl_dir
[0].child
)
7529 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7532 static void register_sched_domain_sysctl(void)
7535 static void unregister_sched_domain_sysctl(void)
7540 static void set_rq_online(struct rq
*rq
)
7543 const struct sched_class
*class;
7545 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7548 for_each_class(class) {
7549 if (class->rq_online
)
7550 class->rq_online(rq
);
7555 static void set_rq_offline(struct rq
*rq
)
7558 const struct sched_class
*class;
7560 for_each_class(class) {
7561 if (class->rq_offline
)
7562 class->rq_offline(rq
);
7565 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7571 * migration_call - callback that gets triggered when a CPU is added.
7572 * Here we can start up the necessary migration thread for the new CPU.
7574 static int __cpuinit
7575 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7577 struct task_struct
*p
;
7578 int cpu
= (long)hcpu
;
7579 unsigned long flags
;
7584 case CPU_UP_PREPARE
:
7585 case CPU_UP_PREPARE_FROZEN
:
7586 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7589 kthread_bind(p
, cpu
);
7590 /* Must be high prio: stop_machine expects to yield to it. */
7591 rq
= task_rq_lock(p
, &flags
);
7592 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7593 task_rq_unlock(rq
, &flags
);
7595 cpu_rq(cpu
)->migration_thread
= p
;
7596 rq
->calc_load_update
= calc_load_update
;
7600 case CPU_ONLINE_FROZEN
:
7601 /* Strictly unnecessary, as first user will wake it. */
7602 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7604 /* Update our root-domain */
7606 spin_lock_irqsave(&rq
->lock
, flags
);
7608 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7612 spin_unlock_irqrestore(&rq
->lock
, flags
);
7615 #ifdef CONFIG_HOTPLUG_CPU
7616 case CPU_UP_CANCELED
:
7617 case CPU_UP_CANCELED_FROZEN
:
7618 if (!cpu_rq(cpu
)->migration_thread
)
7620 /* Unbind it from offline cpu so it can run. Fall thru. */
7621 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7622 cpumask_any(cpu_online_mask
));
7623 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7624 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7625 cpu_rq(cpu
)->migration_thread
= NULL
;
7629 case CPU_DEAD_FROZEN
:
7630 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7631 migrate_live_tasks(cpu
);
7633 kthread_stop(rq
->migration_thread
);
7634 put_task_struct(rq
->migration_thread
);
7635 rq
->migration_thread
= NULL
;
7636 /* Idle task back to normal (off runqueue, low prio) */
7637 spin_lock_irq(&rq
->lock
);
7638 update_rq_clock(rq
);
7639 deactivate_task(rq
, rq
->idle
, 0);
7640 rq
->idle
->static_prio
= MAX_PRIO
;
7641 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7642 rq
->idle
->sched_class
= &idle_sched_class
;
7643 migrate_dead_tasks(cpu
);
7644 spin_unlock_irq(&rq
->lock
);
7646 migrate_nr_uninterruptible(rq
);
7647 BUG_ON(rq
->nr_running
!= 0);
7648 calc_global_load_remove(rq
);
7650 * No need to migrate the tasks: it was best-effort if
7651 * they didn't take sched_hotcpu_mutex. Just wake up
7654 spin_lock_irq(&rq
->lock
);
7655 while (!list_empty(&rq
->migration_queue
)) {
7656 struct migration_req
*req
;
7658 req
= list_entry(rq
->migration_queue
.next
,
7659 struct migration_req
, list
);
7660 list_del_init(&req
->list
);
7661 spin_unlock_irq(&rq
->lock
);
7662 complete(&req
->done
);
7663 spin_lock_irq(&rq
->lock
);
7665 spin_unlock_irq(&rq
->lock
);
7669 case CPU_DYING_FROZEN
:
7670 /* Update our root-domain */
7672 spin_lock_irqsave(&rq
->lock
, flags
);
7674 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7677 spin_unlock_irqrestore(&rq
->lock
, flags
);
7685 * Register at high priority so that task migration (migrate_all_tasks)
7686 * happens before everything else. This has to be lower priority than
7687 * the notifier in the perf_event subsystem, though.
7689 static struct notifier_block __cpuinitdata migration_notifier
= {
7690 .notifier_call
= migration_call
,
7694 static int __init
migration_init(void)
7696 void *cpu
= (void *)(long)smp_processor_id();
7699 /* Start one for the boot CPU: */
7700 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7701 BUG_ON(err
== NOTIFY_BAD
);
7702 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7703 register_cpu_notifier(&migration_notifier
);
7707 early_initcall(migration_init
);
7712 #ifdef CONFIG_SCHED_DEBUG
7714 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7715 struct cpumask
*groupmask
)
7717 struct sched_group
*group
= sd
->groups
;
7720 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7721 cpumask_clear(groupmask
);
7723 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7725 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7726 printk("does not load-balance\n");
7728 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7733 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7735 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7736 printk(KERN_ERR
"ERROR: domain->span does not contain "
7739 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7740 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7744 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7748 printk(KERN_ERR
"ERROR: group is NULL\n");
7752 if (!group
->cpu_power
) {
7753 printk(KERN_CONT
"\n");
7754 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7759 if (!cpumask_weight(sched_group_cpus(group
))) {
7760 printk(KERN_CONT
"\n");
7761 printk(KERN_ERR
"ERROR: empty group\n");
7765 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7766 printk(KERN_CONT
"\n");
7767 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7771 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7773 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7775 printk(KERN_CONT
" %s", str
);
7776 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7777 printk(KERN_CONT
" (cpu_power = %d)",
7781 group
= group
->next
;
7782 } while (group
!= sd
->groups
);
7783 printk(KERN_CONT
"\n");
7785 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7786 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7789 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7790 printk(KERN_ERR
"ERROR: parent span is not a superset "
7791 "of domain->span\n");
7795 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7797 cpumask_var_t groupmask
;
7801 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7805 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7807 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7808 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7813 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7820 free_cpumask_var(groupmask
);
7822 #else /* !CONFIG_SCHED_DEBUG */
7823 # define sched_domain_debug(sd, cpu) do { } while (0)
7824 #endif /* CONFIG_SCHED_DEBUG */
7826 static int sd_degenerate(struct sched_domain
*sd
)
7828 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7831 /* Following flags need at least 2 groups */
7832 if (sd
->flags
& (SD_LOAD_BALANCE
|
7833 SD_BALANCE_NEWIDLE
|
7837 SD_SHARE_PKG_RESOURCES
)) {
7838 if (sd
->groups
!= sd
->groups
->next
)
7842 /* Following flags don't use groups */
7843 if (sd
->flags
& (SD_WAKE_AFFINE
))
7850 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7852 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7854 if (sd_degenerate(parent
))
7857 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7860 /* Flags needing groups don't count if only 1 group in parent */
7861 if (parent
->groups
== parent
->groups
->next
) {
7862 pflags
&= ~(SD_LOAD_BALANCE
|
7863 SD_BALANCE_NEWIDLE
|
7867 SD_SHARE_PKG_RESOURCES
);
7868 if (nr_node_ids
== 1)
7869 pflags
&= ~SD_SERIALIZE
;
7871 if (~cflags
& pflags
)
7877 static void free_rootdomain(struct root_domain
*rd
)
7879 cpupri_cleanup(&rd
->cpupri
);
7881 free_cpumask_var(rd
->rto_mask
);
7882 free_cpumask_var(rd
->online
);
7883 free_cpumask_var(rd
->span
);
7887 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7889 struct root_domain
*old_rd
= NULL
;
7890 unsigned long flags
;
7892 spin_lock_irqsave(&rq
->lock
, flags
);
7897 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7900 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7903 * If we dont want to free the old_rt yet then
7904 * set old_rd to NULL to skip the freeing later
7907 if (!atomic_dec_and_test(&old_rd
->refcount
))
7911 atomic_inc(&rd
->refcount
);
7914 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7915 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
7918 spin_unlock_irqrestore(&rq
->lock
, flags
);
7921 free_rootdomain(old_rd
);
7924 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7926 gfp_t gfp
= GFP_KERNEL
;
7928 memset(rd
, 0, sizeof(*rd
));
7933 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7935 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7937 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7940 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7945 free_cpumask_var(rd
->rto_mask
);
7947 free_cpumask_var(rd
->online
);
7949 free_cpumask_var(rd
->span
);
7954 static void init_defrootdomain(void)
7956 init_rootdomain(&def_root_domain
, true);
7958 atomic_set(&def_root_domain
.refcount
, 1);
7961 static struct root_domain
*alloc_rootdomain(void)
7963 struct root_domain
*rd
;
7965 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7969 if (init_rootdomain(rd
, false) != 0) {
7978 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7979 * hold the hotplug lock.
7982 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7984 struct rq
*rq
= cpu_rq(cpu
);
7985 struct sched_domain
*tmp
;
7987 /* Remove the sched domains which do not contribute to scheduling. */
7988 for (tmp
= sd
; tmp
; ) {
7989 struct sched_domain
*parent
= tmp
->parent
;
7993 if (sd_parent_degenerate(tmp
, parent
)) {
7994 tmp
->parent
= parent
->parent
;
7996 parent
->parent
->child
= tmp
;
8001 if (sd
&& sd_degenerate(sd
)) {
8007 sched_domain_debug(sd
, cpu
);
8009 rq_attach_root(rq
, rd
);
8010 rcu_assign_pointer(rq
->sd
, sd
);
8013 /* cpus with isolated domains */
8014 static cpumask_var_t cpu_isolated_map
;
8016 /* Setup the mask of cpus configured for isolated domains */
8017 static int __init
isolated_cpu_setup(char *str
)
8019 cpulist_parse(str
, cpu_isolated_map
);
8023 __setup("isolcpus=", isolated_cpu_setup
);
8026 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8027 * to a function which identifies what group(along with sched group) a CPU
8028 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8029 * (due to the fact that we keep track of groups covered with a struct cpumask).
8031 * init_sched_build_groups will build a circular linked list of the groups
8032 * covered by the given span, and will set each group's ->cpumask correctly,
8033 * and ->cpu_power to 0.
8036 init_sched_build_groups(const struct cpumask
*span
,
8037 const struct cpumask
*cpu_map
,
8038 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8039 struct sched_group
**sg
,
8040 struct cpumask
*tmpmask
),
8041 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8043 struct sched_group
*first
= NULL
, *last
= NULL
;
8046 cpumask_clear(covered
);
8048 for_each_cpu(i
, span
) {
8049 struct sched_group
*sg
;
8050 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8053 if (cpumask_test_cpu(i
, covered
))
8056 cpumask_clear(sched_group_cpus(sg
));
8059 for_each_cpu(j
, span
) {
8060 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8063 cpumask_set_cpu(j
, covered
);
8064 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8075 #define SD_NODES_PER_DOMAIN 16
8080 * find_next_best_node - find the next node to include in a sched_domain
8081 * @node: node whose sched_domain we're building
8082 * @used_nodes: nodes already in the sched_domain
8084 * Find the next node to include in a given scheduling domain. Simply
8085 * finds the closest node not already in the @used_nodes map.
8087 * Should use nodemask_t.
8089 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8091 int i
, n
, val
, min_val
, best_node
= 0;
8095 for (i
= 0; i
< nr_node_ids
; i
++) {
8096 /* Start at @node */
8097 n
= (node
+ i
) % nr_node_ids
;
8099 if (!nr_cpus_node(n
))
8102 /* Skip already used nodes */
8103 if (node_isset(n
, *used_nodes
))
8106 /* Simple min distance search */
8107 val
= node_distance(node
, n
);
8109 if (val
< min_val
) {
8115 node_set(best_node
, *used_nodes
);
8120 * sched_domain_node_span - get a cpumask for a node's sched_domain
8121 * @node: node whose cpumask we're constructing
8122 * @span: resulting cpumask
8124 * Given a node, construct a good cpumask for its sched_domain to span. It
8125 * should be one that prevents unnecessary balancing, but also spreads tasks
8128 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8130 nodemask_t used_nodes
;
8133 cpumask_clear(span
);
8134 nodes_clear(used_nodes
);
8136 cpumask_or(span
, span
, cpumask_of_node(node
));
8137 node_set(node
, used_nodes
);
8139 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8140 int next_node
= find_next_best_node(node
, &used_nodes
);
8142 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8145 #endif /* CONFIG_NUMA */
8147 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8150 * The cpus mask in sched_group and sched_domain hangs off the end.
8152 * ( See the the comments in include/linux/sched.h:struct sched_group
8153 * and struct sched_domain. )
8155 struct static_sched_group
{
8156 struct sched_group sg
;
8157 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8160 struct static_sched_domain
{
8161 struct sched_domain sd
;
8162 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8168 cpumask_var_t domainspan
;
8169 cpumask_var_t covered
;
8170 cpumask_var_t notcovered
;
8172 cpumask_var_t nodemask
;
8173 cpumask_var_t this_sibling_map
;
8174 cpumask_var_t this_core_map
;
8175 cpumask_var_t send_covered
;
8176 cpumask_var_t tmpmask
;
8177 struct sched_group
**sched_group_nodes
;
8178 struct root_domain
*rd
;
8182 sa_sched_groups
= 0,
8187 sa_this_sibling_map
,
8189 sa_sched_group_nodes
,
8199 * SMT sched-domains:
8201 #ifdef CONFIG_SCHED_SMT
8202 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8203 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8206 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8207 struct sched_group
**sg
, struct cpumask
*unused
)
8210 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8213 #endif /* CONFIG_SCHED_SMT */
8216 * multi-core sched-domains:
8218 #ifdef CONFIG_SCHED_MC
8219 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8220 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8221 #endif /* CONFIG_SCHED_MC */
8223 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8225 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8226 struct sched_group
**sg
, struct cpumask
*mask
)
8230 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8231 group
= cpumask_first(mask
);
8233 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8236 #elif defined(CONFIG_SCHED_MC)
8238 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8239 struct sched_group
**sg
, struct cpumask
*unused
)
8242 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8247 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8248 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8251 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8252 struct sched_group
**sg
, struct cpumask
*mask
)
8255 #ifdef CONFIG_SCHED_MC
8256 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8257 group
= cpumask_first(mask
);
8258 #elif defined(CONFIG_SCHED_SMT)
8259 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8260 group
= cpumask_first(mask
);
8265 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8271 * The init_sched_build_groups can't handle what we want to do with node
8272 * groups, so roll our own. Now each node has its own list of groups which
8273 * gets dynamically allocated.
8275 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8276 static struct sched_group
***sched_group_nodes_bycpu
;
8278 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8279 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8281 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8282 struct sched_group
**sg
,
8283 struct cpumask
*nodemask
)
8287 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8288 group
= cpumask_first(nodemask
);
8291 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8295 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8297 struct sched_group
*sg
= group_head
;
8303 for_each_cpu(j
, sched_group_cpus(sg
)) {
8304 struct sched_domain
*sd
;
8306 sd
= &per_cpu(phys_domains
, j
).sd
;
8307 if (j
!= group_first_cpu(sd
->groups
)) {
8309 * Only add "power" once for each
8315 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8318 } while (sg
!= group_head
);
8321 static int build_numa_sched_groups(struct s_data
*d
,
8322 const struct cpumask
*cpu_map
, int num
)
8324 struct sched_domain
*sd
;
8325 struct sched_group
*sg
, *prev
;
8328 cpumask_clear(d
->covered
);
8329 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8330 if (cpumask_empty(d
->nodemask
)) {
8331 d
->sched_group_nodes
[num
] = NULL
;
8335 sched_domain_node_span(num
, d
->domainspan
);
8336 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8338 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8341 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8345 d
->sched_group_nodes
[num
] = sg
;
8347 for_each_cpu(j
, d
->nodemask
) {
8348 sd
= &per_cpu(node_domains
, j
).sd
;
8353 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8355 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8358 for (j
= 0; j
< nr_node_ids
; j
++) {
8359 n
= (num
+ j
) % nr_node_ids
;
8360 cpumask_complement(d
->notcovered
, d
->covered
);
8361 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8362 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8363 if (cpumask_empty(d
->tmpmask
))
8365 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8366 if (cpumask_empty(d
->tmpmask
))
8368 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8372 "Can not alloc domain group for node %d\n", j
);
8376 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8377 sg
->next
= prev
->next
;
8378 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8385 #endif /* CONFIG_NUMA */
8388 /* Free memory allocated for various sched_group structures */
8389 static void free_sched_groups(const struct cpumask
*cpu_map
,
8390 struct cpumask
*nodemask
)
8394 for_each_cpu(cpu
, cpu_map
) {
8395 struct sched_group
**sched_group_nodes
8396 = sched_group_nodes_bycpu
[cpu
];
8398 if (!sched_group_nodes
)
8401 for (i
= 0; i
< nr_node_ids
; i
++) {
8402 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8404 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8405 if (cpumask_empty(nodemask
))
8415 if (oldsg
!= sched_group_nodes
[i
])
8418 kfree(sched_group_nodes
);
8419 sched_group_nodes_bycpu
[cpu
] = NULL
;
8422 #else /* !CONFIG_NUMA */
8423 static void free_sched_groups(const struct cpumask
*cpu_map
,
8424 struct cpumask
*nodemask
)
8427 #endif /* CONFIG_NUMA */
8430 * Initialize sched groups cpu_power.
8432 * cpu_power indicates the capacity of sched group, which is used while
8433 * distributing the load between different sched groups in a sched domain.
8434 * Typically cpu_power for all the groups in a sched domain will be same unless
8435 * there are asymmetries in the topology. If there are asymmetries, group
8436 * having more cpu_power will pickup more load compared to the group having
8439 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8441 struct sched_domain
*child
;
8442 struct sched_group
*group
;
8446 WARN_ON(!sd
|| !sd
->groups
);
8448 if (cpu
!= group_first_cpu(sd
->groups
))
8453 sd
->groups
->cpu_power
= 0;
8456 power
= SCHED_LOAD_SCALE
;
8457 weight
= cpumask_weight(sched_domain_span(sd
));
8459 * SMT siblings share the power of a single core.
8460 * Usually multiple threads get a better yield out of
8461 * that one core than a single thread would have,
8462 * reflect that in sd->smt_gain.
8464 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8465 power
*= sd
->smt_gain
;
8467 power
>>= SCHED_LOAD_SHIFT
;
8469 sd
->groups
->cpu_power
+= power
;
8474 * Add cpu_power of each child group to this groups cpu_power.
8476 group
= child
->groups
;
8478 sd
->groups
->cpu_power
+= group
->cpu_power
;
8479 group
= group
->next
;
8480 } while (group
!= child
->groups
);
8484 * Initializers for schedule domains
8485 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8488 #ifdef CONFIG_SCHED_DEBUG
8489 # define SD_INIT_NAME(sd, type) sd->name = #type
8491 # define SD_INIT_NAME(sd, type) do { } while (0)
8494 #define SD_INIT(sd, type) sd_init_##type(sd)
8496 #define SD_INIT_FUNC(type) \
8497 static noinline void sd_init_##type(struct sched_domain *sd) \
8499 memset(sd, 0, sizeof(*sd)); \
8500 *sd = SD_##type##_INIT; \
8501 sd->level = SD_LV_##type; \
8502 SD_INIT_NAME(sd, type); \
8507 SD_INIT_FUNC(ALLNODES
)
8510 #ifdef CONFIG_SCHED_SMT
8511 SD_INIT_FUNC(SIBLING
)
8513 #ifdef CONFIG_SCHED_MC
8517 static int default_relax_domain_level
= -1;
8519 static int __init
setup_relax_domain_level(char *str
)
8523 val
= simple_strtoul(str
, NULL
, 0);
8524 if (val
< SD_LV_MAX
)
8525 default_relax_domain_level
= val
;
8529 __setup("relax_domain_level=", setup_relax_domain_level
);
8531 static void set_domain_attribute(struct sched_domain
*sd
,
8532 struct sched_domain_attr
*attr
)
8536 if (!attr
|| attr
->relax_domain_level
< 0) {
8537 if (default_relax_domain_level
< 0)
8540 request
= default_relax_domain_level
;
8542 request
= attr
->relax_domain_level
;
8543 if (request
< sd
->level
) {
8544 /* turn off idle balance on this domain */
8545 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8547 /* turn on idle balance on this domain */
8548 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8552 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8553 const struct cpumask
*cpu_map
)
8556 case sa_sched_groups
:
8557 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8558 d
->sched_group_nodes
= NULL
;
8560 free_rootdomain(d
->rd
); /* fall through */
8562 free_cpumask_var(d
->tmpmask
); /* fall through */
8563 case sa_send_covered
:
8564 free_cpumask_var(d
->send_covered
); /* fall through */
8565 case sa_this_core_map
:
8566 free_cpumask_var(d
->this_core_map
); /* fall through */
8567 case sa_this_sibling_map
:
8568 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8570 free_cpumask_var(d
->nodemask
); /* fall through */
8571 case sa_sched_group_nodes
:
8573 kfree(d
->sched_group_nodes
); /* fall through */
8575 free_cpumask_var(d
->notcovered
); /* fall through */
8577 free_cpumask_var(d
->covered
); /* fall through */
8579 free_cpumask_var(d
->domainspan
); /* fall through */
8586 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8587 const struct cpumask
*cpu_map
)
8590 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8592 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8593 return sa_domainspan
;
8594 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8596 /* Allocate the per-node list of sched groups */
8597 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8598 sizeof(struct sched_group
*), GFP_KERNEL
);
8599 if (!d
->sched_group_nodes
) {
8600 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8601 return sa_notcovered
;
8603 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8605 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8606 return sa_sched_group_nodes
;
8607 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8609 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8610 return sa_this_sibling_map
;
8611 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8612 return sa_this_core_map
;
8613 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8614 return sa_send_covered
;
8615 d
->rd
= alloc_rootdomain();
8617 printk(KERN_WARNING
"Cannot alloc root domain\n");
8620 return sa_rootdomain
;
8623 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8624 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8626 struct sched_domain
*sd
= NULL
;
8628 struct sched_domain
*parent
;
8631 if (cpumask_weight(cpu_map
) >
8632 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8633 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8634 SD_INIT(sd
, ALLNODES
);
8635 set_domain_attribute(sd
, attr
);
8636 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8637 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8642 sd
= &per_cpu(node_domains
, i
).sd
;
8644 set_domain_attribute(sd
, attr
);
8645 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8646 sd
->parent
= parent
;
8649 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8654 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8655 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8656 struct sched_domain
*parent
, int i
)
8658 struct sched_domain
*sd
;
8659 sd
= &per_cpu(phys_domains
, i
).sd
;
8661 set_domain_attribute(sd
, attr
);
8662 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8663 sd
->parent
= parent
;
8666 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8670 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8671 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8672 struct sched_domain
*parent
, int i
)
8674 struct sched_domain
*sd
= parent
;
8675 #ifdef CONFIG_SCHED_MC
8676 sd
= &per_cpu(core_domains
, i
).sd
;
8678 set_domain_attribute(sd
, attr
);
8679 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8680 sd
->parent
= parent
;
8682 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8687 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8688 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8689 struct sched_domain
*parent
, int i
)
8691 struct sched_domain
*sd
= parent
;
8692 #ifdef CONFIG_SCHED_SMT
8693 sd
= &per_cpu(cpu_domains
, i
).sd
;
8694 SD_INIT(sd
, SIBLING
);
8695 set_domain_attribute(sd
, attr
);
8696 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8697 sd
->parent
= parent
;
8699 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8704 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8705 const struct cpumask
*cpu_map
, int cpu
)
8708 #ifdef CONFIG_SCHED_SMT
8709 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8710 cpumask_and(d
->this_sibling_map
, cpu_map
,
8711 topology_thread_cpumask(cpu
));
8712 if (cpu
== cpumask_first(d
->this_sibling_map
))
8713 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8715 d
->send_covered
, d
->tmpmask
);
8718 #ifdef CONFIG_SCHED_MC
8719 case SD_LV_MC
: /* set up multi-core groups */
8720 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8721 if (cpu
== cpumask_first(d
->this_core_map
))
8722 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8724 d
->send_covered
, d
->tmpmask
);
8727 case SD_LV_CPU
: /* set up physical groups */
8728 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8729 if (!cpumask_empty(d
->nodemask
))
8730 init_sched_build_groups(d
->nodemask
, cpu_map
,
8732 d
->send_covered
, d
->tmpmask
);
8735 case SD_LV_ALLNODES
:
8736 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8737 d
->send_covered
, d
->tmpmask
);
8746 * Build sched domains for a given set of cpus and attach the sched domains
8747 * to the individual cpus
8749 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8750 struct sched_domain_attr
*attr
)
8752 enum s_alloc alloc_state
= sa_none
;
8754 struct sched_domain
*sd
;
8760 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8761 if (alloc_state
!= sa_rootdomain
)
8763 alloc_state
= sa_sched_groups
;
8766 * Set up domains for cpus specified by the cpu_map.
8768 for_each_cpu(i
, cpu_map
) {
8769 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8772 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8773 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8774 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8775 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8778 for_each_cpu(i
, cpu_map
) {
8779 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8780 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8783 /* Set up physical groups */
8784 for (i
= 0; i
< nr_node_ids
; i
++)
8785 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8788 /* Set up node groups */
8790 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8792 for (i
= 0; i
< nr_node_ids
; i
++)
8793 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8797 /* Calculate CPU power for physical packages and nodes */
8798 #ifdef CONFIG_SCHED_SMT
8799 for_each_cpu(i
, cpu_map
) {
8800 sd
= &per_cpu(cpu_domains
, i
).sd
;
8801 init_sched_groups_power(i
, sd
);
8804 #ifdef CONFIG_SCHED_MC
8805 for_each_cpu(i
, cpu_map
) {
8806 sd
= &per_cpu(core_domains
, i
).sd
;
8807 init_sched_groups_power(i
, sd
);
8811 for_each_cpu(i
, cpu_map
) {
8812 sd
= &per_cpu(phys_domains
, i
).sd
;
8813 init_sched_groups_power(i
, sd
);
8817 for (i
= 0; i
< nr_node_ids
; i
++)
8818 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8820 if (d
.sd_allnodes
) {
8821 struct sched_group
*sg
;
8823 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8825 init_numa_sched_groups_power(sg
);
8829 /* Attach the domains */
8830 for_each_cpu(i
, cpu_map
) {
8831 #ifdef CONFIG_SCHED_SMT
8832 sd
= &per_cpu(cpu_domains
, i
).sd
;
8833 #elif defined(CONFIG_SCHED_MC)
8834 sd
= &per_cpu(core_domains
, i
).sd
;
8836 sd
= &per_cpu(phys_domains
, i
).sd
;
8838 cpu_attach_domain(sd
, d
.rd
, i
);
8841 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8842 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8846 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8850 static int build_sched_domains(const struct cpumask
*cpu_map
)
8852 return __build_sched_domains(cpu_map
, NULL
);
8855 static struct cpumask
*doms_cur
; /* current sched domains */
8856 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8857 static struct sched_domain_attr
*dattr_cur
;
8858 /* attribues of custom domains in 'doms_cur' */
8861 * Special case: If a kmalloc of a doms_cur partition (array of
8862 * cpumask) fails, then fallback to a single sched domain,
8863 * as determined by the single cpumask fallback_doms.
8865 static cpumask_var_t fallback_doms
;
8868 * arch_update_cpu_topology lets virtualized architectures update the
8869 * cpu core maps. It is supposed to return 1 if the topology changed
8870 * or 0 if it stayed the same.
8872 int __attribute__((weak
)) arch_update_cpu_topology(void)
8878 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8879 * For now this just excludes isolated cpus, but could be used to
8880 * exclude other special cases in the future.
8882 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8886 arch_update_cpu_topology();
8888 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8890 doms_cur
= fallback_doms
;
8891 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8893 err
= build_sched_domains(doms_cur
);
8894 register_sched_domain_sysctl();
8899 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8900 struct cpumask
*tmpmask
)
8902 free_sched_groups(cpu_map
, tmpmask
);
8906 * Detach sched domains from a group of cpus specified in cpu_map
8907 * These cpus will now be attached to the NULL domain
8909 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8911 /* Save because hotplug lock held. */
8912 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8915 for_each_cpu(i
, cpu_map
)
8916 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8917 synchronize_sched();
8918 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8921 /* handle null as "default" */
8922 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8923 struct sched_domain_attr
*new, int idx_new
)
8925 struct sched_domain_attr tmp
;
8932 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8933 new ? (new + idx_new
) : &tmp
,
8934 sizeof(struct sched_domain_attr
));
8938 * Partition sched domains as specified by the 'ndoms_new'
8939 * cpumasks in the array doms_new[] of cpumasks. This compares
8940 * doms_new[] to the current sched domain partitioning, doms_cur[].
8941 * It destroys each deleted domain and builds each new domain.
8943 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8944 * The masks don't intersect (don't overlap.) We should setup one
8945 * sched domain for each mask. CPUs not in any of the cpumasks will
8946 * not be load balanced. If the same cpumask appears both in the
8947 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8950 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8951 * ownership of it and will kfree it when done with it. If the caller
8952 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8953 * ndoms_new == 1, and partition_sched_domains() will fallback to
8954 * the single partition 'fallback_doms', it also forces the domains
8957 * If doms_new == NULL it will be replaced with cpu_online_mask.
8958 * ndoms_new == 0 is a special case for destroying existing domains,
8959 * and it will not create the default domain.
8961 * Call with hotplug lock held
8963 /* FIXME: Change to struct cpumask *doms_new[] */
8964 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8965 struct sched_domain_attr
*dattr_new
)
8970 mutex_lock(&sched_domains_mutex
);
8972 /* always unregister in case we don't destroy any domains */
8973 unregister_sched_domain_sysctl();
8975 /* Let architecture update cpu core mappings. */
8976 new_topology
= arch_update_cpu_topology();
8978 n
= doms_new
? ndoms_new
: 0;
8980 /* Destroy deleted domains */
8981 for (i
= 0; i
< ndoms_cur
; i
++) {
8982 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8983 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8984 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8987 /* no match - a current sched domain not in new doms_new[] */
8988 detach_destroy_domains(doms_cur
+ i
);
8993 if (doms_new
== NULL
) {
8995 doms_new
= fallback_doms
;
8996 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8997 WARN_ON_ONCE(dattr_new
);
9000 /* Build new domains */
9001 for (i
= 0; i
< ndoms_new
; i
++) {
9002 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9003 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9004 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9007 /* no match - add a new doms_new */
9008 __build_sched_domains(doms_new
+ i
,
9009 dattr_new
? dattr_new
+ i
: NULL
);
9014 /* Remember the new sched domains */
9015 if (doms_cur
!= fallback_doms
)
9017 kfree(dattr_cur
); /* kfree(NULL) is safe */
9018 doms_cur
= doms_new
;
9019 dattr_cur
= dattr_new
;
9020 ndoms_cur
= ndoms_new
;
9022 register_sched_domain_sysctl();
9024 mutex_unlock(&sched_domains_mutex
);
9027 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9028 static void arch_reinit_sched_domains(void)
9032 /* Destroy domains first to force the rebuild */
9033 partition_sched_domains(0, NULL
, NULL
);
9035 rebuild_sched_domains();
9039 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9041 unsigned int level
= 0;
9043 if (sscanf(buf
, "%u", &level
) != 1)
9047 * level is always be positive so don't check for
9048 * level < POWERSAVINGS_BALANCE_NONE which is 0
9049 * What happens on 0 or 1 byte write,
9050 * need to check for count as well?
9053 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9057 sched_smt_power_savings
= level
;
9059 sched_mc_power_savings
= level
;
9061 arch_reinit_sched_domains();
9066 #ifdef CONFIG_SCHED_MC
9067 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9070 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9072 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9073 const char *buf
, size_t count
)
9075 return sched_power_savings_store(buf
, count
, 0);
9077 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9078 sched_mc_power_savings_show
,
9079 sched_mc_power_savings_store
);
9082 #ifdef CONFIG_SCHED_SMT
9083 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9086 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9088 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9089 const char *buf
, size_t count
)
9091 return sched_power_savings_store(buf
, count
, 1);
9093 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9094 sched_smt_power_savings_show
,
9095 sched_smt_power_savings_store
);
9098 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9102 #ifdef CONFIG_SCHED_SMT
9104 err
= sysfs_create_file(&cls
->kset
.kobj
,
9105 &attr_sched_smt_power_savings
.attr
);
9107 #ifdef CONFIG_SCHED_MC
9108 if (!err
&& mc_capable())
9109 err
= sysfs_create_file(&cls
->kset
.kobj
,
9110 &attr_sched_mc_power_savings
.attr
);
9114 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9116 #ifndef CONFIG_CPUSETS
9118 * Add online and remove offline CPUs from the scheduler domains.
9119 * When cpusets are enabled they take over this function.
9121 static int update_sched_domains(struct notifier_block
*nfb
,
9122 unsigned long action
, void *hcpu
)
9126 case CPU_ONLINE_FROZEN
:
9128 case CPU_DEAD_FROZEN
:
9129 partition_sched_domains(1, NULL
, NULL
);
9138 static int update_runtime(struct notifier_block
*nfb
,
9139 unsigned long action
, void *hcpu
)
9141 int cpu
= (int)(long)hcpu
;
9144 case CPU_DOWN_PREPARE
:
9145 case CPU_DOWN_PREPARE_FROZEN
:
9146 disable_runtime(cpu_rq(cpu
));
9149 case CPU_DOWN_FAILED
:
9150 case CPU_DOWN_FAILED_FROZEN
:
9152 case CPU_ONLINE_FROZEN
:
9153 enable_runtime(cpu_rq(cpu
));
9161 void __init
sched_init_smp(void)
9163 cpumask_var_t non_isolated_cpus
;
9165 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9166 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9168 #if defined(CONFIG_NUMA)
9169 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9171 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9174 mutex_lock(&sched_domains_mutex
);
9175 arch_init_sched_domains(cpu_online_mask
);
9176 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9177 if (cpumask_empty(non_isolated_cpus
))
9178 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9179 mutex_unlock(&sched_domains_mutex
);
9182 #ifndef CONFIG_CPUSETS
9183 /* XXX: Theoretical race here - CPU may be hotplugged now */
9184 hotcpu_notifier(update_sched_domains
, 0);
9187 /* RT runtime code needs to handle some hotplug events */
9188 hotcpu_notifier(update_runtime
, 0);
9192 /* Move init over to a non-isolated CPU */
9193 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9195 sched_init_granularity();
9196 free_cpumask_var(non_isolated_cpus
);
9198 init_sched_rt_class();
9201 void __init
sched_init_smp(void)
9203 sched_init_granularity();
9205 #endif /* CONFIG_SMP */
9207 const_debug
unsigned int sysctl_timer_migration
= 1;
9209 int in_sched_functions(unsigned long addr
)
9211 return in_lock_functions(addr
) ||
9212 (addr
>= (unsigned long)__sched_text_start
9213 && addr
< (unsigned long)__sched_text_end
);
9216 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9218 cfs_rq
->tasks_timeline
= RB_ROOT
;
9219 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9220 #ifdef CONFIG_FAIR_GROUP_SCHED
9223 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9226 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9228 struct rt_prio_array
*array
;
9231 array
= &rt_rq
->active
;
9232 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9233 INIT_LIST_HEAD(array
->queue
+ i
);
9234 __clear_bit(i
, array
->bitmap
);
9236 /* delimiter for bitsearch: */
9237 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9239 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9240 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9242 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9246 rt_rq
->rt_nr_migratory
= 0;
9247 rt_rq
->overloaded
= 0;
9248 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9252 rt_rq
->rt_throttled
= 0;
9253 rt_rq
->rt_runtime
= 0;
9254 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9256 #ifdef CONFIG_RT_GROUP_SCHED
9257 rt_rq
->rt_nr_boosted
= 0;
9262 #ifdef CONFIG_FAIR_GROUP_SCHED
9263 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9264 struct sched_entity
*se
, int cpu
, int add
,
9265 struct sched_entity
*parent
)
9267 struct rq
*rq
= cpu_rq(cpu
);
9268 tg
->cfs_rq
[cpu
] = cfs_rq
;
9269 init_cfs_rq(cfs_rq
, rq
);
9272 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9275 /* se could be NULL for init_task_group */
9280 se
->cfs_rq
= &rq
->cfs
;
9282 se
->cfs_rq
= parent
->my_q
;
9285 se
->load
.weight
= tg
->shares
;
9286 se
->load
.inv_weight
= 0;
9287 se
->parent
= parent
;
9291 #ifdef CONFIG_RT_GROUP_SCHED
9292 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9293 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9294 struct sched_rt_entity
*parent
)
9296 struct rq
*rq
= cpu_rq(cpu
);
9298 tg
->rt_rq
[cpu
] = rt_rq
;
9299 init_rt_rq(rt_rq
, rq
);
9301 rt_rq
->rt_se
= rt_se
;
9302 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9304 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9306 tg
->rt_se
[cpu
] = rt_se
;
9311 rt_se
->rt_rq
= &rq
->rt
;
9313 rt_se
->rt_rq
= parent
->my_q
;
9315 rt_se
->my_q
= rt_rq
;
9316 rt_se
->parent
= parent
;
9317 INIT_LIST_HEAD(&rt_se
->run_list
);
9321 void __init
sched_init(void)
9324 unsigned long alloc_size
= 0, ptr
;
9326 #ifdef CONFIG_FAIR_GROUP_SCHED
9327 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9329 #ifdef CONFIG_RT_GROUP_SCHED
9330 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9332 #ifdef CONFIG_USER_SCHED
9335 #ifdef CONFIG_CPUMASK_OFFSTACK
9336 alloc_size
+= num_possible_cpus() * cpumask_size();
9339 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9341 #ifdef CONFIG_FAIR_GROUP_SCHED
9342 init_task_group
.se
= (struct sched_entity
**)ptr
;
9343 ptr
+= nr_cpu_ids
* sizeof(void **);
9345 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9346 ptr
+= nr_cpu_ids
* sizeof(void **);
9348 #ifdef CONFIG_USER_SCHED
9349 root_task_group
.se
= (struct sched_entity
**)ptr
;
9350 ptr
+= nr_cpu_ids
* sizeof(void **);
9352 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9353 ptr
+= nr_cpu_ids
* sizeof(void **);
9354 #endif /* CONFIG_USER_SCHED */
9355 #endif /* CONFIG_FAIR_GROUP_SCHED */
9356 #ifdef CONFIG_RT_GROUP_SCHED
9357 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9358 ptr
+= nr_cpu_ids
* sizeof(void **);
9360 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9361 ptr
+= nr_cpu_ids
* sizeof(void **);
9363 #ifdef CONFIG_USER_SCHED
9364 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9365 ptr
+= nr_cpu_ids
* sizeof(void **);
9367 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9368 ptr
+= nr_cpu_ids
* sizeof(void **);
9369 #endif /* CONFIG_USER_SCHED */
9370 #endif /* CONFIG_RT_GROUP_SCHED */
9371 #ifdef CONFIG_CPUMASK_OFFSTACK
9372 for_each_possible_cpu(i
) {
9373 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9374 ptr
+= cpumask_size();
9376 #endif /* CONFIG_CPUMASK_OFFSTACK */
9380 init_defrootdomain();
9383 init_rt_bandwidth(&def_rt_bandwidth
,
9384 global_rt_period(), global_rt_runtime());
9386 #ifdef CONFIG_RT_GROUP_SCHED
9387 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9388 global_rt_period(), global_rt_runtime());
9389 #ifdef CONFIG_USER_SCHED
9390 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9391 global_rt_period(), RUNTIME_INF
);
9392 #endif /* CONFIG_USER_SCHED */
9393 #endif /* CONFIG_RT_GROUP_SCHED */
9395 #ifdef CONFIG_GROUP_SCHED
9396 list_add(&init_task_group
.list
, &task_groups
);
9397 INIT_LIST_HEAD(&init_task_group
.children
);
9399 #ifdef CONFIG_USER_SCHED
9400 INIT_LIST_HEAD(&root_task_group
.children
);
9401 init_task_group
.parent
= &root_task_group
;
9402 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9403 #endif /* CONFIG_USER_SCHED */
9404 #endif /* CONFIG_GROUP_SCHED */
9406 for_each_possible_cpu(i
) {
9410 spin_lock_init(&rq
->lock
);
9412 rq
->calc_load_active
= 0;
9413 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9414 init_cfs_rq(&rq
->cfs
, rq
);
9415 init_rt_rq(&rq
->rt
, rq
);
9416 #ifdef CONFIG_FAIR_GROUP_SCHED
9417 init_task_group
.shares
= init_task_group_load
;
9418 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9419 #ifdef CONFIG_CGROUP_SCHED
9421 * How much cpu bandwidth does init_task_group get?
9423 * In case of task-groups formed thr' the cgroup filesystem, it
9424 * gets 100% of the cpu resources in the system. This overall
9425 * system cpu resource is divided among the tasks of
9426 * init_task_group and its child task-groups in a fair manner,
9427 * based on each entity's (task or task-group's) weight
9428 * (se->load.weight).
9430 * In other words, if init_task_group has 10 tasks of weight
9431 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9432 * then A0's share of the cpu resource is:
9434 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9436 * We achieve this by letting init_task_group's tasks sit
9437 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9439 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9440 #elif defined CONFIG_USER_SCHED
9441 root_task_group
.shares
= NICE_0_LOAD
;
9442 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9444 * In case of task-groups formed thr' the user id of tasks,
9445 * init_task_group represents tasks belonging to root user.
9446 * Hence it forms a sibling of all subsequent groups formed.
9447 * In this case, init_task_group gets only a fraction of overall
9448 * system cpu resource, based on the weight assigned to root
9449 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9450 * by letting tasks of init_task_group sit in a separate cfs_rq
9451 * (init_tg_cfs_rq) and having one entity represent this group of
9452 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9454 init_tg_cfs_entry(&init_task_group
,
9455 &per_cpu(init_tg_cfs_rq
, i
),
9456 &per_cpu(init_sched_entity
, i
), i
, 1,
9457 root_task_group
.se
[i
]);
9460 #endif /* CONFIG_FAIR_GROUP_SCHED */
9462 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9463 #ifdef CONFIG_RT_GROUP_SCHED
9464 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9465 #ifdef CONFIG_CGROUP_SCHED
9466 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9467 #elif defined CONFIG_USER_SCHED
9468 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9469 init_tg_rt_entry(&init_task_group
,
9470 &per_cpu(init_rt_rq
, i
),
9471 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9472 root_task_group
.rt_se
[i
]);
9476 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9477 rq
->cpu_load
[j
] = 0;
9481 rq
->post_schedule
= 0;
9482 rq
->active_balance
= 0;
9483 rq
->next_balance
= jiffies
;
9487 rq
->migration_thread
= NULL
;
9488 INIT_LIST_HEAD(&rq
->migration_queue
);
9489 rq_attach_root(rq
, &def_root_domain
);
9492 atomic_set(&rq
->nr_iowait
, 0);
9495 set_load_weight(&init_task
);
9497 #ifdef CONFIG_PREEMPT_NOTIFIERS
9498 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9502 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9505 #ifdef CONFIG_RT_MUTEXES
9506 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9510 * The boot idle thread does lazy MMU switching as well:
9512 atomic_inc(&init_mm
.mm_count
);
9513 enter_lazy_tlb(&init_mm
, current
);
9516 * Make us the idle thread. Technically, schedule() should not be
9517 * called from this thread, however somewhere below it might be,
9518 * but because we are the idle thread, we just pick up running again
9519 * when this runqueue becomes "idle".
9521 init_idle(current
, smp_processor_id());
9523 calc_load_update
= jiffies
+ LOAD_FREQ
;
9526 * During early bootup we pretend to be a normal task:
9528 current
->sched_class
= &fair_sched_class
;
9530 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9531 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9534 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9535 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9537 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9542 scheduler_running
= 1;
9545 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9546 static inline int preempt_count_equals(int preempt_offset
)
9548 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9550 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9553 void __might_sleep(char *file
, int line
, int preempt_offset
)
9556 static unsigned long prev_jiffy
; /* ratelimiting */
9558 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9559 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9561 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9563 prev_jiffy
= jiffies
;
9566 "BUG: sleeping function called from invalid context at %s:%d\n",
9569 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9570 in_atomic(), irqs_disabled(),
9571 current
->pid
, current
->comm
);
9573 debug_show_held_locks(current
);
9574 if (irqs_disabled())
9575 print_irqtrace_events(current
);
9579 EXPORT_SYMBOL(__might_sleep
);
9582 #ifdef CONFIG_MAGIC_SYSRQ
9583 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9587 update_rq_clock(rq
);
9588 on_rq
= p
->se
.on_rq
;
9590 deactivate_task(rq
, p
, 0);
9591 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9593 activate_task(rq
, p
, 0);
9594 resched_task(rq
->curr
);
9598 void normalize_rt_tasks(void)
9600 struct task_struct
*g
, *p
;
9601 unsigned long flags
;
9604 read_lock_irqsave(&tasklist_lock
, flags
);
9605 do_each_thread(g
, p
) {
9607 * Only normalize user tasks:
9612 p
->se
.exec_start
= 0;
9613 #ifdef CONFIG_SCHEDSTATS
9614 p
->se
.wait_start
= 0;
9615 p
->se
.sleep_start
= 0;
9616 p
->se
.block_start
= 0;
9621 * Renice negative nice level userspace
9624 if (TASK_NICE(p
) < 0 && p
->mm
)
9625 set_user_nice(p
, 0);
9629 spin_lock(&p
->pi_lock
);
9630 rq
= __task_rq_lock(p
);
9632 normalize_task(rq
, p
);
9634 __task_rq_unlock(rq
);
9635 spin_unlock(&p
->pi_lock
);
9636 } while_each_thread(g
, p
);
9638 read_unlock_irqrestore(&tasklist_lock
, flags
);
9641 #endif /* CONFIG_MAGIC_SYSRQ */
9645 * These functions are only useful for the IA64 MCA handling.
9647 * They can only be called when the whole system has been
9648 * stopped - every CPU needs to be quiescent, and no scheduling
9649 * activity can take place. Using them for anything else would
9650 * be a serious bug, and as a result, they aren't even visible
9651 * under any other configuration.
9655 * curr_task - return the current task for a given cpu.
9656 * @cpu: the processor in question.
9658 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9660 struct task_struct
*curr_task(int cpu
)
9662 return cpu_curr(cpu
);
9666 * set_curr_task - set the current task for a given cpu.
9667 * @cpu: the processor in question.
9668 * @p: the task pointer to set.
9670 * Description: This function must only be used when non-maskable interrupts
9671 * are serviced on a separate stack. It allows the architecture to switch the
9672 * notion of the current task on a cpu in a non-blocking manner. This function
9673 * must be called with all CPU's synchronized, and interrupts disabled, the
9674 * and caller must save the original value of the current task (see
9675 * curr_task() above) and restore that value before reenabling interrupts and
9676 * re-starting the system.
9678 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9680 void set_curr_task(int cpu
, struct task_struct
*p
)
9687 #ifdef CONFIG_FAIR_GROUP_SCHED
9688 static void free_fair_sched_group(struct task_group
*tg
)
9692 for_each_possible_cpu(i
) {
9694 kfree(tg
->cfs_rq
[i
]);
9704 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9706 struct cfs_rq
*cfs_rq
;
9707 struct sched_entity
*se
;
9711 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9714 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9718 tg
->shares
= NICE_0_LOAD
;
9720 for_each_possible_cpu(i
) {
9723 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9724 GFP_KERNEL
, cpu_to_node(i
));
9728 se
= kzalloc_node(sizeof(struct sched_entity
),
9729 GFP_KERNEL
, cpu_to_node(i
));
9733 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9742 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9744 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9745 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9748 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9750 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9752 #else /* !CONFG_FAIR_GROUP_SCHED */
9753 static inline void free_fair_sched_group(struct task_group
*tg
)
9758 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9763 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9767 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9770 #endif /* CONFIG_FAIR_GROUP_SCHED */
9772 #ifdef CONFIG_RT_GROUP_SCHED
9773 static void free_rt_sched_group(struct task_group
*tg
)
9777 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9779 for_each_possible_cpu(i
) {
9781 kfree(tg
->rt_rq
[i
]);
9783 kfree(tg
->rt_se
[i
]);
9791 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9793 struct rt_rq
*rt_rq
;
9794 struct sched_rt_entity
*rt_se
;
9798 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9801 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9805 init_rt_bandwidth(&tg
->rt_bandwidth
,
9806 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9808 for_each_possible_cpu(i
) {
9811 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9812 GFP_KERNEL
, cpu_to_node(i
));
9816 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9817 GFP_KERNEL
, cpu_to_node(i
));
9821 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9830 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9832 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9833 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9836 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9838 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9840 #else /* !CONFIG_RT_GROUP_SCHED */
9841 static inline void free_rt_sched_group(struct task_group
*tg
)
9846 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9851 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9855 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9858 #endif /* CONFIG_RT_GROUP_SCHED */
9860 #ifdef CONFIG_GROUP_SCHED
9861 static void free_sched_group(struct task_group
*tg
)
9863 free_fair_sched_group(tg
);
9864 free_rt_sched_group(tg
);
9868 /* allocate runqueue etc for a new task group */
9869 struct task_group
*sched_create_group(struct task_group
*parent
)
9871 struct task_group
*tg
;
9872 unsigned long flags
;
9875 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9877 return ERR_PTR(-ENOMEM
);
9879 if (!alloc_fair_sched_group(tg
, parent
))
9882 if (!alloc_rt_sched_group(tg
, parent
))
9885 spin_lock_irqsave(&task_group_lock
, flags
);
9886 for_each_possible_cpu(i
) {
9887 register_fair_sched_group(tg
, i
);
9888 register_rt_sched_group(tg
, i
);
9890 list_add_rcu(&tg
->list
, &task_groups
);
9892 WARN_ON(!parent
); /* root should already exist */
9894 tg
->parent
= parent
;
9895 INIT_LIST_HEAD(&tg
->children
);
9896 list_add_rcu(&tg
->siblings
, &parent
->children
);
9897 spin_unlock_irqrestore(&task_group_lock
, flags
);
9902 free_sched_group(tg
);
9903 return ERR_PTR(-ENOMEM
);
9906 /* rcu callback to free various structures associated with a task group */
9907 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9909 /* now it should be safe to free those cfs_rqs */
9910 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9913 /* Destroy runqueue etc associated with a task group */
9914 void sched_destroy_group(struct task_group
*tg
)
9916 unsigned long flags
;
9919 spin_lock_irqsave(&task_group_lock
, flags
);
9920 for_each_possible_cpu(i
) {
9921 unregister_fair_sched_group(tg
, i
);
9922 unregister_rt_sched_group(tg
, i
);
9924 list_del_rcu(&tg
->list
);
9925 list_del_rcu(&tg
->siblings
);
9926 spin_unlock_irqrestore(&task_group_lock
, flags
);
9928 /* wait for possible concurrent references to cfs_rqs complete */
9929 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9932 /* change task's runqueue when it moves between groups.
9933 * The caller of this function should have put the task in its new group
9934 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9935 * reflect its new group.
9937 void sched_move_task(struct task_struct
*tsk
)
9940 unsigned long flags
;
9943 rq
= task_rq_lock(tsk
, &flags
);
9945 update_rq_clock(rq
);
9947 running
= task_current(rq
, tsk
);
9948 on_rq
= tsk
->se
.on_rq
;
9951 dequeue_task(rq
, tsk
, 0);
9952 if (unlikely(running
))
9953 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9955 set_task_rq(tsk
, task_cpu(tsk
));
9957 #ifdef CONFIG_FAIR_GROUP_SCHED
9958 if (tsk
->sched_class
->moved_group
)
9959 tsk
->sched_class
->moved_group(tsk
);
9962 if (unlikely(running
))
9963 tsk
->sched_class
->set_curr_task(rq
);
9965 enqueue_task(rq
, tsk
, 0);
9967 task_rq_unlock(rq
, &flags
);
9969 #endif /* CONFIG_GROUP_SCHED */
9971 #ifdef CONFIG_FAIR_GROUP_SCHED
9972 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9974 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9979 dequeue_entity(cfs_rq
, se
, 0);
9981 se
->load
.weight
= shares
;
9982 se
->load
.inv_weight
= 0;
9985 enqueue_entity(cfs_rq
, se
, 0);
9988 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9990 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9991 struct rq
*rq
= cfs_rq
->rq
;
9992 unsigned long flags
;
9994 spin_lock_irqsave(&rq
->lock
, flags
);
9995 __set_se_shares(se
, shares
);
9996 spin_unlock_irqrestore(&rq
->lock
, flags
);
9999 static DEFINE_MUTEX(shares_mutex
);
10001 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10004 unsigned long flags
;
10007 * We can't change the weight of the root cgroup.
10012 if (shares
< MIN_SHARES
)
10013 shares
= MIN_SHARES
;
10014 else if (shares
> MAX_SHARES
)
10015 shares
= MAX_SHARES
;
10017 mutex_lock(&shares_mutex
);
10018 if (tg
->shares
== shares
)
10021 spin_lock_irqsave(&task_group_lock
, flags
);
10022 for_each_possible_cpu(i
)
10023 unregister_fair_sched_group(tg
, i
);
10024 list_del_rcu(&tg
->siblings
);
10025 spin_unlock_irqrestore(&task_group_lock
, flags
);
10027 /* wait for any ongoing reference to this group to finish */
10028 synchronize_sched();
10031 * Now we are free to modify the group's share on each cpu
10032 * w/o tripping rebalance_share or load_balance_fair.
10034 tg
->shares
= shares
;
10035 for_each_possible_cpu(i
) {
10037 * force a rebalance
10039 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10040 set_se_shares(tg
->se
[i
], shares
);
10044 * Enable load balance activity on this group, by inserting it back on
10045 * each cpu's rq->leaf_cfs_rq_list.
10047 spin_lock_irqsave(&task_group_lock
, flags
);
10048 for_each_possible_cpu(i
)
10049 register_fair_sched_group(tg
, i
);
10050 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10051 spin_unlock_irqrestore(&task_group_lock
, flags
);
10053 mutex_unlock(&shares_mutex
);
10057 unsigned long sched_group_shares(struct task_group
*tg
)
10063 #ifdef CONFIG_RT_GROUP_SCHED
10065 * Ensure that the real time constraints are schedulable.
10067 static DEFINE_MUTEX(rt_constraints_mutex
);
10069 static unsigned long to_ratio(u64 period
, u64 runtime
)
10071 if (runtime
== RUNTIME_INF
)
10074 return div64_u64(runtime
<< 20, period
);
10077 /* Must be called with tasklist_lock held */
10078 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10080 struct task_struct
*g
, *p
;
10082 do_each_thread(g
, p
) {
10083 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10085 } while_each_thread(g
, p
);
10090 struct rt_schedulable_data
{
10091 struct task_group
*tg
;
10096 static int tg_schedulable(struct task_group
*tg
, void *data
)
10098 struct rt_schedulable_data
*d
= data
;
10099 struct task_group
*child
;
10100 unsigned long total
, sum
= 0;
10101 u64 period
, runtime
;
10103 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10104 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10107 period
= d
->rt_period
;
10108 runtime
= d
->rt_runtime
;
10111 #ifdef CONFIG_USER_SCHED
10112 if (tg
== &root_task_group
) {
10113 period
= global_rt_period();
10114 runtime
= global_rt_runtime();
10119 * Cannot have more runtime than the period.
10121 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10125 * Ensure we don't starve existing RT tasks.
10127 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10130 total
= to_ratio(period
, runtime
);
10133 * Nobody can have more than the global setting allows.
10135 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10139 * The sum of our children's runtime should not exceed our own.
10141 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10142 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10143 runtime
= child
->rt_bandwidth
.rt_runtime
;
10145 if (child
== d
->tg
) {
10146 period
= d
->rt_period
;
10147 runtime
= d
->rt_runtime
;
10150 sum
+= to_ratio(period
, runtime
);
10159 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10161 struct rt_schedulable_data data
= {
10163 .rt_period
= period
,
10164 .rt_runtime
= runtime
,
10167 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10170 static int tg_set_bandwidth(struct task_group
*tg
,
10171 u64 rt_period
, u64 rt_runtime
)
10175 mutex_lock(&rt_constraints_mutex
);
10176 read_lock(&tasklist_lock
);
10177 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10181 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10182 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10183 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10185 for_each_possible_cpu(i
) {
10186 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10188 spin_lock(&rt_rq
->rt_runtime_lock
);
10189 rt_rq
->rt_runtime
= rt_runtime
;
10190 spin_unlock(&rt_rq
->rt_runtime_lock
);
10192 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10194 read_unlock(&tasklist_lock
);
10195 mutex_unlock(&rt_constraints_mutex
);
10200 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10202 u64 rt_runtime
, rt_period
;
10204 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10205 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10206 if (rt_runtime_us
< 0)
10207 rt_runtime
= RUNTIME_INF
;
10209 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10212 long sched_group_rt_runtime(struct task_group
*tg
)
10216 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10219 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10220 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10221 return rt_runtime_us
;
10224 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10226 u64 rt_runtime
, rt_period
;
10228 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10229 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10231 if (rt_period
== 0)
10234 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10237 long sched_group_rt_period(struct task_group
*tg
)
10241 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10242 do_div(rt_period_us
, NSEC_PER_USEC
);
10243 return rt_period_us
;
10246 static int sched_rt_global_constraints(void)
10248 u64 runtime
, period
;
10251 if (sysctl_sched_rt_period
<= 0)
10254 runtime
= global_rt_runtime();
10255 period
= global_rt_period();
10258 * Sanity check on the sysctl variables.
10260 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10263 mutex_lock(&rt_constraints_mutex
);
10264 read_lock(&tasklist_lock
);
10265 ret
= __rt_schedulable(NULL
, 0, 0);
10266 read_unlock(&tasklist_lock
);
10267 mutex_unlock(&rt_constraints_mutex
);
10272 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10274 /* Don't accept realtime tasks when there is no way for them to run */
10275 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10281 #else /* !CONFIG_RT_GROUP_SCHED */
10282 static int sched_rt_global_constraints(void)
10284 unsigned long flags
;
10287 if (sysctl_sched_rt_period
<= 0)
10291 * There's always some RT tasks in the root group
10292 * -- migration, kstopmachine etc..
10294 if (sysctl_sched_rt_runtime
== 0)
10297 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10298 for_each_possible_cpu(i
) {
10299 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10301 spin_lock(&rt_rq
->rt_runtime_lock
);
10302 rt_rq
->rt_runtime
= global_rt_runtime();
10303 spin_unlock(&rt_rq
->rt_runtime_lock
);
10305 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10309 #endif /* CONFIG_RT_GROUP_SCHED */
10311 int sched_rt_handler(struct ctl_table
*table
, int write
,
10312 void __user
*buffer
, size_t *lenp
,
10316 int old_period
, old_runtime
;
10317 static DEFINE_MUTEX(mutex
);
10319 mutex_lock(&mutex
);
10320 old_period
= sysctl_sched_rt_period
;
10321 old_runtime
= sysctl_sched_rt_runtime
;
10323 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10325 if (!ret
&& write
) {
10326 ret
= sched_rt_global_constraints();
10328 sysctl_sched_rt_period
= old_period
;
10329 sysctl_sched_rt_runtime
= old_runtime
;
10331 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10332 def_rt_bandwidth
.rt_period
=
10333 ns_to_ktime(global_rt_period());
10336 mutex_unlock(&mutex
);
10341 #ifdef CONFIG_CGROUP_SCHED
10343 /* return corresponding task_group object of a cgroup */
10344 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10346 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10347 struct task_group
, css
);
10350 static struct cgroup_subsys_state
*
10351 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10353 struct task_group
*tg
, *parent
;
10355 if (!cgrp
->parent
) {
10356 /* This is early initialization for the top cgroup */
10357 return &init_task_group
.css
;
10360 parent
= cgroup_tg(cgrp
->parent
);
10361 tg
= sched_create_group(parent
);
10363 return ERR_PTR(-ENOMEM
);
10369 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10371 struct task_group
*tg
= cgroup_tg(cgrp
);
10373 sched_destroy_group(tg
);
10377 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10379 #ifdef CONFIG_RT_GROUP_SCHED
10380 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10383 /* We don't support RT-tasks being in separate groups */
10384 if (tsk
->sched_class
!= &fair_sched_class
)
10391 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10392 struct task_struct
*tsk
, bool threadgroup
)
10394 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10398 struct task_struct
*c
;
10400 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10401 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10413 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10414 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10417 sched_move_task(tsk
);
10419 struct task_struct
*c
;
10421 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10422 sched_move_task(c
);
10428 #ifdef CONFIG_FAIR_GROUP_SCHED
10429 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10432 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10435 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10437 struct task_group
*tg
= cgroup_tg(cgrp
);
10439 return (u64
) tg
->shares
;
10441 #endif /* CONFIG_FAIR_GROUP_SCHED */
10443 #ifdef CONFIG_RT_GROUP_SCHED
10444 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10447 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10450 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10452 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10455 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10458 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10461 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10463 return sched_group_rt_period(cgroup_tg(cgrp
));
10465 #endif /* CONFIG_RT_GROUP_SCHED */
10467 static struct cftype cpu_files
[] = {
10468 #ifdef CONFIG_FAIR_GROUP_SCHED
10471 .read_u64
= cpu_shares_read_u64
,
10472 .write_u64
= cpu_shares_write_u64
,
10475 #ifdef CONFIG_RT_GROUP_SCHED
10477 .name
= "rt_runtime_us",
10478 .read_s64
= cpu_rt_runtime_read
,
10479 .write_s64
= cpu_rt_runtime_write
,
10482 .name
= "rt_period_us",
10483 .read_u64
= cpu_rt_period_read_uint
,
10484 .write_u64
= cpu_rt_period_write_uint
,
10489 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10491 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10494 struct cgroup_subsys cpu_cgroup_subsys
= {
10496 .create
= cpu_cgroup_create
,
10497 .destroy
= cpu_cgroup_destroy
,
10498 .can_attach
= cpu_cgroup_can_attach
,
10499 .attach
= cpu_cgroup_attach
,
10500 .populate
= cpu_cgroup_populate
,
10501 .subsys_id
= cpu_cgroup_subsys_id
,
10505 #endif /* CONFIG_CGROUP_SCHED */
10507 #ifdef CONFIG_CGROUP_CPUACCT
10510 * CPU accounting code for task groups.
10512 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10513 * (balbir@in.ibm.com).
10516 /* track cpu usage of a group of tasks and its child groups */
10518 struct cgroup_subsys_state css
;
10519 /* cpuusage holds pointer to a u64-type object on every cpu */
10521 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10522 struct cpuacct
*parent
;
10525 struct cgroup_subsys cpuacct_subsys
;
10527 /* return cpu accounting group corresponding to this container */
10528 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10530 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10531 struct cpuacct
, css
);
10534 /* return cpu accounting group to which this task belongs */
10535 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10537 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10538 struct cpuacct
, css
);
10541 /* create a new cpu accounting group */
10542 static struct cgroup_subsys_state
*cpuacct_create(
10543 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10545 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10551 ca
->cpuusage
= alloc_percpu(u64
);
10555 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10556 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10557 goto out_free_counters
;
10560 ca
->parent
= cgroup_ca(cgrp
->parent
);
10566 percpu_counter_destroy(&ca
->cpustat
[i
]);
10567 free_percpu(ca
->cpuusage
);
10571 return ERR_PTR(-ENOMEM
);
10574 /* destroy an existing cpu accounting group */
10576 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10578 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10581 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10582 percpu_counter_destroy(&ca
->cpustat
[i
]);
10583 free_percpu(ca
->cpuusage
);
10587 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10589 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10592 #ifndef CONFIG_64BIT
10594 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10596 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10598 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10606 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10608 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10610 #ifndef CONFIG_64BIT
10612 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10614 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10616 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10622 /* return total cpu usage (in nanoseconds) of a group */
10623 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10625 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10626 u64 totalcpuusage
= 0;
10629 for_each_present_cpu(i
)
10630 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10632 return totalcpuusage
;
10635 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10638 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10647 for_each_present_cpu(i
)
10648 cpuacct_cpuusage_write(ca
, i
, 0);
10654 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10655 struct seq_file
*m
)
10657 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10661 for_each_present_cpu(i
) {
10662 percpu
= cpuacct_cpuusage_read(ca
, i
);
10663 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10665 seq_printf(m
, "\n");
10669 static const char *cpuacct_stat_desc
[] = {
10670 [CPUACCT_STAT_USER
] = "user",
10671 [CPUACCT_STAT_SYSTEM
] = "system",
10674 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10675 struct cgroup_map_cb
*cb
)
10677 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10680 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10681 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10682 val
= cputime64_to_clock_t(val
);
10683 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10688 static struct cftype files
[] = {
10691 .read_u64
= cpuusage_read
,
10692 .write_u64
= cpuusage_write
,
10695 .name
= "usage_percpu",
10696 .read_seq_string
= cpuacct_percpu_seq_read
,
10700 .read_map
= cpuacct_stats_show
,
10704 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10706 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10710 * charge this task's execution time to its accounting group.
10712 * called with rq->lock held.
10714 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10716 struct cpuacct
*ca
;
10719 if (unlikely(!cpuacct_subsys
.active
))
10722 cpu
= task_cpu(tsk
);
10728 for (; ca
; ca
= ca
->parent
) {
10729 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10730 *cpuusage
+= cputime
;
10737 * Charge the system/user time to the task's accounting group.
10739 static void cpuacct_update_stats(struct task_struct
*tsk
,
10740 enum cpuacct_stat_index idx
, cputime_t val
)
10742 struct cpuacct
*ca
;
10744 if (unlikely(!cpuacct_subsys
.active
))
10751 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10757 struct cgroup_subsys cpuacct_subsys
= {
10759 .create
= cpuacct_create
,
10760 .destroy
= cpuacct_destroy
,
10761 .populate
= cpuacct_populate
,
10762 .subsys_id
= cpuacct_subsys_id
,
10764 #endif /* CONFIG_CGROUP_CPUACCT */
10768 int rcu_expedited_torture_stats(char *page
)
10772 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10774 void synchronize_sched_expedited(void)
10777 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10779 #else /* #ifndef CONFIG_SMP */
10781 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10782 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10784 #define RCU_EXPEDITED_STATE_POST -2
10785 #define RCU_EXPEDITED_STATE_IDLE -1
10787 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10789 int rcu_expedited_torture_stats(char *page
)
10794 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10795 for_each_online_cpu(cpu
) {
10796 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10797 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10799 cnt
+= sprintf(&page
[cnt
], "\n");
10802 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10804 static long synchronize_sched_expedited_count
;
10807 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10808 * approach to force grace period to end quickly. This consumes
10809 * significant time on all CPUs, and is thus not recommended for
10810 * any sort of common-case code.
10812 * Note that it is illegal to call this function while holding any
10813 * lock that is acquired by a CPU-hotplug notifier. Failing to
10814 * observe this restriction will result in deadlock.
10816 void synchronize_sched_expedited(void)
10819 unsigned long flags
;
10820 bool need_full_sync
= 0;
10822 struct migration_req
*req
;
10826 smp_mb(); /* ensure prior mod happens before capturing snap. */
10827 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10829 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10831 if (trycount
++ < 10)
10832 udelay(trycount
* num_online_cpus());
10834 synchronize_sched();
10837 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10838 smp_mb(); /* ensure test happens before caller kfree */
10843 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10844 for_each_online_cpu(cpu
) {
10846 req
= &per_cpu(rcu_migration_req
, cpu
);
10847 init_completion(&req
->done
);
10849 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10850 spin_lock_irqsave(&rq
->lock
, flags
);
10851 list_add(&req
->list
, &rq
->migration_queue
);
10852 spin_unlock_irqrestore(&rq
->lock
, flags
);
10853 wake_up_process(rq
->migration_thread
);
10855 for_each_online_cpu(cpu
) {
10856 rcu_expedited_state
= cpu
;
10857 req
= &per_cpu(rcu_migration_req
, cpu
);
10859 wait_for_completion(&req
->done
);
10860 spin_lock_irqsave(&rq
->lock
, flags
);
10861 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10862 need_full_sync
= 1;
10863 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10864 spin_unlock_irqrestore(&rq
->lock
, flags
);
10866 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10867 mutex_unlock(&rcu_sched_expedited_mutex
);
10869 if (need_full_sync
)
10870 synchronize_sched();
10872 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10874 #endif /* #else #ifndef CONFIG_SMP */