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_counter.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(struct cfs_rq
, init_tg_cfs_rq
) ____cacheline_aligned_in_smp
;
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(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
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
680 * Returns true if the current cpu runqueue is locked.
681 * This interface allows printk to be called with the runqueue lock
682 * held and know whether or not it is OK to wake up the klogd.
684 int runqueue_is_locked(void)
687 struct rq
*rq
= cpu_rq(cpu
);
690 ret
= spin_is_locked(&rq
->lock
);
696 * Debugging: various feature bits
699 #define SCHED_FEAT(name, enabled) \
700 __SCHED_FEAT_##name ,
703 #include "sched_features.h"
708 #define SCHED_FEAT(name, enabled) \
709 (1UL << __SCHED_FEAT_##name) * enabled |
711 const_debug
unsigned int sysctl_sched_features
=
712 #include "sched_features.h"
717 #ifdef CONFIG_SCHED_DEBUG
718 #define SCHED_FEAT(name, enabled) \
721 static __read_mostly
char *sched_feat_names
[] = {
722 #include "sched_features.h"
728 static int sched_feat_show(struct seq_file
*m
, void *v
)
732 for (i
= 0; sched_feat_names
[i
]; i
++) {
733 if (!(sysctl_sched_features
& (1UL << i
)))
735 seq_printf(m
, "%s ", sched_feat_names
[i
]);
743 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
744 size_t cnt
, loff_t
*ppos
)
754 if (copy_from_user(&buf
, ubuf
, cnt
))
759 if (strncmp(buf
, "NO_", 3) == 0) {
764 for (i
= 0; sched_feat_names
[i
]; i
++) {
765 int len
= strlen(sched_feat_names
[i
]);
767 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
769 sysctl_sched_features
&= ~(1UL << i
);
771 sysctl_sched_features
|= (1UL << i
);
776 if (!sched_feat_names
[i
])
784 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
786 return single_open(filp
, sched_feat_show
, NULL
);
789 static struct file_operations sched_feat_fops
= {
790 .open
= sched_feat_open
,
791 .write
= sched_feat_write
,
794 .release
= single_release
,
797 static __init
int sched_init_debug(void)
799 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
804 late_initcall(sched_init_debug
);
808 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
811 * Number of tasks to iterate in a single balance run.
812 * Limited because this is done with IRQs disabled.
814 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
817 * ratelimit for updating the group shares.
820 unsigned int sysctl_sched_shares_ratelimit
= 250000;
823 * Inject some fuzzyness into changing the per-cpu group shares
824 * this avoids remote rq-locks at the expense of fairness.
827 unsigned int sysctl_sched_shares_thresh
= 4;
830 * period over which we average the RT time consumption, measured
835 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
838 * period over which we measure -rt task cpu usage in us.
841 unsigned int sysctl_sched_rt_period
= 1000000;
843 static __read_mostly
int scheduler_running
;
846 * part of the period that we allow rt tasks to run in us.
849 int sysctl_sched_rt_runtime
= 950000;
851 static inline u64
global_rt_period(void)
853 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
856 static inline u64
global_rt_runtime(void)
858 if (sysctl_sched_rt_runtime
< 0)
861 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
864 #ifndef prepare_arch_switch
865 # define prepare_arch_switch(next) do { } while (0)
867 #ifndef finish_arch_switch
868 # define finish_arch_switch(prev) do { } while (0)
871 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
873 return rq
->curr
== p
;
876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
877 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
879 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
888 #ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq
->lock
.owner
= current
;
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
897 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
899 spin_unlock_irq(&rq
->lock
);
902 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
903 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
908 return task_current(rq
, p
);
912 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
922 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq
->lock
);
925 spin_unlock(&rq
->lock
);
929 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
940 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
944 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
950 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
954 struct rq
*rq
= task_rq(p
);
955 spin_lock(&rq
->lock
);
956 if (likely(rq
== task_rq(p
)))
958 spin_unlock(&rq
->lock
);
963 * task_rq_lock - lock the runqueue a given task resides on and disable
964 * interrupts. Note the ordering: we can safely lookup the task_rq without
965 * explicitly disabling preemption.
967 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
973 local_irq_save(*flags
);
975 spin_lock(&rq
->lock
);
976 if (likely(rq
== task_rq(p
)))
978 spin_unlock_irqrestore(&rq
->lock
, *flags
);
982 void task_rq_unlock_wait(struct task_struct
*p
)
984 struct rq
*rq
= task_rq(p
);
986 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
987 spin_unlock_wait(&rq
->lock
);
990 static void __task_rq_unlock(struct rq
*rq
)
993 spin_unlock(&rq
->lock
);
996 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
999 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1003 * this_rq_lock - lock this runqueue and disable interrupts.
1005 static struct rq
*this_rq_lock(void)
1006 __acquires(rq
->lock
)
1010 local_irq_disable();
1012 spin_lock(&rq
->lock
);
1017 #ifdef CONFIG_SCHED_HRTICK
1019 * Use HR-timers to deliver accurate preemption points.
1021 * Its all a bit involved since we cannot program an hrt while holding the
1022 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1025 * When we get rescheduled we reprogram the hrtick_timer outside of the
1031 * - enabled by features
1032 * - hrtimer is actually high res
1034 static inline int hrtick_enabled(struct rq
*rq
)
1036 if (!sched_feat(HRTICK
))
1038 if (!cpu_active(cpu_of(rq
)))
1040 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1043 static void hrtick_clear(struct rq
*rq
)
1045 if (hrtimer_active(&rq
->hrtick_timer
))
1046 hrtimer_cancel(&rq
->hrtick_timer
);
1050 * High-resolution timer tick.
1051 * Runs from hardirq context with interrupts disabled.
1053 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1055 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1057 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1059 spin_lock(&rq
->lock
);
1060 update_rq_clock(rq
);
1061 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1062 spin_unlock(&rq
->lock
);
1064 return HRTIMER_NORESTART
;
1069 * called from hardirq (IPI) context
1071 static void __hrtick_start(void *arg
)
1073 struct rq
*rq
= arg
;
1075 spin_lock(&rq
->lock
);
1076 hrtimer_restart(&rq
->hrtick_timer
);
1077 rq
->hrtick_csd_pending
= 0;
1078 spin_unlock(&rq
->lock
);
1082 * Called to set the hrtick timer state.
1084 * called with rq->lock held and irqs disabled
1086 static void hrtick_start(struct rq
*rq
, u64 delay
)
1088 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1089 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1091 hrtimer_set_expires(timer
, time
);
1093 if (rq
== this_rq()) {
1094 hrtimer_restart(timer
);
1095 } else if (!rq
->hrtick_csd_pending
) {
1096 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1097 rq
->hrtick_csd_pending
= 1;
1102 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1104 int cpu
= (int)(long)hcpu
;
1107 case CPU_UP_CANCELED
:
1108 case CPU_UP_CANCELED_FROZEN
:
1109 case CPU_DOWN_PREPARE
:
1110 case CPU_DOWN_PREPARE_FROZEN
:
1112 case CPU_DEAD_FROZEN
:
1113 hrtick_clear(cpu_rq(cpu
));
1120 static __init
void init_hrtick(void)
1122 hotcpu_notifier(hotplug_hrtick
, 0);
1126 * Called to set the hrtick timer state.
1128 * called with rq->lock held and irqs disabled
1130 static void hrtick_start(struct rq
*rq
, u64 delay
)
1132 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1133 HRTIMER_MODE_REL_PINNED
, 0);
1136 static inline void init_hrtick(void)
1139 #endif /* CONFIG_SMP */
1141 static void init_rq_hrtick(struct rq
*rq
)
1144 rq
->hrtick_csd_pending
= 0;
1146 rq
->hrtick_csd
.flags
= 0;
1147 rq
->hrtick_csd
.func
= __hrtick_start
;
1148 rq
->hrtick_csd
.info
= rq
;
1151 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1152 rq
->hrtick_timer
.function
= hrtick
;
1154 #else /* CONFIG_SCHED_HRTICK */
1155 static inline void hrtick_clear(struct rq
*rq
)
1159 static inline void init_rq_hrtick(struct rq
*rq
)
1163 static inline void init_hrtick(void)
1166 #endif /* CONFIG_SCHED_HRTICK */
1169 * resched_task - mark a task 'to be rescheduled now'.
1171 * On UP this means the setting of the need_resched flag, on SMP it
1172 * might also involve a cross-CPU call to trigger the scheduler on
1177 #ifndef tsk_is_polling
1178 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1181 static void resched_task(struct task_struct
*p
)
1185 assert_spin_locked(&task_rq(p
)->lock
);
1187 if (test_tsk_need_resched(p
))
1190 set_tsk_need_resched(p
);
1193 if (cpu
== smp_processor_id())
1196 /* NEED_RESCHED must be visible before we test polling */
1198 if (!tsk_is_polling(p
))
1199 smp_send_reschedule(cpu
);
1202 static void resched_cpu(int cpu
)
1204 struct rq
*rq
= cpu_rq(cpu
);
1205 unsigned long flags
;
1207 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1209 resched_task(cpu_curr(cpu
));
1210 spin_unlock_irqrestore(&rq
->lock
, flags
);
1215 * When add_timer_on() enqueues a timer into the timer wheel of an
1216 * idle CPU then this timer might expire before the next timer event
1217 * which is scheduled to wake up that CPU. In case of a completely
1218 * idle system the next event might even be infinite time into the
1219 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1220 * leaves the inner idle loop so the newly added timer is taken into
1221 * account when the CPU goes back to idle and evaluates the timer
1222 * wheel for the next timer event.
1224 void wake_up_idle_cpu(int cpu
)
1226 struct rq
*rq
= cpu_rq(cpu
);
1228 if (cpu
== smp_processor_id())
1232 * This is safe, as this function is called with the timer
1233 * wheel base lock of (cpu) held. When the CPU is on the way
1234 * to idle and has not yet set rq->curr to idle then it will
1235 * be serialized on the timer wheel base lock and take the new
1236 * timer into account automatically.
1238 if (rq
->curr
!= rq
->idle
)
1242 * We can set TIF_RESCHED on the idle task of the other CPU
1243 * lockless. The worst case is that the other CPU runs the
1244 * idle task through an additional NOOP schedule()
1246 set_tsk_need_resched(rq
->idle
);
1248 /* NEED_RESCHED must be visible before we test polling */
1250 if (!tsk_is_polling(rq
->idle
))
1251 smp_send_reschedule(cpu
);
1253 #endif /* CONFIG_NO_HZ */
1255 static u64
sched_avg_period(void)
1257 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1260 static void sched_avg_update(struct rq
*rq
)
1262 s64 period
= sched_avg_period();
1264 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1265 rq
->age_stamp
+= period
;
1270 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1272 rq
->rt_avg
+= rt_delta
;
1273 sched_avg_update(rq
);
1276 #else /* !CONFIG_SMP */
1277 static void resched_task(struct task_struct
*p
)
1279 assert_spin_locked(&task_rq(p
)->lock
);
1280 set_tsk_need_resched(p
);
1283 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1306 struct load_weight
*lw
)
1310 if (!lw
->inv_weight
) {
1311 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1314 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1318 tmp
= (u64
)delta_exec
* weight
;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp
> WMULT_CONST
))
1323 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1326 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1328 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1331 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1337 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight
[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult
[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1403 struct rq_iterator
{
1405 struct task_struct
*(*start
)(void *);
1406 struct task_struct
*(*next
)(void *);
1410 static unsigned long
1411 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1412 unsigned long max_load_move
, struct sched_domain
*sd
,
1413 enum cpu_idle_type idle
, int *all_pinned
,
1414 int *this_best_prio
, struct rq_iterator
*iterator
);
1417 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1418 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1419 struct rq_iterator
*iterator
);
1422 /* Time spent by the tasks of the cpu accounting group executing in ... */
1423 enum cpuacct_stat_index
{
1424 CPUACCT_STAT_USER
, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1427 CPUACCT_STAT_NSTATS
,
1430 #ifdef CONFIG_CGROUP_CPUACCT
1431 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1432 static void cpuacct_update_stats(struct task_struct
*tsk
,
1433 enum cpuacct_stat_index idx
, cputime_t val
);
1435 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1436 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1437 enum cpuacct_stat_index idx
, cputime_t val
) {}
1440 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1442 update_load_add(&rq
->load
, load
);
1445 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1447 update_load_sub(&rq
->load
, load
);
1450 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1451 typedef int (*tg_visitor
)(struct task_group
*, void *);
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1457 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1459 struct task_group
*parent
, *child
;
1463 parent
= &root_task_group
;
1465 ret
= (*down
)(parent
, data
);
1468 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1475 ret
= (*up
)(parent
, data
);
1480 parent
= parent
->parent
;
1489 static int tg_nop(struct task_group
*tg
, void *data
)
1496 /* Used instead of source_load when we know the type == 0 */
1497 static unsigned long weighted_cpuload(const int cpu
)
1499 return cpu_rq(cpu
)->load
.weight
;
1503 * Return a low guess at the load of a migration-source cpu weighted
1504 * according to the scheduling class and "nice" value.
1506 * We want to under-estimate the load of migration sources, to
1507 * balance conservatively.
1509 static unsigned long source_load(int cpu
, int type
)
1511 struct rq
*rq
= cpu_rq(cpu
);
1512 unsigned long total
= weighted_cpuload(cpu
);
1514 if (type
== 0 || !sched_feat(LB_BIAS
))
1517 return min(rq
->cpu_load
[type
-1], total
);
1521 * Return a high guess at the load of a migration-target cpu weighted
1522 * according to the scheduling class and "nice" value.
1524 static unsigned long target_load(int cpu
, int type
)
1526 struct rq
*rq
= cpu_rq(cpu
);
1527 unsigned long total
= weighted_cpuload(cpu
);
1529 if (type
== 0 || !sched_feat(LB_BIAS
))
1532 return max(rq
->cpu_load
[type
-1], total
);
1535 static struct sched_group
*group_of(int cpu
)
1537 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1545 static unsigned long power_of(int cpu
)
1547 struct sched_group
*group
= group_of(cpu
);
1550 return SCHED_LOAD_SCALE
;
1552 return group
->cpu_power
;
1555 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1557 static unsigned long cpu_avg_load_per_task(int cpu
)
1559 struct rq
*rq
= cpu_rq(cpu
);
1560 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1563 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1565 rq
->avg_load_per_task
= 0;
1567 return rq
->avg_load_per_task
;
1570 #ifdef CONFIG_FAIR_GROUP_SCHED
1572 struct update_shares_data
{
1573 unsigned long rq_weight
[NR_CPUS
];
1576 static DEFINE_PER_CPU(struct update_shares_data
, update_shares_data
);
1578 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1581 * Calculate and set the cpu's group shares.
1583 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1584 unsigned long sd_shares
,
1585 unsigned long sd_rq_weight
,
1586 struct update_shares_data
*usd
)
1588 unsigned long shares
, rq_weight
;
1591 rq_weight
= usd
->rq_weight
[cpu
];
1594 rq_weight
= NICE_0_LOAD
;
1598 * \Sum_j shares_j * rq_weight_i
1599 * shares_i = -----------------------------
1600 * \Sum_j rq_weight_j
1602 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1603 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1605 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1606 sysctl_sched_shares_thresh
) {
1607 struct rq
*rq
= cpu_rq(cpu
);
1608 unsigned long flags
;
1610 spin_lock_irqsave(&rq
->lock
, flags
);
1611 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1612 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1613 __set_se_shares(tg
->se
[cpu
], shares
);
1614 spin_unlock_irqrestore(&rq
->lock
, flags
);
1619 * Re-compute the task group their per cpu shares over the given domain.
1620 * This needs to be done in a bottom-up fashion because the rq weight of a
1621 * parent group depends on the shares of its child groups.
1623 static int tg_shares_up(struct task_group
*tg
, void *data
)
1625 unsigned long weight
, rq_weight
= 0, shares
= 0;
1626 struct update_shares_data
*usd
;
1627 struct sched_domain
*sd
= data
;
1628 unsigned long flags
;
1634 local_irq_save(flags
);
1635 usd
= &__get_cpu_var(update_shares_data
);
1637 for_each_cpu(i
, sched_domain_span(sd
)) {
1638 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1639 usd
->rq_weight
[i
] = weight
;
1642 * If there are currently no tasks on the cpu pretend there
1643 * is one of average load so that when a new task gets to
1644 * run here it will not get delayed by group starvation.
1647 weight
= NICE_0_LOAD
;
1649 rq_weight
+= weight
;
1650 shares
+= tg
->cfs_rq
[i
]->shares
;
1653 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1654 shares
= tg
->shares
;
1656 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1657 shares
= tg
->shares
;
1659 for_each_cpu(i
, sched_domain_span(sd
))
1660 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd
);
1662 local_irq_restore(flags
);
1668 * Compute the cpu's hierarchical load factor for each task group.
1669 * This needs to be done in a top-down fashion because the load of a child
1670 * group is a fraction of its parents load.
1672 static int tg_load_down(struct task_group
*tg
, void *data
)
1675 long cpu
= (long)data
;
1678 load
= cpu_rq(cpu
)->load
.weight
;
1680 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1681 load
*= tg
->cfs_rq
[cpu
]->shares
;
1682 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1685 tg
->cfs_rq
[cpu
]->h_load
= load
;
1690 static void update_shares(struct sched_domain
*sd
)
1695 if (root_task_group_empty())
1698 now
= cpu_clock(raw_smp_processor_id());
1699 elapsed
= now
- sd
->last_update
;
1701 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1702 sd
->last_update
= now
;
1703 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1707 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1709 if (root_task_group_empty())
1712 spin_unlock(&rq
->lock
);
1714 spin_lock(&rq
->lock
);
1717 static void update_h_load(long cpu
)
1719 if (root_task_group_empty())
1722 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1727 static inline void update_shares(struct sched_domain
*sd
)
1731 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1737 #ifdef CONFIG_PREEMPT
1739 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1742 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1743 * way at the expense of forcing extra atomic operations in all
1744 * invocations. This assures that the double_lock is acquired using the
1745 * same underlying policy as the spinlock_t on this architecture, which
1746 * reduces latency compared to the unfair variant below. However, it
1747 * also adds more overhead and therefore may reduce throughput.
1749 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1750 __releases(this_rq
->lock
)
1751 __acquires(busiest
->lock
)
1752 __acquires(this_rq
->lock
)
1754 spin_unlock(&this_rq
->lock
);
1755 double_rq_lock(this_rq
, busiest
);
1762 * Unfair double_lock_balance: Optimizes throughput at the expense of
1763 * latency by eliminating extra atomic operations when the locks are
1764 * already in proper order on entry. This favors lower cpu-ids and will
1765 * grant the double lock to lower cpus over higher ids under contention,
1766 * regardless of entry order into the function.
1768 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1769 __releases(this_rq
->lock
)
1770 __acquires(busiest
->lock
)
1771 __acquires(this_rq
->lock
)
1775 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1776 if (busiest
< this_rq
) {
1777 spin_unlock(&this_rq
->lock
);
1778 spin_lock(&busiest
->lock
);
1779 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1782 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1787 #endif /* CONFIG_PREEMPT */
1790 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1792 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1794 if (unlikely(!irqs_disabled())) {
1795 /* printk() doesn't work good under rq->lock */
1796 spin_unlock(&this_rq
->lock
);
1800 return _double_lock_balance(this_rq
, busiest
);
1803 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1804 __releases(busiest
->lock
)
1806 spin_unlock(&busiest
->lock
);
1807 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1811 #ifdef CONFIG_FAIR_GROUP_SCHED
1812 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1815 cfs_rq
->shares
= shares
;
1820 static void calc_load_account_active(struct rq
*this_rq
);
1822 #include "sched_stats.h"
1823 #include "sched_idletask.c"
1824 #include "sched_fair.c"
1825 #include "sched_rt.c"
1826 #ifdef CONFIG_SCHED_DEBUG
1827 # include "sched_debug.c"
1830 #define sched_class_highest (&rt_sched_class)
1831 #define for_each_class(class) \
1832 for (class = sched_class_highest; class; class = class->next)
1834 static void inc_nr_running(struct rq
*rq
)
1839 static void dec_nr_running(struct rq
*rq
)
1844 static void set_load_weight(struct task_struct
*p
)
1846 if (task_has_rt_policy(p
)) {
1847 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1848 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1853 * SCHED_IDLE tasks get minimal weight:
1855 if (p
->policy
== SCHED_IDLE
) {
1856 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1857 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1861 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1862 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1865 static void update_avg(u64
*avg
, u64 sample
)
1867 s64 diff
= sample
- *avg
;
1871 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1874 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1876 sched_info_queued(p
);
1877 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1881 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1884 if (p
->se
.last_wakeup
) {
1885 update_avg(&p
->se
.avg_overlap
,
1886 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1887 p
->se
.last_wakeup
= 0;
1889 update_avg(&p
->se
.avg_wakeup
,
1890 sysctl_sched_wakeup_granularity
);
1894 sched_info_dequeued(p
);
1895 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1900 * __normal_prio - return the priority that is based on the static prio
1902 static inline int __normal_prio(struct task_struct
*p
)
1904 return p
->static_prio
;
1908 * Calculate the expected normal priority: i.e. priority
1909 * without taking RT-inheritance into account. Might be
1910 * boosted by interactivity modifiers. Changes upon fork,
1911 * setprio syscalls, and whenever the interactivity
1912 * estimator recalculates.
1914 static inline int normal_prio(struct task_struct
*p
)
1918 if (task_has_rt_policy(p
))
1919 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1921 prio
= __normal_prio(p
);
1926 * Calculate the current priority, i.e. the priority
1927 * taken into account by the scheduler. This value might
1928 * be boosted by RT tasks, or might be boosted by
1929 * interactivity modifiers. Will be RT if the task got
1930 * RT-boosted. If not then it returns p->normal_prio.
1932 static int effective_prio(struct task_struct
*p
)
1934 p
->normal_prio
= normal_prio(p
);
1936 * If we are RT tasks or we were boosted to RT priority,
1937 * keep the priority unchanged. Otherwise, update priority
1938 * to the normal priority:
1940 if (!rt_prio(p
->prio
))
1941 return p
->normal_prio
;
1946 * activate_task - move a task to the runqueue.
1948 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1950 if (task_contributes_to_load(p
))
1951 rq
->nr_uninterruptible
--;
1953 enqueue_task(rq
, p
, wakeup
);
1958 * deactivate_task - remove a task from the runqueue.
1960 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1962 if (task_contributes_to_load(p
))
1963 rq
->nr_uninterruptible
++;
1965 dequeue_task(rq
, p
, sleep
);
1970 * task_curr - is this task currently executing on a CPU?
1971 * @p: the task in question.
1973 inline int task_curr(const struct task_struct
*p
)
1975 return cpu_curr(task_cpu(p
)) == p
;
1978 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1980 set_task_rq(p
, cpu
);
1983 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1984 * successfuly executed on another CPU. We must ensure that updates of
1985 * per-task data have been completed by this moment.
1988 task_thread_info(p
)->cpu
= cpu
;
1992 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1993 const struct sched_class
*prev_class
,
1994 int oldprio
, int running
)
1996 if (prev_class
!= p
->sched_class
) {
1997 if (prev_class
->switched_from
)
1998 prev_class
->switched_from(rq
, p
, running
);
1999 p
->sched_class
->switched_to(rq
, p
, running
);
2001 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2006 * Is this task likely cache-hot:
2009 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2014 * Buddy candidates are cache hot:
2016 if (sched_feat(CACHE_HOT_BUDDY
) &&
2017 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2018 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2021 if (p
->sched_class
!= &fair_sched_class
)
2024 if (sysctl_sched_migration_cost
== -1)
2026 if (sysctl_sched_migration_cost
== 0)
2029 delta
= now
- p
->se
.exec_start
;
2031 return delta
< (s64
)sysctl_sched_migration_cost
;
2035 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2037 int old_cpu
= task_cpu(p
);
2038 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2039 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2040 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2043 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2045 trace_sched_migrate_task(p
, new_cpu
);
2047 #ifdef CONFIG_SCHEDSTATS
2048 if (p
->se
.wait_start
)
2049 p
->se
.wait_start
-= clock_offset
;
2050 if (p
->se
.sleep_start
)
2051 p
->se
.sleep_start
-= clock_offset
;
2052 if (p
->se
.block_start
)
2053 p
->se
.block_start
-= clock_offset
;
2055 if (old_cpu
!= new_cpu
) {
2056 p
->se
.nr_migrations
++;
2057 new_rq
->nr_migrations_in
++;
2058 #ifdef CONFIG_SCHEDSTATS
2059 if (task_hot(p
, old_rq
->clock
, NULL
))
2060 schedstat_inc(p
, se
.nr_forced2_migrations
);
2062 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2065 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2066 new_cfsrq
->min_vruntime
;
2068 __set_task_cpu(p
, new_cpu
);
2071 struct migration_req
{
2072 struct list_head list
;
2074 struct task_struct
*task
;
2077 struct completion done
;
2081 * The task's runqueue lock must be held.
2082 * Returns true if you have to wait for migration thread.
2085 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2087 struct rq
*rq
= task_rq(p
);
2090 * If the task is not on a runqueue (and not running), then
2091 * it is sufficient to simply update the task's cpu field.
2093 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2094 set_task_cpu(p
, dest_cpu
);
2098 init_completion(&req
->done
);
2100 req
->dest_cpu
= dest_cpu
;
2101 list_add(&req
->list
, &rq
->migration_queue
);
2107 * wait_task_context_switch - wait for a thread to complete at least one
2110 * @p must not be current.
2112 void wait_task_context_switch(struct task_struct
*p
)
2114 unsigned long nvcsw
, nivcsw
, flags
;
2122 * The runqueue is assigned before the actual context
2123 * switch. We need to take the runqueue lock.
2125 * We could check initially without the lock but it is
2126 * very likely that we need to take the lock in every
2129 rq
= task_rq_lock(p
, &flags
);
2130 running
= task_running(rq
, p
);
2131 task_rq_unlock(rq
, &flags
);
2133 if (likely(!running
))
2136 * The switch count is incremented before the actual
2137 * context switch. We thus wait for two switches to be
2138 * sure at least one completed.
2140 if ((p
->nvcsw
- nvcsw
) > 1)
2142 if ((p
->nivcsw
- nivcsw
) > 1)
2150 * wait_task_inactive - wait for a thread to unschedule.
2152 * If @match_state is nonzero, it's the @p->state value just checked and
2153 * not expected to change. If it changes, i.e. @p might have woken up,
2154 * then return zero. When we succeed in waiting for @p to be off its CPU,
2155 * we return a positive number (its total switch count). If a second call
2156 * a short while later returns the same number, the caller can be sure that
2157 * @p has remained unscheduled the whole time.
2159 * The caller must ensure that the task *will* unschedule sometime soon,
2160 * else this function might spin for a *long* time. This function can't
2161 * be called with interrupts off, or it may introduce deadlock with
2162 * smp_call_function() if an IPI is sent by the same process we are
2163 * waiting to become inactive.
2165 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2167 unsigned long flags
;
2174 * We do the initial early heuristics without holding
2175 * any task-queue locks at all. We'll only try to get
2176 * the runqueue lock when things look like they will
2182 * If the task is actively running on another CPU
2183 * still, just relax and busy-wait without holding
2186 * NOTE! Since we don't hold any locks, it's not
2187 * even sure that "rq" stays as the right runqueue!
2188 * But we don't care, since "task_running()" will
2189 * return false if the runqueue has changed and p
2190 * is actually now running somewhere else!
2192 while (task_running(rq
, p
)) {
2193 if (match_state
&& unlikely(p
->state
!= match_state
))
2199 * Ok, time to look more closely! We need the rq
2200 * lock now, to be *sure*. If we're wrong, we'll
2201 * just go back and repeat.
2203 rq
= task_rq_lock(p
, &flags
);
2204 trace_sched_wait_task(rq
, p
);
2205 running
= task_running(rq
, p
);
2206 on_rq
= p
->se
.on_rq
;
2208 if (!match_state
|| p
->state
== match_state
)
2209 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2210 task_rq_unlock(rq
, &flags
);
2213 * If it changed from the expected state, bail out now.
2215 if (unlikely(!ncsw
))
2219 * Was it really running after all now that we
2220 * checked with the proper locks actually held?
2222 * Oops. Go back and try again..
2224 if (unlikely(running
)) {
2230 * It's not enough that it's not actively running,
2231 * it must be off the runqueue _entirely_, and not
2234 * So if it was still runnable (but just not actively
2235 * running right now), it's preempted, and we should
2236 * yield - it could be a while.
2238 if (unlikely(on_rq
)) {
2239 schedule_timeout_uninterruptible(1);
2244 * Ahh, all good. It wasn't running, and it wasn't
2245 * runnable, which means that it will never become
2246 * running in the future either. We're all done!
2255 * kick_process - kick a running thread to enter/exit the kernel
2256 * @p: the to-be-kicked thread
2258 * Cause a process which is running on another CPU to enter
2259 * kernel-mode, without any delay. (to get signals handled.)
2261 * NOTE: this function doesnt have to take the runqueue lock,
2262 * because all it wants to ensure is that the remote task enters
2263 * the kernel. If the IPI races and the task has been migrated
2264 * to another CPU then no harm is done and the purpose has been
2267 void kick_process(struct task_struct
*p
)
2273 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2274 smp_send_reschedule(cpu
);
2277 EXPORT_SYMBOL_GPL(kick_process
);
2278 #endif /* CONFIG_SMP */
2281 * task_oncpu_function_call - call a function on the cpu on which a task runs
2282 * @p: the task to evaluate
2283 * @func: the function to be called
2284 * @info: the function call argument
2286 * Calls the function @func when the task is currently running. This might
2287 * be on the current CPU, which just calls the function directly
2289 void task_oncpu_function_call(struct task_struct
*p
,
2290 void (*func
) (void *info
), void *info
)
2297 smp_call_function_single(cpu
, func
, info
, 1);
2302 * try_to_wake_up - wake up a thread
2303 * @p: the to-be-woken-up thread
2304 * @state: the mask of task states that can be woken
2305 * @sync: do a synchronous wakeup?
2307 * Put it on the run-queue if it's not already there. The "current"
2308 * thread is always on the run-queue (except when the actual
2309 * re-schedule is in progress), and as such you're allowed to do
2310 * the simpler "current->state = TASK_RUNNING" to mark yourself
2311 * runnable without the overhead of this.
2313 * returns failure only if the task is already active.
2315 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2318 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2319 unsigned long flags
;
2322 if (!sched_feat(SYNC_WAKEUPS
))
2323 wake_flags
&= ~WF_SYNC
;
2325 this_cpu
= get_cpu();
2328 rq
= task_rq_lock(p
, &flags
);
2329 update_rq_clock(rq
);
2330 if (!(p
->state
& state
))
2340 if (unlikely(task_running(rq
, p
)))
2344 * In order to handle concurrent wakeups and release the rq->lock
2345 * we put the task in TASK_WAKING state.
2347 * First fix up the nr_uninterruptible count:
2349 if (task_contributes_to_load(p
))
2350 rq
->nr_uninterruptible
--;
2351 p
->state
= TASK_WAKING
;
2352 task_rq_unlock(rq
, &flags
);
2354 cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2355 if (cpu
!= orig_cpu
)
2356 set_task_cpu(p
, cpu
);
2358 rq
= task_rq_lock(p
, &flags
);
2359 WARN_ON(p
->state
!= TASK_WAKING
);
2362 #ifdef CONFIG_SCHEDSTATS
2363 schedstat_inc(rq
, ttwu_count
);
2364 if (cpu
== this_cpu
)
2365 schedstat_inc(rq
, ttwu_local
);
2367 struct sched_domain
*sd
;
2368 for_each_domain(this_cpu
, sd
) {
2369 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2370 schedstat_inc(sd
, ttwu_wake_remote
);
2375 #endif /* CONFIG_SCHEDSTATS */
2378 #endif /* CONFIG_SMP */
2379 schedstat_inc(p
, se
.nr_wakeups
);
2380 if (wake_flags
& WF_SYNC
)
2381 schedstat_inc(p
, se
.nr_wakeups_sync
);
2382 if (orig_cpu
!= cpu
)
2383 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2384 if (cpu
== this_cpu
)
2385 schedstat_inc(p
, se
.nr_wakeups_local
);
2387 schedstat_inc(p
, se
.nr_wakeups_remote
);
2388 activate_task(rq
, p
, 1);
2392 * Only attribute actual wakeups done by this task.
2394 if (!in_interrupt()) {
2395 struct sched_entity
*se
= ¤t
->se
;
2396 u64 sample
= se
->sum_exec_runtime
;
2398 if (se
->last_wakeup
)
2399 sample
-= se
->last_wakeup
;
2401 sample
-= se
->start_runtime
;
2402 update_avg(&se
->avg_wakeup
, sample
);
2404 se
->last_wakeup
= se
->sum_exec_runtime
;
2408 trace_sched_wakeup(rq
, p
, success
);
2409 check_preempt_curr(rq
, p
, wake_flags
);
2411 p
->state
= TASK_RUNNING
;
2413 if (p
->sched_class
->task_wake_up
)
2414 p
->sched_class
->task_wake_up(rq
, p
);
2417 task_rq_unlock(rq
, &flags
);
2424 * wake_up_process - Wake up a specific process
2425 * @p: The process to be woken up.
2427 * Attempt to wake up the nominated process and move it to the set of runnable
2428 * processes. Returns 1 if the process was woken up, 0 if it was already
2431 * It may be assumed that this function implies a write memory barrier before
2432 * changing the task state if and only if any tasks are woken up.
2434 int wake_up_process(struct task_struct
*p
)
2436 return try_to_wake_up(p
, TASK_ALL
, 0);
2438 EXPORT_SYMBOL(wake_up_process
);
2440 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2442 return try_to_wake_up(p
, state
, 0);
2446 * Perform scheduler related setup for a newly forked process p.
2447 * p is forked by current.
2449 * __sched_fork() is basic setup used by init_idle() too:
2451 static void __sched_fork(struct task_struct
*p
)
2453 p
->se
.exec_start
= 0;
2454 p
->se
.sum_exec_runtime
= 0;
2455 p
->se
.prev_sum_exec_runtime
= 0;
2456 p
->se
.nr_migrations
= 0;
2457 p
->se
.last_wakeup
= 0;
2458 p
->se
.avg_overlap
= 0;
2459 p
->se
.start_runtime
= 0;
2460 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
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 * Make sure we do not leak PI boosting priority to the child.
2525 p
->prio
= current
->normal_prio
;
2528 * Revert to default priority/policy on fork if requested.
2530 if (unlikely(p
->sched_reset_on_fork
)) {
2531 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
)
2532 p
->policy
= SCHED_NORMAL
;
2534 if (p
->normal_prio
< DEFAULT_PRIO
)
2535 p
->prio
= DEFAULT_PRIO
;
2537 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2538 p
->static_prio
= NICE_TO_PRIO(0);
2543 * We don't need the reset flag anymore after the fork. It has
2544 * fulfilled its duty:
2546 p
->sched_reset_on_fork
= 0;
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 p
->prio
= effective_prio(p
);
2591 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2592 activate_task(rq
, p
, 0);
2595 * Let the scheduling class do new task startup
2596 * management (if any):
2598 p
->sched_class
->task_new(rq
, p
);
2601 trace_sched_wakeup_new(rq
, p
, 1);
2602 check_preempt_curr(rq
, p
, WF_FORK
);
2604 if (p
->sched_class
->task_wake_up
)
2605 p
->sched_class
->task_wake_up(rq
, p
);
2607 task_rq_unlock(rq
, &flags
);
2610 #ifdef CONFIG_PREEMPT_NOTIFIERS
2613 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2614 * @notifier: notifier struct to register
2616 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2618 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2620 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2623 * preempt_notifier_unregister - no longer interested in preemption notifications
2624 * @notifier: notifier struct to unregister
2626 * This is safe to call from within a preemption notifier.
2628 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2630 hlist_del(¬ifier
->link
);
2632 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2634 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2636 struct preempt_notifier
*notifier
;
2637 struct hlist_node
*node
;
2639 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2640 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2644 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2645 struct task_struct
*next
)
2647 struct preempt_notifier
*notifier
;
2648 struct hlist_node
*node
;
2650 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2651 notifier
->ops
->sched_out(notifier
, next
);
2654 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2656 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2661 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2662 struct task_struct
*next
)
2666 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2669 * prepare_task_switch - prepare to switch tasks
2670 * @rq: the runqueue preparing to switch
2671 * @prev: the current task that is being switched out
2672 * @next: the task we are going to switch to.
2674 * This is called with the rq lock held and interrupts off. It must
2675 * be paired with a subsequent finish_task_switch after the context
2678 * prepare_task_switch sets up locking and calls architecture specific
2682 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2683 struct task_struct
*next
)
2685 fire_sched_out_preempt_notifiers(prev
, next
);
2686 prepare_lock_switch(rq
, next
);
2687 prepare_arch_switch(next
);
2691 * finish_task_switch - clean up after a task-switch
2692 * @rq: runqueue associated with task-switch
2693 * @prev: the thread we just switched away from.
2695 * finish_task_switch must be called after the context switch, paired
2696 * with a prepare_task_switch call before the context switch.
2697 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2698 * and do any other architecture-specific cleanup actions.
2700 * Note that we may have delayed dropping an mm in context_switch(). If
2701 * so, we finish that here outside of the runqueue lock. (Doing it
2702 * with the lock held can cause deadlocks; see schedule() for
2705 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2706 __releases(rq
->lock
)
2708 struct mm_struct
*mm
= rq
->prev_mm
;
2714 * A task struct has one reference for the use as "current".
2715 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2716 * schedule one last time. The schedule call will never return, and
2717 * the scheduled task must drop that reference.
2718 * The test for TASK_DEAD must occur while the runqueue locks are
2719 * still held, otherwise prev could be scheduled on another cpu, die
2720 * there before we look at prev->state, and then the reference would
2722 * Manfred Spraul <manfred@colorfullife.com>
2724 prev_state
= prev
->state
;
2725 finish_arch_switch(prev
);
2726 perf_counter_task_sched_in(current
, cpu_of(rq
));
2727 finish_lock_switch(rq
, prev
);
2729 fire_sched_in_preempt_notifiers(current
);
2732 if (unlikely(prev_state
== TASK_DEAD
)) {
2734 * Remove function-return probe instances associated with this
2735 * task and put them back on the free list.
2737 kprobe_flush_task(prev
);
2738 put_task_struct(prev
);
2744 /* assumes rq->lock is held */
2745 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2747 if (prev
->sched_class
->pre_schedule
)
2748 prev
->sched_class
->pre_schedule(rq
, prev
);
2751 /* rq->lock is NOT held, but preemption is disabled */
2752 static inline void post_schedule(struct rq
*rq
)
2754 if (rq
->post_schedule
) {
2755 unsigned long flags
;
2757 spin_lock_irqsave(&rq
->lock
, flags
);
2758 if (rq
->curr
->sched_class
->post_schedule
)
2759 rq
->curr
->sched_class
->post_schedule(rq
);
2760 spin_unlock_irqrestore(&rq
->lock
, flags
);
2762 rq
->post_schedule
= 0;
2768 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2772 static inline void post_schedule(struct rq
*rq
)
2779 * schedule_tail - first thing a freshly forked thread must call.
2780 * @prev: the thread we just switched away from.
2782 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2783 __releases(rq
->lock
)
2785 struct rq
*rq
= this_rq();
2787 finish_task_switch(rq
, prev
);
2790 * FIXME: do we need to worry about rq being invalidated by the
2795 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2796 /* In this case, finish_task_switch does not reenable preemption */
2799 if (current
->set_child_tid
)
2800 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2804 * context_switch - switch to the new MM and the new
2805 * thread's register state.
2808 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2809 struct task_struct
*next
)
2811 struct mm_struct
*mm
, *oldmm
;
2813 prepare_task_switch(rq
, prev
, next
);
2814 trace_sched_switch(rq
, prev
, next
);
2816 oldmm
= prev
->active_mm
;
2818 * For paravirt, this is coupled with an exit in switch_to to
2819 * combine the page table reload and the switch backend into
2822 arch_start_context_switch(prev
);
2824 if (unlikely(!mm
)) {
2825 next
->active_mm
= oldmm
;
2826 atomic_inc(&oldmm
->mm_count
);
2827 enter_lazy_tlb(oldmm
, next
);
2829 switch_mm(oldmm
, mm
, next
);
2831 if (unlikely(!prev
->mm
)) {
2832 prev
->active_mm
= NULL
;
2833 rq
->prev_mm
= oldmm
;
2836 * Since the runqueue lock will be released by the next
2837 * task (which is an invalid locking op but in the case
2838 * of the scheduler it's an obvious special-case), so we
2839 * do an early lockdep release here:
2841 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2842 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2845 /* Here we just switch the register state and the stack. */
2846 switch_to(prev
, next
, prev
);
2850 * this_rq must be evaluated again because prev may have moved
2851 * CPUs since it called schedule(), thus the 'rq' on its stack
2852 * frame will be invalid.
2854 finish_task_switch(this_rq(), prev
);
2858 * nr_running, nr_uninterruptible and nr_context_switches:
2860 * externally visible scheduler statistics: current number of runnable
2861 * threads, current number of uninterruptible-sleeping threads, total
2862 * number of context switches performed since bootup.
2864 unsigned long nr_running(void)
2866 unsigned long i
, sum
= 0;
2868 for_each_online_cpu(i
)
2869 sum
+= cpu_rq(i
)->nr_running
;
2874 unsigned long nr_uninterruptible(void)
2876 unsigned long i
, sum
= 0;
2878 for_each_possible_cpu(i
)
2879 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2882 * Since we read the counters lockless, it might be slightly
2883 * inaccurate. Do not allow it to go below zero though:
2885 if (unlikely((long)sum
< 0))
2891 unsigned long long nr_context_switches(void)
2894 unsigned long long sum
= 0;
2896 for_each_possible_cpu(i
)
2897 sum
+= cpu_rq(i
)->nr_switches
;
2902 unsigned long nr_iowait(void)
2904 unsigned long i
, sum
= 0;
2906 for_each_possible_cpu(i
)
2907 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2912 /* Variables and functions for calc_load */
2913 static atomic_long_t calc_load_tasks
;
2914 static unsigned long calc_load_update
;
2915 unsigned long avenrun
[3];
2916 EXPORT_SYMBOL(avenrun
);
2919 * get_avenrun - get the load average array
2920 * @loads: pointer to dest load array
2921 * @offset: offset to add
2922 * @shift: shift count to shift the result left
2924 * These values are estimates at best, so no need for locking.
2926 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2928 loads
[0] = (avenrun
[0] + offset
) << shift
;
2929 loads
[1] = (avenrun
[1] + offset
) << shift
;
2930 loads
[2] = (avenrun
[2] + offset
) << shift
;
2933 static unsigned long
2934 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2937 load
+= active
* (FIXED_1
- exp
);
2938 return load
>> FSHIFT
;
2942 * calc_load - update the avenrun load estimates 10 ticks after the
2943 * CPUs have updated calc_load_tasks.
2945 void calc_global_load(void)
2947 unsigned long upd
= calc_load_update
+ 10;
2950 if (time_before(jiffies
, upd
))
2953 active
= atomic_long_read(&calc_load_tasks
);
2954 active
= active
> 0 ? active
* FIXED_1
: 0;
2956 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2957 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2958 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2960 calc_load_update
+= LOAD_FREQ
;
2964 * Either called from update_cpu_load() or from a cpu going idle
2966 static void calc_load_account_active(struct rq
*this_rq
)
2968 long nr_active
, delta
;
2970 nr_active
= this_rq
->nr_running
;
2971 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2973 if (nr_active
!= this_rq
->calc_load_active
) {
2974 delta
= nr_active
- this_rq
->calc_load_active
;
2975 this_rq
->calc_load_active
= nr_active
;
2976 atomic_long_add(delta
, &calc_load_tasks
);
2981 * Externally visible per-cpu scheduler statistics:
2982 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2984 u64
cpu_nr_migrations(int cpu
)
2986 return cpu_rq(cpu
)->nr_migrations_in
;
2990 * Update rq->cpu_load[] statistics. This function is usually called every
2991 * scheduler tick (TICK_NSEC).
2993 static void update_cpu_load(struct rq
*this_rq
)
2995 unsigned long this_load
= this_rq
->load
.weight
;
2998 this_rq
->nr_load_updates
++;
3000 /* Update our load: */
3001 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3002 unsigned long old_load
, new_load
;
3004 /* scale is effectively 1 << i now, and >> i divides by scale */
3006 old_load
= this_rq
->cpu_load
[i
];
3007 new_load
= this_load
;
3009 * Round up the averaging division if load is increasing. This
3010 * prevents us from getting stuck on 9 if the load is 10, for
3013 if (new_load
> old_load
)
3014 new_load
+= scale
-1;
3015 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3018 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3019 this_rq
->calc_load_update
+= LOAD_FREQ
;
3020 calc_load_account_active(this_rq
);
3027 * double_rq_lock - safely lock two runqueues
3029 * Note this does not disable interrupts like task_rq_lock,
3030 * you need to do so manually before calling.
3032 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3033 __acquires(rq1
->lock
)
3034 __acquires(rq2
->lock
)
3036 BUG_ON(!irqs_disabled());
3038 spin_lock(&rq1
->lock
);
3039 __acquire(rq2
->lock
); /* Fake it out ;) */
3042 spin_lock(&rq1
->lock
);
3043 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3045 spin_lock(&rq2
->lock
);
3046 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3049 update_rq_clock(rq1
);
3050 update_rq_clock(rq2
);
3054 * double_rq_unlock - safely unlock two runqueues
3056 * Note this does not restore interrupts like task_rq_unlock,
3057 * you need to do so manually after calling.
3059 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3060 __releases(rq1
->lock
)
3061 __releases(rq2
->lock
)
3063 spin_unlock(&rq1
->lock
);
3065 spin_unlock(&rq2
->lock
);
3067 __release(rq2
->lock
);
3071 * If dest_cpu is allowed for this process, migrate the task to it.
3072 * This is accomplished by forcing the cpu_allowed mask to only
3073 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3074 * the cpu_allowed mask is restored.
3076 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3078 struct migration_req req
;
3079 unsigned long flags
;
3082 rq
= task_rq_lock(p
, &flags
);
3083 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3084 || unlikely(!cpu_active(dest_cpu
)))
3087 /* force the process onto the specified CPU */
3088 if (migrate_task(p
, dest_cpu
, &req
)) {
3089 /* Need to wait for migration thread (might exit: take ref). */
3090 struct task_struct
*mt
= rq
->migration_thread
;
3092 get_task_struct(mt
);
3093 task_rq_unlock(rq
, &flags
);
3094 wake_up_process(mt
);
3095 put_task_struct(mt
);
3096 wait_for_completion(&req
.done
);
3101 task_rq_unlock(rq
, &flags
);
3105 * sched_exec - execve() is a valuable balancing opportunity, because at
3106 * this point the task has the smallest effective memory and cache footprint.
3108 void sched_exec(void)
3110 int new_cpu
, this_cpu
= get_cpu();
3111 new_cpu
= current
->sched_class
->select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3113 if (new_cpu
!= this_cpu
)
3114 sched_migrate_task(current
, new_cpu
);
3118 * pull_task - move a task from a remote runqueue to the local runqueue.
3119 * Both runqueues must be locked.
3121 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3122 struct rq
*this_rq
, int this_cpu
)
3124 deactivate_task(src_rq
, p
, 0);
3125 set_task_cpu(p
, this_cpu
);
3126 activate_task(this_rq
, p
, 0);
3128 * Note that idle threads have a prio of MAX_PRIO, for this test
3129 * to be always true for them.
3131 check_preempt_curr(this_rq
, p
, 0);
3135 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3138 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3139 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3142 int tsk_cache_hot
= 0;
3144 * We do not migrate tasks that are:
3145 * 1) running (obviously), or
3146 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3147 * 3) are cache-hot on their current CPU.
3149 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3150 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3155 if (task_running(rq
, p
)) {
3156 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3161 * Aggressive migration if:
3162 * 1) task is cache cold, or
3163 * 2) too many balance attempts have failed.
3166 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3167 if (!tsk_cache_hot
||
3168 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3169 #ifdef CONFIG_SCHEDSTATS
3170 if (tsk_cache_hot
) {
3171 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3172 schedstat_inc(p
, se
.nr_forced_migrations
);
3178 if (tsk_cache_hot
) {
3179 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3185 static unsigned long
3186 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3187 unsigned long max_load_move
, struct sched_domain
*sd
,
3188 enum cpu_idle_type idle
, int *all_pinned
,
3189 int *this_best_prio
, struct rq_iterator
*iterator
)
3191 int loops
= 0, pulled
= 0, pinned
= 0;
3192 struct task_struct
*p
;
3193 long rem_load_move
= max_load_move
;
3195 if (max_load_move
== 0)
3201 * Start the load-balancing iterator:
3203 p
= iterator
->start(iterator
->arg
);
3205 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3208 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3209 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3210 p
= iterator
->next(iterator
->arg
);
3214 pull_task(busiest
, p
, this_rq
, this_cpu
);
3216 rem_load_move
-= p
->se
.load
.weight
;
3218 #ifdef CONFIG_PREEMPT
3220 * NEWIDLE balancing is a source of latency, so preemptible kernels
3221 * will stop after the first task is pulled to minimize the critical
3224 if (idle
== CPU_NEWLY_IDLE
)
3229 * We only want to steal up to the prescribed amount of weighted load.
3231 if (rem_load_move
> 0) {
3232 if (p
->prio
< *this_best_prio
)
3233 *this_best_prio
= p
->prio
;
3234 p
= iterator
->next(iterator
->arg
);
3239 * Right now, this is one of only two places pull_task() is called,
3240 * so we can safely collect pull_task() stats here rather than
3241 * inside pull_task().
3243 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3246 *all_pinned
= pinned
;
3248 return max_load_move
- rem_load_move
;
3252 * move_tasks tries to move up to max_load_move weighted load from busiest to
3253 * this_rq, as part of a balancing operation within domain "sd".
3254 * Returns 1 if successful and 0 otherwise.
3256 * Called with both runqueues locked.
3258 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3259 unsigned long max_load_move
,
3260 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3263 const struct sched_class
*class = sched_class_highest
;
3264 unsigned long total_load_moved
= 0;
3265 int this_best_prio
= this_rq
->curr
->prio
;
3269 class->load_balance(this_rq
, this_cpu
, busiest
,
3270 max_load_move
- total_load_moved
,
3271 sd
, idle
, all_pinned
, &this_best_prio
);
3272 class = class->next
;
3274 #ifdef CONFIG_PREEMPT
3276 * NEWIDLE balancing is a source of latency, so preemptible
3277 * kernels will stop after the first task is pulled to minimize
3278 * the critical section.
3280 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3283 } while (class && max_load_move
> total_load_moved
);
3285 return total_load_moved
> 0;
3289 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3290 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3291 struct rq_iterator
*iterator
)
3293 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3297 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3298 pull_task(busiest
, p
, this_rq
, this_cpu
);
3300 * Right now, this is only the second place pull_task()
3301 * is called, so we can safely collect pull_task()
3302 * stats here rather than inside pull_task().
3304 schedstat_inc(sd
, lb_gained
[idle
]);
3308 p
= iterator
->next(iterator
->arg
);
3315 * move_one_task tries to move exactly one task from busiest to this_rq, as
3316 * part of active balancing operations within "domain".
3317 * Returns 1 if successful and 0 otherwise.
3319 * Called with both runqueues locked.
3321 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3322 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3324 const struct sched_class
*class;
3326 for_each_class(class) {
3327 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3333 /********** Helpers for find_busiest_group ************************/
3335 * sd_lb_stats - Structure to store the statistics of a sched_domain
3336 * during load balancing.
3338 struct sd_lb_stats
{
3339 struct sched_group
*busiest
; /* Busiest group in this sd */
3340 struct sched_group
*this; /* Local group in this sd */
3341 unsigned long total_load
; /* Total load of all groups in sd */
3342 unsigned long total_pwr
; /* Total power of all groups in sd */
3343 unsigned long avg_load
; /* Average load across all groups in sd */
3345 /** Statistics of this group */
3346 unsigned long this_load
;
3347 unsigned long this_load_per_task
;
3348 unsigned long this_nr_running
;
3350 /* Statistics of the busiest group */
3351 unsigned long max_load
;
3352 unsigned long busiest_load_per_task
;
3353 unsigned long busiest_nr_running
;
3355 int group_imb
; /* Is there imbalance in this sd */
3356 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3357 int power_savings_balance
; /* Is powersave balance needed for this sd */
3358 struct sched_group
*group_min
; /* Least loaded group in sd */
3359 struct sched_group
*group_leader
; /* Group which relieves group_min */
3360 unsigned long min_load_per_task
; /* load_per_task in group_min */
3361 unsigned long leader_nr_running
; /* Nr running of group_leader */
3362 unsigned long min_nr_running
; /* Nr running of group_min */
3367 * sg_lb_stats - stats of a sched_group required for load_balancing
3369 struct sg_lb_stats
{
3370 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3371 unsigned long group_load
; /* Total load over the CPUs of the group */
3372 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3373 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3374 unsigned long group_capacity
;
3375 int group_imb
; /* Is there an imbalance in the group ? */
3379 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3380 * @group: The group whose first cpu is to be returned.
3382 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3384 return cpumask_first(sched_group_cpus(group
));
3388 * get_sd_load_idx - Obtain the load index for a given sched domain.
3389 * @sd: The sched_domain whose load_idx is to be obtained.
3390 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3392 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3393 enum cpu_idle_type idle
)
3399 load_idx
= sd
->busy_idx
;
3402 case CPU_NEWLY_IDLE
:
3403 load_idx
= sd
->newidle_idx
;
3406 load_idx
= sd
->idle_idx
;
3414 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3416 * init_sd_power_savings_stats - Initialize power savings statistics for
3417 * the given sched_domain, during load balancing.
3419 * @sd: Sched domain whose power-savings statistics are to be initialized.
3420 * @sds: Variable containing the statistics for sd.
3421 * @idle: Idle status of the CPU at which we're performing load-balancing.
3423 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3424 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3427 * Busy processors will not participate in power savings
3430 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3431 sds
->power_savings_balance
= 0;
3433 sds
->power_savings_balance
= 1;
3434 sds
->min_nr_running
= ULONG_MAX
;
3435 sds
->leader_nr_running
= 0;
3440 * update_sd_power_savings_stats - Update the power saving stats for a
3441 * sched_domain while performing load balancing.
3443 * @group: sched_group belonging to the sched_domain under consideration.
3444 * @sds: Variable containing the statistics of the sched_domain
3445 * @local_group: Does group contain the CPU for which we're performing
3447 * @sgs: Variable containing the statistics of the group.
3449 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3450 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3453 if (!sds
->power_savings_balance
)
3457 * If the local group is idle or completely loaded
3458 * no need to do power savings balance at this domain
3460 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3461 !sds
->this_nr_running
))
3462 sds
->power_savings_balance
= 0;
3465 * If a group is already running at full capacity or idle,
3466 * don't include that group in power savings calculations
3468 if (!sds
->power_savings_balance
||
3469 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3470 !sgs
->sum_nr_running
)
3474 * Calculate the group which has the least non-idle load.
3475 * This is the group from where we need to pick up the load
3478 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3479 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3480 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3481 sds
->group_min
= group
;
3482 sds
->min_nr_running
= sgs
->sum_nr_running
;
3483 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3484 sgs
->sum_nr_running
;
3488 * Calculate the group which is almost near its
3489 * capacity but still has some space to pick up some load
3490 * from other group and save more power
3492 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3495 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3496 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3497 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3498 sds
->group_leader
= group
;
3499 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3504 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3505 * @sds: Variable containing the statistics of the sched_domain
3506 * under consideration.
3507 * @this_cpu: Cpu at which we're currently performing load-balancing.
3508 * @imbalance: Variable to store the imbalance.
3511 * Check if we have potential to perform some power-savings balance.
3512 * If yes, set the busiest group to be the least loaded group in the
3513 * sched_domain, so that it's CPUs can be put to idle.
3515 * Returns 1 if there is potential to perform power-savings balance.
3518 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3519 int this_cpu
, unsigned long *imbalance
)
3521 if (!sds
->power_savings_balance
)
3524 if (sds
->this != sds
->group_leader
||
3525 sds
->group_leader
== sds
->group_min
)
3528 *imbalance
= sds
->min_load_per_task
;
3529 sds
->busiest
= sds
->group_min
;
3534 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3535 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3536 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3541 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3542 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3547 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3548 int this_cpu
, unsigned long *imbalance
)
3552 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3555 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3557 return SCHED_LOAD_SCALE
;
3560 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3562 return default_scale_freq_power(sd
, cpu
);
3565 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3567 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3568 unsigned long smt_gain
= sd
->smt_gain
;
3575 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3577 return default_scale_smt_power(sd
, cpu
);
3580 unsigned long scale_rt_power(int cpu
)
3582 struct rq
*rq
= cpu_rq(cpu
);
3583 u64 total
, available
;
3585 sched_avg_update(rq
);
3587 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3588 available
= total
- rq
->rt_avg
;
3590 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3591 total
= SCHED_LOAD_SCALE
;
3593 total
>>= SCHED_LOAD_SHIFT
;
3595 return div_u64(available
, total
);
3598 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3600 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3601 unsigned long power
= SCHED_LOAD_SCALE
;
3602 struct sched_group
*sdg
= sd
->groups
;
3604 if (sched_feat(ARCH_POWER
))
3605 power
*= arch_scale_freq_power(sd
, cpu
);
3607 power
*= default_scale_freq_power(sd
, cpu
);
3609 power
>>= SCHED_LOAD_SHIFT
;
3611 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3612 if (sched_feat(ARCH_POWER
))
3613 power
*= arch_scale_smt_power(sd
, cpu
);
3615 power
*= default_scale_smt_power(sd
, cpu
);
3617 power
>>= SCHED_LOAD_SHIFT
;
3620 power
*= scale_rt_power(cpu
);
3621 power
>>= SCHED_LOAD_SHIFT
;
3626 sdg
->cpu_power
= power
;
3629 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3631 struct sched_domain
*child
= sd
->child
;
3632 struct sched_group
*group
, *sdg
= sd
->groups
;
3633 unsigned long power
;
3636 update_cpu_power(sd
, cpu
);
3642 group
= child
->groups
;
3644 power
+= group
->cpu_power
;
3645 group
= group
->next
;
3646 } while (group
!= child
->groups
);
3648 sdg
->cpu_power
= power
;
3652 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3653 * @group: sched_group whose statistics are to be updated.
3654 * @this_cpu: Cpu for which load balance is currently performed.
3655 * @idle: Idle status of this_cpu
3656 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3657 * @sd_idle: Idle status of the sched_domain containing group.
3658 * @local_group: Does group contain this_cpu.
3659 * @cpus: Set of cpus considered for load balancing.
3660 * @balance: Should we balance.
3661 * @sgs: variable to hold the statistics for this group.
3663 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3664 struct sched_group
*group
, int this_cpu
,
3665 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3666 int local_group
, const struct cpumask
*cpus
,
3667 int *balance
, struct sg_lb_stats
*sgs
)
3669 unsigned long load
, max_cpu_load
, min_cpu_load
;
3671 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3672 unsigned long sum_avg_load_per_task
;
3673 unsigned long avg_load_per_task
;
3676 balance_cpu
= group_first_cpu(group
);
3677 if (balance_cpu
== this_cpu
)
3678 update_group_power(sd
, this_cpu
);
3681 /* Tally up the load of all CPUs in the group */
3682 sum_avg_load_per_task
= avg_load_per_task
= 0;
3684 min_cpu_load
= ~0UL;
3686 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3687 struct rq
*rq
= cpu_rq(i
);
3689 if (*sd_idle
&& rq
->nr_running
)
3692 /* Bias balancing toward cpus of our domain */
3694 if (idle_cpu(i
) && !first_idle_cpu
) {
3699 load
= target_load(i
, load_idx
);
3701 load
= source_load(i
, load_idx
);
3702 if (load
> max_cpu_load
)
3703 max_cpu_load
= load
;
3704 if (min_cpu_load
> load
)
3705 min_cpu_load
= load
;
3708 sgs
->group_load
+= load
;
3709 sgs
->sum_nr_running
+= rq
->nr_running
;
3710 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3712 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3716 * First idle cpu or the first cpu(busiest) in this sched group
3717 * is eligible for doing load balancing at this and above
3718 * domains. In the newly idle case, we will allow all the cpu's
3719 * to do the newly idle load balance.
3721 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3722 balance_cpu
!= this_cpu
&& balance
) {
3727 /* Adjust by relative CPU power of the group */
3728 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3732 * Consider the group unbalanced when the imbalance is larger
3733 * than the average weight of two tasks.
3735 * APZ: with cgroup the avg task weight can vary wildly and
3736 * might not be a suitable number - should we keep a
3737 * normalized nr_running number somewhere that negates
3740 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3743 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3746 sgs
->group_capacity
=
3747 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3751 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3752 * @sd: sched_domain whose statistics are to be updated.
3753 * @this_cpu: Cpu for which load balance is currently performed.
3754 * @idle: Idle status of this_cpu
3755 * @sd_idle: Idle status of the sched_domain containing group.
3756 * @cpus: Set of cpus considered for load balancing.
3757 * @balance: Should we balance.
3758 * @sds: variable to hold the statistics for this sched_domain.
3760 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3761 enum cpu_idle_type idle
, int *sd_idle
,
3762 const struct cpumask
*cpus
, int *balance
,
3763 struct sd_lb_stats
*sds
)
3765 struct sched_domain
*child
= sd
->child
;
3766 struct sched_group
*group
= sd
->groups
;
3767 struct sg_lb_stats sgs
;
3768 int load_idx
, prefer_sibling
= 0;
3770 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3773 init_sd_power_savings_stats(sd
, sds
, idle
);
3774 load_idx
= get_sd_load_idx(sd
, idle
);
3779 local_group
= cpumask_test_cpu(this_cpu
,
3780 sched_group_cpus(group
));
3781 memset(&sgs
, 0, sizeof(sgs
));
3782 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3783 local_group
, cpus
, balance
, &sgs
);
3785 if (local_group
&& balance
&& !(*balance
))
3788 sds
->total_load
+= sgs
.group_load
;
3789 sds
->total_pwr
+= group
->cpu_power
;
3792 * In case the child domain prefers tasks go to siblings
3793 * first, lower the group capacity to one so that we'll try
3794 * and move all the excess tasks away.
3797 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3800 sds
->this_load
= sgs
.avg_load
;
3802 sds
->this_nr_running
= sgs
.sum_nr_running
;
3803 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3804 } else if (sgs
.avg_load
> sds
->max_load
&&
3805 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3807 sds
->max_load
= sgs
.avg_load
;
3808 sds
->busiest
= group
;
3809 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3810 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3811 sds
->group_imb
= sgs
.group_imb
;
3814 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3815 group
= group
->next
;
3816 } while (group
!= sd
->groups
);
3820 * fix_small_imbalance - Calculate the minor imbalance that exists
3821 * amongst the groups of a sched_domain, during
3823 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3824 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3825 * @imbalance: Variable to store the imbalance.
3827 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3828 int this_cpu
, unsigned long *imbalance
)
3830 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3831 unsigned int imbn
= 2;
3833 if (sds
->this_nr_running
) {
3834 sds
->this_load_per_task
/= sds
->this_nr_running
;
3835 if (sds
->busiest_load_per_task
>
3836 sds
->this_load_per_task
)
3839 sds
->this_load_per_task
=
3840 cpu_avg_load_per_task(this_cpu
);
3842 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3843 sds
->busiest_load_per_task
* imbn
) {
3844 *imbalance
= sds
->busiest_load_per_task
;
3849 * OK, we don't have enough imbalance to justify moving tasks,
3850 * however we may be able to increase total CPU power used by
3854 pwr_now
+= sds
->busiest
->cpu_power
*
3855 min(sds
->busiest_load_per_task
, sds
->max_load
);
3856 pwr_now
+= sds
->this->cpu_power
*
3857 min(sds
->this_load_per_task
, sds
->this_load
);
3858 pwr_now
/= SCHED_LOAD_SCALE
;
3860 /* Amount of load we'd subtract */
3861 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3862 sds
->busiest
->cpu_power
;
3863 if (sds
->max_load
> tmp
)
3864 pwr_move
+= sds
->busiest
->cpu_power
*
3865 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3867 /* Amount of load we'd add */
3868 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3869 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3870 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3871 sds
->this->cpu_power
;
3873 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3874 sds
->this->cpu_power
;
3875 pwr_move
+= sds
->this->cpu_power
*
3876 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3877 pwr_move
/= SCHED_LOAD_SCALE
;
3879 /* Move if we gain throughput */
3880 if (pwr_move
> pwr_now
)
3881 *imbalance
= sds
->busiest_load_per_task
;
3885 * calculate_imbalance - Calculate the amount of imbalance present within the
3886 * groups of a given sched_domain during load balance.
3887 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3888 * @this_cpu: Cpu for which currently load balance is being performed.
3889 * @imbalance: The variable to store the imbalance.
3891 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3892 unsigned long *imbalance
)
3894 unsigned long max_pull
;
3896 * In the presence of smp nice balancing, certain scenarios can have
3897 * max load less than avg load(as we skip the groups at or below
3898 * its cpu_power, while calculating max_load..)
3900 if (sds
->max_load
< sds
->avg_load
) {
3902 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3905 /* Don't want to pull so many tasks that a group would go idle */
3906 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3907 sds
->max_load
- sds
->busiest_load_per_task
);
3909 /* How much load to actually move to equalise the imbalance */
3910 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3911 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3915 * if *imbalance is less than the average load per runnable task
3916 * there is no gaurantee that any tasks will be moved so we'll have
3917 * a think about bumping its value to force at least one task to be
3920 if (*imbalance
< sds
->busiest_load_per_task
)
3921 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3924 /******* find_busiest_group() helpers end here *********************/
3927 * find_busiest_group - Returns the busiest group within the sched_domain
3928 * if there is an imbalance. If there isn't an imbalance, and
3929 * the user has opted for power-savings, it returns a group whose
3930 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3931 * such a group exists.
3933 * Also calculates the amount of weighted load which should be moved
3934 * to restore balance.
3936 * @sd: The sched_domain whose busiest group is to be returned.
3937 * @this_cpu: The cpu for which load balancing is currently being performed.
3938 * @imbalance: Variable which stores amount of weighted load which should
3939 * be moved to restore balance/put a group to idle.
3940 * @idle: The idle status of this_cpu.
3941 * @sd_idle: The idleness of sd
3942 * @cpus: The set of CPUs under consideration for load-balancing.
3943 * @balance: Pointer to a variable indicating if this_cpu
3944 * is the appropriate cpu to perform load balancing at this_level.
3946 * Returns: - the busiest group if imbalance exists.
3947 * - If no imbalance and user has opted for power-savings balance,
3948 * return the least loaded group whose CPUs can be
3949 * put to idle by rebalancing its tasks onto our group.
3951 static struct sched_group
*
3952 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3953 unsigned long *imbalance
, enum cpu_idle_type idle
,
3954 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3956 struct sd_lb_stats sds
;
3958 memset(&sds
, 0, sizeof(sds
));
3961 * Compute the various statistics relavent for load balancing at
3964 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3967 /* Cases where imbalance does not exist from POV of this_cpu */
3968 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3970 * 2) There is no busy sibling group to pull from.
3971 * 3) This group is the busiest group.
3972 * 4) This group is more busy than the avg busieness at this
3974 * 5) The imbalance is within the specified limit.
3975 * 6) Any rebalance would lead to ping-pong
3977 if (balance
&& !(*balance
))
3980 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3983 if (sds
.this_load
>= sds
.max_load
)
3986 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3988 if (sds
.this_load
>= sds
.avg_load
)
3991 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3994 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3996 sds
.busiest_load_per_task
=
3997 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4000 * We're trying to get all the cpus to the average_load, so we don't
4001 * want to push ourselves above the average load, nor do we wish to
4002 * reduce the max loaded cpu below the average load, as either of these
4003 * actions would just result in more rebalancing later, and ping-pong
4004 * tasks around. Thus we look for the minimum possible imbalance.
4005 * Negative imbalances (*we* are more loaded than anyone else) will
4006 * be counted as no imbalance for these purposes -- we can't fix that
4007 * by pulling tasks to us. Be careful of negative numbers as they'll
4008 * appear as very large values with unsigned longs.
4010 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4013 /* Looks like there is an imbalance. Compute it */
4014 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4019 * There is no obvious imbalance. But check if we can do some balancing
4022 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4030 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4033 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4034 unsigned long imbalance
, const struct cpumask
*cpus
)
4036 struct rq
*busiest
= NULL
, *rq
;
4037 unsigned long max_load
= 0;
4040 for_each_cpu(i
, sched_group_cpus(group
)) {
4041 unsigned long power
= power_of(i
);
4042 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4045 if (!cpumask_test_cpu(i
, cpus
))
4049 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4052 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4055 if (wl
> max_load
) {
4065 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4066 * so long as it is large enough.
4068 #define MAX_PINNED_INTERVAL 512
4070 /* Working cpumask for load_balance and load_balance_newidle. */
4071 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4074 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4075 * tasks if there is an imbalance.
4077 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4078 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4081 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4082 struct sched_group
*group
;
4083 unsigned long imbalance
;
4085 unsigned long flags
;
4086 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4088 cpumask_setall(cpus
);
4091 * When power savings policy is enabled for the parent domain, idle
4092 * sibling can pick up load irrespective of busy siblings. In this case,
4093 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4094 * portraying it as CPU_NOT_IDLE.
4096 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4097 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4100 schedstat_inc(sd
, lb_count
[idle
]);
4104 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4111 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4115 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4117 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4121 BUG_ON(busiest
== this_rq
);
4123 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4126 if (busiest
->nr_running
> 1) {
4128 * Attempt to move tasks. If find_busiest_group has found
4129 * an imbalance but busiest->nr_running <= 1, the group is
4130 * still unbalanced. ld_moved simply stays zero, so it is
4131 * correctly treated as an imbalance.
4133 local_irq_save(flags
);
4134 double_rq_lock(this_rq
, busiest
);
4135 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4136 imbalance
, sd
, idle
, &all_pinned
);
4137 double_rq_unlock(this_rq
, busiest
);
4138 local_irq_restore(flags
);
4141 * some other cpu did the load balance for us.
4143 if (ld_moved
&& this_cpu
!= smp_processor_id())
4144 resched_cpu(this_cpu
);
4146 /* All tasks on this runqueue were pinned by CPU affinity */
4147 if (unlikely(all_pinned
)) {
4148 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4149 if (!cpumask_empty(cpus
))
4156 schedstat_inc(sd
, lb_failed
[idle
]);
4157 sd
->nr_balance_failed
++;
4159 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4161 spin_lock_irqsave(&busiest
->lock
, flags
);
4163 /* don't kick the migration_thread, if the curr
4164 * task on busiest cpu can't be moved to this_cpu
4166 if (!cpumask_test_cpu(this_cpu
,
4167 &busiest
->curr
->cpus_allowed
)) {
4168 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4170 goto out_one_pinned
;
4173 if (!busiest
->active_balance
) {
4174 busiest
->active_balance
= 1;
4175 busiest
->push_cpu
= this_cpu
;
4178 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4180 wake_up_process(busiest
->migration_thread
);
4183 * We've kicked active balancing, reset the failure
4186 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4189 sd
->nr_balance_failed
= 0;
4191 if (likely(!active_balance
)) {
4192 /* We were unbalanced, so reset the balancing interval */
4193 sd
->balance_interval
= sd
->min_interval
;
4196 * If we've begun active balancing, start to back off. This
4197 * case may not be covered by the all_pinned logic if there
4198 * is only 1 task on the busy runqueue (because we don't call
4201 if (sd
->balance_interval
< sd
->max_interval
)
4202 sd
->balance_interval
*= 2;
4205 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4206 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4212 schedstat_inc(sd
, lb_balanced
[idle
]);
4214 sd
->nr_balance_failed
= 0;
4217 /* tune up the balancing interval */
4218 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4219 (sd
->balance_interval
< sd
->max_interval
))
4220 sd
->balance_interval
*= 2;
4222 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4223 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4234 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4235 * tasks if there is an imbalance.
4237 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4238 * this_rq is locked.
4241 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4243 struct sched_group
*group
;
4244 struct rq
*busiest
= NULL
;
4245 unsigned long imbalance
;
4249 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4251 cpumask_setall(cpus
);
4254 * When power savings policy is enabled for the parent domain, idle
4255 * sibling can pick up load irrespective of busy siblings. In this case,
4256 * let the state of idle sibling percolate up as IDLE, instead of
4257 * portraying it as CPU_NOT_IDLE.
4259 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4260 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4263 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4265 update_shares_locked(this_rq
, sd
);
4266 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4267 &sd_idle
, cpus
, NULL
);
4269 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4273 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4275 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4279 BUG_ON(busiest
== this_rq
);
4281 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4284 if (busiest
->nr_running
> 1) {
4285 /* Attempt to move tasks */
4286 double_lock_balance(this_rq
, busiest
);
4287 /* this_rq->clock is already updated */
4288 update_rq_clock(busiest
);
4289 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4290 imbalance
, sd
, CPU_NEWLY_IDLE
,
4292 double_unlock_balance(this_rq
, busiest
);
4294 if (unlikely(all_pinned
)) {
4295 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4296 if (!cpumask_empty(cpus
))
4302 int active_balance
= 0;
4304 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4305 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4306 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4309 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4312 if (sd
->nr_balance_failed
++ < 2)
4316 * The only task running in a non-idle cpu can be moved to this
4317 * cpu in an attempt to completely freeup the other CPU
4318 * package. The same method used to move task in load_balance()
4319 * have been extended for load_balance_newidle() to speedup
4320 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4322 * The package power saving logic comes from
4323 * find_busiest_group(). If there are no imbalance, then
4324 * f_b_g() will return NULL. However when sched_mc={1,2} then
4325 * f_b_g() will select a group from which a running task may be
4326 * pulled to this cpu in order to make the other package idle.
4327 * If there is no opportunity to make a package idle and if
4328 * there are no imbalance, then f_b_g() will return NULL and no
4329 * action will be taken in load_balance_newidle().
4331 * Under normal task pull operation due to imbalance, there
4332 * will be more than one task in the source run queue and
4333 * move_tasks() will succeed. ld_moved will be true and this
4334 * active balance code will not be triggered.
4337 /* Lock busiest in correct order while this_rq is held */
4338 double_lock_balance(this_rq
, busiest
);
4341 * don't kick the migration_thread, if the curr
4342 * task on busiest cpu can't be moved to this_cpu
4344 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4345 double_unlock_balance(this_rq
, busiest
);
4350 if (!busiest
->active_balance
) {
4351 busiest
->active_balance
= 1;
4352 busiest
->push_cpu
= this_cpu
;
4356 double_unlock_balance(this_rq
, busiest
);
4358 * Should not call ttwu while holding a rq->lock
4360 spin_unlock(&this_rq
->lock
);
4362 wake_up_process(busiest
->migration_thread
);
4363 spin_lock(&this_rq
->lock
);
4366 sd
->nr_balance_failed
= 0;
4368 update_shares_locked(this_rq
, sd
);
4372 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4373 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4374 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4376 sd
->nr_balance_failed
= 0;
4382 * idle_balance is called by schedule() if this_cpu is about to become
4383 * idle. Attempts to pull tasks from other CPUs.
4385 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4387 struct sched_domain
*sd
;
4388 int pulled_task
= 0;
4389 unsigned long next_balance
= jiffies
+ HZ
;
4391 for_each_domain(this_cpu
, sd
) {
4392 unsigned long interval
;
4394 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4397 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4398 /* If we've pulled tasks over stop searching: */
4399 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4402 interval
= msecs_to_jiffies(sd
->balance_interval
);
4403 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4404 next_balance
= sd
->last_balance
+ interval
;
4408 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4410 * We are going idle. next_balance may be set based on
4411 * a busy processor. So reset next_balance.
4413 this_rq
->next_balance
= next_balance
;
4418 * active_load_balance is run by migration threads. It pushes running tasks
4419 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4420 * running on each physical CPU where possible, and avoids physical /
4421 * logical imbalances.
4423 * Called with busiest_rq locked.
4425 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4427 int target_cpu
= busiest_rq
->push_cpu
;
4428 struct sched_domain
*sd
;
4429 struct rq
*target_rq
;
4431 /* Is there any task to move? */
4432 if (busiest_rq
->nr_running
<= 1)
4435 target_rq
= cpu_rq(target_cpu
);
4438 * This condition is "impossible", if it occurs
4439 * we need to fix it. Originally reported by
4440 * Bjorn Helgaas on a 128-cpu setup.
4442 BUG_ON(busiest_rq
== target_rq
);
4444 /* move a task from busiest_rq to target_rq */
4445 double_lock_balance(busiest_rq
, target_rq
);
4446 update_rq_clock(busiest_rq
);
4447 update_rq_clock(target_rq
);
4449 /* Search for an sd spanning us and the target CPU. */
4450 for_each_domain(target_cpu
, sd
) {
4451 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4452 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4457 schedstat_inc(sd
, alb_count
);
4459 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4461 schedstat_inc(sd
, alb_pushed
);
4463 schedstat_inc(sd
, alb_failed
);
4465 double_unlock_balance(busiest_rq
, target_rq
);
4470 atomic_t load_balancer
;
4471 cpumask_var_t cpu_mask
;
4472 cpumask_var_t ilb_grp_nohz_mask
;
4473 } nohz ____cacheline_aligned
= {
4474 .load_balancer
= ATOMIC_INIT(-1),
4477 int get_nohz_load_balancer(void)
4479 return atomic_read(&nohz
.load_balancer
);
4482 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4484 * lowest_flag_domain - Return lowest sched_domain containing flag.
4485 * @cpu: The cpu whose lowest level of sched domain is to
4487 * @flag: The flag to check for the lowest sched_domain
4488 * for the given cpu.
4490 * Returns the lowest sched_domain of a cpu which contains the given flag.
4492 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4494 struct sched_domain
*sd
;
4496 for_each_domain(cpu
, sd
)
4497 if (sd
&& (sd
->flags
& flag
))
4504 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4505 * @cpu: The cpu whose domains we're iterating over.
4506 * @sd: variable holding the value of the power_savings_sd
4508 * @flag: The flag to filter the sched_domains to be iterated.
4510 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4511 * set, starting from the lowest sched_domain to the highest.
4513 #define for_each_flag_domain(cpu, sd, flag) \
4514 for (sd = lowest_flag_domain(cpu, flag); \
4515 (sd && (sd->flags & flag)); sd = sd->parent)
4518 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4519 * @ilb_group: group to be checked for semi-idleness
4521 * Returns: 1 if the group is semi-idle. 0 otherwise.
4523 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4524 * and atleast one non-idle CPU. This helper function checks if the given
4525 * sched_group is semi-idle or not.
4527 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4529 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4530 sched_group_cpus(ilb_group
));
4533 * A sched_group is semi-idle when it has atleast one busy cpu
4534 * and atleast one idle cpu.
4536 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4539 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4545 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4546 * @cpu: The cpu which is nominating a new idle_load_balancer.
4548 * Returns: Returns the id of the idle load balancer if it exists,
4549 * Else, returns >= nr_cpu_ids.
4551 * This algorithm picks the idle load balancer such that it belongs to a
4552 * semi-idle powersavings sched_domain. The idea is to try and avoid
4553 * completely idle packages/cores just for the purpose of idle load balancing
4554 * when there are other idle cpu's which are better suited for that job.
4556 static int find_new_ilb(int cpu
)
4558 struct sched_domain
*sd
;
4559 struct sched_group
*ilb_group
;
4562 * Have idle load balancer selection from semi-idle packages only
4563 * when power-aware load balancing is enabled
4565 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4569 * Optimize for the case when we have no idle CPUs or only one
4570 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4572 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4575 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4576 ilb_group
= sd
->groups
;
4579 if (is_semi_idle_group(ilb_group
))
4580 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4582 ilb_group
= ilb_group
->next
;
4584 } while (ilb_group
!= sd
->groups
);
4588 return cpumask_first(nohz
.cpu_mask
);
4590 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4591 static inline int find_new_ilb(int call_cpu
)
4593 return cpumask_first(nohz
.cpu_mask
);
4598 * This routine will try to nominate the ilb (idle load balancing)
4599 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4600 * load balancing on behalf of all those cpus. If all the cpus in the system
4601 * go into this tickless mode, then there will be no ilb owner (as there is
4602 * no need for one) and all the cpus will sleep till the next wakeup event
4605 * For the ilb owner, tick is not stopped. And this tick will be used
4606 * for idle load balancing. ilb owner will still be part of
4609 * While stopping the tick, this cpu will become the ilb owner if there
4610 * is no other owner. And will be the owner till that cpu becomes busy
4611 * or if all cpus in the system stop their ticks at which point
4612 * there is no need for ilb owner.
4614 * When the ilb owner becomes busy, it nominates another owner, during the
4615 * next busy scheduler_tick()
4617 int select_nohz_load_balancer(int stop_tick
)
4619 int cpu
= smp_processor_id();
4622 cpu_rq(cpu
)->in_nohz_recently
= 1;
4624 if (!cpu_active(cpu
)) {
4625 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4629 * If we are going offline and still the leader,
4632 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4638 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4640 /* time for ilb owner also to sleep */
4641 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4642 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4643 atomic_set(&nohz
.load_balancer
, -1);
4647 if (atomic_read(&nohz
.load_balancer
) == -1) {
4648 /* make me the ilb owner */
4649 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4651 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4654 if (!(sched_smt_power_savings
||
4655 sched_mc_power_savings
))
4658 * Check to see if there is a more power-efficient
4661 new_ilb
= find_new_ilb(cpu
);
4662 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4663 atomic_set(&nohz
.load_balancer
, -1);
4664 resched_cpu(new_ilb
);
4670 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4673 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4675 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4676 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4683 static DEFINE_SPINLOCK(balancing
);
4686 * It checks each scheduling domain to see if it is due to be balanced,
4687 * and initiates a balancing operation if so.
4689 * Balancing parameters are set up in arch_init_sched_domains.
4691 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4694 struct rq
*rq
= cpu_rq(cpu
);
4695 unsigned long interval
;
4696 struct sched_domain
*sd
;
4697 /* Earliest time when we have to do rebalance again */
4698 unsigned long next_balance
= jiffies
+ 60*HZ
;
4699 int update_next_balance
= 0;
4702 for_each_domain(cpu
, sd
) {
4703 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4706 interval
= sd
->balance_interval
;
4707 if (idle
!= CPU_IDLE
)
4708 interval
*= sd
->busy_factor
;
4710 /* scale ms to jiffies */
4711 interval
= msecs_to_jiffies(interval
);
4712 if (unlikely(!interval
))
4714 if (interval
> HZ
*NR_CPUS
/10)
4715 interval
= HZ
*NR_CPUS
/10;
4717 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4719 if (need_serialize
) {
4720 if (!spin_trylock(&balancing
))
4724 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4725 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4727 * We've pulled tasks over so either we're no
4728 * longer idle, or one of our SMT siblings is
4731 idle
= CPU_NOT_IDLE
;
4733 sd
->last_balance
= jiffies
;
4736 spin_unlock(&balancing
);
4738 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4739 next_balance
= sd
->last_balance
+ interval
;
4740 update_next_balance
= 1;
4744 * Stop the load balance at this level. There is another
4745 * CPU in our sched group which is doing load balancing more
4753 * next_balance will be updated only when there is a need.
4754 * When the cpu is attached to null domain for ex, it will not be
4757 if (likely(update_next_balance
))
4758 rq
->next_balance
= next_balance
;
4762 * run_rebalance_domains is triggered when needed from the scheduler tick.
4763 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4764 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4766 static void run_rebalance_domains(struct softirq_action
*h
)
4768 int this_cpu
= smp_processor_id();
4769 struct rq
*this_rq
= cpu_rq(this_cpu
);
4770 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4771 CPU_IDLE
: CPU_NOT_IDLE
;
4773 rebalance_domains(this_cpu
, idle
);
4777 * If this cpu is the owner for idle load balancing, then do the
4778 * balancing on behalf of the other idle cpus whose ticks are
4781 if (this_rq
->idle_at_tick
&&
4782 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4786 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4787 if (balance_cpu
== this_cpu
)
4791 * If this cpu gets work to do, stop the load balancing
4792 * work being done for other cpus. Next load
4793 * balancing owner will pick it up.
4798 rebalance_domains(balance_cpu
, CPU_IDLE
);
4800 rq
= cpu_rq(balance_cpu
);
4801 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4802 this_rq
->next_balance
= rq
->next_balance
;
4808 static inline int on_null_domain(int cpu
)
4810 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4814 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4816 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4817 * idle load balancing owner or decide to stop the periodic load balancing,
4818 * if the whole system is idle.
4820 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4824 * If we were in the nohz mode recently and busy at the current
4825 * scheduler tick, then check if we need to nominate new idle
4828 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4829 rq
->in_nohz_recently
= 0;
4831 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4832 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4833 atomic_set(&nohz
.load_balancer
, -1);
4836 if (atomic_read(&nohz
.load_balancer
) == -1) {
4837 int ilb
= find_new_ilb(cpu
);
4839 if (ilb
< nr_cpu_ids
)
4845 * If this cpu is idle and doing idle load balancing for all the
4846 * cpus with ticks stopped, is it time for that to stop?
4848 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4849 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4855 * If this cpu is idle and the idle load balancing is done by
4856 * someone else, then no need raise the SCHED_SOFTIRQ
4858 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4859 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4862 /* Don't need to rebalance while attached to NULL domain */
4863 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4864 likely(!on_null_domain(cpu
)))
4865 raise_softirq(SCHED_SOFTIRQ
);
4868 #else /* CONFIG_SMP */
4871 * on UP we do not need to balance between CPUs:
4873 static inline void idle_balance(int cpu
, struct rq
*rq
)
4879 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4881 EXPORT_PER_CPU_SYMBOL(kstat
);
4884 * Return any ns on the sched_clock that have not yet been accounted in
4885 * @p in case that task is currently running.
4887 * Called with task_rq_lock() held on @rq.
4889 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4893 if (task_current(rq
, p
)) {
4894 update_rq_clock(rq
);
4895 ns
= rq
->clock
- p
->se
.exec_start
;
4903 unsigned long long task_delta_exec(struct task_struct
*p
)
4905 unsigned long flags
;
4909 rq
= task_rq_lock(p
, &flags
);
4910 ns
= do_task_delta_exec(p
, rq
);
4911 task_rq_unlock(rq
, &flags
);
4917 * Return accounted runtime for the task.
4918 * In case the task is currently running, return the runtime plus current's
4919 * pending runtime that have not been accounted yet.
4921 unsigned long long task_sched_runtime(struct task_struct
*p
)
4923 unsigned long flags
;
4927 rq
= task_rq_lock(p
, &flags
);
4928 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4929 task_rq_unlock(rq
, &flags
);
4935 * Return sum_exec_runtime for the thread group.
4936 * In case the task is currently running, return the sum plus current's
4937 * pending runtime that have not been accounted yet.
4939 * Note that the thread group might have other running tasks as well,
4940 * so the return value not includes other pending runtime that other
4941 * running tasks might have.
4943 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4945 struct task_cputime totals
;
4946 unsigned long flags
;
4950 rq
= task_rq_lock(p
, &flags
);
4951 thread_group_cputime(p
, &totals
);
4952 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4953 task_rq_unlock(rq
, &flags
);
4959 * Account user cpu time to a process.
4960 * @p: the process that the cpu time gets accounted to
4961 * @cputime: the cpu time spent in user space since the last update
4962 * @cputime_scaled: cputime scaled by cpu frequency
4964 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4965 cputime_t cputime_scaled
)
4967 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4970 /* Add user time to process. */
4971 p
->utime
= cputime_add(p
->utime
, cputime
);
4972 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4973 account_group_user_time(p
, cputime
);
4975 /* Add user time to cpustat. */
4976 tmp
= cputime_to_cputime64(cputime
);
4977 if (TASK_NICE(p
) > 0)
4978 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4980 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4982 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4983 /* Account for user time used */
4984 acct_update_integrals(p
);
4988 * Account guest cpu time to a process.
4989 * @p: the process that the cpu time gets accounted to
4990 * @cputime: the cpu time spent in virtual machine since the last update
4991 * @cputime_scaled: cputime scaled by cpu frequency
4993 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4994 cputime_t cputime_scaled
)
4997 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4999 tmp
= cputime_to_cputime64(cputime
);
5001 /* Add guest time to process. */
5002 p
->utime
= cputime_add(p
->utime
, cputime
);
5003 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5004 account_group_user_time(p
, cputime
);
5005 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5007 /* Add guest time to cpustat. */
5008 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5009 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5013 * Account system cpu time to a process.
5014 * @p: the process that the cpu time gets accounted to
5015 * @hardirq_offset: the offset to subtract from hardirq_count()
5016 * @cputime: the cpu time spent in kernel space since the last update
5017 * @cputime_scaled: cputime scaled by cpu frequency
5019 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5020 cputime_t cputime
, cputime_t cputime_scaled
)
5022 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5025 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5026 account_guest_time(p
, cputime
, cputime_scaled
);
5030 /* Add system time to process. */
5031 p
->stime
= cputime_add(p
->stime
, cputime
);
5032 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5033 account_group_system_time(p
, cputime
);
5035 /* Add system time to cpustat. */
5036 tmp
= cputime_to_cputime64(cputime
);
5037 if (hardirq_count() - hardirq_offset
)
5038 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5039 else if (softirq_count())
5040 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5042 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5044 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5046 /* Account for system time used */
5047 acct_update_integrals(p
);
5051 * Account for involuntary wait time.
5052 * @steal: the cpu time spent in involuntary wait
5054 void account_steal_time(cputime_t cputime
)
5056 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5057 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5059 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5063 * Account for idle time.
5064 * @cputime: the cpu time spent in idle wait
5066 void account_idle_time(cputime_t cputime
)
5068 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5069 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5070 struct rq
*rq
= this_rq();
5072 if (atomic_read(&rq
->nr_iowait
) > 0)
5073 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5075 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5078 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5081 * Account a single tick of cpu time.
5082 * @p: the process that the cpu time gets accounted to
5083 * @user_tick: indicates if the tick is a user or a system tick
5085 void account_process_tick(struct task_struct
*p
, int user_tick
)
5087 cputime_t one_jiffy
= jiffies_to_cputime(1);
5088 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5089 struct rq
*rq
= this_rq();
5092 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5093 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5094 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5097 account_idle_time(one_jiffy
);
5101 * Account multiple ticks of steal time.
5102 * @p: the process from which the cpu time has been stolen
5103 * @ticks: number of stolen ticks
5105 void account_steal_ticks(unsigned long ticks
)
5107 account_steal_time(jiffies_to_cputime(ticks
));
5111 * Account multiple ticks of idle time.
5112 * @ticks: number of stolen ticks
5114 void account_idle_ticks(unsigned long ticks
)
5116 account_idle_time(jiffies_to_cputime(ticks
));
5122 * Use precise platform statistics if available:
5124 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5125 cputime_t
task_utime(struct task_struct
*p
)
5130 cputime_t
task_stime(struct task_struct
*p
)
5135 cputime_t
task_utime(struct task_struct
*p
)
5137 clock_t utime
= cputime_to_clock_t(p
->utime
),
5138 total
= utime
+ cputime_to_clock_t(p
->stime
);
5142 * Use CFS's precise accounting:
5144 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5148 do_div(temp
, total
);
5150 utime
= (clock_t)temp
;
5152 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5153 return p
->prev_utime
;
5156 cputime_t
task_stime(struct task_struct
*p
)
5161 * Use CFS's precise accounting. (we subtract utime from
5162 * the total, to make sure the total observed by userspace
5163 * grows monotonically - apps rely on that):
5165 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5166 cputime_to_clock_t(task_utime(p
));
5169 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5171 return p
->prev_stime
;
5175 inline cputime_t
task_gtime(struct task_struct
*p
)
5181 * This function gets called by the timer code, with HZ frequency.
5182 * We call it with interrupts disabled.
5184 * It also gets called by the fork code, when changing the parent's
5187 void scheduler_tick(void)
5189 int cpu
= smp_processor_id();
5190 struct rq
*rq
= cpu_rq(cpu
);
5191 struct task_struct
*curr
= rq
->curr
;
5195 spin_lock(&rq
->lock
);
5196 update_rq_clock(rq
);
5197 update_cpu_load(rq
);
5198 curr
->sched_class
->task_tick(rq
, curr
, 0);
5199 spin_unlock(&rq
->lock
);
5201 perf_counter_task_tick(curr
, cpu
);
5204 rq
->idle_at_tick
= idle_cpu(cpu
);
5205 trigger_load_balance(rq
, cpu
);
5209 notrace
unsigned long get_parent_ip(unsigned long addr
)
5211 if (in_lock_functions(addr
)) {
5212 addr
= CALLER_ADDR2
;
5213 if (in_lock_functions(addr
))
5214 addr
= CALLER_ADDR3
;
5219 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5220 defined(CONFIG_PREEMPT_TRACER))
5222 void __kprobes
add_preempt_count(int val
)
5224 #ifdef CONFIG_DEBUG_PREEMPT
5228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5231 preempt_count() += val
;
5232 #ifdef CONFIG_DEBUG_PREEMPT
5234 * Spinlock count overflowing soon?
5236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5239 if (preempt_count() == val
)
5240 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5242 EXPORT_SYMBOL(add_preempt_count
);
5244 void __kprobes
sub_preempt_count(int val
)
5246 #ifdef CONFIG_DEBUG_PREEMPT
5250 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5253 * Is the spinlock portion underflowing?
5255 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5256 !(preempt_count() & PREEMPT_MASK
)))
5260 if (preempt_count() == val
)
5261 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5262 preempt_count() -= val
;
5264 EXPORT_SYMBOL(sub_preempt_count
);
5269 * Print scheduling while atomic bug:
5271 static noinline
void __schedule_bug(struct task_struct
*prev
)
5273 struct pt_regs
*regs
= get_irq_regs();
5275 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5276 prev
->comm
, prev
->pid
, preempt_count());
5278 debug_show_held_locks(prev
);
5280 if (irqs_disabled())
5281 print_irqtrace_events(prev
);
5290 * Various schedule()-time debugging checks and statistics:
5292 static inline void schedule_debug(struct task_struct
*prev
)
5295 * Test if we are atomic. Since do_exit() needs to call into
5296 * schedule() atomically, we ignore that path for now.
5297 * Otherwise, whine if we are scheduling when we should not be.
5299 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5300 __schedule_bug(prev
);
5302 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5304 schedstat_inc(this_rq(), sched_count
);
5305 #ifdef CONFIG_SCHEDSTATS
5306 if (unlikely(prev
->lock_depth
>= 0)) {
5307 schedstat_inc(this_rq(), bkl_count
);
5308 schedstat_inc(prev
, sched_info
.bkl_count
);
5313 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5315 if (prev
->state
== TASK_RUNNING
) {
5316 u64 runtime
= prev
->se
.sum_exec_runtime
;
5318 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5319 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5322 * In order to avoid avg_overlap growing stale when we are
5323 * indeed overlapping and hence not getting put to sleep, grow
5324 * the avg_overlap on preemption.
5326 * We use the average preemption runtime because that
5327 * correlates to the amount of cache footprint a task can
5330 update_avg(&prev
->se
.avg_overlap
, runtime
);
5332 prev
->sched_class
->put_prev_task(rq
, prev
);
5336 * Pick up the highest-prio task:
5338 static inline struct task_struct
*
5339 pick_next_task(struct rq
*rq
)
5341 const struct sched_class
*class;
5342 struct task_struct
*p
;
5345 * Optimization: we know that if all tasks are in
5346 * the fair class we can call that function directly:
5348 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5349 p
= fair_sched_class
.pick_next_task(rq
);
5354 class = sched_class_highest
;
5356 p
= class->pick_next_task(rq
);
5360 * Will never be NULL as the idle class always
5361 * returns a non-NULL p:
5363 class = class->next
;
5368 * schedule() is the main scheduler function.
5370 asmlinkage
void __sched
schedule(void)
5372 struct task_struct
*prev
, *next
;
5373 unsigned long *switch_count
;
5379 cpu
= smp_processor_id();
5383 switch_count
= &prev
->nivcsw
;
5385 release_kernel_lock(prev
);
5386 need_resched_nonpreemptible
:
5388 schedule_debug(prev
);
5390 if (sched_feat(HRTICK
))
5393 spin_lock_irq(&rq
->lock
);
5394 update_rq_clock(rq
);
5395 clear_tsk_need_resched(prev
);
5397 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5398 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5399 prev
->state
= TASK_RUNNING
;
5401 deactivate_task(rq
, prev
, 1);
5402 switch_count
= &prev
->nvcsw
;
5405 pre_schedule(rq
, prev
);
5407 if (unlikely(!rq
->nr_running
))
5408 idle_balance(cpu
, rq
);
5410 put_prev_task(rq
, prev
);
5411 next
= pick_next_task(rq
);
5413 if (likely(prev
!= next
)) {
5414 sched_info_switch(prev
, next
);
5415 perf_counter_task_sched_out(prev
, next
, cpu
);
5421 context_switch(rq
, prev
, next
); /* unlocks the rq */
5423 * the context switch might have flipped the stack from under
5424 * us, hence refresh the local variables.
5426 cpu
= smp_processor_id();
5429 spin_unlock_irq(&rq
->lock
);
5433 if (unlikely(reacquire_kernel_lock(current
) < 0))
5434 goto need_resched_nonpreemptible
;
5436 preempt_enable_no_resched();
5440 EXPORT_SYMBOL(schedule
);
5444 * Look out! "owner" is an entirely speculative pointer
5445 * access and not reliable.
5447 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5452 if (!sched_feat(OWNER_SPIN
))
5455 #ifdef CONFIG_DEBUG_PAGEALLOC
5457 * Need to access the cpu field knowing that
5458 * DEBUG_PAGEALLOC could have unmapped it if
5459 * the mutex owner just released it and exited.
5461 if (probe_kernel_address(&owner
->cpu
, cpu
))
5468 * Even if the access succeeded (likely case),
5469 * the cpu field may no longer be valid.
5471 if (cpu
>= nr_cpumask_bits
)
5475 * We need to validate that we can do a
5476 * get_cpu() and that we have the percpu area.
5478 if (!cpu_online(cpu
))
5485 * Owner changed, break to re-assess state.
5487 if (lock
->owner
!= owner
)
5491 * Is that owner really running on that cpu?
5493 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5503 #ifdef CONFIG_PREEMPT
5505 * this is the entry point to schedule() from in-kernel preemption
5506 * off of preempt_enable. Kernel preemptions off return from interrupt
5507 * occur there and call schedule directly.
5509 asmlinkage
void __sched
preempt_schedule(void)
5511 struct thread_info
*ti
= current_thread_info();
5514 * If there is a non-zero preempt_count or interrupts are disabled,
5515 * we do not want to preempt the current task. Just return..
5517 if (likely(ti
->preempt_count
|| irqs_disabled()))
5521 add_preempt_count(PREEMPT_ACTIVE
);
5523 sub_preempt_count(PREEMPT_ACTIVE
);
5526 * Check again in case we missed a preemption opportunity
5527 * between schedule and now.
5530 } while (need_resched());
5532 EXPORT_SYMBOL(preempt_schedule
);
5535 * this is the entry point to schedule() from kernel preemption
5536 * off of irq context.
5537 * Note, that this is called and return with irqs disabled. This will
5538 * protect us against recursive calling from irq.
5540 asmlinkage
void __sched
preempt_schedule_irq(void)
5542 struct thread_info
*ti
= current_thread_info();
5544 /* Catch callers which need to be fixed */
5545 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5548 add_preempt_count(PREEMPT_ACTIVE
);
5551 local_irq_disable();
5552 sub_preempt_count(PREEMPT_ACTIVE
);
5555 * Check again in case we missed a preemption opportunity
5556 * between schedule and now.
5559 } while (need_resched());
5562 #endif /* CONFIG_PREEMPT */
5564 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5567 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5569 EXPORT_SYMBOL(default_wake_function
);
5572 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5573 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5574 * number) then we wake all the non-exclusive tasks and one exclusive task.
5576 * There are circumstances in which we can try to wake a task which has already
5577 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5578 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5580 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5581 int nr_exclusive
, int wake_flags
, void *key
)
5583 wait_queue_t
*curr
, *next
;
5585 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5586 unsigned flags
= curr
->flags
;
5588 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5589 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5595 * __wake_up - wake up threads blocked on a waitqueue.
5597 * @mode: which threads
5598 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5599 * @key: is directly passed to the wakeup function
5601 * It may be assumed that this function implies a write memory barrier before
5602 * changing the task state if and only if any tasks are woken up.
5604 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5605 int nr_exclusive
, void *key
)
5607 unsigned long flags
;
5609 spin_lock_irqsave(&q
->lock
, flags
);
5610 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5611 spin_unlock_irqrestore(&q
->lock
, flags
);
5613 EXPORT_SYMBOL(__wake_up
);
5616 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5618 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5620 __wake_up_common(q
, mode
, 1, 0, NULL
);
5623 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5625 __wake_up_common(q
, mode
, 1, 0, key
);
5629 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5631 * @mode: which threads
5632 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5633 * @key: opaque value to be passed to wakeup targets
5635 * The sync wakeup differs that the waker knows that it will schedule
5636 * away soon, so while the target thread will be woken up, it will not
5637 * be migrated to another CPU - ie. the two threads are 'synchronized'
5638 * with each other. This can prevent needless bouncing between CPUs.
5640 * On UP it can prevent extra preemption.
5642 * It may be assumed that this function implies a write memory barrier before
5643 * changing the task state if and only if any tasks are woken up.
5645 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5646 int nr_exclusive
, void *key
)
5648 unsigned long flags
;
5649 int wake_flags
= WF_SYNC
;
5654 if (unlikely(!nr_exclusive
))
5657 spin_lock_irqsave(&q
->lock
, flags
);
5658 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5659 spin_unlock_irqrestore(&q
->lock
, flags
);
5661 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5664 * __wake_up_sync - see __wake_up_sync_key()
5666 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5668 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5670 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5673 * complete: - signals a single thread waiting on this completion
5674 * @x: holds the state of this particular completion
5676 * This will wake up a single thread waiting on this completion. Threads will be
5677 * awakened in the same order in which they were queued.
5679 * See also complete_all(), wait_for_completion() and related routines.
5681 * It may be assumed that this function implies a write memory barrier before
5682 * changing the task state if and only if any tasks are woken up.
5684 void complete(struct completion
*x
)
5686 unsigned long flags
;
5688 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5690 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5691 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5693 EXPORT_SYMBOL(complete
);
5696 * complete_all: - signals all threads waiting on this completion
5697 * @x: holds the state of this particular completion
5699 * This will wake up all threads waiting on this particular completion event.
5701 * It may be assumed that this function implies a write memory barrier before
5702 * changing the task state if and only if any tasks are woken up.
5704 void complete_all(struct completion
*x
)
5706 unsigned long flags
;
5708 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5709 x
->done
+= UINT_MAX
/2;
5710 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5711 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5713 EXPORT_SYMBOL(complete_all
);
5715 static inline long __sched
5716 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5719 DECLARE_WAITQUEUE(wait
, current
);
5721 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5722 __add_wait_queue_tail(&x
->wait
, &wait
);
5724 if (signal_pending_state(state
, current
)) {
5725 timeout
= -ERESTARTSYS
;
5728 __set_current_state(state
);
5729 spin_unlock_irq(&x
->wait
.lock
);
5730 timeout
= schedule_timeout(timeout
);
5731 spin_lock_irq(&x
->wait
.lock
);
5732 } while (!x
->done
&& timeout
);
5733 __remove_wait_queue(&x
->wait
, &wait
);
5738 return timeout
?: 1;
5742 wait_for_common(struct completion
*x
, long timeout
, int state
)
5746 spin_lock_irq(&x
->wait
.lock
);
5747 timeout
= do_wait_for_common(x
, timeout
, state
);
5748 spin_unlock_irq(&x
->wait
.lock
);
5753 * wait_for_completion: - waits for completion of a task
5754 * @x: holds the state of this particular completion
5756 * This waits to be signaled for completion of a specific task. It is NOT
5757 * interruptible and there is no timeout.
5759 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5760 * and interrupt capability. Also see complete().
5762 void __sched
wait_for_completion(struct completion
*x
)
5764 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5766 EXPORT_SYMBOL(wait_for_completion
);
5769 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5770 * @x: holds the state of this particular completion
5771 * @timeout: timeout value in jiffies
5773 * This waits for either a completion of a specific task to be signaled or for a
5774 * specified timeout to expire. The timeout is in jiffies. It is not
5777 unsigned long __sched
5778 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5780 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5782 EXPORT_SYMBOL(wait_for_completion_timeout
);
5785 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5786 * @x: holds the state of this particular completion
5788 * This waits for completion of a specific task to be signaled. It is
5791 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5793 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5794 if (t
== -ERESTARTSYS
)
5798 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5801 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5802 * @x: holds the state of this particular completion
5803 * @timeout: timeout value in jiffies
5805 * This waits for either a completion of a specific task to be signaled or for a
5806 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5808 unsigned long __sched
5809 wait_for_completion_interruptible_timeout(struct completion
*x
,
5810 unsigned long timeout
)
5812 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5814 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5817 * wait_for_completion_killable: - waits for completion of a task (killable)
5818 * @x: holds the state of this particular completion
5820 * This waits to be signaled for completion of a specific task. It can be
5821 * interrupted by a kill signal.
5823 int __sched
wait_for_completion_killable(struct completion
*x
)
5825 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5826 if (t
== -ERESTARTSYS
)
5830 EXPORT_SYMBOL(wait_for_completion_killable
);
5833 * try_wait_for_completion - try to decrement a completion without blocking
5834 * @x: completion structure
5836 * Returns: 0 if a decrement cannot be done without blocking
5837 * 1 if a decrement succeeded.
5839 * If a completion is being used as a counting completion,
5840 * attempt to decrement the counter without blocking. This
5841 * enables us to avoid waiting if the resource the completion
5842 * is protecting is not available.
5844 bool try_wait_for_completion(struct completion
*x
)
5848 spin_lock_irq(&x
->wait
.lock
);
5853 spin_unlock_irq(&x
->wait
.lock
);
5856 EXPORT_SYMBOL(try_wait_for_completion
);
5859 * completion_done - Test to see if a completion has any waiters
5860 * @x: completion structure
5862 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5863 * 1 if there are no waiters.
5866 bool completion_done(struct completion
*x
)
5870 spin_lock_irq(&x
->wait
.lock
);
5873 spin_unlock_irq(&x
->wait
.lock
);
5876 EXPORT_SYMBOL(completion_done
);
5879 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5881 unsigned long flags
;
5884 init_waitqueue_entry(&wait
, current
);
5886 __set_current_state(state
);
5888 spin_lock_irqsave(&q
->lock
, flags
);
5889 __add_wait_queue(q
, &wait
);
5890 spin_unlock(&q
->lock
);
5891 timeout
= schedule_timeout(timeout
);
5892 spin_lock_irq(&q
->lock
);
5893 __remove_wait_queue(q
, &wait
);
5894 spin_unlock_irqrestore(&q
->lock
, flags
);
5899 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5901 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5903 EXPORT_SYMBOL(interruptible_sleep_on
);
5906 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5908 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5910 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5912 void __sched
sleep_on(wait_queue_head_t
*q
)
5914 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5916 EXPORT_SYMBOL(sleep_on
);
5918 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5920 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5922 EXPORT_SYMBOL(sleep_on_timeout
);
5924 #ifdef CONFIG_RT_MUTEXES
5927 * rt_mutex_setprio - set the current priority of a task
5929 * @prio: prio value (kernel-internal form)
5931 * This function changes the 'effective' priority of a task. It does
5932 * not touch ->normal_prio like __setscheduler().
5934 * Used by the rt_mutex code to implement priority inheritance logic.
5936 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5938 unsigned long flags
;
5939 int oldprio
, on_rq
, running
;
5941 const struct sched_class
*prev_class
= p
->sched_class
;
5943 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5945 rq
= task_rq_lock(p
, &flags
);
5946 update_rq_clock(rq
);
5949 on_rq
= p
->se
.on_rq
;
5950 running
= task_current(rq
, p
);
5952 dequeue_task(rq
, p
, 0);
5954 p
->sched_class
->put_prev_task(rq
, p
);
5957 p
->sched_class
= &rt_sched_class
;
5959 p
->sched_class
= &fair_sched_class
;
5964 p
->sched_class
->set_curr_task(rq
);
5966 enqueue_task(rq
, p
, 0);
5968 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5970 task_rq_unlock(rq
, &flags
);
5975 void set_user_nice(struct task_struct
*p
, long nice
)
5977 int old_prio
, delta
, on_rq
;
5978 unsigned long flags
;
5981 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5984 * We have to be careful, if called from sys_setpriority(),
5985 * the task might be in the middle of scheduling on another CPU.
5987 rq
= task_rq_lock(p
, &flags
);
5988 update_rq_clock(rq
);
5990 * The RT priorities are set via sched_setscheduler(), but we still
5991 * allow the 'normal' nice value to be set - but as expected
5992 * it wont have any effect on scheduling until the task is
5993 * SCHED_FIFO/SCHED_RR:
5995 if (task_has_rt_policy(p
)) {
5996 p
->static_prio
= NICE_TO_PRIO(nice
);
5999 on_rq
= p
->se
.on_rq
;
6001 dequeue_task(rq
, p
, 0);
6003 p
->static_prio
= NICE_TO_PRIO(nice
);
6006 p
->prio
= effective_prio(p
);
6007 delta
= p
->prio
- old_prio
;
6010 enqueue_task(rq
, p
, 0);
6012 * If the task increased its priority or is running and
6013 * lowered its priority, then reschedule its CPU:
6015 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6016 resched_task(rq
->curr
);
6019 task_rq_unlock(rq
, &flags
);
6021 EXPORT_SYMBOL(set_user_nice
);
6024 * can_nice - check if a task can reduce its nice value
6028 int can_nice(const struct task_struct
*p
, const int nice
)
6030 /* convert nice value [19,-20] to rlimit style value [1,40] */
6031 int nice_rlim
= 20 - nice
;
6033 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6034 capable(CAP_SYS_NICE
));
6037 #ifdef __ARCH_WANT_SYS_NICE
6040 * sys_nice - change the priority of the current process.
6041 * @increment: priority increment
6043 * sys_setpriority is a more generic, but much slower function that
6044 * does similar things.
6046 SYSCALL_DEFINE1(nice
, int, increment
)
6051 * Setpriority might change our priority at the same moment.
6052 * We don't have to worry. Conceptually one call occurs first
6053 * and we have a single winner.
6055 if (increment
< -40)
6060 nice
= TASK_NICE(current
) + increment
;
6066 if (increment
< 0 && !can_nice(current
, nice
))
6069 retval
= security_task_setnice(current
, nice
);
6073 set_user_nice(current
, nice
);
6080 * task_prio - return the priority value of a given task.
6081 * @p: the task in question.
6083 * This is the priority value as seen by users in /proc.
6084 * RT tasks are offset by -200. Normal tasks are centered
6085 * around 0, value goes from -16 to +15.
6087 int task_prio(const struct task_struct
*p
)
6089 return p
->prio
- MAX_RT_PRIO
;
6093 * task_nice - return the nice value of a given task.
6094 * @p: the task in question.
6096 int task_nice(const struct task_struct
*p
)
6098 return TASK_NICE(p
);
6100 EXPORT_SYMBOL(task_nice
);
6103 * idle_cpu - is a given cpu idle currently?
6104 * @cpu: the processor in question.
6106 int idle_cpu(int cpu
)
6108 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6112 * idle_task - return the idle task for a given cpu.
6113 * @cpu: the processor in question.
6115 struct task_struct
*idle_task(int cpu
)
6117 return cpu_rq(cpu
)->idle
;
6121 * find_process_by_pid - find a process with a matching PID value.
6122 * @pid: the pid in question.
6124 static struct task_struct
*find_process_by_pid(pid_t pid
)
6126 return pid
? find_task_by_vpid(pid
) : current
;
6129 /* Actually do priority change: must hold rq lock. */
6131 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6133 BUG_ON(p
->se
.on_rq
);
6136 switch (p
->policy
) {
6140 p
->sched_class
= &fair_sched_class
;
6144 p
->sched_class
= &rt_sched_class
;
6148 p
->rt_priority
= prio
;
6149 p
->normal_prio
= normal_prio(p
);
6150 /* we are holding p->pi_lock already */
6151 p
->prio
= rt_mutex_getprio(p
);
6156 * check the target process has a UID that matches the current process's
6158 static bool check_same_owner(struct task_struct
*p
)
6160 const struct cred
*cred
= current_cred(), *pcred
;
6164 pcred
= __task_cred(p
);
6165 match
= (cred
->euid
== pcred
->euid
||
6166 cred
->euid
== pcred
->uid
);
6171 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6172 struct sched_param
*param
, bool user
)
6174 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6175 unsigned long flags
;
6176 const struct sched_class
*prev_class
= p
->sched_class
;
6180 /* may grab non-irq protected spin_locks */
6181 BUG_ON(in_interrupt());
6183 /* double check policy once rq lock held */
6185 reset_on_fork
= p
->sched_reset_on_fork
;
6186 policy
= oldpolicy
= p
->policy
;
6188 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6189 policy
&= ~SCHED_RESET_ON_FORK
;
6191 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6192 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6193 policy
!= SCHED_IDLE
)
6198 * Valid priorities for SCHED_FIFO and SCHED_RR are
6199 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6200 * SCHED_BATCH and SCHED_IDLE is 0.
6202 if (param
->sched_priority
< 0 ||
6203 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6204 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6206 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6210 * Allow unprivileged RT tasks to decrease priority:
6212 if (user
&& !capable(CAP_SYS_NICE
)) {
6213 if (rt_policy(policy
)) {
6214 unsigned long rlim_rtprio
;
6216 if (!lock_task_sighand(p
, &flags
))
6218 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6219 unlock_task_sighand(p
, &flags
);
6221 /* can't set/change the rt policy */
6222 if (policy
!= p
->policy
&& !rlim_rtprio
)
6225 /* can't increase priority */
6226 if (param
->sched_priority
> p
->rt_priority
&&
6227 param
->sched_priority
> rlim_rtprio
)
6231 * Like positive nice levels, dont allow tasks to
6232 * move out of SCHED_IDLE either:
6234 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6237 /* can't change other user's priorities */
6238 if (!check_same_owner(p
))
6241 /* Normal users shall not reset the sched_reset_on_fork flag */
6242 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6247 #ifdef CONFIG_RT_GROUP_SCHED
6249 * Do not allow realtime tasks into groups that have no runtime
6252 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6253 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6257 retval
= security_task_setscheduler(p
, policy
, param
);
6263 * make sure no PI-waiters arrive (or leave) while we are
6264 * changing the priority of the task:
6266 spin_lock_irqsave(&p
->pi_lock
, flags
);
6268 * To be able to change p->policy safely, the apropriate
6269 * runqueue lock must be held.
6271 rq
= __task_rq_lock(p
);
6272 /* recheck policy now with rq lock held */
6273 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6274 policy
= oldpolicy
= -1;
6275 __task_rq_unlock(rq
);
6276 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6279 update_rq_clock(rq
);
6280 on_rq
= p
->se
.on_rq
;
6281 running
= task_current(rq
, p
);
6283 deactivate_task(rq
, p
, 0);
6285 p
->sched_class
->put_prev_task(rq
, p
);
6287 p
->sched_reset_on_fork
= reset_on_fork
;
6290 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6293 p
->sched_class
->set_curr_task(rq
);
6295 activate_task(rq
, p
, 0);
6297 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6299 __task_rq_unlock(rq
);
6300 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6302 rt_mutex_adjust_pi(p
);
6308 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6309 * @p: the task in question.
6310 * @policy: new policy.
6311 * @param: structure containing the new RT priority.
6313 * NOTE that the task may be already dead.
6315 int sched_setscheduler(struct task_struct
*p
, int policy
,
6316 struct sched_param
*param
)
6318 return __sched_setscheduler(p
, policy
, param
, true);
6320 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6323 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6324 * @p: the task in question.
6325 * @policy: new policy.
6326 * @param: structure containing the new RT priority.
6328 * Just like sched_setscheduler, only don't bother checking if the
6329 * current context has permission. For example, this is needed in
6330 * stop_machine(): we create temporary high priority worker threads,
6331 * but our caller might not have that capability.
6333 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6334 struct sched_param
*param
)
6336 return __sched_setscheduler(p
, policy
, param
, false);
6340 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6342 struct sched_param lparam
;
6343 struct task_struct
*p
;
6346 if (!param
|| pid
< 0)
6348 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6353 p
= find_process_by_pid(pid
);
6355 retval
= sched_setscheduler(p
, policy
, &lparam
);
6362 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6363 * @pid: the pid in question.
6364 * @policy: new policy.
6365 * @param: structure containing the new RT priority.
6367 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6368 struct sched_param __user
*, param
)
6370 /* negative values for policy are not valid */
6374 return do_sched_setscheduler(pid
, policy
, param
);
6378 * sys_sched_setparam - set/change the RT priority of a thread
6379 * @pid: the pid in question.
6380 * @param: structure containing the new RT priority.
6382 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6384 return do_sched_setscheduler(pid
, -1, param
);
6388 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6389 * @pid: the pid in question.
6391 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6393 struct task_struct
*p
;
6400 read_lock(&tasklist_lock
);
6401 p
= find_process_by_pid(pid
);
6403 retval
= security_task_getscheduler(p
);
6406 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6408 read_unlock(&tasklist_lock
);
6413 * sys_sched_getparam - get the RT priority of a thread
6414 * @pid: the pid in question.
6415 * @param: structure containing the RT priority.
6417 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6419 struct sched_param lp
;
6420 struct task_struct
*p
;
6423 if (!param
|| pid
< 0)
6426 read_lock(&tasklist_lock
);
6427 p
= find_process_by_pid(pid
);
6432 retval
= security_task_getscheduler(p
);
6436 lp
.sched_priority
= p
->rt_priority
;
6437 read_unlock(&tasklist_lock
);
6440 * This one might sleep, we cannot do it with a spinlock held ...
6442 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6447 read_unlock(&tasklist_lock
);
6451 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6453 cpumask_var_t cpus_allowed
, new_mask
;
6454 struct task_struct
*p
;
6458 read_lock(&tasklist_lock
);
6460 p
= find_process_by_pid(pid
);
6462 read_unlock(&tasklist_lock
);
6468 * It is not safe to call set_cpus_allowed with the
6469 * tasklist_lock held. We will bump the task_struct's
6470 * usage count and then drop tasklist_lock.
6473 read_unlock(&tasklist_lock
);
6475 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6479 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6481 goto out_free_cpus_allowed
;
6484 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6487 retval
= security_task_setscheduler(p
, 0, NULL
);
6491 cpuset_cpus_allowed(p
, cpus_allowed
);
6492 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6494 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6497 cpuset_cpus_allowed(p
, cpus_allowed
);
6498 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6500 * We must have raced with a concurrent cpuset
6501 * update. Just reset the cpus_allowed to the
6502 * cpuset's cpus_allowed
6504 cpumask_copy(new_mask
, cpus_allowed
);
6509 free_cpumask_var(new_mask
);
6510 out_free_cpus_allowed
:
6511 free_cpumask_var(cpus_allowed
);
6518 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6519 struct cpumask
*new_mask
)
6521 if (len
< cpumask_size())
6522 cpumask_clear(new_mask
);
6523 else if (len
> cpumask_size())
6524 len
= cpumask_size();
6526 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6530 * sys_sched_setaffinity - set the cpu affinity of a process
6531 * @pid: pid of the process
6532 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6533 * @user_mask_ptr: user-space pointer to the new cpu mask
6535 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6536 unsigned long __user
*, user_mask_ptr
)
6538 cpumask_var_t new_mask
;
6541 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6544 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6546 retval
= sched_setaffinity(pid
, new_mask
);
6547 free_cpumask_var(new_mask
);
6551 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6553 struct task_struct
*p
;
6557 read_lock(&tasklist_lock
);
6560 p
= find_process_by_pid(pid
);
6564 retval
= security_task_getscheduler(p
);
6568 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6571 read_unlock(&tasklist_lock
);
6578 * sys_sched_getaffinity - get the cpu affinity of a process
6579 * @pid: pid of the process
6580 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6581 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6583 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6584 unsigned long __user
*, user_mask_ptr
)
6589 if (len
< cpumask_size())
6592 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6595 ret
= sched_getaffinity(pid
, mask
);
6597 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6600 ret
= cpumask_size();
6602 free_cpumask_var(mask
);
6608 * sys_sched_yield - yield the current processor to other threads.
6610 * This function yields the current CPU to other tasks. If there are no
6611 * other threads running on this CPU then this function will return.
6613 SYSCALL_DEFINE0(sched_yield
)
6615 struct rq
*rq
= this_rq_lock();
6617 schedstat_inc(rq
, yld_count
);
6618 current
->sched_class
->yield_task(rq
);
6621 * Since we are going to call schedule() anyway, there's
6622 * no need to preempt or enable interrupts:
6624 __release(rq
->lock
);
6625 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6626 _raw_spin_unlock(&rq
->lock
);
6627 preempt_enable_no_resched();
6634 static inline int should_resched(void)
6636 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6639 static void __cond_resched(void)
6641 add_preempt_count(PREEMPT_ACTIVE
);
6643 sub_preempt_count(PREEMPT_ACTIVE
);
6646 int __sched
_cond_resched(void)
6648 if (should_resched()) {
6654 EXPORT_SYMBOL(_cond_resched
);
6657 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6658 * call schedule, and on return reacquire the lock.
6660 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6661 * operations here to prevent schedule() from being called twice (once via
6662 * spin_unlock(), once by hand).
6664 int __cond_resched_lock(spinlock_t
*lock
)
6666 int resched
= should_resched();
6669 lockdep_assert_held(lock
);
6671 if (spin_needbreak(lock
) || resched
) {
6682 EXPORT_SYMBOL(__cond_resched_lock
);
6684 int __sched
__cond_resched_softirq(void)
6686 BUG_ON(!in_softirq());
6688 if (should_resched()) {
6696 EXPORT_SYMBOL(__cond_resched_softirq
);
6699 * yield - yield the current processor to other threads.
6701 * This is a shortcut for kernel-space yielding - it marks the
6702 * thread runnable and calls sys_sched_yield().
6704 void __sched
yield(void)
6706 set_current_state(TASK_RUNNING
);
6709 EXPORT_SYMBOL(yield
);
6712 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6713 * that process accounting knows that this is a task in IO wait state.
6715 * But don't do that if it is a deliberate, throttling IO wait (this task
6716 * has set its backing_dev_info: the queue against which it should throttle)
6718 void __sched
io_schedule(void)
6720 struct rq
*rq
= raw_rq();
6722 delayacct_blkio_start();
6723 atomic_inc(&rq
->nr_iowait
);
6724 current
->in_iowait
= 1;
6726 current
->in_iowait
= 0;
6727 atomic_dec(&rq
->nr_iowait
);
6728 delayacct_blkio_end();
6730 EXPORT_SYMBOL(io_schedule
);
6732 long __sched
io_schedule_timeout(long timeout
)
6734 struct rq
*rq
= raw_rq();
6737 delayacct_blkio_start();
6738 atomic_inc(&rq
->nr_iowait
);
6739 current
->in_iowait
= 1;
6740 ret
= schedule_timeout(timeout
);
6741 current
->in_iowait
= 0;
6742 atomic_dec(&rq
->nr_iowait
);
6743 delayacct_blkio_end();
6748 * sys_sched_get_priority_max - return maximum RT priority.
6749 * @policy: scheduling class.
6751 * this syscall returns the maximum rt_priority that can be used
6752 * by a given scheduling class.
6754 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6761 ret
= MAX_USER_RT_PRIO
-1;
6773 * sys_sched_get_priority_min - return minimum RT priority.
6774 * @policy: scheduling class.
6776 * this syscall returns the minimum rt_priority that can be used
6777 * by a given scheduling class.
6779 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6797 * sys_sched_rr_get_interval - return the default timeslice of a process.
6798 * @pid: pid of the process.
6799 * @interval: userspace pointer to the timeslice value.
6801 * this syscall writes the default timeslice value of a given process
6802 * into the user-space timespec buffer. A value of '0' means infinity.
6804 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6805 struct timespec __user
*, interval
)
6807 struct task_struct
*p
;
6808 unsigned int time_slice
;
6816 read_lock(&tasklist_lock
);
6817 p
= find_process_by_pid(pid
);
6821 retval
= security_task_getscheduler(p
);
6826 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6827 * tasks that are on an otherwise idle runqueue:
6830 if (p
->policy
== SCHED_RR
) {
6831 time_slice
= DEF_TIMESLICE
;
6832 } else if (p
->policy
!= SCHED_FIFO
) {
6833 struct sched_entity
*se
= &p
->se
;
6834 unsigned long flags
;
6837 rq
= task_rq_lock(p
, &flags
);
6838 if (rq
->cfs
.load
.weight
)
6839 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6840 task_rq_unlock(rq
, &flags
);
6842 read_unlock(&tasklist_lock
);
6843 jiffies_to_timespec(time_slice
, &t
);
6844 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6848 read_unlock(&tasklist_lock
);
6852 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6854 void sched_show_task(struct task_struct
*p
)
6856 unsigned long free
= 0;
6859 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6860 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6861 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6862 #if BITS_PER_LONG == 32
6863 if (state
== TASK_RUNNING
)
6864 printk(KERN_CONT
" running ");
6866 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6868 if (state
== TASK_RUNNING
)
6869 printk(KERN_CONT
" running task ");
6871 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6873 #ifdef CONFIG_DEBUG_STACK_USAGE
6874 free
= stack_not_used(p
);
6876 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6877 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6878 (unsigned long)task_thread_info(p
)->flags
);
6880 show_stack(p
, NULL
);
6883 void show_state_filter(unsigned long state_filter
)
6885 struct task_struct
*g
, *p
;
6887 #if BITS_PER_LONG == 32
6889 " task PC stack pid father\n");
6892 " task PC stack pid father\n");
6894 read_lock(&tasklist_lock
);
6895 do_each_thread(g
, p
) {
6897 * reset the NMI-timeout, listing all files on a slow
6898 * console might take alot of time:
6900 touch_nmi_watchdog();
6901 if (!state_filter
|| (p
->state
& state_filter
))
6903 } while_each_thread(g
, p
);
6905 touch_all_softlockup_watchdogs();
6907 #ifdef CONFIG_SCHED_DEBUG
6908 sysrq_sched_debug_show();
6910 read_unlock(&tasklist_lock
);
6912 * Only show locks if all tasks are dumped:
6914 if (state_filter
== -1)
6915 debug_show_all_locks();
6918 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6920 idle
->sched_class
= &idle_sched_class
;
6924 * init_idle - set up an idle thread for a given CPU
6925 * @idle: task in question
6926 * @cpu: cpu the idle task belongs to
6928 * NOTE: this function does not set the idle thread's NEED_RESCHED
6929 * flag, to make booting more robust.
6931 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6933 struct rq
*rq
= cpu_rq(cpu
);
6934 unsigned long flags
;
6936 spin_lock_irqsave(&rq
->lock
, flags
);
6939 idle
->se
.exec_start
= sched_clock();
6941 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6942 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6943 __set_task_cpu(idle
, cpu
);
6945 rq
->curr
= rq
->idle
= idle
;
6946 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6949 spin_unlock_irqrestore(&rq
->lock
, flags
);
6951 /* Set the preempt count _outside_ the spinlocks! */
6952 #if defined(CONFIG_PREEMPT)
6953 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6955 task_thread_info(idle
)->preempt_count
= 0;
6958 * The idle tasks have their own, simple scheduling class:
6960 idle
->sched_class
= &idle_sched_class
;
6961 ftrace_graph_init_task(idle
);
6965 * In a system that switches off the HZ timer nohz_cpu_mask
6966 * indicates which cpus entered this state. This is used
6967 * in the rcu update to wait only for active cpus. For system
6968 * which do not switch off the HZ timer nohz_cpu_mask should
6969 * always be CPU_BITS_NONE.
6971 cpumask_var_t nohz_cpu_mask
;
6974 * Increase the granularity value when there are more CPUs,
6975 * because with more CPUs the 'effective latency' as visible
6976 * to users decreases. But the relationship is not linear,
6977 * so pick a second-best guess by going with the log2 of the
6980 * This idea comes from the SD scheduler of Con Kolivas:
6982 static inline void sched_init_granularity(void)
6984 unsigned int factor
= 1 + ilog2(num_online_cpus());
6985 const unsigned long limit
= 200000000;
6987 sysctl_sched_min_granularity
*= factor
;
6988 if (sysctl_sched_min_granularity
> limit
)
6989 sysctl_sched_min_granularity
= limit
;
6991 sysctl_sched_latency
*= factor
;
6992 if (sysctl_sched_latency
> limit
)
6993 sysctl_sched_latency
= limit
;
6995 sysctl_sched_wakeup_granularity
*= factor
;
6997 sysctl_sched_shares_ratelimit
*= factor
;
7002 * This is how migration works:
7004 * 1) we queue a struct migration_req structure in the source CPU's
7005 * runqueue and wake up that CPU's migration thread.
7006 * 2) we down() the locked semaphore => thread blocks.
7007 * 3) migration thread wakes up (implicitly it forces the migrated
7008 * thread off the CPU)
7009 * 4) it gets the migration request and checks whether the migrated
7010 * task is still in the wrong runqueue.
7011 * 5) if it's in the wrong runqueue then the migration thread removes
7012 * it and puts it into the right queue.
7013 * 6) migration thread up()s the semaphore.
7014 * 7) we wake up and the migration is done.
7018 * Change a given task's CPU affinity. Migrate the thread to a
7019 * proper CPU and schedule it away if the CPU it's executing on
7020 * is removed from the allowed bitmask.
7022 * NOTE: the caller must have a valid reference to the task, the
7023 * task must not exit() & deallocate itself prematurely. The
7024 * call is not atomic; no spinlocks may be held.
7026 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7028 struct migration_req req
;
7029 unsigned long flags
;
7033 rq
= task_rq_lock(p
, &flags
);
7034 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7039 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7040 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7045 if (p
->sched_class
->set_cpus_allowed
)
7046 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7048 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7049 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7052 /* Can the task run on the task's current CPU? If so, we're done */
7053 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7056 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7057 /* Need help from migration thread: drop lock and wait. */
7058 struct task_struct
*mt
= rq
->migration_thread
;
7060 get_task_struct(mt
);
7061 task_rq_unlock(rq
, &flags
);
7062 wake_up_process(rq
->migration_thread
);
7063 put_task_struct(mt
);
7064 wait_for_completion(&req
.done
);
7065 tlb_migrate_finish(p
->mm
);
7069 task_rq_unlock(rq
, &flags
);
7073 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7076 * Move (not current) task off this cpu, onto dest cpu. We're doing
7077 * this because either it can't run here any more (set_cpus_allowed()
7078 * away from this CPU, or CPU going down), or because we're
7079 * attempting to rebalance this task on exec (sched_exec).
7081 * So we race with normal scheduler movements, but that's OK, as long
7082 * as the task is no longer on this CPU.
7084 * Returns non-zero if task was successfully migrated.
7086 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7088 struct rq
*rq_dest
, *rq_src
;
7091 if (unlikely(!cpu_active(dest_cpu
)))
7094 rq_src
= cpu_rq(src_cpu
);
7095 rq_dest
= cpu_rq(dest_cpu
);
7097 double_rq_lock(rq_src
, rq_dest
);
7098 /* Already moved. */
7099 if (task_cpu(p
) != src_cpu
)
7101 /* Affinity changed (again). */
7102 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7105 on_rq
= p
->se
.on_rq
;
7107 deactivate_task(rq_src
, p
, 0);
7109 set_task_cpu(p
, dest_cpu
);
7111 activate_task(rq_dest
, p
, 0);
7112 check_preempt_curr(rq_dest
, p
, 0);
7117 double_rq_unlock(rq_src
, rq_dest
);
7121 #define RCU_MIGRATION_IDLE 0
7122 #define RCU_MIGRATION_NEED_QS 1
7123 #define RCU_MIGRATION_GOT_QS 2
7124 #define RCU_MIGRATION_MUST_SYNC 3
7127 * migration_thread - this is a highprio system thread that performs
7128 * thread migration by bumping thread off CPU then 'pushing' onto
7131 static int migration_thread(void *data
)
7134 int cpu
= (long)data
;
7138 BUG_ON(rq
->migration_thread
!= current
);
7140 set_current_state(TASK_INTERRUPTIBLE
);
7141 while (!kthread_should_stop()) {
7142 struct migration_req
*req
;
7143 struct list_head
*head
;
7145 spin_lock_irq(&rq
->lock
);
7147 if (cpu_is_offline(cpu
)) {
7148 spin_unlock_irq(&rq
->lock
);
7152 if (rq
->active_balance
) {
7153 active_load_balance(rq
, cpu
);
7154 rq
->active_balance
= 0;
7157 head
= &rq
->migration_queue
;
7159 if (list_empty(head
)) {
7160 spin_unlock_irq(&rq
->lock
);
7162 set_current_state(TASK_INTERRUPTIBLE
);
7165 req
= list_entry(head
->next
, struct migration_req
, list
);
7166 list_del_init(head
->next
);
7168 if (req
->task
!= NULL
) {
7169 spin_unlock(&rq
->lock
);
7170 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7171 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7172 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7173 spin_unlock(&rq
->lock
);
7175 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7176 spin_unlock(&rq
->lock
);
7177 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7181 complete(&req
->done
);
7183 __set_current_state(TASK_RUNNING
);
7188 #ifdef CONFIG_HOTPLUG_CPU
7190 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7194 local_irq_disable();
7195 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7201 * Figure out where task on dead CPU should go, use force if necessary.
7203 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7206 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7209 /* Look for allowed, online CPU in same node. */
7210 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7211 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7214 /* Any allowed, online CPU? */
7215 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7216 if (dest_cpu
< nr_cpu_ids
)
7219 /* No more Mr. Nice Guy. */
7220 if (dest_cpu
>= nr_cpu_ids
) {
7221 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7222 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7225 * Don't tell them about moving exiting tasks or
7226 * kernel threads (both mm NULL), since they never
7229 if (p
->mm
&& printk_ratelimit()) {
7230 printk(KERN_INFO
"process %d (%s) no "
7231 "longer affine to cpu%d\n",
7232 task_pid_nr(p
), p
->comm
, dead_cpu
);
7237 /* It can have affinity changed while we were choosing. */
7238 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7243 * While a dead CPU has no uninterruptible tasks queued at this point,
7244 * it might still have a nonzero ->nr_uninterruptible counter, because
7245 * for performance reasons the counter is not stricly tracking tasks to
7246 * their home CPUs. So we just add the counter to another CPU's counter,
7247 * to keep the global sum constant after CPU-down:
7249 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7251 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7252 unsigned long flags
;
7254 local_irq_save(flags
);
7255 double_rq_lock(rq_src
, rq_dest
);
7256 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7257 rq_src
->nr_uninterruptible
= 0;
7258 double_rq_unlock(rq_src
, rq_dest
);
7259 local_irq_restore(flags
);
7262 /* Run through task list and migrate tasks from the dead cpu. */
7263 static void migrate_live_tasks(int src_cpu
)
7265 struct task_struct
*p
, *t
;
7267 read_lock(&tasklist_lock
);
7269 do_each_thread(t
, p
) {
7273 if (task_cpu(p
) == src_cpu
)
7274 move_task_off_dead_cpu(src_cpu
, p
);
7275 } while_each_thread(t
, p
);
7277 read_unlock(&tasklist_lock
);
7281 * Schedules idle task to be the next runnable task on current CPU.
7282 * It does so by boosting its priority to highest possible.
7283 * Used by CPU offline code.
7285 void sched_idle_next(void)
7287 int this_cpu
= smp_processor_id();
7288 struct rq
*rq
= cpu_rq(this_cpu
);
7289 struct task_struct
*p
= rq
->idle
;
7290 unsigned long flags
;
7292 /* cpu has to be offline */
7293 BUG_ON(cpu_online(this_cpu
));
7296 * Strictly not necessary since rest of the CPUs are stopped by now
7297 * and interrupts disabled on the current cpu.
7299 spin_lock_irqsave(&rq
->lock
, flags
);
7301 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7303 update_rq_clock(rq
);
7304 activate_task(rq
, p
, 0);
7306 spin_unlock_irqrestore(&rq
->lock
, flags
);
7310 * Ensures that the idle task is using init_mm right before its cpu goes
7313 void idle_task_exit(void)
7315 struct mm_struct
*mm
= current
->active_mm
;
7317 BUG_ON(cpu_online(smp_processor_id()));
7320 switch_mm(mm
, &init_mm
, current
);
7324 /* called under rq->lock with disabled interrupts */
7325 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7327 struct rq
*rq
= cpu_rq(dead_cpu
);
7329 /* Must be exiting, otherwise would be on tasklist. */
7330 BUG_ON(!p
->exit_state
);
7332 /* Cannot have done final schedule yet: would have vanished. */
7333 BUG_ON(p
->state
== TASK_DEAD
);
7338 * Drop lock around migration; if someone else moves it,
7339 * that's OK. No task can be added to this CPU, so iteration is
7342 spin_unlock_irq(&rq
->lock
);
7343 move_task_off_dead_cpu(dead_cpu
, p
);
7344 spin_lock_irq(&rq
->lock
);
7349 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7350 static void migrate_dead_tasks(unsigned int dead_cpu
)
7352 struct rq
*rq
= cpu_rq(dead_cpu
);
7353 struct task_struct
*next
;
7356 if (!rq
->nr_running
)
7358 update_rq_clock(rq
);
7359 next
= pick_next_task(rq
);
7362 next
->sched_class
->put_prev_task(rq
, next
);
7363 migrate_dead(dead_cpu
, next
);
7369 * remove the tasks which were accounted by rq from calc_load_tasks.
7371 static void calc_global_load_remove(struct rq
*rq
)
7373 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7374 rq
->calc_load_active
= 0;
7376 #endif /* CONFIG_HOTPLUG_CPU */
7378 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7380 static struct ctl_table sd_ctl_dir
[] = {
7382 .procname
= "sched_domain",
7388 static struct ctl_table sd_ctl_root
[] = {
7390 .ctl_name
= CTL_KERN
,
7391 .procname
= "kernel",
7393 .child
= sd_ctl_dir
,
7398 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7400 struct ctl_table
*entry
=
7401 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7406 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7408 struct ctl_table
*entry
;
7411 * In the intermediate directories, both the child directory and
7412 * procname are dynamically allocated and could fail but the mode
7413 * will always be set. In the lowest directory the names are
7414 * static strings and all have proc handlers.
7416 for (entry
= *tablep
; entry
->mode
; entry
++) {
7418 sd_free_ctl_entry(&entry
->child
);
7419 if (entry
->proc_handler
== NULL
)
7420 kfree(entry
->procname
);
7428 set_table_entry(struct ctl_table
*entry
,
7429 const char *procname
, void *data
, int maxlen
,
7430 mode_t mode
, proc_handler
*proc_handler
)
7432 entry
->procname
= procname
;
7434 entry
->maxlen
= maxlen
;
7436 entry
->proc_handler
= proc_handler
;
7439 static struct ctl_table
*
7440 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7442 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7447 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7448 sizeof(long), 0644, proc_doulongvec_minmax
);
7449 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7450 sizeof(long), 0644, proc_doulongvec_minmax
);
7451 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7452 sizeof(int), 0644, proc_dointvec_minmax
);
7453 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7454 sizeof(int), 0644, proc_dointvec_minmax
);
7455 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7456 sizeof(int), 0644, proc_dointvec_minmax
);
7457 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7458 sizeof(int), 0644, proc_dointvec_minmax
);
7459 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7460 sizeof(int), 0644, proc_dointvec_minmax
);
7461 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7462 sizeof(int), 0644, proc_dointvec_minmax
);
7463 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7464 sizeof(int), 0644, proc_dointvec_minmax
);
7465 set_table_entry(&table
[9], "cache_nice_tries",
7466 &sd
->cache_nice_tries
,
7467 sizeof(int), 0644, proc_dointvec_minmax
);
7468 set_table_entry(&table
[10], "flags", &sd
->flags
,
7469 sizeof(int), 0644, proc_dointvec_minmax
);
7470 set_table_entry(&table
[11], "name", sd
->name
,
7471 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7472 /* &table[12] is terminator */
7477 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7479 struct ctl_table
*entry
, *table
;
7480 struct sched_domain
*sd
;
7481 int domain_num
= 0, i
;
7484 for_each_domain(cpu
, sd
)
7486 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7491 for_each_domain(cpu
, sd
) {
7492 snprintf(buf
, 32, "domain%d", i
);
7493 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7495 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7502 static struct ctl_table_header
*sd_sysctl_header
;
7503 static void register_sched_domain_sysctl(void)
7505 int i
, cpu_num
= num_online_cpus();
7506 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7509 WARN_ON(sd_ctl_dir
[0].child
);
7510 sd_ctl_dir
[0].child
= entry
;
7515 for_each_online_cpu(i
) {
7516 snprintf(buf
, 32, "cpu%d", i
);
7517 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7519 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7523 WARN_ON(sd_sysctl_header
);
7524 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7527 /* may be called multiple times per register */
7528 static void unregister_sched_domain_sysctl(void)
7530 if (sd_sysctl_header
)
7531 unregister_sysctl_table(sd_sysctl_header
);
7532 sd_sysctl_header
= NULL
;
7533 if (sd_ctl_dir
[0].child
)
7534 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7537 static void register_sched_domain_sysctl(void)
7540 static void unregister_sched_domain_sysctl(void)
7545 static void set_rq_online(struct rq
*rq
)
7548 const struct sched_class
*class;
7550 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7553 for_each_class(class) {
7554 if (class->rq_online
)
7555 class->rq_online(rq
);
7560 static void set_rq_offline(struct rq
*rq
)
7563 const struct sched_class
*class;
7565 for_each_class(class) {
7566 if (class->rq_offline
)
7567 class->rq_offline(rq
);
7570 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7576 * migration_call - callback that gets triggered when a CPU is added.
7577 * Here we can start up the necessary migration thread for the new CPU.
7579 static int __cpuinit
7580 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7582 struct task_struct
*p
;
7583 int cpu
= (long)hcpu
;
7584 unsigned long flags
;
7589 case CPU_UP_PREPARE
:
7590 case CPU_UP_PREPARE_FROZEN
:
7591 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7594 kthread_bind(p
, cpu
);
7595 /* Must be high prio: stop_machine expects to yield to it. */
7596 rq
= task_rq_lock(p
, &flags
);
7597 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7598 task_rq_unlock(rq
, &flags
);
7600 cpu_rq(cpu
)->migration_thread
= p
;
7601 rq
->calc_load_update
= calc_load_update
;
7605 case CPU_ONLINE_FROZEN
:
7606 /* Strictly unnecessary, as first user will wake it. */
7607 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7609 /* Update our root-domain */
7611 spin_lock_irqsave(&rq
->lock
, flags
);
7613 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7617 spin_unlock_irqrestore(&rq
->lock
, flags
);
7620 #ifdef CONFIG_HOTPLUG_CPU
7621 case CPU_UP_CANCELED
:
7622 case CPU_UP_CANCELED_FROZEN
:
7623 if (!cpu_rq(cpu
)->migration_thread
)
7625 /* Unbind it from offline cpu so it can run. Fall thru. */
7626 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7627 cpumask_any(cpu_online_mask
));
7628 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7629 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7630 cpu_rq(cpu
)->migration_thread
= NULL
;
7634 case CPU_DEAD_FROZEN
:
7635 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7636 migrate_live_tasks(cpu
);
7638 kthread_stop(rq
->migration_thread
);
7639 put_task_struct(rq
->migration_thread
);
7640 rq
->migration_thread
= NULL
;
7641 /* Idle task back to normal (off runqueue, low prio) */
7642 spin_lock_irq(&rq
->lock
);
7643 update_rq_clock(rq
);
7644 deactivate_task(rq
, rq
->idle
, 0);
7645 rq
->idle
->static_prio
= MAX_PRIO
;
7646 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7647 rq
->idle
->sched_class
= &idle_sched_class
;
7648 migrate_dead_tasks(cpu
);
7649 spin_unlock_irq(&rq
->lock
);
7651 migrate_nr_uninterruptible(rq
);
7652 BUG_ON(rq
->nr_running
!= 0);
7653 calc_global_load_remove(rq
);
7655 * No need to migrate the tasks: it was best-effort if
7656 * they didn't take sched_hotcpu_mutex. Just wake up
7659 spin_lock_irq(&rq
->lock
);
7660 while (!list_empty(&rq
->migration_queue
)) {
7661 struct migration_req
*req
;
7663 req
= list_entry(rq
->migration_queue
.next
,
7664 struct migration_req
, list
);
7665 list_del_init(&req
->list
);
7666 spin_unlock_irq(&rq
->lock
);
7667 complete(&req
->done
);
7668 spin_lock_irq(&rq
->lock
);
7670 spin_unlock_irq(&rq
->lock
);
7674 case CPU_DYING_FROZEN
:
7675 /* Update our root-domain */
7677 spin_lock_irqsave(&rq
->lock
, flags
);
7679 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7682 spin_unlock_irqrestore(&rq
->lock
, flags
);
7690 * Register at high priority so that task migration (migrate_all_tasks)
7691 * happens before everything else. This has to be lower priority than
7692 * the notifier in the perf_counter subsystem, though.
7694 static struct notifier_block __cpuinitdata migration_notifier
= {
7695 .notifier_call
= migration_call
,
7699 static int __init
migration_init(void)
7701 void *cpu
= (void *)(long)smp_processor_id();
7704 /* Start one for the boot CPU: */
7705 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7706 BUG_ON(err
== NOTIFY_BAD
);
7707 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7708 register_cpu_notifier(&migration_notifier
);
7712 early_initcall(migration_init
);
7717 #ifdef CONFIG_SCHED_DEBUG
7719 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7720 struct cpumask
*groupmask
)
7722 struct sched_group
*group
= sd
->groups
;
7725 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7726 cpumask_clear(groupmask
);
7728 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7730 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7731 printk("does not load-balance\n");
7733 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7738 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7740 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7741 printk(KERN_ERR
"ERROR: domain->span does not contain "
7744 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7745 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7749 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7753 printk(KERN_ERR
"ERROR: group is NULL\n");
7757 if (!group
->cpu_power
) {
7758 printk(KERN_CONT
"\n");
7759 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7764 if (!cpumask_weight(sched_group_cpus(group
))) {
7765 printk(KERN_CONT
"\n");
7766 printk(KERN_ERR
"ERROR: empty group\n");
7770 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7771 printk(KERN_CONT
"\n");
7772 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7776 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7778 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7780 printk(KERN_CONT
" %s", str
);
7781 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7782 printk(KERN_CONT
" (cpu_power = %d)",
7786 group
= group
->next
;
7787 } while (group
!= sd
->groups
);
7788 printk(KERN_CONT
"\n");
7790 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7791 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7794 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7795 printk(KERN_ERR
"ERROR: parent span is not a superset "
7796 "of domain->span\n");
7800 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7802 cpumask_var_t groupmask
;
7806 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7810 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7812 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7813 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7818 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7825 free_cpumask_var(groupmask
);
7827 #else /* !CONFIG_SCHED_DEBUG */
7828 # define sched_domain_debug(sd, cpu) do { } while (0)
7829 #endif /* CONFIG_SCHED_DEBUG */
7831 static int sd_degenerate(struct sched_domain
*sd
)
7833 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7836 /* Following flags need at least 2 groups */
7837 if (sd
->flags
& (SD_LOAD_BALANCE
|
7838 SD_BALANCE_NEWIDLE
|
7842 SD_SHARE_PKG_RESOURCES
)) {
7843 if (sd
->groups
!= sd
->groups
->next
)
7847 /* Following flags don't use groups */
7848 if (sd
->flags
& (SD_WAKE_AFFINE
))
7855 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7857 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7859 if (sd_degenerate(parent
))
7862 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7865 /* Flags needing groups don't count if only 1 group in parent */
7866 if (parent
->groups
== parent
->groups
->next
) {
7867 pflags
&= ~(SD_LOAD_BALANCE
|
7868 SD_BALANCE_NEWIDLE
|
7872 SD_SHARE_PKG_RESOURCES
);
7873 if (nr_node_ids
== 1)
7874 pflags
&= ~SD_SERIALIZE
;
7876 if (~cflags
& pflags
)
7882 static void free_rootdomain(struct root_domain
*rd
)
7884 cpupri_cleanup(&rd
->cpupri
);
7886 free_cpumask_var(rd
->rto_mask
);
7887 free_cpumask_var(rd
->online
);
7888 free_cpumask_var(rd
->span
);
7892 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7894 struct root_domain
*old_rd
= NULL
;
7895 unsigned long flags
;
7897 spin_lock_irqsave(&rq
->lock
, flags
);
7902 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7905 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7908 * If we dont want to free the old_rt yet then
7909 * set old_rd to NULL to skip the freeing later
7912 if (!atomic_dec_and_test(&old_rd
->refcount
))
7916 atomic_inc(&rd
->refcount
);
7919 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7920 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
7923 spin_unlock_irqrestore(&rq
->lock
, flags
);
7926 free_rootdomain(old_rd
);
7929 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7931 gfp_t gfp
= GFP_KERNEL
;
7933 memset(rd
, 0, sizeof(*rd
));
7938 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7940 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7942 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7945 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7950 free_cpumask_var(rd
->rto_mask
);
7952 free_cpumask_var(rd
->online
);
7954 free_cpumask_var(rd
->span
);
7959 static void init_defrootdomain(void)
7961 init_rootdomain(&def_root_domain
, true);
7963 atomic_set(&def_root_domain
.refcount
, 1);
7966 static struct root_domain
*alloc_rootdomain(void)
7968 struct root_domain
*rd
;
7970 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7974 if (init_rootdomain(rd
, false) != 0) {
7983 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7984 * hold the hotplug lock.
7987 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7989 struct rq
*rq
= cpu_rq(cpu
);
7990 struct sched_domain
*tmp
;
7992 /* Remove the sched domains which do not contribute to scheduling. */
7993 for (tmp
= sd
; tmp
; ) {
7994 struct sched_domain
*parent
= tmp
->parent
;
7998 if (sd_parent_degenerate(tmp
, parent
)) {
7999 tmp
->parent
= parent
->parent
;
8001 parent
->parent
->child
= tmp
;
8006 if (sd
&& sd_degenerate(sd
)) {
8012 sched_domain_debug(sd
, cpu
);
8014 rq_attach_root(rq
, rd
);
8015 rcu_assign_pointer(rq
->sd
, sd
);
8018 /* cpus with isolated domains */
8019 static cpumask_var_t cpu_isolated_map
;
8021 /* Setup the mask of cpus configured for isolated domains */
8022 static int __init
isolated_cpu_setup(char *str
)
8024 cpulist_parse(str
, cpu_isolated_map
);
8028 __setup("isolcpus=", isolated_cpu_setup
);
8031 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8032 * to a function which identifies what group(along with sched group) a CPU
8033 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8034 * (due to the fact that we keep track of groups covered with a struct cpumask).
8036 * init_sched_build_groups will build a circular linked list of the groups
8037 * covered by the given span, and will set each group's ->cpumask correctly,
8038 * and ->cpu_power to 0.
8041 init_sched_build_groups(const struct cpumask
*span
,
8042 const struct cpumask
*cpu_map
,
8043 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8044 struct sched_group
**sg
,
8045 struct cpumask
*tmpmask
),
8046 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8048 struct sched_group
*first
= NULL
, *last
= NULL
;
8051 cpumask_clear(covered
);
8053 for_each_cpu(i
, span
) {
8054 struct sched_group
*sg
;
8055 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8058 if (cpumask_test_cpu(i
, covered
))
8061 cpumask_clear(sched_group_cpus(sg
));
8064 for_each_cpu(j
, span
) {
8065 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8068 cpumask_set_cpu(j
, covered
);
8069 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8080 #define SD_NODES_PER_DOMAIN 16
8085 * find_next_best_node - find the next node to include in a sched_domain
8086 * @node: node whose sched_domain we're building
8087 * @used_nodes: nodes already in the sched_domain
8089 * Find the next node to include in a given scheduling domain. Simply
8090 * finds the closest node not already in the @used_nodes map.
8092 * Should use nodemask_t.
8094 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8096 int i
, n
, val
, min_val
, best_node
= 0;
8100 for (i
= 0; i
< nr_node_ids
; i
++) {
8101 /* Start at @node */
8102 n
= (node
+ i
) % nr_node_ids
;
8104 if (!nr_cpus_node(n
))
8107 /* Skip already used nodes */
8108 if (node_isset(n
, *used_nodes
))
8111 /* Simple min distance search */
8112 val
= node_distance(node
, n
);
8114 if (val
< min_val
) {
8120 node_set(best_node
, *used_nodes
);
8125 * sched_domain_node_span - get a cpumask for a node's sched_domain
8126 * @node: node whose cpumask we're constructing
8127 * @span: resulting cpumask
8129 * Given a node, construct a good cpumask for its sched_domain to span. It
8130 * should be one that prevents unnecessary balancing, but also spreads tasks
8133 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8135 nodemask_t used_nodes
;
8138 cpumask_clear(span
);
8139 nodes_clear(used_nodes
);
8141 cpumask_or(span
, span
, cpumask_of_node(node
));
8142 node_set(node
, used_nodes
);
8144 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8145 int next_node
= find_next_best_node(node
, &used_nodes
);
8147 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8150 #endif /* CONFIG_NUMA */
8152 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8155 * The cpus mask in sched_group and sched_domain hangs off the end.
8157 * ( See the the comments in include/linux/sched.h:struct sched_group
8158 * and struct sched_domain. )
8160 struct static_sched_group
{
8161 struct sched_group sg
;
8162 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8165 struct static_sched_domain
{
8166 struct sched_domain sd
;
8167 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8173 cpumask_var_t domainspan
;
8174 cpumask_var_t covered
;
8175 cpumask_var_t notcovered
;
8177 cpumask_var_t nodemask
;
8178 cpumask_var_t this_sibling_map
;
8179 cpumask_var_t this_core_map
;
8180 cpumask_var_t send_covered
;
8181 cpumask_var_t tmpmask
;
8182 struct sched_group
**sched_group_nodes
;
8183 struct root_domain
*rd
;
8187 sa_sched_groups
= 0,
8192 sa_this_sibling_map
,
8194 sa_sched_group_nodes
,
8204 * SMT sched-domains:
8206 #ifdef CONFIG_SCHED_SMT
8207 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8208 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8211 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8212 struct sched_group
**sg
, struct cpumask
*unused
)
8215 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8218 #endif /* CONFIG_SCHED_SMT */
8221 * multi-core sched-domains:
8223 #ifdef CONFIG_SCHED_MC
8224 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8225 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8226 #endif /* CONFIG_SCHED_MC */
8228 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8230 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8231 struct sched_group
**sg
, struct cpumask
*mask
)
8235 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8236 group
= cpumask_first(mask
);
8238 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8241 #elif defined(CONFIG_SCHED_MC)
8243 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8244 struct sched_group
**sg
, struct cpumask
*unused
)
8247 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8252 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8253 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8256 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8257 struct sched_group
**sg
, struct cpumask
*mask
)
8260 #ifdef CONFIG_SCHED_MC
8261 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8262 group
= cpumask_first(mask
);
8263 #elif defined(CONFIG_SCHED_SMT)
8264 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8265 group
= cpumask_first(mask
);
8270 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8276 * The init_sched_build_groups can't handle what we want to do with node
8277 * groups, so roll our own. Now each node has its own list of groups which
8278 * gets dynamically allocated.
8280 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8281 static struct sched_group
***sched_group_nodes_bycpu
;
8283 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8284 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8286 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8287 struct sched_group
**sg
,
8288 struct cpumask
*nodemask
)
8292 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8293 group
= cpumask_first(nodemask
);
8296 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8300 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8302 struct sched_group
*sg
= group_head
;
8308 for_each_cpu(j
, sched_group_cpus(sg
)) {
8309 struct sched_domain
*sd
;
8311 sd
= &per_cpu(phys_domains
, j
).sd
;
8312 if (j
!= group_first_cpu(sd
->groups
)) {
8314 * Only add "power" once for each
8320 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8323 } while (sg
!= group_head
);
8326 static int build_numa_sched_groups(struct s_data
*d
,
8327 const struct cpumask
*cpu_map
, int num
)
8329 struct sched_domain
*sd
;
8330 struct sched_group
*sg
, *prev
;
8333 cpumask_clear(d
->covered
);
8334 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8335 if (cpumask_empty(d
->nodemask
)) {
8336 d
->sched_group_nodes
[num
] = NULL
;
8340 sched_domain_node_span(num
, d
->domainspan
);
8341 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8343 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8346 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8350 d
->sched_group_nodes
[num
] = sg
;
8352 for_each_cpu(j
, d
->nodemask
) {
8353 sd
= &per_cpu(node_domains
, j
).sd
;
8358 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8360 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8363 for (j
= 0; j
< nr_node_ids
; j
++) {
8364 n
= (num
+ j
) % nr_node_ids
;
8365 cpumask_complement(d
->notcovered
, d
->covered
);
8366 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8367 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8368 if (cpumask_empty(d
->tmpmask
))
8370 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8371 if (cpumask_empty(d
->tmpmask
))
8373 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8377 "Can not alloc domain group for node %d\n", j
);
8381 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8382 sg
->next
= prev
->next
;
8383 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8390 #endif /* CONFIG_NUMA */
8393 /* Free memory allocated for various sched_group structures */
8394 static void free_sched_groups(const struct cpumask
*cpu_map
,
8395 struct cpumask
*nodemask
)
8399 for_each_cpu(cpu
, cpu_map
) {
8400 struct sched_group
**sched_group_nodes
8401 = sched_group_nodes_bycpu
[cpu
];
8403 if (!sched_group_nodes
)
8406 for (i
= 0; i
< nr_node_ids
; i
++) {
8407 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8409 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8410 if (cpumask_empty(nodemask
))
8420 if (oldsg
!= sched_group_nodes
[i
])
8423 kfree(sched_group_nodes
);
8424 sched_group_nodes_bycpu
[cpu
] = NULL
;
8427 #else /* !CONFIG_NUMA */
8428 static void free_sched_groups(const struct cpumask
*cpu_map
,
8429 struct cpumask
*nodemask
)
8432 #endif /* CONFIG_NUMA */
8435 * Initialize sched groups cpu_power.
8437 * cpu_power indicates the capacity of sched group, which is used while
8438 * distributing the load between different sched groups in a sched domain.
8439 * Typically cpu_power for all the groups in a sched domain will be same unless
8440 * there are asymmetries in the topology. If there are asymmetries, group
8441 * having more cpu_power will pickup more load compared to the group having
8444 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8446 struct sched_domain
*child
;
8447 struct sched_group
*group
;
8451 WARN_ON(!sd
|| !sd
->groups
);
8453 if (cpu
!= group_first_cpu(sd
->groups
))
8458 sd
->groups
->cpu_power
= 0;
8461 power
= SCHED_LOAD_SCALE
;
8462 weight
= cpumask_weight(sched_domain_span(sd
));
8464 * SMT siblings share the power of a single core.
8465 * Usually multiple threads get a better yield out of
8466 * that one core than a single thread would have,
8467 * reflect that in sd->smt_gain.
8469 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8470 power
*= sd
->smt_gain
;
8472 power
>>= SCHED_LOAD_SHIFT
;
8474 sd
->groups
->cpu_power
+= power
;
8479 * Add cpu_power of each child group to this groups cpu_power.
8481 group
= child
->groups
;
8483 sd
->groups
->cpu_power
+= group
->cpu_power
;
8484 group
= group
->next
;
8485 } while (group
!= child
->groups
);
8489 * Initializers for schedule domains
8490 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8493 #ifdef CONFIG_SCHED_DEBUG
8494 # define SD_INIT_NAME(sd, type) sd->name = #type
8496 # define SD_INIT_NAME(sd, type) do { } while (0)
8499 #define SD_INIT(sd, type) sd_init_##type(sd)
8501 #define SD_INIT_FUNC(type) \
8502 static noinline void sd_init_##type(struct sched_domain *sd) \
8504 memset(sd, 0, sizeof(*sd)); \
8505 *sd = SD_##type##_INIT; \
8506 sd->level = SD_LV_##type; \
8507 SD_INIT_NAME(sd, type); \
8512 SD_INIT_FUNC(ALLNODES
)
8515 #ifdef CONFIG_SCHED_SMT
8516 SD_INIT_FUNC(SIBLING
)
8518 #ifdef CONFIG_SCHED_MC
8522 static int default_relax_domain_level
= -1;
8524 static int __init
setup_relax_domain_level(char *str
)
8528 val
= simple_strtoul(str
, NULL
, 0);
8529 if (val
< SD_LV_MAX
)
8530 default_relax_domain_level
= val
;
8534 __setup("relax_domain_level=", setup_relax_domain_level
);
8536 static void set_domain_attribute(struct sched_domain
*sd
,
8537 struct sched_domain_attr
*attr
)
8541 if (!attr
|| attr
->relax_domain_level
< 0) {
8542 if (default_relax_domain_level
< 0)
8545 request
= default_relax_domain_level
;
8547 request
= attr
->relax_domain_level
;
8548 if (request
< sd
->level
) {
8549 /* turn off idle balance on this domain */
8550 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8552 /* turn on idle balance on this domain */
8553 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8557 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8558 const struct cpumask
*cpu_map
)
8561 case sa_sched_groups
:
8562 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8563 d
->sched_group_nodes
= NULL
;
8565 free_rootdomain(d
->rd
); /* fall through */
8567 free_cpumask_var(d
->tmpmask
); /* fall through */
8568 case sa_send_covered
:
8569 free_cpumask_var(d
->send_covered
); /* fall through */
8570 case sa_this_core_map
:
8571 free_cpumask_var(d
->this_core_map
); /* fall through */
8572 case sa_this_sibling_map
:
8573 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8575 free_cpumask_var(d
->nodemask
); /* fall through */
8576 case sa_sched_group_nodes
:
8578 kfree(d
->sched_group_nodes
); /* fall through */
8580 free_cpumask_var(d
->notcovered
); /* fall through */
8582 free_cpumask_var(d
->covered
); /* fall through */
8584 free_cpumask_var(d
->domainspan
); /* fall through */
8591 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8592 const struct cpumask
*cpu_map
)
8595 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8597 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8598 return sa_domainspan
;
8599 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8601 /* Allocate the per-node list of sched groups */
8602 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8603 sizeof(struct sched_group
*), GFP_KERNEL
);
8604 if (!d
->sched_group_nodes
) {
8605 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8606 return sa_notcovered
;
8608 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8610 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8611 return sa_sched_group_nodes
;
8612 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8614 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8615 return sa_this_sibling_map
;
8616 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8617 return sa_this_core_map
;
8618 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8619 return sa_send_covered
;
8620 d
->rd
= alloc_rootdomain();
8622 printk(KERN_WARNING
"Cannot alloc root domain\n");
8625 return sa_rootdomain
;
8628 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8629 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8631 struct sched_domain
*sd
= NULL
;
8633 struct sched_domain
*parent
;
8636 if (cpumask_weight(cpu_map
) >
8637 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8638 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8639 SD_INIT(sd
, ALLNODES
);
8640 set_domain_attribute(sd
, attr
);
8641 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8642 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8647 sd
= &per_cpu(node_domains
, i
).sd
;
8649 set_domain_attribute(sd
, attr
);
8650 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8651 sd
->parent
= parent
;
8654 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8659 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8660 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8661 struct sched_domain
*parent
, int i
)
8663 struct sched_domain
*sd
;
8664 sd
= &per_cpu(phys_domains
, i
).sd
;
8666 set_domain_attribute(sd
, attr
);
8667 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8668 sd
->parent
= parent
;
8671 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8675 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8676 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8677 struct sched_domain
*parent
, int i
)
8679 struct sched_domain
*sd
= parent
;
8680 #ifdef CONFIG_SCHED_MC
8681 sd
= &per_cpu(core_domains
, i
).sd
;
8683 set_domain_attribute(sd
, attr
);
8684 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8685 sd
->parent
= parent
;
8687 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8692 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8693 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8694 struct sched_domain
*parent
, int i
)
8696 struct sched_domain
*sd
= parent
;
8697 #ifdef CONFIG_SCHED_SMT
8698 sd
= &per_cpu(cpu_domains
, i
).sd
;
8699 SD_INIT(sd
, SIBLING
);
8700 set_domain_attribute(sd
, attr
);
8701 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8702 sd
->parent
= parent
;
8704 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8709 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8710 const struct cpumask
*cpu_map
, int cpu
)
8713 #ifdef CONFIG_SCHED_SMT
8714 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8715 cpumask_and(d
->this_sibling_map
, cpu_map
,
8716 topology_thread_cpumask(cpu
));
8717 if (cpu
== cpumask_first(d
->this_sibling_map
))
8718 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8720 d
->send_covered
, d
->tmpmask
);
8723 #ifdef CONFIG_SCHED_MC
8724 case SD_LV_MC
: /* set up multi-core groups */
8725 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8726 if (cpu
== cpumask_first(d
->this_core_map
))
8727 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8729 d
->send_covered
, d
->tmpmask
);
8732 case SD_LV_CPU
: /* set up physical groups */
8733 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8734 if (!cpumask_empty(d
->nodemask
))
8735 init_sched_build_groups(d
->nodemask
, cpu_map
,
8737 d
->send_covered
, d
->tmpmask
);
8740 case SD_LV_ALLNODES
:
8741 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8742 d
->send_covered
, d
->tmpmask
);
8751 * Build sched domains for a given set of cpus and attach the sched domains
8752 * to the individual cpus
8754 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8755 struct sched_domain_attr
*attr
)
8757 enum s_alloc alloc_state
= sa_none
;
8759 struct sched_domain
*sd
;
8765 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8766 if (alloc_state
!= sa_rootdomain
)
8768 alloc_state
= sa_sched_groups
;
8771 * Set up domains for cpus specified by the cpu_map.
8773 for_each_cpu(i
, cpu_map
) {
8774 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8777 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8778 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8779 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8780 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8783 for_each_cpu(i
, cpu_map
) {
8784 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8785 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8788 /* Set up physical groups */
8789 for (i
= 0; i
< nr_node_ids
; i
++)
8790 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8793 /* Set up node groups */
8795 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8797 for (i
= 0; i
< nr_node_ids
; i
++)
8798 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8802 /* Calculate CPU power for physical packages and nodes */
8803 #ifdef CONFIG_SCHED_SMT
8804 for_each_cpu(i
, cpu_map
) {
8805 sd
= &per_cpu(cpu_domains
, i
).sd
;
8806 init_sched_groups_power(i
, sd
);
8809 #ifdef CONFIG_SCHED_MC
8810 for_each_cpu(i
, cpu_map
) {
8811 sd
= &per_cpu(core_domains
, i
).sd
;
8812 init_sched_groups_power(i
, sd
);
8816 for_each_cpu(i
, cpu_map
) {
8817 sd
= &per_cpu(phys_domains
, i
).sd
;
8818 init_sched_groups_power(i
, sd
);
8822 for (i
= 0; i
< nr_node_ids
; i
++)
8823 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8825 if (d
.sd_allnodes
) {
8826 struct sched_group
*sg
;
8828 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8830 init_numa_sched_groups_power(sg
);
8834 /* Attach the domains */
8835 for_each_cpu(i
, cpu_map
) {
8836 #ifdef CONFIG_SCHED_SMT
8837 sd
= &per_cpu(cpu_domains
, i
).sd
;
8838 #elif defined(CONFIG_SCHED_MC)
8839 sd
= &per_cpu(core_domains
, i
).sd
;
8841 sd
= &per_cpu(phys_domains
, i
).sd
;
8843 cpu_attach_domain(sd
, d
.rd
, i
);
8846 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8847 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8851 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8855 static int build_sched_domains(const struct cpumask
*cpu_map
)
8857 return __build_sched_domains(cpu_map
, NULL
);
8860 static struct cpumask
*doms_cur
; /* current sched domains */
8861 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8862 static struct sched_domain_attr
*dattr_cur
;
8863 /* attribues of custom domains in 'doms_cur' */
8866 * Special case: If a kmalloc of a doms_cur partition (array of
8867 * cpumask) fails, then fallback to a single sched domain,
8868 * as determined by the single cpumask fallback_doms.
8870 static cpumask_var_t fallback_doms
;
8873 * arch_update_cpu_topology lets virtualized architectures update the
8874 * cpu core maps. It is supposed to return 1 if the topology changed
8875 * or 0 if it stayed the same.
8877 int __attribute__((weak
)) arch_update_cpu_topology(void)
8883 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8884 * For now this just excludes isolated cpus, but could be used to
8885 * exclude other special cases in the future.
8887 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8891 arch_update_cpu_topology();
8893 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8895 doms_cur
= fallback_doms
;
8896 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8898 err
= build_sched_domains(doms_cur
);
8899 register_sched_domain_sysctl();
8904 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8905 struct cpumask
*tmpmask
)
8907 free_sched_groups(cpu_map
, tmpmask
);
8911 * Detach sched domains from a group of cpus specified in cpu_map
8912 * These cpus will now be attached to the NULL domain
8914 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8916 /* Save because hotplug lock held. */
8917 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8920 for_each_cpu(i
, cpu_map
)
8921 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8922 synchronize_sched();
8923 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8926 /* handle null as "default" */
8927 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8928 struct sched_domain_attr
*new, int idx_new
)
8930 struct sched_domain_attr tmp
;
8937 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8938 new ? (new + idx_new
) : &tmp
,
8939 sizeof(struct sched_domain_attr
));
8943 * Partition sched domains as specified by the 'ndoms_new'
8944 * cpumasks in the array doms_new[] of cpumasks. This compares
8945 * doms_new[] to the current sched domain partitioning, doms_cur[].
8946 * It destroys each deleted domain and builds each new domain.
8948 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8949 * The masks don't intersect (don't overlap.) We should setup one
8950 * sched domain for each mask. CPUs not in any of the cpumasks will
8951 * not be load balanced. If the same cpumask appears both in the
8952 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8955 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8956 * ownership of it and will kfree it when done with it. If the caller
8957 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8958 * ndoms_new == 1, and partition_sched_domains() will fallback to
8959 * the single partition 'fallback_doms', it also forces the domains
8962 * If doms_new == NULL it will be replaced with cpu_online_mask.
8963 * ndoms_new == 0 is a special case for destroying existing domains,
8964 * and it will not create the default domain.
8966 * Call with hotplug lock held
8968 /* FIXME: Change to struct cpumask *doms_new[] */
8969 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8970 struct sched_domain_attr
*dattr_new
)
8975 mutex_lock(&sched_domains_mutex
);
8977 /* always unregister in case we don't destroy any domains */
8978 unregister_sched_domain_sysctl();
8980 /* Let architecture update cpu core mappings. */
8981 new_topology
= arch_update_cpu_topology();
8983 n
= doms_new
? ndoms_new
: 0;
8985 /* Destroy deleted domains */
8986 for (i
= 0; i
< ndoms_cur
; i
++) {
8987 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8988 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8989 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8992 /* no match - a current sched domain not in new doms_new[] */
8993 detach_destroy_domains(doms_cur
+ i
);
8998 if (doms_new
== NULL
) {
9000 doms_new
= fallback_doms
;
9001 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
9002 WARN_ON_ONCE(dattr_new
);
9005 /* Build new domains */
9006 for (i
= 0; i
< ndoms_new
; i
++) {
9007 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9008 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9009 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9012 /* no match - add a new doms_new */
9013 __build_sched_domains(doms_new
+ i
,
9014 dattr_new
? dattr_new
+ i
: NULL
);
9019 /* Remember the new sched domains */
9020 if (doms_cur
!= fallback_doms
)
9022 kfree(dattr_cur
); /* kfree(NULL) is safe */
9023 doms_cur
= doms_new
;
9024 dattr_cur
= dattr_new
;
9025 ndoms_cur
= ndoms_new
;
9027 register_sched_domain_sysctl();
9029 mutex_unlock(&sched_domains_mutex
);
9032 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9033 static void arch_reinit_sched_domains(void)
9037 /* Destroy domains first to force the rebuild */
9038 partition_sched_domains(0, NULL
, NULL
);
9040 rebuild_sched_domains();
9044 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9046 unsigned int level
= 0;
9048 if (sscanf(buf
, "%u", &level
) != 1)
9052 * level is always be positive so don't check for
9053 * level < POWERSAVINGS_BALANCE_NONE which is 0
9054 * What happens on 0 or 1 byte write,
9055 * need to check for count as well?
9058 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9062 sched_smt_power_savings
= level
;
9064 sched_mc_power_savings
= level
;
9066 arch_reinit_sched_domains();
9071 #ifdef CONFIG_SCHED_MC
9072 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9075 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9077 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9078 const char *buf
, size_t count
)
9080 return sched_power_savings_store(buf
, count
, 0);
9082 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9083 sched_mc_power_savings_show
,
9084 sched_mc_power_savings_store
);
9087 #ifdef CONFIG_SCHED_SMT
9088 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9091 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9093 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9094 const char *buf
, size_t count
)
9096 return sched_power_savings_store(buf
, count
, 1);
9098 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9099 sched_smt_power_savings_show
,
9100 sched_smt_power_savings_store
);
9103 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9107 #ifdef CONFIG_SCHED_SMT
9109 err
= sysfs_create_file(&cls
->kset
.kobj
,
9110 &attr_sched_smt_power_savings
.attr
);
9112 #ifdef CONFIG_SCHED_MC
9113 if (!err
&& mc_capable())
9114 err
= sysfs_create_file(&cls
->kset
.kobj
,
9115 &attr_sched_mc_power_savings
.attr
);
9119 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9121 #ifndef CONFIG_CPUSETS
9123 * Add online and remove offline CPUs from the scheduler domains.
9124 * When cpusets are enabled they take over this function.
9126 static int update_sched_domains(struct notifier_block
*nfb
,
9127 unsigned long action
, void *hcpu
)
9131 case CPU_ONLINE_FROZEN
:
9133 case CPU_DEAD_FROZEN
:
9134 partition_sched_domains(1, NULL
, NULL
);
9143 static int update_runtime(struct notifier_block
*nfb
,
9144 unsigned long action
, void *hcpu
)
9146 int cpu
= (int)(long)hcpu
;
9149 case CPU_DOWN_PREPARE
:
9150 case CPU_DOWN_PREPARE_FROZEN
:
9151 disable_runtime(cpu_rq(cpu
));
9154 case CPU_DOWN_FAILED
:
9155 case CPU_DOWN_FAILED_FROZEN
:
9157 case CPU_ONLINE_FROZEN
:
9158 enable_runtime(cpu_rq(cpu
));
9166 void __init
sched_init_smp(void)
9168 cpumask_var_t non_isolated_cpus
;
9170 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9172 #if defined(CONFIG_NUMA)
9173 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9175 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9178 mutex_lock(&sched_domains_mutex
);
9179 arch_init_sched_domains(cpu_online_mask
);
9180 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9181 if (cpumask_empty(non_isolated_cpus
))
9182 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9183 mutex_unlock(&sched_domains_mutex
);
9186 #ifndef CONFIG_CPUSETS
9187 /* XXX: Theoretical race here - CPU may be hotplugged now */
9188 hotcpu_notifier(update_sched_domains
, 0);
9191 /* RT runtime code needs to handle some hotplug events */
9192 hotcpu_notifier(update_runtime
, 0);
9196 /* Move init over to a non-isolated CPU */
9197 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9199 sched_init_granularity();
9200 free_cpumask_var(non_isolated_cpus
);
9202 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9203 init_sched_rt_class();
9206 void __init
sched_init_smp(void)
9208 sched_init_granularity();
9210 #endif /* CONFIG_SMP */
9212 const_debug
unsigned int sysctl_timer_migration
= 1;
9214 int in_sched_functions(unsigned long addr
)
9216 return in_lock_functions(addr
) ||
9217 (addr
>= (unsigned long)__sched_text_start
9218 && addr
< (unsigned long)__sched_text_end
);
9221 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9223 cfs_rq
->tasks_timeline
= RB_ROOT
;
9224 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9225 #ifdef CONFIG_FAIR_GROUP_SCHED
9228 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9231 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9233 struct rt_prio_array
*array
;
9236 array
= &rt_rq
->active
;
9237 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9238 INIT_LIST_HEAD(array
->queue
+ i
);
9239 __clear_bit(i
, array
->bitmap
);
9241 /* delimiter for bitsearch: */
9242 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9244 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9245 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9247 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9251 rt_rq
->rt_nr_migratory
= 0;
9252 rt_rq
->overloaded
= 0;
9253 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9257 rt_rq
->rt_throttled
= 0;
9258 rt_rq
->rt_runtime
= 0;
9259 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9261 #ifdef CONFIG_RT_GROUP_SCHED
9262 rt_rq
->rt_nr_boosted
= 0;
9267 #ifdef CONFIG_FAIR_GROUP_SCHED
9268 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9269 struct sched_entity
*se
, int cpu
, int add
,
9270 struct sched_entity
*parent
)
9272 struct rq
*rq
= cpu_rq(cpu
);
9273 tg
->cfs_rq
[cpu
] = cfs_rq
;
9274 init_cfs_rq(cfs_rq
, rq
);
9277 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9280 /* se could be NULL for init_task_group */
9285 se
->cfs_rq
= &rq
->cfs
;
9287 se
->cfs_rq
= parent
->my_q
;
9290 se
->load
.weight
= tg
->shares
;
9291 se
->load
.inv_weight
= 0;
9292 se
->parent
= parent
;
9296 #ifdef CONFIG_RT_GROUP_SCHED
9297 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9298 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9299 struct sched_rt_entity
*parent
)
9301 struct rq
*rq
= cpu_rq(cpu
);
9303 tg
->rt_rq
[cpu
] = rt_rq
;
9304 init_rt_rq(rt_rq
, rq
);
9306 rt_rq
->rt_se
= rt_se
;
9307 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9309 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9311 tg
->rt_se
[cpu
] = rt_se
;
9316 rt_se
->rt_rq
= &rq
->rt
;
9318 rt_se
->rt_rq
= parent
->my_q
;
9320 rt_se
->my_q
= rt_rq
;
9321 rt_se
->parent
= parent
;
9322 INIT_LIST_HEAD(&rt_se
->run_list
);
9326 void __init
sched_init(void)
9329 unsigned long alloc_size
= 0, ptr
;
9331 #ifdef CONFIG_FAIR_GROUP_SCHED
9332 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9334 #ifdef CONFIG_RT_GROUP_SCHED
9335 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9337 #ifdef CONFIG_USER_SCHED
9340 #ifdef CONFIG_CPUMASK_OFFSTACK
9341 alloc_size
+= num_possible_cpus() * cpumask_size();
9344 * As sched_init() is called before page_alloc is setup,
9345 * we use alloc_bootmem().
9348 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9350 #ifdef CONFIG_FAIR_GROUP_SCHED
9351 init_task_group
.se
= (struct sched_entity
**)ptr
;
9352 ptr
+= nr_cpu_ids
* sizeof(void **);
9354 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9355 ptr
+= nr_cpu_ids
* sizeof(void **);
9357 #ifdef CONFIG_USER_SCHED
9358 root_task_group
.se
= (struct sched_entity
**)ptr
;
9359 ptr
+= nr_cpu_ids
* sizeof(void **);
9361 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9362 ptr
+= nr_cpu_ids
* sizeof(void **);
9363 #endif /* CONFIG_USER_SCHED */
9364 #endif /* CONFIG_FAIR_GROUP_SCHED */
9365 #ifdef CONFIG_RT_GROUP_SCHED
9366 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9367 ptr
+= nr_cpu_ids
* sizeof(void **);
9369 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9370 ptr
+= nr_cpu_ids
* sizeof(void **);
9372 #ifdef CONFIG_USER_SCHED
9373 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9374 ptr
+= nr_cpu_ids
* sizeof(void **);
9376 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9377 ptr
+= nr_cpu_ids
* sizeof(void **);
9378 #endif /* CONFIG_USER_SCHED */
9379 #endif /* CONFIG_RT_GROUP_SCHED */
9380 #ifdef CONFIG_CPUMASK_OFFSTACK
9381 for_each_possible_cpu(i
) {
9382 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9383 ptr
+= cpumask_size();
9385 #endif /* CONFIG_CPUMASK_OFFSTACK */
9389 init_defrootdomain();
9392 init_rt_bandwidth(&def_rt_bandwidth
,
9393 global_rt_period(), global_rt_runtime());
9395 #ifdef CONFIG_RT_GROUP_SCHED
9396 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9397 global_rt_period(), global_rt_runtime());
9398 #ifdef CONFIG_USER_SCHED
9399 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9400 global_rt_period(), RUNTIME_INF
);
9401 #endif /* CONFIG_USER_SCHED */
9402 #endif /* CONFIG_RT_GROUP_SCHED */
9404 #ifdef CONFIG_GROUP_SCHED
9405 list_add(&init_task_group
.list
, &task_groups
);
9406 INIT_LIST_HEAD(&init_task_group
.children
);
9408 #ifdef CONFIG_USER_SCHED
9409 INIT_LIST_HEAD(&root_task_group
.children
);
9410 init_task_group
.parent
= &root_task_group
;
9411 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9412 #endif /* CONFIG_USER_SCHED */
9413 #endif /* CONFIG_GROUP_SCHED */
9415 for_each_possible_cpu(i
) {
9419 spin_lock_init(&rq
->lock
);
9421 rq
->calc_load_active
= 0;
9422 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9423 init_cfs_rq(&rq
->cfs
, rq
);
9424 init_rt_rq(&rq
->rt
, rq
);
9425 #ifdef CONFIG_FAIR_GROUP_SCHED
9426 init_task_group
.shares
= init_task_group_load
;
9427 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9428 #ifdef CONFIG_CGROUP_SCHED
9430 * How much cpu bandwidth does init_task_group get?
9432 * In case of task-groups formed thr' the cgroup filesystem, it
9433 * gets 100% of the cpu resources in the system. This overall
9434 * system cpu resource is divided among the tasks of
9435 * init_task_group and its child task-groups in a fair manner,
9436 * based on each entity's (task or task-group's) weight
9437 * (se->load.weight).
9439 * In other words, if init_task_group has 10 tasks of weight
9440 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9441 * then A0's share of the cpu resource is:
9443 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9445 * We achieve this by letting init_task_group's tasks sit
9446 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9448 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9449 #elif defined CONFIG_USER_SCHED
9450 root_task_group
.shares
= NICE_0_LOAD
;
9451 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9453 * In case of task-groups formed thr' the user id of tasks,
9454 * init_task_group represents tasks belonging to root user.
9455 * Hence it forms a sibling of all subsequent groups formed.
9456 * In this case, init_task_group gets only a fraction of overall
9457 * system cpu resource, based on the weight assigned to root
9458 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9459 * by letting tasks of init_task_group sit in a separate cfs_rq
9460 * (init_tg_cfs_rq) and having one entity represent this group of
9461 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9463 init_tg_cfs_entry(&init_task_group
,
9464 &per_cpu(init_tg_cfs_rq
, i
),
9465 &per_cpu(init_sched_entity
, i
), i
, 1,
9466 root_task_group
.se
[i
]);
9469 #endif /* CONFIG_FAIR_GROUP_SCHED */
9471 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9472 #ifdef CONFIG_RT_GROUP_SCHED
9473 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9474 #ifdef CONFIG_CGROUP_SCHED
9475 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9476 #elif defined CONFIG_USER_SCHED
9477 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9478 init_tg_rt_entry(&init_task_group
,
9479 &per_cpu(init_rt_rq
, i
),
9480 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9481 root_task_group
.rt_se
[i
]);
9485 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9486 rq
->cpu_load
[j
] = 0;
9490 rq
->post_schedule
= 0;
9491 rq
->active_balance
= 0;
9492 rq
->next_balance
= jiffies
;
9496 rq
->migration_thread
= NULL
;
9497 INIT_LIST_HEAD(&rq
->migration_queue
);
9498 rq_attach_root(rq
, &def_root_domain
);
9501 atomic_set(&rq
->nr_iowait
, 0);
9504 set_load_weight(&init_task
);
9506 #ifdef CONFIG_PREEMPT_NOTIFIERS
9507 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9511 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9514 #ifdef CONFIG_RT_MUTEXES
9515 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9519 * The boot idle thread does lazy MMU switching as well:
9521 atomic_inc(&init_mm
.mm_count
);
9522 enter_lazy_tlb(&init_mm
, current
);
9525 * Make us the idle thread. Technically, schedule() should not be
9526 * called from this thread, however somewhere below it might be,
9527 * but because we are the idle thread, we just pick up running again
9528 * when this runqueue becomes "idle".
9530 init_idle(current
, smp_processor_id());
9532 calc_load_update
= jiffies
+ LOAD_FREQ
;
9535 * During early bootup we pretend to be a normal task:
9537 current
->sched_class
= &fair_sched_class
;
9539 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9540 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9543 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9544 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9546 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9549 perf_counter_init();
9551 scheduler_running
= 1;
9554 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9555 static inline int preempt_count_equals(int preempt_offset
)
9557 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9559 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9562 void __might_sleep(char *file
, int line
, int preempt_offset
)
9565 static unsigned long prev_jiffy
; /* ratelimiting */
9567 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9568 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9570 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9572 prev_jiffy
= jiffies
;
9575 "BUG: sleeping function called from invalid context at %s:%d\n",
9578 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9579 in_atomic(), irqs_disabled(),
9580 current
->pid
, current
->comm
);
9582 debug_show_held_locks(current
);
9583 if (irqs_disabled())
9584 print_irqtrace_events(current
);
9588 EXPORT_SYMBOL(__might_sleep
);
9591 #ifdef CONFIG_MAGIC_SYSRQ
9592 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9596 update_rq_clock(rq
);
9597 on_rq
= p
->se
.on_rq
;
9599 deactivate_task(rq
, p
, 0);
9600 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9602 activate_task(rq
, p
, 0);
9603 resched_task(rq
->curr
);
9607 void normalize_rt_tasks(void)
9609 struct task_struct
*g
, *p
;
9610 unsigned long flags
;
9613 read_lock_irqsave(&tasklist_lock
, flags
);
9614 do_each_thread(g
, p
) {
9616 * Only normalize user tasks:
9621 p
->se
.exec_start
= 0;
9622 #ifdef CONFIG_SCHEDSTATS
9623 p
->se
.wait_start
= 0;
9624 p
->se
.sleep_start
= 0;
9625 p
->se
.block_start
= 0;
9630 * Renice negative nice level userspace
9633 if (TASK_NICE(p
) < 0 && p
->mm
)
9634 set_user_nice(p
, 0);
9638 spin_lock(&p
->pi_lock
);
9639 rq
= __task_rq_lock(p
);
9641 normalize_task(rq
, p
);
9643 __task_rq_unlock(rq
);
9644 spin_unlock(&p
->pi_lock
);
9645 } while_each_thread(g
, p
);
9647 read_unlock_irqrestore(&tasklist_lock
, flags
);
9650 #endif /* CONFIG_MAGIC_SYSRQ */
9654 * These functions are only useful for the IA64 MCA handling.
9656 * They can only be called when the whole system has been
9657 * stopped - every CPU needs to be quiescent, and no scheduling
9658 * activity can take place. Using them for anything else would
9659 * be a serious bug, and as a result, they aren't even visible
9660 * under any other configuration.
9664 * curr_task - return the current task for a given cpu.
9665 * @cpu: the processor in question.
9667 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9669 struct task_struct
*curr_task(int cpu
)
9671 return cpu_curr(cpu
);
9675 * set_curr_task - set the current task for a given cpu.
9676 * @cpu: the processor in question.
9677 * @p: the task pointer to set.
9679 * Description: This function must only be used when non-maskable interrupts
9680 * are serviced on a separate stack. It allows the architecture to switch the
9681 * notion of the current task on a cpu in a non-blocking manner. This function
9682 * must be called with all CPU's synchronized, and interrupts disabled, the
9683 * and caller must save the original value of the current task (see
9684 * curr_task() above) and restore that value before reenabling interrupts and
9685 * re-starting the system.
9687 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9689 void set_curr_task(int cpu
, struct task_struct
*p
)
9696 #ifdef CONFIG_FAIR_GROUP_SCHED
9697 static void free_fair_sched_group(struct task_group
*tg
)
9701 for_each_possible_cpu(i
) {
9703 kfree(tg
->cfs_rq
[i
]);
9713 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9715 struct cfs_rq
*cfs_rq
;
9716 struct sched_entity
*se
;
9720 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9723 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9727 tg
->shares
= NICE_0_LOAD
;
9729 for_each_possible_cpu(i
) {
9732 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9733 GFP_KERNEL
, cpu_to_node(i
));
9737 se
= kzalloc_node(sizeof(struct sched_entity
),
9738 GFP_KERNEL
, cpu_to_node(i
));
9742 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9751 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9753 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9754 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9757 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9759 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9761 #else /* !CONFG_FAIR_GROUP_SCHED */
9762 static inline void free_fair_sched_group(struct task_group
*tg
)
9767 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9772 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9776 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9779 #endif /* CONFIG_FAIR_GROUP_SCHED */
9781 #ifdef CONFIG_RT_GROUP_SCHED
9782 static void free_rt_sched_group(struct task_group
*tg
)
9786 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9788 for_each_possible_cpu(i
) {
9790 kfree(tg
->rt_rq
[i
]);
9792 kfree(tg
->rt_se
[i
]);
9800 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9802 struct rt_rq
*rt_rq
;
9803 struct sched_rt_entity
*rt_se
;
9807 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9810 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9814 init_rt_bandwidth(&tg
->rt_bandwidth
,
9815 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9817 for_each_possible_cpu(i
) {
9820 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9821 GFP_KERNEL
, cpu_to_node(i
));
9825 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9826 GFP_KERNEL
, cpu_to_node(i
));
9830 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9839 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9841 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9842 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9845 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9847 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9849 #else /* !CONFIG_RT_GROUP_SCHED */
9850 static inline void free_rt_sched_group(struct task_group
*tg
)
9855 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9860 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9864 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9867 #endif /* CONFIG_RT_GROUP_SCHED */
9869 #ifdef CONFIG_GROUP_SCHED
9870 static void free_sched_group(struct task_group
*tg
)
9872 free_fair_sched_group(tg
);
9873 free_rt_sched_group(tg
);
9877 /* allocate runqueue etc for a new task group */
9878 struct task_group
*sched_create_group(struct task_group
*parent
)
9880 struct task_group
*tg
;
9881 unsigned long flags
;
9884 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9886 return ERR_PTR(-ENOMEM
);
9888 if (!alloc_fair_sched_group(tg
, parent
))
9891 if (!alloc_rt_sched_group(tg
, parent
))
9894 spin_lock_irqsave(&task_group_lock
, flags
);
9895 for_each_possible_cpu(i
) {
9896 register_fair_sched_group(tg
, i
);
9897 register_rt_sched_group(tg
, i
);
9899 list_add_rcu(&tg
->list
, &task_groups
);
9901 WARN_ON(!parent
); /* root should already exist */
9903 tg
->parent
= parent
;
9904 INIT_LIST_HEAD(&tg
->children
);
9905 list_add_rcu(&tg
->siblings
, &parent
->children
);
9906 spin_unlock_irqrestore(&task_group_lock
, flags
);
9911 free_sched_group(tg
);
9912 return ERR_PTR(-ENOMEM
);
9915 /* rcu callback to free various structures associated with a task group */
9916 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9918 /* now it should be safe to free those cfs_rqs */
9919 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9922 /* Destroy runqueue etc associated with a task group */
9923 void sched_destroy_group(struct task_group
*tg
)
9925 unsigned long flags
;
9928 spin_lock_irqsave(&task_group_lock
, flags
);
9929 for_each_possible_cpu(i
) {
9930 unregister_fair_sched_group(tg
, i
);
9931 unregister_rt_sched_group(tg
, i
);
9933 list_del_rcu(&tg
->list
);
9934 list_del_rcu(&tg
->siblings
);
9935 spin_unlock_irqrestore(&task_group_lock
, flags
);
9937 /* wait for possible concurrent references to cfs_rqs complete */
9938 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9941 /* change task's runqueue when it moves between groups.
9942 * The caller of this function should have put the task in its new group
9943 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9944 * reflect its new group.
9946 void sched_move_task(struct task_struct
*tsk
)
9949 unsigned long flags
;
9952 rq
= task_rq_lock(tsk
, &flags
);
9954 update_rq_clock(rq
);
9956 running
= task_current(rq
, tsk
);
9957 on_rq
= tsk
->se
.on_rq
;
9960 dequeue_task(rq
, tsk
, 0);
9961 if (unlikely(running
))
9962 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9964 set_task_rq(tsk
, task_cpu(tsk
));
9966 #ifdef CONFIG_FAIR_GROUP_SCHED
9967 if (tsk
->sched_class
->moved_group
)
9968 tsk
->sched_class
->moved_group(tsk
);
9971 if (unlikely(running
))
9972 tsk
->sched_class
->set_curr_task(rq
);
9974 enqueue_task(rq
, tsk
, 0);
9976 task_rq_unlock(rq
, &flags
);
9978 #endif /* CONFIG_GROUP_SCHED */
9980 #ifdef CONFIG_FAIR_GROUP_SCHED
9981 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9983 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9988 dequeue_entity(cfs_rq
, se
, 0);
9990 se
->load
.weight
= shares
;
9991 se
->load
.inv_weight
= 0;
9994 enqueue_entity(cfs_rq
, se
, 0);
9997 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9999 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10000 struct rq
*rq
= cfs_rq
->rq
;
10001 unsigned long flags
;
10003 spin_lock_irqsave(&rq
->lock
, flags
);
10004 __set_se_shares(se
, shares
);
10005 spin_unlock_irqrestore(&rq
->lock
, flags
);
10008 static DEFINE_MUTEX(shares_mutex
);
10010 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10013 unsigned long flags
;
10016 * We can't change the weight of the root cgroup.
10021 if (shares
< MIN_SHARES
)
10022 shares
= MIN_SHARES
;
10023 else if (shares
> MAX_SHARES
)
10024 shares
= MAX_SHARES
;
10026 mutex_lock(&shares_mutex
);
10027 if (tg
->shares
== shares
)
10030 spin_lock_irqsave(&task_group_lock
, flags
);
10031 for_each_possible_cpu(i
)
10032 unregister_fair_sched_group(tg
, i
);
10033 list_del_rcu(&tg
->siblings
);
10034 spin_unlock_irqrestore(&task_group_lock
, flags
);
10036 /* wait for any ongoing reference to this group to finish */
10037 synchronize_sched();
10040 * Now we are free to modify the group's share on each cpu
10041 * w/o tripping rebalance_share or load_balance_fair.
10043 tg
->shares
= shares
;
10044 for_each_possible_cpu(i
) {
10046 * force a rebalance
10048 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10049 set_se_shares(tg
->se
[i
], shares
);
10053 * Enable load balance activity on this group, by inserting it back on
10054 * each cpu's rq->leaf_cfs_rq_list.
10056 spin_lock_irqsave(&task_group_lock
, flags
);
10057 for_each_possible_cpu(i
)
10058 register_fair_sched_group(tg
, i
);
10059 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10060 spin_unlock_irqrestore(&task_group_lock
, flags
);
10062 mutex_unlock(&shares_mutex
);
10066 unsigned long sched_group_shares(struct task_group
*tg
)
10072 #ifdef CONFIG_RT_GROUP_SCHED
10074 * Ensure that the real time constraints are schedulable.
10076 static DEFINE_MUTEX(rt_constraints_mutex
);
10078 static unsigned long to_ratio(u64 period
, u64 runtime
)
10080 if (runtime
== RUNTIME_INF
)
10083 return div64_u64(runtime
<< 20, period
);
10086 /* Must be called with tasklist_lock held */
10087 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10089 struct task_struct
*g
, *p
;
10091 do_each_thread(g
, p
) {
10092 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10094 } while_each_thread(g
, p
);
10099 struct rt_schedulable_data
{
10100 struct task_group
*tg
;
10105 static int tg_schedulable(struct task_group
*tg
, void *data
)
10107 struct rt_schedulable_data
*d
= data
;
10108 struct task_group
*child
;
10109 unsigned long total
, sum
= 0;
10110 u64 period
, runtime
;
10112 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10113 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10116 period
= d
->rt_period
;
10117 runtime
= d
->rt_runtime
;
10120 #ifdef CONFIG_USER_SCHED
10121 if (tg
== &root_task_group
) {
10122 period
= global_rt_period();
10123 runtime
= global_rt_runtime();
10128 * Cannot have more runtime than the period.
10130 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10134 * Ensure we don't starve existing RT tasks.
10136 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10139 total
= to_ratio(period
, runtime
);
10142 * Nobody can have more than the global setting allows.
10144 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10148 * The sum of our children's runtime should not exceed our own.
10150 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10151 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10152 runtime
= child
->rt_bandwidth
.rt_runtime
;
10154 if (child
== d
->tg
) {
10155 period
= d
->rt_period
;
10156 runtime
= d
->rt_runtime
;
10159 sum
+= to_ratio(period
, runtime
);
10168 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10170 struct rt_schedulable_data data
= {
10172 .rt_period
= period
,
10173 .rt_runtime
= runtime
,
10176 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10179 static int tg_set_bandwidth(struct task_group
*tg
,
10180 u64 rt_period
, u64 rt_runtime
)
10184 mutex_lock(&rt_constraints_mutex
);
10185 read_lock(&tasklist_lock
);
10186 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10190 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10191 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10192 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10194 for_each_possible_cpu(i
) {
10195 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10197 spin_lock(&rt_rq
->rt_runtime_lock
);
10198 rt_rq
->rt_runtime
= rt_runtime
;
10199 spin_unlock(&rt_rq
->rt_runtime_lock
);
10201 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10203 read_unlock(&tasklist_lock
);
10204 mutex_unlock(&rt_constraints_mutex
);
10209 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10211 u64 rt_runtime
, rt_period
;
10213 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10214 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10215 if (rt_runtime_us
< 0)
10216 rt_runtime
= RUNTIME_INF
;
10218 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10221 long sched_group_rt_runtime(struct task_group
*tg
)
10225 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10228 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10229 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10230 return rt_runtime_us
;
10233 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10235 u64 rt_runtime
, rt_period
;
10237 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10238 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10240 if (rt_period
== 0)
10243 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10246 long sched_group_rt_period(struct task_group
*tg
)
10250 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10251 do_div(rt_period_us
, NSEC_PER_USEC
);
10252 return rt_period_us
;
10255 static int sched_rt_global_constraints(void)
10257 u64 runtime
, period
;
10260 if (sysctl_sched_rt_period
<= 0)
10263 runtime
= global_rt_runtime();
10264 period
= global_rt_period();
10267 * Sanity check on the sysctl variables.
10269 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10272 mutex_lock(&rt_constraints_mutex
);
10273 read_lock(&tasklist_lock
);
10274 ret
= __rt_schedulable(NULL
, 0, 0);
10275 read_unlock(&tasklist_lock
);
10276 mutex_unlock(&rt_constraints_mutex
);
10281 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10283 /* Don't accept realtime tasks when there is no way for them to run */
10284 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10290 #else /* !CONFIG_RT_GROUP_SCHED */
10291 static int sched_rt_global_constraints(void)
10293 unsigned long flags
;
10296 if (sysctl_sched_rt_period
<= 0)
10300 * There's always some RT tasks in the root group
10301 * -- migration, kstopmachine etc..
10303 if (sysctl_sched_rt_runtime
== 0)
10306 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10307 for_each_possible_cpu(i
) {
10308 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10310 spin_lock(&rt_rq
->rt_runtime_lock
);
10311 rt_rq
->rt_runtime
= global_rt_runtime();
10312 spin_unlock(&rt_rq
->rt_runtime_lock
);
10314 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10318 #endif /* CONFIG_RT_GROUP_SCHED */
10320 int sched_rt_handler(struct ctl_table
*table
, int write
,
10321 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10325 int old_period
, old_runtime
;
10326 static DEFINE_MUTEX(mutex
);
10328 mutex_lock(&mutex
);
10329 old_period
= sysctl_sched_rt_period
;
10330 old_runtime
= sysctl_sched_rt_runtime
;
10332 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10334 if (!ret
&& write
) {
10335 ret
= sched_rt_global_constraints();
10337 sysctl_sched_rt_period
= old_period
;
10338 sysctl_sched_rt_runtime
= old_runtime
;
10340 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10341 def_rt_bandwidth
.rt_period
=
10342 ns_to_ktime(global_rt_period());
10345 mutex_unlock(&mutex
);
10350 #ifdef CONFIG_CGROUP_SCHED
10352 /* return corresponding task_group object of a cgroup */
10353 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10355 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10356 struct task_group
, css
);
10359 static struct cgroup_subsys_state
*
10360 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10362 struct task_group
*tg
, *parent
;
10364 if (!cgrp
->parent
) {
10365 /* This is early initialization for the top cgroup */
10366 return &init_task_group
.css
;
10369 parent
= cgroup_tg(cgrp
->parent
);
10370 tg
= sched_create_group(parent
);
10372 return ERR_PTR(-ENOMEM
);
10378 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10380 struct task_group
*tg
= cgroup_tg(cgrp
);
10382 sched_destroy_group(tg
);
10386 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10387 struct task_struct
*tsk
)
10389 #ifdef CONFIG_RT_GROUP_SCHED
10390 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10393 /* We don't support RT-tasks being in separate groups */
10394 if (tsk
->sched_class
!= &fair_sched_class
)
10402 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10403 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10405 sched_move_task(tsk
);
10408 #ifdef CONFIG_FAIR_GROUP_SCHED
10409 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10412 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10415 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10417 struct task_group
*tg
= cgroup_tg(cgrp
);
10419 return (u64
) tg
->shares
;
10421 #endif /* CONFIG_FAIR_GROUP_SCHED */
10423 #ifdef CONFIG_RT_GROUP_SCHED
10424 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10427 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10430 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10432 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10435 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10438 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10441 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10443 return sched_group_rt_period(cgroup_tg(cgrp
));
10445 #endif /* CONFIG_RT_GROUP_SCHED */
10447 static struct cftype cpu_files
[] = {
10448 #ifdef CONFIG_FAIR_GROUP_SCHED
10451 .read_u64
= cpu_shares_read_u64
,
10452 .write_u64
= cpu_shares_write_u64
,
10455 #ifdef CONFIG_RT_GROUP_SCHED
10457 .name
= "rt_runtime_us",
10458 .read_s64
= cpu_rt_runtime_read
,
10459 .write_s64
= cpu_rt_runtime_write
,
10462 .name
= "rt_period_us",
10463 .read_u64
= cpu_rt_period_read_uint
,
10464 .write_u64
= cpu_rt_period_write_uint
,
10469 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10471 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10474 struct cgroup_subsys cpu_cgroup_subsys
= {
10476 .create
= cpu_cgroup_create
,
10477 .destroy
= cpu_cgroup_destroy
,
10478 .can_attach
= cpu_cgroup_can_attach
,
10479 .attach
= cpu_cgroup_attach
,
10480 .populate
= cpu_cgroup_populate
,
10481 .subsys_id
= cpu_cgroup_subsys_id
,
10485 #endif /* CONFIG_CGROUP_SCHED */
10487 #ifdef CONFIG_CGROUP_CPUACCT
10490 * CPU accounting code for task groups.
10492 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10493 * (balbir@in.ibm.com).
10496 /* track cpu usage of a group of tasks and its child groups */
10498 struct cgroup_subsys_state css
;
10499 /* cpuusage holds pointer to a u64-type object on every cpu */
10501 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10502 struct cpuacct
*parent
;
10505 struct cgroup_subsys cpuacct_subsys
;
10507 /* return cpu accounting group corresponding to this container */
10508 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10510 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10511 struct cpuacct
, css
);
10514 /* return cpu accounting group to which this task belongs */
10515 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10517 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10518 struct cpuacct
, css
);
10521 /* create a new cpu accounting group */
10522 static struct cgroup_subsys_state
*cpuacct_create(
10523 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10525 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10531 ca
->cpuusage
= alloc_percpu(u64
);
10535 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10536 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10537 goto out_free_counters
;
10540 ca
->parent
= cgroup_ca(cgrp
->parent
);
10546 percpu_counter_destroy(&ca
->cpustat
[i
]);
10547 free_percpu(ca
->cpuusage
);
10551 return ERR_PTR(-ENOMEM
);
10554 /* destroy an existing cpu accounting group */
10556 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10558 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10561 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10562 percpu_counter_destroy(&ca
->cpustat
[i
]);
10563 free_percpu(ca
->cpuusage
);
10567 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10569 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10572 #ifndef CONFIG_64BIT
10574 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10576 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10578 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10586 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10588 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10590 #ifndef CONFIG_64BIT
10592 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10594 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10596 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10602 /* return total cpu usage (in nanoseconds) of a group */
10603 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10605 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10606 u64 totalcpuusage
= 0;
10609 for_each_present_cpu(i
)
10610 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10612 return totalcpuusage
;
10615 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10618 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10627 for_each_present_cpu(i
)
10628 cpuacct_cpuusage_write(ca
, i
, 0);
10634 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10635 struct seq_file
*m
)
10637 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10641 for_each_present_cpu(i
) {
10642 percpu
= cpuacct_cpuusage_read(ca
, i
);
10643 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10645 seq_printf(m
, "\n");
10649 static const char *cpuacct_stat_desc
[] = {
10650 [CPUACCT_STAT_USER
] = "user",
10651 [CPUACCT_STAT_SYSTEM
] = "system",
10654 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10655 struct cgroup_map_cb
*cb
)
10657 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10660 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10661 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10662 val
= cputime64_to_clock_t(val
);
10663 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10668 static struct cftype files
[] = {
10671 .read_u64
= cpuusage_read
,
10672 .write_u64
= cpuusage_write
,
10675 .name
= "usage_percpu",
10676 .read_seq_string
= cpuacct_percpu_seq_read
,
10680 .read_map
= cpuacct_stats_show
,
10684 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10686 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10690 * charge this task's execution time to its accounting group.
10692 * called with rq->lock held.
10694 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10696 struct cpuacct
*ca
;
10699 if (unlikely(!cpuacct_subsys
.active
))
10702 cpu
= task_cpu(tsk
);
10708 for (; ca
; ca
= ca
->parent
) {
10709 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10710 *cpuusage
+= cputime
;
10717 * Charge the system/user time to the task's accounting group.
10719 static void cpuacct_update_stats(struct task_struct
*tsk
,
10720 enum cpuacct_stat_index idx
, cputime_t val
)
10722 struct cpuacct
*ca
;
10724 if (unlikely(!cpuacct_subsys
.active
))
10731 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10737 struct cgroup_subsys cpuacct_subsys
= {
10739 .create
= cpuacct_create
,
10740 .destroy
= cpuacct_destroy
,
10741 .populate
= cpuacct_populate
,
10742 .subsys_id
= cpuacct_subsys_id
,
10744 #endif /* CONFIG_CGROUP_CPUACCT */
10748 int rcu_expedited_torture_stats(char *page
)
10752 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10754 void synchronize_sched_expedited(void)
10757 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10759 #else /* #ifndef CONFIG_SMP */
10761 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10762 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10764 #define RCU_EXPEDITED_STATE_POST -2
10765 #define RCU_EXPEDITED_STATE_IDLE -1
10767 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10769 int rcu_expedited_torture_stats(char *page
)
10774 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10775 for_each_online_cpu(cpu
) {
10776 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10777 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10779 cnt
+= sprintf(&page
[cnt
], "\n");
10782 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10784 static long synchronize_sched_expedited_count
;
10787 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10788 * approach to force grace period to end quickly. This consumes
10789 * significant time on all CPUs, and is thus not recommended for
10790 * any sort of common-case code.
10792 * Note that it is illegal to call this function while holding any
10793 * lock that is acquired by a CPU-hotplug notifier. Failing to
10794 * observe this restriction will result in deadlock.
10796 void synchronize_sched_expedited(void)
10799 unsigned long flags
;
10800 bool need_full_sync
= 0;
10802 struct migration_req
*req
;
10806 smp_mb(); /* ensure prior mod happens before capturing snap. */
10807 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10809 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10811 if (trycount
++ < 10)
10812 udelay(trycount
* num_online_cpus());
10814 synchronize_sched();
10817 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
10818 smp_mb(); /* ensure test happens before caller kfree */
10823 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
10824 for_each_online_cpu(cpu
) {
10826 req
= &per_cpu(rcu_migration_req
, cpu
);
10827 init_completion(&req
->done
);
10829 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
10830 spin_lock_irqsave(&rq
->lock
, flags
);
10831 list_add(&req
->list
, &rq
->migration_queue
);
10832 spin_unlock_irqrestore(&rq
->lock
, flags
);
10833 wake_up_process(rq
->migration_thread
);
10835 for_each_online_cpu(cpu
) {
10836 rcu_expedited_state
= cpu
;
10837 req
= &per_cpu(rcu_migration_req
, cpu
);
10839 wait_for_completion(&req
->done
);
10840 spin_lock_irqsave(&rq
->lock
, flags
);
10841 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
10842 need_full_sync
= 1;
10843 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
10844 spin_unlock_irqrestore(&rq
->lock
, flags
);
10846 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10847 mutex_unlock(&rcu_sched_expedited_mutex
);
10849 if (need_full_sync
)
10850 synchronize_sched();
10852 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10854 #endif /* #else #ifndef CONFIG_SMP */