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/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.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/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
, 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
;
497 struct plist_head pushable_tasks
;
502 /* Nests inside the rq lock: */
503 spinlock_t rt_runtime_lock
;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 unsigned long rt_nr_boosted
;
509 struct list_head leaf_rt_rq_list
;
510 struct task_group
*tg
;
511 struct sched_rt_entity
*rt_se
;
518 * We add the notion of a root-domain which will be used to define per-domain
519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
521 * exclusive cpuset is created, we also create and attach a new root-domain
528 cpumask_var_t online
;
531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
534 cpumask_var_t rto_mask
;
537 struct cpupri cpupri
;
539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545 unsigned int sched_mc_preferred_wakeup_cpu
;
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
553 static struct root_domain def_root_domain
;
558 * This is the main, per-CPU runqueue data structure.
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
572 unsigned long nr_running
;
573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
576 unsigned long last_tick_seen
;
577 unsigned char in_nohz_recently
;
579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load
;
581 unsigned long nr_load_updates
;
587 #ifdef CONFIG_FAIR_GROUP_SCHED
588 /* list of leaf cfs_rq on this cpu: */
589 struct list_head leaf_cfs_rq_list
;
591 #ifdef CONFIG_RT_GROUP_SCHED
592 struct list_head leaf_rt_rq_list
;
596 * This is part of a global counter where only the total sum
597 * over all CPUs matters. A task can increase this counter on
598 * one CPU and if it got migrated afterwards it may decrease
599 * it on another CPU. Always updated under the runqueue lock:
601 unsigned long nr_uninterruptible
;
603 struct task_struct
*curr
, *idle
;
604 unsigned long next_balance
;
605 struct mm_struct
*prev_mm
;
612 struct root_domain
*rd
;
613 struct sched_domain
*sd
;
615 unsigned char idle_at_tick
;
616 /* For active balancing */
619 /* cpu of this runqueue: */
623 unsigned long avg_load_per_task
;
625 struct task_struct
*migration_thread
;
626 struct list_head migration_queue
;
629 #ifdef CONFIG_SCHED_HRTICK
631 int hrtick_csd_pending
;
632 struct call_single_data hrtick_csd
;
634 struct hrtimer hrtick_timer
;
637 #ifdef CONFIG_SCHEDSTATS
639 struct sched_info rq_sched_info
;
640 unsigned long long rq_cpu_time
;
641 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
643 /* sys_sched_yield() stats */
644 unsigned int yld_count
;
646 /* schedule() stats */
647 unsigned int sched_switch
;
648 unsigned int sched_count
;
649 unsigned int sched_goidle
;
651 /* try_to_wake_up() stats */
652 unsigned int ttwu_count
;
653 unsigned int ttwu_local
;
656 unsigned int bkl_count
;
660 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
662 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
664 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
667 static inline int cpu_of(struct rq
*rq
)
677 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
678 * See detach_destroy_domains: synchronize_sched for details.
680 * The domain tree of any CPU may only be accessed from within
681 * preempt-disabled sections.
683 #define for_each_domain(cpu, __sd) \
684 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
686 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
687 #define this_rq() (&__get_cpu_var(runqueues))
688 #define task_rq(p) cpu_rq(task_cpu(p))
689 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
691 static inline void update_rq_clock(struct rq
*rq
)
693 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
697 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
699 #ifdef CONFIG_SCHED_DEBUG
700 # define const_debug __read_mostly
702 # define const_debug static const
708 * Returns true if the current cpu runqueue is locked.
709 * This interface allows printk to be called with the runqueue lock
710 * held and know whether or not it is OK to wake up the klogd.
712 int runqueue_is_locked(void)
715 struct rq
*rq
= cpu_rq(cpu
);
718 ret
= spin_is_locked(&rq
->lock
);
724 * Debugging: various feature bits
727 #define SCHED_FEAT(name, enabled) \
728 __SCHED_FEAT_##name ,
731 #include "sched_features.h"
736 #define SCHED_FEAT(name, enabled) \
737 (1UL << __SCHED_FEAT_##name) * enabled |
739 const_debug
unsigned int sysctl_sched_features
=
740 #include "sched_features.h"
745 #ifdef CONFIG_SCHED_DEBUG
746 #define SCHED_FEAT(name, enabled) \
749 static __read_mostly
char *sched_feat_names
[] = {
750 #include "sched_features.h"
756 static int sched_feat_show(struct seq_file
*m
, void *v
)
760 for (i
= 0; sched_feat_names
[i
]; i
++) {
761 if (!(sysctl_sched_features
& (1UL << i
)))
763 seq_printf(m
, "%s ", sched_feat_names
[i
]);
771 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
772 size_t cnt
, loff_t
*ppos
)
782 if (copy_from_user(&buf
, ubuf
, cnt
))
787 if (strncmp(buf
, "NO_", 3) == 0) {
792 for (i
= 0; sched_feat_names
[i
]; i
++) {
793 int len
= strlen(sched_feat_names
[i
]);
795 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
797 sysctl_sched_features
&= ~(1UL << i
);
799 sysctl_sched_features
|= (1UL << i
);
804 if (!sched_feat_names
[i
])
812 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
814 return single_open(filp
, sched_feat_show
, NULL
);
817 static struct file_operations sched_feat_fops
= {
818 .open
= sched_feat_open
,
819 .write
= sched_feat_write
,
822 .release
= single_release
,
825 static __init
int sched_init_debug(void)
827 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
832 late_initcall(sched_init_debug
);
836 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
839 * Number of tasks to iterate in a single balance run.
840 * Limited because this is done with IRQs disabled.
842 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
845 * ratelimit for updating the group shares.
848 unsigned int sysctl_sched_shares_ratelimit
= 250000;
851 * Inject some fuzzyness into changing the per-cpu group shares
852 * this avoids remote rq-locks at the expense of fairness.
855 unsigned int sysctl_sched_shares_thresh
= 4;
858 * period over which we measure -rt task cpu usage in us.
861 unsigned int sysctl_sched_rt_period
= 1000000;
863 static __read_mostly
int scheduler_running
;
866 * part of the period that we allow rt tasks to run in us.
869 int sysctl_sched_rt_runtime
= 950000;
871 static inline u64
global_rt_period(void)
873 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
876 static inline u64
global_rt_runtime(void)
878 if (sysctl_sched_rt_runtime
< 0)
881 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
884 #ifndef prepare_arch_switch
885 # define prepare_arch_switch(next) do { } while (0)
887 #ifndef finish_arch_switch
888 # define finish_arch_switch(prev) do { } while (0)
891 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
893 return rq
->curr
== p
;
896 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
897 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
899 return task_current(rq
, p
);
902 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
906 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
908 #ifdef CONFIG_DEBUG_SPINLOCK
909 /* this is a valid case when another task releases the spinlock */
910 rq
->lock
.owner
= current
;
913 * If we are tracking spinlock dependencies then we have to
914 * fix up the runqueue lock - which gets 'carried over' from
917 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
919 spin_unlock_irq(&rq
->lock
);
922 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
923 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
928 return task_current(rq
, p
);
932 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
936 * We can optimise this out completely for !SMP, because the
937 * SMP rebalancing from interrupt is the only thing that cares
942 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 spin_unlock_irq(&rq
->lock
);
945 spin_unlock(&rq
->lock
);
949 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
953 * After ->oncpu is cleared, the task can be moved to a different CPU.
954 * We must ensure this doesn't happen until the switch is completely
960 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
964 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
967 * __task_rq_lock - lock the runqueue a given task resides on.
968 * Must be called interrupts disabled.
970 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
974 struct rq
*rq
= task_rq(p
);
975 spin_lock(&rq
->lock
);
976 if (likely(rq
== task_rq(p
)))
978 spin_unlock(&rq
->lock
);
983 * task_rq_lock - lock the runqueue a given task resides on and disable
984 * interrupts. Note the ordering: we can safely lookup the task_rq without
985 * explicitly disabling preemption.
987 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
993 local_irq_save(*flags
);
995 spin_lock(&rq
->lock
);
996 if (likely(rq
== task_rq(p
)))
998 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1002 void task_rq_unlock_wait(struct task_struct
*p
)
1004 struct rq
*rq
= task_rq(p
);
1006 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1007 spin_unlock_wait(&rq
->lock
);
1010 static void __task_rq_unlock(struct rq
*rq
)
1011 __releases(rq
->lock
)
1013 spin_unlock(&rq
->lock
);
1016 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1017 __releases(rq
->lock
)
1019 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1023 * this_rq_lock - lock this runqueue and disable interrupts.
1025 static struct rq
*this_rq_lock(void)
1026 __acquires(rq
->lock
)
1030 local_irq_disable();
1032 spin_lock(&rq
->lock
);
1037 #ifdef CONFIG_SCHED_HRTICK
1039 * Use HR-timers to deliver accurate preemption points.
1041 * Its all a bit involved since we cannot program an hrt while holding the
1042 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1045 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 * - enabled by features
1052 * - hrtimer is actually high res
1054 static inline int hrtick_enabled(struct rq
*rq
)
1056 if (!sched_feat(HRTICK
))
1058 if (!cpu_active(cpu_of(rq
)))
1060 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1063 static void hrtick_clear(struct rq
*rq
)
1065 if (hrtimer_active(&rq
->hrtick_timer
))
1066 hrtimer_cancel(&rq
->hrtick_timer
);
1070 * High-resolution timer tick.
1071 * Runs from hardirq context with interrupts disabled.
1073 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1075 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1077 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1079 spin_lock(&rq
->lock
);
1080 update_rq_clock(rq
);
1081 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1082 spin_unlock(&rq
->lock
);
1084 return HRTIMER_NORESTART
;
1089 * called from hardirq (IPI) context
1091 static void __hrtick_start(void *arg
)
1093 struct rq
*rq
= arg
;
1095 spin_lock(&rq
->lock
);
1096 hrtimer_restart(&rq
->hrtick_timer
);
1097 rq
->hrtick_csd_pending
= 0;
1098 spin_unlock(&rq
->lock
);
1102 * Called to set the hrtick timer state.
1104 * called with rq->lock held and irqs disabled
1106 static void hrtick_start(struct rq
*rq
, u64 delay
)
1108 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1109 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1111 hrtimer_set_expires(timer
, time
);
1113 if (rq
== this_rq()) {
1114 hrtimer_restart(timer
);
1115 } else if (!rq
->hrtick_csd_pending
) {
1116 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1117 rq
->hrtick_csd_pending
= 1;
1122 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1124 int cpu
= (int)(long)hcpu
;
1127 case CPU_UP_CANCELED
:
1128 case CPU_UP_CANCELED_FROZEN
:
1129 case CPU_DOWN_PREPARE
:
1130 case CPU_DOWN_PREPARE_FROZEN
:
1132 case CPU_DEAD_FROZEN
:
1133 hrtick_clear(cpu_rq(cpu
));
1140 static __init
void init_hrtick(void)
1142 hotcpu_notifier(hotplug_hrtick
, 0);
1146 * Called to set the hrtick timer state.
1148 * called with rq->lock held and irqs disabled
1150 static void hrtick_start(struct rq
*rq
, u64 delay
)
1152 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1153 HRTIMER_MODE_REL
, 0);
1156 static inline void init_hrtick(void)
1159 #endif /* CONFIG_SMP */
1161 static void init_rq_hrtick(struct rq
*rq
)
1164 rq
->hrtick_csd_pending
= 0;
1166 rq
->hrtick_csd
.flags
= 0;
1167 rq
->hrtick_csd
.func
= __hrtick_start
;
1168 rq
->hrtick_csd
.info
= rq
;
1171 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1172 rq
->hrtick_timer
.function
= hrtick
;
1174 #else /* CONFIG_SCHED_HRTICK */
1175 static inline void hrtick_clear(struct rq
*rq
)
1179 static inline void init_rq_hrtick(struct rq
*rq
)
1183 static inline void init_hrtick(void)
1186 #endif /* CONFIG_SCHED_HRTICK */
1189 * resched_task - mark a task 'to be rescheduled now'.
1191 * On UP this means the setting of the need_resched flag, on SMP it
1192 * might also involve a cross-CPU call to trigger the scheduler on
1197 #ifndef tsk_is_polling
1198 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1201 static void resched_task(struct task_struct
*p
)
1205 assert_spin_locked(&task_rq(p
)->lock
);
1207 if (test_tsk_need_resched(p
))
1210 set_tsk_need_resched(p
);
1213 if (cpu
== smp_processor_id())
1216 /* NEED_RESCHED must be visible before we test polling */
1218 if (!tsk_is_polling(p
))
1219 smp_send_reschedule(cpu
);
1222 static void resched_cpu(int cpu
)
1224 struct rq
*rq
= cpu_rq(cpu
);
1225 unsigned long flags
;
1227 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1229 resched_task(cpu_curr(cpu
));
1230 spin_unlock_irqrestore(&rq
->lock
, flags
);
1235 * When add_timer_on() enqueues a timer into the timer wheel of an
1236 * idle CPU then this timer might expire before the next timer event
1237 * which is scheduled to wake up that CPU. In case of a completely
1238 * idle system the next event might even be infinite time into the
1239 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1240 * leaves the inner idle loop so the newly added timer is taken into
1241 * account when the CPU goes back to idle and evaluates the timer
1242 * wheel for the next timer event.
1244 void wake_up_idle_cpu(int cpu
)
1246 struct rq
*rq
= cpu_rq(cpu
);
1248 if (cpu
== smp_processor_id())
1252 * This is safe, as this function is called with the timer
1253 * wheel base lock of (cpu) held. When the CPU is on the way
1254 * to idle and has not yet set rq->curr to idle then it will
1255 * be serialized on the timer wheel base lock and take the new
1256 * timer into account automatically.
1258 if (rq
->curr
!= rq
->idle
)
1262 * We can set TIF_RESCHED on the idle task of the other CPU
1263 * lockless. The worst case is that the other CPU runs the
1264 * idle task through an additional NOOP schedule()
1266 set_tsk_need_resched(rq
->idle
);
1268 /* NEED_RESCHED must be visible before we test polling */
1270 if (!tsk_is_polling(rq
->idle
))
1271 smp_send_reschedule(cpu
);
1273 #endif /* CONFIG_NO_HZ */
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct
*p
)
1278 assert_spin_locked(&task_rq(p
)->lock
);
1279 set_tsk_need_resched(p
);
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1301 struct load_weight
*lw
)
1305 if (!lw
->inv_weight
) {
1306 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1309 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1313 tmp
= (u64
)delta_exec
* weight
;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp
> WMULT_CONST
))
1318 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1321 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1323 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1326 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1332 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight
[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult
[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator
{
1400 struct task_struct
*(*start
)(void *);
1401 struct task_struct
*(*next
)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1407 unsigned long max_load_move
, struct sched_domain
*sd
,
1408 enum cpu_idle_type idle
, int *all_pinned
,
1409 int *this_best_prio
, struct rq_iterator
*iterator
);
1412 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1413 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1414 struct rq_iterator
*iterator
);
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index
{
1419 CPUACCT_STAT_USER
, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS
,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1427 static void cpuacct_update_stats(struct task_struct
*tsk
,
1428 enum cpuacct_stat_index idx
, cputime_t val
);
1430 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1431 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1432 enum cpuacct_stat_index idx
, cputime_t val
) {}
1435 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1437 update_load_add(&rq
->load
, load
);
1440 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1442 update_load_sub(&rq
->load
, load
);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor
)(struct task_group
*, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1454 struct task_group
*parent
, *child
;
1458 parent
= &root_task_group
;
1460 ret
= (*down
)(parent
, data
);
1463 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1470 ret
= (*up
)(parent
, data
);
1475 parent
= parent
->parent
;
1484 static int tg_nop(struct task_group
*tg
, void *data
)
1491 static unsigned long source_load(int cpu
, int type
);
1492 static unsigned long target_load(int cpu
, int type
);
1493 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1495 static unsigned long cpu_avg_load_per_task(int cpu
)
1497 struct rq
*rq
= cpu_rq(cpu
);
1498 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1501 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1503 rq
->avg_load_per_task
= 0;
1505 return rq
->avg_load_per_task
;
1508 #ifdef CONFIG_FAIR_GROUP_SCHED
1510 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1513 * Calculate and set the cpu's group shares.
1516 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1517 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1519 unsigned long shares
;
1520 unsigned long rq_weight
;
1525 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1528 * \Sum shares * rq_weight
1529 * shares = -----------------------
1533 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1534 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1536 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1537 sysctl_sched_shares_thresh
) {
1538 struct rq
*rq
= cpu_rq(cpu
);
1539 unsigned long flags
;
1541 spin_lock_irqsave(&rq
->lock
, flags
);
1542 tg
->cfs_rq
[cpu
]->shares
= shares
;
1544 __set_se_shares(tg
->se
[cpu
], shares
);
1545 spin_unlock_irqrestore(&rq
->lock
, flags
);
1550 * Re-compute the task group their per cpu shares over the given domain.
1551 * This needs to be done in a bottom-up fashion because the rq weight of a
1552 * parent group depends on the shares of its child groups.
1554 static int tg_shares_up(struct task_group
*tg
, void *data
)
1556 unsigned long weight
, rq_weight
= 0;
1557 unsigned long shares
= 0;
1558 struct sched_domain
*sd
= data
;
1561 for_each_cpu(i
, sched_domain_span(sd
)) {
1563 * If there are currently no tasks on the cpu pretend there
1564 * is one of average load so that when a new task gets to
1565 * run here it will not get delayed by group starvation.
1567 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1569 weight
= NICE_0_LOAD
;
1571 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1572 rq_weight
+= weight
;
1573 shares
+= tg
->cfs_rq
[i
]->shares
;
1576 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1577 shares
= tg
->shares
;
1579 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1580 shares
= tg
->shares
;
1582 for_each_cpu(i
, sched_domain_span(sd
))
1583 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1589 * Compute the cpu's hierarchical load factor for each task group.
1590 * This needs to be done in a top-down fashion because the load of a child
1591 * group is a fraction of its parents load.
1593 static int tg_load_down(struct task_group
*tg
, void *data
)
1596 long cpu
= (long)data
;
1599 load
= cpu_rq(cpu
)->load
.weight
;
1601 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1602 load
*= tg
->cfs_rq
[cpu
]->shares
;
1603 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1606 tg
->cfs_rq
[cpu
]->h_load
= load
;
1611 static void update_shares(struct sched_domain
*sd
)
1613 u64 now
= cpu_clock(raw_smp_processor_id());
1614 s64 elapsed
= now
- sd
->last_update
;
1616 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1617 sd
->last_update
= now
;
1618 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1622 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1624 spin_unlock(&rq
->lock
);
1626 spin_lock(&rq
->lock
);
1629 static void update_h_load(long cpu
)
1631 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1636 static inline void update_shares(struct sched_domain
*sd
)
1640 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1646 #ifdef CONFIG_PREEMPT
1649 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1650 * way at the expense of forcing extra atomic operations in all
1651 * invocations. This assures that the double_lock is acquired using the
1652 * same underlying policy as the spinlock_t on this architecture, which
1653 * reduces latency compared to the unfair variant below. However, it
1654 * also adds more overhead and therefore may reduce throughput.
1656 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1657 __releases(this_rq
->lock
)
1658 __acquires(busiest
->lock
)
1659 __acquires(this_rq
->lock
)
1661 spin_unlock(&this_rq
->lock
);
1662 double_rq_lock(this_rq
, busiest
);
1669 * Unfair double_lock_balance: Optimizes throughput at the expense of
1670 * latency by eliminating extra atomic operations when the locks are
1671 * already in proper order on entry. This favors lower cpu-ids and will
1672 * grant the double lock to lower cpus over higher ids under contention,
1673 * regardless of entry order into the function.
1675 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1676 __releases(this_rq
->lock
)
1677 __acquires(busiest
->lock
)
1678 __acquires(this_rq
->lock
)
1682 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1683 if (busiest
< this_rq
) {
1684 spin_unlock(&this_rq
->lock
);
1685 spin_lock(&busiest
->lock
);
1686 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1689 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1694 #endif /* CONFIG_PREEMPT */
1697 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1699 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1701 if (unlikely(!irqs_disabled())) {
1702 /* printk() doesn't work good under rq->lock */
1703 spin_unlock(&this_rq
->lock
);
1707 return _double_lock_balance(this_rq
, busiest
);
1710 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1711 __releases(busiest
->lock
)
1713 spin_unlock(&busiest
->lock
);
1714 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1718 #ifdef CONFIG_FAIR_GROUP_SCHED
1719 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1722 cfs_rq
->shares
= shares
;
1727 #include "sched_stats.h"
1728 #include "sched_idletask.c"
1729 #include "sched_fair.c"
1730 #include "sched_rt.c"
1731 #ifdef CONFIG_SCHED_DEBUG
1732 # include "sched_debug.c"
1735 #define sched_class_highest (&rt_sched_class)
1736 #define for_each_class(class) \
1737 for (class = sched_class_highest; class; class = class->next)
1739 static void inc_nr_running(struct rq
*rq
)
1744 static void dec_nr_running(struct rq
*rq
)
1749 static void set_load_weight(struct task_struct
*p
)
1751 if (task_has_rt_policy(p
)) {
1752 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1753 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1758 * SCHED_IDLE tasks get minimal weight:
1760 if (p
->policy
== SCHED_IDLE
) {
1761 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1762 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1766 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1767 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1770 static void update_avg(u64
*avg
, u64 sample
)
1772 s64 diff
= sample
- *avg
;
1776 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1779 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1781 sched_info_queued(p
);
1782 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1786 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1789 if (p
->se
.last_wakeup
) {
1790 update_avg(&p
->se
.avg_overlap
,
1791 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1792 p
->se
.last_wakeup
= 0;
1794 update_avg(&p
->se
.avg_wakeup
,
1795 sysctl_sched_wakeup_granularity
);
1799 sched_info_dequeued(p
);
1800 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1805 * __normal_prio - return the priority that is based on the static prio
1807 static inline int __normal_prio(struct task_struct
*p
)
1809 return p
->static_prio
;
1813 * Calculate the expected normal priority: i.e. priority
1814 * without taking RT-inheritance into account. Might be
1815 * boosted by interactivity modifiers. Changes upon fork,
1816 * setprio syscalls, and whenever the interactivity
1817 * estimator recalculates.
1819 static inline int normal_prio(struct task_struct
*p
)
1823 if (task_has_rt_policy(p
))
1824 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1826 prio
= __normal_prio(p
);
1831 * Calculate the current priority, i.e. the priority
1832 * taken into account by the scheduler. This value might
1833 * be boosted by RT tasks, or might be boosted by
1834 * interactivity modifiers. Will be RT if the task got
1835 * RT-boosted. If not then it returns p->normal_prio.
1837 static int effective_prio(struct task_struct
*p
)
1839 p
->normal_prio
= normal_prio(p
);
1841 * If we are RT tasks or we were boosted to RT priority,
1842 * keep the priority unchanged. Otherwise, update priority
1843 * to the normal priority:
1845 if (!rt_prio(p
->prio
))
1846 return p
->normal_prio
;
1851 * activate_task - move a task to the runqueue.
1853 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1855 if (task_contributes_to_load(p
))
1856 rq
->nr_uninterruptible
--;
1858 enqueue_task(rq
, p
, wakeup
);
1863 * deactivate_task - remove a task from the runqueue.
1865 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1867 if (task_contributes_to_load(p
))
1868 rq
->nr_uninterruptible
++;
1870 dequeue_task(rq
, p
, sleep
);
1875 * task_curr - is this task currently executing on a CPU?
1876 * @p: the task in question.
1878 inline int task_curr(const struct task_struct
*p
)
1880 return cpu_curr(task_cpu(p
)) == p
;
1883 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1885 set_task_rq(p
, cpu
);
1888 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1889 * successfuly executed on another CPU. We must ensure that updates of
1890 * per-task data have been completed by this moment.
1893 task_thread_info(p
)->cpu
= cpu
;
1897 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1898 const struct sched_class
*prev_class
,
1899 int oldprio
, int running
)
1901 if (prev_class
!= p
->sched_class
) {
1902 if (prev_class
->switched_from
)
1903 prev_class
->switched_from(rq
, p
, running
);
1904 p
->sched_class
->switched_to(rq
, p
, running
);
1906 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1911 /* Used instead of source_load when we know the type == 0 */
1912 static unsigned long weighted_cpuload(const int cpu
)
1914 return cpu_rq(cpu
)->load
.weight
;
1918 * Is this task likely cache-hot:
1921 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1926 * Buddy candidates are cache hot:
1928 if (sched_feat(CACHE_HOT_BUDDY
) &&
1929 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1930 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1933 if (p
->sched_class
!= &fair_sched_class
)
1936 if (sysctl_sched_migration_cost
== -1)
1938 if (sysctl_sched_migration_cost
== 0)
1941 delta
= now
- p
->se
.exec_start
;
1943 return delta
< (s64
)sysctl_sched_migration_cost
;
1947 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1949 int old_cpu
= task_cpu(p
);
1950 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1951 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1952 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1955 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1957 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1959 #ifdef CONFIG_SCHEDSTATS
1960 if (p
->se
.wait_start
)
1961 p
->se
.wait_start
-= clock_offset
;
1962 if (p
->se
.sleep_start
)
1963 p
->se
.sleep_start
-= clock_offset
;
1964 if (p
->se
.block_start
)
1965 p
->se
.block_start
-= clock_offset
;
1966 if (old_cpu
!= new_cpu
) {
1967 schedstat_inc(p
, se
.nr_migrations
);
1968 if (task_hot(p
, old_rq
->clock
, NULL
))
1969 schedstat_inc(p
, se
.nr_forced2_migrations
);
1972 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1973 new_cfsrq
->min_vruntime
;
1975 __set_task_cpu(p
, new_cpu
);
1978 struct migration_req
{
1979 struct list_head list
;
1981 struct task_struct
*task
;
1984 struct completion done
;
1988 * The task's runqueue lock must be held.
1989 * Returns true if you have to wait for migration thread.
1992 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1994 struct rq
*rq
= task_rq(p
);
1997 * If the task is not on a runqueue (and not running), then
1998 * it is sufficient to simply update the task's cpu field.
2000 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2001 set_task_cpu(p
, dest_cpu
);
2005 init_completion(&req
->done
);
2007 req
->dest_cpu
= dest_cpu
;
2008 list_add(&req
->list
, &rq
->migration_queue
);
2014 * wait_task_inactive - wait for a thread to unschedule.
2016 * If @match_state is nonzero, it's the @p->state value just checked and
2017 * not expected to change. If it changes, i.e. @p might have woken up,
2018 * then return zero. When we succeed in waiting for @p to be off its CPU,
2019 * we return a positive number (its total switch count). If a second call
2020 * a short while later returns the same number, the caller can be sure that
2021 * @p has remained unscheduled the whole time.
2023 * The caller must ensure that the task *will* unschedule sometime soon,
2024 * else this function might spin for a *long* time. This function can't
2025 * be called with interrupts off, or it may introduce deadlock with
2026 * smp_call_function() if an IPI is sent by the same process we are
2027 * waiting to become inactive.
2029 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2031 unsigned long flags
;
2038 * We do the initial early heuristics without holding
2039 * any task-queue locks at all. We'll only try to get
2040 * the runqueue lock when things look like they will
2046 * If the task is actively running on another CPU
2047 * still, just relax and busy-wait without holding
2050 * NOTE! Since we don't hold any locks, it's not
2051 * even sure that "rq" stays as the right runqueue!
2052 * But we don't care, since "task_running()" will
2053 * return false if the runqueue has changed and p
2054 * is actually now running somewhere else!
2056 while (task_running(rq
, p
)) {
2057 if (match_state
&& unlikely(p
->state
!= match_state
))
2063 * Ok, time to look more closely! We need the rq
2064 * lock now, to be *sure*. If we're wrong, we'll
2065 * just go back and repeat.
2067 rq
= task_rq_lock(p
, &flags
);
2068 trace_sched_wait_task(rq
, p
);
2069 running
= task_running(rq
, p
);
2070 on_rq
= p
->se
.on_rq
;
2072 if (!match_state
|| p
->state
== match_state
)
2073 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2074 task_rq_unlock(rq
, &flags
);
2077 * If it changed from the expected state, bail out now.
2079 if (unlikely(!ncsw
))
2083 * Was it really running after all now that we
2084 * checked with the proper locks actually held?
2086 * Oops. Go back and try again..
2088 if (unlikely(running
)) {
2094 * It's not enough that it's not actively running,
2095 * it must be off the runqueue _entirely_, and not
2098 * So if it was still runnable (but just not actively
2099 * running right now), it's preempted, and we should
2100 * yield - it could be a while.
2102 if (unlikely(on_rq
)) {
2103 schedule_timeout_uninterruptible(1);
2108 * Ahh, all good. It wasn't running, and it wasn't
2109 * runnable, which means that it will never become
2110 * running in the future either. We're all done!
2119 * kick_process - kick a running thread to enter/exit the kernel
2120 * @p: the to-be-kicked thread
2122 * Cause a process which is running on another CPU to enter
2123 * kernel-mode, without any delay. (to get signals handled.)
2125 * NOTE: this function doesnt have to take the runqueue lock,
2126 * because all it wants to ensure is that the remote task enters
2127 * the kernel. If the IPI races and the task has been migrated
2128 * to another CPU then no harm is done and the purpose has been
2131 void kick_process(struct task_struct
*p
)
2137 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2138 smp_send_reschedule(cpu
);
2143 * Return a low guess at the load of a migration-source cpu weighted
2144 * according to the scheduling class and "nice" value.
2146 * We want to under-estimate the load of migration sources, to
2147 * balance conservatively.
2149 static unsigned long source_load(int cpu
, int type
)
2151 struct rq
*rq
= cpu_rq(cpu
);
2152 unsigned long total
= weighted_cpuload(cpu
);
2154 if (type
== 0 || !sched_feat(LB_BIAS
))
2157 return min(rq
->cpu_load
[type
-1], total
);
2161 * Return a high guess at the load of a migration-target cpu weighted
2162 * according to the scheduling class and "nice" value.
2164 static unsigned long target_load(int cpu
, int type
)
2166 struct rq
*rq
= cpu_rq(cpu
);
2167 unsigned long total
= weighted_cpuload(cpu
);
2169 if (type
== 0 || !sched_feat(LB_BIAS
))
2172 return max(rq
->cpu_load
[type
-1], total
);
2176 * find_idlest_group finds and returns the least busy CPU group within the
2179 static struct sched_group
*
2180 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2182 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2183 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2184 int load_idx
= sd
->forkexec_idx
;
2185 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2188 unsigned long load
, avg_load
;
2192 /* Skip over this group if it has no CPUs allowed */
2193 if (!cpumask_intersects(sched_group_cpus(group
),
2197 local_group
= cpumask_test_cpu(this_cpu
,
2198 sched_group_cpus(group
));
2200 /* Tally up the load of all CPUs in the group */
2203 for_each_cpu(i
, sched_group_cpus(group
)) {
2204 /* Bias balancing toward cpus of our domain */
2206 load
= source_load(i
, load_idx
);
2208 load
= target_load(i
, load_idx
);
2213 /* Adjust by relative CPU power of the group */
2214 avg_load
= sg_div_cpu_power(group
,
2215 avg_load
* SCHED_LOAD_SCALE
);
2218 this_load
= avg_load
;
2220 } else if (avg_load
< min_load
) {
2221 min_load
= avg_load
;
2224 } while (group
= group
->next
, group
!= sd
->groups
);
2226 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2232 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2235 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2237 unsigned long load
, min_load
= ULONG_MAX
;
2241 /* Traverse only the allowed CPUs */
2242 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2243 load
= weighted_cpuload(i
);
2245 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2255 * sched_balance_self: balance the current task (running on cpu) in domains
2256 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2259 * Balance, ie. select the least loaded group.
2261 * Returns the target CPU number, or the same CPU if no balancing is needed.
2263 * preempt must be disabled.
2265 static int sched_balance_self(int cpu
, int flag
)
2267 struct task_struct
*t
= current
;
2268 struct sched_domain
*tmp
, *sd
= NULL
;
2270 for_each_domain(cpu
, tmp
) {
2272 * If power savings logic is enabled for a domain, stop there.
2274 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2276 if (tmp
->flags
& flag
)
2284 struct sched_group
*group
;
2285 int new_cpu
, weight
;
2287 if (!(sd
->flags
& flag
)) {
2292 group
= find_idlest_group(sd
, t
, cpu
);
2298 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2299 if (new_cpu
== -1 || new_cpu
== cpu
) {
2300 /* Now try balancing at a lower domain level of cpu */
2305 /* Now try balancing at a lower domain level of new_cpu */
2307 weight
= cpumask_weight(sched_domain_span(sd
));
2309 for_each_domain(cpu
, tmp
) {
2310 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2312 if (tmp
->flags
& flag
)
2315 /* while loop will break here if sd == NULL */
2321 #endif /* CONFIG_SMP */
2324 * try_to_wake_up - wake up a thread
2325 * @p: the to-be-woken-up thread
2326 * @state: the mask of task states that can be woken
2327 * @sync: do a synchronous wakeup?
2329 * Put it on the run-queue if it's not already there. The "current"
2330 * thread is always on the run-queue (except when the actual
2331 * re-schedule is in progress), and as such you're allowed to do
2332 * the simpler "current->state = TASK_RUNNING" to mark yourself
2333 * runnable without the overhead of this.
2335 * returns failure only if the task is already active.
2337 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2339 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2340 unsigned long flags
;
2344 if (!sched_feat(SYNC_WAKEUPS
))
2348 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2349 struct sched_domain
*sd
;
2351 this_cpu
= raw_smp_processor_id();
2354 for_each_domain(this_cpu
, sd
) {
2355 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2364 rq
= task_rq_lock(p
, &flags
);
2365 update_rq_clock(rq
);
2366 old_state
= p
->state
;
2367 if (!(old_state
& state
))
2375 this_cpu
= smp_processor_id();
2378 if (unlikely(task_running(rq
, p
)))
2381 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2382 if (cpu
!= orig_cpu
) {
2383 set_task_cpu(p
, cpu
);
2384 task_rq_unlock(rq
, &flags
);
2385 /* might preempt at this point */
2386 rq
= task_rq_lock(p
, &flags
);
2387 old_state
= p
->state
;
2388 if (!(old_state
& state
))
2393 this_cpu
= smp_processor_id();
2397 #ifdef CONFIG_SCHEDSTATS
2398 schedstat_inc(rq
, ttwu_count
);
2399 if (cpu
== this_cpu
)
2400 schedstat_inc(rq
, ttwu_local
);
2402 struct sched_domain
*sd
;
2403 for_each_domain(this_cpu
, sd
) {
2404 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2405 schedstat_inc(sd
, ttwu_wake_remote
);
2410 #endif /* CONFIG_SCHEDSTATS */
2413 #endif /* CONFIG_SMP */
2414 schedstat_inc(p
, se
.nr_wakeups
);
2416 schedstat_inc(p
, se
.nr_wakeups_sync
);
2417 if (orig_cpu
!= cpu
)
2418 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2419 if (cpu
== this_cpu
)
2420 schedstat_inc(p
, se
.nr_wakeups_local
);
2422 schedstat_inc(p
, se
.nr_wakeups_remote
);
2423 activate_task(rq
, p
, 1);
2427 * Only attribute actual wakeups done by this task.
2429 if (!in_interrupt()) {
2430 struct sched_entity
*se
= ¤t
->se
;
2431 u64 sample
= se
->sum_exec_runtime
;
2433 if (se
->last_wakeup
)
2434 sample
-= se
->last_wakeup
;
2436 sample
-= se
->start_runtime
;
2437 update_avg(&se
->avg_wakeup
, sample
);
2439 se
->last_wakeup
= se
->sum_exec_runtime
;
2443 trace_sched_wakeup(rq
, p
, success
);
2444 check_preempt_curr(rq
, p
, sync
);
2446 p
->state
= TASK_RUNNING
;
2448 if (p
->sched_class
->task_wake_up
)
2449 p
->sched_class
->task_wake_up(rq
, p
);
2452 task_rq_unlock(rq
, &flags
);
2457 int wake_up_process(struct task_struct
*p
)
2459 return try_to_wake_up(p
, TASK_ALL
, 0);
2461 EXPORT_SYMBOL(wake_up_process
);
2463 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2465 return try_to_wake_up(p
, state
, 0);
2469 * Perform scheduler related setup for a newly forked process p.
2470 * p is forked by current.
2472 * __sched_fork() is basic setup used by init_idle() too:
2474 static void __sched_fork(struct task_struct
*p
)
2476 p
->se
.exec_start
= 0;
2477 p
->se
.sum_exec_runtime
= 0;
2478 p
->se
.prev_sum_exec_runtime
= 0;
2479 p
->se
.last_wakeup
= 0;
2480 p
->se
.avg_overlap
= 0;
2481 p
->se
.start_runtime
= 0;
2482 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2484 #ifdef CONFIG_SCHEDSTATS
2485 p
->se
.wait_start
= 0;
2486 p
->se
.sum_sleep_runtime
= 0;
2487 p
->se
.sleep_start
= 0;
2488 p
->se
.block_start
= 0;
2489 p
->se
.sleep_max
= 0;
2490 p
->se
.block_max
= 0;
2492 p
->se
.slice_max
= 0;
2496 INIT_LIST_HEAD(&p
->rt
.run_list
);
2498 INIT_LIST_HEAD(&p
->se
.group_node
);
2500 #ifdef CONFIG_PREEMPT_NOTIFIERS
2501 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2505 * We mark the process as running here, but have not actually
2506 * inserted it onto the runqueue yet. This guarantees that
2507 * nobody will actually run it, and a signal or other external
2508 * event cannot wake it up and insert it on the runqueue either.
2510 p
->state
= TASK_RUNNING
;
2514 * fork()/clone()-time setup:
2516 void sched_fork(struct task_struct
*p
, int clone_flags
)
2518 int cpu
= get_cpu();
2523 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2525 set_task_cpu(p
, cpu
);
2528 * Make sure we do not leak PI boosting priority to the child:
2530 p
->prio
= current
->normal_prio
;
2531 if (!rt_prio(p
->prio
))
2532 p
->sched_class
= &fair_sched_class
;
2534 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2535 if (likely(sched_info_on()))
2536 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2538 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2541 #ifdef CONFIG_PREEMPT
2542 /* Want to start with kernel preemption disabled. */
2543 task_thread_info(p
)->preempt_count
= 1;
2545 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2551 * wake_up_new_task - wake up a newly created task for the first time.
2553 * This function will do some initial scheduler statistics housekeeping
2554 * that must be done for every newly created context, then puts the task
2555 * on the runqueue and wakes it.
2557 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2559 unsigned long flags
;
2562 rq
= task_rq_lock(p
, &flags
);
2563 BUG_ON(p
->state
!= TASK_RUNNING
);
2564 update_rq_clock(rq
);
2566 p
->prio
= effective_prio(p
);
2568 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2569 activate_task(rq
, p
, 0);
2572 * Let the scheduling class do new task startup
2573 * management (if any):
2575 p
->sched_class
->task_new(rq
, p
);
2578 trace_sched_wakeup_new(rq
, p
, 1);
2579 check_preempt_curr(rq
, p
, 0);
2581 if (p
->sched_class
->task_wake_up
)
2582 p
->sched_class
->task_wake_up(rq
, p
);
2584 task_rq_unlock(rq
, &flags
);
2587 #ifdef CONFIG_PREEMPT_NOTIFIERS
2590 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2591 * @notifier: notifier struct to register
2593 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2595 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2597 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2600 * preempt_notifier_unregister - no longer interested in preemption notifications
2601 * @notifier: notifier struct to unregister
2603 * This is safe to call from within a preemption notifier.
2605 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2607 hlist_del(¬ifier
->link
);
2609 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2611 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2613 struct preempt_notifier
*notifier
;
2614 struct hlist_node
*node
;
2616 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2617 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2621 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2622 struct task_struct
*next
)
2624 struct preempt_notifier
*notifier
;
2625 struct hlist_node
*node
;
2627 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2628 notifier
->ops
->sched_out(notifier
, next
);
2631 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2633 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2638 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2639 struct task_struct
*next
)
2643 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2646 * prepare_task_switch - prepare to switch tasks
2647 * @rq: the runqueue preparing to switch
2648 * @prev: the current task that is being switched out
2649 * @next: the task we are going to switch to.
2651 * This is called with the rq lock held and interrupts off. It must
2652 * be paired with a subsequent finish_task_switch after the context
2655 * prepare_task_switch sets up locking and calls architecture specific
2659 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2660 struct task_struct
*next
)
2662 fire_sched_out_preempt_notifiers(prev
, next
);
2663 prepare_lock_switch(rq
, next
);
2664 prepare_arch_switch(next
);
2668 * finish_task_switch - clean up after a task-switch
2669 * @rq: runqueue associated with task-switch
2670 * @prev: the thread we just switched away from.
2672 * finish_task_switch must be called after the context switch, paired
2673 * with a prepare_task_switch call before the context switch.
2674 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2675 * and do any other architecture-specific cleanup actions.
2677 * Note that we may have delayed dropping an mm in context_switch(). If
2678 * so, we finish that here outside of the runqueue lock. (Doing it
2679 * with the lock held can cause deadlocks; see schedule() for
2682 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2683 __releases(rq
->lock
)
2685 struct mm_struct
*mm
= rq
->prev_mm
;
2688 int post_schedule
= 0;
2690 if (current
->sched_class
->needs_post_schedule
)
2691 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2697 * A task struct has one reference for the use as "current".
2698 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2699 * schedule one last time. The schedule call will never return, and
2700 * the scheduled task must drop that reference.
2701 * The test for TASK_DEAD must occur while the runqueue locks are
2702 * still held, otherwise prev could be scheduled on another cpu, die
2703 * there before we look at prev->state, and then the reference would
2705 * Manfred Spraul <manfred@colorfullife.com>
2707 prev_state
= prev
->state
;
2708 finish_arch_switch(prev
);
2709 finish_lock_switch(rq
, prev
);
2712 current
->sched_class
->post_schedule(rq
);
2715 fire_sched_in_preempt_notifiers(current
);
2718 if (unlikely(prev_state
== TASK_DEAD
)) {
2720 * Remove function-return probe instances associated with this
2721 * task and put them back on the free list.
2723 kprobe_flush_task(prev
);
2724 put_task_struct(prev
);
2729 * schedule_tail - first thing a freshly forked thread must call.
2730 * @prev: the thread we just switched away from.
2732 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2733 __releases(rq
->lock
)
2735 struct rq
*rq
= this_rq();
2737 finish_task_switch(rq
, prev
);
2738 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2739 /* In this case, finish_task_switch does not reenable preemption */
2742 if (current
->set_child_tid
)
2743 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2747 * context_switch - switch to the new MM and the new
2748 * thread's register state.
2751 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2752 struct task_struct
*next
)
2754 struct mm_struct
*mm
, *oldmm
;
2756 prepare_task_switch(rq
, prev
, next
);
2757 trace_sched_switch(rq
, prev
, next
);
2759 oldmm
= prev
->active_mm
;
2761 * For paravirt, this is coupled with an exit in switch_to to
2762 * combine the page table reload and the switch backend into
2765 arch_enter_lazy_cpu_mode();
2767 if (unlikely(!mm
)) {
2768 next
->active_mm
= oldmm
;
2769 atomic_inc(&oldmm
->mm_count
);
2770 enter_lazy_tlb(oldmm
, next
);
2772 switch_mm(oldmm
, mm
, next
);
2774 if (unlikely(!prev
->mm
)) {
2775 prev
->active_mm
= NULL
;
2776 rq
->prev_mm
= oldmm
;
2779 * Since the runqueue lock will be released by the next
2780 * task (which is an invalid locking op but in the case
2781 * of the scheduler it's an obvious special-case), so we
2782 * do an early lockdep release here:
2784 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2785 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2788 /* Here we just switch the register state and the stack. */
2789 switch_to(prev
, next
, prev
);
2793 * this_rq must be evaluated again because prev may have moved
2794 * CPUs since it called schedule(), thus the 'rq' on its stack
2795 * frame will be invalid.
2797 finish_task_switch(this_rq(), prev
);
2801 * nr_running, nr_uninterruptible and nr_context_switches:
2803 * externally visible scheduler statistics: current number of runnable
2804 * threads, current number of uninterruptible-sleeping threads, total
2805 * number of context switches performed since bootup.
2807 unsigned long nr_running(void)
2809 unsigned long i
, sum
= 0;
2811 for_each_online_cpu(i
)
2812 sum
+= cpu_rq(i
)->nr_running
;
2817 unsigned long nr_uninterruptible(void)
2819 unsigned long i
, sum
= 0;
2821 for_each_possible_cpu(i
)
2822 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2825 * Since we read the counters lockless, it might be slightly
2826 * inaccurate. Do not allow it to go below zero though:
2828 if (unlikely((long)sum
< 0))
2834 unsigned long long nr_context_switches(void)
2837 unsigned long long sum
= 0;
2839 for_each_possible_cpu(i
)
2840 sum
+= cpu_rq(i
)->nr_switches
;
2845 unsigned long nr_iowait(void)
2847 unsigned long i
, sum
= 0;
2849 for_each_possible_cpu(i
)
2850 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2855 unsigned long nr_active(void)
2857 unsigned long i
, running
= 0, uninterruptible
= 0;
2859 for_each_online_cpu(i
) {
2860 running
+= cpu_rq(i
)->nr_running
;
2861 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2864 if (unlikely((long)uninterruptible
< 0))
2865 uninterruptible
= 0;
2867 return running
+ uninterruptible
;
2871 * Update rq->cpu_load[] statistics. This function is usually called every
2872 * scheduler tick (TICK_NSEC).
2874 static void update_cpu_load(struct rq
*this_rq
)
2876 unsigned long this_load
= this_rq
->load
.weight
;
2879 this_rq
->nr_load_updates
++;
2881 /* Update our load: */
2882 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2883 unsigned long old_load
, new_load
;
2885 /* scale is effectively 1 << i now, and >> i divides by scale */
2887 old_load
= this_rq
->cpu_load
[i
];
2888 new_load
= this_load
;
2890 * Round up the averaging division if load is increasing. This
2891 * prevents us from getting stuck on 9 if the load is 10, for
2894 if (new_load
> old_load
)
2895 new_load
+= scale
-1;
2896 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2903 * double_rq_lock - safely lock two runqueues
2905 * Note this does not disable interrupts like task_rq_lock,
2906 * you need to do so manually before calling.
2908 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2909 __acquires(rq1
->lock
)
2910 __acquires(rq2
->lock
)
2912 BUG_ON(!irqs_disabled());
2914 spin_lock(&rq1
->lock
);
2915 __acquire(rq2
->lock
); /* Fake it out ;) */
2918 spin_lock(&rq1
->lock
);
2919 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2921 spin_lock(&rq2
->lock
);
2922 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2925 update_rq_clock(rq1
);
2926 update_rq_clock(rq2
);
2930 * double_rq_unlock - safely unlock two runqueues
2932 * Note this does not restore interrupts like task_rq_unlock,
2933 * you need to do so manually after calling.
2935 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2936 __releases(rq1
->lock
)
2937 __releases(rq2
->lock
)
2939 spin_unlock(&rq1
->lock
);
2941 spin_unlock(&rq2
->lock
);
2943 __release(rq2
->lock
);
2947 * If dest_cpu is allowed for this process, migrate the task to it.
2948 * This is accomplished by forcing the cpu_allowed mask to only
2949 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2950 * the cpu_allowed mask is restored.
2952 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2954 struct migration_req req
;
2955 unsigned long flags
;
2958 rq
= task_rq_lock(p
, &flags
);
2959 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2960 || unlikely(!cpu_active(dest_cpu
)))
2963 /* force the process onto the specified CPU */
2964 if (migrate_task(p
, dest_cpu
, &req
)) {
2965 /* Need to wait for migration thread (might exit: take ref). */
2966 struct task_struct
*mt
= rq
->migration_thread
;
2968 get_task_struct(mt
);
2969 task_rq_unlock(rq
, &flags
);
2970 wake_up_process(mt
);
2971 put_task_struct(mt
);
2972 wait_for_completion(&req
.done
);
2977 task_rq_unlock(rq
, &flags
);
2981 * sched_exec - execve() is a valuable balancing opportunity, because at
2982 * this point the task has the smallest effective memory and cache footprint.
2984 void sched_exec(void)
2986 int new_cpu
, this_cpu
= get_cpu();
2987 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2989 if (new_cpu
!= this_cpu
)
2990 sched_migrate_task(current
, new_cpu
);
2994 * pull_task - move a task from a remote runqueue to the local runqueue.
2995 * Both runqueues must be locked.
2997 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2998 struct rq
*this_rq
, int this_cpu
)
3000 deactivate_task(src_rq
, p
, 0);
3001 set_task_cpu(p
, this_cpu
);
3002 activate_task(this_rq
, p
, 0);
3004 * Note that idle threads have a prio of MAX_PRIO, for this test
3005 * to be always true for them.
3007 check_preempt_curr(this_rq
, p
, 0);
3011 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3014 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3015 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3018 int tsk_cache_hot
= 0;
3020 * We do not migrate tasks that are:
3021 * 1) running (obviously), or
3022 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3023 * 3) are cache-hot on their current CPU.
3025 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3026 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3031 if (task_running(rq
, p
)) {
3032 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3037 * Aggressive migration if:
3038 * 1) task is cache cold, or
3039 * 2) too many balance attempts have failed.
3042 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3043 if (!tsk_cache_hot
||
3044 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3045 #ifdef CONFIG_SCHEDSTATS
3046 if (tsk_cache_hot
) {
3047 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3048 schedstat_inc(p
, se
.nr_forced_migrations
);
3054 if (tsk_cache_hot
) {
3055 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3061 static unsigned long
3062 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3063 unsigned long max_load_move
, struct sched_domain
*sd
,
3064 enum cpu_idle_type idle
, int *all_pinned
,
3065 int *this_best_prio
, struct rq_iterator
*iterator
)
3067 int loops
= 0, pulled
= 0, pinned
= 0;
3068 struct task_struct
*p
;
3069 long rem_load_move
= max_load_move
;
3071 if (max_load_move
== 0)
3077 * Start the load-balancing iterator:
3079 p
= iterator
->start(iterator
->arg
);
3081 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3084 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3085 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3086 p
= iterator
->next(iterator
->arg
);
3090 pull_task(busiest
, p
, this_rq
, this_cpu
);
3092 rem_load_move
-= p
->se
.load
.weight
;
3094 #ifdef CONFIG_PREEMPT
3096 * NEWIDLE balancing is a source of latency, so preemptible kernels
3097 * will stop after the first task is pulled to minimize the critical
3100 if (idle
== CPU_NEWLY_IDLE
)
3105 * We only want to steal up to the prescribed amount of weighted load.
3107 if (rem_load_move
> 0) {
3108 if (p
->prio
< *this_best_prio
)
3109 *this_best_prio
= p
->prio
;
3110 p
= iterator
->next(iterator
->arg
);
3115 * Right now, this is one of only two places pull_task() is called,
3116 * so we can safely collect pull_task() stats here rather than
3117 * inside pull_task().
3119 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3122 *all_pinned
= pinned
;
3124 return max_load_move
- rem_load_move
;
3128 * move_tasks tries to move up to max_load_move weighted load from busiest to
3129 * this_rq, as part of a balancing operation within domain "sd".
3130 * Returns 1 if successful and 0 otherwise.
3132 * Called with both runqueues locked.
3134 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3135 unsigned long max_load_move
,
3136 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3139 const struct sched_class
*class = sched_class_highest
;
3140 unsigned long total_load_moved
= 0;
3141 int this_best_prio
= this_rq
->curr
->prio
;
3145 class->load_balance(this_rq
, this_cpu
, busiest
,
3146 max_load_move
- total_load_moved
,
3147 sd
, idle
, all_pinned
, &this_best_prio
);
3148 class = class->next
;
3150 #ifdef CONFIG_PREEMPT
3152 * NEWIDLE balancing is a source of latency, so preemptible
3153 * kernels will stop after the first task is pulled to minimize
3154 * the critical section.
3156 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3159 } while (class && max_load_move
> total_load_moved
);
3161 return total_load_moved
> 0;
3165 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3166 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3167 struct rq_iterator
*iterator
)
3169 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3173 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3174 pull_task(busiest
, p
, this_rq
, this_cpu
);
3176 * Right now, this is only the second place pull_task()
3177 * is called, so we can safely collect pull_task()
3178 * stats here rather than inside pull_task().
3180 schedstat_inc(sd
, lb_gained
[idle
]);
3184 p
= iterator
->next(iterator
->arg
);
3191 * move_one_task tries to move exactly one task from busiest to this_rq, as
3192 * part of active balancing operations within "domain".
3193 * Returns 1 if successful and 0 otherwise.
3195 * Called with both runqueues locked.
3197 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3198 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3200 const struct sched_class
*class;
3202 for (class = sched_class_highest
; class; class = class->next
)
3203 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3208 /********** Helpers for find_busiest_group ************************/
3210 * sd_lb_stats - Structure to store the statistics of a sched_domain
3211 * during load balancing.
3213 struct sd_lb_stats
{
3214 struct sched_group
*busiest
; /* Busiest group in this sd */
3215 struct sched_group
*this; /* Local group in this sd */
3216 unsigned long total_load
; /* Total load of all groups in sd */
3217 unsigned long total_pwr
; /* Total power of all groups in sd */
3218 unsigned long avg_load
; /* Average load across all groups in sd */
3220 /** Statistics of this group */
3221 unsigned long this_load
;
3222 unsigned long this_load_per_task
;
3223 unsigned long this_nr_running
;
3225 /* Statistics of the busiest group */
3226 unsigned long max_load
;
3227 unsigned long busiest_load_per_task
;
3228 unsigned long busiest_nr_running
;
3230 int group_imb
; /* Is there imbalance in this sd */
3231 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3232 int power_savings_balance
; /* Is powersave balance needed for this sd */
3233 struct sched_group
*group_min
; /* Least loaded group in sd */
3234 struct sched_group
*group_leader
; /* Group which relieves group_min */
3235 unsigned long min_load_per_task
; /* load_per_task in group_min */
3236 unsigned long leader_nr_running
; /* Nr running of group_leader */
3237 unsigned long min_nr_running
; /* Nr running of group_min */
3242 * sg_lb_stats - stats of a sched_group required for load_balancing
3244 struct sg_lb_stats
{
3245 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3246 unsigned long group_load
; /* Total load over the CPUs of the group */
3247 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3248 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3249 unsigned long group_capacity
;
3250 int group_imb
; /* Is there an imbalance in the group ? */
3254 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3255 * @group: The group whose first cpu is to be returned.
3257 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3259 return cpumask_first(sched_group_cpus(group
));
3263 * get_sd_load_idx - Obtain the load index for a given sched domain.
3264 * @sd: The sched_domain whose load_idx is to be obtained.
3265 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3267 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3268 enum cpu_idle_type idle
)
3274 load_idx
= sd
->busy_idx
;
3277 case CPU_NEWLY_IDLE
:
3278 load_idx
= sd
->newidle_idx
;
3281 load_idx
= sd
->idle_idx
;
3289 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3291 * init_sd_power_savings_stats - Initialize power savings statistics for
3292 * the given sched_domain, during load balancing.
3294 * @sd: Sched domain whose power-savings statistics are to be initialized.
3295 * @sds: Variable containing the statistics for sd.
3296 * @idle: Idle status of the CPU at which we're performing load-balancing.
3298 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3299 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3302 * Busy processors will not participate in power savings
3305 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3306 sds
->power_savings_balance
= 0;
3308 sds
->power_savings_balance
= 1;
3309 sds
->min_nr_running
= ULONG_MAX
;
3310 sds
->leader_nr_running
= 0;
3315 * update_sd_power_savings_stats - Update the power saving stats for a
3316 * sched_domain while performing load balancing.
3318 * @group: sched_group belonging to the sched_domain under consideration.
3319 * @sds: Variable containing the statistics of the sched_domain
3320 * @local_group: Does group contain the CPU for which we're performing
3322 * @sgs: Variable containing the statistics of the group.
3324 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3325 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3328 if (!sds
->power_savings_balance
)
3332 * If the local group is idle or completely loaded
3333 * no need to do power savings balance at this domain
3335 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3336 !sds
->this_nr_running
))
3337 sds
->power_savings_balance
= 0;
3340 * If a group is already running at full capacity or idle,
3341 * don't include that group in power savings calculations
3343 if (!sds
->power_savings_balance
||
3344 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3345 !sgs
->sum_nr_running
)
3349 * Calculate the group which has the least non-idle load.
3350 * This is the group from where we need to pick up the load
3353 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3354 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3355 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3356 sds
->group_min
= group
;
3357 sds
->min_nr_running
= sgs
->sum_nr_running
;
3358 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3359 sgs
->sum_nr_running
;
3363 * Calculate the group which is almost near its
3364 * capacity but still has some space to pick up some load
3365 * from other group and save more power
3367 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3370 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3371 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3372 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3373 sds
->group_leader
= group
;
3374 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3379 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3380 * @sds: Variable containing the statistics of the sched_domain
3381 * under consideration.
3382 * @this_cpu: Cpu at which we're currently performing load-balancing.
3383 * @imbalance: Variable to store the imbalance.
3386 * Check if we have potential to perform some power-savings balance.
3387 * If yes, set the busiest group to be the least loaded group in the
3388 * sched_domain, so that it's CPUs can be put to idle.
3390 * Returns 1 if there is potential to perform power-savings balance.
3393 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3394 int this_cpu
, unsigned long *imbalance
)
3396 if (!sds
->power_savings_balance
)
3399 if (sds
->this != sds
->group_leader
||
3400 sds
->group_leader
== sds
->group_min
)
3403 *imbalance
= sds
->min_load_per_task
;
3404 sds
->busiest
= sds
->group_min
;
3406 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3407 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3408 group_first_cpu(sds
->group_leader
);
3414 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3415 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3416 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3421 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3422 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3427 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3428 int this_cpu
, unsigned long *imbalance
)
3432 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3436 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3437 * @group: sched_group whose statistics are to be updated.
3438 * @this_cpu: Cpu for which load balance is currently performed.
3439 * @idle: Idle status of this_cpu
3440 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3441 * @sd_idle: Idle status of the sched_domain containing group.
3442 * @local_group: Does group contain this_cpu.
3443 * @cpus: Set of cpus considered for load balancing.
3444 * @balance: Should we balance.
3445 * @sgs: variable to hold the statistics for this group.
3447 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3448 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3449 int local_group
, const struct cpumask
*cpus
,
3450 int *balance
, struct sg_lb_stats
*sgs
)
3452 unsigned long load
, max_cpu_load
, min_cpu_load
;
3454 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3455 unsigned long sum_avg_load_per_task
;
3456 unsigned long avg_load_per_task
;
3459 balance_cpu
= group_first_cpu(group
);
3461 /* Tally up the load of all CPUs in the group */
3462 sum_avg_load_per_task
= avg_load_per_task
= 0;
3464 min_cpu_load
= ~0UL;
3466 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3467 struct rq
*rq
= cpu_rq(i
);
3469 if (*sd_idle
&& rq
->nr_running
)
3472 /* Bias balancing toward cpus of our domain */
3474 if (idle_cpu(i
) && !first_idle_cpu
) {
3479 load
= target_load(i
, load_idx
);
3481 load
= source_load(i
, load_idx
);
3482 if (load
> max_cpu_load
)
3483 max_cpu_load
= load
;
3484 if (min_cpu_load
> load
)
3485 min_cpu_load
= load
;
3488 sgs
->group_load
+= load
;
3489 sgs
->sum_nr_running
+= rq
->nr_running
;
3490 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3492 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3496 * First idle cpu or the first cpu(busiest) in this sched group
3497 * is eligible for doing load balancing at this and above
3498 * domains. In the newly idle case, we will allow all the cpu's
3499 * to do the newly idle load balance.
3501 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3502 balance_cpu
!= this_cpu
&& balance
) {
3507 /* Adjust by relative CPU power of the group */
3508 sgs
->avg_load
= sg_div_cpu_power(group
,
3509 sgs
->group_load
* SCHED_LOAD_SCALE
);
3513 * Consider the group unbalanced when the imbalance is larger
3514 * than the average weight of two tasks.
3516 * APZ: with cgroup the avg task weight can vary wildly and
3517 * might not be a suitable number - should we keep a
3518 * normalized nr_running number somewhere that negates
3521 avg_load_per_task
= sg_div_cpu_power(group
,
3522 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3524 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3527 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3532 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3533 * @sd: sched_domain whose statistics are to be updated.
3534 * @this_cpu: Cpu for which load balance is currently performed.
3535 * @idle: Idle status of this_cpu
3536 * @sd_idle: Idle status of the sched_domain containing group.
3537 * @cpus: Set of cpus considered for load balancing.
3538 * @balance: Should we balance.
3539 * @sds: variable to hold the statistics for this sched_domain.
3541 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3542 enum cpu_idle_type idle
, int *sd_idle
,
3543 const struct cpumask
*cpus
, int *balance
,
3544 struct sd_lb_stats
*sds
)
3546 struct sched_group
*group
= sd
->groups
;
3547 struct sg_lb_stats sgs
;
3550 init_sd_power_savings_stats(sd
, sds
, idle
);
3551 load_idx
= get_sd_load_idx(sd
, idle
);
3556 local_group
= cpumask_test_cpu(this_cpu
,
3557 sched_group_cpus(group
));
3558 memset(&sgs
, 0, sizeof(sgs
));
3559 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3560 local_group
, cpus
, balance
, &sgs
);
3562 if (local_group
&& balance
&& !(*balance
))
3565 sds
->total_load
+= sgs
.group_load
;
3566 sds
->total_pwr
+= group
->__cpu_power
;
3569 sds
->this_load
= sgs
.avg_load
;
3571 sds
->this_nr_running
= sgs
.sum_nr_running
;
3572 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3573 } else if (sgs
.avg_load
> sds
->max_load
&&
3574 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3576 sds
->max_load
= sgs
.avg_load
;
3577 sds
->busiest
= group
;
3578 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3579 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3580 sds
->group_imb
= sgs
.group_imb
;
3583 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3584 group
= group
->next
;
3585 } while (group
!= sd
->groups
);
3590 * fix_small_imbalance - Calculate the minor imbalance that exists
3591 * amongst the groups of a sched_domain, during
3593 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3594 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3595 * @imbalance: Variable to store the imbalance.
3597 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3598 int this_cpu
, unsigned long *imbalance
)
3600 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3601 unsigned int imbn
= 2;
3603 if (sds
->this_nr_running
) {
3604 sds
->this_load_per_task
/= sds
->this_nr_running
;
3605 if (sds
->busiest_load_per_task
>
3606 sds
->this_load_per_task
)
3609 sds
->this_load_per_task
=
3610 cpu_avg_load_per_task(this_cpu
);
3612 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3613 sds
->busiest_load_per_task
* imbn
) {
3614 *imbalance
= sds
->busiest_load_per_task
;
3619 * OK, we don't have enough imbalance to justify moving tasks,
3620 * however we may be able to increase total CPU power used by
3624 pwr_now
+= sds
->busiest
->__cpu_power
*
3625 min(sds
->busiest_load_per_task
, sds
->max_load
);
3626 pwr_now
+= sds
->this->__cpu_power
*
3627 min(sds
->this_load_per_task
, sds
->this_load
);
3628 pwr_now
/= SCHED_LOAD_SCALE
;
3630 /* Amount of load we'd subtract */
3631 tmp
= sg_div_cpu_power(sds
->busiest
,
3632 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3633 if (sds
->max_load
> tmp
)
3634 pwr_move
+= sds
->busiest
->__cpu_power
*
3635 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3637 /* Amount of load we'd add */
3638 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3639 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3640 tmp
= sg_div_cpu_power(sds
->this,
3641 sds
->max_load
* sds
->busiest
->__cpu_power
);
3643 tmp
= sg_div_cpu_power(sds
->this,
3644 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3645 pwr_move
+= sds
->this->__cpu_power
*
3646 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3647 pwr_move
/= SCHED_LOAD_SCALE
;
3649 /* Move if we gain throughput */
3650 if (pwr_move
> pwr_now
)
3651 *imbalance
= sds
->busiest_load_per_task
;
3655 * calculate_imbalance - Calculate the amount of imbalance present within the
3656 * groups of a given sched_domain during load balance.
3657 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3658 * @this_cpu: Cpu for which currently load balance is being performed.
3659 * @imbalance: The variable to store the imbalance.
3661 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3662 unsigned long *imbalance
)
3664 unsigned long max_pull
;
3666 * In the presence of smp nice balancing, certain scenarios can have
3667 * max load less than avg load(as we skip the groups at or below
3668 * its cpu_power, while calculating max_load..)
3670 if (sds
->max_load
< sds
->avg_load
) {
3672 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3675 /* Don't want to pull so many tasks that a group would go idle */
3676 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3677 sds
->max_load
- sds
->busiest_load_per_task
);
3679 /* How much load to actually move to equalise the imbalance */
3680 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3681 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3685 * if *imbalance is less than the average load per runnable task
3686 * there is no gaurantee that any tasks will be moved so we'll have
3687 * a think about bumping its value to force at least one task to be
3690 if (*imbalance
< sds
->busiest_load_per_task
)
3691 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3694 /******* find_busiest_group() helpers end here *********************/
3697 * find_busiest_group - Returns the busiest group within the sched_domain
3698 * if there is an imbalance. If there isn't an imbalance, and
3699 * the user has opted for power-savings, it returns a group whose
3700 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3701 * such a group exists.
3703 * Also calculates the amount of weighted load which should be moved
3704 * to restore balance.
3706 * @sd: The sched_domain whose busiest group is to be returned.
3707 * @this_cpu: The cpu for which load balancing is currently being performed.
3708 * @imbalance: Variable which stores amount of weighted load which should
3709 * be moved to restore balance/put a group to idle.
3710 * @idle: The idle status of this_cpu.
3711 * @sd_idle: The idleness of sd
3712 * @cpus: The set of CPUs under consideration for load-balancing.
3713 * @balance: Pointer to a variable indicating if this_cpu
3714 * is the appropriate cpu to perform load balancing at this_level.
3716 * Returns: - the busiest group if imbalance exists.
3717 * - If no imbalance and user has opted for power-savings balance,
3718 * return the least loaded group whose CPUs can be
3719 * put to idle by rebalancing its tasks onto our group.
3721 static struct sched_group
*
3722 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3723 unsigned long *imbalance
, enum cpu_idle_type idle
,
3724 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3726 struct sd_lb_stats sds
;
3728 memset(&sds
, 0, sizeof(sds
));
3731 * Compute the various statistics relavent for load balancing at
3734 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3737 /* Cases where imbalance does not exist from POV of this_cpu */
3738 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3740 * 2) There is no busy sibling group to pull from.
3741 * 3) This group is the busiest group.
3742 * 4) This group is more busy than the avg busieness at this
3744 * 5) The imbalance is within the specified limit.
3745 * 6) Any rebalance would lead to ping-pong
3747 if (balance
&& !(*balance
))
3750 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3753 if (sds
.this_load
>= sds
.max_load
)
3756 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3758 if (sds
.this_load
>= sds
.avg_load
)
3761 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3764 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3766 sds
.busiest_load_per_task
=
3767 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3770 * We're trying to get all the cpus to the average_load, so we don't
3771 * want to push ourselves above the average load, nor do we wish to
3772 * reduce the max loaded cpu below the average load, as either of these
3773 * actions would just result in more rebalancing later, and ping-pong
3774 * tasks around. Thus we look for the minimum possible imbalance.
3775 * Negative imbalances (*we* are more loaded than anyone else) will
3776 * be counted as no imbalance for these purposes -- we can't fix that
3777 * by pulling tasks to us. Be careful of negative numbers as they'll
3778 * appear as very large values with unsigned longs.
3780 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3783 /* Looks like there is an imbalance. Compute it */
3784 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3789 * There is no obvious imbalance. But check if we can do some balancing
3792 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3800 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3803 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3804 unsigned long imbalance
, const struct cpumask
*cpus
)
3806 struct rq
*busiest
= NULL
, *rq
;
3807 unsigned long max_load
= 0;
3810 for_each_cpu(i
, sched_group_cpus(group
)) {
3813 if (!cpumask_test_cpu(i
, cpus
))
3817 wl
= weighted_cpuload(i
);
3819 if (rq
->nr_running
== 1 && wl
> imbalance
)
3822 if (wl
> max_load
) {
3832 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3833 * so long as it is large enough.
3835 #define MAX_PINNED_INTERVAL 512
3837 /* Working cpumask for load_balance and load_balance_newidle. */
3838 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3841 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3842 * tasks if there is an imbalance.
3844 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3845 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3848 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3849 struct sched_group
*group
;
3850 unsigned long imbalance
;
3852 unsigned long flags
;
3853 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3855 cpumask_setall(cpus
);
3858 * When power savings policy is enabled for the parent domain, idle
3859 * sibling can pick up load irrespective of busy siblings. In this case,
3860 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3861 * portraying it as CPU_NOT_IDLE.
3863 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3864 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3867 schedstat_inc(sd
, lb_count
[idle
]);
3871 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3878 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3882 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3884 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3888 BUG_ON(busiest
== this_rq
);
3890 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3893 if (busiest
->nr_running
> 1) {
3895 * Attempt to move tasks. If find_busiest_group has found
3896 * an imbalance but busiest->nr_running <= 1, the group is
3897 * still unbalanced. ld_moved simply stays zero, so it is
3898 * correctly treated as an imbalance.
3900 local_irq_save(flags
);
3901 double_rq_lock(this_rq
, busiest
);
3902 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3903 imbalance
, sd
, idle
, &all_pinned
);
3904 double_rq_unlock(this_rq
, busiest
);
3905 local_irq_restore(flags
);
3908 * some other cpu did the load balance for us.
3910 if (ld_moved
&& this_cpu
!= smp_processor_id())
3911 resched_cpu(this_cpu
);
3913 /* All tasks on this runqueue were pinned by CPU affinity */
3914 if (unlikely(all_pinned
)) {
3915 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3916 if (!cpumask_empty(cpus
))
3923 schedstat_inc(sd
, lb_failed
[idle
]);
3924 sd
->nr_balance_failed
++;
3926 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3928 spin_lock_irqsave(&busiest
->lock
, flags
);
3930 /* don't kick the migration_thread, if the curr
3931 * task on busiest cpu can't be moved to this_cpu
3933 if (!cpumask_test_cpu(this_cpu
,
3934 &busiest
->curr
->cpus_allowed
)) {
3935 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3937 goto out_one_pinned
;
3940 if (!busiest
->active_balance
) {
3941 busiest
->active_balance
= 1;
3942 busiest
->push_cpu
= this_cpu
;
3945 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3947 wake_up_process(busiest
->migration_thread
);
3950 * We've kicked active balancing, reset the failure
3953 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3956 sd
->nr_balance_failed
= 0;
3958 if (likely(!active_balance
)) {
3959 /* We were unbalanced, so reset the balancing interval */
3960 sd
->balance_interval
= sd
->min_interval
;
3963 * If we've begun active balancing, start to back off. This
3964 * case may not be covered by the all_pinned logic if there
3965 * is only 1 task on the busy runqueue (because we don't call
3968 if (sd
->balance_interval
< sd
->max_interval
)
3969 sd
->balance_interval
*= 2;
3972 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3973 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3979 schedstat_inc(sd
, lb_balanced
[idle
]);
3981 sd
->nr_balance_failed
= 0;
3984 /* tune up the balancing interval */
3985 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3986 (sd
->balance_interval
< sd
->max_interval
))
3987 sd
->balance_interval
*= 2;
3989 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3990 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4001 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4002 * tasks if there is an imbalance.
4004 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4005 * this_rq is locked.
4008 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4010 struct sched_group
*group
;
4011 struct rq
*busiest
= NULL
;
4012 unsigned long imbalance
;
4016 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4018 cpumask_setall(cpus
);
4021 * When power savings policy is enabled for the parent domain, idle
4022 * sibling can pick up load irrespective of busy siblings. In this case,
4023 * let the state of idle sibling percolate up as IDLE, instead of
4024 * portraying it as CPU_NOT_IDLE.
4026 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4027 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4030 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4032 update_shares_locked(this_rq
, sd
);
4033 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4034 &sd_idle
, cpus
, NULL
);
4036 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4040 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4042 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4046 BUG_ON(busiest
== this_rq
);
4048 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4051 if (busiest
->nr_running
> 1) {
4052 /* Attempt to move tasks */
4053 double_lock_balance(this_rq
, busiest
);
4054 /* this_rq->clock is already updated */
4055 update_rq_clock(busiest
);
4056 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4057 imbalance
, sd
, CPU_NEWLY_IDLE
,
4059 double_unlock_balance(this_rq
, busiest
);
4061 if (unlikely(all_pinned
)) {
4062 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4063 if (!cpumask_empty(cpus
))
4069 int active_balance
= 0;
4071 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4072 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4073 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4076 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4079 if (sd
->nr_balance_failed
++ < 2)
4083 * The only task running in a non-idle cpu can be moved to this
4084 * cpu in an attempt to completely freeup the other CPU
4085 * package. The same method used to move task in load_balance()
4086 * have been extended for load_balance_newidle() to speedup
4087 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4089 * The package power saving logic comes from
4090 * find_busiest_group(). If there are no imbalance, then
4091 * f_b_g() will return NULL. However when sched_mc={1,2} then
4092 * f_b_g() will select a group from which a running task may be
4093 * pulled to this cpu in order to make the other package idle.
4094 * If there is no opportunity to make a package idle and if
4095 * there are no imbalance, then f_b_g() will return NULL and no
4096 * action will be taken in load_balance_newidle().
4098 * Under normal task pull operation due to imbalance, there
4099 * will be more than one task in the source run queue and
4100 * move_tasks() will succeed. ld_moved will be true and this
4101 * active balance code will not be triggered.
4104 /* Lock busiest in correct order while this_rq is held */
4105 double_lock_balance(this_rq
, busiest
);
4108 * don't kick the migration_thread, if the curr
4109 * task on busiest cpu can't be moved to this_cpu
4111 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4112 double_unlock_balance(this_rq
, busiest
);
4117 if (!busiest
->active_balance
) {
4118 busiest
->active_balance
= 1;
4119 busiest
->push_cpu
= this_cpu
;
4123 double_unlock_balance(this_rq
, busiest
);
4125 * Should not call ttwu while holding a rq->lock
4127 spin_unlock(&this_rq
->lock
);
4129 wake_up_process(busiest
->migration_thread
);
4130 spin_lock(&this_rq
->lock
);
4133 sd
->nr_balance_failed
= 0;
4135 update_shares_locked(this_rq
, sd
);
4139 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4140 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4141 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4143 sd
->nr_balance_failed
= 0;
4149 * idle_balance is called by schedule() if this_cpu is about to become
4150 * idle. Attempts to pull tasks from other CPUs.
4152 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4154 struct sched_domain
*sd
;
4155 int pulled_task
= 0;
4156 unsigned long next_balance
= jiffies
+ HZ
;
4158 for_each_domain(this_cpu
, sd
) {
4159 unsigned long interval
;
4161 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4164 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4165 /* If we've pulled tasks over stop searching: */
4166 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4169 interval
= msecs_to_jiffies(sd
->balance_interval
);
4170 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4171 next_balance
= sd
->last_balance
+ interval
;
4175 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4177 * We are going idle. next_balance may be set based on
4178 * a busy processor. So reset next_balance.
4180 this_rq
->next_balance
= next_balance
;
4185 * active_load_balance is run by migration threads. It pushes running tasks
4186 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4187 * running on each physical CPU where possible, and avoids physical /
4188 * logical imbalances.
4190 * Called with busiest_rq locked.
4192 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4194 int target_cpu
= busiest_rq
->push_cpu
;
4195 struct sched_domain
*sd
;
4196 struct rq
*target_rq
;
4198 /* Is there any task to move? */
4199 if (busiest_rq
->nr_running
<= 1)
4202 target_rq
= cpu_rq(target_cpu
);
4205 * This condition is "impossible", if it occurs
4206 * we need to fix it. Originally reported by
4207 * Bjorn Helgaas on a 128-cpu setup.
4209 BUG_ON(busiest_rq
== target_rq
);
4211 /* move a task from busiest_rq to target_rq */
4212 double_lock_balance(busiest_rq
, target_rq
);
4213 update_rq_clock(busiest_rq
);
4214 update_rq_clock(target_rq
);
4216 /* Search for an sd spanning us and the target CPU. */
4217 for_each_domain(target_cpu
, sd
) {
4218 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4219 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4224 schedstat_inc(sd
, alb_count
);
4226 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4228 schedstat_inc(sd
, alb_pushed
);
4230 schedstat_inc(sd
, alb_failed
);
4232 double_unlock_balance(busiest_rq
, target_rq
);
4237 atomic_t load_balancer
;
4238 cpumask_var_t cpu_mask
;
4239 } nohz ____cacheline_aligned
= {
4240 .load_balancer
= ATOMIC_INIT(-1),
4244 * This routine will try to nominate the ilb (idle load balancing)
4245 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4246 * load balancing on behalf of all those cpus. If all the cpus in the system
4247 * go into this tickless mode, then there will be no ilb owner (as there is
4248 * no need for one) and all the cpus will sleep till the next wakeup event
4251 * For the ilb owner, tick is not stopped. And this tick will be used
4252 * for idle load balancing. ilb owner will still be part of
4255 * While stopping the tick, this cpu will become the ilb owner if there
4256 * is no other owner. And will be the owner till that cpu becomes busy
4257 * or if all cpus in the system stop their ticks at which point
4258 * there is no need for ilb owner.
4260 * When the ilb owner becomes busy, it nominates another owner, during the
4261 * next busy scheduler_tick()
4263 int select_nohz_load_balancer(int stop_tick
)
4265 int cpu
= smp_processor_id();
4268 cpu_rq(cpu
)->in_nohz_recently
= 1;
4270 if (!cpu_active(cpu
)) {
4271 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4275 * If we are going offline and still the leader,
4278 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4284 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4286 /* time for ilb owner also to sleep */
4287 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4288 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4289 atomic_set(&nohz
.load_balancer
, -1);
4293 if (atomic_read(&nohz
.load_balancer
) == -1) {
4294 /* make me the ilb owner */
4295 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4297 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4300 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4303 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4305 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4306 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4313 static DEFINE_SPINLOCK(balancing
);
4316 * It checks each scheduling domain to see if it is due to be balanced,
4317 * and initiates a balancing operation if so.
4319 * Balancing parameters are set up in arch_init_sched_domains.
4321 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4324 struct rq
*rq
= cpu_rq(cpu
);
4325 unsigned long interval
;
4326 struct sched_domain
*sd
;
4327 /* Earliest time when we have to do rebalance again */
4328 unsigned long next_balance
= jiffies
+ 60*HZ
;
4329 int update_next_balance
= 0;
4332 for_each_domain(cpu
, sd
) {
4333 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4336 interval
= sd
->balance_interval
;
4337 if (idle
!= CPU_IDLE
)
4338 interval
*= sd
->busy_factor
;
4340 /* scale ms to jiffies */
4341 interval
= msecs_to_jiffies(interval
);
4342 if (unlikely(!interval
))
4344 if (interval
> HZ
*NR_CPUS
/10)
4345 interval
= HZ
*NR_CPUS
/10;
4347 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4349 if (need_serialize
) {
4350 if (!spin_trylock(&balancing
))
4354 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4355 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4357 * We've pulled tasks over so either we're no
4358 * longer idle, or one of our SMT siblings is
4361 idle
= CPU_NOT_IDLE
;
4363 sd
->last_balance
= jiffies
;
4366 spin_unlock(&balancing
);
4368 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4369 next_balance
= sd
->last_balance
+ interval
;
4370 update_next_balance
= 1;
4374 * Stop the load balance at this level. There is another
4375 * CPU in our sched group which is doing load balancing more
4383 * next_balance will be updated only when there is a need.
4384 * When the cpu is attached to null domain for ex, it will not be
4387 if (likely(update_next_balance
))
4388 rq
->next_balance
= next_balance
;
4392 * run_rebalance_domains is triggered when needed from the scheduler tick.
4393 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4394 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4396 static void run_rebalance_domains(struct softirq_action
*h
)
4398 int this_cpu
= smp_processor_id();
4399 struct rq
*this_rq
= cpu_rq(this_cpu
);
4400 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4401 CPU_IDLE
: CPU_NOT_IDLE
;
4403 rebalance_domains(this_cpu
, idle
);
4407 * If this cpu is the owner for idle load balancing, then do the
4408 * balancing on behalf of the other idle cpus whose ticks are
4411 if (this_rq
->idle_at_tick
&&
4412 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4416 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4417 if (balance_cpu
== this_cpu
)
4421 * If this cpu gets work to do, stop the load balancing
4422 * work being done for other cpus. Next load
4423 * balancing owner will pick it up.
4428 rebalance_domains(balance_cpu
, CPU_IDLE
);
4430 rq
= cpu_rq(balance_cpu
);
4431 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4432 this_rq
->next_balance
= rq
->next_balance
;
4438 static inline int on_null_domain(int cpu
)
4440 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4444 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4446 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4447 * idle load balancing owner or decide to stop the periodic load balancing,
4448 * if the whole system is idle.
4450 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4454 * If we were in the nohz mode recently and busy at the current
4455 * scheduler tick, then check if we need to nominate new idle
4458 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4459 rq
->in_nohz_recently
= 0;
4461 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4462 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4463 atomic_set(&nohz
.load_balancer
, -1);
4466 if (atomic_read(&nohz
.load_balancer
) == -1) {
4468 * simple selection for now: Nominate the
4469 * first cpu in the nohz list to be the next
4472 * TBD: Traverse the sched domains and nominate
4473 * the nearest cpu in the nohz.cpu_mask.
4475 int ilb
= cpumask_first(nohz
.cpu_mask
);
4477 if (ilb
< nr_cpu_ids
)
4483 * If this cpu is idle and doing idle load balancing for all the
4484 * cpus with ticks stopped, is it time for that to stop?
4486 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4487 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4493 * If this cpu is idle and the idle load balancing is done by
4494 * someone else, then no need raise the SCHED_SOFTIRQ
4496 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4497 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4500 /* Don't need to rebalance while attached to NULL domain */
4501 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4502 likely(!on_null_domain(cpu
)))
4503 raise_softirq(SCHED_SOFTIRQ
);
4506 #else /* CONFIG_SMP */
4509 * on UP we do not need to balance between CPUs:
4511 static inline void idle_balance(int cpu
, struct rq
*rq
)
4517 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4519 EXPORT_PER_CPU_SYMBOL(kstat
);
4522 * Return any ns on the sched_clock that have not yet been accounted in
4523 * @p in case that task is currently running.
4525 * Called with task_rq_lock() held on @rq.
4527 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4531 if (task_current(rq
, p
)) {
4532 update_rq_clock(rq
);
4533 ns
= rq
->clock
- p
->se
.exec_start
;
4541 unsigned long long task_delta_exec(struct task_struct
*p
)
4543 unsigned long flags
;
4547 rq
= task_rq_lock(p
, &flags
);
4548 ns
= do_task_delta_exec(p
, rq
);
4549 task_rq_unlock(rq
, &flags
);
4555 * Return accounted runtime for the task.
4556 * In case the task is currently running, return the runtime plus current's
4557 * pending runtime that have not been accounted yet.
4559 unsigned long long task_sched_runtime(struct task_struct
*p
)
4561 unsigned long flags
;
4565 rq
= task_rq_lock(p
, &flags
);
4566 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4567 task_rq_unlock(rq
, &flags
);
4573 * Return sum_exec_runtime for the thread group.
4574 * In case the task is currently running, return the sum plus current's
4575 * pending runtime that have not been accounted yet.
4577 * Note that the thread group might have other running tasks as well,
4578 * so the return value not includes other pending runtime that other
4579 * running tasks might have.
4581 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4583 struct task_cputime totals
;
4584 unsigned long flags
;
4588 rq
= task_rq_lock(p
, &flags
);
4589 thread_group_cputime(p
, &totals
);
4590 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4591 task_rq_unlock(rq
, &flags
);
4597 * Account user cpu time to a process.
4598 * @p: the process that the cpu time gets accounted to
4599 * @cputime: the cpu time spent in user space since the last update
4600 * @cputime_scaled: cputime scaled by cpu frequency
4602 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4603 cputime_t cputime_scaled
)
4605 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4608 /* Add user time to process. */
4609 p
->utime
= cputime_add(p
->utime
, cputime
);
4610 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4611 account_group_user_time(p
, cputime
);
4613 /* Add user time to cpustat. */
4614 tmp
= cputime_to_cputime64(cputime
);
4615 if (TASK_NICE(p
) > 0)
4616 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4618 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4620 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4621 /* Account for user time used */
4622 acct_update_integrals(p
);
4626 * Account guest cpu time to a process.
4627 * @p: the process that the cpu time gets accounted to
4628 * @cputime: the cpu time spent in virtual machine since the last update
4629 * @cputime_scaled: cputime scaled by cpu frequency
4631 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4632 cputime_t cputime_scaled
)
4635 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4637 tmp
= cputime_to_cputime64(cputime
);
4639 /* Add guest time to process. */
4640 p
->utime
= cputime_add(p
->utime
, cputime
);
4641 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4642 account_group_user_time(p
, cputime
);
4643 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4645 /* Add guest time to cpustat. */
4646 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4647 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4651 * Account system cpu time to a process.
4652 * @p: the process that the cpu time gets accounted to
4653 * @hardirq_offset: the offset to subtract from hardirq_count()
4654 * @cputime: the cpu time spent in kernel space since the last update
4655 * @cputime_scaled: cputime scaled by cpu frequency
4657 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4658 cputime_t cputime
, cputime_t cputime_scaled
)
4660 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4663 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4664 account_guest_time(p
, cputime
, cputime_scaled
);
4668 /* Add system time to process. */
4669 p
->stime
= cputime_add(p
->stime
, cputime
);
4670 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4671 account_group_system_time(p
, cputime
);
4673 /* Add system time to cpustat. */
4674 tmp
= cputime_to_cputime64(cputime
);
4675 if (hardirq_count() - hardirq_offset
)
4676 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4677 else if (softirq_count())
4678 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4680 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4682 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4684 /* Account for system time used */
4685 acct_update_integrals(p
);
4689 * Account for involuntary wait time.
4690 * @steal: the cpu time spent in involuntary wait
4692 void account_steal_time(cputime_t cputime
)
4694 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4695 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4697 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4701 * Account for idle time.
4702 * @cputime: the cpu time spent in idle wait
4704 void account_idle_time(cputime_t cputime
)
4706 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4707 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4708 struct rq
*rq
= this_rq();
4710 if (atomic_read(&rq
->nr_iowait
) > 0)
4711 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4713 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4716 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4719 * Account a single tick of cpu time.
4720 * @p: the process that the cpu time gets accounted to
4721 * @user_tick: indicates if the tick is a user or a system tick
4723 void account_process_tick(struct task_struct
*p
, int user_tick
)
4725 cputime_t one_jiffy
= jiffies_to_cputime(1);
4726 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4727 struct rq
*rq
= this_rq();
4730 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4731 else if (p
!= rq
->idle
)
4732 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4735 account_idle_time(one_jiffy
);
4739 * Account multiple ticks of steal time.
4740 * @p: the process from which the cpu time has been stolen
4741 * @ticks: number of stolen ticks
4743 void account_steal_ticks(unsigned long ticks
)
4745 account_steal_time(jiffies_to_cputime(ticks
));
4749 * Account multiple ticks of idle time.
4750 * @ticks: number of stolen ticks
4752 void account_idle_ticks(unsigned long ticks
)
4754 account_idle_time(jiffies_to_cputime(ticks
));
4760 * Use precise platform statistics if available:
4762 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4763 cputime_t
task_utime(struct task_struct
*p
)
4768 cputime_t
task_stime(struct task_struct
*p
)
4773 cputime_t
task_utime(struct task_struct
*p
)
4775 clock_t utime
= cputime_to_clock_t(p
->utime
),
4776 total
= utime
+ cputime_to_clock_t(p
->stime
);
4780 * Use CFS's precise accounting:
4782 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4786 do_div(temp
, total
);
4788 utime
= (clock_t)temp
;
4790 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4791 return p
->prev_utime
;
4794 cputime_t
task_stime(struct task_struct
*p
)
4799 * Use CFS's precise accounting. (we subtract utime from
4800 * the total, to make sure the total observed by userspace
4801 * grows monotonically - apps rely on that):
4803 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4804 cputime_to_clock_t(task_utime(p
));
4807 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4809 return p
->prev_stime
;
4813 inline cputime_t
task_gtime(struct task_struct
*p
)
4819 * This function gets called by the timer code, with HZ frequency.
4820 * We call it with interrupts disabled.
4822 * It also gets called by the fork code, when changing the parent's
4825 void scheduler_tick(void)
4827 int cpu
= smp_processor_id();
4828 struct rq
*rq
= cpu_rq(cpu
);
4829 struct task_struct
*curr
= rq
->curr
;
4833 spin_lock(&rq
->lock
);
4834 update_rq_clock(rq
);
4835 update_cpu_load(rq
);
4836 curr
->sched_class
->task_tick(rq
, curr
, 0);
4837 spin_unlock(&rq
->lock
);
4840 rq
->idle_at_tick
= idle_cpu(cpu
);
4841 trigger_load_balance(rq
, cpu
);
4845 unsigned long get_parent_ip(unsigned long addr
)
4847 if (in_lock_functions(addr
)) {
4848 addr
= CALLER_ADDR2
;
4849 if (in_lock_functions(addr
))
4850 addr
= CALLER_ADDR3
;
4855 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4856 defined(CONFIG_PREEMPT_TRACER))
4858 void __kprobes
add_preempt_count(int val
)
4860 #ifdef CONFIG_DEBUG_PREEMPT
4864 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4867 preempt_count() += val
;
4868 #ifdef CONFIG_DEBUG_PREEMPT
4870 * Spinlock count overflowing soon?
4872 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4875 if (preempt_count() == val
)
4876 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4878 EXPORT_SYMBOL(add_preempt_count
);
4880 void __kprobes
sub_preempt_count(int val
)
4882 #ifdef CONFIG_DEBUG_PREEMPT
4886 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4889 * Is the spinlock portion underflowing?
4891 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4892 !(preempt_count() & PREEMPT_MASK
)))
4896 if (preempt_count() == val
)
4897 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4898 preempt_count() -= val
;
4900 EXPORT_SYMBOL(sub_preempt_count
);
4905 * Print scheduling while atomic bug:
4907 static noinline
void __schedule_bug(struct task_struct
*prev
)
4909 struct pt_regs
*regs
= get_irq_regs();
4911 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4912 prev
->comm
, prev
->pid
, preempt_count());
4914 debug_show_held_locks(prev
);
4916 if (irqs_disabled())
4917 print_irqtrace_events(prev
);
4926 * Various schedule()-time debugging checks and statistics:
4928 static inline void schedule_debug(struct task_struct
*prev
)
4931 * Test if we are atomic. Since do_exit() needs to call into
4932 * schedule() atomically, we ignore that path for now.
4933 * Otherwise, whine if we are scheduling when we should not be.
4935 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4936 __schedule_bug(prev
);
4938 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4940 schedstat_inc(this_rq(), sched_count
);
4941 #ifdef CONFIG_SCHEDSTATS
4942 if (unlikely(prev
->lock_depth
>= 0)) {
4943 schedstat_inc(this_rq(), bkl_count
);
4944 schedstat_inc(prev
, sched_info
.bkl_count
);
4949 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4951 if (prev
->state
== TASK_RUNNING
) {
4952 u64 runtime
= prev
->se
.sum_exec_runtime
;
4954 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4955 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
4958 * In order to avoid avg_overlap growing stale when we are
4959 * indeed overlapping and hence not getting put to sleep, grow
4960 * the avg_overlap on preemption.
4962 * We use the average preemption runtime because that
4963 * correlates to the amount of cache footprint a task can
4966 update_avg(&prev
->se
.avg_overlap
, runtime
);
4968 prev
->sched_class
->put_prev_task(rq
, prev
);
4972 * Pick up the highest-prio task:
4974 static inline struct task_struct
*
4975 pick_next_task(struct rq
*rq
)
4977 const struct sched_class
*class;
4978 struct task_struct
*p
;
4981 * Optimization: we know that if all tasks are in
4982 * the fair class we can call that function directly:
4984 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4985 p
= fair_sched_class
.pick_next_task(rq
);
4990 class = sched_class_highest
;
4992 p
= class->pick_next_task(rq
);
4996 * Will never be NULL as the idle class always
4997 * returns a non-NULL p:
4999 class = class->next
;
5004 * schedule() is the main scheduler function.
5006 asmlinkage
void __sched
__schedule(void)
5008 struct task_struct
*prev
, *next
;
5009 unsigned long *switch_count
;
5013 cpu
= smp_processor_id();
5017 switch_count
= &prev
->nivcsw
;
5019 release_kernel_lock(prev
);
5020 need_resched_nonpreemptible
:
5022 schedule_debug(prev
);
5024 if (sched_feat(HRTICK
))
5027 spin_lock_irq(&rq
->lock
);
5028 update_rq_clock(rq
);
5029 clear_tsk_need_resched(prev
);
5031 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5032 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5033 prev
->state
= TASK_RUNNING
;
5035 deactivate_task(rq
, prev
, 1);
5036 switch_count
= &prev
->nvcsw
;
5040 if (prev
->sched_class
->pre_schedule
)
5041 prev
->sched_class
->pre_schedule(rq
, prev
);
5044 if (unlikely(!rq
->nr_running
))
5045 idle_balance(cpu
, rq
);
5047 put_prev_task(rq
, prev
);
5048 next
= pick_next_task(rq
);
5050 if (likely(prev
!= next
)) {
5051 sched_info_switch(prev
, next
);
5057 context_switch(rq
, prev
, next
); /* unlocks the rq */
5059 * the context switch might have flipped the stack from under
5060 * us, hence refresh the local variables.
5062 cpu
= smp_processor_id();
5065 spin_unlock_irq(&rq
->lock
);
5067 if (unlikely(reacquire_kernel_lock(current
) < 0))
5068 goto need_resched_nonpreemptible
;
5071 asmlinkage
void __sched
schedule(void)
5076 preempt_enable_no_resched();
5077 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5080 EXPORT_SYMBOL(schedule
);
5084 * Look out! "owner" is an entirely speculative pointer
5085 * access and not reliable.
5087 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5092 if (!sched_feat(OWNER_SPIN
))
5095 #ifdef CONFIG_DEBUG_PAGEALLOC
5097 * Need to access the cpu field knowing that
5098 * DEBUG_PAGEALLOC could have unmapped it if
5099 * the mutex owner just released it and exited.
5101 if (probe_kernel_address(&owner
->cpu
, cpu
))
5108 * Even if the access succeeded (likely case),
5109 * the cpu field may no longer be valid.
5111 if (cpu
>= nr_cpumask_bits
)
5115 * We need to validate that we can do a
5116 * get_cpu() and that we have the percpu area.
5118 if (!cpu_online(cpu
))
5125 * Owner changed, break to re-assess state.
5127 if (lock
->owner
!= owner
)
5131 * Is that owner really running on that cpu?
5133 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5143 #ifdef CONFIG_PREEMPT
5145 * this is the entry point to schedule() from in-kernel preemption
5146 * off of preempt_enable. Kernel preemptions off return from interrupt
5147 * occur there and call schedule directly.
5149 asmlinkage
void __sched
preempt_schedule(void)
5151 struct thread_info
*ti
= current_thread_info();
5154 * If there is a non-zero preempt_count or interrupts are disabled,
5155 * we do not want to preempt the current task. Just return..
5157 if (likely(ti
->preempt_count
|| irqs_disabled()))
5161 add_preempt_count(PREEMPT_ACTIVE
);
5163 sub_preempt_count(PREEMPT_ACTIVE
);
5166 * Check again in case we missed a preemption opportunity
5167 * between schedule and now.
5170 } while (need_resched());
5172 EXPORT_SYMBOL(preempt_schedule
);
5175 * this is the entry point to schedule() from kernel preemption
5176 * off of irq context.
5177 * Note, that this is called and return with irqs disabled. This will
5178 * protect us against recursive calling from irq.
5180 asmlinkage
void __sched
preempt_schedule_irq(void)
5182 struct thread_info
*ti
= current_thread_info();
5184 /* Catch callers which need to be fixed */
5185 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5188 add_preempt_count(PREEMPT_ACTIVE
);
5191 local_irq_disable();
5192 sub_preempt_count(PREEMPT_ACTIVE
);
5195 * Check again in case we missed a preemption opportunity
5196 * between schedule and now.
5199 } while (need_resched());
5202 #endif /* CONFIG_PREEMPT */
5204 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5207 return try_to_wake_up(curr
->private, mode
, sync
);
5209 EXPORT_SYMBOL(default_wake_function
);
5212 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5213 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5214 * number) then we wake all the non-exclusive tasks and one exclusive task.
5216 * There are circumstances in which we can try to wake a task which has already
5217 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5218 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5220 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5221 int nr_exclusive
, int sync
, void *key
)
5223 wait_queue_t
*curr
, *next
;
5225 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5226 unsigned flags
= curr
->flags
;
5228 if (curr
->func(curr
, mode
, sync
, key
) &&
5229 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5235 * __wake_up - wake up threads blocked on a waitqueue.
5237 * @mode: which threads
5238 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5239 * @key: is directly passed to the wakeup function
5241 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5242 int nr_exclusive
, void *key
)
5244 unsigned long flags
;
5246 spin_lock_irqsave(&q
->lock
, flags
);
5247 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5248 spin_unlock_irqrestore(&q
->lock
, flags
);
5250 EXPORT_SYMBOL(__wake_up
);
5253 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5255 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5257 __wake_up_common(q
, mode
, 1, 0, NULL
);
5260 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5262 __wake_up_common(q
, mode
, 1, 0, key
);
5266 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5268 * @mode: which threads
5269 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5270 * @key: opaque value to be passed to wakeup targets
5272 * The sync wakeup differs that the waker knows that it will schedule
5273 * away soon, so while the target thread will be woken up, it will not
5274 * be migrated to another CPU - ie. the two threads are 'synchronized'
5275 * with each other. This can prevent needless bouncing between CPUs.
5277 * On UP it can prevent extra preemption.
5279 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5280 int nr_exclusive
, void *key
)
5282 unsigned long flags
;
5288 if (unlikely(!nr_exclusive
))
5291 spin_lock_irqsave(&q
->lock
, flags
);
5292 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5293 spin_unlock_irqrestore(&q
->lock
, flags
);
5295 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5298 * __wake_up_sync - see __wake_up_sync_key()
5300 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5302 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5304 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5307 * complete: - signals a single thread waiting on this completion
5308 * @x: holds the state of this particular completion
5310 * This will wake up a single thread waiting on this completion. Threads will be
5311 * awakened in the same order in which they were queued.
5313 * See also complete_all(), wait_for_completion() and related routines.
5315 void complete(struct completion
*x
)
5317 unsigned long flags
;
5319 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5321 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5322 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5324 EXPORT_SYMBOL(complete
);
5327 * complete_all: - signals all threads waiting on this completion
5328 * @x: holds the state of this particular completion
5330 * This will wake up all threads waiting on this particular completion event.
5332 void complete_all(struct completion
*x
)
5334 unsigned long flags
;
5336 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5337 x
->done
+= UINT_MAX
/2;
5338 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5339 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5341 EXPORT_SYMBOL(complete_all
);
5343 static inline long __sched
5344 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5347 DECLARE_WAITQUEUE(wait
, current
);
5349 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5350 __add_wait_queue_tail(&x
->wait
, &wait
);
5352 if (signal_pending_state(state
, current
)) {
5353 timeout
= -ERESTARTSYS
;
5356 __set_current_state(state
);
5357 spin_unlock_irq(&x
->wait
.lock
);
5358 timeout
= schedule_timeout(timeout
);
5359 spin_lock_irq(&x
->wait
.lock
);
5360 } while (!x
->done
&& timeout
);
5361 __remove_wait_queue(&x
->wait
, &wait
);
5366 return timeout
?: 1;
5370 wait_for_common(struct completion
*x
, long timeout
, int state
)
5374 spin_lock_irq(&x
->wait
.lock
);
5375 timeout
= do_wait_for_common(x
, timeout
, state
);
5376 spin_unlock_irq(&x
->wait
.lock
);
5381 * wait_for_completion: - waits for completion of a task
5382 * @x: holds the state of this particular completion
5384 * This waits to be signaled for completion of a specific task. It is NOT
5385 * interruptible and there is no timeout.
5387 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5388 * and interrupt capability. Also see complete().
5390 void __sched
wait_for_completion(struct completion
*x
)
5392 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5394 EXPORT_SYMBOL(wait_for_completion
);
5397 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5398 * @x: holds the state of this particular completion
5399 * @timeout: timeout value in jiffies
5401 * This waits for either a completion of a specific task to be signaled or for a
5402 * specified timeout to expire. The timeout is in jiffies. It is not
5405 unsigned long __sched
5406 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5408 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5410 EXPORT_SYMBOL(wait_for_completion_timeout
);
5413 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5414 * @x: holds the state of this particular completion
5416 * This waits for completion of a specific task to be signaled. It is
5419 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5421 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5422 if (t
== -ERESTARTSYS
)
5426 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5429 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5430 * @x: holds the state of this particular completion
5431 * @timeout: timeout value in jiffies
5433 * This waits for either a completion of a specific task to be signaled or for a
5434 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5436 unsigned long __sched
5437 wait_for_completion_interruptible_timeout(struct completion
*x
,
5438 unsigned long timeout
)
5440 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5442 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5445 * wait_for_completion_killable: - waits for completion of a task (killable)
5446 * @x: holds the state of this particular completion
5448 * This waits to be signaled for completion of a specific task. It can be
5449 * interrupted by a kill signal.
5451 int __sched
wait_for_completion_killable(struct completion
*x
)
5453 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5454 if (t
== -ERESTARTSYS
)
5458 EXPORT_SYMBOL(wait_for_completion_killable
);
5461 * try_wait_for_completion - try to decrement a completion without blocking
5462 * @x: completion structure
5464 * Returns: 0 if a decrement cannot be done without blocking
5465 * 1 if a decrement succeeded.
5467 * If a completion is being used as a counting completion,
5468 * attempt to decrement the counter without blocking. This
5469 * enables us to avoid waiting if the resource the completion
5470 * is protecting is not available.
5472 bool try_wait_for_completion(struct completion
*x
)
5476 spin_lock_irq(&x
->wait
.lock
);
5481 spin_unlock_irq(&x
->wait
.lock
);
5484 EXPORT_SYMBOL(try_wait_for_completion
);
5487 * completion_done - Test to see if a completion has any waiters
5488 * @x: completion structure
5490 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5491 * 1 if there are no waiters.
5494 bool completion_done(struct completion
*x
)
5498 spin_lock_irq(&x
->wait
.lock
);
5501 spin_unlock_irq(&x
->wait
.lock
);
5504 EXPORT_SYMBOL(completion_done
);
5507 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5509 unsigned long flags
;
5512 init_waitqueue_entry(&wait
, current
);
5514 __set_current_state(state
);
5516 spin_lock_irqsave(&q
->lock
, flags
);
5517 __add_wait_queue(q
, &wait
);
5518 spin_unlock(&q
->lock
);
5519 timeout
= schedule_timeout(timeout
);
5520 spin_lock_irq(&q
->lock
);
5521 __remove_wait_queue(q
, &wait
);
5522 spin_unlock_irqrestore(&q
->lock
, flags
);
5527 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5529 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5531 EXPORT_SYMBOL(interruptible_sleep_on
);
5534 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5536 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5538 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5540 void __sched
sleep_on(wait_queue_head_t
*q
)
5542 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5544 EXPORT_SYMBOL(sleep_on
);
5546 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5548 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5550 EXPORT_SYMBOL(sleep_on_timeout
);
5552 #ifdef CONFIG_RT_MUTEXES
5555 * rt_mutex_setprio - set the current priority of a task
5557 * @prio: prio value (kernel-internal form)
5559 * This function changes the 'effective' priority of a task. It does
5560 * not touch ->normal_prio like __setscheduler().
5562 * Used by the rt_mutex code to implement priority inheritance logic.
5564 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5566 unsigned long flags
;
5567 int oldprio
, on_rq
, running
;
5569 const struct sched_class
*prev_class
= p
->sched_class
;
5571 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5573 rq
= task_rq_lock(p
, &flags
);
5574 update_rq_clock(rq
);
5577 on_rq
= p
->se
.on_rq
;
5578 running
= task_current(rq
, p
);
5580 dequeue_task(rq
, p
, 0);
5582 p
->sched_class
->put_prev_task(rq
, p
);
5585 p
->sched_class
= &rt_sched_class
;
5587 p
->sched_class
= &fair_sched_class
;
5592 p
->sched_class
->set_curr_task(rq
);
5594 enqueue_task(rq
, p
, 0);
5596 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5598 task_rq_unlock(rq
, &flags
);
5603 void set_user_nice(struct task_struct
*p
, long nice
)
5605 int old_prio
, delta
, on_rq
;
5606 unsigned long flags
;
5609 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5612 * We have to be careful, if called from sys_setpriority(),
5613 * the task might be in the middle of scheduling on another CPU.
5615 rq
= task_rq_lock(p
, &flags
);
5616 update_rq_clock(rq
);
5618 * The RT priorities are set via sched_setscheduler(), but we still
5619 * allow the 'normal' nice value to be set - but as expected
5620 * it wont have any effect on scheduling until the task is
5621 * SCHED_FIFO/SCHED_RR:
5623 if (task_has_rt_policy(p
)) {
5624 p
->static_prio
= NICE_TO_PRIO(nice
);
5627 on_rq
= p
->se
.on_rq
;
5629 dequeue_task(rq
, p
, 0);
5631 p
->static_prio
= NICE_TO_PRIO(nice
);
5634 p
->prio
= effective_prio(p
);
5635 delta
= p
->prio
- old_prio
;
5638 enqueue_task(rq
, p
, 0);
5640 * If the task increased its priority or is running and
5641 * lowered its priority, then reschedule its CPU:
5643 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5644 resched_task(rq
->curr
);
5647 task_rq_unlock(rq
, &flags
);
5649 EXPORT_SYMBOL(set_user_nice
);
5652 * can_nice - check if a task can reduce its nice value
5656 int can_nice(const struct task_struct
*p
, const int nice
)
5658 /* convert nice value [19,-20] to rlimit style value [1,40] */
5659 int nice_rlim
= 20 - nice
;
5661 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5662 capable(CAP_SYS_NICE
));
5665 #ifdef __ARCH_WANT_SYS_NICE
5668 * sys_nice - change the priority of the current process.
5669 * @increment: priority increment
5671 * sys_setpriority is a more generic, but much slower function that
5672 * does similar things.
5674 SYSCALL_DEFINE1(nice
, int, increment
)
5679 * Setpriority might change our priority at the same moment.
5680 * We don't have to worry. Conceptually one call occurs first
5681 * and we have a single winner.
5683 if (increment
< -40)
5688 nice
= TASK_NICE(current
) + increment
;
5694 if (increment
< 0 && !can_nice(current
, nice
))
5697 retval
= security_task_setnice(current
, nice
);
5701 set_user_nice(current
, nice
);
5708 * task_prio - return the priority value of a given task.
5709 * @p: the task in question.
5711 * This is the priority value as seen by users in /proc.
5712 * RT tasks are offset by -200. Normal tasks are centered
5713 * around 0, value goes from -16 to +15.
5715 int task_prio(const struct task_struct
*p
)
5717 return p
->prio
- MAX_RT_PRIO
;
5721 * task_nice - return the nice value of a given task.
5722 * @p: the task in question.
5724 int task_nice(const struct task_struct
*p
)
5726 return TASK_NICE(p
);
5728 EXPORT_SYMBOL(task_nice
);
5731 * idle_cpu - is a given cpu idle currently?
5732 * @cpu: the processor in question.
5734 int idle_cpu(int cpu
)
5736 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5740 * idle_task - return the idle task for a given cpu.
5741 * @cpu: the processor in question.
5743 struct task_struct
*idle_task(int cpu
)
5745 return cpu_rq(cpu
)->idle
;
5749 * find_process_by_pid - find a process with a matching PID value.
5750 * @pid: the pid in question.
5752 static struct task_struct
*find_process_by_pid(pid_t pid
)
5754 return pid
? find_task_by_vpid(pid
) : current
;
5757 /* Actually do priority change: must hold rq lock. */
5759 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5761 BUG_ON(p
->se
.on_rq
);
5764 switch (p
->policy
) {
5768 p
->sched_class
= &fair_sched_class
;
5772 p
->sched_class
= &rt_sched_class
;
5776 p
->rt_priority
= prio
;
5777 p
->normal_prio
= normal_prio(p
);
5778 /* we are holding p->pi_lock already */
5779 p
->prio
= rt_mutex_getprio(p
);
5784 * check the target process has a UID that matches the current process's
5786 static bool check_same_owner(struct task_struct
*p
)
5788 const struct cred
*cred
= current_cred(), *pcred
;
5792 pcred
= __task_cred(p
);
5793 match
= (cred
->euid
== pcred
->euid
||
5794 cred
->euid
== pcred
->uid
);
5799 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5800 struct sched_param
*param
, bool user
)
5802 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5803 unsigned long flags
;
5804 const struct sched_class
*prev_class
= p
->sched_class
;
5807 /* may grab non-irq protected spin_locks */
5808 BUG_ON(in_interrupt());
5810 /* double check policy once rq lock held */
5812 policy
= oldpolicy
= p
->policy
;
5813 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5814 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5815 policy
!= SCHED_IDLE
)
5818 * Valid priorities for SCHED_FIFO and SCHED_RR are
5819 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5820 * SCHED_BATCH and SCHED_IDLE is 0.
5822 if (param
->sched_priority
< 0 ||
5823 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5824 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5826 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5830 * Allow unprivileged RT tasks to decrease priority:
5832 if (user
&& !capable(CAP_SYS_NICE
)) {
5833 if (rt_policy(policy
)) {
5834 unsigned long rlim_rtprio
;
5836 if (!lock_task_sighand(p
, &flags
))
5838 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5839 unlock_task_sighand(p
, &flags
);
5841 /* can't set/change the rt policy */
5842 if (policy
!= p
->policy
&& !rlim_rtprio
)
5845 /* can't increase priority */
5846 if (param
->sched_priority
> p
->rt_priority
&&
5847 param
->sched_priority
> rlim_rtprio
)
5851 * Like positive nice levels, dont allow tasks to
5852 * move out of SCHED_IDLE either:
5854 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5857 /* can't change other user's priorities */
5858 if (!check_same_owner(p
))
5863 #ifdef CONFIG_RT_GROUP_SCHED
5865 * Do not allow realtime tasks into groups that have no runtime
5868 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5869 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5873 retval
= security_task_setscheduler(p
, policy
, param
);
5879 * make sure no PI-waiters arrive (or leave) while we are
5880 * changing the priority of the task:
5882 spin_lock_irqsave(&p
->pi_lock
, flags
);
5884 * To be able to change p->policy safely, the apropriate
5885 * runqueue lock must be held.
5887 rq
= __task_rq_lock(p
);
5888 /* recheck policy now with rq lock held */
5889 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5890 policy
= oldpolicy
= -1;
5891 __task_rq_unlock(rq
);
5892 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5895 update_rq_clock(rq
);
5896 on_rq
= p
->se
.on_rq
;
5897 running
= task_current(rq
, p
);
5899 deactivate_task(rq
, p
, 0);
5901 p
->sched_class
->put_prev_task(rq
, p
);
5904 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5907 p
->sched_class
->set_curr_task(rq
);
5909 activate_task(rq
, p
, 0);
5911 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5913 __task_rq_unlock(rq
);
5914 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5916 rt_mutex_adjust_pi(p
);
5922 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5923 * @p: the task in question.
5924 * @policy: new policy.
5925 * @param: structure containing the new RT priority.
5927 * NOTE that the task may be already dead.
5929 int sched_setscheduler(struct task_struct
*p
, int policy
,
5930 struct sched_param
*param
)
5932 return __sched_setscheduler(p
, policy
, param
, true);
5934 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5937 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5938 * @p: the task in question.
5939 * @policy: new policy.
5940 * @param: structure containing the new RT priority.
5942 * Just like sched_setscheduler, only don't bother checking if the
5943 * current context has permission. For example, this is needed in
5944 * stop_machine(): we create temporary high priority worker threads,
5945 * but our caller might not have that capability.
5947 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5948 struct sched_param
*param
)
5950 return __sched_setscheduler(p
, policy
, param
, false);
5954 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5956 struct sched_param lparam
;
5957 struct task_struct
*p
;
5960 if (!param
|| pid
< 0)
5962 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5967 p
= find_process_by_pid(pid
);
5969 retval
= sched_setscheduler(p
, policy
, &lparam
);
5976 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5977 * @pid: the pid in question.
5978 * @policy: new policy.
5979 * @param: structure containing the new RT priority.
5981 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5982 struct sched_param __user
*, param
)
5984 /* negative values for policy are not valid */
5988 return do_sched_setscheduler(pid
, policy
, param
);
5992 * sys_sched_setparam - set/change the RT priority of a thread
5993 * @pid: the pid in question.
5994 * @param: structure containing the new RT priority.
5996 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5998 return do_sched_setscheduler(pid
, -1, param
);
6002 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6003 * @pid: the pid in question.
6005 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6007 struct task_struct
*p
;
6014 read_lock(&tasklist_lock
);
6015 p
= find_process_by_pid(pid
);
6017 retval
= security_task_getscheduler(p
);
6021 read_unlock(&tasklist_lock
);
6026 * sys_sched_getscheduler - get the RT priority of a thread
6027 * @pid: the pid in question.
6028 * @param: structure containing the RT priority.
6030 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6032 struct sched_param lp
;
6033 struct task_struct
*p
;
6036 if (!param
|| pid
< 0)
6039 read_lock(&tasklist_lock
);
6040 p
= find_process_by_pid(pid
);
6045 retval
= security_task_getscheduler(p
);
6049 lp
.sched_priority
= p
->rt_priority
;
6050 read_unlock(&tasklist_lock
);
6053 * This one might sleep, we cannot do it with a spinlock held ...
6055 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6060 read_unlock(&tasklist_lock
);
6064 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6066 cpumask_var_t cpus_allowed
, new_mask
;
6067 struct task_struct
*p
;
6071 read_lock(&tasklist_lock
);
6073 p
= find_process_by_pid(pid
);
6075 read_unlock(&tasklist_lock
);
6081 * It is not safe to call set_cpus_allowed with the
6082 * tasklist_lock held. We will bump the task_struct's
6083 * usage count and then drop tasklist_lock.
6086 read_unlock(&tasklist_lock
);
6088 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6092 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6094 goto out_free_cpus_allowed
;
6097 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6100 retval
= security_task_setscheduler(p
, 0, NULL
);
6104 cpuset_cpus_allowed(p
, cpus_allowed
);
6105 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6107 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6110 cpuset_cpus_allowed(p
, cpus_allowed
);
6111 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6113 * We must have raced with a concurrent cpuset
6114 * update. Just reset the cpus_allowed to the
6115 * cpuset's cpus_allowed
6117 cpumask_copy(new_mask
, cpus_allowed
);
6122 free_cpumask_var(new_mask
);
6123 out_free_cpus_allowed
:
6124 free_cpumask_var(cpus_allowed
);
6131 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6132 struct cpumask
*new_mask
)
6134 if (len
< cpumask_size())
6135 cpumask_clear(new_mask
);
6136 else if (len
> cpumask_size())
6137 len
= cpumask_size();
6139 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6143 * sys_sched_setaffinity - set the cpu affinity of a process
6144 * @pid: pid of the process
6145 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6146 * @user_mask_ptr: user-space pointer to the new cpu mask
6148 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6149 unsigned long __user
*, user_mask_ptr
)
6151 cpumask_var_t new_mask
;
6154 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6157 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6159 retval
= sched_setaffinity(pid
, new_mask
);
6160 free_cpumask_var(new_mask
);
6164 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6166 struct task_struct
*p
;
6170 read_lock(&tasklist_lock
);
6173 p
= find_process_by_pid(pid
);
6177 retval
= security_task_getscheduler(p
);
6181 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6184 read_unlock(&tasklist_lock
);
6191 * sys_sched_getaffinity - get the cpu affinity of a process
6192 * @pid: pid of the process
6193 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6194 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6196 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6197 unsigned long __user
*, user_mask_ptr
)
6202 if (len
< cpumask_size())
6205 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6208 ret
= sched_getaffinity(pid
, mask
);
6210 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6213 ret
= cpumask_size();
6215 free_cpumask_var(mask
);
6221 * sys_sched_yield - yield the current processor to other threads.
6223 * This function yields the current CPU to other tasks. If there are no
6224 * other threads running on this CPU then this function will return.
6226 SYSCALL_DEFINE0(sched_yield
)
6228 struct rq
*rq
= this_rq_lock();
6230 schedstat_inc(rq
, yld_count
);
6231 current
->sched_class
->yield_task(rq
);
6234 * Since we are going to call schedule() anyway, there's
6235 * no need to preempt or enable interrupts:
6237 __release(rq
->lock
);
6238 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6239 _raw_spin_unlock(&rq
->lock
);
6240 preempt_enable_no_resched();
6247 static void __cond_resched(void)
6249 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6250 __might_sleep(__FILE__
, __LINE__
);
6253 * The BKS might be reacquired before we have dropped
6254 * PREEMPT_ACTIVE, which could trigger a second
6255 * cond_resched() call.
6258 add_preempt_count(PREEMPT_ACTIVE
);
6260 sub_preempt_count(PREEMPT_ACTIVE
);
6261 } while (need_resched());
6264 int __sched
_cond_resched(void)
6266 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6267 system_state
== SYSTEM_RUNNING
) {
6273 EXPORT_SYMBOL(_cond_resched
);
6276 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6277 * call schedule, and on return reacquire the lock.
6279 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6280 * operations here to prevent schedule() from being called twice (once via
6281 * spin_unlock(), once by hand).
6283 int cond_resched_lock(spinlock_t
*lock
)
6285 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6288 if (spin_needbreak(lock
) || resched
) {
6290 if (resched
&& need_resched())
6299 EXPORT_SYMBOL(cond_resched_lock
);
6301 int __sched
cond_resched_softirq(void)
6303 BUG_ON(!in_softirq());
6305 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6313 EXPORT_SYMBOL(cond_resched_softirq
);
6316 * yield - yield the current processor to other threads.
6318 * This is a shortcut for kernel-space yielding - it marks the
6319 * thread runnable and calls sys_sched_yield().
6321 void __sched
yield(void)
6323 set_current_state(TASK_RUNNING
);
6326 EXPORT_SYMBOL(yield
);
6329 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6330 * that process accounting knows that this is a task in IO wait state.
6332 * But don't do that if it is a deliberate, throttling IO wait (this task
6333 * has set its backing_dev_info: the queue against which it should throttle)
6335 void __sched
io_schedule(void)
6337 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6339 delayacct_blkio_start();
6340 atomic_inc(&rq
->nr_iowait
);
6342 atomic_dec(&rq
->nr_iowait
);
6343 delayacct_blkio_end();
6345 EXPORT_SYMBOL(io_schedule
);
6347 long __sched
io_schedule_timeout(long timeout
)
6349 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6352 delayacct_blkio_start();
6353 atomic_inc(&rq
->nr_iowait
);
6354 ret
= schedule_timeout(timeout
);
6355 atomic_dec(&rq
->nr_iowait
);
6356 delayacct_blkio_end();
6361 * sys_sched_get_priority_max - return maximum RT priority.
6362 * @policy: scheduling class.
6364 * this syscall returns the maximum rt_priority that can be used
6365 * by a given scheduling class.
6367 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6374 ret
= MAX_USER_RT_PRIO
-1;
6386 * sys_sched_get_priority_min - return minimum RT priority.
6387 * @policy: scheduling class.
6389 * this syscall returns the minimum rt_priority that can be used
6390 * by a given scheduling class.
6392 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6410 * sys_sched_rr_get_interval - return the default timeslice of a process.
6411 * @pid: pid of the process.
6412 * @interval: userspace pointer to the timeslice value.
6414 * this syscall writes the default timeslice value of a given process
6415 * into the user-space timespec buffer. A value of '0' means infinity.
6417 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6418 struct timespec __user
*, interval
)
6420 struct task_struct
*p
;
6421 unsigned int time_slice
;
6429 read_lock(&tasklist_lock
);
6430 p
= find_process_by_pid(pid
);
6434 retval
= security_task_getscheduler(p
);
6439 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6440 * tasks that are on an otherwise idle runqueue:
6443 if (p
->policy
== SCHED_RR
) {
6444 time_slice
= DEF_TIMESLICE
;
6445 } else if (p
->policy
!= SCHED_FIFO
) {
6446 struct sched_entity
*se
= &p
->se
;
6447 unsigned long flags
;
6450 rq
= task_rq_lock(p
, &flags
);
6451 if (rq
->cfs
.load
.weight
)
6452 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6453 task_rq_unlock(rq
, &flags
);
6455 read_unlock(&tasklist_lock
);
6456 jiffies_to_timespec(time_slice
, &t
);
6457 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6461 read_unlock(&tasklist_lock
);
6465 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6467 void sched_show_task(struct task_struct
*p
)
6469 unsigned long free
= 0;
6472 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6473 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6474 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6475 #if BITS_PER_LONG == 32
6476 if (state
== TASK_RUNNING
)
6477 printk(KERN_CONT
" running ");
6479 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6481 if (state
== TASK_RUNNING
)
6482 printk(KERN_CONT
" running task ");
6484 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6486 #ifdef CONFIG_DEBUG_STACK_USAGE
6487 free
= stack_not_used(p
);
6489 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6490 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6492 show_stack(p
, NULL
);
6495 void show_state_filter(unsigned long state_filter
)
6497 struct task_struct
*g
, *p
;
6499 #if BITS_PER_LONG == 32
6501 " task PC stack pid father\n");
6504 " task PC stack pid father\n");
6506 read_lock(&tasklist_lock
);
6507 do_each_thread(g
, p
) {
6509 * reset the NMI-timeout, listing all files on a slow
6510 * console might take alot of time:
6512 touch_nmi_watchdog();
6513 if (!state_filter
|| (p
->state
& state_filter
))
6515 } while_each_thread(g
, p
);
6517 touch_all_softlockup_watchdogs();
6519 #ifdef CONFIG_SCHED_DEBUG
6520 sysrq_sched_debug_show();
6522 read_unlock(&tasklist_lock
);
6524 * Only show locks if all tasks are dumped:
6526 if (state_filter
== -1)
6527 debug_show_all_locks();
6530 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6532 idle
->sched_class
= &idle_sched_class
;
6536 * init_idle - set up an idle thread for a given CPU
6537 * @idle: task in question
6538 * @cpu: cpu the idle task belongs to
6540 * NOTE: this function does not set the idle thread's NEED_RESCHED
6541 * flag, to make booting more robust.
6543 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6545 struct rq
*rq
= cpu_rq(cpu
);
6546 unsigned long flags
;
6548 spin_lock_irqsave(&rq
->lock
, flags
);
6551 idle
->se
.exec_start
= sched_clock();
6553 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6554 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6555 __set_task_cpu(idle
, cpu
);
6557 rq
->curr
= rq
->idle
= idle
;
6558 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6561 spin_unlock_irqrestore(&rq
->lock
, flags
);
6563 /* Set the preempt count _outside_ the spinlocks! */
6564 #if defined(CONFIG_PREEMPT)
6565 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6567 task_thread_info(idle
)->preempt_count
= 0;
6570 * The idle tasks have their own, simple scheduling class:
6572 idle
->sched_class
= &idle_sched_class
;
6573 ftrace_graph_init_task(idle
);
6577 * In a system that switches off the HZ timer nohz_cpu_mask
6578 * indicates which cpus entered this state. This is used
6579 * in the rcu update to wait only for active cpus. For system
6580 * which do not switch off the HZ timer nohz_cpu_mask should
6581 * always be CPU_BITS_NONE.
6583 cpumask_var_t nohz_cpu_mask
;
6586 * Increase the granularity value when there are more CPUs,
6587 * because with more CPUs the 'effective latency' as visible
6588 * to users decreases. But the relationship is not linear,
6589 * so pick a second-best guess by going with the log2 of the
6592 * This idea comes from the SD scheduler of Con Kolivas:
6594 static inline void sched_init_granularity(void)
6596 unsigned int factor
= 1 + ilog2(num_online_cpus());
6597 const unsigned long limit
= 200000000;
6599 sysctl_sched_min_granularity
*= factor
;
6600 if (sysctl_sched_min_granularity
> limit
)
6601 sysctl_sched_min_granularity
= limit
;
6603 sysctl_sched_latency
*= factor
;
6604 if (sysctl_sched_latency
> limit
)
6605 sysctl_sched_latency
= limit
;
6607 sysctl_sched_wakeup_granularity
*= factor
;
6609 sysctl_sched_shares_ratelimit
*= factor
;
6614 * This is how migration works:
6616 * 1) we queue a struct migration_req structure in the source CPU's
6617 * runqueue and wake up that CPU's migration thread.
6618 * 2) we down() the locked semaphore => thread blocks.
6619 * 3) migration thread wakes up (implicitly it forces the migrated
6620 * thread off the CPU)
6621 * 4) it gets the migration request and checks whether the migrated
6622 * task is still in the wrong runqueue.
6623 * 5) if it's in the wrong runqueue then the migration thread removes
6624 * it and puts it into the right queue.
6625 * 6) migration thread up()s the semaphore.
6626 * 7) we wake up and the migration is done.
6630 * Change a given task's CPU affinity. Migrate the thread to a
6631 * proper CPU and schedule it away if the CPU it's executing on
6632 * is removed from the allowed bitmask.
6634 * NOTE: the caller must have a valid reference to the task, the
6635 * task must not exit() & deallocate itself prematurely. The
6636 * call is not atomic; no spinlocks may be held.
6638 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6640 struct migration_req req
;
6641 unsigned long flags
;
6645 rq
= task_rq_lock(p
, &flags
);
6646 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6651 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6652 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6657 if (p
->sched_class
->set_cpus_allowed
)
6658 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6660 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6661 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6664 /* Can the task run on the task's current CPU? If so, we're done */
6665 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6668 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6669 /* Need help from migration thread: drop lock and wait. */
6670 task_rq_unlock(rq
, &flags
);
6671 wake_up_process(rq
->migration_thread
);
6672 wait_for_completion(&req
.done
);
6673 tlb_migrate_finish(p
->mm
);
6677 task_rq_unlock(rq
, &flags
);
6681 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6684 * Move (not current) task off this cpu, onto dest cpu. We're doing
6685 * this because either it can't run here any more (set_cpus_allowed()
6686 * away from this CPU, or CPU going down), or because we're
6687 * attempting to rebalance this task on exec (sched_exec).
6689 * So we race with normal scheduler movements, but that's OK, as long
6690 * as the task is no longer on this CPU.
6692 * Returns non-zero if task was successfully migrated.
6694 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6696 struct rq
*rq_dest
, *rq_src
;
6699 if (unlikely(!cpu_active(dest_cpu
)))
6702 rq_src
= cpu_rq(src_cpu
);
6703 rq_dest
= cpu_rq(dest_cpu
);
6705 double_rq_lock(rq_src
, rq_dest
);
6706 /* Already moved. */
6707 if (task_cpu(p
) != src_cpu
)
6709 /* Affinity changed (again). */
6710 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6713 on_rq
= p
->se
.on_rq
;
6715 deactivate_task(rq_src
, p
, 0);
6717 set_task_cpu(p
, dest_cpu
);
6719 activate_task(rq_dest
, p
, 0);
6720 check_preempt_curr(rq_dest
, p
, 0);
6725 double_rq_unlock(rq_src
, rq_dest
);
6730 * migration_thread - this is a highprio system thread that performs
6731 * thread migration by bumping thread off CPU then 'pushing' onto
6734 static int migration_thread(void *data
)
6736 int cpu
= (long)data
;
6740 BUG_ON(rq
->migration_thread
!= current
);
6742 set_current_state(TASK_INTERRUPTIBLE
);
6743 while (!kthread_should_stop()) {
6744 struct migration_req
*req
;
6745 struct list_head
*head
;
6747 spin_lock_irq(&rq
->lock
);
6749 if (cpu_is_offline(cpu
)) {
6750 spin_unlock_irq(&rq
->lock
);
6754 if (rq
->active_balance
) {
6755 active_load_balance(rq
, cpu
);
6756 rq
->active_balance
= 0;
6759 head
= &rq
->migration_queue
;
6761 if (list_empty(head
)) {
6762 spin_unlock_irq(&rq
->lock
);
6764 set_current_state(TASK_INTERRUPTIBLE
);
6767 req
= list_entry(head
->next
, struct migration_req
, list
);
6768 list_del_init(head
->next
);
6770 spin_unlock(&rq
->lock
);
6771 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6774 complete(&req
->done
);
6776 __set_current_state(TASK_RUNNING
);
6780 /* Wait for kthread_stop */
6781 set_current_state(TASK_INTERRUPTIBLE
);
6782 while (!kthread_should_stop()) {
6784 set_current_state(TASK_INTERRUPTIBLE
);
6786 __set_current_state(TASK_RUNNING
);
6790 #ifdef CONFIG_HOTPLUG_CPU
6792 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6796 local_irq_disable();
6797 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6803 * Figure out where task on dead CPU should go, use force if necessary.
6805 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6808 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6811 /* Look for allowed, online CPU in same node. */
6812 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6813 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6816 /* Any allowed, online CPU? */
6817 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6818 if (dest_cpu
< nr_cpu_ids
)
6821 /* No more Mr. Nice Guy. */
6822 if (dest_cpu
>= nr_cpu_ids
) {
6823 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6824 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6827 * Don't tell them about moving exiting tasks or
6828 * kernel threads (both mm NULL), since they never
6831 if (p
->mm
&& printk_ratelimit()) {
6832 printk(KERN_INFO
"process %d (%s) no "
6833 "longer affine to cpu%d\n",
6834 task_pid_nr(p
), p
->comm
, dead_cpu
);
6839 /* It can have affinity changed while we were choosing. */
6840 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6845 * While a dead CPU has no uninterruptible tasks queued at this point,
6846 * it might still have a nonzero ->nr_uninterruptible counter, because
6847 * for performance reasons the counter is not stricly tracking tasks to
6848 * their home CPUs. So we just add the counter to another CPU's counter,
6849 * to keep the global sum constant after CPU-down:
6851 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6853 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6854 unsigned long flags
;
6856 local_irq_save(flags
);
6857 double_rq_lock(rq_src
, rq_dest
);
6858 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6859 rq_src
->nr_uninterruptible
= 0;
6860 double_rq_unlock(rq_src
, rq_dest
);
6861 local_irq_restore(flags
);
6864 /* Run through task list and migrate tasks from the dead cpu. */
6865 static void migrate_live_tasks(int src_cpu
)
6867 struct task_struct
*p
, *t
;
6869 read_lock(&tasklist_lock
);
6871 do_each_thread(t
, p
) {
6875 if (task_cpu(p
) == src_cpu
)
6876 move_task_off_dead_cpu(src_cpu
, p
);
6877 } while_each_thread(t
, p
);
6879 read_unlock(&tasklist_lock
);
6883 * Schedules idle task to be the next runnable task on current CPU.
6884 * It does so by boosting its priority to highest possible.
6885 * Used by CPU offline code.
6887 void sched_idle_next(void)
6889 int this_cpu
= smp_processor_id();
6890 struct rq
*rq
= cpu_rq(this_cpu
);
6891 struct task_struct
*p
= rq
->idle
;
6892 unsigned long flags
;
6894 /* cpu has to be offline */
6895 BUG_ON(cpu_online(this_cpu
));
6898 * Strictly not necessary since rest of the CPUs are stopped by now
6899 * and interrupts disabled on the current cpu.
6901 spin_lock_irqsave(&rq
->lock
, flags
);
6903 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6905 update_rq_clock(rq
);
6906 activate_task(rq
, p
, 0);
6908 spin_unlock_irqrestore(&rq
->lock
, flags
);
6912 * Ensures that the idle task is using init_mm right before its cpu goes
6915 void idle_task_exit(void)
6917 struct mm_struct
*mm
= current
->active_mm
;
6919 BUG_ON(cpu_online(smp_processor_id()));
6922 switch_mm(mm
, &init_mm
, current
);
6926 /* called under rq->lock with disabled interrupts */
6927 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6929 struct rq
*rq
= cpu_rq(dead_cpu
);
6931 /* Must be exiting, otherwise would be on tasklist. */
6932 BUG_ON(!p
->exit_state
);
6934 /* Cannot have done final schedule yet: would have vanished. */
6935 BUG_ON(p
->state
== TASK_DEAD
);
6940 * Drop lock around migration; if someone else moves it,
6941 * that's OK. No task can be added to this CPU, so iteration is
6944 spin_unlock_irq(&rq
->lock
);
6945 move_task_off_dead_cpu(dead_cpu
, p
);
6946 spin_lock_irq(&rq
->lock
);
6951 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6952 static void migrate_dead_tasks(unsigned int dead_cpu
)
6954 struct rq
*rq
= cpu_rq(dead_cpu
);
6955 struct task_struct
*next
;
6958 if (!rq
->nr_running
)
6960 update_rq_clock(rq
);
6961 next
= pick_next_task(rq
);
6964 next
->sched_class
->put_prev_task(rq
, next
);
6965 migrate_dead(dead_cpu
, next
);
6969 #endif /* CONFIG_HOTPLUG_CPU */
6971 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6973 static struct ctl_table sd_ctl_dir
[] = {
6975 .procname
= "sched_domain",
6981 static struct ctl_table sd_ctl_root
[] = {
6983 .ctl_name
= CTL_KERN
,
6984 .procname
= "kernel",
6986 .child
= sd_ctl_dir
,
6991 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6993 struct ctl_table
*entry
=
6994 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6999 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7001 struct ctl_table
*entry
;
7004 * In the intermediate directories, both the child directory and
7005 * procname are dynamically allocated and could fail but the mode
7006 * will always be set. In the lowest directory the names are
7007 * static strings and all have proc handlers.
7009 for (entry
= *tablep
; entry
->mode
; entry
++) {
7011 sd_free_ctl_entry(&entry
->child
);
7012 if (entry
->proc_handler
== NULL
)
7013 kfree(entry
->procname
);
7021 set_table_entry(struct ctl_table
*entry
,
7022 const char *procname
, void *data
, int maxlen
,
7023 mode_t mode
, proc_handler
*proc_handler
)
7025 entry
->procname
= procname
;
7027 entry
->maxlen
= maxlen
;
7029 entry
->proc_handler
= proc_handler
;
7032 static struct ctl_table
*
7033 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7035 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7040 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7041 sizeof(long), 0644, proc_doulongvec_minmax
);
7042 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7043 sizeof(long), 0644, proc_doulongvec_minmax
);
7044 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7045 sizeof(int), 0644, proc_dointvec_minmax
);
7046 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7047 sizeof(int), 0644, proc_dointvec_minmax
);
7048 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7049 sizeof(int), 0644, proc_dointvec_minmax
);
7050 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7051 sizeof(int), 0644, proc_dointvec_minmax
);
7052 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7053 sizeof(int), 0644, proc_dointvec_minmax
);
7054 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7055 sizeof(int), 0644, proc_dointvec_minmax
);
7056 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7057 sizeof(int), 0644, proc_dointvec_minmax
);
7058 set_table_entry(&table
[9], "cache_nice_tries",
7059 &sd
->cache_nice_tries
,
7060 sizeof(int), 0644, proc_dointvec_minmax
);
7061 set_table_entry(&table
[10], "flags", &sd
->flags
,
7062 sizeof(int), 0644, proc_dointvec_minmax
);
7063 set_table_entry(&table
[11], "name", sd
->name
,
7064 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7065 /* &table[12] is terminator */
7070 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7072 struct ctl_table
*entry
, *table
;
7073 struct sched_domain
*sd
;
7074 int domain_num
= 0, i
;
7077 for_each_domain(cpu
, sd
)
7079 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7084 for_each_domain(cpu
, sd
) {
7085 snprintf(buf
, 32, "domain%d", i
);
7086 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7088 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7095 static struct ctl_table_header
*sd_sysctl_header
;
7096 static void register_sched_domain_sysctl(void)
7098 int i
, cpu_num
= num_online_cpus();
7099 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7102 WARN_ON(sd_ctl_dir
[0].child
);
7103 sd_ctl_dir
[0].child
= entry
;
7108 for_each_online_cpu(i
) {
7109 snprintf(buf
, 32, "cpu%d", i
);
7110 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7112 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7116 WARN_ON(sd_sysctl_header
);
7117 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7120 /* may be called multiple times per register */
7121 static void unregister_sched_domain_sysctl(void)
7123 if (sd_sysctl_header
)
7124 unregister_sysctl_table(sd_sysctl_header
);
7125 sd_sysctl_header
= NULL
;
7126 if (sd_ctl_dir
[0].child
)
7127 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7130 static void register_sched_domain_sysctl(void)
7133 static void unregister_sched_domain_sysctl(void)
7138 static void set_rq_online(struct rq
*rq
)
7141 const struct sched_class
*class;
7143 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7146 for_each_class(class) {
7147 if (class->rq_online
)
7148 class->rq_online(rq
);
7153 static void set_rq_offline(struct rq
*rq
)
7156 const struct sched_class
*class;
7158 for_each_class(class) {
7159 if (class->rq_offline
)
7160 class->rq_offline(rq
);
7163 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7169 * migration_call - callback that gets triggered when a CPU is added.
7170 * Here we can start up the necessary migration thread for the new CPU.
7172 static int __cpuinit
7173 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7175 struct task_struct
*p
;
7176 int cpu
= (long)hcpu
;
7177 unsigned long flags
;
7182 case CPU_UP_PREPARE
:
7183 case CPU_UP_PREPARE_FROZEN
:
7184 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7187 kthread_bind(p
, cpu
);
7188 /* Must be high prio: stop_machine expects to yield to it. */
7189 rq
= task_rq_lock(p
, &flags
);
7190 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7191 task_rq_unlock(rq
, &flags
);
7192 cpu_rq(cpu
)->migration_thread
= p
;
7196 case CPU_ONLINE_FROZEN
:
7197 /* Strictly unnecessary, as first user will wake it. */
7198 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7200 /* Update our root-domain */
7202 spin_lock_irqsave(&rq
->lock
, flags
);
7204 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7208 spin_unlock_irqrestore(&rq
->lock
, flags
);
7211 #ifdef CONFIG_HOTPLUG_CPU
7212 case CPU_UP_CANCELED
:
7213 case CPU_UP_CANCELED_FROZEN
:
7214 if (!cpu_rq(cpu
)->migration_thread
)
7216 /* Unbind it from offline cpu so it can run. Fall thru. */
7217 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7218 cpumask_any(cpu_online_mask
));
7219 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7220 cpu_rq(cpu
)->migration_thread
= NULL
;
7224 case CPU_DEAD_FROZEN
:
7225 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7226 migrate_live_tasks(cpu
);
7228 kthread_stop(rq
->migration_thread
);
7229 rq
->migration_thread
= NULL
;
7230 /* Idle task back to normal (off runqueue, low prio) */
7231 spin_lock_irq(&rq
->lock
);
7232 update_rq_clock(rq
);
7233 deactivate_task(rq
, rq
->idle
, 0);
7234 rq
->idle
->static_prio
= MAX_PRIO
;
7235 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7236 rq
->idle
->sched_class
= &idle_sched_class
;
7237 migrate_dead_tasks(cpu
);
7238 spin_unlock_irq(&rq
->lock
);
7240 migrate_nr_uninterruptible(rq
);
7241 BUG_ON(rq
->nr_running
!= 0);
7244 * No need to migrate the tasks: it was best-effort if
7245 * they didn't take sched_hotcpu_mutex. Just wake up
7248 spin_lock_irq(&rq
->lock
);
7249 while (!list_empty(&rq
->migration_queue
)) {
7250 struct migration_req
*req
;
7252 req
= list_entry(rq
->migration_queue
.next
,
7253 struct migration_req
, list
);
7254 list_del_init(&req
->list
);
7255 spin_unlock_irq(&rq
->lock
);
7256 complete(&req
->done
);
7257 spin_lock_irq(&rq
->lock
);
7259 spin_unlock_irq(&rq
->lock
);
7263 case CPU_DYING_FROZEN
:
7264 /* Update our root-domain */
7266 spin_lock_irqsave(&rq
->lock
, flags
);
7268 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7271 spin_unlock_irqrestore(&rq
->lock
, flags
);
7278 /* Register at highest priority so that task migration (migrate_all_tasks)
7279 * happens before everything else.
7281 static struct notifier_block __cpuinitdata migration_notifier
= {
7282 .notifier_call
= migration_call
,
7286 static int __init
migration_init(void)
7288 void *cpu
= (void *)(long)smp_processor_id();
7291 /* Start one for the boot CPU: */
7292 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7293 BUG_ON(err
== NOTIFY_BAD
);
7294 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7295 register_cpu_notifier(&migration_notifier
);
7299 early_initcall(migration_init
);
7304 #ifdef CONFIG_SCHED_DEBUG
7306 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7307 struct cpumask
*groupmask
)
7309 struct sched_group
*group
= sd
->groups
;
7312 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7313 cpumask_clear(groupmask
);
7315 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7317 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7318 printk("does not load-balance\n");
7320 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7325 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7327 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7328 printk(KERN_ERR
"ERROR: domain->span does not contain "
7331 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7332 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7336 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7340 printk(KERN_ERR
"ERROR: group is NULL\n");
7344 if (!group
->__cpu_power
) {
7345 printk(KERN_CONT
"\n");
7346 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7351 if (!cpumask_weight(sched_group_cpus(group
))) {
7352 printk(KERN_CONT
"\n");
7353 printk(KERN_ERR
"ERROR: empty group\n");
7357 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7358 printk(KERN_CONT
"\n");
7359 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7363 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7365 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7366 printk(KERN_CONT
" %s (__cpu_power = %d)", str
,
7367 group
->__cpu_power
);
7369 group
= group
->next
;
7370 } while (group
!= sd
->groups
);
7371 printk(KERN_CONT
"\n");
7373 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7374 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7377 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7378 printk(KERN_ERR
"ERROR: parent span is not a superset "
7379 "of domain->span\n");
7383 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7385 cpumask_var_t groupmask
;
7389 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7393 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7395 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7396 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7401 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7408 free_cpumask_var(groupmask
);
7410 #else /* !CONFIG_SCHED_DEBUG */
7411 # define sched_domain_debug(sd, cpu) do { } while (0)
7412 #endif /* CONFIG_SCHED_DEBUG */
7414 static int sd_degenerate(struct sched_domain
*sd
)
7416 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7419 /* Following flags need at least 2 groups */
7420 if (sd
->flags
& (SD_LOAD_BALANCE
|
7421 SD_BALANCE_NEWIDLE
|
7425 SD_SHARE_PKG_RESOURCES
)) {
7426 if (sd
->groups
!= sd
->groups
->next
)
7430 /* Following flags don't use groups */
7431 if (sd
->flags
& (SD_WAKE_IDLE
|
7440 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7442 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7444 if (sd_degenerate(parent
))
7447 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7450 /* Does parent contain flags not in child? */
7451 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7452 if (cflags
& SD_WAKE_AFFINE
)
7453 pflags
&= ~SD_WAKE_BALANCE
;
7454 /* Flags needing groups don't count if only 1 group in parent */
7455 if (parent
->groups
== parent
->groups
->next
) {
7456 pflags
&= ~(SD_LOAD_BALANCE
|
7457 SD_BALANCE_NEWIDLE
|
7461 SD_SHARE_PKG_RESOURCES
);
7462 if (nr_node_ids
== 1)
7463 pflags
&= ~SD_SERIALIZE
;
7465 if (~cflags
& pflags
)
7471 static void free_rootdomain(struct root_domain
*rd
)
7473 cpupri_cleanup(&rd
->cpupri
);
7475 free_cpumask_var(rd
->rto_mask
);
7476 free_cpumask_var(rd
->online
);
7477 free_cpumask_var(rd
->span
);
7481 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7483 struct root_domain
*old_rd
= NULL
;
7484 unsigned long flags
;
7486 spin_lock_irqsave(&rq
->lock
, flags
);
7491 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7494 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7497 * If we dont want to free the old_rt yet then
7498 * set old_rd to NULL to skip the freeing later
7501 if (!atomic_dec_and_test(&old_rd
->refcount
))
7505 atomic_inc(&rd
->refcount
);
7508 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7509 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7512 spin_unlock_irqrestore(&rq
->lock
, flags
);
7515 free_rootdomain(old_rd
);
7518 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7520 memset(rd
, 0, sizeof(*rd
));
7523 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7524 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7525 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7526 cpupri_init(&rd
->cpupri
, true);
7530 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7532 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7534 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7537 if (cpupri_init(&rd
->cpupri
, false) != 0)
7542 free_cpumask_var(rd
->rto_mask
);
7544 free_cpumask_var(rd
->online
);
7546 free_cpumask_var(rd
->span
);
7551 static void init_defrootdomain(void)
7553 init_rootdomain(&def_root_domain
, true);
7555 atomic_set(&def_root_domain
.refcount
, 1);
7558 static struct root_domain
*alloc_rootdomain(void)
7560 struct root_domain
*rd
;
7562 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7566 if (init_rootdomain(rd
, false) != 0) {
7575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7576 * hold the hotplug lock.
7579 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7581 struct rq
*rq
= cpu_rq(cpu
);
7582 struct sched_domain
*tmp
;
7584 /* Remove the sched domains which do not contribute to scheduling. */
7585 for (tmp
= sd
; tmp
; ) {
7586 struct sched_domain
*parent
= tmp
->parent
;
7590 if (sd_parent_degenerate(tmp
, parent
)) {
7591 tmp
->parent
= parent
->parent
;
7593 parent
->parent
->child
= tmp
;
7598 if (sd
&& sd_degenerate(sd
)) {
7604 sched_domain_debug(sd
, cpu
);
7606 rq_attach_root(rq
, rd
);
7607 rcu_assign_pointer(rq
->sd
, sd
);
7610 /* cpus with isolated domains */
7611 static cpumask_var_t cpu_isolated_map
;
7613 /* Setup the mask of cpus configured for isolated domains */
7614 static int __init
isolated_cpu_setup(char *str
)
7616 cpulist_parse(str
, cpu_isolated_map
);
7620 __setup("isolcpus=", isolated_cpu_setup
);
7623 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7624 * to a function which identifies what group(along with sched group) a CPU
7625 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7626 * (due to the fact that we keep track of groups covered with a struct cpumask).
7628 * init_sched_build_groups will build a circular linked list of the groups
7629 * covered by the given span, and will set each group's ->cpumask correctly,
7630 * and ->cpu_power to 0.
7633 init_sched_build_groups(const struct cpumask
*span
,
7634 const struct cpumask
*cpu_map
,
7635 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7636 struct sched_group
**sg
,
7637 struct cpumask
*tmpmask
),
7638 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7640 struct sched_group
*first
= NULL
, *last
= NULL
;
7643 cpumask_clear(covered
);
7645 for_each_cpu(i
, span
) {
7646 struct sched_group
*sg
;
7647 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7650 if (cpumask_test_cpu(i
, covered
))
7653 cpumask_clear(sched_group_cpus(sg
));
7654 sg
->__cpu_power
= 0;
7656 for_each_cpu(j
, span
) {
7657 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7660 cpumask_set_cpu(j
, covered
);
7661 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7672 #define SD_NODES_PER_DOMAIN 16
7677 * find_next_best_node - find the next node to include in a sched_domain
7678 * @node: node whose sched_domain we're building
7679 * @used_nodes: nodes already in the sched_domain
7681 * Find the next node to include in a given scheduling domain. Simply
7682 * finds the closest node not already in the @used_nodes map.
7684 * Should use nodemask_t.
7686 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7688 int i
, n
, val
, min_val
, best_node
= 0;
7692 for (i
= 0; i
< nr_node_ids
; i
++) {
7693 /* Start at @node */
7694 n
= (node
+ i
) % nr_node_ids
;
7696 if (!nr_cpus_node(n
))
7699 /* Skip already used nodes */
7700 if (node_isset(n
, *used_nodes
))
7703 /* Simple min distance search */
7704 val
= node_distance(node
, n
);
7706 if (val
< min_val
) {
7712 node_set(best_node
, *used_nodes
);
7717 * sched_domain_node_span - get a cpumask for a node's sched_domain
7718 * @node: node whose cpumask we're constructing
7719 * @span: resulting cpumask
7721 * Given a node, construct a good cpumask for its sched_domain to span. It
7722 * should be one that prevents unnecessary balancing, but also spreads tasks
7725 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7727 nodemask_t used_nodes
;
7730 cpumask_clear(span
);
7731 nodes_clear(used_nodes
);
7733 cpumask_or(span
, span
, cpumask_of_node(node
));
7734 node_set(node
, used_nodes
);
7736 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7737 int next_node
= find_next_best_node(node
, &used_nodes
);
7739 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7742 #endif /* CONFIG_NUMA */
7744 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7747 * The cpus mask in sched_group and sched_domain hangs off the end.
7748 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7749 * for nr_cpu_ids < CONFIG_NR_CPUS.
7751 struct static_sched_group
{
7752 struct sched_group sg
;
7753 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7756 struct static_sched_domain
{
7757 struct sched_domain sd
;
7758 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7762 * SMT sched-domains:
7764 #ifdef CONFIG_SCHED_SMT
7765 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7766 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7769 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7770 struct sched_group
**sg
, struct cpumask
*unused
)
7773 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7776 #endif /* CONFIG_SCHED_SMT */
7779 * multi-core sched-domains:
7781 #ifdef CONFIG_SCHED_MC
7782 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7783 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7784 #endif /* CONFIG_SCHED_MC */
7786 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7788 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7789 struct sched_group
**sg
, struct cpumask
*mask
)
7793 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7794 group
= cpumask_first(mask
);
7796 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7799 #elif defined(CONFIG_SCHED_MC)
7801 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7802 struct sched_group
**sg
, struct cpumask
*unused
)
7805 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7810 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7811 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7814 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7815 struct sched_group
**sg
, struct cpumask
*mask
)
7818 #ifdef CONFIG_SCHED_MC
7819 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7820 group
= cpumask_first(mask
);
7821 #elif defined(CONFIG_SCHED_SMT)
7822 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7823 group
= cpumask_first(mask
);
7828 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7834 * The init_sched_build_groups can't handle what we want to do with node
7835 * groups, so roll our own. Now each node has its own list of groups which
7836 * gets dynamically allocated.
7838 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7839 static struct sched_group
***sched_group_nodes_bycpu
;
7841 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7842 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7844 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7845 struct sched_group
**sg
,
7846 struct cpumask
*nodemask
)
7850 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7851 group
= cpumask_first(nodemask
);
7854 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7858 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7860 struct sched_group
*sg
= group_head
;
7866 for_each_cpu(j
, sched_group_cpus(sg
)) {
7867 struct sched_domain
*sd
;
7869 sd
= &per_cpu(phys_domains
, j
).sd
;
7870 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7872 * Only add "power" once for each
7878 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7881 } while (sg
!= group_head
);
7883 #endif /* CONFIG_NUMA */
7886 /* Free memory allocated for various sched_group structures */
7887 static void free_sched_groups(const struct cpumask
*cpu_map
,
7888 struct cpumask
*nodemask
)
7892 for_each_cpu(cpu
, cpu_map
) {
7893 struct sched_group
**sched_group_nodes
7894 = sched_group_nodes_bycpu
[cpu
];
7896 if (!sched_group_nodes
)
7899 for (i
= 0; i
< nr_node_ids
; i
++) {
7900 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7902 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7903 if (cpumask_empty(nodemask
))
7913 if (oldsg
!= sched_group_nodes
[i
])
7916 kfree(sched_group_nodes
);
7917 sched_group_nodes_bycpu
[cpu
] = NULL
;
7920 #else /* !CONFIG_NUMA */
7921 static void free_sched_groups(const struct cpumask
*cpu_map
,
7922 struct cpumask
*nodemask
)
7925 #endif /* CONFIG_NUMA */
7928 * Initialize sched groups cpu_power.
7930 * cpu_power indicates the capacity of sched group, which is used while
7931 * distributing the load between different sched groups in a sched domain.
7932 * Typically cpu_power for all the groups in a sched domain will be same unless
7933 * there are asymmetries in the topology. If there are asymmetries, group
7934 * having more cpu_power will pickup more load compared to the group having
7937 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7938 * the maximum number of tasks a group can handle in the presence of other idle
7939 * or lightly loaded groups in the same sched domain.
7941 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7943 struct sched_domain
*child
;
7944 struct sched_group
*group
;
7946 WARN_ON(!sd
|| !sd
->groups
);
7948 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7953 sd
->groups
->__cpu_power
= 0;
7956 * For perf policy, if the groups in child domain share resources
7957 * (for example cores sharing some portions of the cache hierarchy
7958 * or SMT), then set this domain groups cpu_power such that each group
7959 * can handle only one task, when there are other idle groups in the
7960 * same sched domain.
7962 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7964 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7965 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7970 * add cpu_power of each child group to this groups cpu_power
7972 group
= child
->groups
;
7974 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7975 group
= group
->next
;
7976 } while (group
!= child
->groups
);
7980 * Initializers for schedule domains
7981 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7984 #ifdef CONFIG_SCHED_DEBUG
7985 # define SD_INIT_NAME(sd, type) sd->name = #type
7987 # define SD_INIT_NAME(sd, type) do { } while (0)
7990 #define SD_INIT(sd, type) sd_init_##type(sd)
7992 #define SD_INIT_FUNC(type) \
7993 static noinline void sd_init_##type(struct sched_domain *sd) \
7995 memset(sd, 0, sizeof(*sd)); \
7996 *sd = SD_##type##_INIT; \
7997 sd->level = SD_LV_##type; \
7998 SD_INIT_NAME(sd, type); \
8003 SD_INIT_FUNC(ALLNODES
)
8006 #ifdef CONFIG_SCHED_SMT
8007 SD_INIT_FUNC(SIBLING
)
8009 #ifdef CONFIG_SCHED_MC
8013 static int default_relax_domain_level
= -1;
8015 static int __init
setup_relax_domain_level(char *str
)
8019 val
= simple_strtoul(str
, NULL
, 0);
8020 if (val
< SD_LV_MAX
)
8021 default_relax_domain_level
= val
;
8025 __setup("relax_domain_level=", setup_relax_domain_level
);
8027 static void set_domain_attribute(struct sched_domain
*sd
,
8028 struct sched_domain_attr
*attr
)
8032 if (!attr
|| attr
->relax_domain_level
< 0) {
8033 if (default_relax_domain_level
< 0)
8036 request
= default_relax_domain_level
;
8038 request
= attr
->relax_domain_level
;
8039 if (request
< sd
->level
) {
8040 /* turn off idle balance on this domain */
8041 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8043 /* turn on idle balance on this domain */
8044 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8049 * Build sched domains for a given set of cpus and attach the sched domains
8050 * to the individual cpus
8052 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8053 struct sched_domain_attr
*attr
)
8055 int i
, err
= -ENOMEM
;
8056 struct root_domain
*rd
;
8057 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8060 cpumask_var_t domainspan
, covered
, notcovered
;
8061 struct sched_group
**sched_group_nodes
= NULL
;
8062 int sd_allnodes
= 0;
8064 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8066 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8067 goto free_domainspan
;
8068 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8072 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8073 goto free_notcovered
;
8074 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8076 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8077 goto free_this_sibling_map
;
8078 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8079 goto free_this_core_map
;
8080 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8081 goto free_send_covered
;
8085 * Allocate the per-node list of sched groups
8087 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8089 if (!sched_group_nodes
) {
8090 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8095 rd
= alloc_rootdomain();
8097 printk(KERN_WARNING
"Cannot alloc root domain\n");
8098 goto free_sched_groups
;
8102 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8106 * Set up domains for cpus specified by the cpu_map.
8108 for_each_cpu(i
, cpu_map
) {
8109 struct sched_domain
*sd
= NULL
, *p
;
8111 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8114 if (cpumask_weight(cpu_map
) >
8115 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8116 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8117 SD_INIT(sd
, ALLNODES
);
8118 set_domain_attribute(sd
, attr
);
8119 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8120 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8126 sd
= &per_cpu(node_domains
, i
).sd
;
8128 set_domain_attribute(sd
, attr
);
8129 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8133 cpumask_and(sched_domain_span(sd
),
8134 sched_domain_span(sd
), cpu_map
);
8138 sd
= &per_cpu(phys_domains
, i
).sd
;
8140 set_domain_attribute(sd
, attr
);
8141 cpumask_copy(sched_domain_span(sd
), nodemask
);
8145 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8147 #ifdef CONFIG_SCHED_MC
8149 sd
= &per_cpu(core_domains
, i
).sd
;
8151 set_domain_attribute(sd
, attr
);
8152 cpumask_and(sched_domain_span(sd
), cpu_map
,
8153 cpu_coregroup_mask(i
));
8156 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8159 #ifdef CONFIG_SCHED_SMT
8161 sd
= &per_cpu(cpu_domains
, i
).sd
;
8162 SD_INIT(sd
, SIBLING
);
8163 set_domain_attribute(sd
, attr
);
8164 cpumask_and(sched_domain_span(sd
),
8165 topology_thread_cpumask(i
), cpu_map
);
8168 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8172 #ifdef CONFIG_SCHED_SMT
8173 /* Set up CPU (sibling) groups */
8174 for_each_cpu(i
, cpu_map
) {
8175 cpumask_and(this_sibling_map
,
8176 topology_thread_cpumask(i
), cpu_map
);
8177 if (i
!= cpumask_first(this_sibling_map
))
8180 init_sched_build_groups(this_sibling_map
, cpu_map
,
8182 send_covered
, tmpmask
);
8186 #ifdef CONFIG_SCHED_MC
8187 /* Set up multi-core groups */
8188 for_each_cpu(i
, cpu_map
) {
8189 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8190 if (i
!= cpumask_first(this_core_map
))
8193 init_sched_build_groups(this_core_map
, cpu_map
,
8195 send_covered
, tmpmask
);
8199 /* Set up physical groups */
8200 for (i
= 0; i
< nr_node_ids
; i
++) {
8201 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8202 if (cpumask_empty(nodemask
))
8205 init_sched_build_groups(nodemask
, cpu_map
,
8207 send_covered
, tmpmask
);
8211 /* Set up node groups */
8213 init_sched_build_groups(cpu_map
, cpu_map
,
8214 &cpu_to_allnodes_group
,
8215 send_covered
, tmpmask
);
8218 for (i
= 0; i
< nr_node_ids
; i
++) {
8219 /* Set up node groups */
8220 struct sched_group
*sg
, *prev
;
8223 cpumask_clear(covered
);
8224 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8225 if (cpumask_empty(nodemask
)) {
8226 sched_group_nodes
[i
] = NULL
;
8230 sched_domain_node_span(i
, domainspan
);
8231 cpumask_and(domainspan
, domainspan
, cpu_map
);
8233 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8236 printk(KERN_WARNING
"Can not alloc domain group for "
8240 sched_group_nodes
[i
] = sg
;
8241 for_each_cpu(j
, nodemask
) {
8242 struct sched_domain
*sd
;
8244 sd
= &per_cpu(node_domains
, j
).sd
;
8247 sg
->__cpu_power
= 0;
8248 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8250 cpumask_or(covered
, covered
, nodemask
);
8253 for (j
= 0; j
< nr_node_ids
; j
++) {
8254 int n
= (i
+ j
) % nr_node_ids
;
8256 cpumask_complement(notcovered
, covered
);
8257 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8258 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8259 if (cpumask_empty(tmpmask
))
8262 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8263 if (cpumask_empty(tmpmask
))
8266 sg
= kmalloc_node(sizeof(struct sched_group
) +
8271 "Can not alloc domain group for node %d\n", j
);
8274 sg
->__cpu_power
= 0;
8275 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8276 sg
->next
= prev
->next
;
8277 cpumask_or(covered
, covered
, tmpmask
);
8284 /* Calculate CPU power for physical packages and nodes */
8285 #ifdef CONFIG_SCHED_SMT
8286 for_each_cpu(i
, cpu_map
) {
8287 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8289 init_sched_groups_power(i
, sd
);
8292 #ifdef CONFIG_SCHED_MC
8293 for_each_cpu(i
, cpu_map
) {
8294 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8296 init_sched_groups_power(i
, sd
);
8300 for_each_cpu(i
, cpu_map
) {
8301 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8303 init_sched_groups_power(i
, sd
);
8307 for (i
= 0; i
< nr_node_ids
; i
++)
8308 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8311 struct sched_group
*sg
;
8313 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8315 init_numa_sched_groups_power(sg
);
8319 /* Attach the domains */
8320 for_each_cpu(i
, cpu_map
) {
8321 struct sched_domain
*sd
;
8322 #ifdef CONFIG_SCHED_SMT
8323 sd
= &per_cpu(cpu_domains
, i
).sd
;
8324 #elif defined(CONFIG_SCHED_MC)
8325 sd
= &per_cpu(core_domains
, i
).sd
;
8327 sd
= &per_cpu(phys_domains
, i
).sd
;
8329 cpu_attach_domain(sd
, rd
, i
);
8335 free_cpumask_var(tmpmask
);
8337 free_cpumask_var(send_covered
);
8339 free_cpumask_var(this_core_map
);
8340 free_this_sibling_map
:
8341 free_cpumask_var(this_sibling_map
);
8343 free_cpumask_var(nodemask
);
8346 free_cpumask_var(notcovered
);
8348 free_cpumask_var(covered
);
8350 free_cpumask_var(domainspan
);
8357 kfree(sched_group_nodes
);
8363 free_sched_groups(cpu_map
, tmpmask
);
8364 free_rootdomain(rd
);
8369 static int build_sched_domains(const struct cpumask
*cpu_map
)
8371 return __build_sched_domains(cpu_map
, NULL
);
8374 static struct cpumask
*doms_cur
; /* current sched domains */
8375 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8376 static struct sched_domain_attr
*dattr_cur
;
8377 /* attribues of custom domains in 'doms_cur' */
8380 * Special case: If a kmalloc of a doms_cur partition (array of
8381 * cpumask) fails, then fallback to a single sched domain,
8382 * as determined by the single cpumask fallback_doms.
8384 static cpumask_var_t fallback_doms
;
8387 * arch_update_cpu_topology lets virtualized architectures update the
8388 * cpu core maps. It is supposed to return 1 if the topology changed
8389 * or 0 if it stayed the same.
8391 int __attribute__((weak
)) arch_update_cpu_topology(void)
8397 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8398 * For now this just excludes isolated cpus, but could be used to
8399 * exclude other special cases in the future.
8401 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8405 arch_update_cpu_topology();
8407 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8409 doms_cur
= fallback_doms
;
8410 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8412 err
= build_sched_domains(doms_cur
);
8413 register_sched_domain_sysctl();
8418 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8419 struct cpumask
*tmpmask
)
8421 free_sched_groups(cpu_map
, tmpmask
);
8425 * Detach sched domains from a group of cpus specified in cpu_map
8426 * These cpus will now be attached to the NULL domain
8428 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8430 /* Save because hotplug lock held. */
8431 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8434 for_each_cpu(i
, cpu_map
)
8435 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8436 synchronize_sched();
8437 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8440 /* handle null as "default" */
8441 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8442 struct sched_domain_attr
*new, int idx_new
)
8444 struct sched_domain_attr tmp
;
8451 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8452 new ? (new + idx_new
) : &tmp
,
8453 sizeof(struct sched_domain_attr
));
8457 * Partition sched domains as specified by the 'ndoms_new'
8458 * cpumasks in the array doms_new[] of cpumasks. This compares
8459 * doms_new[] to the current sched domain partitioning, doms_cur[].
8460 * It destroys each deleted domain and builds each new domain.
8462 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8463 * The masks don't intersect (don't overlap.) We should setup one
8464 * sched domain for each mask. CPUs not in any of the cpumasks will
8465 * not be load balanced. If the same cpumask appears both in the
8466 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8469 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8470 * ownership of it and will kfree it when done with it. If the caller
8471 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8472 * ndoms_new == 1, and partition_sched_domains() will fallback to
8473 * the single partition 'fallback_doms', it also forces the domains
8476 * If doms_new == NULL it will be replaced with cpu_online_mask.
8477 * ndoms_new == 0 is a special case for destroying existing domains,
8478 * and it will not create the default domain.
8480 * Call with hotplug lock held
8482 /* FIXME: Change to struct cpumask *doms_new[] */
8483 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8484 struct sched_domain_attr
*dattr_new
)
8489 mutex_lock(&sched_domains_mutex
);
8491 /* always unregister in case we don't destroy any domains */
8492 unregister_sched_domain_sysctl();
8494 /* Let architecture update cpu core mappings. */
8495 new_topology
= arch_update_cpu_topology();
8497 n
= doms_new
? ndoms_new
: 0;
8499 /* Destroy deleted domains */
8500 for (i
= 0; i
< ndoms_cur
; i
++) {
8501 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8502 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8503 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8506 /* no match - a current sched domain not in new doms_new[] */
8507 detach_destroy_domains(doms_cur
+ i
);
8512 if (doms_new
== NULL
) {
8514 doms_new
= fallback_doms
;
8515 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8516 WARN_ON_ONCE(dattr_new
);
8519 /* Build new domains */
8520 for (i
= 0; i
< ndoms_new
; i
++) {
8521 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8522 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8523 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8526 /* no match - add a new doms_new */
8527 __build_sched_domains(doms_new
+ i
,
8528 dattr_new
? dattr_new
+ i
: NULL
);
8533 /* Remember the new sched domains */
8534 if (doms_cur
!= fallback_doms
)
8536 kfree(dattr_cur
); /* kfree(NULL) is safe */
8537 doms_cur
= doms_new
;
8538 dattr_cur
= dattr_new
;
8539 ndoms_cur
= ndoms_new
;
8541 register_sched_domain_sysctl();
8543 mutex_unlock(&sched_domains_mutex
);
8546 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8547 static void arch_reinit_sched_domains(void)
8551 /* Destroy domains first to force the rebuild */
8552 partition_sched_domains(0, NULL
, NULL
);
8554 rebuild_sched_domains();
8558 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8560 unsigned int level
= 0;
8562 if (sscanf(buf
, "%u", &level
) != 1)
8566 * level is always be positive so don't check for
8567 * level < POWERSAVINGS_BALANCE_NONE which is 0
8568 * What happens on 0 or 1 byte write,
8569 * need to check for count as well?
8572 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8576 sched_smt_power_savings
= level
;
8578 sched_mc_power_savings
= level
;
8580 arch_reinit_sched_domains();
8585 #ifdef CONFIG_SCHED_MC
8586 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8589 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8591 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8592 const char *buf
, size_t count
)
8594 return sched_power_savings_store(buf
, count
, 0);
8596 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8597 sched_mc_power_savings_show
,
8598 sched_mc_power_savings_store
);
8601 #ifdef CONFIG_SCHED_SMT
8602 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8605 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8607 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8608 const char *buf
, size_t count
)
8610 return sched_power_savings_store(buf
, count
, 1);
8612 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8613 sched_smt_power_savings_show
,
8614 sched_smt_power_savings_store
);
8617 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8621 #ifdef CONFIG_SCHED_SMT
8623 err
= sysfs_create_file(&cls
->kset
.kobj
,
8624 &attr_sched_smt_power_savings
.attr
);
8626 #ifdef CONFIG_SCHED_MC
8627 if (!err
&& mc_capable())
8628 err
= sysfs_create_file(&cls
->kset
.kobj
,
8629 &attr_sched_mc_power_savings
.attr
);
8633 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8635 #ifndef CONFIG_CPUSETS
8637 * Add online and remove offline CPUs from the scheduler domains.
8638 * When cpusets are enabled they take over this function.
8640 static int update_sched_domains(struct notifier_block
*nfb
,
8641 unsigned long action
, void *hcpu
)
8645 case CPU_ONLINE_FROZEN
:
8647 case CPU_DEAD_FROZEN
:
8648 partition_sched_domains(1, NULL
, NULL
);
8657 static int update_runtime(struct notifier_block
*nfb
,
8658 unsigned long action
, void *hcpu
)
8660 int cpu
= (int)(long)hcpu
;
8663 case CPU_DOWN_PREPARE
:
8664 case CPU_DOWN_PREPARE_FROZEN
:
8665 disable_runtime(cpu_rq(cpu
));
8668 case CPU_DOWN_FAILED
:
8669 case CPU_DOWN_FAILED_FROZEN
:
8671 case CPU_ONLINE_FROZEN
:
8672 enable_runtime(cpu_rq(cpu
));
8680 void __init
sched_init_smp(void)
8682 cpumask_var_t non_isolated_cpus
;
8684 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8686 #if defined(CONFIG_NUMA)
8687 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8689 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8692 mutex_lock(&sched_domains_mutex
);
8693 arch_init_sched_domains(cpu_online_mask
);
8694 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8695 if (cpumask_empty(non_isolated_cpus
))
8696 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8697 mutex_unlock(&sched_domains_mutex
);
8700 #ifndef CONFIG_CPUSETS
8701 /* XXX: Theoretical race here - CPU may be hotplugged now */
8702 hotcpu_notifier(update_sched_domains
, 0);
8705 /* RT runtime code needs to handle some hotplug events */
8706 hotcpu_notifier(update_runtime
, 0);
8710 /* Move init over to a non-isolated CPU */
8711 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8713 sched_init_granularity();
8714 free_cpumask_var(non_isolated_cpus
);
8716 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8717 init_sched_rt_class();
8720 void __init
sched_init_smp(void)
8722 sched_init_granularity();
8724 #endif /* CONFIG_SMP */
8726 int in_sched_functions(unsigned long addr
)
8728 return in_lock_functions(addr
) ||
8729 (addr
>= (unsigned long)__sched_text_start
8730 && addr
< (unsigned long)__sched_text_end
);
8733 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8735 cfs_rq
->tasks_timeline
= RB_ROOT
;
8736 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8737 #ifdef CONFIG_FAIR_GROUP_SCHED
8740 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8743 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8745 struct rt_prio_array
*array
;
8748 array
= &rt_rq
->active
;
8749 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8750 INIT_LIST_HEAD(array
->queue
+ i
);
8751 __clear_bit(i
, array
->bitmap
);
8753 /* delimiter for bitsearch: */
8754 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8756 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8757 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8759 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8763 rt_rq
->rt_nr_migratory
= 0;
8764 rt_rq
->overloaded
= 0;
8765 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8769 rt_rq
->rt_throttled
= 0;
8770 rt_rq
->rt_runtime
= 0;
8771 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8773 #ifdef CONFIG_RT_GROUP_SCHED
8774 rt_rq
->rt_nr_boosted
= 0;
8779 #ifdef CONFIG_FAIR_GROUP_SCHED
8780 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8781 struct sched_entity
*se
, int cpu
, int add
,
8782 struct sched_entity
*parent
)
8784 struct rq
*rq
= cpu_rq(cpu
);
8785 tg
->cfs_rq
[cpu
] = cfs_rq
;
8786 init_cfs_rq(cfs_rq
, rq
);
8789 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8792 /* se could be NULL for init_task_group */
8797 se
->cfs_rq
= &rq
->cfs
;
8799 se
->cfs_rq
= parent
->my_q
;
8802 se
->load
.weight
= tg
->shares
;
8803 se
->load
.inv_weight
= 0;
8804 se
->parent
= parent
;
8808 #ifdef CONFIG_RT_GROUP_SCHED
8809 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8810 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8811 struct sched_rt_entity
*parent
)
8813 struct rq
*rq
= cpu_rq(cpu
);
8815 tg
->rt_rq
[cpu
] = rt_rq
;
8816 init_rt_rq(rt_rq
, rq
);
8818 rt_rq
->rt_se
= rt_se
;
8819 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8821 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8823 tg
->rt_se
[cpu
] = rt_se
;
8828 rt_se
->rt_rq
= &rq
->rt
;
8830 rt_se
->rt_rq
= parent
->my_q
;
8832 rt_se
->my_q
= rt_rq
;
8833 rt_se
->parent
= parent
;
8834 INIT_LIST_HEAD(&rt_se
->run_list
);
8838 void __init
sched_init(void)
8841 unsigned long alloc_size
= 0, ptr
;
8843 #ifdef CONFIG_FAIR_GROUP_SCHED
8844 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8846 #ifdef CONFIG_RT_GROUP_SCHED
8847 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8849 #ifdef CONFIG_USER_SCHED
8852 #ifdef CONFIG_CPUMASK_OFFSTACK
8853 alloc_size
+= num_possible_cpus() * cpumask_size();
8856 * As sched_init() is called before page_alloc is setup,
8857 * we use alloc_bootmem().
8860 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8862 #ifdef CONFIG_FAIR_GROUP_SCHED
8863 init_task_group
.se
= (struct sched_entity
**)ptr
;
8864 ptr
+= nr_cpu_ids
* sizeof(void **);
8866 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8867 ptr
+= nr_cpu_ids
* sizeof(void **);
8869 #ifdef CONFIG_USER_SCHED
8870 root_task_group
.se
= (struct sched_entity
**)ptr
;
8871 ptr
+= nr_cpu_ids
* sizeof(void **);
8873 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8874 ptr
+= nr_cpu_ids
* sizeof(void **);
8875 #endif /* CONFIG_USER_SCHED */
8876 #endif /* CONFIG_FAIR_GROUP_SCHED */
8877 #ifdef CONFIG_RT_GROUP_SCHED
8878 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8879 ptr
+= nr_cpu_ids
* sizeof(void **);
8881 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8882 ptr
+= nr_cpu_ids
* sizeof(void **);
8884 #ifdef CONFIG_USER_SCHED
8885 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8886 ptr
+= nr_cpu_ids
* sizeof(void **);
8888 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8889 ptr
+= nr_cpu_ids
* sizeof(void **);
8890 #endif /* CONFIG_USER_SCHED */
8891 #endif /* CONFIG_RT_GROUP_SCHED */
8892 #ifdef CONFIG_CPUMASK_OFFSTACK
8893 for_each_possible_cpu(i
) {
8894 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8895 ptr
+= cpumask_size();
8897 #endif /* CONFIG_CPUMASK_OFFSTACK */
8901 init_defrootdomain();
8904 init_rt_bandwidth(&def_rt_bandwidth
,
8905 global_rt_period(), global_rt_runtime());
8907 #ifdef CONFIG_RT_GROUP_SCHED
8908 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8909 global_rt_period(), global_rt_runtime());
8910 #ifdef CONFIG_USER_SCHED
8911 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8912 global_rt_period(), RUNTIME_INF
);
8913 #endif /* CONFIG_USER_SCHED */
8914 #endif /* CONFIG_RT_GROUP_SCHED */
8916 #ifdef CONFIG_GROUP_SCHED
8917 list_add(&init_task_group
.list
, &task_groups
);
8918 INIT_LIST_HEAD(&init_task_group
.children
);
8920 #ifdef CONFIG_USER_SCHED
8921 INIT_LIST_HEAD(&root_task_group
.children
);
8922 init_task_group
.parent
= &root_task_group
;
8923 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8924 #endif /* CONFIG_USER_SCHED */
8925 #endif /* CONFIG_GROUP_SCHED */
8927 for_each_possible_cpu(i
) {
8931 spin_lock_init(&rq
->lock
);
8933 init_cfs_rq(&rq
->cfs
, rq
);
8934 init_rt_rq(&rq
->rt
, rq
);
8935 #ifdef CONFIG_FAIR_GROUP_SCHED
8936 init_task_group
.shares
= init_task_group_load
;
8937 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8938 #ifdef CONFIG_CGROUP_SCHED
8940 * How much cpu bandwidth does init_task_group get?
8942 * In case of task-groups formed thr' the cgroup filesystem, it
8943 * gets 100% of the cpu resources in the system. This overall
8944 * system cpu resource is divided among the tasks of
8945 * init_task_group and its child task-groups in a fair manner,
8946 * based on each entity's (task or task-group's) weight
8947 * (se->load.weight).
8949 * In other words, if init_task_group has 10 tasks of weight
8950 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8951 * then A0's share of the cpu resource is:
8953 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8955 * We achieve this by letting init_task_group's tasks sit
8956 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8958 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8959 #elif defined CONFIG_USER_SCHED
8960 root_task_group
.shares
= NICE_0_LOAD
;
8961 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8963 * In case of task-groups formed thr' the user id of tasks,
8964 * init_task_group represents tasks belonging to root user.
8965 * Hence it forms a sibling of all subsequent groups formed.
8966 * In this case, init_task_group gets only a fraction of overall
8967 * system cpu resource, based on the weight assigned to root
8968 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8969 * by letting tasks of init_task_group sit in a separate cfs_rq
8970 * (init_cfs_rq) and having one entity represent this group of
8971 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8973 init_tg_cfs_entry(&init_task_group
,
8974 &per_cpu(init_cfs_rq
, i
),
8975 &per_cpu(init_sched_entity
, i
), i
, 1,
8976 root_task_group
.se
[i
]);
8979 #endif /* CONFIG_FAIR_GROUP_SCHED */
8981 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8982 #ifdef CONFIG_RT_GROUP_SCHED
8983 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8984 #ifdef CONFIG_CGROUP_SCHED
8985 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8986 #elif defined CONFIG_USER_SCHED
8987 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8988 init_tg_rt_entry(&init_task_group
,
8989 &per_cpu(init_rt_rq
, i
),
8990 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8991 root_task_group
.rt_se
[i
]);
8995 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8996 rq
->cpu_load
[j
] = 0;
9000 rq
->active_balance
= 0;
9001 rq
->next_balance
= jiffies
;
9005 rq
->migration_thread
= NULL
;
9006 INIT_LIST_HEAD(&rq
->migration_queue
);
9007 rq_attach_root(rq
, &def_root_domain
);
9010 atomic_set(&rq
->nr_iowait
, 0);
9013 set_load_weight(&init_task
);
9015 #ifdef CONFIG_PREEMPT_NOTIFIERS
9016 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9020 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9023 #ifdef CONFIG_RT_MUTEXES
9024 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9028 * The boot idle thread does lazy MMU switching as well:
9030 atomic_inc(&init_mm
.mm_count
);
9031 enter_lazy_tlb(&init_mm
, current
);
9034 * Make us the idle thread. Technically, schedule() should not be
9035 * called from this thread, however somewhere below it might be,
9036 * but because we are the idle thread, we just pick up running again
9037 * when this runqueue becomes "idle".
9039 init_idle(current
, smp_processor_id());
9041 * During early bootup we pretend to be a normal task:
9043 current
->sched_class
= &fair_sched_class
;
9045 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9046 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9049 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9051 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9054 scheduler_running
= 1;
9057 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9058 void __might_sleep(char *file
, int line
)
9061 static unsigned long prev_jiffy
; /* ratelimiting */
9063 if ((!in_atomic() && !irqs_disabled()) ||
9064 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9066 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9068 prev_jiffy
= jiffies
;
9071 "BUG: sleeping function called from invalid context at %s:%d\n",
9074 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9075 in_atomic(), irqs_disabled(),
9076 current
->pid
, current
->comm
);
9078 debug_show_held_locks(current
);
9079 if (irqs_disabled())
9080 print_irqtrace_events(current
);
9084 EXPORT_SYMBOL(__might_sleep
);
9087 #ifdef CONFIG_MAGIC_SYSRQ
9088 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9092 update_rq_clock(rq
);
9093 on_rq
= p
->se
.on_rq
;
9095 deactivate_task(rq
, p
, 0);
9096 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9098 activate_task(rq
, p
, 0);
9099 resched_task(rq
->curr
);
9103 void normalize_rt_tasks(void)
9105 struct task_struct
*g
, *p
;
9106 unsigned long flags
;
9109 read_lock_irqsave(&tasklist_lock
, flags
);
9110 do_each_thread(g
, p
) {
9112 * Only normalize user tasks:
9117 p
->se
.exec_start
= 0;
9118 #ifdef CONFIG_SCHEDSTATS
9119 p
->se
.wait_start
= 0;
9120 p
->se
.sleep_start
= 0;
9121 p
->se
.block_start
= 0;
9126 * Renice negative nice level userspace
9129 if (TASK_NICE(p
) < 0 && p
->mm
)
9130 set_user_nice(p
, 0);
9134 spin_lock(&p
->pi_lock
);
9135 rq
= __task_rq_lock(p
);
9137 normalize_task(rq
, p
);
9139 __task_rq_unlock(rq
);
9140 spin_unlock(&p
->pi_lock
);
9141 } while_each_thread(g
, p
);
9143 read_unlock_irqrestore(&tasklist_lock
, flags
);
9146 #endif /* CONFIG_MAGIC_SYSRQ */
9150 * These functions are only useful for the IA64 MCA handling.
9152 * They can only be called when the whole system has been
9153 * stopped - every CPU needs to be quiescent, and no scheduling
9154 * activity can take place. Using them for anything else would
9155 * be a serious bug, and as a result, they aren't even visible
9156 * under any other configuration.
9160 * curr_task - return the current task for a given cpu.
9161 * @cpu: the processor in question.
9163 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9165 struct task_struct
*curr_task(int cpu
)
9167 return cpu_curr(cpu
);
9171 * set_curr_task - set the current task for a given cpu.
9172 * @cpu: the processor in question.
9173 * @p: the task pointer to set.
9175 * Description: This function must only be used when non-maskable interrupts
9176 * are serviced on a separate stack. It allows the architecture to switch the
9177 * notion of the current task on a cpu in a non-blocking manner. This function
9178 * must be called with all CPU's synchronized, and interrupts disabled, the
9179 * and caller must save the original value of the current task (see
9180 * curr_task() above) and restore that value before reenabling interrupts and
9181 * re-starting the system.
9183 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9185 void set_curr_task(int cpu
, struct task_struct
*p
)
9192 #ifdef CONFIG_FAIR_GROUP_SCHED
9193 static void free_fair_sched_group(struct task_group
*tg
)
9197 for_each_possible_cpu(i
) {
9199 kfree(tg
->cfs_rq
[i
]);
9209 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9211 struct cfs_rq
*cfs_rq
;
9212 struct sched_entity
*se
;
9216 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9219 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9223 tg
->shares
= NICE_0_LOAD
;
9225 for_each_possible_cpu(i
) {
9228 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9229 GFP_KERNEL
, cpu_to_node(i
));
9233 se
= kzalloc_node(sizeof(struct sched_entity
),
9234 GFP_KERNEL
, cpu_to_node(i
));
9238 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9247 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9249 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9250 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9253 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9255 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9257 #else /* !CONFG_FAIR_GROUP_SCHED */
9258 static inline void free_fair_sched_group(struct task_group
*tg
)
9263 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9268 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9272 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9275 #endif /* CONFIG_FAIR_GROUP_SCHED */
9277 #ifdef CONFIG_RT_GROUP_SCHED
9278 static void free_rt_sched_group(struct task_group
*tg
)
9282 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9284 for_each_possible_cpu(i
) {
9286 kfree(tg
->rt_rq
[i
]);
9288 kfree(tg
->rt_se
[i
]);
9296 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9298 struct rt_rq
*rt_rq
;
9299 struct sched_rt_entity
*rt_se
;
9303 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9306 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9310 init_rt_bandwidth(&tg
->rt_bandwidth
,
9311 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9313 for_each_possible_cpu(i
) {
9316 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9317 GFP_KERNEL
, cpu_to_node(i
));
9321 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9322 GFP_KERNEL
, cpu_to_node(i
));
9326 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9335 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9337 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9338 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9341 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9343 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9345 #else /* !CONFIG_RT_GROUP_SCHED */
9346 static inline void free_rt_sched_group(struct task_group
*tg
)
9351 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9356 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9360 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9363 #endif /* CONFIG_RT_GROUP_SCHED */
9365 #ifdef CONFIG_GROUP_SCHED
9366 static void free_sched_group(struct task_group
*tg
)
9368 free_fair_sched_group(tg
);
9369 free_rt_sched_group(tg
);
9373 /* allocate runqueue etc for a new task group */
9374 struct task_group
*sched_create_group(struct task_group
*parent
)
9376 struct task_group
*tg
;
9377 unsigned long flags
;
9380 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9382 return ERR_PTR(-ENOMEM
);
9384 if (!alloc_fair_sched_group(tg
, parent
))
9387 if (!alloc_rt_sched_group(tg
, parent
))
9390 spin_lock_irqsave(&task_group_lock
, flags
);
9391 for_each_possible_cpu(i
) {
9392 register_fair_sched_group(tg
, i
);
9393 register_rt_sched_group(tg
, i
);
9395 list_add_rcu(&tg
->list
, &task_groups
);
9397 WARN_ON(!parent
); /* root should already exist */
9399 tg
->parent
= parent
;
9400 INIT_LIST_HEAD(&tg
->children
);
9401 list_add_rcu(&tg
->siblings
, &parent
->children
);
9402 spin_unlock_irqrestore(&task_group_lock
, flags
);
9407 free_sched_group(tg
);
9408 return ERR_PTR(-ENOMEM
);
9411 /* rcu callback to free various structures associated with a task group */
9412 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9414 /* now it should be safe to free those cfs_rqs */
9415 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9418 /* Destroy runqueue etc associated with a task group */
9419 void sched_destroy_group(struct task_group
*tg
)
9421 unsigned long flags
;
9424 spin_lock_irqsave(&task_group_lock
, flags
);
9425 for_each_possible_cpu(i
) {
9426 unregister_fair_sched_group(tg
, i
);
9427 unregister_rt_sched_group(tg
, i
);
9429 list_del_rcu(&tg
->list
);
9430 list_del_rcu(&tg
->siblings
);
9431 spin_unlock_irqrestore(&task_group_lock
, flags
);
9433 /* wait for possible concurrent references to cfs_rqs complete */
9434 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9437 /* change task's runqueue when it moves between groups.
9438 * The caller of this function should have put the task in its new group
9439 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9440 * reflect its new group.
9442 void sched_move_task(struct task_struct
*tsk
)
9445 unsigned long flags
;
9448 rq
= task_rq_lock(tsk
, &flags
);
9450 update_rq_clock(rq
);
9452 running
= task_current(rq
, tsk
);
9453 on_rq
= tsk
->se
.on_rq
;
9456 dequeue_task(rq
, tsk
, 0);
9457 if (unlikely(running
))
9458 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9460 set_task_rq(tsk
, task_cpu(tsk
));
9462 #ifdef CONFIG_FAIR_GROUP_SCHED
9463 if (tsk
->sched_class
->moved_group
)
9464 tsk
->sched_class
->moved_group(tsk
);
9467 if (unlikely(running
))
9468 tsk
->sched_class
->set_curr_task(rq
);
9470 enqueue_task(rq
, tsk
, 0);
9472 task_rq_unlock(rq
, &flags
);
9474 #endif /* CONFIG_GROUP_SCHED */
9476 #ifdef CONFIG_FAIR_GROUP_SCHED
9477 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9479 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9484 dequeue_entity(cfs_rq
, se
, 0);
9486 se
->load
.weight
= shares
;
9487 se
->load
.inv_weight
= 0;
9490 enqueue_entity(cfs_rq
, se
, 0);
9493 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9495 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9496 struct rq
*rq
= cfs_rq
->rq
;
9497 unsigned long flags
;
9499 spin_lock_irqsave(&rq
->lock
, flags
);
9500 __set_se_shares(se
, shares
);
9501 spin_unlock_irqrestore(&rq
->lock
, flags
);
9504 static DEFINE_MUTEX(shares_mutex
);
9506 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9509 unsigned long flags
;
9512 * We can't change the weight of the root cgroup.
9517 if (shares
< MIN_SHARES
)
9518 shares
= MIN_SHARES
;
9519 else if (shares
> MAX_SHARES
)
9520 shares
= MAX_SHARES
;
9522 mutex_lock(&shares_mutex
);
9523 if (tg
->shares
== shares
)
9526 spin_lock_irqsave(&task_group_lock
, flags
);
9527 for_each_possible_cpu(i
)
9528 unregister_fair_sched_group(tg
, i
);
9529 list_del_rcu(&tg
->siblings
);
9530 spin_unlock_irqrestore(&task_group_lock
, flags
);
9532 /* wait for any ongoing reference to this group to finish */
9533 synchronize_sched();
9536 * Now we are free to modify the group's share on each cpu
9537 * w/o tripping rebalance_share or load_balance_fair.
9539 tg
->shares
= shares
;
9540 for_each_possible_cpu(i
) {
9544 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9545 set_se_shares(tg
->se
[i
], shares
);
9549 * Enable load balance activity on this group, by inserting it back on
9550 * each cpu's rq->leaf_cfs_rq_list.
9552 spin_lock_irqsave(&task_group_lock
, flags
);
9553 for_each_possible_cpu(i
)
9554 register_fair_sched_group(tg
, i
);
9555 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9556 spin_unlock_irqrestore(&task_group_lock
, flags
);
9558 mutex_unlock(&shares_mutex
);
9562 unsigned long sched_group_shares(struct task_group
*tg
)
9568 #ifdef CONFIG_RT_GROUP_SCHED
9570 * Ensure that the real time constraints are schedulable.
9572 static DEFINE_MUTEX(rt_constraints_mutex
);
9574 static unsigned long to_ratio(u64 period
, u64 runtime
)
9576 if (runtime
== RUNTIME_INF
)
9579 return div64_u64(runtime
<< 20, period
);
9582 /* Must be called with tasklist_lock held */
9583 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9585 struct task_struct
*g
, *p
;
9587 do_each_thread(g
, p
) {
9588 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9590 } while_each_thread(g
, p
);
9595 struct rt_schedulable_data
{
9596 struct task_group
*tg
;
9601 static int tg_schedulable(struct task_group
*tg
, void *data
)
9603 struct rt_schedulable_data
*d
= data
;
9604 struct task_group
*child
;
9605 unsigned long total
, sum
= 0;
9606 u64 period
, runtime
;
9608 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9609 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9612 period
= d
->rt_period
;
9613 runtime
= d
->rt_runtime
;
9616 #ifdef CONFIG_USER_SCHED
9617 if (tg
== &root_task_group
) {
9618 period
= global_rt_period();
9619 runtime
= global_rt_runtime();
9624 * Cannot have more runtime than the period.
9626 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9630 * Ensure we don't starve existing RT tasks.
9632 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9635 total
= to_ratio(period
, runtime
);
9638 * Nobody can have more than the global setting allows.
9640 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9644 * The sum of our children's runtime should not exceed our own.
9646 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9647 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9648 runtime
= child
->rt_bandwidth
.rt_runtime
;
9650 if (child
== d
->tg
) {
9651 period
= d
->rt_period
;
9652 runtime
= d
->rt_runtime
;
9655 sum
+= to_ratio(period
, runtime
);
9664 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9666 struct rt_schedulable_data data
= {
9668 .rt_period
= period
,
9669 .rt_runtime
= runtime
,
9672 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9675 static int tg_set_bandwidth(struct task_group
*tg
,
9676 u64 rt_period
, u64 rt_runtime
)
9680 mutex_lock(&rt_constraints_mutex
);
9681 read_lock(&tasklist_lock
);
9682 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9686 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9687 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9688 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9690 for_each_possible_cpu(i
) {
9691 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9693 spin_lock(&rt_rq
->rt_runtime_lock
);
9694 rt_rq
->rt_runtime
= rt_runtime
;
9695 spin_unlock(&rt_rq
->rt_runtime_lock
);
9697 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9699 read_unlock(&tasklist_lock
);
9700 mutex_unlock(&rt_constraints_mutex
);
9705 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9707 u64 rt_runtime
, rt_period
;
9709 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9710 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9711 if (rt_runtime_us
< 0)
9712 rt_runtime
= RUNTIME_INF
;
9714 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9717 long sched_group_rt_runtime(struct task_group
*tg
)
9721 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9724 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9725 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9726 return rt_runtime_us
;
9729 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9731 u64 rt_runtime
, rt_period
;
9733 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9734 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9739 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9742 long sched_group_rt_period(struct task_group
*tg
)
9746 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9747 do_div(rt_period_us
, NSEC_PER_USEC
);
9748 return rt_period_us
;
9751 static int sched_rt_global_constraints(void)
9753 u64 runtime
, period
;
9756 if (sysctl_sched_rt_period
<= 0)
9759 runtime
= global_rt_runtime();
9760 period
= global_rt_period();
9763 * Sanity check on the sysctl variables.
9765 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9768 mutex_lock(&rt_constraints_mutex
);
9769 read_lock(&tasklist_lock
);
9770 ret
= __rt_schedulable(NULL
, 0, 0);
9771 read_unlock(&tasklist_lock
);
9772 mutex_unlock(&rt_constraints_mutex
);
9777 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9779 /* Don't accept realtime tasks when there is no way for them to run */
9780 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9786 #else /* !CONFIG_RT_GROUP_SCHED */
9787 static int sched_rt_global_constraints(void)
9789 unsigned long flags
;
9792 if (sysctl_sched_rt_period
<= 0)
9795 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9796 for_each_possible_cpu(i
) {
9797 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9799 spin_lock(&rt_rq
->rt_runtime_lock
);
9800 rt_rq
->rt_runtime
= global_rt_runtime();
9801 spin_unlock(&rt_rq
->rt_runtime_lock
);
9803 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9807 #endif /* CONFIG_RT_GROUP_SCHED */
9809 int sched_rt_handler(struct ctl_table
*table
, int write
,
9810 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9814 int old_period
, old_runtime
;
9815 static DEFINE_MUTEX(mutex
);
9818 old_period
= sysctl_sched_rt_period
;
9819 old_runtime
= sysctl_sched_rt_runtime
;
9821 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9823 if (!ret
&& write
) {
9824 ret
= sched_rt_global_constraints();
9826 sysctl_sched_rt_period
= old_period
;
9827 sysctl_sched_rt_runtime
= old_runtime
;
9829 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9830 def_rt_bandwidth
.rt_period
=
9831 ns_to_ktime(global_rt_period());
9834 mutex_unlock(&mutex
);
9839 #ifdef CONFIG_CGROUP_SCHED
9841 /* return corresponding task_group object of a cgroup */
9842 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9844 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9845 struct task_group
, css
);
9848 static struct cgroup_subsys_state
*
9849 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9851 struct task_group
*tg
, *parent
;
9853 if (!cgrp
->parent
) {
9854 /* This is early initialization for the top cgroup */
9855 return &init_task_group
.css
;
9858 parent
= cgroup_tg(cgrp
->parent
);
9859 tg
= sched_create_group(parent
);
9861 return ERR_PTR(-ENOMEM
);
9867 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9869 struct task_group
*tg
= cgroup_tg(cgrp
);
9871 sched_destroy_group(tg
);
9875 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9876 struct task_struct
*tsk
)
9878 #ifdef CONFIG_RT_GROUP_SCHED
9879 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9882 /* We don't support RT-tasks being in separate groups */
9883 if (tsk
->sched_class
!= &fair_sched_class
)
9891 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9892 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9894 sched_move_task(tsk
);
9897 #ifdef CONFIG_FAIR_GROUP_SCHED
9898 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9901 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9904 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9906 struct task_group
*tg
= cgroup_tg(cgrp
);
9908 return (u64
) tg
->shares
;
9910 #endif /* CONFIG_FAIR_GROUP_SCHED */
9912 #ifdef CONFIG_RT_GROUP_SCHED
9913 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9916 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9919 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9921 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9924 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9927 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9930 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9932 return sched_group_rt_period(cgroup_tg(cgrp
));
9934 #endif /* CONFIG_RT_GROUP_SCHED */
9936 static struct cftype cpu_files
[] = {
9937 #ifdef CONFIG_FAIR_GROUP_SCHED
9940 .read_u64
= cpu_shares_read_u64
,
9941 .write_u64
= cpu_shares_write_u64
,
9944 #ifdef CONFIG_RT_GROUP_SCHED
9946 .name
= "rt_runtime_us",
9947 .read_s64
= cpu_rt_runtime_read
,
9948 .write_s64
= cpu_rt_runtime_write
,
9951 .name
= "rt_period_us",
9952 .read_u64
= cpu_rt_period_read_uint
,
9953 .write_u64
= cpu_rt_period_write_uint
,
9958 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9960 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9963 struct cgroup_subsys cpu_cgroup_subsys
= {
9965 .create
= cpu_cgroup_create
,
9966 .destroy
= cpu_cgroup_destroy
,
9967 .can_attach
= cpu_cgroup_can_attach
,
9968 .attach
= cpu_cgroup_attach
,
9969 .populate
= cpu_cgroup_populate
,
9970 .subsys_id
= cpu_cgroup_subsys_id
,
9974 #endif /* CONFIG_CGROUP_SCHED */
9976 #ifdef CONFIG_CGROUP_CPUACCT
9979 * CPU accounting code for task groups.
9981 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9982 * (balbir@in.ibm.com).
9985 /* track cpu usage of a group of tasks and its child groups */
9987 struct cgroup_subsys_state css
;
9988 /* cpuusage holds pointer to a u64-type object on every cpu */
9990 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9991 struct cpuacct
*parent
;
9994 struct cgroup_subsys cpuacct_subsys
;
9996 /* return cpu accounting group corresponding to this container */
9997 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9999 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10000 struct cpuacct
, css
);
10003 /* return cpu accounting group to which this task belongs */
10004 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10006 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10007 struct cpuacct
, css
);
10010 /* create a new cpu accounting group */
10011 static struct cgroup_subsys_state
*cpuacct_create(
10012 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10014 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10020 ca
->cpuusage
= alloc_percpu(u64
);
10024 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10025 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10026 goto out_free_counters
;
10029 ca
->parent
= cgroup_ca(cgrp
->parent
);
10035 percpu_counter_destroy(&ca
->cpustat
[i
]);
10036 free_percpu(ca
->cpuusage
);
10040 return ERR_PTR(-ENOMEM
);
10043 /* destroy an existing cpu accounting group */
10045 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10047 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10050 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10051 percpu_counter_destroy(&ca
->cpustat
[i
]);
10052 free_percpu(ca
->cpuusage
);
10056 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10058 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10061 #ifndef CONFIG_64BIT
10063 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10065 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10067 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10075 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10077 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10079 #ifndef CONFIG_64BIT
10081 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10083 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10085 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10091 /* return total cpu usage (in nanoseconds) of a group */
10092 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10094 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10095 u64 totalcpuusage
= 0;
10098 for_each_present_cpu(i
)
10099 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10101 return totalcpuusage
;
10104 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10107 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10116 for_each_present_cpu(i
)
10117 cpuacct_cpuusage_write(ca
, i
, 0);
10123 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10124 struct seq_file
*m
)
10126 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10130 for_each_present_cpu(i
) {
10131 percpu
= cpuacct_cpuusage_read(ca
, i
);
10132 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10134 seq_printf(m
, "\n");
10138 static const char *cpuacct_stat_desc
[] = {
10139 [CPUACCT_STAT_USER
] = "user",
10140 [CPUACCT_STAT_SYSTEM
] = "system",
10143 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10144 struct cgroup_map_cb
*cb
)
10146 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10149 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10150 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10151 val
= cputime64_to_clock_t(val
);
10152 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10157 static struct cftype files
[] = {
10160 .read_u64
= cpuusage_read
,
10161 .write_u64
= cpuusage_write
,
10164 .name
= "usage_percpu",
10165 .read_seq_string
= cpuacct_percpu_seq_read
,
10169 .read_map
= cpuacct_stats_show
,
10173 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10175 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10179 * charge this task's execution time to its accounting group.
10181 * called with rq->lock held.
10183 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10185 struct cpuacct
*ca
;
10188 if (unlikely(!cpuacct_subsys
.active
))
10191 cpu
= task_cpu(tsk
);
10197 for (; ca
; ca
= ca
->parent
) {
10198 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10199 *cpuusage
+= cputime
;
10206 * Charge the system/user time to the task's accounting group.
10208 static void cpuacct_update_stats(struct task_struct
*tsk
,
10209 enum cpuacct_stat_index idx
, cputime_t val
)
10211 struct cpuacct
*ca
;
10213 if (unlikely(!cpuacct_subsys
.active
))
10220 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10226 struct cgroup_subsys cpuacct_subsys
= {
10228 .create
= cpuacct_create
,
10229 .destroy
= cpuacct_destroy
,
10230 .populate
= cpuacct_populate
,
10231 .subsys_id
= cpuacct_subsys_id
,
10233 #endif /* CONFIG_CGROUP_CPUACCT */