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/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
133 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
142 sg
->__cpu_power
+= val
;
143 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
147 static inline int rt_policy(int policy
)
149 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
154 static inline int task_has_rt_policy(struct task_struct
*p
)
156 return rt_policy(p
->policy
);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array
{
163 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
164 struct list_head queue
[MAX_RT_PRIO
];
167 struct rt_bandwidth
{
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock
;
172 struct hrtimer rt_period_timer
;
175 static struct rt_bandwidth def_rt_bandwidth
;
177 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
179 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
181 struct rt_bandwidth
*rt_b
=
182 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
188 now
= hrtimer_cb_get_time(timer
);
189 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
194 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
197 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
201 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
203 rt_b
->rt_period
= ns_to_ktime(period
);
204 rt_b
->rt_runtime
= runtime
;
206 spin_lock_init(&rt_b
->rt_runtime_lock
);
208 hrtimer_init(&rt_b
->rt_period_timer
,
209 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
210 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime
>= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
222 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
225 if (hrtimer_active(&rt_b
->rt_period_timer
))
228 spin_lock(&rt_b
->rt_runtime_lock
);
233 if (hrtimer_active(&rt_b
->rt_period_timer
))
236 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
237 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
239 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
240 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
241 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
242 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
243 HRTIMER_MODE_ABS_PINNED
, 0);
245 spin_unlock(&rt_b
->rt_runtime_lock
);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
251 hrtimer_cancel(&rt_b
->rt_period_timer
);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex
);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups
);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css
;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity
**se
;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq
**cfs_rq
;
284 unsigned long shares
;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity
**rt_se
;
289 struct rt_rq
**rt_rq
;
291 struct rt_bandwidth rt_bandwidth
;
295 struct list_head list
;
297 struct task_group
*parent
;
298 struct list_head siblings
;
299 struct list_head children
;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct
*user
)
307 user
->tg
->uid
= user
->uid
;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group
;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
326 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock
);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group
.children
);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group
;
370 /* return group to which a task belongs */
371 static inline struct task_group
*task_group(struct task_struct
*p
)
373 struct task_group
*tg
;
375 #ifdef CONFIG_USER_SCHED
377 tg
= __task_cred(p
)->user
->tg
;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
381 struct task_group
, css
);
383 tg
= &init_task_group
;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
393 p
->se
.parent
= task_group(p
)->se
[cpu
];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
398 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
412 static inline struct task_group
*task_group(struct task_struct
*p
)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load
;
422 unsigned long nr_running
;
427 struct rb_root tasks_timeline
;
428 struct rb_node
*rb_leftmost
;
430 struct list_head tasks
;
431 struct list_head
*balance_iterator
;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity
*curr
, *next
, *last
;
439 unsigned int nr_spread_over
;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list
;
453 struct task_group
*tg
; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight
;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load
;
470 * this cpu's part of tg->shares
472 unsigned long shares
;
475 * load.weight at the time we set shares
477 unsigned long rq_weight
;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active
;
485 unsigned long rt_nr_running
;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr
; /* highest queued rt task prio */
490 int next
; /* next highest */
495 unsigned long rt_nr_migratory
;
496 unsigned long rt_nr_total
;
498 struct plist_head pushable_tasks
;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock
;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted
;
510 struct list_head leaf_rt_rq_list
;
511 struct task_group
*tg
;
512 struct sched_rt_entity
*rt_se
;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online
;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask
;
538 struct cpupri cpupri
;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu
;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain
;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running
;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
577 unsigned long last_tick_seen
;
578 unsigned char in_nohz_recently
;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load
;
582 unsigned long nr_load_updates
;
584 u64 nr_migrations_in
;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list
;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list
;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible
;
605 struct task_struct
*curr
, *idle
;
606 unsigned long next_balance
;
607 struct mm_struct
*prev_mm
;
614 struct root_domain
*rd
;
615 struct sched_domain
*sd
;
617 unsigned char idle_at_tick
;
618 /* For active balancing */
621 /* cpu of this runqueue: */
625 unsigned long avg_load_per_task
;
627 struct task_struct
*migration_thread
;
628 struct list_head migration_queue
;
631 /* calc_load related fields */
632 unsigned long calc_load_update
;
633 long calc_load_active
;
635 #ifdef CONFIG_SCHED_HRTICK
637 int hrtick_csd_pending
;
638 struct call_single_data hrtick_csd
;
640 struct hrtimer hrtick_timer
;
643 #ifdef CONFIG_SCHEDSTATS
645 struct sched_info rq_sched_info
;
646 unsigned long long rq_cpu_time
;
647 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
649 /* sys_sched_yield() stats */
650 unsigned int yld_count
;
652 /* schedule() stats */
653 unsigned int sched_switch
;
654 unsigned int sched_count
;
655 unsigned int sched_goidle
;
657 /* try_to_wake_up() stats */
658 unsigned int ttwu_count
;
659 unsigned int ttwu_local
;
662 unsigned int bkl_count
;
666 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
668 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
670 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
673 static inline int cpu_of(struct rq
*rq
)
683 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684 * See detach_destroy_domains: synchronize_sched for details.
686 * The domain tree of any CPU may only be accessed from within
687 * preempt-disabled sections.
689 #define for_each_domain(cpu, __sd) \
690 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
692 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
693 #define this_rq() (&__get_cpu_var(runqueues))
694 #define task_rq(p) cpu_rq(task_cpu(p))
695 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 #define raw_rq() (&__raw_get_cpu_var(runqueues))
698 inline void update_rq_clock(struct rq
*rq
)
700 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
704 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
706 #ifdef CONFIG_SCHED_DEBUG
707 # define const_debug __read_mostly
709 # define const_debug static const
715 * Returns true if the current cpu runqueue is locked.
716 * This interface allows printk to be called with the runqueue lock
717 * held and know whether or not it is OK to wake up the klogd.
719 int runqueue_is_locked(void)
722 struct rq
*rq
= cpu_rq(cpu
);
725 ret
= spin_is_locked(&rq
->lock
);
731 * Debugging: various feature bits
734 #define SCHED_FEAT(name, enabled) \
735 __SCHED_FEAT_##name ,
738 #include "sched_features.h"
743 #define SCHED_FEAT(name, enabled) \
744 (1UL << __SCHED_FEAT_##name) * enabled |
746 const_debug
unsigned int sysctl_sched_features
=
747 #include "sched_features.h"
752 #ifdef CONFIG_SCHED_DEBUG
753 #define SCHED_FEAT(name, enabled) \
756 static __read_mostly
char *sched_feat_names
[] = {
757 #include "sched_features.h"
763 static int sched_feat_show(struct seq_file
*m
, void *v
)
767 for (i
= 0; sched_feat_names
[i
]; i
++) {
768 if (!(sysctl_sched_features
& (1UL << i
)))
770 seq_printf(m
, "%s ", sched_feat_names
[i
]);
778 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
779 size_t cnt
, loff_t
*ppos
)
789 if (copy_from_user(&buf
, ubuf
, cnt
))
794 if (strncmp(buf
, "NO_", 3) == 0) {
799 for (i
= 0; sched_feat_names
[i
]; i
++) {
800 int len
= strlen(sched_feat_names
[i
]);
802 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
804 sysctl_sched_features
&= ~(1UL << i
);
806 sysctl_sched_features
|= (1UL << i
);
811 if (!sched_feat_names
[i
])
819 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
821 return single_open(filp
, sched_feat_show
, NULL
);
824 static struct file_operations sched_feat_fops
= {
825 .open
= sched_feat_open
,
826 .write
= sched_feat_write
,
829 .release
= single_release
,
832 static __init
int sched_init_debug(void)
834 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
839 late_initcall(sched_init_debug
);
843 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
846 * Number of tasks to iterate in a single balance run.
847 * Limited because this is done with IRQs disabled.
849 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
852 * ratelimit for updating the group shares.
855 unsigned int sysctl_sched_shares_ratelimit
= 250000;
858 * Inject some fuzzyness into changing the per-cpu group shares
859 * this avoids remote rq-locks at the expense of fairness.
862 unsigned int sysctl_sched_shares_thresh
= 4;
865 * period over which we measure -rt task cpu usage in us.
868 unsigned int sysctl_sched_rt_period
= 1000000;
870 static __read_mostly
int scheduler_running
;
873 * part of the period that we allow rt tasks to run in us.
876 int sysctl_sched_rt_runtime
= 950000;
878 static inline u64
global_rt_period(void)
880 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
883 static inline u64
global_rt_runtime(void)
885 if (sysctl_sched_rt_runtime
< 0)
888 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
891 #ifndef prepare_arch_switch
892 # define prepare_arch_switch(next) do { } while (0)
894 #ifndef finish_arch_switch
895 # define finish_arch_switch(prev) do { } while (0)
898 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
900 return rq
->curr
== p
;
903 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
904 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
906 return task_current(rq
, p
);
909 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
913 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
915 #ifdef CONFIG_DEBUG_SPINLOCK
916 /* this is a valid case when another task releases the spinlock */
917 rq
->lock
.owner
= current
;
920 * If we are tracking spinlock dependencies then we have to
921 * fix up the runqueue lock - which gets 'carried over' from
924 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
926 spin_unlock_irq(&rq
->lock
);
929 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
930 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
935 return task_current(rq
, p
);
939 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
943 * We can optimise this out completely for !SMP, because the
944 * SMP rebalancing from interrupt is the only thing that cares
949 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
950 spin_unlock_irq(&rq
->lock
);
952 spin_unlock(&rq
->lock
);
956 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
960 * After ->oncpu is cleared, the task can be moved to a different CPU.
961 * We must ensure this doesn't happen until the switch is completely
967 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
971 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
974 * __task_rq_lock - lock the runqueue a given task resides on.
975 * Must be called interrupts disabled.
977 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
981 struct rq
*rq
= task_rq(p
);
982 spin_lock(&rq
->lock
);
983 if (likely(rq
== task_rq(p
)))
985 spin_unlock(&rq
->lock
);
990 * task_rq_lock - lock the runqueue a given task resides on and disable
991 * interrupts. Note the ordering: we can safely lookup the task_rq without
992 * explicitly disabling preemption.
994 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1000 local_irq_save(*flags
);
1002 spin_lock(&rq
->lock
);
1003 if (likely(rq
== task_rq(p
)))
1005 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1009 void task_rq_unlock_wait(struct task_struct
*p
)
1011 struct rq
*rq
= task_rq(p
);
1013 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1014 spin_unlock_wait(&rq
->lock
);
1017 static void __task_rq_unlock(struct rq
*rq
)
1018 __releases(rq
->lock
)
1020 spin_unlock(&rq
->lock
);
1023 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1024 __releases(rq
->lock
)
1026 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1030 * this_rq_lock - lock this runqueue and disable interrupts.
1032 static struct rq
*this_rq_lock(void)
1033 __acquires(rq
->lock
)
1037 local_irq_disable();
1039 spin_lock(&rq
->lock
);
1044 #ifdef CONFIG_SCHED_HRTICK
1046 * Use HR-timers to deliver accurate preemption points.
1048 * Its all a bit involved since we cannot program an hrt while holding the
1049 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1052 * When we get rescheduled we reprogram the hrtick_timer outside of the
1058 * - enabled by features
1059 * - hrtimer is actually high res
1061 static inline int hrtick_enabled(struct rq
*rq
)
1063 if (!sched_feat(HRTICK
))
1065 if (!cpu_active(cpu_of(rq
)))
1067 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1070 static void hrtick_clear(struct rq
*rq
)
1072 if (hrtimer_active(&rq
->hrtick_timer
))
1073 hrtimer_cancel(&rq
->hrtick_timer
);
1077 * High-resolution timer tick.
1078 * Runs from hardirq context with interrupts disabled.
1080 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1082 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1084 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1086 spin_lock(&rq
->lock
);
1087 update_rq_clock(rq
);
1088 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1089 spin_unlock(&rq
->lock
);
1091 return HRTIMER_NORESTART
;
1096 * called from hardirq (IPI) context
1098 static void __hrtick_start(void *arg
)
1100 struct rq
*rq
= arg
;
1102 spin_lock(&rq
->lock
);
1103 hrtimer_restart(&rq
->hrtick_timer
);
1104 rq
->hrtick_csd_pending
= 0;
1105 spin_unlock(&rq
->lock
);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq
*rq
, u64 delay
)
1115 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1116 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1118 hrtimer_set_expires(timer
, time
);
1120 if (rq
== this_rq()) {
1121 hrtimer_restart(timer
);
1122 } else if (!rq
->hrtick_csd_pending
) {
1123 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1124 rq
->hrtick_csd_pending
= 1;
1129 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1131 int cpu
= (int)(long)hcpu
;
1134 case CPU_UP_CANCELED
:
1135 case CPU_UP_CANCELED_FROZEN
:
1136 case CPU_DOWN_PREPARE
:
1137 case CPU_DOWN_PREPARE_FROZEN
:
1139 case CPU_DEAD_FROZEN
:
1140 hrtick_clear(cpu_rq(cpu
));
1147 static __init
void init_hrtick(void)
1149 hotcpu_notifier(hotplug_hrtick
, 0);
1153 * Called to set the hrtick timer state.
1155 * called with rq->lock held and irqs disabled
1157 static void hrtick_start(struct rq
*rq
, u64 delay
)
1159 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1160 HRTIMER_MODE_REL_PINNED
, 0);
1163 static inline void init_hrtick(void)
1166 #endif /* CONFIG_SMP */
1168 static void init_rq_hrtick(struct rq
*rq
)
1171 rq
->hrtick_csd_pending
= 0;
1173 rq
->hrtick_csd
.flags
= 0;
1174 rq
->hrtick_csd
.func
= __hrtick_start
;
1175 rq
->hrtick_csd
.info
= rq
;
1178 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1179 rq
->hrtick_timer
.function
= hrtick
;
1181 #else /* CONFIG_SCHED_HRTICK */
1182 static inline void hrtick_clear(struct rq
*rq
)
1186 static inline void init_rq_hrtick(struct rq
*rq
)
1190 static inline void init_hrtick(void)
1193 #endif /* CONFIG_SCHED_HRTICK */
1196 * resched_task - mark a task 'to be rescheduled now'.
1198 * On UP this means the setting of the need_resched flag, on SMP it
1199 * might also involve a cross-CPU call to trigger the scheduler on
1204 #ifndef tsk_is_polling
1205 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1208 static void resched_task(struct task_struct
*p
)
1212 assert_spin_locked(&task_rq(p
)->lock
);
1214 if (test_tsk_need_resched(p
))
1217 set_tsk_need_resched(p
);
1220 if (cpu
== smp_processor_id())
1223 /* NEED_RESCHED must be visible before we test polling */
1225 if (!tsk_is_polling(p
))
1226 smp_send_reschedule(cpu
);
1229 static void resched_cpu(int cpu
)
1231 struct rq
*rq
= cpu_rq(cpu
);
1232 unsigned long flags
;
1234 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1236 resched_task(cpu_curr(cpu
));
1237 spin_unlock_irqrestore(&rq
->lock
, flags
);
1242 * When add_timer_on() enqueues a timer into the timer wheel of an
1243 * idle CPU then this timer might expire before the next timer event
1244 * which is scheduled to wake up that CPU. In case of a completely
1245 * idle system the next event might even be infinite time into the
1246 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1247 * leaves the inner idle loop so the newly added timer is taken into
1248 * account when the CPU goes back to idle and evaluates the timer
1249 * wheel for the next timer event.
1251 void wake_up_idle_cpu(int cpu
)
1253 struct rq
*rq
= cpu_rq(cpu
);
1255 if (cpu
== smp_processor_id())
1259 * This is safe, as this function is called with the timer
1260 * wheel base lock of (cpu) held. When the CPU is on the way
1261 * to idle and has not yet set rq->curr to idle then it will
1262 * be serialized on the timer wheel base lock and take the new
1263 * timer into account automatically.
1265 if (rq
->curr
!= rq
->idle
)
1269 * We can set TIF_RESCHED on the idle task of the other CPU
1270 * lockless. The worst case is that the other CPU runs the
1271 * idle task through an additional NOOP schedule()
1273 set_tsk_need_resched(rq
->idle
);
1275 /* NEED_RESCHED must be visible before we test polling */
1277 if (!tsk_is_polling(rq
->idle
))
1278 smp_send_reschedule(cpu
);
1280 #endif /* CONFIG_NO_HZ */
1282 #else /* !CONFIG_SMP */
1283 static void resched_task(struct task_struct
*p
)
1285 assert_spin_locked(&task_rq(p
)->lock
);
1286 set_tsk_need_resched(p
);
1288 #endif /* CONFIG_SMP */
1290 #if BITS_PER_LONG == 32
1291 # define WMULT_CONST (~0UL)
1293 # define WMULT_CONST (1UL << 32)
1296 #define WMULT_SHIFT 32
1299 * Shift right and round:
1301 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1304 * delta *= weight / lw
1306 static unsigned long
1307 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1308 struct load_weight
*lw
)
1312 if (!lw
->inv_weight
) {
1313 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1316 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1320 tmp
= (u64
)delta_exec
* weight
;
1322 * Check whether we'd overflow the 64-bit multiplication:
1324 if (unlikely(tmp
> WMULT_CONST
))
1325 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1328 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1330 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1333 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1339 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1346 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1347 * of tasks with abnormal "nice" values across CPUs the contribution that
1348 * each task makes to its run queue's load is weighted according to its
1349 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1350 * scaled version of the new time slice allocation that they receive on time
1354 #define WEIGHT_IDLEPRIO 3
1355 #define WMULT_IDLEPRIO 1431655765
1358 * Nice levels are multiplicative, with a gentle 10% change for every
1359 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1360 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1361 * that remained on nice 0.
1363 * The "10% effect" is relative and cumulative: from _any_ nice level,
1364 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1365 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1366 * If a task goes up by ~10% and another task goes down by ~10% then
1367 * the relative distance between them is ~25%.)
1369 static const int prio_to_weight
[40] = {
1370 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1371 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1372 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1373 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1374 /* 0 */ 1024, 820, 655, 526, 423,
1375 /* 5 */ 335, 272, 215, 172, 137,
1376 /* 10 */ 110, 87, 70, 56, 45,
1377 /* 15 */ 36, 29, 23, 18, 15,
1381 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1383 * In cases where the weight does not change often, we can use the
1384 * precalculated inverse to speed up arithmetics by turning divisions
1385 * into multiplications:
1387 static const u32 prio_to_wmult
[40] = {
1388 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1389 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1390 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1391 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1392 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1393 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1394 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1395 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1398 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1401 * runqueue iterator, to support SMP load-balancing between different
1402 * scheduling classes, without having to expose their internal data
1403 * structures to the load-balancing proper:
1405 struct rq_iterator
{
1407 struct task_struct
*(*start
)(void *);
1408 struct task_struct
*(*next
)(void *);
1412 static unsigned long
1413 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1414 unsigned long max_load_move
, struct sched_domain
*sd
,
1415 enum cpu_idle_type idle
, int *all_pinned
,
1416 int *this_best_prio
, struct rq_iterator
*iterator
);
1419 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1420 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1421 struct rq_iterator
*iterator
);
1424 /* Time spent by the tasks of the cpu accounting group executing in ... */
1425 enum cpuacct_stat_index
{
1426 CPUACCT_STAT_USER
, /* ... user mode */
1427 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1429 CPUACCT_STAT_NSTATS
,
1432 #ifdef CONFIG_CGROUP_CPUACCT
1433 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1434 static void cpuacct_update_stats(struct task_struct
*tsk
,
1435 enum cpuacct_stat_index idx
, cputime_t val
);
1437 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1438 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1439 enum cpuacct_stat_index idx
, cputime_t val
) {}
1442 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1444 update_load_add(&rq
->load
, load
);
1447 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1449 update_load_sub(&rq
->load
, load
);
1452 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1453 typedef int (*tg_visitor
)(struct task_group
*, void *);
1456 * Iterate the full tree, calling @down when first entering a node and @up when
1457 * leaving it for the final time.
1459 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1461 struct task_group
*parent
, *child
;
1465 parent
= &root_task_group
;
1467 ret
= (*down
)(parent
, data
);
1470 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1477 ret
= (*up
)(parent
, data
);
1482 parent
= parent
->parent
;
1491 static int tg_nop(struct task_group
*tg
, void *data
)
1498 static unsigned long source_load(int cpu
, int type
);
1499 static unsigned long target_load(int cpu
, int type
);
1500 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1502 static unsigned long cpu_avg_load_per_task(int cpu
)
1504 struct rq
*rq
= cpu_rq(cpu
);
1505 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1508 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1510 rq
->avg_load_per_task
= 0;
1512 return rq
->avg_load_per_task
;
1515 #ifdef CONFIG_FAIR_GROUP_SCHED
1517 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1520 * Calculate and set the cpu's group shares.
1523 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1524 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1526 unsigned long shares
;
1527 unsigned long rq_weight
;
1532 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1535 * \Sum shares * rq_weight
1536 * shares = -----------------------
1540 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1541 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1543 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1544 sysctl_sched_shares_thresh
) {
1545 struct rq
*rq
= cpu_rq(cpu
);
1546 unsigned long flags
;
1548 spin_lock_irqsave(&rq
->lock
, flags
);
1549 tg
->cfs_rq
[cpu
]->shares
= shares
;
1551 __set_se_shares(tg
->se
[cpu
], shares
);
1552 spin_unlock_irqrestore(&rq
->lock
, flags
);
1557 * Re-compute the task group their per cpu shares over the given domain.
1558 * This needs to be done in a bottom-up fashion because the rq weight of a
1559 * parent group depends on the shares of its child groups.
1561 static int tg_shares_up(struct task_group
*tg
, void *data
)
1563 unsigned long weight
, rq_weight
= 0;
1564 unsigned long shares
= 0;
1565 struct sched_domain
*sd
= data
;
1568 for_each_cpu(i
, sched_domain_span(sd
)) {
1570 * If there are currently no tasks on the cpu pretend there
1571 * is one of average load so that when a new task gets to
1572 * run here it will not get delayed by group starvation.
1574 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1576 weight
= NICE_0_LOAD
;
1578 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1579 rq_weight
+= weight
;
1580 shares
+= tg
->cfs_rq
[i
]->shares
;
1583 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1584 shares
= tg
->shares
;
1586 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1587 shares
= tg
->shares
;
1589 for_each_cpu(i
, sched_domain_span(sd
))
1590 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1596 * Compute the cpu's hierarchical load factor for each task group.
1597 * This needs to be done in a top-down fashion because the load of a child
1598 * group is a fraction of its parents load.
1600 static int tg_load_down(struct task_group
*tg
, void *data
)
1603 long cpu
= (long)data
;
1606 load
= cpu_rq(cpu
)->load
.weight
;
1608 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1609 load
*= tg
->cfs_rq
[cpu
]->shares
;
1610 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1613 tg
->cfs_rq
[cpu
]->h_load
= load
;
1618 static void update_shares(struct sched_domain
*sd
)
1620 u64 now
= cpu_clock(raw_smp_processor_id());
1621 s64 elapsed
= now
- sd
->last_update
;
1623 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1624 sd
->last_update
= now
;
1625 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1629 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1631 spin_unlock(&rq
->lock
);
1633 spin_lock(&rq
->lock
);
1636 static void update_h_load(long cpu
)
1638 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1643 static inline void update_shares(struct sched_domain
*sd
)
1647 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1653 #ifdef CONFIG_PREEMPT
1656 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1657 * way at the expense of forcing extra atomic operations in all
1658 * invocations. This assures that the double_lock is acquired using the
1659 * same underlying policy as the spinlock_t on this architecture, which
1660 * reduces latency compared to the unfair variant below. However, it
1661 * also adds more overhead and therefore may reduce throughput.
1663 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1664 __releases(this_rq
->lock
)
1665 __acquires(busiest
->lock
)
1666 __acquires(this_rq
->lock
)
1668 spin_unlock(&this_rq
->lock
);
1669 double_rq_lock(this_rq
, busiest
);
1676 * Unfair double_lock_balance: Optimizes throughput at the expense of
1677 * latency by eliminating extra atomic operations when the locks are
1678 * already in proper order on entry. This favors lower cpu-ids and will
1679 * grant the double lock to lower cpus over higher ids under contention,
1680 * regardless of entry order into the function.
1682 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1683 __releases(this_rq
->lock
)
1684 __acquires(busiest
->lock
)
1685 __acquires(this_rq
->lock
)
1689 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1690 if (busiest
< this_rq
) {
1691 spin_unlock(&this_rq
->lock
);
1692 spin_lock(&busiest
->lock
);
1693 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1696 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1701 #endif /* CONFIG_PREEMPT */
1704 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1706 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1708 if (unlikely(!irqs_disabled())) {
1709 /* printk() doesn't work good under rq->lock */
1710 spin_unlock(&this_rq
->lock
);
1714 return _double_lock_balance(this_rq
, busiest
);
1717 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1718 __releases(busiest
->lock
)
1720 spin_unlock(&busiest
->lock
);
1721 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1725 #ifdef CONFIG_FAIR_GROUP_SCHED
1726 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1729 cfs_rq
->shares
= shares
;
1734 static void calc_load_account_active(struct rq
*this_rq
);
1736 #include "sched_stats.h"
1737 #include "sched_idletask.c"
1738 #include "sched_fair.c"
1739 #include "sched_rt.c"
1740 #ifdef CONFIG_SCHED_DEBUG
1741 # include "sched_debug.c"
1744 #define sched_class_highest (&rt_sched_class)
1745 #define for_each_class(class) \
1746 for (class = sched_class_highest; class; class = class->next)
1748 static void inc_nr_running(struct rq
*rq
)
1753 static void dec_nr_running(struct rq
*rq
)
1758 static void set_load_weight(struct task_struct
*p
)
1760 if (task_has_rt_policy(p
)) {
1761 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1762 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1767 * SCHED_IDLE tasks get minimal weight:
1769 if (p
->policy
== SCHED_IDLE
) {
1770 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1771 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1775 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1776 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1779 static void update_avg(u64
*avg
, u64 sample
)
1781 s64 diff
= sample
- *avg
;
1785 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1788 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1790 sched_info_queued(p
);
1791 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1795 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1798 if (p
->se
.last_wakeup
) {
1799 update_avg(&p
->se
.avg_overlap
,
1800 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1801 p
->se
.last_wakeup
= 0;
1803 update_avg(&p
->se
.avg_wakeup
,
1804 sysctl_sched_wakeup_granularity
);
1808 sched_info_dequeued(p
);
1809 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1814 * __normal_prio - return the priority that is based on the static prio
1816 static inline int __normal_prio(struct task_struct
*p
)
1818 return p
->static_prio
;
1822 * Calculate the expected normal priority: i.e. priority
1823 * without taking RT-inheritance into account. Might be
1824 * boosted by interactivity modifiers. Changes upon fork,
1825 * setprio syscalls, and whenever the interactivity
1826 * estimator recalculates.
1828 static inline int normal_prio(struct task_struct
*p
)
1832 if (task_has_rt_policy(p
))
1833 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1835 prio
= __normal_prio(p
);
1840 * Calculate the current priority, i.e. the priority
1841 * taken into account by the scheduler. This value might
1842 * be boosted by RT tasks, or might be boosted by
1843 * interactivity modifiers. Will be RT if the task got
1844 * RT-boosted. If not then it returns p->normal_prio.
1846 static int effective_prio(struct task_struct
*p
)
1848 p
->normal_prio
= normal_prio(p
);
1850 * If we are RT tasks or we were boosted to RT priority,
1851 * keep the priority unchanged. Otherwise, update priority
1852 * to the normal priority:
1854 if (!rt_prio(p
->prio
))
1855 return p
->normal_prio
;
1860 * activate_task - move a task to the runqueue.
1862 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1864 if (task_contributes_to_load(p
))
1865 rq
->nr_uninterruptible
--;
1867 enqueue_task(rq
, p
, wakeup
);
1872 * deactivate_task - remove a task from the runqueue.
1874 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1876 if (task_contributes_to_load(p
))
1877 rq
->nr_uninterruptible
++;
1879 dequeue_task(rq
, p
, sleep
);
1884 * task_curr - is this task currently executing on a CPU?
1885 * @p: the task in question.
1887 inline int task_curr(const struct task_struct
*p
)
1889 return cpu_curr(task_cpu(p
)) == p
;
1892 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1894 set_task_rq(p
, cpu
);
1897 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1898 * successfuly executed on another CPU. We must ensure that updates of
1899 * per-task data have been completed by this moment.
1902 task_thread_info(p
)->cpu
= cpu
;
1906 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1907 const struct sched_class
*prev_class
,
1908 int oldprio
, int running
)
1910 if (prev_class
!= p
->sched_class
) {
1911 if (prev_class
->switched_from
)
1912 prev_class
->switched_from(rq
, p
, running
);
1913 p
->sched_class
->switched_to(rq
, p
, running
);
1915 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1920 /* Used instead of source_load when we know the type == 0 */
1921 static unsigned long weighted_cpuload(const int cpu
)
1923 return cpu_rq(cpu
)->load
.weight
;
1927 * Is this task likely cache-hot:
1930 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1935 * Buddy candidates are cache hot:
1937 if (sched_feat(CACHE_HOT_BUDDY
) &&
1938 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1939 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1942 if (p
->sched_class
!= &fair_sched_class
)
1945 if (sysctl_sched_migration_cost
== -1)
1947 if (sysctl_sched_migration_cost
== 0)
1950 delta
= now
- p
->se
.exec_start
;
1952 return delta
< (s64
)sysctl_sched_migration_cost
;
1956 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1958 int old_cpu
= task_cpu(p
);
1959 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1960 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1961 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1964 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1966 trace_sched_migrate_task(p
, new_cpu
);
1968 #ifdef CONFIG_SCHEDSTATS
1969 if (p
->se
.wait_start
)
1970 p
->se
.wait_start
-= clock_offset
;
1971 if (p
->se
.sleep_start
)
1972 p
->se
.sleep_start
-= clock_offset
;
1973 if (p
->se
.block_start
)
1974 p
->se
.block_start
-= clock_offset
;
1976 if (old_cpu
!= new_cpu
) {
1977 p
->se
.nr_migrations
++;
1978 new_rq
->nr_migrations_in
++;
1979 #ifdef CONFIG_SCHEDSTATS
1980 if (task_hot(p
, old_rq
->clock
, NULL
))
1981 schedstat_inc(p
, se
.nr_forced2_migrations
);
1983 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
1986 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1987 new_cfsrq
->min_vruntime
;
1989 __set_task_cpu(p
, new_cpu
);
1992 struct migration_req
{
1993 struct list_head list
;
1995 struct task_struct
*task
;
1998 struct completion done
;
2002 * The task's runqueue lock must be held.
2003 * Returns true if you have to wait for migration thread.
2006 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2008 struct rq
*rq
= task_rq(p
);
2011 * If the task is not on a runqueue (and not running), then
2012 * it is sufficient to simply update the task's cpu field.
2014 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2015 set_task_cpu(p
, dest_cpu
);
2019 init_completion(&req
->done
);
2021 req
->dest_cpu
= dest_cpu
;
2022 list_add(&req
->list
, &rq
->migration_queue
);
2028 * wait_task_context_switch - wait for a thread to complete at least one
2031 * @p must not be current.
2033 void wait_task_context_switch(struct task_struct
*p
)
2035 unsigned long nvcsw
, nivcsw
, flags
;
2043 * The runqueue is assigned before the actual context
2044 * switch. We need to take the runqueue lock.
2046 * We could check initially without the lock but it is
2047 * very likely that we need to take the lock in every
2050 rq
= task_rq_lock(p
, &flags
);
2051 running
= task_running(rq
, p
);
2052 task_rq_unlock(rq
, &flags
);
2054 if (likely(!running
))
2057 * The switch count is incremented before the actual
2058 * context switch. We thus wait for two switches to be
2059 * sure at least one completed.
2061 if ((p
->nvcsw
- nvcsw
) > 1)
2063 if ((p
->nivcsw
- nivcsw
) > 1)
2071 * wait_task_inactive - wait for a thread to unschedule.
2073 * If @match_state is nonzero, it's the @p->state value just checked and
2074 * not expected to change. If it changes, i.e. @p might have woken up,
2075 * then return zero. When we succeed in waiting for @p to be off its CPU,
2076 * we return a positive number (its total switch count). If a second call
2077 * a short while later returns the same number, the caller can be sure that
2078 * @p has remained unscheduled the whole time.
2080 * The caller must ensure that the task *will* unschedule sometime soon,
2081 * else this function might spin for a *long* time. This function can't
2082 * be called with interrupts off, or it may introduce deadlock with
2083 * smp_call_function() if an IPI is sent by the same process we are
2084 * waiting to become inactive.
2086 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2088 unsigned long flags
;
2095 * We do the initial early heuristics without holding
2096 * any task-queue locks at all. We'll only try to get
2097 * the runqueue lock when things look like they will
2103 * If the task is actively running on another CPU
2104 * still, just relax and busy-wait without holding
2107 * NOTE! Since we don't hold any locks, it's not
2108 * even sure that "rq" stays as the right runqueue!
2109 * But we don't care, since "task_running()" will
2110 * return false if the runqueue has changed and p
2111 * is actually now running somewhere else!
2113 while (task_running(rq
, p
)) {
2114 if (match_state
&& unlikely(p
->state
!= match_state
))
2120 * Ok, time to look more closely! We need the rq
2121 * lock now, to be *sure*. If we're wrong, we'll
2122 * just go back and repeat.
2124 rq
= task_rq_lock(p
, &flags
);
2125 trace_sched_wait_task(rq
, p
);
2126 running
= task_running(rq
, p
);
2127 on_rq
= p
->se
.on_rq
;
2129 if (!match_state
|| p
->state
== match_state
)
2130 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2131 task_rq_unlock(rq
, &flags
);
2134 * If it changed from the expected state, bail out now.
2136 if (unlikely(!ncsw
))
2140 * Was it really running after all now that we
2141 * checked with the proper locks actually held?
2143 * Oops. Go back and try again..
2145 if (unlikely(running
)) {
2151 * It's not enough that it's not actively running,
2152 * it must be off the runqueue _entirely_, and not
2155 * So if it was still runnable (but just not actively
2156 * running right now), it's preempted, and we should
2157 * yield - it could be a while.
2159 if (unlikely(on_rq
)) {
2160 schedule_timeout_uninterruptible(1);
2165 * Ahh, all good. It wasn't running, and it wasn't
2166 * runnable, which means that it will never become
2167 * running in the future either. We're all done!
2176 * kick_process - kick a running thread to enter/exit the kernel
2177 * @p: the to-be-kicked thread
2179 * Cause a process which is running on another CPU to enter
2180 * kernel-mode, without any delay. (to get signals handled.)
2182 * NOTE: this function doesnt have to take the runqueue lock,
2183 * because all it wants to ensure is that the remote task enters
2184 * the kernel. If the IPI races and the task has been migrated
2185 * to another CPU then no harm is done and the purpose has been
2188 void kick_process(struct task_struct
*p
)
2194 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2195 smp_send_reschedule(cpu
);
2198 EXPORT_SYMBOL_GPL(kick_process
);
2201 * Return a low guess at the load of a migration-source cpu weighted
2202 * according to the scheduling class and "nice" value.
2204 * We want to under-estimate the load of migration sources, to
2205 * balance conservatively.
2207 static unsigned long source_load(int cpu
, int type
)
2209 struct rq
*rq
= cpu_rq(cpu
);
2210 unsigned long total
= weighted_cpuload(cpu
);
2212 if (type
== 0 || !sched_feat(LB_BIAS
))
2215 return min(rq
->cpu_load
[type
-1], total
);
2219 * Return a high guess at the load of a migration-target cpu weighted
2220 * according to the scheduling class and "nice" value.
2222 static unsigned long target_load(int cpu
, int type
)
2224 struct rq
*rq
= cpu_rq(cpu
);
2225 unsigned long total
= weighted_cpuload(cpu
);
2227 if (type
== 0 || !sched_feat(LB_BIAS
))
2230 return max(rq
->cpu_load
[type
-1], total
);
2234 * find_idlest_group finds and returns the least busy CPU group within the
2237 static struct sched_group
*
2238 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2240 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2241 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2242 int load_idx
= sd
->forkexec_idx
;
2243 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2246 unsigned long load
, avg_load
;
2250 /* Skip over this group if it has no CPUs allowed */
2251 if (!cpumask_intersects(sched_group_cpus(group
),
2255 local_group
= cpumask_test_cpu(this_cpu
,
2256 sched_group_cpus(group
));
2258 /* Tally up the load of all CPUs in the group */
2261 for_each_cpu(i
, sched_group_cpus(group
)) {
2262 /* Bias balancing toward cpus of our domain */
2264 load
= source_load(i
, load_idx
);
2266 load
= target_load(i
, load_idx
);
2271 /* Adjust by relative CPU power of the group */
2272 avg_load
= sg_div_cpu_power(group
,
2273 avg_load
* SCHED_LOAD_SCALE
);
2276 this_load
= avg_load
;
2278 } else if (avg_load
< min_load
) {
2279 min_load
= avg_load
;
2282 } while (group
= group
->next
, group
!= sd
->groups
);
2284 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2290 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2293 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2295 unsigned long load
, min_load
= ULONG_MAX
;
2299 /* Traverse only the allowed CPUs */
2300 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2301 load
= weighted_cpuload(i
);
2303 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2313 * sched_balance_self: balance the current task (running on cpu) in domains
2314 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2317 * Balance, ie. select the least loaded group.
2319 * Returns the target CPU number, or the same CPU if no balancing is needed.
2321 * preempt must be disabled.
2323 static int sched_balance_self(int cpu
, int flag
)
2325 struct task_struct
*t
= current
;
2326 struct sched_domain
*tmp
, *sd
= NULL
;
2328 for_each_domain(cpu
, tmp
) {
2330 * If power savings logic is enabled for a domain, stop there.
2332 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2334 if (tmp
->flags
& flag
)
2342 struct sched_group
*group
;
2343 int new_cpu
, weight
;
2345 if (!(sd
->flags
& flag
)) {
2350 group
= find_idlest_group(sd
, t
, cpu
);
2356 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2357 if (new_cpu
== -1 || new_cpu
== cpu
) {
2358 /* Now try balancing at a lower domain level of cpu */
2363 /* Now try balancing at a lower domain level of new_cpu */
2365 weight
= cpumask_weight(sched_domain_span(sd
));
2367 for_each_domain(cpu
, tmp
) {
2368 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2370 if (tmp
->flags
& flag
)
2373 /* while loop will break here if sd == NULL */
2379 #endif /* CONFIG_SMP */
2382 * task_oncpu_function_call - call a function on the cpu on which a task runs
2383 * @p: the task to evaluate
2384 * @func: the function to be called
2385 * @info: the function call argument
2387 * Calls the function @func when the task is currently running. This might
2388 * be on the current CPU, which just calls the function directly
2390 void task_oncpu_function_call(struct task_struct
*p
,
2391 void (*func
) (void *info
), void *info
)
2398 smp_call_function_single(cpu
, func
, info
, 1);
2403 * try_to_wake_up - wake up a thread
2404 * @p: the to-be-woken-up thread
2405 * @state: the mask of task states that can be woken
2406 * @sync: do a synchronous wakeup?
2408 * Put it on the run-queue if it's not already there. The "current"
2409 * thread is always on the run-queue (except when the actual
2410 * re-schedule is in progress), and as such you're allowed to do
2411 * the simpler "current->state = TASK_RUNNING" to mark yourself
2412 * runnable without the overhead of this.
2414 * returns failure only if the task is already active.
2416 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2418 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2419 unsigned long flags
;
2423 if (!sched_feat(SYNC_WAKEUPS
))
2427 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2428 struct sched_domain
*sd
;
2430 this_cpu
= raw_smp_processor_id();
2433 for_each_domain(this_cpu
, sd
) {
2434 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2443 rq
= task_rq_lock(p
, &flags
);
2444 update_rq_clock(rq
);
2445 old_state
= p
->state
;
2446 if (!(old_state
& state
))
2454 this_cpu
= smp_processor_id();
2457 if (unlikely(task_running(rq
, p
)))
2460 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2461 if (cpu
!= orig_cpu
) {
2462 set_task_cpu(p
, cpu
);
2463 task_rq_unlock(rq
, &flags
);
2464 /* might preempt at this point */
2465 rq
= task_rq_lock(p
, &flags
);
2466 old_state
= p
->state
;
2467 if (!(old_state
& state
))
2472 this_cpu
= smp_processor_id();
2476 #ifdef CONFIG_SCHEDSTATS
2477 schedstat_inc(rq
, ttwu_count
);
2478 if (cpu
== this_cpu
)
2479 schedstat_inc(rq
, ttwu_local
);
2481 struct sched_domain
*sd
;
2482 for_each_domain(this_cpu
, sd
) {
2483 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2484 schedstat_inc(sd
, ttwu_wake_remote
);
2489 #endif /* CONFIG_SCHEDSTATS */
2492 #endif /* CONFIG_SMP */
2493 schedstat_inc(p
, se
.nr_wakeups
);
2495 schedstat_inc(p
, se
.nr_wakeups_sync
);
2496 if (orig_cpu
!= cpu
)
2497 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2498 if (cpu
== this_cpu
)
2499 schedstat_inc(p
, se
.nr_wakeups_local
);
2501 schedstat_inc(p
, se
.nr_wakeups_remote
);
2502 activate_task(rq
, p
, 1);
2506 * Only attribute actual wakeups done by this task.
2508 if (!in_interrupt()) {
2509 struct sched_entity
*se
= ¤t
->se
;
2510 u64 sample
= se
->sum_exec_runtime
;
2512 if (se
->last_wakeup
)
2513 sample
-= se
->last_wakeup
;
2515 sample
-= se
->start_runtime
;
2516 update_avg(&se
->avg_wakeup
, sample
);
2518 se
->last_wakeup
= se
->sum_exec_runtime
;
2522 trace_sched_wakeup(rq
, p
, success
);
2523 check_preempt_curr(rq
, p
, sync
);
2525 p
->state
= TASK_RUNNING
;
2527 if (p
->sched_class
->task_wake_up
)
2528 p
->sched_class
->task_wake_up(rq
, p
);
2531 task_rq_unlock(rq
, &flags
);
2537 * wake_up_process - Wake up a specific process
2538 * @p: The process to be woken up.
2540 * Attempt to wake up the nominated process and move it to the set of runnable
2541 * processes. Returns 1 if the process was woken up, 0 if it was already
2544 * It may be assumed that this function implies a write memory barrier before
2545 * changing the task state if and only if any tasks are woken up.
2547 int wake_up_process(struct task_struct
*p
)
2549 return try_to_wake_up(p
, TASK_ALL
, 0);
2551 EXPORT_SYMBOL(wake_up_process
);
2553 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2555 return try_to_wake_up(p
, state
, 0);
2559 * Perform scheduler related setup for a newly forked process p.
2560 * p is forked by current.
2562 * __sched_fork() is basic setup used by init_idle() too:
2564 static void __sched_fork(struct task_struct
*p
)
2566 p
->se
.exec_start
= 0;
2567 p
->se
.sum_exec_runtime
= 0;
2568 p
->se
.prev_sum_exec_runtime
= 0;
2569 p
->se
.nr_migrations
= 0;
2570 p
->se
.last_wakeup
= 0;
2571 p
->se
.avg_overlap
= 0;
2572 p
->se
.start_runtime
= 0;
2573 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2575 #ifdef CONFIG_SCHEDSTATS
2576 p
->se
.wait_start
= 0;
2578 p
->se
.wait_count
= 0;
2581 p
->se
.sleep_start
= 0;
2582 p
->se
.sleep_max
= 0;
2583 p
->se
.sum_sleep_runtime
= 0;
2585 p
->se
.block_start
= 0;
2586 p
->se
.block_max
= 0;
2588 p
->se
.slice_max
= 0;
2590 p
->se
.nr_migrations_cold
= 0;
2591 p
->se
.nr_failed_migrations_affine
= 0;
2592 p
->se
.nr_failed_migrations_running
= 0;
2593 p
->se
.nr_failed_migrations_hot
= 0;
2594 p
->se
.nr_forced_migrations
= 0;
2595 p
->se
.nr_forced2_migrations
= 0;
2597 p
->se
.nr_wakeups
= 0;
2598 p
->se
.nr_wakeups_sync
= 0;
2599 p
->se
.nr_wakeups_migrate
= 0;
2600 p
->se
.nr_wakeups_local
= 0;
2601 p
->se
.nr_wakeups_remote
= 0;
2602 p
->se
.nr_wakeups_affine
= 0;
2603 p
->se
.nr_wakeups_affine_attempts
= 0;
2604 p
->se
.nr_wakeups_passive
= 0;
2605 p
->se
.nr_wakeups_idle
= 0;
2609 INIT_LIST_HEAD(&p
->rt
.run_list
);
2611 INIT_LIST_HEAD(&p
->se
.group_node
);
2613 #ifdef CONFIG_PREEMPT_NOTIFIERS
2614 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2618 * We mark the process as running here, but have not actually
2619 * inserted it onto the runqueue yet. This guarantees that
2620 * nobody will actually run it, and a signal or other external
2621 * event cannot wake it up and insert it on the runqueue either.
2623 p
->state
= TASK_RUNNING
;
2627 * fork()/clone()-time setup:
2629 void sched_fork(struct task_struct
*p
, int clone_flags
)
2631 int cpu
= get_cpu();
2636 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2638 set_task_cpu(p
, cpu
);
2641 * Make sure we do not leak PI boosting priority to the child.
2643 p
->prio
= current
->normal_prio
;
2646 * Revert to default priority/policy on fork if requested.
2648 if (unlikely(p
->sched_reset_on_fork
)) {
2649 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
)
2650 p
->policy
= SCHED_NORMAL
;
2652 if (p
->normal_prio
< DEFAULT_PRIO
)
2653 p
->prio
= DEFAULT_PRIO
;
2655 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2656 p
->static_prio
= NICE_TO_PRIO(0);
2661 * We don't need the reset flag anymore after the fork. It has
2662 * fulfilled its duty:
2664 p
->sched_reset_on_fork
= 0;
2667 if (!rt_prio(p
->prio
))
2668 p
->sched_class
= &fair_sched_class
;
2670 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2671 if (likely(sched_info_on()))
2672 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2674 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2677 #ifdef CONFIG_PREEMPT
2678 /* Want to start with kernel preemption disabled. */
2679 task_thread_info(p
)->preempt_count
= 1;
2681 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2687 * wake_up_new_task - wake up a newly created task for the first time.
2689 * This function will do some initial scheduler statistics housekeeping
2690 * that must be done for every newly created context, then puts the task
2691 * on the runqueue and wakes it.
2693 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2695 unsigned long flags
;
2698 rq
= task_rq_lock(p
, &flags
);
2699 BUG_ON(p
->state
!= TASK_RUNNING
);
2700 update_rq_clock(rq
);
2702 p
->prio
= effective_prio(p
);
2704 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2705 activate_task(rq
, p
, 0);
2708 * Let the scheduling class do new task startup
2709 * management (if any):
2711 p
->sched_class
->task_new(rq
, p
);
2714 trace_sched_wakeup_new(rq
, p
, 1);
2715 check_preempt_curr(rq
, p
, 0);
2717 if (p
->sched_class
->task_wake_up
)
2718 p
->sched_class
->task_wake_up(rq
, p
);
2720 task_rq_unlock(rq
, &flags
);
2723 #ifdef CONFIG_PREEMPT_NOTIFIERS
2726 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2727 * @notifier: notifier struct to register
2729 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2731 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2733 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2736 * preempt_notifier_unregister - no longer interested in preemption notifications
2737 * @notifier: notifier struct to unregister
2739 * This is safe to call from within a preemption notifier.
2741 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2743 hlist_del(¬ifier
->link
);
2745 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2747 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2749 struct preempt_notifier
*notifier
;
2750 struct hlist_node
*node
;
2752 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2753 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2757 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2758 struct task_struct
*next
)
2760 struct preempt_notifier
*notifier
;
2761 struct hlist_node
*node
;
2763 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2764 notifier
->ops
->sched_out(notifier
, next
);
2767 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2769 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2774 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2775 struct task_struct
*next
)
2779 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2782 * prepare_task_switch - prepare to switch tasks
2783 * @rq: the runqueue preparing to switch
2784 * @prev: the current task that is being switched out
2785 * @next: the task we are going to switch to.
2787 * This is called with the rq lock held and interrupts off. It must
2788 * be paired with a subsequent finish_task_switch after the context
2791 * prepare_task_switch sets up locking and calls architecture specific
2795 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2796 struct task_struct
*next
)
2798 fire_sched_out_preempt_notifiers(prev
, next
);
2799 prepare_lock_switch(rq
, next
);
2800 prepare_arch_switch(next
);
2804 * finish_task_switch - clean up after a task-switch
2805 * @rq: runqueue associated with task-switch
2806 * @prev: the thread we just switched away from.
2808 * finish_task_switch must be called after the context switch, paired
2809 * with a prepare_task_switch call before the context switch.
2810 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2811 * and do any other architecture-specific cleanup actions.
2813 * Note that we may have delayed dropping an mm in context_switch(). If
2814 * so, we finish that here outside of the runqueue lock. (Doing it
2815 * with the lock held can cause deadlocks; see schedule() for
2818 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2819 __releases(rq
->lock
)
2821 struct mm_struct
*mm
= rq
->prev_mm
;
2824 int post_schedule
= 0;
2826 if (current
->sched_class
->needs_post_schedule
)
2827 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2833 * A task struct has one reference for the use as "current".
2834 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2835 * schedule one last time. The schedule call will never return, and
2836 * the scheduled task must drop that reference.
2837 * The test for TASK_DEAD must occur while the runqueue locks are
2838 * still held, otherwise prev could be scheduled on another cpu, die
2839 * there before we look at prev->state, and then the reference would
2841 * Manfred Spraul <manfred@colorfullife.com>
2843 prev_state
= prev
->state
;
2844 finish_arch_switch(prev
);
2845 perf_counter_task_sched_in(current
, cpu_of(rq
));
2846 finish_lock_switch(rq
, prev
);
2849 current
->sched_class
->post_schedule(rq
);
2852 fire_sched_in_preempt_notifiers(current
);
2855 if (unlikely(prev_state
== TASK_DEAD
)) {
2857 * Remove function-return probe instances associated with this
2858 * task and put them back on the free list.
2860 kprobe_flush_task(prev
);
2861 put_task_struct(prev
);
2866 * schedule_tail - first thing a freshly forked thread must call.
2867 * @prev: the thread we just switched away from.
2869 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2870 __releases(rq
->lock
)
2872 struct rq
*rq
= this_rq();
2874 finish_task_switch(rq
, prev
);
2875 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2876 /* In this case, finish_task_switch does not reenable preemption */
2879 if (current
->set_child_tid
)
2880 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2884 * context_switch - switch to the new MM and the new
2885 * thread's register state.
2888 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2889 struct task_struct
*next
)
2891 struct mm_struct
*mm
, *oldmm
;
2893 prepare_task_switch(rq
, prev
, next
);
2894 trace_sched_switch(rq
, prev
, next
);
2896 oldmm
= prev
->active_mm
;
2898 * For paravirt, this is coupled with an exit in switch_to to
2899 * combine the page table reload and the switch backend into
2902 arch_start_context_switch(prev
);
2904 if (unlikely(!mm
)) {
2905 next
->active_mm
= oldmm
;
2906 atomic_inc(&oldmm
->mm_count
);
2907 enter_lazy_tlb(oldmm
, next
);
2909 switch_mm(oldmm
, mm
, next
);
2911 if (unlikely(!prev
->mm
)) {
2912 prev
->active_mm
= NULL
;
2913 rq
->prev_mm
= oldmm
;
2916 * Since the runqueue lock will be released by the next
2917 * task (which is an invalid locking op but in the case
2918 * of the scheduler it's an obvious special-case), so we
2919 * do an early lockdep release here:
2921 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2922 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2925 /* Here we just switch the register state and the stack. */
2926 switch_to(prev
, next
, prev
);
2930 * this_rq must be evaluated again because prev may have moved
2931 * CPUs since it called schedule(), thus the 'rq' on its stack
2932 * frame will be invalid.
2934 finish_task_switch(this_rq(), prev
);
2938 * nr_running, nr_uninterruptible and nr_context_switches:
2940 * externally visible scheduler statistics: current number of runnable
2941 * threads, current number of uninterruptible-sleeping threads, total
2942 * number of context switches performed since bootup.
2944 unsigned long nr_running(void)
2946 unsigned long i
, sum
= 0;
2948 for_each_online_cpu(i
)
2949 sum
+= cpu_rq(i
)->nr_running
;
2954 unsigned long nr_uninterruptible(void)
2956 unsigned long i
, sum
= 0;
2958 for_each_possible_cpu(i
)
2959 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2962 * Since we read the counters lockless, it might be slightly
2963 * inaccurate. Do not allow it to go below zero though:
2965 if (unlikely((long)sum
< 0))
2971 unsigned long long nr_context_switches(void)
2974 unsigned long long sum
= 0;
2976 for_each_possible_cpu(i
)
2977 sum
+= cpu_rq(i
)->nr_switches
;
2982 unsigned long nr_iowait(void)
2984 unsigned long i
, sum
= 0;
2986 for_each_possible_cpu(i
)
2987 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2992 /* Variables and functions for calc_load */
2993 static atomic_long_t calc_load_tasks
;
2994 static unsigned long calc_load_update
;
2995 unsigned long avenrun
[3];
2996 EXPORT_SYMBOL(avenrun
);
2999 * get_avenrun - get the load average array
3000 * @loads: pointer to dest load array
3001 * @offset: offset to add
3002 * @shift: shift count to shift the result left
3004 * These values are estimates at best, so no need for locking.
3006 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3008 loads
[0] = (avenrun
[0] + offset
) << shift
;
3009 loads
[1] = (avenrun
[1] + offset
) << shift
;
3010 loads
[2] = (avenrun
[2] + offset
) << shift
;
3013 static unsigned long
3014 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3017 load
+= active
* (FIXED_1
- exp
);
3018 return load
>> FSHIFT
;
3022 * calc_load - update the avenrun load estimates 10 ticks after the
3023 * CPUs have updated calc_load_tasks.
3025 void calc_global_load(void)
3027 unsigned long upd
= calc_load_update
+ 10;
3030 if (time_before(jiffies
, upd
))
3033 active
= atomic_long_read(&calc_load_tasks
);
3034 active
= active
> 0 ? active
* FIXED_1
: 0;
3036 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3037 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3038 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3040 calc_load_update
+= LOAD_FREQ
;
3044 * Either called from update_cpu_load() or from a cpu going idle
3046 static void calc_load_account_active(struct rq
*this_rq
)
3048 long nr_active
, delta
;
3050 nr_active
= this_rq
->nr_running
;
3051 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3053 if (nr_active
!= this_rq
->calc_load_active
) {
3054 delta
= nr_active
- this_rq
->calc_load_active
;
3055 this_rq
->calc_load_active
= nr_active
;
3056 atomic_long_add(delta
, &calc_load_tasks
);
3061 * Externally visible per-cpu scheduler statistics:
3062 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3064 u64
cpu_nr_migrations(int cpu
)
3066 return cpu_rq(cpu
)->nr_migrations_in
;
3070 * Update rq->cpu_load[] statistics. This function is usually called every
3071 * scheduler tick (TICK_NSEC).
3073 static void update_cpu_load(struct rq
*this_rq
)
3075 unsigned long this_load
= this_rq
->load
.weight
;
3078 this_rq
->nr_load_updates
++;
3080 /* Update our load: */
3081 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3082 unsigned long old_load
, new_load
;
3084 /* scale is effectively 1 << i now, and >> i divides by scale */
3086 old_load
= this_rq
->cpu_load
[i
];
3087 new_load
= this_load
;
3089 * Round up the averaging division if load is increasing. This
3090 * prevents us from getting stuck on 9 if the load is 10, for
3093 if (new_load
> old_load
)
3094 new_load
+= scale
-1;
3095 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3098 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3099 this_rq
->calc_load_update
+= LOAD_FREQ
;
3100 calc_load_account_active(this_rq
);
3107 * double_rq_lock - safely lock two runqueues
3109 * Note this does not disable interrupts like task_rq_lock,
3110 * you need to do so manually before calling.
3112 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3113 __acquires(rq1
->lock
)
3114 __acquires(rq2
->lock
)
3116 BUG_ON(!irqs_disabled());
3118 spin_lock(&rq1
->lock
);
3119 __acquire(rq2
->lock
); /* Fake it out ;) */
3122 spin_lock(&rq1
->lock
);
3123 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3125 spin_lock(&rq2
->lock
);
3126 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3129 update_rq_clock(rq1
);
3130 update_rq_clock(rq2
);
3134 * double_rq_unlock - safely unlock two runqueues
3136 * Note this does not restore interrupts like task_rq_unlock,
3137 * you need to do so manually after calling.
3139 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3140 __releases(rq1
->lock
)
3141 __releases(rq2
->lock
)
3143 spin_unlock(&rq1
->lock
);
3145 spin_unlock(&rq2
->lock
);
3147 __release(rq2
->lock
);
3151 * If dest_cpu is allowed for this process, migrate the task to it.
3152 * This is accomplished by forcing the cpu_allowed mask to only
3153 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3154 * the cpu_allowed mask is restored.
3156 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3158 struct migration_req req
;
3159 unsigned long flags
;
3162 rq
= task_rq_lock(p
, &flags
);
3163 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3164 || unlikely(!cpu_active(dest_cpu
)))
3167 /* force the process onto the specified CPU */
3168 if (migrate_task(p
, dest_cpu
, &req
)) {
3169 /* Need to wait for migration thread (might exit: take ref). */
3170 struct task_struct
*mt
= rq
->migration_thread
;
3172 get_task_struct(mt
);
3173 task_rq_unlock(rq
, &flags
);
3174 wake_up_process(mt
);
3175 put_task_struct(mt
);
3176 wait_for_completion(&req
.done
);
3181 task_rq_unlock(rq
, &flags
);
3185 * sched_exec - execve() is a valuable balancing opportunity, because at
3186 * this point the task has the smallest effective memory and cache footprint.
3188 void sched_exec(void)
3190 int new_cpu
, this_cpu
= get_cpu();
3191 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3193 if (new_cpu
!= this_cpu
)
3194 sched_migrate_task(current
, new_cpu
);
3198 * pull_task - move a task from a remote runqueue to the local runqueue.
3199 * Both runqueues must be locked.
3201 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3202 struct rq
*this_rq
, int this_cpu
)
3204 deactivate_task(src_rq
, p
, 0);
3205 set_task_cpu(p
, this_cpu
);
3206 activate_task(this_rq
, p
, 0);
3208 * Note that idle threads have a prio of MAX_PRIO, for this test
3209 * to be always true for them.
3211 check_preempt_curr(this_rq
, p
, 0);
3215 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3218 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3219 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3222 int tsk_cache_hot
= 0;
3224 * We do not migrate tasks that are:
3225 * 1) running (obviously), or
3226 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3227 * 3) are cache-hot on their current CPU.
3229 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3230 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3235 if (task_running(rq
, p
)) {
3236 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3241 * Aggressive migration if:
3242 * 1) task is cache cold, or
3243 * 2) too many balance attempts have failed.
3246 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3247 if (!tsk_cache_hot
||
3248 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3249 #ifdef CONFIG_SCHEDSTATS
3250 if (tsk_cache_hot
) {
3251 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3252 schedstat_inc(p
, se
.nr_forced_migrations
);
3258 if (tsk_cache_hot
) {
3259 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3265 static unsigned long
3266 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3267 unsigned long max_load_move
, struct sched_domain
*sd
,
3268 enum cpu_idle_type idle
, int *all_pinned
,
3269 int *this_best_prio
, struct rq_iterator
*iterator
)
3271 int loops
= 0, pulled
= 0, pinned
= 0;
3272 struct task_struct
*p
;
3273 long rem_load_move
= max_load_move
;
3275 if (max_load_move
== 0)
3281 * Start the load-balancing iterator:
3283 p
= iterator
->start(iterator
->arg
);
3285 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3288 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3289 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3290 p
= iterator
->next(iterator
->arg
);
3294 pull_task(busiest
, p
, this_rq
, this_cpu
);
3296 rem_load_move
-= p
->se
.load
.weight
;
3298 #ifdef CONFIG_PREEMPT
3300 * NEWIDLE balancing is a source of latency, so preemptible kernels
3301 * will stop after the first task is pulled to minimize the critical
3304 if (idle
== CPU_NEWLY_IDLE
)
3309 * We only want to steal up to the prescribed amount of weighted load.
3311 if (rem_load_move
> 0) {
3312 if (p
->prio
< *this_best_prio
)
3313 *this_best_prio
= p
->prio
;
3314 p
= iterator
->next(iterator
->arg
);
3319 * Right now, this is one of only two places pull_task() is called,
3320 * so we can safely collect pull_task() stats here rather than
3321 * inside pull_task().
3323 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3326 *all_pinned
= pinned
;
3328 return max_load_move
- rem_load_move
;
3332 * move_tasks tries to move up to max_load_move weighted load from busiest to
3333 * this_rq, as part of a balancing operation within domain "sd".
3334 * Returns 1 if successful and 0 otherwise.
3336 * Called with both runqueues locked.
3338 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3339 unsigned long max_load_move
,
3340 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3343 const struct sched_class
*class = sched_class_highest
;
3344 unsigned long total_load_moved
= 0;
3345 int this_best_prio
= this_rq
->curr
->prio
;
3349 class->load_balance(this_rq
, this_cpu
, busiest
,
3350 max_load_move
- total_load_moved
,
3351 sd
, idle
, all_pinned
, &this_best_prio
);
3352 class = class->next
;
3354 #ifdef CONFIG_PREEMPT
3356 * NEWIDLE balancing is a source of latency, so preemptible
3357 * kernels will stop after the first task is pulled to minimize
3358 * the critical section.
3360 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3363 } while (class && max_load_move
> total_load_moved
);
3365 return total_load_moved
> 0;
3369 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3370 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3371 struct rq_iterator
*iterator
)
3373 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3377 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3378 pull_task(busiest
, p
, this_rq
, this_cpu
);
3380 * Right now, this is only the second place pull_task()
3381 * is called, so we can safely collect pull_task()
3382 * stats here rather than inside pull_task().
3384 schedstat_inc(sd
, lb_gained
[idle
]);
3388 p
= iterator
->next(iterator
->arg
);
3395 * move_one_task tries to move exactly one task from busiest to this_rq, as
3396 * part of active balancing operations within "domain".
3397 * Returns 1 if successful and 0 otherwise.
3399 * Called with both runqueues locked.
3401 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3402 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3404 const struct sched_class
*class;
3406 for (class = sched_class_highest
; class; class = class->next
)
3407 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3412 /********** Helpers for find_busiest_group ************************/
3414 * sd_lb_stats - Structure to store the statistics of a sched_domain
3415 * during load balancing.
3417 struct sd_lb_stats
{
3418 struct sched_group
*busiest
; /* Busiest group in this sd */
3419 struct sched_group
*this; /* Local group in this sd */
3420 unsigned long total_load
; /* Total load of all groups in sd */
3421 unsigned long total_pwr
; /* Total power of all groups in sd */
3422 unsigned long avg_load
; /* Average load across all groups in sd */
3424 /** Statistics of this group */
3425 unsigned long this_load
;
3426 unsigned long this_load_per_task
;
3427 unsigned long this_nr_running
;
3429 /* Statistics of the busiest group */
3430 unsigned long max_load
;
3431 unsigned long busiest_load_per_task
;
3432 unsigned long busiest_nr_running
;
3434 int group_imb
; /* Is there imbalance in this sd */
3435 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3436 int power_savings_balance
; /* Is powersave balance needed for this sd */
3437 struct sched_group
*group_min
; /* Least loaded group in sd */
3438 struct sched_group
*group_leader
; /* Group which relieves group_min */
3439 unsigned long min_load_per_task
; /* load_per_task in group_min */
3440 unsigned long leader_nr_running
; /* Nr running of group_leader */
3441 unsigned long min_nr_running
; /* Nr running of group_min */
3446 * sg_lb_stats - stats of a sched_group required for load_balancing
3448 struct sg_lb_stats
{
3449 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3450 unsigned long group_load
; /* Total load over the CPUs of the group */
3451 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3452 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3453 unsigned long group_capacity
;
3454 int group_imb
; /* Is there an imbalance in the group ? */
3458 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3459 * @group: The group whose first cpu is to be returned.
3461 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3463 return cpumask_first(sched_group_cpus(group
));
3467 * get_sd_load_idx - Obtain the load index for a given sched domain.
3468 * @sd: The sched_domain whose load_idx is to be obtained.
3469 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3471 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3472 enum cpu_idle_type idle
)
3478 load_idx
= sd
->busy_idx
;
3481 case CPU_NEWLY_IDLE
:
3482 load_idx
= sd
->newidle_idx
;
3485 load_idx
= sd
->idle_idx
;
3493 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3495 * init_sd_power_savings_stats - Initialize power savings statistics for
3496 * the given sched_domain, during load balancing.
3498 * @sd: Sched domain whose power-savings statistics are to be initialized.
3499 * @sds: Variable containing the statistics for sd.
3500 * @idle: Idle status of the CPU at which we're performing load-balancing.
3502 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3503 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3506 * Busy processors will not participate in power savings
3509 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3510 sds
->power_savings_balance
= 0;
3512 sds
->power_savings_balance
= 1;
3513 sds
->min_nr_running
= ULONG_MAX
;
3514 sds
->leader_nr_running
= 0;
3519 * update_sd_power_savings_stats - Update the power saving stats for a
3520 * sched_domain while performing load balancing.
3522 * @group: sched_group belonging to the sched_domain under consideration.
3523 * @sds: Variable containing the statistics of the sched_domain
3524 * @local_group: Does group contain the CPU for which we're performing
3526 * @sgs: Variable containing the statistics of the group.
3528 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3529 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3532 if (!sds
->power_savings_balance
)
3536 * If the local group is idle or completely loaded
3537 * no need to do power savings balance at this domain
3539 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3540 !sds
->this_nr_running
))
3541 sds
->power_savings_balance
= 0;
3544 * If a group is already running at full capacity or idle,
3545 * don't include that group in power savings calculations
3547 if (!sds
->power_savings_balance
||
3548 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3549 !sgs
->sum_nr_running
)
3553 * Calculate the group which has the least non-idle load.
3554 * This is the group from where we need to pick up the load
3557 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3558 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3559 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3560 sds
->group_min
= group
;
3561 sds
->min_nr_running
= sgs
->sum_nr_running
;
3562 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3563 sgs
->sum_nr_running
;
3567 * Calculate the group which is almost near its
3568 * capacity but still has some space to pick up some load
3569 * from other group and save more power
3571 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3574 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3575 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3576 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3577 sds
->group_leader
= group
;
3578 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3583 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3584 * @sds: Variable containing the statistics of the sched_domain
3585 * under consideration.
3586 * @this_cpu: Cpu at which we're currently performing load-balancing.
3587 * @imbalance: Variable to store the imbalance.
3590 * Check if we have potential to perform some power-savings balance.
3591 * If yes, set the busiest group to be the least loaded group in the
3592 * sched_domain, so that it's CPUs can be put to idle.
3594 * Returns 1 if there is potential to perform power-savings balance.
3597 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3598 int this_cpu
, unsigned long *imbalance
)
3600 if (!sds
->power_savings_balance
)
3603 if (sds
->this != sds
->group_leader
||
3604 sds
->group_leader
== sds
->group_min
)
3607 *imbalance
= sds
->min_load_per_task
;
3608 sds
->busiest
= sds
->group_min
;
3610 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3611 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3612 group_first_cpu(sds
->group_leader
);
3618 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3619 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3620 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3625 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3626 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3631 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3632 int this_cpu
, unsigned long *imbalance
)
3636 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3640 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3641 * @group: sched_group whose statistics are to be updated.
3642 * @this_cpu: Cpu for which load balance is currently performed.
3643 * @idle: Idle status of this_cpu
3644 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3645 * @sd_idle: Idle status of the sched_domain containing group.
3646 * @local_group: Does group contain this_cpu.
3647 * @cpus: Set of cpus considered for load balancing.
3648 * @balance: Should we balance.
3649 * @sgs: variable to hold the statistics for this group.
3651 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3652 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3653 int local_group
, const struct cpumask
*cpus
,
3654 int *balance
, struct sg_lb_stats
*sgs
)
3656 unsigned long load
, max_cpu_load
, min_cpu_load
;
3658 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3659 unsigned long sum_avg_load_per_task
;
3660 unsigned long avg_load_per_task
;
3663 balance_cpu
= group_first_cpu(group
);
3665 /* Tally up the load of all CPUs in the group */
3666 sum_avg_load_per_task
= avg_load_per_task
= 0;
3668 min_cpu_load
= ~0UL;
3670 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3671 struct rq
*rq
= cpu_rq(i
);
3673 if (*sd_idle
&& rq
->nr_running
)
3676 /* Bias balancing toward cpus of our domain */
3678 if (idle_cpu(i
) && !first_idle_cpu
) {
3683 load
= target_load(i
, load_idx
);
3685 load
= source_load(i
, load_idx
);
3686 if (load
> max_cpu_load
)
3687 max_cpu_load
= load
;
3688 if (min_cpu_load
> load
)
3689 min_cpu_load
= load
;
3692 sgs
->group_load
+= load
;
3693 sgs
->sum_nr_running
+= rq
->nr_running
;
3694 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3696 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3700 * First idle cpu or the first cpu(busiest) in this sched group
3701 * is eligible for doing load balancing at this and above
3702 * domains. In the newly idle case, we will allow all the cpu's
3703 * to do the newly idle load balance.
3705 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3706 balance_cpu
!= this_cpu
&& balance
) {
3711 /* Adjust by relative CPU power of the group */
3712 sgs
->avg_load
= sg_div_cpu_power(group
,
3713 sgs
->group_load
* SCHED_LOAD_SCALE
);
3717 * Consider the group unbalanced when the imbalance is larger
3718 * than the average weight of two tasks.
3720 * APZ: with cgroup the avg task weight can vary wildly and
3721 * might not be a suitable number - should we keep a
3722 * normalized nr_running number somewhere that negates
3725 avg_load_per_task
= sg_div_cpu_power(group
,
3726 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3728 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3731 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3736 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3737 * @sd: sched_domain whose statistics are to be updated.
3738 * @this_cpu: Cpu for which load balance is currently performed.
3739 * @idle: Idle status of this_cpu
3740 * @sd_idle: Idle status of the sched_domain containing group.
3741 * @cpus: Set of cpus considered for load balancing.
3742 * @balance: Should we balance.
3743 * @sds: variable to hold the statistics for this sched_domain.
3745 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3746 enum cpu_idle_type idle
, int *sd_idle
,
3747 const struct cpumask
*cpus
, int *balance
,
3748 struct sd_lb_stats
*sds
)
3750 struct sched_group
*group
= sd
->groups
;
3751 struct sg_lb_stats sgs
;
3754 init_sd_power_savings_stats(sd
, sds
, idle
);
3755 load_idx
= get_sd_load_idx(sd
, idle
);
3760 local_group
= cpumask_test_cpu(this_cpu
,
3761 sched_group_cpus(group
));
3762 memset(&sgs
, 0, sizeof(sgs
));
3763 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3764 local_group
, cpus
, balance
, &sgs
);
3766 if (local_group
&& balance
&& !(*balance
))
3769 sds
->total_load
+= sgs
.group_load
;
3770 sds
->total_pwr
+= group
->__cpu_power
;
3773 sds
->this_load
= sgs
.avg_load
;
3775 sds
->this_nr_running
= sgs
.sum_nr_running
;
3776 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3777 } else if (sgs
.avg_load
> sds
->max_load
&&
3778 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3780 sds
->max_load
= sgs
.avg_load
;
3781 sds
->busiest
= group
;
3782 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3783 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3784 sds
->group_imb
= sgs
.group_imb
;
3787 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3788 group
= group
->next
;
3789 } while (group
!= sd
->groups
);
3794 * fix_small_imbalance - Calculate the minor imbalance that exists
3795 * amongst the groups of a sched_domain, during
3797 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3798 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3799 * @imbalance: Variable to store the imbalance.
3801 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3802 int this_cpu
, unsigned long *imbalance
)
3804 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3805 unsigned int imbn
= 2;
3807 if (sds
->this_nr_running
) {
3808 sds
->this_load_per_task
/= sds
->this_nr_running
;
3809 if (sds
->busiest_load_per_task
>
3810 sds
->this_load_per_task
)
3813 sds
->this_load_per_task
=
3814 cpu_avg_load_per_task(this_cpu
);
3816 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3817 sds
->busiest_load_per_task
* imbn
) {
3818 *imbalance
= sds
->busiest_load_per_task
;
3823 * OK, we don't have enough imbalance to justify moving tasks,
3824 * however we may be able to increase total CPU power used by
3828 pwr_now
+= sds
->busiest
->__cpu_power
*
3829 min(sds
->busiest_load_per_task
, sds
->max_load
);
3830 pwr_now
+= sds
->this->__cpu_power
*
3831 min(sds
->this_load_per_task
, sds
->this_load
);
3832 pwr_now
/= SCHED_LOAD_SCALE
;
3834 /* Amount of load we'd subtract */
3835 tmp
= sg_div_cpu_power(sds
->busiest
,
3836 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3837 if (sds
->max_load
> tmp
)
3838 pwr_move
+= sds
->busiest
->__cpu_power
*
3839 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3841 /* Amount of load we'd add */
3842 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3843 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3844 tmp
= sg_div_cpu_power(sds
->this,
3845 sds
->max_load
* sds
->busiest
->__cpu_power
);
3847 tmp
= sg_div_cpu_power(sds
->this,
3848 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3849 pwr_move
+= sds
->this->__cpu_power
*
3850 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3851 pwr_move
/= SCHED_LOAD_SCALE
;
3853 /* Move if we gain throughput */
3854 if (pwr_move
> pwr_now
)
3855 *imbalance
= sds
->busiest_load_per_task
;
3859 * calculate_imbalance - Calculate the amount of imbalance present within the
3860 * groups of a given sched_domain during load balance.
3861 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3862 * @this_cpu: Cpu for which currently load balance is being performed.
3863 * @imbalance: The variable to store the imbalance.
3865 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3866 unsigned long *imbalance
)
3868 unsigned long max_pull
;
3870 * In the presence of smp nice balancing, certain scenarios can have
3871 * max load less than avg load(as we skip the groups at or below
3872 * its cpu_power, while calculating max_load..)
3874 if (sds
->max_load
< sds
->avg_load
) {
3876 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3879 /* Don't want to pull so many tasks that a group would go idle */
3880 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3881 sds
->max_load
- sds
->busiest_load_per_task
);
3883 /* How much load to actually move to equalise the imbalance */
3884 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3885 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3889 * if *imbalance is less than the average load per runnable task
3890 * there is no gaurantee that any tasks will be moved so we'll have
3891 * a think about bumping its value to force at least one task to be
3894 if (*imbalance
< sds
->busiest_load_per_task
)
3895 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3898 /******* find_busiest_group() helpers end here *********************/
3901 * find_busiest_group - Returns the busiest group within the sched_domain
3902 * if there is an imbalance. If there isn't an imbalance, and
3903 * the user has opted for power-savings, it returns a group whose
3904 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3905 * such a group exists.
3907 * Also calculates the amount of weighted load which should be moved
3908 * to restore balance.
3910 * @sd: The sched_domain whose busiest group is to be returned.
3911 * @this_cpu: The cpu for which load balancing is currently being performed.
3912 * @imbalance: Variable which stores amount of weighted load which should
3913 * be moved to restore balance/put a group to idle.
3914 * @idle: The idle status of this_cpu.
3915 * @sd_idle: The idleness of sd
3916 * @cpus: The set of CPUs under consideration for load-balancing.
3917 * @balance: Pointer to a variable indicating if this_cpu
3918 * is the appropriate cpu to perform load balancing at this_level.
3920 * Returns: - the busiest group if imbalance exists.
3921 * - If no imbalance and user has opted for power-savings balance,
3922 * return the least loaded group whose CPUs can be
3923 * put to idle by rebalancing its tasks onto our group.
3925 static struct sched_group
*
3926 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3927 unsigned long *imbalance
, enum cpu_idle_type idle
,
3928 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3930 struct sd_lb_stats sds
;
3932 memset(&sds
, 0, sizeof(sds
));
3935 * Compute the various statistics relavent for load balancing at
3938 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3941 /* Cases where imbalance does not exist from POV of this_cpu */
3942 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3944 * 2) There is no busy sibling group to pull from.
3945 * 3) This group is the busiest group.
3946 * 4) This group is more busy than the avg busieness at this
3948 * 5) The imbalance is within the specified limit.
3949 * 6) Any rebalance would lead to ping-pong
3951 if (balance
&& !(*balance
))
3954 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3957 if (sds
.this_load
>= sds
.max_load
)
3960 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3962 if (sds
.this_load
>= sds
.avg_load
)
3965 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3968 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3970 sds
.busiest_load_per_task
=
3971 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3974 * We're trying to get all the cpus to the average_load, so we don't
3975 * want to push ourselves above the average load, nor do we wish to
3976 * reduce the max loaded cpu below the average load, as either of these
3977 * actions would just result in more rebalancing later, and ping-pong
3978 * tasks around. Thus we look for the minimum possible imbalance.
3979 * Negative imbalances (*we* are more loaded than anyone else) will
3980 * be counted as no imbalance for these purposes -- we can't fix that
3981 * by pulling tasks to us. Be careful of negative numbers as they'll
3982 * appear as very large values with unsigned longs.
3984 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3987 /* Looks like there is an imbalance. Compute it */
3988 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3993 * There is no obvious imbalance. But check if we can do some balancing
3996 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4004 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4007 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4008 unsigned long imbalance
, const struct cpumask
*cpus
)
4010 struct rq
*busiest
= NULL
, *rq
;
4011 unsigned long max_load
= 0;
4014 for_each_cpu(i
, sched_group_cpus(group
)) {
4017 if (!cpumask_test_cpu(i
, cpus
))
4021 wl
= weighted_cpuload(i
);
4023 if (rq
->nr_running
== 1 && wl
> imbalance
)
4026 if (wl
> max_load
) {
4036 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4037 * so long as it is large enough.
4039 #define MAX_PINNED_INTERVAL 512
4041 /* Working cpumask for load_balance and load_balance_newidle. */
4042 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4045 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4046 * tasks if there is an imbalance.
4048 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4049 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4052 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4053 struct sched_group
*group
;
4054 unsigned long imbalance
;
4056 unsigned long flags
;
4057 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4059 cpumask_setall(cpus
);
4062 * When power savings policy is enabled for the parent domain, idle
4063 * sibling can pick up load irrespective of busy siblings. In this case,
4064 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4065 * portraying it as CPU_NOT_IDLE.
4067 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4068 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4071 schedstat_inc(sd
, lb_count
[idle
]);
4075 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4082 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4086 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4088 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4092 BUG_ON(busiest
== this_rq
);
4094 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4097 if (busiest
->nr_running
> 1) {
4099 * Attempt to move tasks. If find_busiest_group has found
4100 * an imbalance but busiest->nr_running <= 1, the group is
4101 * still unbalanced. ld_moved simply stays zero, so it is
4102 * correctly treated as an imbalance.
4104 local_irq_save(flags
);
4105 double_rq_lock(this_rq
, busiest
);
4106 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4107 imbalance
, sd
, idle
, &all_pinned
);
4108 double_rq_unlock(this_rq
, busiest
);
4109 local_irq_restore(flags
);
4112 * some other cpu did the load balance for us.
4114 if (ld_moved
&& this_cpu
!= smp_processor_id())
4115 resched_cpu(this_cpu
);
4117 /* All tasks on this runqueue were pinned by CPU affinity */
4118 if (unlikely(all_pinned
)) {
4119 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4120 if (!cpumask_empty(cpus
))
4127 schedstat_inc(sd
, lb_failed
[idle
]);
4128 sd
->nr_balance_failed
++;
4130 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4132 spin_lock_irqsave(&busiest
->lock
, flags
);
4134 /* don't kick the migration_thread, if the curr
4135 * task on busiest cpu can't be moved to this_cpu
4137 if (!cpumask_test_cpu(this_cpu
,
4138 &busiest
->curr
->cpus_allowed
)) {
4139 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4141 goto out_one_pinned
;
4144 if (!busiest
->active_balance
) {
4145 busiest
->active_balance
= 1;
4146 busiest
->push_cpu
= this_cpu
;
4149 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4151 wake_up_process(busiest
->migration_thread
);
4154 * We've kicked active balancing, reset the failure
4157 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4160 sd
->nr_balance_failed
= 0;
4162 if (likely(!active_balance
)) {
4163 /* We were unbalanced, so reset the balancing interval */
4164 sd
->balance_interval
= sd
->min_interval
;
4167 * If we've begun active balancing, start to back off. This
4168 * case may not be covered by the all_pinned logic if there
4169 * is only 1 task on the busy runqueue (because we don't call
4172 if (sd
->balance_interval
< sd
->max_interval
)
4173 sd
->balance_interval
*= 2;
4176 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4177 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4183 schedstat_inc(sd
, lb_balanced
[idle
]);
4185 sd
->nr_balance_failed
= 0;
4188 /* tune up the balancing interval */
4189 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4190 (sd
->balance_interval
< sd
->max_interval
))
4191 sd
->balance_interval
*= 2;
4193 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4194 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4205 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4206 * tasks if there is an imbalance.
4208 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4209 * this_rq is locked.
4212 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4214 struct sched_group
*group
;
4215 struct rq
*busiest
= NULL
;
4216 unsigned long imbalance
;
4220 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4222 cpumask_setall(cpus
);
4225 * When power savings policy is enabled for the parent domain, idle
4226 * sibling can pick up load irrespective of busy siblings. In this case,
4227 * let the state of idle sibling percolate up as IDLE, instead of
4228 * portraying it as CPU_NOT_IDLE.
4230 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4231 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4234 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4236 update_shares_locked(this_rq
, sd
);
4237 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4238 &sd_idle
, cpus
, NULL
);
4240 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4244 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4246 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4250 BUG_ON(busiest
== this_rq
);
4252 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4255 if (busiest
->nr_running
> 1) {
4256 /* Attempt to move tasks */
4257 double_lock_balance(this_rq
, busiest
);
4258 /* this_rq->clock is already updated */
4259 update_rq_clock(busiest
);
4260 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4261 imbalance
, sd
, CPU_NEWLY_IDLE
,
4263 double_unlock_balance(this_rq
, busiest
);
4265 if (unlikely(all_pinned
)) {
4266 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4267 if (!cpumask_empty(cpus
))
4273 int active_balance
= 0;
4275 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4276 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4277 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4280 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4283 if (sd
->nr_balance_failed
++ < 2)
4287 * The only task running in a non-idle cpu can be moved to this
4288 * cpu in an attempt to completely freeup the other CPU
4289 * package. The same method used to move task in load_balance()
4290 * have been extended for load_balance_newidle() to speedup
4291 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4293 * The package power saving logic comes from
4294 * find_busiest_group(). If there are no imbalance, then
4295 * f_b_g() will return NULL. However when sched_mc={1,2} then
4296 * f_b_g() will select a group from which a running task may be
4297 * pulled to this cpu in order to make the other package idle.
4298 * If there is no opportunity to make a package idle and if
4299 * there are no imbalance, then f_b_g() will return NULL and no
4300 * action will be taken in load_balance_newidle().
4302 * Under normal task pull operation due to imbalance, there
4303 * will be more than one task in the source run queue and
4304 * move_tasks() will succeed. ld_moved will be true and this
4305 * active balance code will not be triggered.
4308 /* Lock busiest in correct order while this_rq is held */
4309 double_lock_balance(this_rq
, busiest
);
4312 * don't kick the migration_thread, if the curr
4313 * task on busiest cpu can't be moved to this_cpu
4315 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4316 double_unlock_balance(this_rq
, busiest
);
4321 if (!busiest
->active_balance
) {
4322 busiest
->active_balance
= 1;
4323 busiest
->push_cpu
= this_cpu
;
4327 double_unlock_balance(this_rq
, busiest
);
4329 * Should not call ttwu while holding a rq->lock
4331 spin_unlock(&this_rq
->lock
);
4333 wake_up_process(busiest
->migration_thread
);
4334 spin_lock(&this_rq
->lock
);
4337 sd
->nr_balance_failed
= 0;
4339 update_shares_locked(this_rq
, sd
);
4343 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4344 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4345 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4347 sd
->nr_balance_failed
= 0;
4353 * idle_balance is called by schedule() if this_cpu is about to become
4354 * idle. Attempts to pull tasks from other CPUs.
4356 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4358 struct sched_domain
*sd
;
4359 int pulled_task
= 0;
4360 unsigned long next_balance
= jiffies
+ HZ
;
4362 for_each_domain(this_cpu
, sd
) {
4363 unsigned long interval
;
4365 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4368 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4369 /* If we've pulled tasks over stop searching: */
4370 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4373 interval
= msecs_to_jiffies(sd
->balance_interval
);
4374 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4375 next_balance
= sd
->last_balance
+ interval
;
4379 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4381 * We are going idle. next_balance may be set based on
4382 * a busy processor. So reset next_balance.
4384 this_rq
->next_balance
= next_balance
;
4389 * active_load_balance is run by migration threads. It pushes running tasks
4390 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4391 * running on each physical CPU where possible, and avoids physical /
4392 * logical imbalances.
4394 * Called with busiest_rq locked.
4396 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4398 int target_cpu
= busiest_rq
->push_cpu
;
4399 struct sched_domain
*sd
;
4400 struct rq
*target_rq
;
4402 /* Is there any task to move? */
4403 if (busiest_rq
->nr_running
<= 1)
4406 target_rq
= cpu_rq(target_cpu
);
4409 * This condition is "impossible", if it occurs
4410 * we need to fix it. Originally reported by
4411 * Bjorn Helgaas on a 128-cpu setup.
4413 BUG_ON(busiest_rq
== target_rq
);
4415 /* move a task from busiest_rq to target_rq */
4416 double_lock_balance(busiest_rq
, target_rq
);
4417 update_rq_clock(busiest_rq
);
4418 update_rq_clock(target_rq
);
4420 /* Search for an sd spanning us and the target CPU. */
4421 for_each_domain(target_cpu
, sd
) {
4422 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4423 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4428 schedstat_inc(sd
, alb_count
);
4430 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4432 schedstat_inc(sd
, alb_pushed
);
4434 schedstat_inc(sd
, alb_failed
);
4436 double_unlock_balance(busiest_rq
, target_rq
);
4441 atomic_t load_balancer
;
4442 cpumask_var_t cpu_mask
;
4443 cpumask_var_t ilb_grp_nohz_mask
;
4444 } nohz ____cacheline_aligned
= {
4445 .load_balancer
= ATOMIC_INIT(-1),
4448 int get_nohz_load_balancer(void)
4450 return atomic_read(&nohz
.load_balancer
);
4453 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4455 * lowest_flag_domain - Return lowest sched_domain containing flag.
4456 * @cpu: The cpu whose lowest level of sched domain is to
4458 * @flag: The flag to check for the lowest sched_domain
4459 * for the given cpu.
4461 * Returns the lowest sched_domain of a cpu which contains the given flag.
4463 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4465 struct sched_domain
*sd
;
4467 for_each_domain(cpu
, sd
)
4468 if (sd
&& (sd
->flags
& flag
))
4475 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4476 * @cpu: The cpu whose domains we're iterating over.
4477 * @sd: variable holding the value of the power_savings_sd
4479 * @flag: The flag to filter the sched_domains to be iterated.
4481 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4482 * set, starting from the lowest sched_domain to the highest.
4484 #define for_each_flag_domain(cpu, sd, flag) \
4485 for (sd = lowest_flag_domain(cpu, flag); \
4486 (sd && (sd->flags & flag)); sd = sd->parent)
4489 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4490 * @ilb_group: group to be checked for semi-idleness
4492 * Returns: 1 if the group is semi-idle. 0 otherwise.
4494 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4495 * and atleast one non-idle CPU. This helper function checks if the given
4496 * sched_group is semi-idle or not.
4498 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4500 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4501 sched_group_cpus(ilb_group
));
4504 * A sched_group is semi-idle when it has atleast one busy cpu
4505 * and atleast one idle cpu.
4507 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4510 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4516 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4517 * @cpu: The cpu which is nominating a new idle_load_balancer.
4519 * Returns: Returns the id of the idle load balancer if it exists,
4520 * Else, returns >= nr_cpu_ids.
4522 * This algorithm picks the idle load balancer such that it belongs to a
4523 * semi-idle powersavings sched_domain. The idea is to try and avoid
4524 * completely idle packages/cores just for the purpose of idle load balancing
4525 * when there are other idle cpu's which are better suited for that job.
4527 static int find_new_ilb(int cpu
)
4529 struct sched_domain
*sd
;
4530 struct sched_group
*ilb_group
;
4533 * Have idle load balancer selection from semi-idle packages only
4534 * when power-aware load balancing is enabled
4536 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4540 * Optimize for the case when we have no idle CPUs or only one
4541 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4543 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4546 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4547 ilb_group
= sd
->groups
;
4550 if (is_semi_idle_group(ilb_group
))
4551 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4553 ilb_group
= ilb_group
->next
;
4555 } while (ilb_group
!= sd
->groups
);
4559 return cpumask_first(nohz
.cpu_mask
);
4561 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4562 static inline int find_new_ilb(int call_cpu
)
4564 return cpumask_first(nohz
.cpu_mask
);
4569 * This routine will try to nominate the ilb (idle load balancing)
4570 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4571 * load balancing on behalf of all those cpus. If all the cpus in the system
4572 * go into this tickless mode, then there will be no ilb owner (as there is
4573 * no need for one) and all the cpus will sleep till the next wakeup event
4576 * For the ilb owner, tick is not stopped. And this tick will be used
4577 * for idle load balancing. ilb owner will still be part of
4580 * While stopping the tick, this cpu will become the ilb owner if there
4581 * is no other owner. And will be the owner till that cpu becomes busy
4582 * or if all cpus in the system stop their ticks at which point
4583 * there is no need for ilb owner.
4585 * When the ilb owner becomes busy, it nominates another owner, during the
4586 * next busy scheduler_tick()
4588 int select_nohz_load_balancer(int stop_tick
)
4590 int cpu
= smp_processor_id();
4593 cpu_rq(cpu
)->in_nohz_recently
= 1;
4595 if (!cpu_active(cpu
)) {
4596 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4600 * If we are going offline and still the leader,
4603 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4609 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4611 /* time for ilb owner also to sleep */
4612 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4613 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4614 atomic_set(&nohz
.load_balancer
, -1);
4618 if (atomic_read(&nohz
.load_balancer
) == -1) {
4619 /* make me the ilb owner */
4620 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4622 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4625 if (!(sched_smt_power_savings
||
4626 sched_mc_power_savings
))
4629 * Check to see if there is a more power-efficient
4632 new_ilb
= find_new_ilb(cpu
);
4633 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4634 atomic_set(&nohz
.load_balancer
, -1);
4635 resched_cpu(new_ilb
);
4641 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4644 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4646 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4647 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4654 static DEFINE_SPINLOCK(balancing
);
4657 * It checks each scheduling domain to see if it is due to be balanced,
4658 * and initiates a balancing operation if so.
4660 * Balancing parameters are set up in arch_init_sched_domains.
4662 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4665 struct rq
*rq
= cpu_rq(cpu
);
4666 unsigned long interval
;
4667 struct sched_domain
*sd
;
4668 /* Earliest time when we have to do rebalance again */
4669 unsigned long next_balance
= jiffies
+ 60*HZ
;
4670 int update_next_balance
= 0;
4673 for_each_domain(cpu
, sd
) {
4674 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4677 interval
= sd
->balance_interval
;
4678 if (idle
!= CPU_IDLE
)
4679 interval
*= sd
->busy_factor
;
4681 /* scale ms to jiffies */
4682 interval
= msecs_to_jiffies(interval
);
4683 if (unlikely(!interval
))
4685 if (interval
> HZ
*NR_CPUS
/10)
4686 interval
= HZ
*NR_CPUS
/10;
4688 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4690 if (need_serialize
) {
4691 if (!spin_trylock(&balancing
))
4695 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4696 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4698 * We've pulled tasks over so either we're no
4699 * longer idle, or one of our SMT siblings is
4702 idle
= CPU_NOT_IDLE
;
4704 sd
->last_balance
= jiffies
;
4707 spin_unlock(&balancing
);
4709 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4710 next_balance
= sd
->last_balance
+ interval
;
4711 update_next_balance
= 1;
4715 * Stop the load balance at this level. There is another
4716 * CPU in our sched group which is doing load balancing more
4724 * next_balance will be updated only when there is a need.
4725 * When the cpu is attached to null domain for ex, it will not be
4728 if (likely(update_next_balance
))
4729 rq
->next_balance
= next_balance
;
4733 * run_rebalance_domains is triggered when needed from the scheduler tick.
4734 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4735 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4737 static void run_rebalance_domains(struct softirq_action
*h
)
4739 int this_cpu
= smp_processor_id();
4740 struct rq
*this_rq
= cpu_rq(this_cpu
);
4741 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4742 CPU_IDLE
: CPU_NOT_IDLE
;
4744 rebalance_domains(this_cpu
, idle
);
4748 * If this cpu is the owner for idle load balancing, then do the
4749 * balancing on behalf of the other idle cpus whose ticks are
4752 if (this_rq
->idle_at_tick
&&
4753 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4757 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4758 if (balance_cpu
== this_cpu
)
4762 * If this cpu gets work to do, stop the load balancing
4763 * work being done for other cpus. Next load
4764 * balancing owner will pick it up.
4769 rebalance_domains(balance_cpu
, CPU_IDLE
);
4771 rq
= cpu_rq(balance_cpu
);
4772 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4773 this_rq
->next_balance
= rq
->next_balance
;
4779 static inline int on_null_domain(int cpu
)
4781 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4785 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4787 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4788 * idle load balancing owner or decide to stop the periodic load balancing,
4789 * if the whole system is idle.
4791 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4795 * If we were in the nohz mode recently and busy at the current
4796 * scheduler tick, then check if we need to nominate new idle
4799 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4800 rq
->in_nohz_recently
= 0;
4802 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4803 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4804 atomic_set(&nohz
.load_balancer
, -1);
4807 if (atomic_read(&nohz
.load_balancer
) == -1) {
4808 int ilb
= find_new_ilb(cpu
);
4810 if (ilb
< nr_cpu_ids
)
4816 * If this cpu is idle and doing idle load balancing for all the
4817 * cpus with ticks stopped, is it time for that to stop?
4819 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4820 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4826 * If this cpu is idle and the idle load balancing is done by
4827 * someone else, then no need raise the SCHED_SOFTIRQ
4829 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4830 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4833 /* Don't need to rebalance while attached to NULL domain */
4834 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4835 likely(!on_null_domain(cpu
)))
4836 raise_softirq(SCHED_SOFTIRQ
);
4839 #else /* CONFIG_SMP */
4842 * on UP we do not need to balance between CPUs:
4844 static inline void idle_balance(int cpu
, struct rq
*rq
)
4850 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4852 EXPORT_PER_CPU_SYMBOL(kstat
);
4855 * Return any ns on the sched_clock that have not yet been accounted in
4856 * @p in case that task is currently running.
4858 * Called with task_rq_lock() held on @rq.
4860 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4864 if (task_current(rq
, p
)) {
4865 update_rq_clock(rq
);
4866 ns
= rq
->clock
- p
->se
.exec_start
;
4874 unsigned long long task_delta_exec(struct task_struct
*p
)
4876 unsigned long flags
;
4880 rq
= task_rq_lock(p
, &flags
);
4881 ns
= do_task_delta_exec(p
, rq
);
4882 task_rq_unlock(rq
, &flags
);
4888 * Return accounted runtime for the task.
4889 * In case the task is currently running, return the runtime plus current's
4890 * pending runtime that have not been accounted yet.
4892 unsigned long long task_sched_runtime(struct task_struct
*p
)
4894 unsigned long flags
;
4898 rq
= task_rq_lock(p
, &flags
);
4899 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4900 task_rq_unlock(rq
, &flags
);
4906 * Return sum_exec_runtime for the thread group.
4907 * In case the task is currently running, return the sum plus current's
4908 * pending runtime that have not been accounted yet.
4910 * Note that the thread group might have other running tasks as well,
4911 * so the return value not includes other pending runtime that other
4912 * running tasks might have.
4914 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4916 struct task_cputime totals
;
4917 unsigned long flags
;
4921 rq
= task_rq_lock(p
, &flags
);
4922 thread_group_cputime(p
, &totals
);
4923 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4924 task_rq_unlock(rq
, &flags
);
4930 * Account user cpu time to a process.
4931 * @p: the process that the cpu time gets accounted to
4932 * @cputime: the cpu time spent in user space since the last update
4933 * @cputime_scaled: cputime scaled by cpu frequency
4935 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4936 cputime_t cputime_scaled
)
4938 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4941 /* Add user time to process. */
4942 p
->utime
= cputime_add(p
->utime
, cputime
);
4943 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4944 account_group_user_time(p
, cputime
);
4946 /* Add user time to cpustat. */
4947 tmp
= cputime_to_cputime64(cputime
);
4948 if (TASK_NICE(p
) > 0)
4949 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4951 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4953 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4954 /* Account for user time used */
4955 acct_update_integrals(p
);
4959 * Account guest cpu time to a process.
4960 * @p: the process that the cpu time gets accounted to
4961 * @cputime: the cpu time spent in virtual machine since the last update
4962 * @cputime_scaled: cputime scaled by cpu frequency
4964 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4965 cputime_t cputime_scaled
)
4968 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4970 tmp
= cputime_to_cputime64(cputime
);
4972 /* Add guest time to process. */
4973 p
->utime
= cputime_add(p
->utime
, cputime
);
4974 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4975 account_group_user_time(p
, cputime
);
4976 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4978 /* Add guest time to cpustat. */
4979 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4980 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4984 * Account system cpu time to a process.
4985 * @p: the process that the cpu time gets accounted to
4986 * @hardirq_offset: the offset to subtract from hardirq_count()
4987 * @cputime: the cpu time spent in kernel space since the last update
4988 * @cputime_scaled: cputime scaled by cpu frequency
4990 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4991 cputime_t cputime
, cputime_t cputime_scaled
)
4993 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4996 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4997 account_guest_time(p
, cputime
, cputime_scaled
);
5001 /* Add system time to process. */
5002 p
->stime
= cputime_add(p
->stime
, cputime
);
5003 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5004 account_group_system_time(p
, cputime
);
5006 /* Add system time to cpustat. */
5007 tmp
= cputime_to_cputime64(cputime
);
5008 if (hardirq_count() - hardirq_offset
)
5009 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5010 else if (softirq_count())
5011 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5013 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5015 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5017 /* Account for system time used */
5018 acct_update_integrals(p
);
5022 * Account for involuntary wait time.
5023 * @steal: the cpu time spent in involuntary wait
5025 void account_steal_time(cputime_t cputime
)
5027 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5028 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5030 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5034 * Account for idle time.
5035 * @cputime: the cpu time spent in idle wait
5037 void account_idle_time(cputime_t cputime
)
5039 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5040 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5041 struct rq
*rq
= this_rq();
5043 if (atomic_read(&rq
->nr_iowait
) > 0)
5044 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5046 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5049 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5052 * Account a single tick of cpu time.
5053 * @p: the process that the cpu time gets accounted to
5054 * @user_tick: indicates if the tick is a user or a system tick
5056 void account_process_tick(struct task_struct
*p
, int user_tick
)
5058 cputime_t one_jiffy
= jiffies_to_cputime(1);
5059 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5060 struct rq
*rq
= this_rq();
5063 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5064 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5065 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5068 account_idle_time(one_jiffy
);
5072 * Account multiple ticks of steal time.
5073 * @p: the process from which the cpu time has been stolen
5074 * @ticks: number of stolen ticks
5076 void account_steal_ticks(unsigned long ticks
)
5078 account_steal_time(jiffies_to_cputime(ticks
));
5082 * Account multiple ticks of idle time.
5083 * @ticks: number of stolen ticks
5085 void account_idle_ticks(unsigned long ticks
)
5087 account_idle_time(jiffies_to_cputime(ticks
));
5093 * Use precise platform statistics if available:
5095 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5096 cputime_t
task_utime(struct task_struct
*p
)
5101 cputime_t
task_stime(struct task_struct
*p
)
5106 cputime_t
task_utime(struct task_struct
*p
)
5108 clock_t utime
= cputime_to_clock_t(p
->utime
),
5109 total
= utime
+ cputime_to_clock_t(p
->stime
);
5113 * Use CFS's precise accounting:
5115 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5119 do_div(temp
, total
);
5121 utime
= (clock_t)temp
;
5123 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5124 return p
->prev_utime
;
5127 cputime_t
task_stime(struct task_struct
*p
)
5132 * Use CFS's precise accounting. (we subtract utime from
5133 * the total, to make sure the total observed by userspace
5134 * grows monotonically - apps rely on that):
5136 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5137 cputime_to_clock_t(task_utime(p
));
5140 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5142 return p
->prev_stime
;
5146 inline cputime_t
task_gtime(struct task_struct
*p
)
5152 * This function gets called by the timer code, with HZ frequency.
5153 * We call it with interrupts disabled.
5155 * It also gets called by the fork code, when changing the parent's
5158 void scheduler_tick(void)
5160 int cpu
= smp_processor_id();
5161 struct rq
*rq
= cpu_rq(cpu
);
5162 struct task_struct
*curr
= rq
->curr
;
5166 spin_lock(&rq
->lock
);
5167 update_rq_clock(rq
);
5168 update_cpu_load(rq
);
5169 curr
->sched_class
->task_tick(rq
, curr
, 0);
5170 spin_unlock(&rq
->lock
);
5172 perf_counter_task_tick(curr
, cpu
);
5175 rq
->idle_at_tick
= idle_cpu(cpu
);
5176 trigger_load_balance(rq
, cpu
);
5180 notrace
unsigned long get_parent_ip(unsigned long addr
)
5182 if (in_lock_functions(addr
)) {
5183 addr
= CALLER_ADDR2
;
5184 if (in_lock_functions(addr
))
5185 addr
= CALLER_ADDR3
;
5190 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5191 defined(CONFIG_PREEMPT_TRACER))
5193 void __kprobes
add_preempt_count(int val
)
5195 #ifdef CONFIG_DEBUG_PREEMPT
5199 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5202 preempt_count() += val
;
5203 #ifdef CONFIG_DEBUG_PREEMPT
5205 * Spinlock count overflowing soon?
5207 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5210 if (preempt_count() == val
)
5211 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5213 EXPORT_SYMBOL(add_preempt_count
);
5215 void __kprobes
sub_preempt_count(int val
)
5217 #ifdef CONFIG_DEBUG_PREEMPT
5221 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5224 * Is the spinlock portion underflowing?
5226 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5227 !(preempt_count() & PREEMPT_MASK
)))
5231 if (preempt_count() == val
)
5232 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5233 preempt_count() -= val
;
5235 EXPORT_SYMBOL(sub_preempt_count
);
5240 * Print scheduling while atomic bug:
5242 static noinline
void __schedule_bug(struct task_struct
*prev
)
5244 struct pt_regs
*regs
= get_irq_regs();
5246 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5247 prev
->comm
, prev
->pid
, preempt_count());
5249 debug_show_held_locks(prev
);
5251 if (irqs_disabled())
5252 print_irqtrace_events(prev
);
5261 * Various schedule()-time debugging checks and statistics:
5263 static inline void schedule_debug(struct task_struct
*prev
)
5266 * Test if we are atomic. Since do_exit() needs to call into
5267 * schedule() atomically, we ignore that path for now.
5268 * Otherwise, whine if we are scheduling when we should not be.
5270 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5271 __schedule_bug(prev
);
5273 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5275 schedstat_inc(this_rq(), sched_count
);
5276 #ifdef CONFIG_SCHEDSTATS
5277 if (unlikely(prev
->lock_depth
>= 0)) {
5278 schedstat_inc(this_rq(), bkl_count
);
5279 schedstat_inc(prev
, sched_info
.bkl_count
);
5284 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5286 if (prev
->state
== TASK_RUNNING
) {
5287 u64 runtime
= prev
->se
.sum_exec_runtime
;
5289 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5290 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5293 * In order to avoid avg_overlap growing stale when we are
5294 * indeed overlapping and hence not getting put to sleep, grow
5295 * the avg_overlap on preemption.
5297 * We use the average preemption runtime because that
5298 * correlates to the amount of cache footprint a task can
5301 update_avg(&prev
->se
.avg_overlap
, runtime
);
5303 prev
->sched_class
->put_prev_task(rq
, prev
);
5307 * Pick up the highest-prio task:
5309 static inline struct task_struct
*
5310 pick_next_task(struct rq
*rq
)
5312 const struct sched_class
*class;
5313 struct task_struct
*p
;
5316 * Optimization: we know that if all tasks are in
5317 * the fair class we can call that function directly:
5319 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5320 p
= fair_sched_class
.pick_next_task(rq
);
5325 class = sched_class_highest
;
5327 p
= class->pick_next_task(rq
);
5331 * Will never be NULL as the idle class always
5332 * returns a non-NULL p:
5334 class = class->next
;
5339 * schedule() is the main scheduler function.
5341 asmlinkage
void __sched
schedule(void)
5343 struct task_struct
*prev
, *next
;
5344 unsigned long *switch_count
;
5350 cpu
= smp_processor_id();
5354 switch_count
= &prev
->nivcsw
;
5356 release_kernel_lock(prev
);
5357 need_resched_nonpreemptible
:
5359 schedule_debug(prev
);
5361 if (sched_feat(HRTICK
))
5364 spin_lock_irq(&rq
->lock
);
5365 update_rq_clock(rq
);
5366 clear_tsk_need_resched(prev
);
5368 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5369 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5370 prev
->state
= TASK_RUNNING
;
5372 deactivate_task(rq
, prev
, 1);
5373 switch_count
= &prev
->nvcsw
;
5377 if (prev
->sched_class
->pre_schedule
)
5378 prev
->sched_class
->pre_schedule(rq
, prev
);
5381 if (unlikely(!rq
->nr_running
))
5382 idle_balance(cpu
, rq
);
5384 put_prev_task(rq
, prev
);
5385 next
= pick_next_task(rq
);
5387 if (likely(prev
!= next
)) {
5388 sched_info_switch(prev
, next
);
5389 perf_counter_task_sched_out(prev
, next
, cpu
);
5395 context_switch(rq
, prev
, next
); /* unlocks the rq */
5397 * the context switch might have flipped the stack from under
5398 * us, hence refresh the local variables.
5400 cpu
= smp_processor_id();
5403 spin_unlock_irq(&rq
->lock
);
5405 if (unlikely(reacquire_kernel_lock(current
) < 0))
5406 goto need_resched_nonpreemptible
;
5408 preempt_enable_no_resched();
5412 EXPORT_SYMBOL(schedule
);
5416 * Look out! "owner" is an entirely speculative pointer
5417 * access and not reliable.
5419 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5424 if (!sched_feat(OWNER_SPIN
))
5427 #ifdef CONFIG_DEBUG_PAGEALLOC
5429 * Need to access the cpu field knowing that
5430 * DEBUG_PAGEALLOC could have unmapped it if
5431 * the mutex owner just released it and exited.
5433 if (probe_kernel_address(&owner
->cpu
, cpu
))
5440 * Even if the access succeeded (likely case),
5441 * the cpu field may no longer be valid.
5443 if (cpu
>= nr_cpumask_bits
)
5447 * We need to validate that we can do a
5448 * get_cpu() and that we have the percpu area.
5450 if (!cpu_online(cpu
))
5457 * Owner changed, break to re-assess state.
5459 if (lock
->owner
!= owner
)
5463 * Is that owner really running on that cpu?
5465 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5475 #ifdef CONFIG_PREEMPT
5477 * this is the entry point to schedule() from in-kernel preemption
5478 * off of preempt_enable. Kernel preemptions off return from interrupt
5479 * occur there and call schedule directly.
5481 asmlinkage
void __sched
preempt_schedule(void)
5483 struct thread_info
*ti
= current_thread_info();
5486 * If there is a non-zero preempt_count or interrupts are disabled,
5487 * we do not want to preempt the current task. Just return..
5489 if (likely(ti
->preempt_count
|| irqs_disabled()))
5493 add_preempt_count(PREEMPT_ACTIVE
);
5495 sub_preempt_count(PREEMPT_ACTIVE
);
5498 * Check again in case we missed a preemption opportunity
5499 * between schedule and now.
5502 } while (need_resched());
5504 EXPORT_SYMBOL(preempt_schedule
);
5507 * this is the entry point to schedule() from kernel preemption
5508 * off of irq context.
5509 * Note, that this is called and return with irqs disabled. This will
5510 * protect us against recursive calling from irq.
5512 asmlinkage
void __sched
preempt_schedule_irq(void)
5514 struct thread_info
*ti
= current_thread_info();
5516 /* Catch callers which need to be fixed */
5517 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5520 add_preempt_count(PREEMPT_ACTIVE
);
5523 local_irq_disable();
5524 sub_preempt_count(PREEMPT_ACTIVE
);
5527 * Check again in case we missed a preemption opportunity
5528 * between schedule and now.
5531 } while (need_resched());
5534 #endif /* CONFIG_PREEMPT */
5536 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5539 return try_to_wake_up(curr
->private, mode
, sync
);
5541 EXPORT_SYMBOL(default_wake_function
);
5544 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5545 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5546 * number) then we wake all the non-exclusive tasks and one exclusive task.
5548 * There are circumstances in which we can try to wake a task which has already
5549 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5550 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5552 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5553 int nr_exclusive
, int sync
, void *key
)
5555 wait_queue_t
*curr
, *next
;
5557 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5558 unsigned flags
= curr
->flags
;
5560 if (curr
->func(curr
, mode
, sync
, key
) &&
5561 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5567 * __wake_up - wake up threads blocked on a waitqueue.
5569 * @mode: which threads
5570 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5571 * @key: is directly passed to the wakeup function
5573 * It may be assumed that this function implies a write memory barrier before
5574 * changing the task state if and only if any tasks are woken up.
5576 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5577 int nr_exclusive
, void *key
)
5579 unsigned long flags
;
5581 spin_lock_irqsave(&q
->lock
, flags
);
5582 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5583 spin_unlock_irqrestore(&q
->lock
, flags
);
5585 EXPORT_SYMBOL(__wake_up
);
5588 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5590 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5592 __wake_up_common(q
, mode
, 1, 0, NULL
);
5595 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5597 __wake_up_common(q
, mode
, 1, 0, key
);
5601 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5603 * @mode: which threads
5604 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5605 * @key: opaque value to be passed to wakeup targets
5607 * The sync wakeup differs that the waker knows that it will schedule
5608 * away soon, so while the target thread will be woken up, it will not
5609 * be migrated to another CPU - ie. the two threads are 'synchronized'
5610 * with each other. This can prevent needless bouncing between CPUs.
5612 * On UP it can prevent extra preemption.
5614 * It may be assumed that this function implies a write memory barrier before
5615 * changing the task state if and only if any tasks are woken up.
5617 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5618 int nr_exclusive
, void *key
)
5620 unsigned long flags
;
5626 if (unlikely(!nr_exclusive
))
5629 spin_lock_irqsave(&q
->lock
, flags
);
5630 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5631 spin_unlock_irqrestore(&q
->lock
, flags
);
5633 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5636 * __wake_up_sync - see __wake_up_sync_key()
5638 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5640 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5642 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5645 * complete: - signals a single thread waiting on this completion
5646 * @x: holds the state of this particular completion
5648 * This will wake up a single thread waiting on this completion. Threads will be
5649 * awakened in the same order in which they were queued.
5651 * See also complete_all(), wait_for_completion() and related routines.
5653 * It may be assumed that this function implies a write memory barrier before
5654 * changing the task state if and only if any tasks are woken up.
5656 void complete(struct completion
*x
)
5658 unsigned long flags
;
5660 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5662 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5663 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5665 EXPORT_SYMBOL(complete
);
5668 * complete_all: - signals all threads waiting on this completion
5669 * @x: holds the state of this particular completion
5671 * This will wake up all threads waiting on this particular completion event.
5673 * It may be assumed that this function implies a write memory barrier before
5674 * changing the task state if and only if any tasks are woken up.
5676 void complete_all(struct completion
*x
)
5678 unsigned long flags
;
5680 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5681 x
->done
+= UINT_MAX
/2;
5682 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5683 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5685 EXPORT_SYMBOL(complete_all
);
5687 static inline long __sched
5688 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5691 DECLARE_WAITQUEUE(wait
, current
);
5693 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5694 __add_wait_queue_tail(&x
->wait
, &wait
);
5696 if (signal_pending_state(state
, current
)) {
5697 timeout
= -ERESTARTSYS
;
5700 __set_current_state(state
);
5701 spin_unlock_irq(&x
->wait
.lock
);
5702 timeout
= schedule_timeout(timeout
);
5703 spin_lock_irq(&x
->wait
.lock
);
5704 } while (!x
->done
&& timeout
);
5705 __remove_wait_queue(&x
->wait
, &wait
);
5710 return timeout
?: 1;
5714 wait_for_common(struct completion
*x
, long timeout
, int state
)
5718 spin_lock_irq(&x
->wait
.lock
);
5719 timeout
= do_wait_for_common(x
, timeout
, state
);
5720 spin_unlock_irq(&x
->wait
.lock
);
5725 * wait_for_completion: - waits for completion of a task
5726 * @x: holds the state of this particular completion
5728 * This waits to be signaled for completion of a specific task. It is NOT
5729 * interruptible and there is no timeout.
5731 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5732 * and interrupt capability. Also see complete().
5734 void __sched
wait_for_completion(struct completion
*x
)
5736 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5738 EXPORT_SYMBOL(wait_for_completion
);
5741 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5742 * @x: holds the state of this particular completion
5743 * @timeout: timeout value in jiffies
5745 * This waits for either a completion of a specific task to be signaled or for a
5746 * specified timeout to expire. The timeout is in jiffies. It is not
5749 unsigned long __sched
5750 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5752 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5754 EXPORT_SYMBOL(wait_for_completion_timeout
);
5757 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5758 * @x: holds the state of this particular completion
5760 * This waits for completion of a specific task to be signaled. It is
5763 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5765 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5766 if (t
== -ERESTARTSYS
)
5770 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5773 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5774 * @x: holds the state of this particular completion
5775 * @timeout: timeout value in jiffies
5777 * This waits for either a completion of a specific task to be signaled or for a
5778 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5780 unsigned long __sched
5781 wait_for_completion_interruptible_timeout(struct completion
*x
,
5782 unsigned long timeout
)
5784 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5786 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5789 * wait_for_completion_killable: - waits for completion of a task (killable)
5790 * @x: holds the state of this particular completion
5792 * This waits to be signaled for completion of a specific task. It can be
5793 * interrupted by a kill signal.
5795 int __sched
wait_for_completion_killable(struct completion
*x
)
5797 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5798 if (t
== -ERESTARTSYS
)
5802 EXPORT_SYMBOL(wait_for_completion_killable
);
5805 * try_wait_for_completion - try to decrement a completion without blocking
5806 * @x: completion structure
5808 * Returns: 0 if a decrement cannot be done without blocking
5809 * 1 if a decrement succeeded.
5811 * If a completion is being used as a counting completion,
5812 * attempt to decrement the counter without blocking. This
5813 * enables us to avoid waiting if the resource the completion
5814 * is protecting is not available.
5816 bool try_wait_for_completion(struct completion
*x
)
5820 spin_lock_irq(&x
->wait
.lock
);
5825 spin_unlock_irq(&x
->wait
.lock
);
5828 EXPORT_SYMBOL(try_wait_for_completion
);
5831 * completion_done - Test to see if a completion has any waiters
5832 * @x: completion structure
5834 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5835 * 1 if there are no waiters.
5838 bool completion_done(struct completion
*x
)
5842 spin_lock_irq(&x
->wait
.lock
);
5845 spin_unlock_irq(&x
->wait
.lock
);
5848 EXPORT_SYMBOL(completion_done
);
5851 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5853 unsigned long flags
;
5856 init_waitqueue_entry(&wait
, current
);
5858 __set_current_state(state
);
5860 spin_lock_irqsave(&q
->lock
, flags
);
5861 __add_wait_queue(q
, &wait
);
5862 spin_unlock(&q
->lock
);
5863 timeout
= schedule_timeout(timeout
);
5864 spin_lock_irq(&q
->lock
);
5865 __remove_wait_queue(q
, &wait
);
5866 spin_unlock_irqrestore(&q
->lock
, flags
);
5871 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5873 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5875 EXPORT_SYMBOL(interruptible_sleep_on
);
5878 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5880 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5882 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5884 void __sched
sleep_on(wait_queue_head_t
*q
)
5886 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5888 EXPORT_SYMBOL(sleep_on
);
5890 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5892 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5894 EXPORT_SYMBOL(sleep_on_timeout
);
5896 #ifdef CONFIG_RT_MUTEXES
5899 * rt_mutex_setprio - set the current priority of a task
5901 * @prio: prio value (kernel-internal form)
5903 * This function changes the 'effective' priority of a task. It does
5904 * not touch ->normal_prio like __setscheduler().
5906 * Used by the rt_mutex code to implement priority inheritance logic.
5908 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5910 unsigned long flags
;
5911 int oldprio
, on_rq
, running
;
5913 const struct sched_class
*prev_class
= p
->sched_class
;
5915 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5917 rq
= task_rq_lock(p
, &flags
);
5918 update_rq_clock(rq
);
5921 on_rq
= p
->se
.on_rq
;
5922 running
= task_current(rq
, p
);
5924 dequeue_task(rq
, p
, 0);
5926 p
->sched_class
->put_prev_task(rq
, p
);
5929 p
->sched_class
= &rt_sched_class
;
5931 p
->sched_class
= &fair_sched_class
;
5936 p
->sched_class
->set_curr_task(rq
);
5938 enqueue_task(rq
, p
, 0);
5940 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5942 task_rq_unlock(rq
, &flags
);
5947 void set_user_nice(struct task_struct
*p
, long nice
)
5949 int old_prio
, delta
, on_rq
;
5950 unsigned long flags
;
5953 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5956 * We have to be careful, if called from sys_setpriority(),
5957 * the task might be in the middle of scheduling on another CPU.
5959 rq
= task_rq_lock(p
, &flags
);
5960 update_rq_clock(rq
);
5962 * The RT priorities are set via sched_setscheduler(), but we still
5963 * allow the 'normal' nice value to be set - but as expected
5964 * it wont have any effect on scheduling until the task is
5965 * SCHED_FIFO/SCHED_RR:
5967 if (task_has_rt_policy(p
)) {
5968 p
->static_prio
= NICE_TO_PRIO(nice
);
5971 on_rq
= p
->se
.on_rq
;
5973 dequeue_task(rq
, p
, 0);
5975 p
->static_prio
= NICE_TO_PRIO(nice
);
5978 p
->prio
= effective_prio(p
);
5979 delta
= p
->prio
- old_prio
;
5982 enqueue_task(rq
, p
, 0);
5984 * If the task increased its priority or is running and
5985 * lowered its priority, then reschedule its CPU:
5987 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5988 resched_task(rq
->curr
);
5991 task_rq_unlock(rq
, &flags
);
5993 EXPORT_SYMBOL(set_user_nice
);
5996 * can_nice - check if a task can reduce its nice value
6000 int can_nice(const struct task_struct
*p
, const int nice
)
6002 /* convert nice value [19,-20] to rlimit style value [1,40] */
6003 int nice_rlim
= 20 - nice
;
6005 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6006 capable(CAP_SYS_NICE
));
6009 #ifdef __ARCH_WANT_SYS_NICE
6012 * sys_nice - change the priority of the current process.
6013 * @increment: priority increment
6015 * sys_setpriority is a more generic, but much slower function that
6016 * does similar things.
6018 SYSCALL_DEFINE1(nice
, int, increment
)
6023 * Setpriority might change our priority at the same moment.
6024 * We don't have to worry. Conceptually one call occurs first
6025 * and we have a single winner.
6027 if (increment
< -40)
6032 nice
= TASK_NICE(current
) + increment
;
6038 if (increment
< 0 && !can_nice(current
, nice
))
6041 retval
= security_task_setnice(current
, nice
);
6045 set_user_nice(current
, nice
);
6052 * task_prio - return the priority value of a given task.
6053 * @p: the task in question.
6055 * This is the priority value as seen by users in /proc.
6056 * RT tasks are offset by -200. Normal tasks are centered
6057 * around 0, value goes from -16 to +15.
6059 int task_prio(const struct task_struct
*p
)
6061 return p
->prio
- MAX_RT_PRIO
;
6065 * task_nice - return the nice value of a given task.
6066 * @p: the task in question.
6068 int task_nice(const struct task_struct
*p
)
6070 return TASK_NICE(p
);
6072 EXPORT_SYMBOL(task_nice
);
6075 * idle_cpu - is a given cpu idle currently?
6076 * @cpu: the processor in question.
6078 int idle_cpu(int cpu
)
6080 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6084 * idle_task - return the idle task for a given cpu.
6085 * @cpu: the processor in question.
6087 struct task_struct
*idle_task(int cpu
)
6089 return cpu_rq(cpu
)->idle
;
6093 * find_process_by_pid - find a process with a matching PID value.
6094 * @pid: the pid in question.
6096 static struct task_struct
*find_process_by_pid(pid_t pid
)
6098 return pid
? find_task_by_vpid(pid
) : current
;
6101 /* Actually do priority change: must hold rq lock. */
6103 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6105 BUG_ON(p
->se
.on_rq
);
6108 switch (p
->policy
) {
6112 p
->sched_class
= &fair_sched_class
;
6116 p
->sched_class
= &rt_sched_class
;
6120 p
->rt_priority
= prio
;
6121 p
->normal_prio
= normal_prio(p
);
6122 /* we are holding p->pi_lock already */
6123 p
->prio
= rt_mutex_getprio(p
);
6128 * check the target process has a UID that matches the current process's
6130 static bool check_same_owner(struct task_struct
*p
)
6132 const struct cred
*cred
= current_cred(), *pcred
;
6136 pcred
= __task_cred(p
);
6137 match
= (cred
->euid
== pcred
->euid
||
6138 cred
->euid
== pcred
->uid
);
6143 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6144 struct sched_param
*param
, bool user
)
6146 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6147 unsigned long flags
;
6148 const struct sched_class
*prev_class
= p
->sched_class
;
6152 /* may grab non-irq protected spin_locks */
6153 BUG_ON(in_interrupt());
6155 /* double check policy once rq lock held */
6157 reset_on_fork
= p
->sched_reset_on_fork
;
6158 policy
= oldpolicy
= p
->policy
;
6160 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6161 policy
&= ~SCHED_RESET_ON_FORK
;
6163 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6164 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6165 policy
!= SCHED_IDLE
)
6170 * Valid priorities for SCHED_FIFO and SCHED_RR are
6171 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6172 * SCHED_BATCH and SCHED_IDLE is 0.
6174 if (param
->sched_priority
< 0 ||
6175 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6176 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6178 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6182 * Allow unprivileged RT tasks to decrease priority:
6184 if (user
&& !capable(CAP_SYS_NICE
)) {
6185 if (rt_policy(policy
)) {
6186 unsigned long rlim_rtprio
;
6188 if (!lock_task_sighand(p
, &flags
))
6190 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6191 unlock_task_sighand(p
, &flags
);
6193 /* can't set/change the rt policy */
6194 if (policy
!= p
->policy
&& !rlim_rtprio
)
6197 /* can't increase priority */
6198 if (param
->sched_priority
> p
->rt_priority
&&
6199 param
->sched_priority
> rlim_rtprio
)
6203 * Like positive nice levels, dont allow tasks to
6204 * move out of SCHED_IDLE either:
6206 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6209 /* can't change other user's priorities */
6210 if (!check_same_owner(p
))
6213 /* Normal users shall not reset the sched_reset_on_fork flag */
6214 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6219 #ifdef CONFIG_RT_GROUP_SCHED
6221 * Do not allow realtime tasks into groups that have no runtime
6224 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6225 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6229 retval
= security_task_setscheduler(p
, policy
, param
);
6235 * make sure no PI-waiters arrive (or leave) while we are
6236 * changing the priority of the task:
6238 spin_lock_irqsave(&p
->pi_lock
, flags
);
6240 * To be able to change p->policy safely, the apropriate
6241 * runqueue lock must be held.
6243 rq
= __task_rq_lock(p
);
6244 /* recheck policy now with rq lock held */
6245 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6246 policy
= oldpolicy
= -1;
6247 __task_rq_unlock(rq
);
6248 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6251 update_rq_clock(rq
);
6252 on_rq
= p
->se
.on_rq
;
6253 running
= task_current(rq
, p
);
6255 deactivate_task(rq
, p
, 0);
6257 p
->sched_class
->put_prev_task(rq
, p
);
6259 p
->sched_reset_on_fork
= reset_on_fork
;
6262 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6265 p
->sched_class
->set_curr_task(rq
);
6267 activate_task(rq
, p
, 0);
6269 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6271 __task_rq_unlock(rq
);
6272 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6274 rt_mutex_adjust_pi(p
);
6280 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6281 * @p: the task in question.
6282 * @policy: new policy.
6283 * @param: structure containing the new RT priority.
6285 * NOTE that the task may be already dead.
6287 int sched_setscheduler(struct task_struct
*p
, int policy
,
6288 struct sched_param
*param
)
6290 return __sched_setscheduler(p
, policy
, param
, true);
6292 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6295 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6296 * @p: the task in question.
6297 * @policy: new policy.
6298 * @param: structure containing the new RT priority.
6300 * Just like sched_setscheduler, only don't bother checking if the
6301 * current context has permission. For example, this is needed in
6302 * stop_machine(): we create temporary high priority worker threads,
6303 * but our caller might not have that capability.
6305 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6306 struct sched_param
*param
)
6308 return __sched_setscheduler(p
, policy
, param
, false);
6312 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6314 struct sched_param lparam
;
6315 struct task_struct
*p
;
6318 if (!param
|| pid
< 0)
6320 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6325 p
= find_process_by_pid(pid
);
6327 retval
= sched_setscheduler(p
, policy
, &lparam
);
6334 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6335 * @pid: the pid in question.
6336 * @policy: new policy.
6337 * @param: structure containing the new RT priority.
6339 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6340 struct sched_param __user
*, param
)
6342 /* negative values for policy are not valid */
6346 return do_sched_setscheduler(pid
, policy
, param
);
6350 * sys_sched_setparam - set/change the RT priority of a thread
6351 * @pid: the pid in question.
6352 * @param: structure containing the new RT priority.
6354 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6356 return do_sched_setscheduler(pid
, -1, param
);
6360 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6361 * @pid: the pid in question.
6363 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6365 struct task_struct
*p
;
6372 read_lock(&tasklist_lock
);
6373 p
= find_process_by_pid(pid
);
6375 retval
= security_task_getscheduler(p
);
6378 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6380 read_unlock(&tasklist_lock
);
6385 * sys_sched_getparam - get the RT priority of a thread
6386 * @pid: the pid in question.
6387 * @param: structure containing the RT priority.
6389 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6391 struct sched_param lp
;
6392 struct task_struct
*p
;
6395 if (!param
|| pid
< 0)
6398 read_lock(&tasklist_lock
);
6399 p
= find_process_by_pid(pid
);
6404 retval
= security_task_getscheduler(p
);
6408 lp
.sched_priority
= p
->rt_priority
;
6409 read_unlock(&tasklist_lock
);
6412 * This one might sleep, we cannot do it with a spinlock held ...
6414 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6419 read_unlock(&tasklist_lock
);
6423 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6425 cpumask_var_t cpus_allowed
, new_mask
;
6426 struct task_struct
*p
;
6430 read_lock(&tasklist_lock
);
6432 p
= find_process_by_pid(pid
);
6434 read_unlock(&tasklist_lock
);
6440 * It is not safe to call set_cpus_allowed with the
6441 * tasklist_lock held. We will bump the task_struct's
6442 * usage count and then drop tasklist_lock.
6445 read_unlock(&tasklist_lock
);
6447 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6451 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6453 goto out_free_cpus_allowed
;
6456 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6459 retval
= security_task_setscheduler(p
, 0, NULL
);
6463 cpuset_cpus_allowed(p
, cpus_allowed
);
6464 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6466 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6469 cpuset_cpus_allowed(p
, cpus_allowed
);
6470 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6472 * We must have raced with a concurrent cpuset
6473 * update. Just reset the cpus_allowed to the
6474 * cpuset's cpus_allowed
6476 cpumask_copy(new_mask
, cpus_allowed
);
6481 free_cpumask_var(new_mask
);
6482 out_free_cpus_allowed
:
6483 free_cpumask_var(cpus_allowed
);
6490 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6491 struct cpumask
*new_mask
)
6493 if (len
< cpumask_size())
6494 cpumask_clear(new_mask
);
6495 else if (len
> cpumask_size())
6496 len
= cpumask_size();
6498 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6502 * sys_sched_setaffinity - set the cpu affinity of a process
6503 * @pid: pid of the process
6504 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6505 * @user_mask_ptr: user-space pointer to the new cpu mask
6507 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6508 unsigned long __user
*, user_mask_ptr
)
6510 cpumask_var_t new_mask
;
6513 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6516 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6518 retval
= sched_setaffinity(pid
, new_mask
);
6519 free_cpumask_var(new_mask
);
6523 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6525 struct task_struct
*p
;
6529 read_lock(&tasklist_lock
);
6532 p
= find_process_by_pid(pid
);
6536 retval
= security_task_getscheduler(p
);
6540 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6543 read_unlock(&tasklist_lock
);
6550 * sys_sched_getaffinity - get the cpu affinity of a process
6551 * @pid: pid of the process
6552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6553 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6555 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6556 unsigned long __user
*, user_mask_ptr
)
6561 if (len
< cpumask_size())
6564 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6567 ret
= sched_getaffinity(pid
, mask
);
6569 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6572 ret
= cpumask_size();
6574 free_cpumask_var(mask
);
6580 * sys_sched_yield - yield the current processor to other threads.
6582 * This function yields the current CPU to other tasks. If there are no
6583 * other threads running on this CPU then this function will return.
6585 SYSCALL_DEFINE0(sched_yield
)
6587 struct rq
*rq
= this_rq_lock();
6589 schedstat_inc(rq
, yld_count
);
6590 current
->sched_class
->yield_task(rq
);
6593 * Since we are going to call schedule() anyway, there's
6594 * no need to preempt or enable interrupts:
6596 __release(rq
->lock
);
6597 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6598 _raw_spin_unlock(&rq
->lock
);
6599 preempt_enable_no_resched();
6606 static inline int should_resched(void)
6608 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6611 static void __cond_resched(void)
6613 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6614 __might_sleep(__FILE__
, __LINE__
);
6617 * The BKS might be reacquired before we have dropped
6618 * PREEMPT_ACTIVE, which could trigger a second
6619 * cond_resched() call.
6621 add_preempt_count(PREEMPT_ACTIVE
);
6623 sub_preempt_count(PREEMPT_ACTIVE
);
6626 int __sched
_cond_resched(void)
6628 if (should_resched()) {
6634 EXPORT_SYMBOL(_cond_resched
);
6637 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6638 * call schedule, and on return reacquire the lock.
6640 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6641 * operations here to prevent schedule() from being called twice (once via
6642 * spin_unlock(), once by hand).
6644 int cond_resched_lock(spinlock_t
*lock
)
6646 int resched
= should_resched();
6649 if (spin_needbreak(lock
) || resched
) {
6660 EXPORT_SYMBOL(cond_resched_lock
);
6662 int __sched
cond_resched_softirq(void)
6664 BUG_ON(!in_softirq());
6666 if (should_resched()) {
6674 EXPORT_SYMBOL(cond_resched_softirq
);
6677 * yield - yield the current processor to other threads.
6679 * This is a shortcut for kernel-space yielding - it marks the
6680 * thread runnable and calls sys_sched_yield().
6682 void __sched
yield(void)
6684 set_current_state(TASK_RUNNING
);
6687 EXPORT_SYMBOL(yield
);
6690 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6691 * that process accounting knows that this is a task in IO wait state.
6693 * But don't do that if it is a deliberate, throttling IO wait (this task
6694 * has set its backing_dev_info: the queue against which it should throttle)
6696 void __sched
io_schedule(void)
6698 struct rq
*rq
= raw_rq();
6700 delayacct_blkio_start();
6701 atomic_inc(&rq
->nr_iowait
);
6703 atomic_dec(&rq
->nr_iowait
);
6704 delayacct_blkio_end();
6706 EXPORT_SYMBOL(io_schedule
);
6708 long __sched
io_schedule_timeout(long timeout
)
6710 struct rq
*rq
= raw_rq();
6713 delayacct_blkio_start();
6714 atomic_inc(&rq
->nr_iowait
);
6715 ret
= schedule_timeout(timeout
);
6716 atomic_dec(&rq
->nr_iowait
);
6717 delayacct_blkio_end();
6722 * sys_sched_get_priority_max - return maximum RT priority.
6723 * @policy: scheduling class.
6725 * this syscall returns the maximum rt_priority that can be used
6726 * by a given scheduling class.
6728 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6735 ret
= MAX_USER_RT_PRIO
-1;
6747 * sys_sched_get_priority_min - return minimum RT priority.
6748 * @policy: scheduling class.
6750 * this syscall returns the minimum rt_priority that can be used
6751 * by a given scheduling class.
6753 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6771 * sys_sched_rr_get_interval - return the default timeslice of a process.
6772 * @pid: pid of the process.
6773 * @interval: userspace pointer to the timeslice value.
6775 * this syscall writes the default timeslice value of a given process
6776 * into the user-space timespec buffer. A value of '0' means infinity.
6778 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6779 struct timespec __user
*, interval
)
6781 struct task_struct
*p
;
6782 unsigned int time_slice
;
6790 read_lock(&tasklist_lock
);
6791 p
= find_process_by_pid(pid
);
6795 retval
= security_task_getscheduler(p
);
6800 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6801 * tasks that are on an otherwise idle runqueue:
6804 if (p
->policy
== SCHED_RR
) {
6805 time_slice
= DEF_TIMESLICE
;
6806 } else if (p
->policy
!= SCHED_FIFO
) {
6807 struct sched_entity
*se
= &p
->se
;
6808 unsigned long flags
;
6811 rq
= task_rq_lock(p
, &flags
);
6812 if (rq
->cfs
.load
.weight
)
6813 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6814 task_rq_unlock(rq
, &flags
);
6816 read_unlock(&tasklist_lock
);
6817 jiffies_to_timespec(time_slice
, &t
);
6818 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6822 read_unlock(&tasklist_lock
);
6826 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6828 void sched_show_task(struct task_struct
*p
)
6830 unsigned long free
= 0;
6833 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6834 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6835 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6836 #if BITS_PER_LONG == 32
6837 if (state
== TASK_RUNNING
)
6838 printk(KERN_CONT
" running ");
6840 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6842 if (state
== TASK_RUNNING
)
6843 printk(KERN_CONT
" running task ");
6845 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6847 #ifdef CONFIG_DEBUG_STACK_USAGE
6848 free
= stack_not_used(p
);
6850 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
6851 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6852 (unsigned long)task_thread_info(p
)->flags
);
6854 show_stack(p
, NULL
);
6857 void show_state_filter(unsigned long state_filter
)
6859 struct task_struct
*g
, *p
;
6861 #if BITS_PER_LONG == 32
6863 " task PC stack pid father\n");
6866 " task PC stack pid father\n");
6868 read_lock(&tasklist_lock
);
6869 do_each_thread(g
, p
) {
6871 * reset the NMI-timeout, listing all files on a slow
6872 * console might take alot of time:
6874 touch_nmi_watchdog();
6875 if (!state_filter
|| (p
->state
& state_filter
))
6877 } while_each_thread(g
, p
);
6879 touch_all_softlockup_watchdogs();
6881 #ifdef CONFIG_SCHED_DEBUG
6882 sysrq_sched_debug_show();
6884 read_unlock(&tasklist_lock
);
6886 * Only show locks if all tasks are dumped:
6888 if (state_filter
== -1)
6889 debug_show_all_locks();
6892 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6894 idle
->sched_class
= &idle_sched_class
;
6898 * init_idle - set up an idle thread for a given CPU
6899 * @idle: task in question
6900 * @cpu: cpu the idle task belongs to
6902 * NOTE: this function does not set the idle thread's NEED_RESCHED
6903 * flag, to make booting more robust.
6905 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6907 struct rq
*rq
= cpu_rq(cpu
);
6908 unsigned long flags
;
6910 spin_lock_irqsave(&rq
->lock
, flags
);
6913 idle
->se
.exec_start
= sched_clock();
6915 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6916 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6917 __set_task_cpu(idle
, cpu
);
6919 rq
->curr
= rq
->idle
= idle
;
6920 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6923 spin_unlock_irqrestore(&rq
->lock
, flags
);
6925 /* Set the preempt count _outside_ the spinlocks! */
6926 #if defined(CONFIG_PREEMPT)
6927 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6929 task_thread_info(idle
)->preempt_count
= 0;
6932 * The idle tasks have their own, simple scheduling class:
6934 idle
->sched_class
= &idle_sched_class
;
6935 ftrace_graph_init_task(idle
);
6939 * In a system that switches off the HZ timer nohz_cpu_mask
6940 * indicates which cpus entered this state. This is used
6941 * in the rcu update to wait only for active cpus. For system
6942 * which do not switch off the HZ timer nohz_cpu_mask should
6943 * always be CPU_BITS_NONE.
6945 cpumask_var_t nohz_cpu_mask
;
6948 * Increase the granularity value when there are more CPUs,
6949 * because with more CPUs the 'effective latency' as visible
6950 * to users decreases. But the relationship is not linear,
6951 * so pick a second-best guess by going with the log2 of the
6954 * This idea comes from the SD scheduler of Con Kolivas:
6956 static inline void sched_init_granularity(void)
6958 unsigned int factor
= 1 + ilog2(num_online_cpus());
6959 const unsigned long limit
= 200000000;
6961 sysctl_sched_min_granularity
*= factor
;
6962 if (sysctl_sched_min_granularity
> limit
)
6963 sysctl_sched_min_granularity
= limit
;
6965 sysctl_sched_latency
*= factor
;
6966 if (sysctl_sched_latency
> limit
)
6967 sysctl_sched_latency
= limit
;
6969 sysctl_sched_wakeup_granularity
*= factor
;
6971 sysctl_sched_shares_ratelimit
*= factor
;
6976 * This is how migration works:
6978 * 1) we queue a struct migration_req structure in the source CPU's
6979 * runqueue and wake up that CPU's migration thread.
6980 * 2) we down() the locked semaphore => thread blocks.
6981 * 3) migration thread wakes up (implicitly it forces the migrated
6982 * thread off the CPU)
6983 * 4) it gets the migration request and checks whether the migrated
6984 * task is still in the wrong runqueue.
6985 * 5) if it's in the wrong runqueue then the migration thread removes
6986 * it and puts it into the right queue.
6987 * 6) migration thread up()s the semaphore.
6988 * 7) we wake up and the migration is done.
6992 * Change a given task's CPU affinity. Migrate the thread to a
6993 * proper CPU and schedule it away if the CPU it's executing on
6994 * is removed from the allowed bitmask.
6996 * NOTE: the caller must have a valid reference to the task, the
6997 * task must not exit() & deallocate itself prematurely. The
6998 * call is not atomic; no spinlocks may be held.
7000 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7002 struct migration_req req
;
7003 unsigned long flags
;
7007 rq
= task_rq_lock(p
, &flags
);
7008 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7013 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7014 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7019 if (p
->sched_class
->set_cpus_allowed
)
7020 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7022 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7023 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7026 /* Can the task run on the task's current CPU? If so, we're done */
7027 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7030 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7031 /* Need help from migration thread: drop lock and wait. */
7032 task_rq_unlock(rq
, &flags
);
7033 wake_up_process(rq
->migration_thread
);
7034 wait_for_completion(&req
.done
);
7035 tlb_migrate_finish(p
->mm
);
7039 task_rq_unlock(rq
, &flags
);
7043 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7046 * Move (not current) task off this cpu, onto dest cpu. We're doing
7047 * this because either it can't run here any more (set_cpus_allowed()
7048 * away from this CPU, or CPU going down), or because we're
7049 * attempting to rebalance this task on exec (sched_exec).
7051 * So we race with normal scheduler movements, but that's OK, as long
7052 * as the task is no longer on this CPU.
7054 * Returns non-zero if task was successfully migrated.
7056 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7058 struct rq
*rq_dest
, *rq_src
;
7061 if (unlikely(!cpu_active(dest_cpu
)))
7064 rq_src
= cpu_rq(src_cpu
);
7065 rq_dest
= cpu_rq(dest_cpu
);
7067 double_rq_lock(rq_src
, rq_dest
);
7068 /* Already moved. */
7069 if (task_cpu(p
) != src_cpu
)
7071 /* Affinity changed (again). */
7072 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7075 on_rq
= p
->se
.on_rq
;
7077 deactivate_task(rq_src
, p
, 0);
7079 set_task_cpu(p
, dest_cpu
);
7081 activate_task(rq_dest
, p
, 0);
7082 check_preempt_curr(rq_dest
, p
, 0);
7087 double_rq_unlock(rq_src
, rq_dest
);
7092 * migration_thread - this is a highprio system thread that performs
7093 * thread migration by bumping thread off CPU then 'pushing' onto
7096 static int migration_thread(void *data
)
7098 int cpu
= (long)data
;
7102 BUG_ON(rq
->migration_thread
!= current
);
7104 set_current_state(TASK_INTERRUPTIBLE
);
7105 while (!kthread_should_stop()) {
7106 struct migration_req
*req
;
7107 struct list_head
*head
;
7109 spin_lock_irq(&rq
->lock
);
7111 if (cpu_is_offline(cpu
)) {
7112 spin_unlock_irq(&rq
->lock
);
7116 if (rq
->active_balance
) {
7117 active_load_balance(rq
, cpu
);
7118 rq
->active_balance
= 0;
7121 head
= &rq
->migration_queue
;
7123 if (list_empty(head
)) {
7124 spin_unlock_irq(&rq
->lock
);
7126 set_current_state(TASK_INTERRUPTIBLE
);
7129 req
= list_entry(head
->next
, struct migration_req
, list
);
7130 list_del_init(head
->next
);
7132 spin_unlock(&rq
->lock
);
7133 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7136 complete(&req
->done
);
7138 __set_current_state(TASK_RUNNING
);
7143 #ifdef CONFIG_HOTPLUG_CPU
7145 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7149 local_irq_disable();
7150 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7156 * Figure out where task on dead CPU should go, use force if necessary.
7158 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7161 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7164 /* Look for allowed, online CPU in same node. */
7165 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7166 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7169 /* Any allowed, online CPU? */
7170 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7171 if (dest_cpu
< nr_cpu_ids
)
7174 /* No more Mr. Nice Guy. */
7175 if (dest_cpu
>= nr_cpu_ids
) {
7176 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7177 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7180 * Don't tell them about moving exiting tasks or
7181 * kernel threads (both mm NULL), since they never
7184 if (p
->mm
&& printk_ratelimit()) {
7185 printk(KERN_INFO
"process %d (%s) no "
7186 "longer affine to cpu%d\n",
7187 task_pid_nr(p
), p
->comm
, dead_cpu
);
7192 /* It can have affinity changed while we were choosing. */
7193 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7198 * While a dead CPU has no uninterruptible tasks queued at this point,
7199 * it might still have a nonzero ->nr_uninterruptible counter, because
7200 * for performance reasons the counter is not stricly tracking tasks to
7201 * their home CPUs. So we just add the counter to another CPU's counter,
7202 * to keep the global sum constant after CPU-down:
7204 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7206 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7207 unsigned long flags
;
7209 local_irq_save(flags
);
7210 double_rq_lock(rq_src
, rq_dest
);
7211 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7212 rq_src
->nr_uninterruptible
= 0;
7213 double_rq_unlock(rq_src
, rq_dest
);
7214 local_irq_restore(flags
);
7217 /* Run through task list and migrate tasks from the dead cpu. */
7218 static void migrate_live_tasks(int src_cpu
)
7220 struct task_struct
*p
, *t
;
7222 read_lock(&tasklist_lock
);
7224 do_each_thread(t
, p
) {
7228 if (task_cpu(p
) == src_cpu
)
7229 move_task_off_dead_cpu(src_cpu
, p
);
7230 } while_each_thread(t
, p
);
7232 read_unlock(&tasklist_lock
);
7236 * Schedules idle task to be the next runnable task on current CPU.
7237 * It does so by boosting its priority to highest possible.
7238 * Used by CPU offline code.
7240 void sched_idle_next(void)
7242 int this_cpu
= smp_processor_id();
7243 struct rq
*rq
= cpu_rq(this_cpu
);
7244 struct task_struct
*p
= rq
->idle
;
7245 unsigned long flags
;
7247 /* cpu has to be offline */
7248 BUG_ON(cpu_online(this_cpu
));
7251 * Strictly not necessary since rest of the CPUs are stopped by now
7252 * and interrupts disabled on the current cpu.
7254 spin_lock_irqsave(&rq
->lock
, flags
);
7256 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7258 update_rq_clock(rq
);
7259 activate_task(rq
, p
, 0);
7261 spin_unlock_irqrestore(&rq
->lock
, flags
);
7265 * Ensures that the idle task is using init_mm right before its cpu goes
7268 void idle_task_exit(void)
7270 struct mm_struct
*mm
= current
->active_mm
;
7272 BUG_ON(cpu_online(smp_processor_id()));
7275 switch_mm(mm
, &init_mm
, current
);
7279 /* called under rq->lock with disabled interrupts */
7280 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7282 struct rq
*rq
= cpu_rq(dead_cpu
);
7284 /* Must be exiting, otherwise would be on tasklist. */
7285 BUG_ON(!p
->exit_state
);
7287 /* Cannot have done final schedule yet: would have vanished. */
7288 BUG_ON(p
->state
== TASK_DEAD
);
7293 * Drop lock around migration; if someone else moves it,
7294 * that's OK. No task can be added to this CPU, so iteration is
7297 spin_unlock_irq(&rq
->lock
);
7298 move_task_off_dead_cpu(dead_cpu
, p
);
7299 spin_lock_irq(&rq
->lock
);
7304 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7305 static void migrate_dead_tasks(unsigned int dead_cpu
)
7307 struct rq
*rq
= cpu_rq(dead_cpu
);
7308 struct task_struct
*next
;
7311 if (!rq
->nr_running
)
7313 update_rq_clock(rq
);
7314 next
= pick_next_task(rq
);
7317 next
->sched_class
->put_prev_task(rq
, next
);
7318 migrate_dead(dead_cpu
, next
);
7324 * remove the tasks which were accounted by rq from calc_load_tasks.
7326 static void calc_global_load_remove(struct rq
*rq
)
7328 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7330 #endif /* CONFIG_HOTPLUG_CPU */
7332 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7334 static struct ctl_table sd_ctl_dir
[] = {
7336 .procname
= "sched_domain",
7342 static struct ctl_table sd_ctl_root
[] = {
7344 .ctl_name
= CTL_KERN
,
7345 .procname
= "kernel",
7347 .child
= sd_ctl_dir
,
7352 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7354 struct ctl_table
*entry
=
7355 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7360 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7362 struct ctl_table
*entry
;
7365 * In the intermediate directories, both the child directory and
7366 * procname are dynamically allocated and could fail but the mode
7367 * will always be set. In the lowest directory the names are
7368 * static strings and all have proc handlers.
7370 for (entry
= *tablep
; entry
->mode
; entry
++) {
7372 sd_free_ctl_entry(&entry
->child
);
7373 if (entry
->proc_handler
== NULL
)
7374 kfree(entry
->procname
);
7382 set_table_entry(struct ctl_table
*entry
,
7383 const char *procname
, void *data
, int maxlen
,
7384 mode_t mode
, proc_handler
*proc_handler
)
7386 entry
->procname
= procname
;
7388 entry
->maxlen
= maxlen
;
7390 entry
->proc_handler
= proc_handler
;
7393 static struct ctl_table
*
7394 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7396 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7401 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7402 sizeof(long), 0644, proc_doulongvec_minmax
);
7403 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7404 sizeof(long), 0644, proc_doulongvec_minmax
);
7405 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7406 sizeof(int), 0644, proc_dointvec_minmax
);
7407 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7408 sizeof(int), 0644, proc_dointvec_minmax
);
7409 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7410 sizeof(int), 0644, proc_dointvec_minmax
);
7411 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7412 sizeof(int), 0644, proc_dointvec_minmax
);
7413 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7414 sizeof(int), 0644, proc_dointvec_minmax
);
7415 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7416 sizeof(int), 0644, proc_dointvec_minmax
);
7417 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7418 sizeof(int), 0644, proc_dointvec_minmax
);
7419 set_table_entry(&table
[9], "cache_nice_tries",
7420 &sd
->cache_nice_tries
,
7421 sizeof(int), 0644, proc_dointvec_minmax
);
7422 set_table_entry(&table
[10], "flags", &sd
->flags
,
7423 sizeof(int), 0644, proc_dointvec_minmax
);
7424 set_table_entry(&table
[11], "name", sd
->name
,
7425 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7426 /* &table[12] is terminator */
7431 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7433 struct ctl_table
*entry
, *table
;
7434 struct sched_domain
*sd
;
7435 int domain_num
= 0, i
;
7438 for_each_domain(cpu
, sd
)
7440 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7445 for_each_domain(cpu
, sd
) {
7446 snprintf(buf
, 32, "domain%d", i
);
7447 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7449 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7456 static struct ctl_table_header
*sd_sysctl_header
;
7457 static void register_sched_domain_sysctl(void)
7459 int i
, cpu_num
= num_online_cpus();
7460 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7463 WARN_ON(sd_ctl_dir
[0].child
);
7464 sd_ctl_dir
[0].child
= entry
;
7469 for_each_online_cpu(i
) {
7470 snprintf(buf
, 32, "cpu%d", i
);
7471 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7473 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7477 WARN_ON(sd_sysctl_header
);
7478 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7481 /* may be called multiple times per register */
7482 static void unregister_sched_domain_sysctl(void)
7484 if (sd_sysctl_header
)
7485 unregister_sysctl_table(sd_sysctl_header
);
7486 sd_sysctl_header
= NULL
;
7487 if (sd_ctl_dir
[0].child
)
7488 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7491 static void register_sched_domain_sysctl(void)
7494 static void unregister_sched_domain_sysctl(void)
7499 static void set_rq_online(struct rq
*rq
)
7502 const struct sched_class
*class;
7504 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7507 for_each_class(class) {
7508 if (class->rq_online
)
7509 class->rq_online(rq
);
7514 static void set_rq_offline(struct rq
*rq
)
7517 const struct sched_class
*class;
7519 for_each_class(class) {
7520 if (class->rq_offline
)
7521 class->rq_offline(rq
);
7524 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7530 * migration_call - callback that gets triggered when a CPU is added.
7531 * Here we can start up the necessary migration thread for the new CPU.
7533 static int __cpuinit
7534 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7536 struct task_struct
*p
;
7537 int cpu
= (long)hcpu
;
7538 unsigned long flags
;
7543 case CPU_UP_PREPARE
:
7544 case CPU_UP_PREPARE_FROZEN
:
7545 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7548 kthread_bind(p
, cpu
);
7549 /* Must be high prio: stop_machine expects to yield to it. */
7550 rq
= task_rq_lock(p
, &flags
);
7551 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7552 task_rq_unlock(rq
, &flags
);
7554 cpu_rq(cpu
)->migration_thread
= p
;
7558 case CPU_ONLINE_FROZEN
:
7559 /* Strictly unnecessary, as first user will wake it. */
7560 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7562 /* Update our root-domain */
7564 spin_lock_irqsave(&rq
->lock
, flags
);
7565 rq
->calc_load_update
= calc_load_update
;
7566 rq
->calc_load_active
= 0;
7568 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7572 spin_unlock_irqrestore(&rq
->lock
, flags
);
7575 #ifdef CONFIG_HOTPLUG_CPU
7576 case CPU_UP_CANCELED
:
7577 case CPU_UP_CANCELED_FROZEN
:
7578 if (!cpu_rq(cpu
)->migration_thread
)
7580 /* Unbind it from offline cpu so it can run. Fall thru. */
7581 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7582 cpumask_any(cpu_online_mask
));
7583 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7584 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7585 cpu_rq(cpu
)->migration_thread
= NULL
;
7589 case CPU_DEAD_FROZEN
:
7590 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7591 migrate_live_tasks(cpu
);
7593 kthread_stop(rq
->migration_thread
);
7594 put_task_struct(rq
->migration_thread
);
7595 rq
->migration_thread
= NULL
;
7596 /* Idle task back to normal (off runqueue, low prio) */
7597 spin_lock_irq(&rq
->lock
);
7598 update_rq_clock(rq
);
7599 deactivate_task(rq
, rq
->idle
, 0);
7600 rq
->idle
->static_prio
= MAX_PRIO
;
7601 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7602 rq
->idle
->sched_class
= &idle_sched_class
;
7603 migrate_dead_tasks(cpu
);
7604 spin_unlock_irq(&rq
->lock
);
7606 migrate_nr_uninterruptible(rq
);
7607 BUG_ON(rq
->nr_running
!= 0);
7608 calc_global_load_remove(rq
);
7610 * No need to migrate the tasks: it was best-effort if
7611 * they didn't take sched_hotcpu_mutex. Just wake up
7614 spin_lock_irq(&rq
->lock
);
7615 while (!list_empty(&rq
->migration_queue
)) {
7616 struct migration_req
*req
;
7618 req
= list_entry(rq
->migration_queue
.next
,
7619 struct migration_req
, list
);
7620 list_del_init(&req
->list
);
7621 spin_unlock_irq(&rq
->lock
);
7622 complete(&req
->done
);
7623 spin_lock_irq(&rq
->lock
);
7625 spin_unlock_irq(&rq
->lock
);
7629 case CPU_DYING_FROZEN
:
7630 /* Update our root-domain */
7632 spin_lock_irqsave(&rq
->lock
, flags
);
7634 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7637 spin_unlock_irqrestore(&rq
->lock
, flags
);
7645 * Register at high priority so that task migration (migrate_all_tasks)
7646 * happens before everything else. This has to be lower priority than
7647 * the notifier in the perf_counter subsystem, though.
7649 static struct notifier_block __cpuinitdata migration_notifier
= {
7650 .notifier_call
= migration_call
,
7654 static int __init
migration_init(void)
7656 void *cpu
= (void *)(long)smp_processor_id();
7659 /* Start one for the boot CPU: */
7660 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7661 BUG_ON(err
== NOTIFY_BAD
);
7662 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7663 register_cpu_notifier(&migration_notifier
);
7667 early_initcall(migration_init
);
7672 #ifdef CONFIG_SCHED_DEBUG
7674 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7675 struct cpumask
*groupmask
)
7677 struct sched_group
*group
= sd
->groups
;
7680 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7681 cpumask_clear(groupmask
);
7683 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7685 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7686 printk("does not load-balance\n");
7688 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7693 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7695 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7696 printk(KERN_ERR
"ERROR: domain->span does not contain "
7699 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7700 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7704 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7708 printk(KERN_ERR
"ERROR: group is NULL\n");
7712 if (!group
->__cpu_power
) {
7713 printk(KERN_CONT
"\n");
7714 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7719 if (!cpumask_weight(sched_group_cpus(group
))) {
7720 printk(KERN_CONT
"\n");
7721 printk(KERN_ERR
"ERROR: empty group\n");
7725 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7726 printk(KERN_CONT
"\n");
7727 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7731 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7733 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7735 printk(KERN_CONT
" %s", str
);
7736 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7737 printk(KERN_CONT
" (__cpu_power = %d)",
7738 group
->__cpu_power
);
7741 group
= group
->next
;
7742 } while (group
!= sd
->groups
);
7743 printk(KERN_CONT
"\n");
7745 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7746 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7749 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7750 printk(KERN_ERR
"ERROR: parent span is not a superset "
7751 "of domain->span\n");
7755 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7757 cpumask_var_t groupmask
;
7761 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7765 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7767 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7768 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7773 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7780 free_cpumask_var(groupmask
);
7782 #else /* !CONFIG_SCHED_DEBUG */
7783 # define sched_domain_debug(sd, cpu) do { } while (0)
7784 #endif /* CONFIG_SCHED_DEBUG */
7786 static int sd_degenerate(struct sched_domain
*sd
)
7788 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7791 /* Following flags need at least 2 groups */
7792 if (sd
->flags
& (SD_LOAD_BALANCE
|
7793 SD_BALANCE_NEWIDLE
|
7797 SD_SHARE_PKG_RESOURCES
)) {
7798 if (sd
->groups
!= sd
->groups
->next
)
7802 /* Following flags don't use groups */
7803 if (sd
->flags
& (SD_WAKE_IDLE
|
7812 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7814 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7816 if (sd_degenerate(parent
))
7819 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7822 /* Does parent contain flags not in child? */
7823 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7824 if (cflags
& SD_WAKE_AFFINE
)
7825 pflags
&= ~SD_WAKE_BALANCE
;
7826 /* Flags needing groups don't count if only 1 group in parent */
7827 if (parent
->groups
== parent
->groups
->next
) {
7828 pflags
&= ~(SD_LOAD_BALANCE
|
7829 SD_BALANCE_NEWIDLE
|
7833 SD_SHARE_PKG_RESOURCES
);
7834 if (nr_node_ids
== 1)
7835 pflags
&= ~SD_SERIALIZE
;
7837 if (~cflags
& pflags
)
7843 static void free_rootdomain(struct root_domain
*rd
)
7845 cpupri_cleanup(&rd
->cpupri
);
7847 free_cpumask_var(rd
->rto_mask
);
7848 free_cpumask_var(rd
->online
);
7849 free_cpumask_var(rd
->span
);
7853 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7855 struct root_domain
*old_rd
= NULL
;
7856 unsigned long flags
;
7858 spin_lock_irqsave(&rq
->lock
, flags
);
7863 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7866 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7869 * If we dont want to free the old_rt yet then
7870 * set old_rd to NULL to skip the freeing later
7873 if (!atomic_dec_and_test(&old_rd
->refcount
))
7877 atomic_inc(&rd
->refcount
);
7880 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7881 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7884 spin_unlock_irqrestore(&rq
->lock
, flags
);
7887 free_rootdomain(old_rd
);
7890 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7892 gfp_t gfp
= GFP_KERNEL
;
7894 memset(rd
, 0, sizeof(*rd
));
7899 if (!alloc_cpumask_var(&rd
->span
, gfp
))
7901 if (!alloc_cpumask_var(&rd
->online
, gfp
))
7903 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
7906 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
7911 free_cpumask_var(rd
->rto_mask
);
7913 free_cpumask_var(rd
->online
);
7915 free_cpumask_var(rd
->span
);
7920 static void init_defrootdomain(void)
7922 init_rootdomain(&def_root_domain
, true);
7924 atomic_set(&def_root_domain
.refcount
, 1);
7927 static struct root_domain
*alloc_rootdomain(void)
7929 struct root_domain
*rd
;
7931 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7935 if (init_rootdomain(rd
, false) != 0) {
7944 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7945 * hold the hotplug lock.
7948 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7950 struct rq
*rq
= cpu_rq(cpu
);
7951 struct sched_domain
*tmp
;
7953 /* Remove the sched domains which do not contribute to scheduling. */
7954 for (tmp
= sd
; tmp
; ) {
7955 struct sched_domain
*parent
= tmp
->parent
;
7959 if (sd_parent_degenerate(tmp
, parent
)) {
7960 tmp
->parent
= parent
->parent
;
7962 parent
->parent
->child
= tmp
;
7967 if (sd
&& sd_degenerate(sd
)) {
7973 sched_domain_debug(sd
, cpu
);
7975 rq_attach_root(rq
, rd
);
7976 rcu_assign_pointer(rq
->sd
, sd
);
7979 /* cpus with isolated domains */
7980 static cpumask_var_t cpu_isolated_map
;
7982 /* Setup the mask of cpus configured for isolated domains */
7983 static int __init
isolated_cpu_setup(char *str
)
7985 cpulist_parse(str
, cpu_isolated_map
);
7989 __setup("isolcpus=", isolated_cpu_setup
);
7992 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7993 * to a function which identifies what group(along with sched group) a CPU
7994 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7995 * (due to the fact that we keep track of groups covered with a struct cpumask).
7997 * init_sched_build_groups will build a circular linked list of the groups
7998 * covered by the given span, and will set each group's ->cpumask correctly,
7999 * and ->cpu_power to 0.
8002 init_sched_build_groups(const struct cpumask
*span
,
8003 const struct cpumask
*cpu_map
,
8004 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8005 struct sched_group
**sg
,
8006 struct cpumask
*tmpmask
),
8007 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8009 struct sched_group
*first
= NULL
, *last
= NULL
;
8012 cpumask_clear(covered
);
8014 for_each_cpu(i
, span
) {
8015 struct sched_group
*sg
;
8016 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8019 if (cpumask_test_cpu(i
, covered
))
8022 cpumask_clear(sched_group_cpus(sg
));
8023 sg
->__cpu_power
= 0;
8025 for_each_cpu(j
, span
) {
8026 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8029 cpumask_set_cpu(j
, covered
);
8030 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8041 #define SD_NODES_PER_DOMAIN 16
8046 * find_next_best_node - find the next node to include in a sched_domain
8047 * @node: node whose sched_domain we're building
8048 * @used_nodes: nodes already in the sched_domain
8050 * Find the next node to include in a given scheduling domain. Simply
8051 * finds the closest node not already in the @used_nodes map.
8053 * Should use nodemask_t.
8055 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8057 int i
, n
, val
, min_val
, best_node
= 0;
8061 for (i
= 0; i
< nr_node_ids
; i
++) {
8062 /* Start at @node */
8063 n
= (node
+ i
) % nr_node_ids
;
8065 if (!nr_cpus_node(n
))
8068 /* Skip already used nodes */
8069 if (node_isset(n
, *used_nodes
))
8072 /* Simple min distance search */
8073 val
= node_distance(node
, n
);
8075 if (val
< min_val
) {
8081 node_set(best_node
, *used_nodes
);
8086 * sched_domain_node_span - get a cpumask for a node's sched_domain
8087 * @node: node whose cpumask we're constructing
8088 * @span: resulting cpumask
8090 * Given a node, construct a good cpumask for its sched_domain to span. It
8091 * should be one that prevents unnecessary balancing, but also spreads tasks
8094 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8096 nodemask_t used_nodes
;
8099 cpumask_clear(span
);
8100 nodes_clear(used_nodes
);
8102 cpumask_or(span
, span
, cpumask_of_node(node
));
8103 node_set(node
, used_nodes
);
8105 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8106 int next_node
= find_next_best_node(node
, &used_nodes
);
8108 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8111 #endif /* CONFIG_NUMA */
8113 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8116 * The cpus mask in sched_group and sched_domain hangs off the end.
8118 * ( See the the comments in include/linux/sched.h:struct sched_group
8119 * and struct sched_domain. )
8121 struct static_sched_group
{
8122 struct sched_group sg
;
8123 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8126 struct static_sched_domain
{
8127 struct sched_domain sd
;
8128 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8132 * SMT sched-domains:
8134 #ifdef CONFIG_SCHED_SMT
8135 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8136 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8139 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8140 struct sched_group
**sg
, struct cpumask
*unused
)
8143 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8146 #endif /* CONFIG_SCHED_SMT */
8149 * multi-core sched-domains:
8151 #ifdef CONFIG_SCHED_MC
8152 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8153 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8154 #endif /* CONFIG_SCHED_MC */
8156 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8158 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8159 struct sched_group
**sg
, struct cpumask
*mask
)
8163 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8164 group
= cpumask_first(mask
);
8166 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8169 #elif defined(CONFIG_SCHED_MC)
8171 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8172 struct sched_group
**sg
, struct cpumask
*unused
)
8175 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8180 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8181 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8184 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8185 struct sched_group
**sg
, struct cpumask
*mask
)
8188 #ifdef CONFIG_SCHED_MC
8189 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8190 group
= cpumask_first(mask
);
8191 #elif defined(CONFIG_SCHED_SMT)
8192 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8193 group
= cpumask_first(mask
);
8198 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8204 * The init_sched_build_groups can't handle what we want to do with node
8205 * groups, so roll our own. Now each node has its own list of groups which
8206 * gets dynamically allocated.
8208 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8209 static struct sched_group
***sched_group_nodes_bycpu
;
8211 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8212 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8214 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8215 struct sched_group
**sg
,
8216 struct cpumask
*nodemask
)
8220 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8221 group
= cpumask_first(nodemask
);
8224 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8228 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8230 struct sched_group
*sg
= group_head
;
8236 for_each_cpu(j
, sched_group_cpus(sg
)) {
8237 struct sched_domain
*sd
;
8239 sd
= &per_cpu(phys_domains
, j
).sd
;
8240 if (j
!= group_first_cpu(sd
->groups
)) {
8242 * Only add "power" once for each
8248 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
8251 } while (sg
!= group_head
);
8253 #endif /* CONFIG_NUMA */
8256 /* Free memory allocated for various sched_group structures */
8257 static void free_sched_groups(const struct cpumask
*cpu_map
,
8258 struct cpumask
*nodemask
)
8262 for_each_cpu(cpu
, cpu_map
) {
8263 struct sched_group
**sched_group_nodes
8264 = sched_group_nodes_bycpu
[cpu
];
8266 if (!sched_group_nodes
)
8269 for (i
= 0; i
< nr_node_ids
; i
++) {
8270 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8272 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8273 if (cpumask_empty(nodemask
))
8283 if (oldsg
!= sched_group_nodes
[i
])
8286 kfree(sched_group_nodes
);
8287 sched_group_nodes_bycpu
[cpu
] = NULL
;
8290 #else /* !CONFIG_NUMA */
8291 static void free_sched_groups(const struct cpumask
*cpu_map
,
8292 struct cpumask
*nodemask
)
8295 #endif /* CONFIG_NUMA */
8298 * Initialize sched groups cpu_power.
8300 * cpu_power indicates the capacity of sched group, which is used while
8301 * distributing the load between different sched groups in a sched domain.
8302 * Typically cpu_power for all the groups in a sched domain will be same unless
8303 * there are asymmetries in the topology. If there are asymmetries, group
8304 * having more cpu_power will pickup more load compared to the group having
8307 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8308 * the maximum number of tasks a group can handle in the presence of other idle
8309 * or lightly loaded groups in the same sched domain.
8311 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8313 struct sched_domain
*child
;
8314 struct sched_group
*group
;
8316 WARN_ON(!sd
|| !sd
->groups
);
8318 if (cpu
!= group_first_cpu(sd
->groups
))
8323 sd
->groups
->__cpu_power
= 0;
8326 * For perf policy, if the groups in child domain share resources
8327 * (for example cores sharing some portions of the cache hierarchy
8328 * or SMT), then set this domain groups cpu_power such that each group
8329 * can handle only one task, when there are other idle groups in the
8330 * same sched domain.
8332 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8334 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8335 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8340 * add cpu_power of each child group to this groups cpu_power
8342 group
= child
->groups
;
8344 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8345 group
= group
->next
;
8346 } while (group
!= child
->groups
);
8350 * Initializers for schedule domains
8351 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8354 #ifdef CONFIG_SCHED_DEBUG
8355 # define SD_INIT_NAME(sd, type) sd->name = #type
8357 # define SD_INIT_NAME(sd, type) do { } while (0)
8360 #define SD_INIT(sd, type) sd_init_##type(sd)
8362 #define SD_INIT_FUNC(type) \
8363 static noinline void sd_init_##type(struct sched_domain *sd) \
8365 memset(sd, 0, sizeof(*sd)); \
8366 *sd = SD_##type##_INIT; \
8367 sd->level = SD_LV_##type; \
8368 SD_INIT_NAME(sd, type); \
8373 SD_INIT_FUNC(ALLNODES
)
8376 #ifdef CONFIG_SCHED_SMT
8377 SD_INIT_FUNC(SIBLING
)
8379 #ifdef CONFIG_SCHED_MC
8383 static int default_relax_domain_level
= -1;
8385 static int __init
setup_relax_domain_level(char *str
)
8389 val
= simple_strtoul(str
, NULL
, 0);
8390 if (val
< SD_LV_MAX
)
8391 default_relax_domain_level
= val
;
8395 __setup("relax_domain_level=", setup_relax_domain_level
);
8397 static void set_domain_attribute(struct sched_domain
*sd
,
8398 struct sched_domain_attr
*attr
)
8402 if (!attr
|| attr
->relax_domain_level
< 0) {
8403 if (default_relax_domain_level
< 0)
8406 request
= default_relax_domain_level
;
8408 request
= attr
->relax_domain_level
;
8409 if (request
< sd
->level
) {
8410 /* turn off idle balance on this domain */
8411 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8413 /* turn on idle balance on this domain */
8414 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8419 * Build sched domains for a given set of cpus and attach the sched domains
8420 * to the individual cpus
8422 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8423 struct sched_domain_attr
*attr
)
8425 int i
, err
= -ENOMEM
;
8426 struct root_domain
*rd
;
8427 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8430 cpumask_var_t domainspan
, covered
, notcovered
;
8431 struct sched_group
**sched_group_nodes
= NULL
;
8432 int sd_allnodes
= 0;
8434 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8436 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8437 goto free_domainspan
;
8438 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8442 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8443 goto free_notcovered
;
8444 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8446 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8447 goto free_this_sibling_map
;
8448 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8449 goto free_this_core_map
;
8450 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8451 goto free_send_covered
;
8455 * Allocate the per-node list of sched groups
8457 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8459 if (!sched_group_nodes
) {
8460 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8465 rd
= alloc_rootdomain();
8467 printk(KERN_WARNING
"Cannot alloc root domain\n");
8468 goto free_sched_groups
;
8472 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8476 * Set up domains for cpus specified by the cpu_map.
8478 for_each_cpu(i
, cpu_map
) {
8479 struct sched_domain
*sd
= NULL
, *p
;
8481 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8484 if (cpumask_weight(cpu_map
) >
8485 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8486 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8487 SD_INIT(sd
, ALLNODES
);
8488 set_domain_attribute(sd
, attr
);
8489 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8490 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8496 sd
= &per_cpu(node_domains
, i
).sd
;
8498 set_domain_attribute(sd
, attr
);
8499 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8503 cpumask_and(sched_domain_span(sd
),
8504 sched_domain_span(sd
), cpu_map
);
8508 sd
= &per_cpu(phys_domains
, i
).sd
;
8510 set_domain_attribute(sd
, attr
);
8511 cpumask_copy(sched_domain_span(sd
), nodemask
);
8515 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8517 #ifdef CONFIG_SCHED_MC
8519 sd
= &per_cpu(core_domains
, i
).sd
;
8521 set_domain_attribute(sd
, attr
);
8522 cpumask_and(sched_domain_span(sd
), cpu_map
,
8523 cpu_coregroup_mask(i
));
8526 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8529 #ifdef CONFIG_SCHED_SMT
8531 sd
= &per_cpu(cpu_domains
, i
).sd
;
8532 SD_INIT(sd
, SIBLING
);
8533 set_domain_attribute(sd
, attr
);
8534 cpumask_and(sched_domain_span(sd
),
8535 topology_thread_cpumask(i
), cpu_map
);
8538 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8542 #ifdef CONFIG_SCHED_SMT
8543 /* Set up CPU (sibling) groups */
8544 for_each_cpu(i
, cpu_map
) {
8545 cpumask_and(this_sibling_map
,
8546 topology_thread_cpumask(i
), cpu_map
);
8547 if (i
!= cpumask_first(this_sibling_map
))
8550 init_sched_build_groups(this_sibling_map
, cpu_map
,
8552 send_covered
, tmpmask
);
8556 #ifdef CONFIG_SCHED_MC
8557 /* Set up multi-core groups */
8558 for_each_cpu(i
, cpu_map
) {
8559 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8560 if (i
!= cpumask_first(this_core_map
))
8563 init_sched_build_groups(this_core_map
, cpu_map
,
8565 send_covered
, tmpmask
);
8569 /* Set up physical groups */
8570 for (i
= 0; i
< nr_node_ids
; i
++) {
8571 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8572 if (cpumask_empty(nodemask
))
8575 init_sched_build_groups(nodemask
, cpu_map
,
8577 send_covered
, tmpmask
);
8581 /* Set up node groups */
8583 init_sched_build_groups(cpu_map
, cpu_map
,
8584 &cpu_to_allnodes_group
,
8585 send_covered
, tmpmask
);
8588 for (i
= 0; i
< nr_node_ids
; i
++) {
8589 /* Set up node groups */
8590 struct sched_group
*sg
, *prev
;
8593 cpumask_clear(covered
);
8594 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8595 if (cpumask_empty(nodemask
)) {
8596 sched_group_nodes
[i
] = NULL
;
8600 sched_domain_node_span(i
, domainspan
);
8601 cpumask_and(domainspan
, domainspan
, cpu_map
);
8603 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8606 printk(KERN_WARNING
"Can not alloc domain group for "
8610 sched_group_nodes
[i
] = sg
;
8611 for_each_cpu(j
, nodemask
) {
8612 struct sched_domain
*sd
;
8614 sd
= &per_cpu(node_domains
, j
).sd
;
8617 sg
->__cpu_power
= 0;
8618 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8620 cpumask_or(covered
, covered
, nodemask
);
8623 for (j
= 0; j
< nr_node_ids
; j
++) {
8624 int n
= (i
+ j
) % nr_node_ids
;
8626 cpumask_complement(notcovered
, covered
);
8627 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8628 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8629 if (cpumask_empty(tmpmask
))
8632 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8633 if (cpumask_empty(tmpmask
))
8636 sg
= kmalloc_node(sizeof(struct sched_group
) +
8641 "Can not alloc domain group for node %d\n", j
);
8644 sg
->__cpu_power
= 0;
8645 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8646 sg
->next
= prev
->next
;
8647 cpumask_or(covered
, covered
, tmpmask
);
8654 /* Calculate CPU power for physical packages and nodes */
8655 #ifdef CONFIG_SCHED_SMT
8656 for_each_cpu(i
, cpu_map
) {
8657 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8659 init_sched_groups_power(i
, sd
);
8662 #ifdef CONFIG_SCHED_MC
8663 for_each_cpu(i
, cpu_map
) {
8664 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8666 init_sched_groups_power(i
, sd
);
8670 for_each_cpu(i
, cpu_map
) {
8671 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8673 init_sched_groups_power(i
, sd
);
8677 for (i
= 0; i
< nr_node_ids
; i
++)
8678 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8681 struct sched_group
*sg
;
8683 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8685 init_numa_sched_groups_power(sg
);
8689 /* Attach the domains */
8690 for_each_cpu(i
, cpu_map
) {
8691 struct sched_domain
*sd
;
8692 #ifdef CONFIG_SCHED_SMT
8693 sd
= &per_cpu(cpu_domains
, i
).sd
;
8694 #elif defined(CONFIG_SCHED_MC)
8695 sd
= &per_cpu(core_domains
, i
).sd
;
8697 sd
= &per_cpu(phys_domains
, i
).sd
;
8699 cpu_attach_domain(sd
, rd
, i
);
8705 free_cpumask_var(tmpmask
);
8707 free_cpumask_var(send_covered
);
8709 free_cpumask_var(this_core_map
);
8710 free_this_sibling_map
:
8711 free_cpumask_var(this_sibling_map
);
8713 free_cpumask_var(nodemask
);
8716 free_cpumask_var(notcovered
);
8718 free_cpumask_var(covered
);
8720 free_cpumask_var(domainspan
);
8727 kfree(sched_group_nodes
);
8733 free_sched_groups(cpu_map
, tmpmask
);
8734 free_rootdomain(rd
);
8739 static int build_sched_domains(const struct cpumask
*cpu_map
)
8741 return __build_sched_domains(cpu_map
, NULL
);
8744 static struct cpumask
*doms_cur
; /* current sched domains */
8745 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8746 static struct sched_domain_attr
*dattr_cur
;
8747 /* attribues of custom domains in 'doms_cur' */
8750 * Special case: If a kmalloc of a doms_cur partition (array of
8751 * cpumask) fails, then fallback to a single sched domain,
8752 * as determined by the single cpumask fallback_doms.
8754 static cpumask_var_t fallback_doms
;
8757 * arch_update_cpu_topology lets virtualized architectures update the
8758 * cpu core maps. It is supposed to return 1 if the topology changed
8759 * or 0 if it stayed the same.
8761 int __attribute__((weak
)) arch_update_cpu_topology(void)
8767 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8768 * For now this just excludes isolated cpus, but could be used to
8769 * exclude other special cases in the future.
8771 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8775 arch_update_cpu_topology();
8777 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8779 doms_cur
= fallback_doms
;
8780 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8782 err
= build_sched_domains(doms_cur
);
8783 register_sched_domain_sysctl();
8788 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8789 struct cpumask
*tmpmask
)
8791 free_sched_groups(cpu_map
, tmpmask
);
8795 * Detach sched domains from a group of cpus specified in cpu_map
8796 * These cpus will now be attached to the NULL domain
8798 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8800 /* Save because hotplug lock held. */
8801 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8804 for_each_cpu(i
, cpu_map
)
8805 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8806 synchronize_sched();
8807 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8810 /* handle null as "default" */
8811 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8812 struct sched_domain_attr
*new, int idx_new
)
8814 struct sched_domain_attr tmp
;
8821 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8822 new ? (new + idx_new
) : &tmp
,
8823 sizeof(struct sched_domain_attr
));
8827 * Partition sched domains as specified by the 'ndoms_new'
8828 * cpumasks in the array doms_new[] of cpumasks. This compares
8829 * doms_new[] to the current sched domain partitioning, doms_cur[].
8830 * It destroys each deleted domain and builds each new domain.
8832 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8833 * The masks don't intersect (don't overlap.) We should setup one
8834 * sched domain for each mask. CPUs not in any of the cpumasks will
8835 * not be load balanced. If the same cpumask appears both in the
8836 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8839 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8840 * ownership of it and will kfree it when done with it. If the caller
8841 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8842 * ndoms_new == 1, and partition_sched_domains() will fallback to
8843 * the single partition 'fallback_doms', it also forces the domains
8846 * If doms_new == NULL it will be replaced with cpu_online_mask.
8847 * ndoms_new == 0 is a special case for destroying existing domains,
8848 * and it will not create the default domain.
8850 * Call with hotplug lock held
8852 /* FIXME: Change to struct cpumask *doms_new[] */
8853 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8854 struct sched_domain_attr
*dattr_new
)
8859 mutex_lock(&sched_domains_mutex
);
8861 /* always unregister in case we don't destroy any domains */
8862 unregister_sched_domain_sysctl();
8864 /* Let architecture update cpu core mappings. */
8865 new_topology
= arch_update_cpu_topology();
8867 n
= doms_new
? ndoms_new
: 0;
8869 /* Destroy deleted domains */
8870 for (i
= 0; i
< ndoms_cur
; i
++) {
8871 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8872 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8873 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8876 /* no match - a current sched domain not in new doms_new[] */
8877 detach_destroy_domains(doms_cur
+ i
);
8882 if (doms_new
== NULL
) {
8884 doms_new
= fallback_doms
;
8885 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8886 WARN_ON_ONCE(dattr_new
);
8889 /* Build new domains */
8890 for (i
= 0; i
< ndoms_new
; i
++) {
8891 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8892 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8893 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8896 /* no match - add a new doms_new */
8897 __build_sched_domains(doms_new
+ i
,
8898 dattr_new
? dattr_new
+ i
: NULL
);
8903 /* Remember the new sched domains */
8904 if (doms_cur
!= fallback_doms
)
8906 kfree(dattr_cur
); /* kfree(NULL) is safe */
8907 doms_cur
= doms_new
;
8908 dattr_cur
= dattr_new
;
8909 ndoms_cur
= ndoms_new
;
8911 register_sched_domain_sysctl();
8913 mutex_unlock(&sched_domains_mutex
);
8916 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8917 static void arch_reinit_sched_domains(void)
8921 /* Destroy domains first to force the rebuild */
8922 partition_sched_domains(0, NULL
, NULL
);
8924 rebuild_sched_domains();
8928 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8930 unsigned int level
= 0;
8932 if (sscanf(buf
, "%u", &level
) != 1)
8936 * level is always be positive so don't check for
8937 * level < POWERSAVINGS_BALANCE_NONE which is 0
8938 * What happens on 0 or 1 byte write,
8939 * need to check for count as well?
8942 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8946 sched_smt_power_savings
= level
;
8948 sched_mc_power_savings
= level
;
8950 arch_reinit_sched_domains();
8955 #ifdef CONFIG_SCHED_MC
8956 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8959 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8961 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8962 const char *buf
, size_t count
)
8964 return sched_power_savings_store(buf
, count
, 0);
8966 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8967 sched_mc_power_savings_show
,
8968 sched_mc_power_savings_store
);
8971 #ifdef CONFIG_SCHED_SMT
8972 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8975 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8977 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8978 const char *buf
, size_t count
)
8980 return sched_power_savings_store(buf
, count
, 1);
8982 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8983 sched_smt_power_savings_show
,
8984 sched_smt_power_savings_store
);
8987 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8991 #ifdef CONFIG_SCHED_SMT
8993 err
= sysfs_create_file(&cls
->kset
.kobj
,
8994 &attr_sched_smt_power_savings
.attr
);
8996 #ifdef CONFIG_SCHED_MC
8997 if (!err
&& mc_capable())
8998 err
= sysfs_create_file(&cls
->kset
.kobj
,
8999 &attr_sched_mc_power_savings
.attr
);
9003 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9005 #ifndef CONFIG_CPUSETS
9007 * Add online and remove offline CPUs from the scheduler domains.
9008 * When cpusets are enabled they take over this function.
9010 static int update_sched_domains(struct notifier_block
*nfb
,
9011 unsigned long action
, void *hcpu
)
9015 case CPU_ONLINE_FROZEN
:
9017 case CPU_DEAD_FROZEN
:
9018 partition_sched_domains(1, NULL
, NULL
);
9027 static int update_runtime(struct notifier_block
*nfb
,
9028 unsigned long action
, void *hcpu
)
9030 int cpu
= (int)(long)hcpu
;
9033 case CPU_DOWN_PREPARE
:
9034 case CPU_DOWN_PREPARE_FROZEN
:
9035 disable_runtime(cpu_rq(cpu
));
9038 case CPU_DOWN_FAILED
:
9039 case CPU_DOWN_FAILED_FROZEN
:
9041 case CPU_ONLINE_FROZEN
:
9042 enable_runtime(cpu_rq(cpu
));
9050 void __init
sched_init_smp(void)
9052 cpumask_var_t non_isolated_cpus
;
9054 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9056 #if defined(CONFIG_NUMA)
9057 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9059 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9062 mutex_lock(&sched_domains_mutex
);
9063 arch_init_sched_domains(cpu_online_mask
);
9064 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9065 if (cpumask_empty(non_isolated_cpus
))
9066 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9067 mutex_unlock(&sched_domains_mutex
);
9070 #ifndef CONFIG_CPUSETS
9071 /* XXX: Theoretical race here - CPU may be hotplugged now */
9072 hotcpu_notifier(update_sched_domains
, 0);
9075 /* RT runtime code needs to handle some hotplug events */
9076 hotcpu_notifier(update_runtime
, 0);
9080 /* Move init over to a non-isolated CPU */
9081 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9083 sched_init_granularity();
9084 free_cpumask_var(non_isolated_cpus
);
9086 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9087 init_sched_rt_class();
9090 void __init
sched_init_smp(void)
9092 sched_init_granularity();
9094 #endif /* CONFIG_SMP */
9096 const_debug
unsigned int sysctl_timer_migration
= 1;
9098 int in_sched_functions(unsigned long addr
)
9100 return in_lock_functions(addr
) ||
9101 (addr
>= (unsigned long)__sched_text_start
9102 && addr
< (unsigned long)__sched_text_end
);
9105 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9107 cfs_rq
->tasks_timeline
= RB_ROOT
;
9108 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9109 #ifdef CONFIG_FAIR_GROUP_SCHED
9112 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9115 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9117 struct rt_prio_array
*array
;
9120 array
= &rt_rq
->active
;
9121 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9122 INIT_LIST_HEAD(array
->queue
+ i
);
9123 __clear_bit(i
, array
->bitmap
);
9125 /* delimiter for bitsearch: */
9126 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9128 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9129 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9131 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9135 rt_rq
->rt_nr_migratory
= 0;
9136 rt_rq
->overloaded
= 0;
9137 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9141 rt_rq
->rt_throttled
= 0;
9142 rt_rq
->rt_runtime
= 0;
9143 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9145 #ifdef CONFIG_RT_GROUP_SCHED
9146 rt_rq
->rt_nr_boosted
= 0;
9151 #ifdef CONFIG_FAIR_GROUP_SCHED
9152 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9153 struct sched_entity
*se
, int cpu
, int add
,
9154 struct sched_entity
*parent
)
9156 struct rq
*rq
= cpu_rq(cpu
);
9157 tg
->cfs_rq
[cpu
] = cfs_rq
;
9158 init_cfs_rq(cfs_rq
, rq
);
9161 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9164 /* se could be NULL for init_task_group */
9169 se
->cfs_rq
= &rq
->cfs
;
9171 se
->cfs_rq
= parent
->my_q
;
9174 se
->load
.weight
= tg
->shares
;
9175 se
->load
.inv_weight
= 0;
9176 se
->parent
= parent
;
9180 #ifdef CONFIG_RT_GROUP_SCHED
9181 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9182 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9183 struct sched_rt_entity
*parent
)
9185 struct rq
*rq
= cpu_rq(cpu
);
9187 tg
->rt_rq
[cpu
] = rt_rq
;
9188 init_rt_rq(rt_rq
, rq
);
9190 rt_rq
->rt_se
= rt_se
;
9191 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9193 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9195 tg
->rt_se
[cpu
] = rt_se
;
9200 rt_se
->rt_rq
= &rq
->rt
;
9202 rt_se
->rt_rq
= parent
->my_q
;
9204 rt_se
->my_q
= rt_rq
;
9205 rt_se
->parent
= parent
;
9206 INIT_LIST_HEAD(&rt_se
->run_list
);
9210 void __init
sched_init(void)
9213 unsigned long alloc_size
= 0, ptr
;
9215 #ifdef CONFIG_FAIR_GROUP_SCHED
9216 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9218 #ifdef CONFIG_RT_GROUP_SCHED
9219 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9221 #ifdef CONFIG_USER_SCHED
9224 #ifdef CONFIG_CPUMASK_OFFSTACK
9225 alloc_size
+= num_possible_cpus() * cpumask_size();
9228 * As sched_init() is called before page_alloc is setup,
9229 * we use alloc_bootmem().
9232 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9234 #ifdef CONFIG_FAIR_GROUP_SCHED
9235 init_task_group
.se
= (struct sched_entity
**)ptr
;
9236 ptr
+= nr_cpu_ids
* sizeof(void **);
9238 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9239 ptr
+= nr_cpu_ids
* sizeof(void **);
9241 #ifdef CONFIG_USER_SCHED
9242 root_task_group
.se
= (struct sched_entity
**)ptr
;
9243 ptr
+= nr_cpu_ids
* sizeof(void **);
9245 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9246 ptr
+= nr_cpu_ids
* sizeof(void **);
9247 #endif /* CONFIG_USER_SCHED */
9248 #endif /* CONFIG_FAIR_GROUP_SCHED */
9249 #ifdef CONFIG_RT_GROUP_SCHED
9250 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9251 ptr
+= nr_cpu_ids
* sizeof(void **);
9253 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9254 ptr
+= nr_cpu_ids
* sizeof(void **);
9256 #ifdef CONFIG_USER_SCHED
9257 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9258 ptr
+= nr_cpu_ids
* sizeof(void **);
9260 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9261 ptr
+= nr_cpu_ids
* sizeof(void **);
9262 #endif /* CONFIG_USER_SCHED */
9263 #endif /* CONFIG_RT_GROUP_SCHED */
9264 #ifdef CONFIG_CPUMASK_OFFSTACK
9265 for_each_possible_cpu(i
) {
9266 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9267 ptr
+= cpumask_size();
9269 #endif /* CONFIG_CPUMASK_OFFSTACK */
9273 init_defrootdomain();
9276 init_rt_bandwidth(&def_rt_bandwidth
,
9277 global_rt_period(), global_rt_runtime());
9279 #ifdef CONFIG_RT_GROUP_SCHED
9280 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9281 global_rt_period(), global_rt_runtime());
9282 #ifdef CONFIG_USER_SCHED
9283 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9284 global_rt_period(), RUNTIME_INF
);
9285 #endif /* CONFIG_USER_SCHED */
9286 #endif /* CONFIG_RT_GROUP_SCHED */
9288 #ifdef CONFIG_GROUP_SCHED
9289 list_add(&init_task_group
.list
, &task_groups
);
9290 INIT_LIST_HEAD(&init_task_group
.children
);
9292 #ifdef CONFIG_USER_SCHED
9293 INIT_LIST_HEAD(&root_task_group
.children
);
9294 init_task_group
.parent
= &root_task_group
;
9295 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9296 #endif /* CONFIG_USER_SCHED */
9297 #endif /* CONFIG_GROUP_SCHED */
9299 for_each_possible_cpu(i
) {
9303 spin_lock_init(&rq
->lock
);
9305 rq
->calc_load_active
= 0;
9306 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9307 init_cfs_rq(&rq
->cfs
, rq
);
9308 init_rt_rq(&rq
->rt
, rq
);
9309 #ifdef CONFIG_FAIR_GROUP_SCHED
9310 init_task_group
.shares
= init_task_group_load
;
9311 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9312 #ifdef CONFIG_CGROUP_SCHED
9314 * How much cpu bandwidth does init_task_group get?
9316 * In case of task-groups formed thr' the cgroup filesystem, it
9317 * gets 100% of the cpu resources in the system. This overall
9318 * system cpu resource is divided among the tasks of
9319 * init_task_group and its child task-groups in a fair manner,
9320 * based on each entity's (task or task-group's) weight
9321 * (se->load.weight).
9323 * In other words, if init_task_group has 10 tasks of weight
9324 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9325 * then A0's share of the cpu resource is:
9327 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9329 * We achieve this by letting init_task_group's tasks sit
9330 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9332 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9333 #elif defined CONFIG_USER_SCHED
9334 root_task_group
.shares
= NICE_0_LOAD
;
9335 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9337 * In case of task-groups formed thr' the user id of tasks,
9338 * init_task_group represents tasks belonging to root user.
9339 * Hence it forms a sibling of all subsequent groups formed.
9340 * In this case, init_task_group gets only a fraction of overall
9341 * system cpu resource, based on the weight assigned to root
9342 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9343 * by letting tasks of init_task_group sit in a separate cfs_rq
9344 * (init_cfs_rq) and having one entity represent this group of
9345 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9347 init_tg_cfs_entry(&init_task_group
,
9348 &per_cpu(init_cfs_rq
, i
),
9349 &per_cpu(init_sched_entity
, i
), i
, 1,
9350 root_task_group
.se
[i
]);
9353 #endif /* CONFIG_FAIR_GROUP_SCHED */
9355 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9356 #ifdef CONFIG_RT_GROUP_SCHED
9357 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9358 #ifdef CONFIG_CGROUP_SCHED
9359 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9360 #elif defined CONFIG_USER_SCHED
9361 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9362 init_tg_rt_entry(&init_task_group
,
9363 &per_cpu(init_rt_rq
, i
),
9364 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9365 root_task_group
.rt_se
[i
]);
9369 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9370 rq
->cpu_load
[j
] = 0;
9374 rq
->active_balance
= 0;
9375 rq
->next_balance
= jiffies
;
9379 rq
->migration_thread
= NULL
;
9380 INIT_LIST_HEAD(&rq
->migration_queue
);
9381 rq_attach_root(rq
, &def_root_domain
);
9384 atomic_set(&rq
->nr_iowait
, 0);
9387 set_load_weight(&init_task
);
9389 #ifdef CONFIG_PREEMPT_NOTIFIERS
9390 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9394 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9397 #ifdef CONFIG_RT_MUTEXES
9398 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9402 * The boot idle thread does lazy MMU switching as well:
9404 atomic_inc(&init_mm
.mm_count
);
9405 enter_lazy_tlb(&init_mm
, current
);
9408 * Make us the idle thread. Technically, schedule() should not be
9409 * called from this thread, however somewhere below it might be,
9410 * but because we are the idle thread, we just pick up running again
9411 * when this runqueue becomes "idle".
9413 init_idle(current
, smp_processor_id());
9415 calc_load_update
= jiffies
+ LOAD_FREQ
;
9418 * During early bootup we pretend to be a normal task:
9420 current
->sched_class
= &fair_sched_class
;
9422 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9423 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9426 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9427 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9429 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9432 perf_counter_init();
9434 scheduler_running
= 1;
9437 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9438 void __might_sleep(char *file
, int line
)
9441 static unsigned long prev_jiffy
; /* ratelimiting */
9443 if ((!in_atomic() && !irqs_disabled()) ||
9444 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9446 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9448 prev_jiffy
= jiffies
;
9451 "BUG: sleeping function called from invalid context at %s:%d\n",
9454 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9455 in_atomic(), irqs_disabled(),
9456 current
->pid
, current
->comm
);
9458 debug_show_held_locks(current
);
9459 if (irqs_disabled())
9460 print_irqtrace_events(current
);
9464 EXPORT_SYMBOL(__might_sleep
);
9467 #ifdef CONFIG_MAGIC_SYSRQ
9468 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9472 update_rq_clock(rq
);
9473 on_rq
= p
->se
.on_rq
;
9475 deactivate_task(rq
, p
, 0);
9476 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9478 activate_task(rq
, p
, 0);
9479 resched_task(rq
->curr
);
9483 void normalize_rt_tasks(void)
9485 struct task_struct
*g
, *p
;
9486 unsigned long flags
;
9489 read_lock_irqsave(&tasklist_lock
, flags
);
9490 do_each_thread(g
, p
) {
9492 * Only normalize user tasks:
9497 p
->se
.exec_start
= 0;
9498 #ifdef CONFIG_SCHEDSTATS
9499 p
->se
.wait_start
= 0;
9500 p
->se
.sleep_start
= 0;
9501 p
->se
.block_start
= 0;
9506 * Renice negative nice level userspace
9509 if (TASK_NICE(p
) < 0 && p
->mm
)
9510 set_user_nice(p
, 0);
9514 spin_lock(&p
->pi_lock
);
9515 rq
= __task_rq_lock(p
);
9517 normalize_task(rq
, p
);
9519 __task_rq_unlock(rq
);
9520 spin_unlock(&p
->pi_lock
);
9521 } while_each_thread(g
, p
);
9523 read_unlock_irqrestore(&tasklist_lock
, flags
);
9526 #endif /* CONFIG_MAGIC_SYSRQ */
9530 * These functions are only useful for the IA64 MCA handling.
9532 * They can only be called when the whole system has been
9533 * stopped - every CPU needs to be quiescent, and no scheduling
9534 * activity can take place. Using them for anything else would
9535 * be a serious bug, and as a result, they aren't even visible
9536 * under any other configuration.
9540 * curr_task - return the current task for a given cpu.
9541 * @cpu: the processor in question.
9543 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9545 struct task_struct
*curr_task(int cpu
)
9547 return cpu_curr(cpu
);
9551 * set_curr_task - set the current task for a given cpu.
9552 * @cpu: the processor in question.
9553 * @p: the task pointer to set.
9555 * Description: This function must only be used when non-maskable interrupts
9556 * are serviced on a separate stack. It allows the architecture to switch the
9557 * notion of the current task on a cpu in a non-blocking manner. This function
9558 * must be called with all CPU's synchronized, and interrupts disabled, the
9559 * and caller must save the original value of the current task (see
9560 * curr_task() above) and restore that value before reenabling interrupts and
9561 * re-starting the system.
9563 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9565 void set_curr_task(int cpu
, struct task_struct
*p
)
9572 #ifdef CONFIG_FAIR_GROUP_SCHED
9573 static void free_fair_sched_group(struct task_group
*tg
)
9577 for_each_possible_cpu(i
) {
9579 kfree(tg
->cfs_rq
[i
]);
9589 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9591 struct cfs_rq
*cfs_rq
;
9592 struct sched_entity
*se
;
9596 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9599 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9603 tg
->shares
= NICE_0_LOAD
;
9605 for_each_possible_cpu(i
) {
9608 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9609 GFP_KERNEL
, cpu_to_node(i
));
9613 se
= kzalloc_node(sizeof(struct sched_entity
),
9614 GFP_KERNEL
, cpu_to_node(i
));
9618 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9627 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9629 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9630 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9633 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9635 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9637 #else /* !CONFG_FAIR_GROUP_SCHED */
9638 static inline void free_fair_sched_group(struct task_group
*tg
)
9643 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9648 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9652 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9655 #endif /* CONFIG_FAIR_GROUP_SCHED */
9657 #ifdef CONFIG_RT_GROUP_SCHED
9658 static void free_rt_sched_group(struct task_group
*tg
)
9662 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9664 for_each_possible_cpu(i
) {
9666 kfree(tg
->rt_rq
[i
]);
9668 kfree(tg
->rt_se
[i
]);
9676 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9678 struct rt_rq
*rt_rq
;
9679 struct sched_rt_entity
*rt_se
;
9683 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9686 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9690 init_rt_bandwidth(&tg
->rt_bandwidth
,
9691 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9693 for_each_possible_cpu(i
) {
9696 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9697 GFP_KERNEL
, cpu_to_node(i
));
9701 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9702 GFP_KERNEL
, cpu_to_node(i
));
9706 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9715 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9717 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9718 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9721 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9723 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9725 #else /* !CONFIG_RT_GROUP_SCHED */
9726 static inline void free_rt_sched_group(struct task_group
*tg
)
9731 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9736 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9740 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9743 #endif /* CONFIG_RT_GROUP_SCHED */
9745 #ifdef CONFIG_GROUP_SCHED
9746 static void free_sched_group(struct task_group
*tg
)
9748 free_fair_sched_group(tg
);
9749 free_rt_sched_group(tg
);
9753 /* allocate runqueue etc for a new task group */
9754 struct task_group
*sched_create_group(struct task_group
*parent
)
9756 struct task_group
*tg
;
9757 unsigned long flags
;
9760 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9762 return ERR_PTR(-ENOMEM
);
9764 if (!alloc_fair_sched_group(tg
, parent
))
9767 if (!alloc_rt_sched_group(tg
, parent
))
9770 spin_lock_irqsave(&task_group_lock
, flags
);
9771 for_each_possible_cpu(i
) {
9772 register_fair_sched_group(tg
, i
);
9773 register_rt_sched_group(tg
, i
);
9775 list_add_rcu(&tg
->list
, &task_groups
);
9777 WARN_ON(!parent
); /* root should already exist */
9779 tg
->parent
= parent
;
9780 INIT_LIST_HEAD(&tg
->children
);
9781 list_add_rcu(&tg
->siblings
, &parent
->children
);
9782 spin_unlock_irqrestore(&task_group_lock
, flags
);
9787 free_sched_group(tg
);
9788 return ERR_PTR(-ENOMEM
);
9791 /* rcu callback to free various structures associated with a task group */
9792 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9794 /* now it should be safe to free those cfs_rqs */
9795 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9798 /* Destroy runqueue etc associated with a task group */
9799 void sched_destroy_group(struct task_group
*tg
)
9801 unsigned long flags
;
9804 spin_lock_irqsave(&task_group_lock
, flags
);
9805 for_each_possible_cpu(i
) {
9806 unregister_fair_sched_group(tg
, i
);
9807 unregister_rt_sched_group(tg
, i
);
9809 list_del_rcu(&tg
->list
);
9810 list_del_rcu(&tg
->siblings
);
9811 spin_unlock_irqrestore(&task_group_lock
, flags
);
9813 /* wait for possible concurrent references to cfs_rqs complete */
9814 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9817 /* change task's runqueue when it moves between groups.
9818 * The caller of this function should have put the task in its new group
9819 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9820 * reflect its new group.
9822 void sched_move_task(struct task_struct
*tsk
)
9825 unsigned long flags
;
9828 rq
= task_rq_lock(tsk
, &flags
);
9830 update_rq_clock(rq
);
9832 running
= task_current(rq
, tsk
);
9833 on_rq
= tsk
->se
.on_rq
;
9836 dequeue_task(rq
, tsk
, 0);
9837 if (unlikely(running
))
9838 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9840 set_task_rq(tsk
, task_cpu(tsk
));
9842 #ifdef CONFIG_FAIR_GROUP_SCHED
9843 if (tsk
->sched_class
->moved_group
)
9844 tsk
->sched_class
->moved_group(tsk
);
9847 if (unlikely(running
))
9848 tsk
->sched_class
->set_curr_task(rq
);
9850 enqueue_task(rq
, tsk
, 0);
9852 task_rq_unlock(rq
, &flags
);
9854 #endif /* CONFIG_GROUP_SCHED */
9856 #ifdef CONFIG_FAIR_GROUP_SCHED
9857 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9859 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9864 dequeue_entity(cfs_rq
, se
, 0);
9866 se
->load
.weight
= shares
;
9867 se
->load
.inv_weight
= 0;
9870 enqueue_entity(cfs_rq
, se
, 0);
9873 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9875 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9876 struct rq
*rq
= cfs_rq
->rq
;
9877 unsigned long flags
;
9879 spin_lock_irqsave(&rq
->lock
, flags
);
9880 __set_se_shares(se
, shares
);
9881 spin_unlock_irqrestore(&rq
->lock
, flags
);
9884 static DEFINE_MUTEX(shares_mutex
);
9886 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9889 unsigned long flags
;
9892 * We can't change the weight of the root cgroup.
9897 if (shares
< MIN_SHARES
)
9898 shares
= MIN_SHARES
;
9899 else if (shares
> MAX_SHARES
)
9900 shares
= MAX_SHARES
;
9902 mutex_lock(&shares_mutex
);
9903 if (tg
->shares
== shares
)
9906 spin_lock_irqsave(&task_group_lock
, flags
);
9907 for_each_possible_cpu(i
)
9908 unregister_fair_sched_group(tg
, i
);
9909 list_del_rcu(&tg
->siblings
);
9910 spin_unlock_irqrestore(&task_group_lock
, flags
);
9912 /* wait for any ongoing reference to this group to finish */
9913 synchronize_sched();
9916 * Now we are free to modify the group's share on each cpu
9917 * w/o tripping rebalance_share or load_balance_fair.
9919 tg
->shares
= shares
;
9920 for_each_possible_cpu(i
) {
9924 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9925 set_se_shares(tg
->se
[i
], shares
);
9929 * Enable load balance activity on this group, by inserting it back on
9930 * each cpu's rq->leaf_cfs_rq_list.
9932 spin_lock_irqsave(&task_group_lock
, flags
);
9933 for_each_possible_cpu(i
)
9934 register_fair_sched_group(tg
, i
);
9935 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9936 spin_unlock_irqrestore(&task_group_lock
, flags
);
9938 mutex_unlock(&shares_mutex
);
9942 unsigned long sched_group_shares(struct task_group
*tg
)
9948 #ifdef CONFIG_RT_GROUP_SCHED
9950 * Ensure that the real time constraints are schedulable.
9952 static DEFINE_MUTEX(rt_constraints_mutex
);
9954 static unsigned long to_ratio(u64 period
, u64 runtime
)
9956 if (runtime
== RUNTIME_INF
)
9959 return div64_u64(runtime
<< 20, period
);
9962 /* Must be called with tasklist_lock held */
9963 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9965 struct task_struct
*g
, *p
;
9967 do_each_thread(g
, p
) {
9968 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9970 } while_each_thread(g
, p
);
9975 struct rt_schedulable_data
{
9976 struct task_group
*tg
;
9981 static int tg_schedulable(struct task_group
*tg
, void *data
)
9983 struct rt_schedulable_data
*d
= data
;
9984 struct task_group
*child
;
9985 unsigned long total
, sum
= 0;
9986 u64 period
, runtime
;
9988 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9989 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9992 period
= d
->rt_period
;
9993 runtime
= d
->rt_runtime
;
9996 #ifdef CONFIG_USER_SCHED
9997 if (tg
== &root_task_group
) {
9998 period
= global_rt_period();
9999 runtime
= global_rt_runtime();
10004 * Cannot have more runtime than the period.
10006 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10010 * Ensure we don't starve existing RT tasks.
10012 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10015 total
= to_ratio(period
, runtime
);
10018 * Nobody can have more than the global setting allows.
10020 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10024 * The sum of our children's runtime should not exceed our own.
10026 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10027 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10028 runtime
= child
->rt_bandwidth
.rt_runtime
;
10030 if (child
== d
->tg
) {
10031 period
= d
->rt_period
;
10032 runtime
= d
->rt_runtime
;
10035 sum
+= to_ratio(period
, runtime
);
10044 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10046 struct rt_schedulable_data data
= {
10048 .rt_period
= period
,
10049 .rt_runtime
= runtime
,
10052 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10055 static int tg_set_bandwidth(struct task_group
*tg
,
10056 u64 rt_period
, u64 rt_runtime
)
10060 mutex_lock(&rt_constraints_mutex
);
10061 read_lock(&tasklist_lock
);
10062 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10066 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10067 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10068 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10070 for_each_possible_cpu(i
) {
10071 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10073 spin_lock(&rt_rq
->rt_runtime_lock
);
10074 rt_rq
->rt_runtime
= rt_runtime
;
10075 spin_unlock(&rt_rq
->rt_runtime_lock
);
10077 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10079 read_unlock(&tasklist_lock
);
10080 mutex_unlock(&rt_constraints_mutex
);
10085 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10087 u64 rt_runtime
, rt_period
;
10089 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10090 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10091 if (rt_runtime_us
< 0)
10092 rt_runtime
= RUNTIME_INF
;
10094 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10097 long sched_group_rt_runtime(struct task_group
*tg
)
10101 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10104 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10105 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10106 return rt_runtime_us
;
10109 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10111 u64 rt_runtime
, rt_period
;
10113 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10114 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10116 if (rt_period
== 0)
10119 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10122 long sched_group_rt_period(struct task_group
*tg
)
10126 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10127 do_div(rt_period_us
, NSEC_PER_USEC
);
10128 return rt_period_us
;
10131 static int sched_rt_global_constraints(void)
10133 u64 runtime
, period
;
10136 if (sysctl_sched_rt_period
<= 0)
10139 runtime
= global_rt_runtime();
10140 period
= global_rt_period();
10143 * Sanity check on the sysctl variables.
10145 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10148 mutex_lock(&rt_constraints_mutex
);
10149 read_lock(&tasklist_lock
);
10150 ret
= __rt_schedulable(NULL
, 0, 0);
10151 read_unlock(&tasklist_lock
);
10152 mutex_unlock(&rt_constraints_mutex
);
10157 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10159 /* Don't accept realtime tasks when there is no way for them to run */
10160 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10166 #else /* !CONFIG_RT_GROUP_SCHED */
10167 static int sched_rt_global_constraints(void)
10169 unsigned long flags
;
10172 if (sysctl_sched_rt_period
<= 0)
10176 * There's always some RT tasks in the root group
10177 * -- migration, kstopmachine etc..
10179 if (sysctl_sched_rt_runtime
== 0)
10182 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10183 for_each_possible_cpu(i
) {
10184 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10186 spin_lock(&rt_rq
->rt_runtime_lock
);
10187 rt_rq
->rt_runtime
= global_rt_runtime();
10188 spin_unlock(&rt_rq
->rt_runtime_lock
);
10190 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10194 #endif /* CONFIG_RT_GROUP_SCHED */
10196 int sched_rt_handler(struct ctl_table
*table
, int write
,
10197 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10201 int old_period
, old_runtime
;
10202 static DEFINE_MUTEX(mutex
);
10204 mutex_lock(&mutex
);
10205 old_period
= sysctl_sched_rt_period
;
10206 old_runtime
= sysctl_sched_rt_runtime
;
10208 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10210 if (!ret
&& write
) {
10211 ret
= sched_rt_global_constraints();
10213 sysctl_sched_rt_period
= old_period
;
10214 sysctl_sched_rt_runtime
= old_runtime
;
10216 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10217 def_rt_bandwidth
.rt_period
=
10218 ns_to_ktime(global_rt_period());
10221 mutex_unlock(&mutex
);
10226 #ifdef CONFIG_CGROUP_SCHED
10228 /* return corresponding task_group object of a cgroup */
10229 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10231 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10232 struct task_group
, css
);
10235 static struct cgroup_subsys_state
*
10236 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10238 struct task_group
*tg
, *parent
;
10240 if (!cgrp
->parent
) {
10241 /* This is early initialization for the top cgroup */
10242 return &init_task_group
.css
;
10245 parent
= cgroup_tg(cgrp
->parent
);
10246 tg
= sched_create_group(parent
);
10248 return ERR_PTR(-ENOMEM
);
10254 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10256 struct task_group
*tg
= cgroup_tg(cgrp
);
10258 sched_destroy_group(tg
);
10262 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10263 struct task_struct
*tsk
)
10265 #ifdef CONFIG_RT_GROUP_SCHED
10266 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10269 /* We don't support RT-tasks being in separate groups */
10270 if (tsk
->sched_class
!= &fair_sched_class
)
10278 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10279 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10281 sched_move_task(tsk
);
10284 #ifdef CONFIG_FAIR_GROUP_SCHED
10285 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10288 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10291 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10293 struct task_group
*tg
= cgroup_tg(cgrp
);
10295 return (u64
) tg
->shares
;
10297 #endif /* CONFIG_FAIR_GROUP_SCHED */
10299 #ifdef CONFIG_RT_GROUP_SCHED
10300 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10303 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10306 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10308 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10311 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10314 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10317 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10319 return sched_group_rt_period(cgroup_tg(cgrp
));
10321 #endif /* CONFIG_RT_GROUP_SCHED */
10323 static struct cftype cpu_files
[] = {
10324 #ifdef CONFIG_FAIR_GROUP_SCHED
10327 .read_u64
= cpu_shares_read_u64
,
10328 .write_u64
= cpu_shares_write_u64
,
10331 #ifdef CONFIG_RT_GROUP_SCHED
10333 .name
= "rt_runtime_us",
10334 .read_s64
= cpu_rt_runtime_read
,
10335 .write_s64
= cpu_rt_runtime_write
,
10338 .name
= "rt_period_us",
10339 .read_u64
= cpu_rt_period_read_uint
,
10340 .write_u64
= cpu_rt_period_write_uint
,
10345 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10347 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10350 struct cgroup_subsys cpu_cgroup_subsys
= {
10352 .create
= cpu_cgroup_create
,
10353 .destroy
= cpu_cgroup_destroy
,
10354 .can_attach
= cpu_cgroup_can_attach
,
10355 .attach
= cpu_cgroup_attach
,
10356 .populate
= cpu_cgroup_populate
,
10357 .subsys_id
= cpu_cgroup_subsys_id
,
10361 #endif /* CONFIG_CGROUP_SCHED */
10363 #ifdef CONFIG_CGROUP_CPUACCT
10366 * CPU accounting code for task groups.
10368 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10369 * (balbir@in.ibm.com).
10372 /* track cpu usage of a group of tasks and its child groups */
10374 struct cgroup_subsys_state css
;
10375 /* cpuusage holds pointer to a u64-type object on every cpu */
10377 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10378 struct cpuacct
*parent
;
10381 struct cgroup_subsys cpuacct_subsys
;
10383 /* return cpu accounting group corresponding to this container */
10384 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10386 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10387 struct cpuacct
, css
);
10390 /* return cpu accounting group to which this task belongs */
10391 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10393 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10394 struct cpuacct
, css
);
10397 /* create a new cpu accounting group */
10398 static struct cgroup_subsys_state
*cpuacct_create(
10399 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10401 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10407 ca
->cpuusage
= alloc_percpu(u64
);
10411 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10412 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10413 goto out_free_counters
;
10416 ca
->parent
= cgroup_ca(cgrp
->parent
);
10422 percpu_counter_destroy(&ca
->cpustat
[i
]);
10423 free_percpu(ca
->cpuusage
);
10427 return ERR_PTR(-ENOMEM
);
10430 /* destroy an existing cpu accounting group */
10432 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10434 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10437 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10438 percpu_counter_destroy(&ca
->cpustat
[i
]);
10439 free_percpu(ca
->cpuusage
);
10443 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10445 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10448 #ifndef CONFIG_64BIT
10450 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10452 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10454 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10462 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10464 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10466 #ifndef CONFIG_64BIT
10468 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10470 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10472 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10478 /* return total cpu usage (in nanoseconds) of a group */
10479 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10481 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10482 u64 totalcpuusage
= 0;
10485 for_each_present_cpu(i
)
10486 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10488 return totalcpuusage
;
10491 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10494 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10503 for_each_present_cpu(i
)
10504 cpuacct_cpuusage_write(ca
, i
, 0);
10510 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10511 struct seq_file
*m
)
10513 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10517 for_each_present_cpu(i
) {
10518 percpu
= cpuacct_cpuusage_read(ca
, i
);
10519 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10521 seq_printf(m
, "\n");
10525 static const char *cpuacct_stat_desc
[] = {
10526 [CPUACCT_STAT_USER
] = "user",
10527 [CPUACCT_STAT_SYSTEM
] = "system",
10530 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10531 struct cgroup_map_cb
*cb
)
10533 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10536 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10537 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10538 val
= cputime64_to_clock_t(val
);
10539 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10544 static struct cftype files
[] = {
10547 .read_u64
= cpuusage_read
,
10548 .write_u64
= cpuusage_write
,
10551 .name
= "usage_percpu",
10552 .read_seq_string
= cpuacct_percpu_seq_read
,
10556 .read_map
= cpuacct_stats_show
,
10560 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10562 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10566 * charge this task's execution time to its accounting group.
10568 * called with rq->lock held.
10570 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10572 struct cpuacct
*ca
;
10575 if (unlikely(!cpuacct_subsys
.active
))
10578 cpu
= task_cpu(tsk
);
10584 for (; ca
; ca
= ca
->parent
) {
10585 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10586 *cpuusage
+= cputime
;
10593 * Charge the system/user time to the task's accounting group.
10595 static void cpuacct_update_stats(struct task_struct
*tsk
,
10596 enum cpuacct_stat_index idx
, cputime_t val
)
10598 struct cpuacct
*ca
;
10600 if (unlikely(!cpuacct_subsys
.active
))
10607 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10613 struct cgroup_subsys cpuacct_subsys
= {
10615 .create
= cpuacct_create
,
10616 .destroy
= cpuacct_destroy
,
10617 .populate
= cpuacct_populate
,
10618 .subsys_id
= cpuacct_subsys_id
,
10620 #endif /* CONFIG_CGROUP_CPUACCT */