4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/bootmem.h>
73 #include <linux/debugfs.h>
74 #include <linux/ctype.h>
75 #include <linux/ftrace.h>
76 #include <trace/sched.h>
79 #include <asm/irq_regs.h>
81 #include "sched_cpupri.h"
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 DEFINE_TRACE(sched_wait_task
);
123 DEFINE_TRACE(sched_wakeup
);
124 DEFINE_TRACE(sched_wakeup_new
);
125 DEFINE_TRACE(sched_switch
);
126 DEFINE_TRACE(sched_migrate_task
);
130 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
133 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
134 * Since cpu_power is a 'constant', we can use a reciprocal divide.
136 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
138 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
142 * Each time a sched group cpu_power is changed,
143 * we must compute its reciprocal value
145 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
147 sg
->__cpu_power
+= val
;
148 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
152 static inline int rt_policy(int policy
)
154 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
159 static inline int task_has_rt_policy(struct task_struct
*p
)
161 return rt_policy(p
->policy
);
165 * This is the priority-queue data structure of the RT scheduling class:
167 struct rt_prio_array
{
168 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
169 struct list_head queue
[MAX_RT_PRIO
];
172 struct rt_bandwidth
{
173 /* nests inside the rq lock: */
174 spinlock_t rt_runtime_lock
;
177 struct hrtimer rt_period_timer
;
180 static struct rt_bandwidth def_rt_bandwidth
;
182 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
184 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
186 struct rt_bandwidth
*rt_b
=
187 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
193 now
= hrtimer_cb_get_time(timer
);
194 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
199 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
202 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
206 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
208 rt_b
->rt_period
= ns_to_ktime(period
);
209 rt_b
->rt_runtime
= runtime
;
211 spin_lock_init(&rt_b
->rt_runtime_lock
);
213 hrtimer_init(&rt_b
->rt_period_timer
,
214 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
215 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
218 static inline int rt_bandwidth_enabled(void)
220 return sysctl_sched_rt_runtime
>= 0;
223 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
227 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
230 if (hrtimer_active(&rt_b
->rt_period_timer
))
233 spin_lock(&rt_b
->rt_runtime_lock
);
238 if (hrtimer_active(&rt_b
->rt_period_timer
))
241 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
242 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
244 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
245 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
246 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
247 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
248 HRTIMER_MODE_ABS
, 0);
250 spin_unlock(&rt_b
->rt_runtime_lock
);
253 #ifdef CONFIG_RT_GROUP_SCHED
254 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
256 hrtimer_cancel(&rt_b
->rt_period_timer
);
261 * sched_domains_mutex serializes calls to arch_init_sched_domains,
262 * detach_destroy_domains and partition_sched_domains.
264 static DEFINE_MUTEX(sched_domains_mutex
);
266 #ifdef CONFIG_GROUP_SCHED
268 #include <linux/cgroup.h>
272 static LIST_HEAD(task_groups
);
274 /* task group related information */
276 #ifdef CONFIG_CGROUP_SCHED
277 struct cgroup_subsys_state css
;
280 #ifdef CONFIG_USER_SCHED
284 #ifdef CONFIG_FAIR_GROUP_SCHED
285 /* schedulable entities of this group on each cpu */
286 struct sched_entity
**se
;
287 /* runqueue "owned" by this group on each cpu */
288 struct cfs_rq
**cfs_rq
;
289 unsigned long shares
;
292 #ifdef CONFIG_RT_GROUP_SCHED
293 struct sched_rt_entity
**rt_se
;
294 struct rt_rq
**rt_rq
;
296 struct rt_bandwidth rt_bandwidth
;
300 struct list_head list
;
302 struct task_group
*parent
;
303 struct list_head siblings
;
304 struct list_head children
;
307 #ifdef CONFIG_USER_SCHED
309 /* Helper function to pass uid information to create_sched_user() */
310 void set_tg_uid(struct user_struct
*user
)
312 user
->tg
->uid
= user
->uid
;
317 * Every UID task group (including init_task_group aka UID-0) will
318 * be a child to this group.
320 struct task_group root_task_group
;
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 /* Default task group's sched entity on each cpu */
324 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
325 /* Default task group's cfs_rq on each cpu */
326 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
327 #endif /* CONFIG_FAIR_GROUP_SCHED */
329 #ifdef CONFIG_RT_GROUP_SCHED
330 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
331 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
332 #endif /* CONFIG_RT_GROUP_SCHED */
333 #else /* !CONFIG_USER_SCHED */
334 #define root_task_group init_task_group
335 #endif /* CONFIG_USER_SCHED */
337 /* task_group_lock serializes add/remove of task groups and also changes to
338 * a task group's cpu shares.
340 static DEFINE_SPINLOCK(task_group_lock
);
343 static int root_task_group_empty(void)
345 return list_empty(&root_task_group
.children
);
349 #ifdef CONFIG_FAIR_GROUP_SCHED
350 #ifdef CONFIG_USER_SCHED
351 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
352 #else /* !CONFIG_USER_SCHED */
353 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
354 #endif /* CONFIG_USER_SCHED */
357 * A weight of 0 or 1 can cause arithmetics problems.
358 * A weight of a cfs_rq is the sum of weights of which entities
359 * are queued on this cfs_rq, so a weight of a entity should not be
360 * too large, so as the shares value of a task group.
361 * (The default weight is 1024 - so there's no practical
362 * limitation from this.)
365 #define MAX_SHARES (1UL << 18)
367 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
370 /* Default task group.
371 * Every task in system belong to this group at bootup.
373 struct task_group init_task_group
;
375 /* return group to which a task belongs */
376 static inline struct task_group
*task_group(struct task_struct
*p
)
378 struct task_group
*tg
;
380 #ifdef CONFIG_USER_SCHED
382 tg
= __task_cred(p
)->user
->tg
;
384 #elif defined(CONFIG_CGROUP_SCHED)
385 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
386 struct task_group
, css
);
388 tg
= &init_task_group
;
393 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
394 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
396 #ifdef CONFIG_FAIR_GROUP_SCHED
397 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
398 p
->se
.parent
= task_group(p
)->se
[cpu
];
401 #ifdef CONFIG_RT_GROUP_SCHED
402 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
403 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
410 static int root_task_group_empty(void)
416 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
417 static inline struct task_group
*task_group(struct task_struct
*p
)
422 #endif /* CONFIG_GROUP_SCHED */
424 /* CFS-related fields in a runqueue */
426 struct load_weight load
;
427 unsigned long nr_running
;
432 struct rb_root tasks_timeline
;
433 struct rb_node
*rb_leftmost
;
435 struct list_head tasks
;
436 struct list_head
*balance_iterator
;
439 * 'curr' points to currently running entity on this cfs_rq.
440 * It is set to NULL otherwise (i.e when none are currently running).
442 struct sched_entity
*curr
, *next
, *last
;
444 unsigned int nr_spread_over
;
446 #ifdef CONFIG_FAIR_GROUP_SCHED
447 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
450 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
451 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
452 * (like users, containers etc.)
454 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
455 * list is used during load balance.
457 struct list_head leaf_cfs_rq_list
;
458 struct task_group
*tg
; /* group that "owns" this runqueue */
462 * the part of load.weight contributed by tasks
464 unsigned long task_weight
;
467 * h_load = weight * f(tg)
469 * Where f(tg) is the recursive weight fraction assigned to
472 unsigned long h_load
;
475 * this cpu's part of tg->shares
477 unsigned long shares
;
480 * load.weight at the time we set shares
482 unsigned long rq_weight
;
487 /* Real-Time classes' related field in a runqueue: */
489 struct rt_prio_array active
;
490 unsigned long rt_nr_running
;
491 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
493 int curr
; /* highest queued rt task prio */
495 int next
; /* next highest */
500 unsigned long rt_nr_migratory
;
502 struct plist_head pushable_tasks
;
507 /* Nests inside the rq lock: */
508 spinlock_t rt_runtime_lock
;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 unsigned long rt_nr_boosted
;
514 struct list_head leaf_rt_rq_list
;
515 struct task_group
*tg
;
516 struct sched_rt_entity
*rt_se
;
523 * We add the notion of a root-domain which will be used to define per-domain
524 * variables. Each exclusive cpuset essentially defines an island domain by
525 * fully partitioning the member cpus from any other cpuset. Whenever a new
526 * exclusive cpuset is created, we also create and attach a new root-domain
533 cpumask_var_t online
;
536 * The "RT overload" flag: it gets set if a CPU has more than
537 * one runnable RT task.
539 cpumask_var_t rto_mask
;
542 struct cpupri cpupri
;
544 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
546 * Preferred wake up cpu nominated by sched_mc balance that will be
547 * used when most cpus are idle in the system indicating overall very
548 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
550 unsigned int sched_mc_preferred_wakeup_cpu
;
555 * By default the system creates a single root-domain with all cpus as
556 * members (mimicking the global state we have today).
558 static struct root_domain def_root_domain
;
563 * This is the main, per-CPU runqueue data structure.
565 * Locking rule: those places that want to lock multiple runqueues
566 * (such as the load balancing or the thread migration code), lock
567 * acquire operations must be ordered by ascending &runqueue.
574 * nr_running and cpu_load should be in the same cacheline because
575 * remote CPUs use both these fields when doing load calculation.
577 unsigned long nr_running
;
578 #define CPU_LOAD_IDX_MAX 5
579 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
581 unsigned long last_tick_seen
;
582 unsigned char in_nohz_recently
;
584 /* capture load from *all* tasks on this cpu: */
585 struct load_weight load
;
586 unsigned long nr_load_updates
;
588 u64 nr_migrations_in
;
593 #ifdef CONFIG_FAIR_GROUP_SCHED
594 /* list of leaf cfs_rq on this cpu: */
595 struct list_head leaf_cfs_rq_list
;
597 #ifdef CONFIG_RT_GROUP_SCHED
598 struct list_head leaf_rt_rq_list
;
602 * This is part of a global counter where only the total sum
603 * over all CPUs matters. A task can increase this counter on
604 * one CPU and if it got migrated afterwards it may decrease
605 * it on another CPU. Always updated under the runqueue lock:
607 unsigned long nr_uninterruptible
;
609 struct task_struct
*curr
, *idle
;
610 unsigned long next_balance
;
611 struct mm_struct
*prev_mm
;
618 struct root_domain
*rd
;
619 struct sched_domain
*sd
;
621 unsigned char idle_at_tick
;
622 /* For active balancing */
625 /* cpu of this runqueue: */
629 unsigned long avg_load_per_task
;
631 struct task_struct
*migration_thread
;
632 struct list_head migration_queue
;
635 #ifdef CONFIG_SCHED_HRTICK
637 int hrtick_csd_pending
;
638 struct call_single_data hrtick_csd
;
640 struct hrtimer hrtick_timer
;
643 #ifdef CONFIG_SCHEDSTATS
645 struct sched_info rq_sched_info
;
646 unsigned long long rq_cpu_time
;
647 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
649 /* sys_sched_yield() stats */
650 unsigned int yld_count
;
652 /* schedule() stats */
653 unsigned int sched_switch
;
654 unsigned int sched_count
;
655 unsigned int sched_goidle
;
657 /* try_to_wake_up() stats */
658 unsigned int ttwu_count
;
659 unsigned int ttwu_local
;
662 unsigned int bkl_count
;
666 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
668 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
670 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
673 static inline int cpu_of(struct rq
*rq
)
683 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
684 * See detach_destroy_domains: synchronize_sched for details.
686 * The domain tree of any CPU may only be accessed from within
687 * preempt-disabled sections.
689 #define for_each_domain(cpu, __sd) \
690 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
692 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
693 #define this_rq() (&__get_cpu_var(runqueues))
694 #define task_rq(p) cpu_rq(task_cpu(p))
695 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
697 inline void update_rq_clock(struct rq
*rq
)
699 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
703 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
705 #ifdef CONFIG_SCHED_DEBUG
706 # define const_debug __read_mostly
708 # define const_debug static const
714 * Returns true if the current cpu runqueue is locked.
715 * This interface allows printk to be called with the runqueue lock
716 * held and know whether or not it is OK to wake up the klogd.
718 int runqueue_is_locked(void)
721 struct rq
*rq
= cpu_rq(cpu
);
724 ret
= spin_is_locked(&rq
->lock
);
730 * Debugging: various feature bits
733 #define SCHED_FEAT(name, enabled) \
734 __SCHED_FEAT_##name ,
737 #include "sched_features.h"
742 #define SCHED_FEAT(name, enabled) \
743 (1UL << __SCHED_FEAT_##name) * enabled |
745 const_debug
unsigned int sysctl_sched_features
=
746 #include "sched_features.h"
751 #ifdef CONFIG_SCHED_DEBUG
752 #define SCHED_FEAT(name, enabled) \
755 static __read_mostly
char *sched_feat_names
[] = {
756 #include "sched_features.h"
762 static int sched_feat_show(struct seq_file
*m
, void *v
)
766 for (i
= 0; sched_feat_names
[i
]; i
++) {
767 if (!(sysctl_sched_features
& (1UL << i
)))
769 seq_printf(m
, "%s ", sched_feat_names
[i
]);
777 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
778 size_t cnt
, loff_t
*ppos
)
788 if (copy_from_user(&buf
, ubuf
, cnt
))
793 if (strncmp(buf
, "NO_", 3) == 0) {
798 for (i
= 0; sched_feat_names
[i
]; i
++) {
799 int len
= strlen(sched_feat_names
[i
]);
801 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
803 sysctl_sched_features
&= ~(1UL << i
);
805 sysctl_sched_features
|= (1UL << i
);
810 if (!sched_feat_names
[i
])
818 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
820 return single_open(filp
, sched_feat_show
, NULL
);
823 static struct file_operations sched_feat_fops
= {
824 .open
= sched_feat_open
,
825 .write
= sched_feat_write
,
828 .release
= single_release
,
831 static __init
int sched_init_debug(void)
833 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
838 late_initcall(sched_init_debug
);
842 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
845 * Number of tasks to iterate in a single balance run.
846 * Limited because this is done with IRQs disabled.
848 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
851 * ratelimit for updating the group shares.
854 unsigned int sysctl_sched_shares_ratelimit
= 250000;
857 * Inject some fuzzyness into changing the per-cpu group shares
858 * this avoids remote rq-locks at the expense of fairness.
861 unsigned int sysctl_sched_shares_thresh
= 4;
864 * period over which we measure -rt task cpu usage in us.
867 unsigned int sysctl_sched_rt_period
= 1000000;
869 static __read_mostly
int scheduler_running
;
872 * part of the period that we allow rt tasks to run in us.
875 int sysctl_sched_rt_runtime
= 950000;
877 static inline u64
global_rt_period(void)
879 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
882 static inline u64
global_rt_runtime(void)
884 if (sysctl_sched_rt_runtime
< 0)
887 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
890 #ifndef prepare_arch_switch
891 # define prepare_arch_switch(next) do { } while (0)
893 #ifndef finish_arch_switch
894 # define finish_arch_switch(prev) do { } while (0)
897 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
899 return rq
->curr
== p
;
902 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
903 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
905 return task_current(rq
, p
);
908 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
912 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
914 #ifdef CONFIG_DEBUG_SPINLOCK
915 /* this is a valid case when another task releases the spinlock */
916 rq
->lock
.owner
= current
;
919 * If we are tracking spinlock dependencies then we have to
920 * fix up the runqueue lock - which gets 'carried over' from
923 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
925 spin_unlock_irq(&rq
->lock
);
928 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
929 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
934 return task_current(rq
, p
);
938 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
942 * We can optimise this out completely for !SMP, because the
943 * SMP rebalancing from interrupt is the only thing that cares
948 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
949 spin_unlock_irq(&rq
->lock
);
951 spin_unlock(&rq
->lock
);
955 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
959 * After ->oncpu is cleared, the task can be moved to a different CPU.
960 * We must ensure this doesn't happen until the switch is completely
966 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
970 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
973 * __task_rq_lock - lock the runqueue a given task resides on.
974 * Must be called interrupts disabled.
976 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
980 struct rq
*rq
= task_rq(p
);
981 spin_lock(&rq
->lock
);
982 if (likely(rq
== task_rq(p
)))
984 spin_unlock(&rq
->lock
);
989 * task_rq_lock - lock the runqueue a given task resides on and disable
990 * interrupts. Note the ordering: we can safely lookup the task_rq without
991 * explicitly disabling preemption.
993 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
999 local_irq_save(*flags
);
1001 spin_lock(&rq
->lock
);
1002 if (likely(rq
== task_rq(p
)))
1004 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1008 void task_rq_unlock_wait(struct task_struct
*p
)
1010 struct rq
*rq
= task_rq(p
);
1012 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1013 spin_unlock_wait(&rq
->lock
);
1016 static void __task_rq_unlock(struct rq
*rq
)
1017 __releases(rq
->lock
)
1019 spin_unlock(&rq
->lock
);
1022 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1023 __releases(rq
->lock
)
1025 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1029 * this_rq_lock - lock this runqueue and disable interrupts.
1031 static struct rq
*this_rq_lock(void)
1032 __acquires(rq
->lock
)
1036 local_irq_disable();
1038 spin_lock(&rq
->lock
);
1043 #ifdef CONFIG_SCHED_HRTICK
1045 * Use HR-timers to deliver accurate preemption points.
1047 * Its all a bit involved since we cannot program an hrt while holding the
1048 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1051 * When we get rescheduled we reprogram the hrtick_timer outside of the
1057 * - enabled by features
1058 * - hrtimer is actually high res
1060 static inline int hrtick_enabled(struct rq
*rq
)
1062 if (!sched_feat(HRTICK
))
1064 if (!cpu_active(cpu_of(rq
)))
1066 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1069 static void hrtick_clear(struct rq
*rq
)
1071 if (hrtimer_active(&rq
->hrtick_timer
))
1072 hrtimer_cancel(&rq
->hrtick_timer
);
1076 * High-resolution timer tick.
1077 * Runs from hardirq context with interrupts disabled.
1079 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1081 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1083 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1085 spin_lock(&rq
->lock
);
1086 update_rq_clock(rq
);
1087 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1088 spin_unlock(&rq
->lock
);
1090 return HRTIMER_NORESTART
;
1095 * called from hardirq (IPI) context
1097 static void __hrtick_start(void *arg
)
1099 struct rq
*rq
= arg
;
1101 spin_lock(&rq
->lock
);
1102 hrtimer_restart(&rq
->hrtick_timer
);
1103 rq
->hrtick_csd_pending
= 0;
1104 spin_unlock(&rq
->lock
);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq
*rq
, u64 delay
)
1114 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1115 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1117 hrtimer_set_expires(timer
, time
);
1119 if (rq
== this_rq()) {
1120 hrtimer_restart(timer
);
1121 } else if (!rq
->hrtick_csd_pending
) {
1122 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1123 rq
->hrtick_csd_pending
= 1;
1128 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1130 int cpu
= (int)(long)hcpu
;
1133 case CPU_UP_CANCELED
:
1134 case CPU_UP_CANCELED_FROZEN
:
1135 case CPU_DOWN_PREPARE
:
1136 case CPU_DOWN_PREPARE_FROZEN
:
1138 case CPU_DEAD_FROZEN
:
1139 hrtick_clear(cpu_rq(cpu
));
1146 static __init
void init_hrtick(void)
1148 hotcpu_notifier(hotplug_hrtick
, 0);
1152 * Called to set the hrtick timer state.
1154 * called with rq->lock held and irqs disabled
1156 static void hrtick_start(struct rq
*rq
, u64 delay
)
1158 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1159 HRTIMER_MODE_REL
, 0);
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SMP */
1167 static void init_rq_hrtick(struct rq
*rq
)
1170 rq
->hrtick_csd_pending
= 0;
1172 rq
->hrtick_csd
.flags
= 0;
1173 rq
->hrtick_csd
.func
= __hrtick_start
;
1174 rq
->hrtick_csd
.info
= rq
;
1177 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1178 rq
->hrtick_timer
.function
= hrtick
;
1180 #else /* CONFIG_SCHED_HRTICK */
1181 static inline void hrtick_clear(struct rq
*rq
)
1185 static inline void init_rq_hrtick(struct rq
*rq
)
1189 static inline void init_hrtick(void)
1192 #endif /* CONFIG_SCHED_HRTICK */
1195 * resched_task - mark a task 'to be rescheduled now'.
1197 * On UP this means the setting of the need_resched flag, on SMP it
1198 * might also involve a cross-CPU call to trigger the scheduler on
1203 #ifndef tsk_is_polling
1204 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1207 static void resched_task(struct task_struct
*p
)
1211 assert_spin_locked(&task_rq(p
)->lock
);
1213 if (test_tsk_need_resched(p
))
1216 set_tsk_need_resched(p
);
1219 if (cpu
== smp_processor_id())
1222 /* NEED_RESCHED must be visible before we test polling */
1224 if (!tsk_is_polling(p
))
1225 smp_send_reschedule(cpu
);
1228 static void resched_cpu(int cpu
)
1230 struct rq
*rq
= cpu_rq(cpu
);
1231 unsigned long flags
;
1233 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1235 resched_task(cpu_curr(cpu
));
1236 spin_unlock_irqrestore(&rq
->lock
, flags
);
1241 * When add_timer_on() enqueues a timer into the timer wheel of an
1242 * idle CPU then this timer might expire before the next timer event
1243 * which is scheduled to wake up that CPU. In case of a completely
1244 * idle system the next event might even be infinite time into the
1245 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1246 * leaves the inner idle loop so the newly added timer is taken into
1247 * account when the CPU goes back to idle and evaluates the timer
1248 * wheel for the next timer event.
1250 void wake_up_idle_cpu(int cpu
)
1252 struct rq
*rq
= cpu_rq(cpu
);
1254 if (cpu
== smp_processor_id())
1258 * This is safe, as this function is called with the timer
1259 * wheel base lock of (cpu) held. When the CPU is on the way
1260 * to idle and has not yet set rq->curr to idle then it will
1261 * be serialized on the timer wheel base lock and take the new
1262 * timer into account automatically.
1264 if (rq
->curr
!= rq
->idle
)
1268 * We can set TIF_RESCHED on the idle task of the other CPU
1269 * lockless. The worst case is that the other CPU runs the
1270 * idle task through an additional NOOP schedule()
1272 set_tsk_need_resched(rq
->idle
);
1274 /* NEED_RESCHED must be visible before we test polling */
1276 if (!tsk_is_polling(rq
->idle
))
1277 smp_send_reschedule(cpu
);
1279 #endif /* CONFIG_NO_HZ */
1281 #else /* !CONFIG_SMP */
1282 static void resched_task(struct task_struct
*p
)
1284 assert_spin_locked(&task_rq(p
)->lock
);
1285 set_tsk_need_resched(p
);
1287 #endif /* CONFIG_SMP */
1289 #if BITS_PER_LONG == 32
1290 # define WMULT_CONST (~0UL)
1292 # define WMULT_CONST (1UL << 32)
1295 #define WMULT_SHIFT 32
1298 * Shift right and round:
1300 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1303 * delta *= weight / lw
1305 static unsigned long
1306 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1307 struct load_weight
*lw
)
1311 if (!lw
->inv_weight
) {
1312 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1315 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1319 tmp
= (u64
)delta_exec
* weight
;
1321 * Check whether we'd overflow the 64-bit multiplication:
1323 if (unlikely(tmp
> WMULT_CONST
))
1324 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1327 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1329 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1332 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1338 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1345 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1346 * of tasks with abnormal "nice" values across CPUs the contribution that
1347 * each task makes to its run queue's load is weighted according to its
1348 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1349 * scaled version of the new time slice allocation that they receive on time
1353 #define WEIGHT_IDLEPRIO 3
1354 #define WMULT_IDLEPRIO 1431655765
1357 * Nice levels are multiplicative, with a gentle 10% change for every
1358 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1359 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1360 * that remained on nice 0.
1362 * The "10% effect" is relative and cumulative: from _any_ nice level,
1363 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1364 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1365 * If a task goes up by ~10% and another task goes down by ~10% then
1366 * the relative distance between them is ~25%.)
1368 static const int prio_to_weight
[40] = {
1369 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1370 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1371 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1372 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1373 /* 0 */ 1024, 820, 655, 526, 423,
1374 /* 5 */ 335, 272, 215, 172, 137,
1375 /* 10 */ 110, 87, 70, 56, 45,
1376 /* 15 */ 36, 29, 23, 18, 15,
1380 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1382 * In cases where the weight does not change often, we can use the
1383 * precalculated inverse to speed up arithmetics by turning divisions
1384 * into multiplications:
1386 static const u32 prio_to_wmult
[40] = {
1387 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1388 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1389 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1390 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1391 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1392 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1393 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1394 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1397 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1400 * runqueue iterator, to support SMP load-balancing between different
1401 * scheduling classes, without having to expose their internal data
1402 * structures to the load-balancing proper:
1404 struct rq_iterator
{
1406 struct task_struct
*(*start
)(void *);
1407 struct task_struct
*(*next
)(void *);
1411 static unsigned long
1412 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1413 unsigned long max_load_move
, struct sched_domain
*sd
,
1414 enum cpu_idle_type idle
, int *all_pinned
,
1415 int *this_best_prio
, struct rq_iterator
*iterator
);
1418 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1419 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1420 struct rq_iterator
*iterator
);
1423 /* Time spent by the tasks of the cpu accounting group executing in ... */
1424 enum cpuacct_stat_index
{
1425 CPUACCT_STAT_USER
, /* ... user mode */
1426 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1428 CPUACCT_STAT_NSTATS
,
1431 #ifdef CONFIG_CGROUP_CPUACCT
1432 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1433 static void cpuacct_update_stats(struct task_struct
*tsk
,
1434 enum cpuacct_stat_index idx
, cputime_t val
);
1436 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1437 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1438 enum cpuacct_stat_index idx
, cputime_t val
) {}
1441 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1443 update_load_add(&rq
->load
, load
);
1446 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1448 update_load_sub(&rq
->load
, load
);
1451 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1452 typedef int (*tg_visitor
)(struct task_group
*, void *);
1455 * Iterate the full tree, calling @down when first entering a node and @up when
1456 * leaving it for the final time.
1458 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1460 struct task_group
*parent
, *child
;
1464 parent
= &root_task_group
;
1466 ret
= (*down
)(parent
, data
);
1469 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1476 ret
= (*up
)(parent
, data
);
1481 parent
= parent
->parent
;
1490 static int tg_nop(struct task_group
*tg
, void *data
)
1497 static unsigned long source_load(int cpu
, int type
);
1498 static unsigned long target_load(int cpu
, int type
);
1499 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1501 static unsigned long cpu_avg_load_per_task(int cpu
)
1503 struct rq
*rq
= cpu_rq(cpu
);
1504 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1507 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1509 rq
->avg_load_per_task
= 0;
1511 return rq
->avg_load_per_task
;
1514 #ifdef CONFIG_FAIR_GROUP_SCHED
1516 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1519 * Calculate and set the cpu's group shares.
1522 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1523 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1525 unsigned long shares
;
1526 unsigned long rq_weight
;
1531 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1534 * \Sum shares * rq_weight
1535 * shares = -----------------------
1539 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1540 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1542 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1543 sysctl_sched_shares_thresh
) {
1544 struct rq
*rq
= cpu_rq(cpu
);
1545 unsigned long flags
;
1547 spin_lock_irqsave(&rq
->lock
, flags
);
1548 tg
->cfs_rq
[cpu
]->shares
= shares
;
1550 __set_se_shares(tg
->se
[cpu
], shares
);
1551 spin_unlock_irqrestore(&rq
->lock
, flags
);
1556 * Re-compute the task group their per cpu shares over the given domain.
1557 * This needs to be done in a bottom-up fashion because the rq weight of a
1558 * parent group depends on the shares of its child groups.
1560 static int tg_shares_up(struct task_group
*tg
, void *data
)
1562 unsigned long weight
, rq_weight
= 0;
1563 unsigned long shares
= 0;
1564 struct sched_domain
*sd
= data
;
1567 for_each_cpu(i
, sched_domain_span(sd
)) {
1569 * If there are currently no tasks on the cpu pretend there
1570 * is one of average load so that when a new task gets to
1571 * run here it will not get delayed by group starvation.
1573 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1575 weight
= NICE_0_LOAD
;
1577 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1578 rq_weight
+= weight
;
1579 shares
+= tg
->cfs_rq
[i
]->shares
;
1582 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1583 shares
= tg
->shares
;
1585 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1586 shares
= tg
->shares
;
1588 for_each_cpu(i
, sched_domain_span(sd
))
1589 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1595 * Compute the cpu's hierarchical load factor for each task group.
1596 * This needs to be done in a top-down fashion because the load of a child
1597 * group is a fraction of its parents load.
1599 static int tg_load_down(struct task_group
*tg
, void *data
)
1602 long cpu
= (long)data
;
1605 load
= cpu_rq(cpu
)->load
.weight
;
1607 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1608 load
*= tg
->cfs_rq
[cpu
]->shares
;
1609 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1612 tg
->cfs_rq
[cpu
]->h_load
= load
;
1617 static void update_shares(struct sched_domain
*sd
)
1619 u64 now
= cpu_clock(raw_smp_processor_id());
1620 s64 elapsed
= now
- sd
->last_update
;
1622 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1623 sd
->last_update
= now
;
1624 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1628 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1630 spin_unlock(&rq
->lock
);
1632 spin_lock(&rq
->lock
);
1635 static void update_h_load(long cpu
)
1637 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1642 static inline void update_shares(struct sched_domain
*sd
)
1646 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1652 #ifdef CONFIG_PREEMPT
1655 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1656 * way at the expense of forcing extra atomic operations in all
1657 * invocations. This assures that the double_lock is acquired using the
1658 * same underlying policy as the spinlock_t on this architecture, which
1659 * reduces latency compared to the unfair variant below. However, it
1660 * also adds more overhead and therefore may reduce throughput.
1662 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1663 __releases(this_rq
->lock
)
1664 __acquires(busiest
->lock
)
1665 __acquires(this_rq
->lock
)
1667 spin_unlock(&this_rq
->lock
);
1668 double_rq_lock(this_rq
, busiest
);
1675 * Unfair double_lock_balance: Optimizes throughput at the expense of
1676 * latency by eliminating extra atomic operations when the locks are
1677 * already in proper order on entry. This favors lower cpu-ids and will
1678 * grant the double lock to lower cpus over higher ids under contention,
1679 * regardless of entry order into the function.
1681 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1682 __releases(this_rq
->lock
)
1683 __acquires(busiest
->lock
)
1684 __acquires(this_rq
->lock
)
1688 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1689 if (busiest
< this_rq
) {
1690 spin_unlock(&this_rq
->lock
);
1691 spin_lock(&busiest
->lock
);
1692 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1695 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1700 #endif /* CONFIG_PREEMPT */
1703 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1705 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1707 if (unlikely(!irqs_disabled())) {
1708 /* printk() doesn't work good under rq->lock */
1709 spin_unlock(&this_rq
->lock
);
1713 return _double_lock_balance(this_rq
, busiest
);
1716 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1717 __releases(busiest
->lock
)
1719 spin_unlock(&busiest
->lock
);
1720 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1724 #ifdef CONFIG_FAIR_GROUP_SCHED
1725 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1728 cfs_rq
->shares
= shares
;
1733 #include "sched_stats.h"
1734 #include "sched_idletask.c"
1735 #include "sched_fair.c"
1736 #include "sched_rt.c"
1737 #ifdef CONFIG_SCHED_DEBUG
1738 # include "sched_debug.c"
1741 #define sched_class_highest (&rt_sched_class)
1742 #define for_each_class(class) \
1743 for (class = sched_class_highest; class; class = class->next)
1745 static void inc_nr_running(struct rq
*rq
)
1750 static void dec_nr_running(struct rq
*rq
)
1755 static void set_load_weight(struct task_struct
*p
)
1757 if (task_has_rt_policy(p
)) {
1758 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1759 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1764 * SCHED_IDLE tasks get minimal weight:
1766 if (p
->policy
== SCHED_IDLE
) {
1767 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1768 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1772 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1773 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1776 static void update_avg(u64
*avg
, u64 sample
)
1778 s64 diff
= sample
- *avg
;
1782 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1785 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1787 sched_info_queued(p
);
1788 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1792 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1795 if (p
->se
.last_wakeup
) {
1796 update_avg(&p
->se
.avg_overlap
,
1797 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1798 p
->se
.last_wakeup
= 0;
1800 update_avg(&p
->se
.avg_wakeup
,
1801 sysctl_sched_wakeup_granularity
);
1805 sched_info_dequeued(p
);
1806 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1811 * __normal_prio - return the priority that is based on the static prio
1813 static inline int __normal_prio(struct task_struct
*p
)
1815 return p
->static_prio
;
1819 * Calculate the expected normal priority: i.e. priority
1820 * without taking RT-inheritance into account. Might be
1821 * boosted by interactivity modifiers. Changes upon fork,
1822 * setprio syscalls, and whenever the interactivity
1823 * estimator recalculates.
1825 static inline int normal_prio(struct task_struct
*p
)
1829 if (task_has_rt_policy(p
))
1830 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1832 prio
= __normal_prio(p
);
1837 * Calculate the current priority, i.e. the priority
1838 * taken into account by the scheduler. This value might
1839 * be boosted by RT tasks, or might be boosted by
1840 * interactivity modifiers. Will be RT if the task got
1841 * RT-boosted. If not then it returns p->normal_prio.
1843 static int effective_prio(struct task_struct
*p
)
1845 p
->normal_prio
= normal_prio(p
);
1847 * If we are RT tasks or we were boosted to RT priority,
1848 * keep the priority unchanged. Otherwise, update priority
1849 * to the normal priority:
1851 if (!rt_prio(p
->prio
))
1852 return p
->normal_prio
;
1857 * activate_task - move a task to the runqueue.
1859 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1861 if (task_contributes_to_load(p
))
1862 rq
->nr_uninterruptible
--;
1864 enqueue_task(rq
, p
, wakeup
);
1869 * deactivate_task - remove a task from the runqueue.
1871 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1873 if (task_contributes_to_load(p
))
1874 rq
->nr_uninterruptible
++;
1876 dequeue_task(rq
, p
, sleep
);
1881 * task_curr - is this task currently executing on a CPU?
1882 * @p: the task in question.
1884 inline int task_curr(const struct task_struct
*p
)
1886 return cpu_curr(task_cpu(p
)) == p
;
1889 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1891 set_task_rq(p
, cpu
);
1894 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1895 * successfuly executed on another CPU. We must ensure that updates of
1896 * per-task data have been completed by this moment.
1899 task_thread_info(p
)->cpu
= cpu
;
1903 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1904 const struct sched_class
*prev_class
,
1905 int oldprio
, int running
)
1907 if (prev_class
!= p
->sched_class
) {
1908 if (prev_class
->switched_from
)
1909 prev_class
->switched_from(rq
, p
, running
);
1910 p
->sched_class
->switched_to(rq
, p
, running
);
1912 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1917 /* Used instead of source_load when we know the type == 0 */
1918 static unsigned long weighted_cpuload(const int cpu
)
1920 return cpu_rq(cpu
)->load
.weight
;
1924 * Is this task likely cache-hot:
1927 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1932 * Buddy candidates are cache hot:
1934 if (sched_feat(CACHE_HOT_BUDDY
) &&
1935 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1936 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1939 if (p
->sched_class
!= &fair_sched_class
)
1942 if (sysctl_sched_migration_cost
== -1)
1944 if (sysctl_sched_migration_cost
== 0)
1947 delta
= now
- p
->se
.exec_start
;
1949 return delta
< (s64
)sysctl_sched_migration_cost
;
1953 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1955 int old_cpu
= task_cpu(p
);
1956 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1957 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1958 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1961 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1963 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1965 #ifdef CONFIG_SCHEDSTATS
1966 if (p
->se
.wait_start
)
1967 p
->se
.wait_start
-= clock_offset
;
1968 if (p
->se
.sleep_start
)
1969 p
->se
.sleep_start
-= clock_offset
;
1970 if (p
->se
.block_start
)
1971 p
->se
.block_start
-= clock_offset
;
1973 if (old_cpu
!= new_cpu
) {
1974 p
->se
.nr_migrations
++;
1975 new_rq
->nr_migrations_in
++;
1976 #ifdef CONFIG_SCHEDSTATS
1977 if (task_hot(p
, old_rq
->clock
, NULL
))
1978 schedstat_inc(p
, se
.nr_forced2_migrations
);
1980 perf_counter_task_migration(p
, new_cpu
);
1982 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1983 new_cfsrq
->min_vruntime
;
1985 __set_task_cpu(p
, new_cpu
);
1988 struct migration_req
{
1989 struct list_head list
;
1991 struct task_struct
*task
;
1994 struct completion done
;
1998 * The task's runqueue lock must be held.
1999 * Returns true if you have to wait for migration thread.
2002 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2004 struct rq
*rq
= task_rq(p
);
2007 * If the task is not on a runqueue (and not running), then
2008 * it is sufficient to simply update the task's cpu field.
2010 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2011 set_task_cpu(p
, dest_cpu
);
2015 init_completion(&req
->done
);
2017 req
->dest_cpu
= dest_cpu
;
2018 list_add(&req
->list
, &rq
->migration_queue
);
2024 * wait_task_inactive - wait for a thread to unschedule.
2026 * If @match_state is nonzero, it's the @p->state value just checked and
2027 * not expected to change. If it changes, i.e. @p might have woken up,
2028 * then return zero. When we succeed in waiting for @p to be off its CPU,
2029 * we return a positive number (its total switch count). If a second call
2030 * a short while later returns the same number, the caller can be sure that
2031 * @p has remained unscheduled the whole time.
2033 * The caller must ensure that the task *will* unschedule sometime soon,
2034 * else this function might spin for a *long* time. This function can't
2035 * be called with interrupts off, or it may introduce deadlock with
2036 * smp_call_function() if an IPI is sent by the same process we are
2037 * waiting to become inactive.
2039 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2041 unsigned long flags
;
2048 * We do the initial early heuristics without holding
2049 * any task-queue locks at all. We'll only try to get
2050 * the runqueue lock when things look like they will
2056 * If the task is actively running on another CPU
2057 * still, just relax and busy-wait without holding
2060 * NOTE! Since we don't hold any locks, it's not
2061 * even sure that "rq" stays as the right runqueue!
2062 * But we don't care, since "task_running()" will
2063 * return false if the runqueue has changed and p
2064 * is actually now running somewhere else!
2066 while (task_running(rq
, p
)) {
2067 if (match_state
&& unlikely(p
->state
!= match_state
))
2073 * Ok, time to look more closely! We need the rq
2074 * lock now, to be *sure*. If we're wrong, we'll
2075 * just go back and repeat.
2077 rq
= task_rq_lock(p
, &flags
);
2078 trace_sched_wait_task(rq
, p
);
2079 running
= task_running(rq
, p
);
2080 on_rq
= p
->se
.on_rq
;
2082 if (!match_state
|| p
->state
== match_state
)
2083 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2084 task_rq_unlock(rq
, &flags
);
2087 * If it changed from the expected state, bail out now.
2089 if (unlikely(!ncsw
))
2093 * Was it really running after all now that we
2094 * checked with the proper locks actually held?
2096 * Oops. Go back and try again..
2098 if (unlikely(running
)) {
2104 * It's not enough that it's not actively running,
2105 * it must be off the runqueue _entirely_, and not
2108 * So if it was still runnable (but just not actively
2109 * running right now), it's preempted, and we should
2110 * yield - it could be a while.
2112 if (unlikely(on_rq
)) {
2113 schedule_timeout_uninterruptible(1);
2118 * Ahh, all good. It wasn't running, and it wasn't
2119 * runnable, which means that it will never become
2120 * running in the future either. We're all done!
2129 * kick_process - kick a running thread to enter/exit the kernel
2130 * @p: the to-be-kicked thread
2132 * Cause a process which is running on another CPU to enter
2133 * kernel-mode, without any delay. (to get signals handled.)
2135 * NOTE: this function doesnt have to take the runqueue lock,
2136 * because all it wants to ensure is that the remote task enters
2137 * the kernel. If the IPI races and the task has been migrated
2138 * to another CPU then no harm is done and the purpose has been
2141 void kick_process(struct task_struct
*p
)
2147 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2148 smp_send_reschedule(cpu
);
2153 * Return a low guess at the load of a migration-source cpu weighted
2154 * according to the scheduling class and "nice" value.
2156 * We want to under-estimate the load of migration sources, to
2157 * balance conservatively.
2159 static unsigned long source_load(int cpu
, int type
)
2161 struct rq
*rq
= cpu_rq(cpu
);
2162 unsigned long total
= weighted_cpuload(cpu
);
2164 if (type
== 0 || !sched_feat(LB_BIAS
))
2167 return min(rq
->cpu_load
[type
-1], total
);
2171 * Return a high guess at the load of a migration-target cpu weighted
2172 * according to the scheduling class and "nice" value.
2174 static unsigned long target_load(int cpu
, int type
)
2176 struct rq
*rq
= cpu_rq(cpu
);
2177 unsigned long total
= weighted_cpuload(cpu
);
2179 if (type
== 0 || !sched_feat(LB_BIAS
))
2182 return max(rq
->cpu_load
[type
-1], total
);
2186 * find_idlest_group finds and returns the least busy CPU group within the
2189 static struct sched_group
*
2190 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2192 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2193 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2194 int load_idx
= sd
->forkexec_idx
;
2195 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2198 unsigned long load
, avg_load
;
2202 /* Skip over this group if it has no CPUs allowed */
2203 if (!cpumask_intersects(sched_group_cpus(group
),
2207 local_group
= cpumask_test_cpu(this_cpu
,
2208 sched_group_cpus(group
));
2210 /* Tally up the load of all CPUs in the group */
2213 for_each_cpu(i
, sched_group_cpus(group
)) {
2214 /* Bias balancing toward cpus of our domain */
2216 load
= source_load(i
, load_idx
);
2218 load
= target_load(i
, load_idx
);
2223 /* Adjust by relative CPU power of the group */
2224 avg_load
= sg_div_cpu_power(group
,
2225 avg_load
* SCHED_LOAD_SCALE
);
2228 this_load
= avg_load
;
2230 } else if (avg_load
< min_load
) {
2231 min_load
= avg_load
;
2234 } while (group
= group
->next
, group
!= sd
->groups
);
2236 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2242 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2245 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2247 unsigned long load
, min_load
= ULONG_MAX
;
2251 /* Traverse only the allowed CPUs */
2252 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2253 load
= weighted_cpuload(i
);
2255 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2265 * sched_balance_self: balance the current task (running on cpu) in domains
2266 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2269 * Balance, ie. select the least loaded group.
2271 * Returns the target CPU number, or the same CPU if no balancing is needed.
2273 * preempt must be disabled.
2275 static int sched_balance_self(int cpu
, int flag
)
2277 struct task_struct
*t
= current
;
2278 struct sched_domain
*tmp
, *sd
= NULL
;
2280 for_each_domain(cpu
, tmp
) {
2282 * If power savings logic is enabled for a domain, stop there.
2284 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2286 if (tmp
->flags
& flag
)
2294 struct sched_group
*group
;
2295 int new_cpu
, weight
;
2297 if (!(sd
->flags
& flag
)) {
2302 group
= find_idlest_group(sd
, t
, cpu
);
2308 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2309 if (new_cpu
== -1 || new_cpu
== cpu
) {
2310 /* Now try balancing at a lower domain level of cpu */
2315 /* Now try balancing at a lower domain level of new_cpu */
2317 weight
= cpumask_weight(sched_domain_span(sd
));
2319 for_each_domain(cpu
, tmp
) {
2320 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2322 if (tmp
->flags
& flag
)
2325 /* while loop will break here if sd == NULL */
2331 #endif /* CONFIG_SMP */
2334 * task_oncpu_function_call - call a function on the cpu on which a task runs
2335 * @p: the task to evaluate
2336 * @func: the function to be called
2337 * @info: the function call argument
2339 * Calls the function @func when the task is currently running. This might
2340 * be on the current CPU, which just calls the function directly
2342 void task_oncpu_function_call(struct task_struct
*p
,
2343 void (*func
) (void *info
), void *info
)
2350 smp_call_function_single(cpu
, func
, info
, 1);
2355 * try_to_wake_up - wake up a thread
2356 * @p: the to-be-woken-up thread
2357 * @state: the mask of task states that can be woken
2358 * @sync: do a synchronous wakeup?
2360 * Put it on the run-queue if it's not already there. The "current"
2361 * thread is always on the run-queue (except when the actual
2362 * re-schedule is in progress), and as such you're allowed to do
2363 * the simpler "current->state = TASK_RUNNING" to mark yourself
2364 * runnable without the overhead of this.
2366 * returns failure only if the task is already active.
2368 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2370 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2371 unsigned long flags
;
2375 if (!sched_feat(SYNC_WAKEUPS
))
2379 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2380 struct sched_domain
*sd
;
2382 this_cpu
= raw_smp_processor_id();
2385 for_each_domain(this_cpu
, sd
) {
2386 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2395 rq
= task_rq_lock(p
, &flags
);
2396 update_rq_clock(rq
);
2397 old_state
= p
->state
;
2398 if (!(old_state
& state
))
2406 this_cpu
= smp_processor_id();
2409 if (unlikely(task_running(rq
, p
)))
2412 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2413 if (cpu
!= orig_cpu
) {
2414 set_task_cpu(p
, cpu
);
2415 task_rq_unlock(rq
, &flags
);
2416 /* might preempt at this point */
2417 rq
= task_rq_lock(p
, &flags
);
2418 old_state
= p
->state
;
2419 if (!(old_state
& state
))
2424 this_cpu
= smp_processor_id();
2428 #ifdef CONFIG_SCHEDSTATS
2429 schedstat_inc(rq
, ttwu_count
);
2430 if (cpu
== this_cpu
)
2431 schedstat_inc(rq
, ttwu_local
);
2433 struct sched_domain
*sd
;
2434 for_each_domain(this_cpu
, sd
) {
2435 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2436 schedstat_inc(sd
, ttwu_wake_remote
);
2441 #endif /* CONFIG_SCHEDSTATS */
2444 #endif /* CONFIG_SMP */
2445 schedstat_inc(p
, se
.nr_wakeups
);
2447 schedstat_inc(p
, se
.nr_wakeups_sync
);
2448 if (orig_cpu
!= cpu
)
2449 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2450 if (cpu
== this_cpu
)
2451 schedstat_inc(p
, se
.nr_wakeups_local
);
2453 schedstat_inc(p
, se
.nr_wakeups_remote
);
2454 activate_task(rq
, p
, 1);
2458 * Only attribute actual wakeups done by this task.
2460 if (!in_interrupt()) {
2461 struct sched_entity
*se
= ¤t
->se
;
2462 u64 sample
= se
->sum_exec_runtime
;
2464 if (se
->last_wakeup
)
2465 sample
-= se
->last_wakeup
;
2467 sample
-= se
->start_runtime
;
2468 update_avg(&se
->avg_wakeup
, sample
);
2470 se
->last_wakeup
= se
->sum_exec_runtime
;
2474 trace_sched_wakeup(rq
, p
, success
);
2475 check_preempt_curr(rq
, p
, sync
);
2477 p
->state
= TASK_RUNNING
;
2479 if (p
->sched_class
->task_wake_up
)
2480 p
->sched_class
->task_wake_up(rq
, p
);
2483 task_rq_unlock(rq
, &flags
);
2488 int wake_up_process(struct task_struct
*p
)
2490 return try_to_wake_up(p
, TASK_ALL
, 0);
2492 EXPORT_SYMBOL(wake_up_process
);
2494 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2496 return try_to_wake_up(p
, state
, 0);
2500 * Perform scheduler related setup for a newly forked process p.
2501 * p is forked by current.
2503 * __sched_fork() is basic setup used by init_idle() too:
2505 static void __sched_fork(struct task_struct
*p
)
2507 p
->se
.exec_start
= 0;
2508 p
->se
.sum_exec_runtime
= 0;
2509 p
->se
.prev_sum_exec_runtime
= 0;
2510 p
->se
.nr_migrations
= 0;
2511 p
->se
.last_wakeup
= 0;
2512 p
->se
.avg_overlap
= 0;
2513 p
->se
.start_runtime
= 0;
2514 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2516 #ifdef CONFIG_SCHEDSTATS
2517 p
->se
.wait_start
= 0;
2518 p
->se
.sum_sleep_runtime
= 0;
2519 p
->se
.sleep_start
= 0;
2520 p
->se
.block_start
= 0;
2521 p
->se
.sleep_max
= 0;
2522 p
->se
.block_max
= 0;
2524 p
->se
.slice_max
= 0;
2528 INIT_LIST_HEAD(&p
->rt
.run_list
);
2530 INIT_LIST_HEAD(&p
->se
.group_node
);
2532 #ifdef CONFIG_PREEMPT_NOTIFIERS
2533 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2537 * We mark the process as running here, but have not actually
2538 * inserted it onto the runqueue yet. This guarantees that
2539 * nobody will actually run it, and a signal or other external
2540 * event cannot wake it up and insert it on the runqueue either.
2542 p
->state
= TASK_RUNNING
;
2546 * fork()/clone()-time setup:
2548 void sched_fork(struct task_struct
*p
, int clone_flags
)
2550 int cpu
= get_cpu();
2555 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2557 set_task_cpu(p
, cpu
);
2560 * Make sure we do not leak PI boosting priority to the child:
2562 p
->prio
= current
->normal_prio
;
2563 if (!rt_prio(p
->prio
))
2564 p
->sched_class
= &fair_sched_class
;
2566 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2567 if (likely(sched_info_on()))
2568 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2570 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2573 #ifdef CONFIG_PREEMPT
2574 /* Want to start with kernel preemption disabled. */
2575 task_thread_info(p
)->preempt_count
= 1;
2577 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2583 * wake_up_new_task - wake up a newly created task for the first time.
2585 * This function will do some initial scheduler statistics housekeeping
2586 * that must be done for every newly created context, then puts the task
2587 * on the runqueue and wakes it.
2589 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2591 unsigned long flags
;
2594 rq
= task_rq_lock(p
, &flags
);
2595 BUG_ON(p
->state
!= TASK_RUNNING
);
2596 update_rq_clock(rq
);
2598 p
->prio
= effective_prio(p
);
2600 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2601 activate_task(rq
, p
, 0);
2604 * Let the scheduling class do new task startup
2605 * management (if any):
2607 p
->sched_class
->task_new(rq
, p
);
2610 trace_sched_wakeup_new(rq
, p
, 1);
2611 check_preempt_curr(rq
, p
, 0);
2613 if (p
->sched_class
->task_wake_up
)
2614 p
->sched_class
->task_wake_up(rq
, p
);
2616 task_rq_unlock(rq
, &flags
);
2619 #ifdef CONFIG_PREEMPT_NOTIFIERS
2622 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2623 * @notifier: notifier struct to register
2625 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2627 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2629 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2632 * preempt_notifier_unregister - no longer interested in preemption notifications
2633 * @notifier: notifier struct to unregister
2635 * This is safe to call from within a preemption notifier.
2637 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2639 hlist_del(¬ifier
->link
);
2641 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2643 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2645 struct preempt_notifier
*notifier
;
2646 struct hlist_node
*node
;
2648 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2649 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2653 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2654 struct task_struct
*next
)
2656 struct preempt_notifier
*notifier
;
2657 struct hlist_node
*node
;
2659 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2660 notifier
->ops
->sched_out(notifier
, next
);
2663 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2665 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2670 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2671 struct task_struct
*next
)
2675 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2678 * prepare_task_switch - prepare to switch tasks
2679 * @rq: the runqueue preparing to switch
2680 * @prev: the current task that is being switched out
2681 * @next: the task we are going to switch to.
2683 * This is called with the rq lock held and interrupts off. It must
2684 * be paired with a subsequent finish_task_switch after the context
2687 * prepare_task_switch sets up locking and calls architecture specific
2691 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2692 struct task_struct
*next
)
2694 fire_sched_out_preempt_notifiers(prev
, next
);
2695 prepare_lock_switch(rq
, next
);
2696 prepare_arch_switch(next
);
2700 * finish_task_switch - clean up after a task-switch
2701 * @rq: runqueue associated with task-switch
2702 * @prev: the thread we just switched away from.
2704 * finish_task_switch must be called after the context switch, paired
2705 * with a prepare_task_switch call before the context switch.
2706 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2707 * and do any other architecture-specific cleanup actions.
2709 * Note that we may have delayed dropping an mm in context_switch(). If
2710 * so, we finish that here outside of the runqueue lock. (Doing it
2711 * with the lock held can cause deadlocks; see schedule() for
2714 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2715 __releases(rq
->lock
)
2717 struct mm_struct
*mm
= rq
->prev_mm
;
2720 int post_schedule
= 0;
2722 if (current
->sched_class
->needs_post_schedule
)
2723 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2729 * A task struct has one reference for the use as "current".
2730 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2731 * schedule one last time. The schedule call will never return, and
2732 * the scheduled task must drop that reference.
2733 * The test for TASK_DEAD must occur while the runqueue locks are
2734 * still held, otherwise prev could be scheduled on another cpu, die
2735 * there before we look at prev->state, and then the reference would
2737 * Manfred Spraul <manfred@colorfullife.com>
2739 prev_state
= prev
->state
;
2740 finish_arch_switch(prev
);
2741 perf_counter_task_sched_in(current
, cpu_of(rq
));
2742 finish_lock_switch(rq
, prev
);
2745 current
->sched_class
->post_schedule(rq
);
2748 fire_sched_in_preempt_notifiers(current
);
2751 if (unlikely(prev_state
== TASK_DEAD
)) {
2753 * Remove function-return probe instances associated with this
2754 * task and put them back on the free list.
2756 kprobe_flush_task(prev
);
2757 put_task_struct(prev
);
2762 * schedule_tail - first thing a freshly forked thread must call.
2763 * @prev: the thread we just switched away from.
2765 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2766 __releases(rq
->lock
)
2768 struct rq
*rq
= this_rq();
2770 finish_task_switch(rq
, prev
);
2771 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2772 /* In this case, finish_task_switch does not reenable preemption */
2775 if (current
->set_child_tid
)
2776 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2780 * context_switch - switch to the new MM and the new
2781 * thread's register state.
2784 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2785 struct task_struct
*next
)
2787 struct mm_struct
*mm
, *oldmm
;
2789 prepare_task_switch(rq
, prev
, next
);
2790 trace_sched_switch(rq
, prev
, next
);
2792 oldmm
= prev
->active_mm
;
2794 * For paravirt, this is coupled with an exit in switch_to to
2795 * combine the page table reload and the switch backend into
2798 arch_enter_lazy_cpu_mode();
2800 if (unlikely(!mm
)) {
2801 next
->active_mm
= oldmm
;
2802 atomic_inc(&oldmm
->mm_count
);
2803 enter_lazy_tlb(oldmm
, next
);
2805 switch_mm(oldmm
, mm
, next
);
2807 if (unlikely(!prev
->mm
)) {
2808 prev
->active_mm
= NULL
;
2809 rq
->prev_mm
= oldmm
;
2812 * Since the runqueue lock will be released by the next
2813 * task (which is an invalid locking op but in the case
2814 * of the scheduler it's an obvious special-case), so we
2815 * do an early lockdep release here:
2817 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2818 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2821 /* Here we just switch the register state and the stack. */
2822 switch_to(prev
, next
, prev
);
2826 * this_rq must be evaluated again because prev may have moved
2827 * CPUs since it called schedule(), thus the 'rq' on its stack
2828 * frame will be invalid.
2830 finish_task_switch(this_rq(), prev
);
2834 * nr_running, nr_uninterruptible and nr_context_switches:
2836 * externally visible scheduler statistics: current number of runnable
2837 * threads, current number of uninterruptible-sleeping threads, total
2838 * number of context switches performed since bootup.
2840 unsigned long nr_running(void)
2842 unsigned long i
, sum
= 0;
2844 for_each_online_cpu(i
)
2845 sum
+= cpu_rq(i
)->nr_running
;
2850 unsigned long nr_uninterruptible(void)
2852 unsigned long i
, sum
= 0;
2854 for_each_possible_cpu(i
)
2855 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2858 * Since we read the counters lockless, it might be slightly
2859 * inaccurate. Do not allow it to go below zero though:
2861 if (unlikely((long)sum
< 0))
2867 unsigned long long nr_context_switches(void)
2870 unsigned long long sum
= 0;
2872 for_each_possible_cpu(i
)
2873 sum
+= cpu_rq(i
)->nr_switches
;
2878 unsigned long nr_iowait(void)
2880 unsigned long i
, sum
= 0;
2882 for_each_possible_cpu(i
)
2883 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2888 unsigned long nr_active(void)
2890 unsigned long i
, running
= 0, uninterruptible
= 0;
2892 for_each_online_cpu(i
) {
2893 running
+= cpu_rq(i
)->nr_running
;
2894 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2897 if (unlikely((long)uninterruptible
< 0))
2898 uninterruptible
= 0;
2900 return running
+ uninterruptible
;
2904 * Externally visible per-cpu scheduler statistics:
2905 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2907 u64
cpu_nr_migrations(int cpu
)
2909 return cpu_rq(cpu
)->nr_migrations_in
;
2913 * Update rq->cpu_load[] statistics. This function is usually called every
2914 * scheduler tick (TICK_NSEC).
2916 static void update_cpu_load(struct rq
*this_rq
)
2918 unsigned long this_load
= this_rq
->load
.weight
;
2921 this_rq
->nr_load_updates
++;
2923 /* Update our load: */
2924 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2925 unsigned long old_load
, new_load
;
2927 /* scale is effectively 1 << i now, and >> i divides by scale */
2929 old_load
= this_rq
->cpu_load
[i
];
2930 new_load
= this_load
;
2932 * Round up the averaging division if load is increasing. This
2933 * prevents us from getting stuck on 9 if the load is 10, for
2936 if (new_load
> old_load
)
2937 new_load
+= scale
-1;
2938 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2945 * double_rq_lock - safely lock two runqueues
2947 * Note this does not disable interrupts like task_rq_lock,
2948 * you need to do so manually before calling.
2950 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2951 __acquires(rq1
->lock
)
2952 __acquires(rq2
->lock
)
2954 BUG_ON(!irqs_disabled());
2956 spin_lock(&rq1
->lock
);
2957 __acquire(rq2
->lock
); /* Fake it out ;) */
2960 spin_lock(&rq1
->lock
);
2961 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2963 spin_lock(&rq2
->lock
);
2964 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2967 update_rq_clock(rq1
);
2968 update_rq_clock(rq2
);
2972 * double_rq_unlock - safely unlock two runqueues
2974 * Note this does not restore interrupts like task_rq_unlock,
2975 * you need to do so manually after calling.
2977 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2978 __releases(rq1
->lock
)
2979 __releases(rq2
->lock
)
2981 spin_unlock(&rq1
->lock
);
2983 spin_unlock(&rq2
->lock
);
2985 __release(rq2
->lock
);
2989 * If dest_cpu is allowed for this process, migrate the task to it.
2990 * This is accomplished by forcing the cpu_allowed mask to only
2991 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2992 * the cpu_allowed mask is restored.
2994 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2996 struct migration_req req
;
2997 unsigned long flags
;
3000 rq
= task_rq_lock(p
, &flags
);
3001 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3002 || unlikely(!cpu_active(dest_cpu
)))
3005 /* force the process onto the specified CPU */
3006 if (migrate_task(p
, dest_cpu
, &req
)) {
3007 /* Need to wait for migration thread (might exit: take ref). */
3008 struct task_struct
*mt
= rq
->migration_thread
;
3010 get_task_struct(mt
);
3011 task_rq_unlock(rq
, &flags
);
3012 wake_up_process(mt
);
3013 put_task_struct(mt
);
3014 wait_for_completion(&req
.done
);
3019 task_rq_unlock(rq
, &flags
);
3023 * sched_exec - execve() is a valuable balancing opportunity, because at
3024 * this point the task has the smallest effective memory and cache footprint.
3026 void sched_exec(void)
3028 int new_cpu
, this_cpu
= get_cpu();
3029 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3031 if (new_cpu
!= this_cpu
)
3032 sched_migrate_task(current
, new_cpu
);
3036 * pull_task - move a task from a remote runqueue to the local runqueue.
3037 * Both runqueues must be locked.
3039 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3040 struct rq
*this_rq
, int this_cpu
)
3042 deactivate_task(src_rq
, p
, 0);
3043 set_task_cpu(p
, this_cpu
);
3044 activate_task(this_rq
, p
, 0);
3046 * Note that idle threads have a prio of MAX_PRIO, for this test
3047 * to be always true for them.
3049 check_preempt_curr(this_rq
, p
, 0);
3053 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3056 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3057 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3060 int tsk_cache_hot
= 0;
3062 * We do not migrate tasks that are:
3063 * 1) running (obviously), or
3064 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3065 * 3) are cache-hot on their current CPU.
3067 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3068 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3073 if (task_running(rq
, p
)) {
3074 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3079 * Aggressive migration if:
3080 * 1) task is cache cold, or
3081 * 2) too many balance attempts have failed.
3084 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3085 if (!tsk_cache_hot
||
3086 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3087 #ifdef CONFIG_SCHEDSTATS
3088 if (tsk_cache_hot
) {
3089 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3090 schedstat_inc(p
, se
.nr_forced_migrations
);
3096 if (tsk_cache_hot
) {
3097 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3103 static unsigned long
3104 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3105 unsigned long max_load_move
, struct sched_domain
*sd
,
3106 enum cpu_idle_type idle
, int *all_pinned
,
3107 int *this_best_prio
, struct rq_iterator
*iterator
)
3109 int loops
= 0, pulled
= 0, pinned
= 0;
3110 struct task_struct
*p
;
3111 long rem_load_move
= max_load_move
;
3113 if (max_load_move
== 0)
3119 * Start the load-balancing iterator:
3121 p
= iterator
->start(iterator
->arg
);
3123 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3126 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3127 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3128 p
= iterator
->next(iterator
->arg
);
3132 pull_task(busiest
, p
, this_rq
, this_cpu
);
3134 rem_load_move
-= p
->se
.load
.weight
;
3136 #ifdef CONFIG_PREEMPT
3138 * NEWIDLE balancing is a source of latency, so preemptible kernels
3139 * will stop after the first task is pulled to minimize the critical
3142 if (idle
== CPU_NEWLY_IDLE
)
3147 * We only want to steal up to the prescribed amount of weighted load.
3149 if (rem_load_move
> 0) {
3150 if (p
->prio
< *this_best_prio
)
3151 *this_best_prio
= p
->prio
;
3152 p
= iterator
->next(iterator
->arg
);
3157 * Right now, this is one of only two places pull_task() is called,
3158 * so we can safely collect pull_task() stats here rather than
3159 * inside pull_task().
3161 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3164 *all_pinned
= pinned
;
3166 return max_load_move
- rem_load_move
;
3170 * move_tasks tries to move up to max_load_move weighted load from busiest to
3171 * this_rq, as part of a balancing operation within domain "sd".
3172 * Returns 1 if successful and 0 otherwise.
3174 * Called with both runqueues locked.
3176 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3177 unsigned long max_load_move
,
3178 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3181 const struct sched_class
*class = sched_class_highest
;
3182 unsigned long total_load_moved
= 0;
3183 int this_best_prio
= this_rq
->curr
->prio
;
3187 class->load_balance(this_rq
, this_cpu
, busiest
,
3188 max_load_move
- total_load_moved
,
3189 sd
, idle
, all_pinned
, &this_best_prio
);
3190 class = class->next
;
3192 #ifdef CONFIG_PREEMPT
3194 * NEWIDLE balancing is a source of latency, so preemptible
3195 * kernels will stop after the first task is pulled to minimize
3196 * the critical section.
3198 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3201 } while (class && max_load_move
> total_load_moved
);
3203 return total_load_moved
> 0;
3207 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3208 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3209 struct rq_iterator
*iterator
)
3211 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3215 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3216 pull_task(busiest
, p
, this_rq
, this_cpu
);
3218 * Right now, this is only the second place pull_task()
3219 * is called, so we can safely collect pull_task()
3220 * stats here rather than inside pull_task().
3222 schedstat_inc(sd
, lb_gained
[idle
]);
3226 p
= iterator
->next(iterator
->arg
);
3233 * move_one_task tries to move exactly one task from busiest to this_rq, as
3234 * part of active balancing operations within "domain".
3235 * Returns 1 if successful and 0 otherwise.
3237 * Called with both runqueues locked.
3239 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3240 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3242 const struct sched_class
*class;
3244 for (class = sched_class_highest
; class; class = class->next
)
3245 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3250 /********** Helpers for find_busiest_group ************************/
3252 * sd_lb_stats - Structure to store the statistics of a sched_domain
3253 * during load balancing.
3255 struct sd_lb_stats
{
3256 struct sched_group
*busiest
; /* Busiest group in this sd */
3257 struct sched_group
*this; /* Local group in this sd */
3258 unsigned long total_load
; /* Total load of all groups in sd */
3259 unsigned long total_pwr
; /* Total power of all groups in sd */
3260 unsigned long avg_load
; /* Average load across all groups in sd */
3262 /** Statistics of this group */
3263 unsigned long this_load
;
3264 unsigned long this_load_per_task
;
3265 unsigned long this_nr_running
;
3267 /* Statistics of the busiest group */
3268 unsigned long max_load
;
3269 unsigned long busiest_load_per_task
;
3270 unsigned long busiest_nr_running
;
3272 int group_imb
; /* Is there imbalance in this sd */
3273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3274 int power_savings_balance
; /* Is powersave balance needed for this sd */
3275 struct sched_group
*group_min
; /* Least loaded group in sd */
3276 struct sched_group
*group_leader
; /* Group which relieves group_min */
3277 unsigned long min_load_per_task
; /* load_per_task in group_min */
3278 unsigned long leader_nr_running
; /* Nr running of group_leader */
3279 unsigned long min_nr_running
; /* Nr running of group_min */
3284 * sg_lb_stats - stats of a sched_group required for load_balancing
3286 struct sg_lb_stats
{
3287 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3288 unsigned long group_load
; /* Total load over the CPUs of the group */
3289 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3290 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3291 unsigned long group_capacity
;
3292 int group_imb
; /* Is there an imbalance in the group ? */
3296 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3297 * @group: The group whose first cpu is to be returned.
3299 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3301 return cpumask_first(sched_group_cpus(group
));
3305 * get_sd_load_idx - Obtain the load index for a given sched domain.
3306 * @sd: The sched_domain whose load_idx is to be obtained.
3307 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3309 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3310 enum cpu_idle_type idle
)
3316 load_idx
= sd
->busy_idx
;
3319 case CPU_NEWLY_IDLE
:
3320 load_idx
= sd
->newidle_idx
;
3323 load_idx
= sd
->idle_idx
;
3331 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3333 * init_sd_power_savings_stats - Initialize power savings statistics for
3334 * the given sched_domain, during load balancing.
3336 * @sd: Sched domain whose power-savings statistics are to be initialized.
3337 * @sds: Variable containing the statistics for sd.
3338 * @idle: Idle status of the CPU at which we're performing load-balancing.
3340 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3341 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3344 * Busy processors will not participate in power savings
3347 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3348 sds
->power_savings_balance
= 0;
3350 sds
->power_savings_balance
= 1;
3351 sds
->min_nr_running
= ULONG_MAX
;
3352 sds
->leader_nr_running
= 0;
3357 * update_sd_power_savings_stats - Update the power saving stats for a
3358 * sched_domain while performing load balancing.
3360 * @group: sched_group belonging to the sched_domain under consideration.
3361 * @sds: Variable containing the statistics of the sched_domain
3362 * @local_group: Does group contain the CPU for which we're performing
3364 * @sgs: Variable containing the statistics of the group.
3366 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3367 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3370 if (!sds
->power_savings_balance
)
3374 * If the local group is idle or completely loaded
3375 * no need to do power savings balance at this domain
3377 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3378 !sds
->this_nr_running
))
3379 sds
->power_savings_balance
= 0;
3382 * If a group is already running at full capacity or idle,
3383 * don't include that group in power savings calculations
3385 if (!sds
->power_savings_balance
||
3386 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3387 !sgs
->sum_nr_running
)
3391 * Calculate the group which has the least non-idle load.
3392 * This is the group from where we need to pick up the load
3395 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3396 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3397 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3398 sds
->group_min
= group
;
3399 sds
->min_nr_running
= sgs
->sum_nr_running
;
3400 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3401 sgs
->sum_nr_running
;
3405 * Calculate the group which is almost near its
3406 * capacity but still has some space to pick up some load
3407 * from other group and save more power
3409 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3412 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3413 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3414 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3415 sds
->group_leader
= group
;
3416 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3421 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3422 * @sds: Variable containing the statistics of the sched_domain
3423 * under consideration.
3424 * @this_cpu: Cpu at which we're currently performing load-balancing.
3425 * @imbalance: Variable to store the imbalance.
3428 * Check if we have potential to perform some power-savings balance.
3429 * If yes, set the busiest group to be the least loaded group in the
3430 * sched_domain, so that it's CPUs can be put to idle.
3432 * Returns 1 if there is potential to perform power-savings balance.
3435 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3436 int this_cpu
, unsigned long *imbalance
)
3438 if (!sds
->power_savings_balance
)
3441 if (sds
->this != sds
->group_leader
||
3442 sds
->group_leader
== sds
->group_min
)
3445 *imbalance
= sds
->min_load_per_task
;
3446 sds
->busiest
= sds
->group_min
;
3448 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3449 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3450 group_first_cpu(sds
->group_leader
);
3456 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3457 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3458 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3463 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3464 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3469 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3470 int this_cpu
, unsigned long *imbalance
)
3474 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3478 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3479 * @group: sched_group whose statistics are to be updated.
3480 * @this_cpu: Cpu for which load balance is currently performed.
3481 * @idle: Idle status of this_cpu
3482 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3483 * @sd_idle: Idle status of the sched_domain containing group.
3484 * @local_group: Does group contain this_cpu.
3485 * @cpus: Set of cpus considered for load balancing.
3486 * @balance: Should we balance.
3487 * @sgs: variable to hold the statistics for this group.
3489 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3490 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3491 int local_group
, const struct cpumask
*cpus
,
3492 int *balance
, struct sg_lb_stats
*sgs
)
3494 unsigned long load
, max_cpu_load
, min_cpu_load
;
3496 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3497 unsigned long sum_avg_load_per_task
;
3498 unsigned long avg_load_per_task
;
3501 balance_cpu
= group_first_cpu(group
);
3503 /* Tally up the load of all CPUs in the group */
3504 sum_avg_load_per_task
= avg_load_per_task
= 0;
3506 min_cpu_load
= ~0UL;
3508 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3509 struct rq
*rq
= cpu_rq(i
);
3511 if (*sd_idle
&& rq
->nr_running
)
3514 /* Bias balancing toward cpus of our domain */
3516 if (idle_cpu(i
) && !first_idle_cpu
) {
3521 load
= target_load(i
, load_idx
);
3523 load
= source_load(i
, load_idx
);
3524 if (load
> max_cpu_load
)
3525 max_cpu_load
= load
;
3526 if (min_cpu_load
> load
)
3527 min_cpu_load
= load
;
3530 sgs
->group_load
+= load
;
3531 sgs
->sum_nr_running
+= rq
->nr_running
;
3532 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3534 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3538 * First idle cpu or the first cpu(busiest) in this sched group
3539 * is eligible for doing load balancing at this and above
3540 * domains. In the newly idle case, we will allow all the cpu's
3541 * to do the newly idle load balance.
3543 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3544 balance_cpu
!= this_cpu
&& balance
) {
3549 /* Adjust by relative CPU power of the group */
3550 sgs
->avg_load
= sg_div_cpu_power(group
,
3551 sgs
->group_load
* SCHED_LOAD_SCALE
);
3555 * Consider the group unbalanced when the imbalance is larger
3556 * than the average weight of two tasks.
3558 * APZ: with cgroup the avg task weight can vary wildly and
3559 * might not be a suitable number - should we keep a
3560 * normalized nr_running number somewhere that negates
3563 avg_load_per_task
= sg_div_cpu_power(group
,
3564 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3566 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3569 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3574 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3575 * @sd: sched_domain whose statistics are to be updated.
3576 * @this_cpu: Cpu for which load balance is currently performed.
3577 * @idle: Idle status of this_cpu
3578 * @sd_idle: Idle status of the sched_domain containing group.
3579 * @cpus: Set of cpus considered for load balancing.
3580 * @balance: Should we balance.
3581 * @sds: variable to hold the statistics for this sched_domain.
3583 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3584 enum cpu_idle_type idle
, int *sd_idle
,
3585 const struct cpumask
*cpus
, int *balance
,
3586 struct sd_lb_stats
*sds
)
3588 struct sched_group
*group
= sd
->groups
;
3589 struct sg_lb_stats sgs
;
3592 init_sd_power_savings_stats(sd
, sds
, idle
);
3593 load_idx
= get_sd_load_idx(sd
, idle
);
3598 local_group
= cpumask_test_cpu(this_cpu
,
3599 sched_group_cpus(group
));
3600 memset(&sgs
, 0, sizeof(sgs
));
3601 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3602 local_group
, cpus
, balance
, &sgs
);
3604 if (local_group
&& balance
&& !(*balance
))
3607 sds
->total_load
+= sgs
.group_load
;
3608 sds
->total_pwr
+= group
->__cpu_power
;
3611 sds
->this_load
= sgs
.avg_load
;
3613 sds
->this_nr_running
= sgs
.sum_nr_running
;
3614 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3615 } else if (sgs
.avg_load
> sds
->max_load
&&
3616 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3618 sds
->max_load
= sgs
.avg_load
;
3619 sds
->busiest
= group
;
3620 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3621 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3622 sds
->group_imb
= sgs
.group_imb
;
3625 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3626 group
= group
->next
;
3627 } while (group
!= sd
->groups
);
3632 * fix_small_imbalance - Calculate the minor imbalance that exists
3633 * amongst the groups of a sched_domain, during
3635 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3636 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3637 * @imbalance: Variable to store the imbalance.
3639 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3640 int this_cpu
, unsigned long *imbalance
)
3642 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3643 unsigned int imbn
= 2;
3645 if (sds
->this_nr_running
) {
3646 sds
->this_load_per_task
/= sds
->this_nr_running
;
3647 if (sds
->busiest_load_per_task
>
3648 sds
->this_load_per_task
)
3651 sds
->this_load_per_task
=
3652 cpu_avg_load_per_task(this_cpu
);
3654 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3655 sds
->busiest_load_per_task
* imbn
) {
3656 *imbalance
= sds
->busiest_load_per_task
;
3661 * OK, we don't have enough imbalance to justify moving tasks,
3662 * however we may be able to increase total CPU power used by
3666 pwr_now
+= sds
->busiest
->__cpu_power
*
3667 min(sds
->busiest_load_per_task
, sds
->max_load
);
3668 pwr_now
+= sds
->this->__cpu_power
*
3669 min(sds
->this_load_per_task
, sds
->this_load
);
3670 pwr_now
/= SCHED_LOAD_SCALE
;
3672 /* Amount of load we'd subtract */
3673 tmp
= sg_div_cpu_power(sds
->busiest
,
3674 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3675 if (sds
->max_load
> tmp
)
3676 pwr_move
+= sds
->busiest
->__cpu_power
*
3677 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3679 /* Amount of load we'd add */
3680 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3681 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3682 tmp
= sg_div_cpu_power(sds
->this,
3683 sds
->max_load
* sds
->busiest
->__cpu_power
);
3685 tmp
= sg_div_cpu_power(sds
->this,
3686 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3687 pwr_move
+= sds
->this->__cpu_power
*
3688 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3689 pwr_move
/= SCHED_LOAD_SCALE
;
3691 /* Move if we gain throughput */
3692 if (pwr_move
> pwr_now
)
3693 *imbalance
= sds
->busiest_load_per_task
;
3697 * calculate_imbalance - Calculate the amount of imbalance present within the
3698 * groups of a given sched_domain during load balance.
3699 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3700 * @this_cpu: Cpu for which currently load balance is being performed.
3701 * @imbalance: The variable to store the imbalance.
3703 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3704 unsigned long *imbalance
)
3706 unsigned long max_pull
;
3708 * In the presence of smp nice balancing, certain scenarios can have
3709 * max load less than avg load(as we skip the groups at or below
3710 * its cpu_power, while calculating max_load..)
3712 if (sds
->max_load
< sds
->avg_load
) {
3714 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3717 /* Don't want to pull so many tasks that a group would go idle */
3718 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3719 sds
->max_load
- sds
->busiest_load_per_task
);
3721 /* How much load to actually move to equalise the imbalance */
3722 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3723 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3727 * if *imbalance is less than the average load per runnable task
3728 * there is no gaurantee that any tasks will be moved so we'll have
3729 * a think about bumping its value to force at least one task to be
3732 if (*imbalance
< sds
->busiest_load_per_task
)
3733 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3736 /******* find_busiest_group() helpers end here *********************/
3739 * find_busiest_group - Returns the busiest group within the sched_domain
3740 * if there is an imbalance. If there isn't an imbalance, and
3741 * the user has opted for power-savings, it returns a group whose
3742 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3743 * such a group exists.
3745 * Also calculates the amount of weighted load which should be moved
3746 * to restore balance.
3748 * @sd: The sched_domain whose busiest group is to be returned.
3749 * @this_cpu: The cpu for which load balancing is currently being performed.
3750 * @imbalance: Variable which stores amount of weighted load which should
3751 * be moved to restore balance/put a group to idle.
3752 * @idle: The idle status of this_cpu.
3753 * @sd_idle: The idleness of sd
3754 * @cpus: The set of CPUs under consideration for load-balancing.
3755 * @balance: Pointer to a variable indicating if this_cpu
3756 * is the appropriate cpu to perform load balancing at this_level.
3758 * Returns: - the busiest group if imbalance exists.
3759 * - If no imbalance and user has opted for power-savings balance,
3760 * return the least loaded group whose CPUs can be
3761 * put to idle by rebalancing its tasks onto our group.
3763 static struct sched_group
*
3764 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3765 unsigned long *imbalance
, enum cpu_idle_type idle
,
3766 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3768 struct sd_lb_stats sds
;
3770 memset(&sds
, 0, sizeof(sds
));
3773 * Compute the various statistics relavent for load balancing at
3776 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3779 /* Cases where imbalance does not exist from POV of this_cpu */
3780 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3782 * 2) There is no busy sibling group to pull from.
3783 * 3) This group is the busiest group.
3784 * 4) This group is more busy than the avg busieness at this
3786 * 5) The imbalance is within the specified limit.
3787 * 6) Any rebalance would lead to ping-pong
3789 if (balance
&& !(*balance
))
3792 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3795 if (sds
.this_load
>= sds
.max_load
)
3798 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3800 if (sds
.this_load
>= sds
.avg_load
)
3803 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3806 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3808 sds
.busiest_load_per_task
=
3809 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3812 * We're trying to get all the cpus to the average_load, so we don't
3813 * want to push ourselves above the average load, nor do we wish to
3814 * reduce the max loaded cpu below the average load, as either of these
3815 * actions would just result in more rebalancing later, and ping-pong
3816 * tasks around. Thus we look for the minimum possible imbalance.
3817 * Negative imbalances (*we* are more loaded than anyone else) will
3818 * be counted as no imbalance for these purposes -- we can't fix that
3819 * by pulling tasks to us. Be careful of negative numbers as they'll
3820 * appear as very large values with unsigned longs.
3822 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3825 /* Looks like there is an imbalance. Compute it */
3826 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3831 * There is no obvious imbalance. But check if we can do some balancing
3834 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3842 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3845 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3846 unsigned long imbalance
, const struct cpumask
*cpus
)
3848 struct rq
*busiest
= NULL
, *rq
;
3849 unsigned long max_load
= 0;
3852 for_each_cpu(i
, sched_group_cpus(group
)) {
3855 if (!cpumask_test_cpu(i
, cpus
))
3859 wl
= weighted_cpuload(i
);
3861 if (rq
->nr_running
== 1 && wl
> imbalance
)
3864 if (wl
> max_load
) {
3874 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3875 * so long as it is large enough.
3877 #define MAX_PINNED_INTERVAL 512
3879 /* Working cpumask for load_balance and load_balance_newidle. */
3880 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3883 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3884 * tasks if there is an imbalance.
3886 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3887 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3890 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3891 struct sched_group
*group
;
3892 unsigned long imbalance
;
3894 unsigned long flags
;
3895 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3897 cpumask_setall(cpus
);
3900 * When power savings policy is enabled for the parent domain, idle
3901 * sibling can pick up load irrespective of busy siblings. In this case,
3902 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3903 * portraying it as CPU_NOT_IDLE.
3905 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3906 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3909 schedstat_inc(sd
, lb_count
[idle
]);
3913 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3920 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3924 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3926 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3930 BUG_ON(busiest
== this_rq
);
3932 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3935 if (busiest
->nr_running
> 1) {
3937 * Attempt to move tasks. If find_busiest_group has found
3938 * an imbalance but busiest->nr_running <= 1, the group is
3939 * still unbalanced. ld_moved simply stays zero, so it is
3940 * correctly treated as an imbalance.
3942 local_irq_save(flags
);
3943 double_rq_lock(this_rq
, busiest
);
3944 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3945 imbalance
, sd
, idle
, &all_pinned
);
3946 double_rq_unlock(this_rq
, busiest
);
3947 local_irq_restore(flags
);
3950 * some other cpu did the load balance for us.
3952 if (ld_moved
&& this_cpu
!= smp_processor_id())
3953 resched_cpu(this_cpu
);
3955 /* All tasks on this runqueue were pinned by CPU affinity */
3956 if (unlikely(all_pinned
)) {
3957 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3958 if (!cpumask_empty(cpus
))
3965 schedstat_inc(sd
, lb_failed
[idle
]);
3966 sd
->nr_balance_failed
++;
3968 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3970 spin_lock_irqsave(&busiest
->lock
, flags
);
3972 /* don't kick the migration_thread, if the curr
3973 * task on busiest cpu can't be moved to this_cpu
3975 if (!cpumask_test_cpu(this_cpu
,
3976 &busiest
->curr
->cpus_allowed
)) {
3977 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3979 goto out_one_pinned
;
3982 if (!busiest
->active_balance
) {
3983 busiest
->active_balance
= 1;
3984 busiest
->push_cpu
= this_cpu
;
3987 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3989 wake_up_process(busiest
->migration_thread
);
3992 * We've kicked active balancing, reset the failure
3995 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3998 sd
->nr_balance_failed
= 0;
4000 if (likely(!active_balance
)) {
4001 /* We were unbalanced, so reset the balancing interval */
4002 sd
->balance_interval
= sd
->min_interval
;
4005 * If we've begun active balancing, start to back off. This
4006 * case may not be covered by the all_pinned logic if there
4007 * is only 1 task on the busy runqueue (because we don't call
4010 if (sd
->balance_interval
< sd
->max_interval
)
4011 sd
->balance_interval
*= 2;
4014 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4015 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4021 schedstat_inc(sd
, lb_balanced
[idle
]);
4023 sd
->nr_balance_failed
= 0;
4026 /* tune up the balancing interval */
4027 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4028 (sd
->balance_interval
< sd
->max_interval
))
4029 sd
->balance_interval
*= 2;
4031 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4032 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4043 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4044 * tasks if there is an imbalance.
4046 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4047 * this_rq is locked.
4050 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4052 struct sched_group
*group
;
4053 struct rq
*busiest
= NULL
;
4054 unsigned long imbalance
;
4058 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4060 cpumask_setall(cpus
);
4063 * When power savings policy is enabled for the parent domain, idle
4064 * sibling can pick up load irrespective of busy siblings. In this case,
4065 * let the state of idle sibling percolate up as IDLE, instead of
4066 * portraying it as CPU_NOT_IDLE.
4068 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4069 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4072 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4074 update_shares_locked(this_rq
, sd
);
4075 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4076 &sd_idle
, cpus
, NULL
);
4078 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4082 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4084 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4088 BUG_ON(busiest
== this_rq
);
4090 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4093 if (busiest
->nr_running
> 1) {
4094 /* Attempt to move tasks */
4095 double_lock_balance(this_rq
, busiest
);
4096 /* this_rq->clock is already updated */
4097 update_rq_clock(busiest
);
4098 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4099 imbalance
, sd
, CPU_NEWLY_IDLE
,
4101 double_unlock_balance(this_rq
, busiest
);
4103 if (unlikely(all_pinned
)) {
4104 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4105 if (!cpumask_empty(cpus
))
4111 int active_balance
= 0;
4113 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4114 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4115 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4118 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4121 if (sd
->nr_balance_failed
++ < 2)
4125 * The only task running in a non-idle cpu can be moved to this
4126 * cpu in an attempt to completely freeup the other CPU
4127 * package. The same method used to move task in load_balance()
4128 * have been extended for load_balance_newidle() to speedup
4129 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4131 * The package power saving logic comes from
4132 * find_busiest_group(). If there are no imbalance, then
4133 * f_b_g() will return NULL. However when sched_mc={1,2} then
4134 * f_b_g() will select a group from which a running task may be
4135 * pulled to this cpu in order to make the other package idle.
4136 * If there is no opportunity to make a package idle and if
4137 * there are no imbalance, then f_b_g() will return NULL and no
4138 * action will be taken in load_balance_newidle().
4140 * Under normal task pull operation due to imbalance, there
4141 * will be more than one task in the source run queue and
4142 * move_tasks() will succeed. ld_moved will be true and this
4143 * active balance code will not be triggered.
4146 /* Lock busiest in correct order while this_rq is held */
4147 double_lock_balance(this_rq
, busiest
);
4150 * don't kick the migration_thread, if the curr
4151 * task on busiest cpu can't be moved to this_cpu
4153 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4154 double_unlock_balance(this_rq
, busiest
);
4159 if (!busiest
->active_balance
) {
4160 busiest
->active_balance
= 1;
4161 busiest
->push_cpu
= this_cpu
;
4165 double_unlock_balance(this_rq
, busiest
);
4167 * Should not call ttwu while holding a rq->lock
4169 spin_unlock(&this_rq
->lock
);
4171 wake_up_process(busiest
->migration_thread
);
4172 spin_lock(&this_rq
->lock
);
4175 sd
->nr_balance_failed
= 0;
4177 update_shares_locked(this_rq
, sd
);
4181 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4182 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4183 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4185 sd
->nr_balance_failed
= 0;
4191 * idle_balance is called by schedule() if this_cpu is about to become
4192 * idle. Attempts to pull tasks from other CPUs.
4194 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4196 struct sched_domain
*sd
;
4197 int pulled_task
= 0;
4198 unsigned long next_balance
= jiffies
+ HZ
;
4200 for_each_domain(this_cpu
, sd
) {
4201 unsigned long interval
;
4203 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4206 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4207 /* If we've pulled tasks over stop searching: */
4208 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4211 interval
= msecs_to_jiffies(sd
->balance_interval
);
4212 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4213 next_balance
= sd
->last_balance
+ interval
;
4217 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4219 * We are going idle. next_balance may be set based on
4220 * a busy processor. So reset next_balance.
4222 this_rq
->next_balance
= next_balance
;
4227 * active_load_balance is run by migration threads. It pushes running tasks
4228 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4229 * running on each physical CPU where possible, and avoids physical /
4230 * logical imbalances.
4232 * Called with busiest_rq locked.
4234 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4236 int target_cpu
= busiest_rq
->push_cpu
;
4237 struct sched_domain
*sd
;
4238 struct rq
*target_rq
;
4240 /* Is there any task to move? */
4241 if (busiest_rq
->nr_running
<= 1)
4244 target_rq
= cpu_rq(target_cpu
);
4247 * This condition is "impossible", if it occurs
4248 * we need to fix it. Originally reported by
4249 * Bjorn Helgaas on a 128-cpu setup.
4251 BUG_ON(busiest_rq
== target_rq
);
4253 /* move a task from busiest_rq to target_rq */
4254 double_lock_balance(busiest_rq
, target_rq
);
4255 update_rq_clock(busiest_rq
);
4256 update_rq_clock(target_rq
);
4258 /* Search for an sd spanning us and the target CPU. */
4259 for_each_domain(target_cpu
, sd
) {
4260 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4261 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4266 schedstat_inc(sd
, alb_count
);
4268 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4270 schedstat_inc(sd
, alb_pushed
);
4272 schedstat_inc(sd
, alb_failed
);
4274 double_unlock_balance(busiest_rq
, target_rq
);
4279 atomic_t load_balancer
;
4280 cpumask_var_t cpu_mask
;
4281 } nohz ____cacheline_aligned
= {
4282 .load_balancer
= ATOMIC_INIT(-1),
4286 * This routine will try to nominate the ilb (idle load balancing)
4287 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4288 * load balancing on behalf of all those cpus. If all the cpus in the system
4289 * go into this tickless mode, then there will be no ilb owner (as there is
4290 * no need for one) and all the cpus will sleep till the next wakeup event
4293 * For the ilb owner, tick is not stopped. And this tick will be used
4294 * for idle load balancing. ilb owner will still be part of
4297 * While stopping the tick, this cpu will become the ilb owner if there
4298 * is no other owner. And will be the owner till that cpu becomes busy
4299 * or if all cpus in the system stop their ticks at which point
4300 * there is no need for ilb owner.
4302 * When the ilb owner becomes busy, it nominates another owner, during the
4303 * next busy scheduler_tick()
4305 int select_nohz_load_balancer(int stop_tick
)
4307 int cpu
= smp_processor_id();
4310 cpu_rq(cpu
)->in_nohz_recently
= 1;
4312 if (!cpu_active(cpu
)) {
4313 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4317 * If we are going offline and still the leader,
4320 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4326 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4328 /* time for ilb owner also to sleep */
4329 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4330 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4331 atomic_set(&nohz
.load_balancer
, -1);
4335 if (atomic_read(&nohz
.load_balancer
) == -1) {
4336 /* make me the ilb owner */
4337 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4339 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4342 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4345 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4347 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4348 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4355 static DEFINE_SPINLOCK(balancing
);
4358 * It checks each scheduling domain to see if it is due to be balanced,
4359 * and initiates a balancing operation if so.
4361 * Balancing parameters are set up in arch_init_sched_domains.
4363 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4366 struct rq
*rq
= cpu_rq(cpu
);
4367 unsigned long interval
;
4368 struct sched_domain
*sd
;
4369 /* Earliest time when we have to do rebalance again */
4370 unsigned long next_balance
= jiffies
+ 60*HZ
;
4371 int update_next_balance
= 0;
4374 for_each_domain(cpu
, sd
) {
4375 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4378 interval
= sd
->balance_interval
;
4379 if (idle
!= CPU_IDLE
)
4380 interval
*= sd
->busy_factor
;
4382 /* scale ms to jiffies */
4383 interval
= msecs_to_jiffies(interval
);
4384 if (unlikely(!interval
))
4386 if (interval
> HZ
*NR_CPUS
/10)
4387 interval
= HZ
*NR_CPUS
/10;
4389 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4391 if (need_serialize
) {
4392 if (!spin_trylock(&balancing
))
4396 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4397 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4399 * We've pulled tasks over so either we're no
4400 * longer idle, or one of our SMT siblings is
4403 idle
= CPU_NOT_IDLE
;
4405 sd
->last_balance
= jiffies
;
4408 spin_unlock(&balancing
);
4410 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4411 next_balance
= sd
->last_balance
+ interval
;
4412 update_next_balance
= 1;
4416 * Stop the load balance at this level. There is another
4417 * CPU in our sched group which is doing load balancing more
4425 * next_balance will be updated only when there is a need.
4426 * When the cpu is attached to null domain for ex, it will not be
4429 if (likely(update_next_balance
))
4430 rq
->next_balance
= next_balance
;
4434 * run_rebalance_domains is triggered when needed from the scheduler tick.
4435 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4436 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4438 static void run_rebalance_domains(struct softirq_action
*h
)
4440 int this_cpu
= smp_processor_id();
4441 struct rq
*this_rq
= cpu_rq(this_cpu
);
4442 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4443 CPU_IDLE
: CPU_NOT_IDLE
;
4445 rebalance_domains(this_cpu
, idle
);
4449 * If this cpu is the owner for idle load balancing, then do the
4450 * balancing on behalf of the other idle cpus whose ticks are
4453 if (this_rq
->idle_at_tick
&&
4454 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4458 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4459 if (balance_cpu
== this_cpu
)
4463 * If this cpu gets work to do, stop the load balancing
4464 * work being done for other cpus. Next load
4465 * balancing owner will pick it up.
4470 rebalance_domains(balance_cpu
, CPU_IDLE
);
4472 rq
= cpu_rq(balance_cpu
);
4473 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4474 this_rq
->next_balance
= rq
->next_balance
;
4480 static inline int on_null_domain(int cpu
)
4482 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4486 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4488 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4489 * idle load balancing owner or decide to stop the periodic load balancing,
4490 * if the whole system is idle.
4492 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4496 * If we were in the nohz mode recently and busy at the current
4497 * scheduler tick, then check if we need to nominate new idle
4500 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4501 rq
->in_nohz_recently
= 0;
4503 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4504 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4505 atomic_set(&nohz
.load_balancer
, -1);
4508 if (atomic_read(&nohz
.load_balancer
) == -1) {
4510 * simple selection for now: Nominate the
4511 * first cpu in the nohz list to be the next
4514 * TBD: Traverse the sched domains and nominate
4515 * the nearest cpu in the nohz.cpu_mask.
4517 int ilb
= cpumask_first(nohz
.cpu_mask
);
4519 if (ilb
< nr_cpu_ids
)
4525 * If this cpu is idle and doing idle load balancing for all the
4526 * cpus with ticks stopped, is it time for that to stop?
4528 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4529 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4535 * If this cpu is idle and the idle load balancing is done by
4536 * someone else, then no need raise the SCHED_SOFTIRQ
4538 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4539 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4542 /* Don't need to rebalance while attached to NULL domain */
4543 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4544 likely(!on_null_domain(cpu
)))
4545 raise_softirq(SCHED_SOFTIRQ
);
4548 #else /* CONFIG_SMP */
4551 * on UP we do not need to balance between CPUs:
4553 static inline void idle_balance(int cpu
, struct rq
*rq
)
4559 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4561 EXPORT_PER_CPU_SYMBOL(kstat
);
4564 * Return any ns on the sched_clock that have not yet been accounted in
4565 * @p in case that task is currently running.
4567 * Called with task_rq_lock() held on @rq.
4569 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4573 if (task_current(rq
, p
)) {
4574 update_rq_clock(rq
);
4575 ns
= rq
->clock
- p
->se
.exec_start
;
4583 unsigned long long task_delta_exec(struct task_struct
*p
)
4585 unsigned long flags
;
4589 rq
= task_rq_lock(p
, &flags
);
4590 ns
= do_task_delta_exec(p
, rq
);
4591 task_rq_unlock(rq
, &flags
);
4597 * Return accounted runtime for the task.
4598 * In case the task is currently running, return the runtime plus current's
4599 * pending runtime that have not been accounted yet.
4601 unsigned long long task_sched_runtime(struct task_struct
*p
)
4603 unsigned long flags
;
4607 rq
= task_rq_lock(p
, &flags
);
4608 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4609 task_rq_unlock(rq
, &flags
);
4615 * Return sum_exec_runtime for the thread group.
4616 * In case the task is currently running, return the sum plus current's
4617 * pending runtime that have not been accounted yet.
4619 * Note that the thread group might have other running tasks as well,
4620 * so the return value not includes other pending runtime that other
4621 * running tasks might have.
4623 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4625 struct task_cputime totals
;
4626 unsigned long flags
;
4630 rq
= task_rq_lock(p
, &flags
);
4631 thread_group_cputime(p
, &totals
);
4632 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4633 task_rq_unlock(rq
, &flags
);
4639 * Account user cpu time to a process.
4640 * @p: the process that the cpu time gets accounted to
4641 * @cputime: the cpu time spent in user space since the last update
4642 * @cputime_scaled: cputime scaled by cpu frequency
4644 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4645 cputime_t cputime_scaled
)
4647 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4650 /* Add user time to process. */
4651 p
->utime
= cputime_add(p
->utime
, cputime
);
4652 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4653 account_group_user_time(p
, cputime
);
4655 /* Add user time to cpustat. */
4656 tmp
= cputime_to_cputime64(cputime
);
4657 if (TASK_NICE(p
) > 0)
4658 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4660 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4662 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4663 /* Account for user time used */
4664 acct_update_integrals(p
);
4668 * Account guest cpu time to a process.
4669 * @p: the process that the cpu time gets accounted to
4670 * @cputime: the cpu time spent in virtual machine since the last update
4671 * @cputime_scaled: cputime scaled by cpu frequency
4673 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4674 cputime_t cputime_scaled
)
4677 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4679 tmp
= cputime_to_cputime64(cputime
);
4681 /* Add guest time to process. */
4682 p
->utime
= cputime_add(p
->utime
, cputime
);
4683 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4684 account_group_user_time(p
, cputime
);
4685 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4687 /* Add guest time to cpustat. */
4688 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4689 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4693 * Account system cpu time to a process.
4694 * @p: the process that the cpu time gets accounted to
4695 * @hardirq_offset: the offset to subtract from hardirq_count()
4696 * @cputime: the cpu time spent in kernel space since the last update
4697 * @cputime_scaled: cputime scaled by cpu frequency
4699 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4700 cputime_t cputime
, cputime_t cputime_scaled
)
4702 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4705 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4706 account_guest_time(p
, cputime
, cputime_scaled
);
4710 /* Add system time to process. */
4711 p
->stime
= cputime_add(p
->stime
, cputime
);
4712 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4713 account_group_system_time(p
, cputime
);
4715 /* Add system time to cpustat. */
4716 tmp
= cputime_to_cputime64(cputime
);
4717 if (hardirq_count() - hardirq_offset
)
4718 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4719 else if (softirq_count())
4720 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4722 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4724 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4726 /* Account for system time used */
4727 acct_update_integrals(p
);
4731 * Account for involuntary wait time.
4732 * @steal: the cpu time spent in involuntary wait
4734 void account_steal_time(cputime_t cputime
)
4736 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4737 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4739 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4743 * Account for idle time.
4744 * @cputime: the cpu time spent in idle wait
4746 void account_idle_time(cputime_t cputime
)
4748 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4749 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4750 struct rq
*rq
= this_rq();
4752 if (atomic_read(&rq
->nr_iowait
) > 0)
4753 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4755 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4758 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4761 * Account a single tick of cpu time.
4762 * @p: the process that the cpu time gets accounted to
4763 * @user_tick: indicates if the tick is a user or a system tick
4765 void account_process_tick(struct task_struct
*p
, int user_tick
)
4767 cputime_t one_jiffy
= jiffies_to_cputime(1);
4768 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4769 struct rq
*rq
= this_rq();
4772 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4773 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
4774 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4777 account_idle_time(one_jiffy
);
4781 * Account multiple ticks of steal time.
4782 * @p: the process from which the cpu time has been stolen
4783 * @ticks: number of stolen ticks
4785 void account_steal_ticks(unsigned long ticks
)
4787 account_steal_time(jiffies_to_cputime(ticks
));
4791 * Account multiple ticks of idle time.
4792 * @ticks: number of stolen ticks
4794 void account_idle_ticks(unsigned long ticks
)
4796 account_idle_time(jiffies_to_cputime(ticks
));
4802 * Use precise platform statistics if available:
4804 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4805 cputime_t
task_utime(struct task_struct
*p
)
4810 cputime_t
task_stime(struct task_struct
*p
)
4815 cputime_t
task_utime(struct task_struct
*p
)
4817 clock_t utime
= cputime_to_clock_t(p
->utime
),
4818 total
= utime
+ cputime_to_clock_t(p
->stime
);
4822 * Use CFS's precise accounting:
4824 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4828 do_div(temp
, total
);
4830 utime
= (clock_t)temp
;
4832 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4833 return p
->prev_utime
;
4836 cputime_t
task_stime(struct task_struct
*p
)
4841 * Use CFS's precise accounting. (we subtract utime from
4842 * the total, to make sure the total observed by userspace
4843 * grows monotonically - apps rely on that):
4845 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4846 cputime_to_clock_t(task_utime(p
));
4849 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4851 return p
->prev_stime
;
4855 inline cputime_t
task_gtime(struct task_struct
*p
)
4861 * This function gets called by the timer code, with HZ frequency.
4862 * We call it with interrupts disabled.
4864 * It also gets called by the fork code, when changing the parent's
4867 void scheduler_tick(void)
4869 int cpu
= smp_processor_id();
4870 struct rq
*rq
= cpu_rq(cpu
);
4871 struct task_struct
*curr
= rq
->curr
;
4875 spin_lock(&rq
->lock
);
4876 update_rq_clock(rq
);
4877 update_cpu_load(rq
);
4878 curr
->sched_class
->task_tick(rq
, curr
, 0);
4879 spin_unlock(&rq
->lock
);
4881 perf_counter_task_tick(curr
, cpu
);
4884 rq
->idle_at_tick
= idle_cpu(cpu
);
4885 trigger_load_balance(rq
, cpu
);
4889 notrace
unsigned long get_parent_ip(unsigned long addr
)
4891 if (in_lock_functions(addr
)) {
4892 addr
= CALLER_ADDR2
;
4893 if (in_lock_functions(addr
))
4894 addr
= CALLER_ADDR3
;
4899 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4900 defined(CONFIG_PREEMPT_TRACER))
4902 void __kprobes
add_preempt_count(int val
)
4904 #ifdef CONFIG_DEBUG_PREEMPT
4908 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4911 preempt_count() += val
;
4912 #ifdef CONFIG_DEBUG_PREEMPT
4914 * Spinlock count overflowing soon?
4916 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4919 if (preempt_count() == val
)
4920 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4922 EXPORT_SYMBOL(add_preempt_count
);
4924 void __kprobes
sub_preempt_count(int val
)
4926 #ifdef CONFIG_DEBUG_PREEMPT
4930 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4933 * Is the spinlock portion underflowing?
4935 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4936 !(preempt_count() & PREEMPT_MASK
)))
4940 if (preempt_count() == val
)
4941 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4942 preempt_count() -= val
;
4944 EXPORT_SYMBOL(sub_preempt_count
);
4949 * Print scheduling while atomic bug:
4951 static noinline
void __schedule_bug(struct task_struct
*prev
)
4953 struct pt_regs
*regs
= get_irq_regs();
4955 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4956 prev
->comm
, prev
->pid
, preempt_count());
4958 debug_show_held_locks(prev
);
4960 if (irqs_disabled())
4961 print_irqtrace_events(prev
);
4970 * Various schedule()-time debugging checks and statistics:
4972 static inline void schedule_debug(struct task_struct
*prev
)
4975 * Test if we are atomic. Since do_exit() needs to call into
4976 * schedule() atomically, we ignore that path for now.
4977 * Otherwise, whine if we are scheduling when we should not be.
4979 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4980 __schedule_bug(prev
);
4982 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4984 schedstat_inc(this_rq(), sched_count
);
4985 #ifdef CONFIG_SCHEDSTATS
4986 if (unlikely(prev
->lock_depth
>= 0)) {
4987 schedstat_inc(this_rq(), bkl_count
);
4988 schedstat_inc(prev
, sched_info
.bkl_count
);
4993 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4995 if (prev
->state
== TASK_RUNNING
) {
4996 u64 runtime
= prev
->se
.sum_exec_runtime
;
4998 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4999 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5002 * In order to avoid avg_overlap growing stale when we are
5003 * indeed overlapping and hence not getting put to sleep, grow
5004 * the avg_overlap on preemption.
5006 * We use the average preemption runtime because that
5007 * correlates to the amount of cache footprint a task can
5010 update_avg(&prev
->se
.avg_overlap
, runtime
);
5012 prev
->sched_class
->put_prev_task(rq
, prev
);
5016 * Pick up the highest-prio task:
5018 static inline struct task_struct
*
5019 pick_next_task(struct rq
*rq
)
5021 const struct sched_class
*class;
5022 struct task_struct
*p
;
5025 * Optimization: we know that if all tasks are in
5026 * the fair class we can call that function directly:
5028 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5029 p
= fair_sched_class
.pick_next_task(rq
);
5034 class = sched_class_highest
;
5036 p
= class->pick_next_task(rq
);
5040 * Will never be NULL as the idle class always
5041 * returns a non-NULL p:
5043 class = class->next
;
5048 * schedule() is the main scheduler function.
5050 asmlinkage
void __sched
__schedule(void)
5052 struct task_struct
*prev
, *next
;
5053 unsigned long *switch_count
;
5057 cpu
= smp_processor_id();
5061 switch_count
= &prev
->nivcsw
;
5063 release_kernel_lock(prev
);
5064 need_resched_nonpreemptible
:
5066 schedule_debug(prev
);
5068 if (sched_feat(HRTICK
))
5071 spin_lock_irq(&rq
->lock
);
5072 update_rq_clock(rq
);
5073 clear_tsk_need_resched(prev
);
5075 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5076 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5077 prev
->state
= TASK_RUNNING
;
5079 deactivate_task(rq
, prev
, 1);
5080 switch_count
= &prev
->nvcsw
;
5084 if (prev
->sched_class
->pre_schedule
)
5085 prev
->sched_class
->pre_schedule(rq
, prev
);
5088 if (unlikely(!rq
->nr_running
))
5089 idle_balance(cpu
, rq
);
5091 put_prev_task(rq
, prev
);
5092 next
= pick_next_task(rq
);
5094 if (likely(prev
!= next
)) {
5095 sched_info_switch(prev
, next
);
5096 perf_counter_task_sched_out(prev
, next
, cpu
);
5102 context_switch(rq
, prev
, next
); /* unlocks the rq */
5104 * the context switch might have flipped the stack from under
5105 * us, hence refresh the local variables.
5107 cpu
= smp_processor_id();
5110 spin_unlock_irq(&rq
->lock
);
5112 if (unlikely(reacquire_kernel_lock(current
) < 0))
5113 goto need_resched_nonpreemptible
;
5116 asmlinkage
void __sched
schedule(void)
5121 preempt_enable_no_resched();
5122 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5125 EXPORT_SYMBOL(schedule
);
5129 * Look out! "owner" is an entirely speculative pointer
5130 * access and not reliable.
5132 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5137 if (!sched_feat(OWNER_SPIN
))
5140 #ifdef CONFIG_DEBUG_PAGEALLOC
5142 * Need to access the cpu field knowing that
5143 * DEBUG_PAGEALLOC could have unmapped it if
5144 * the mutex owner just released it and exited.
5146 if (probe_kernel_address(&owner
->cpu
, cpu
))
5153 * Even if the access succeeded (likely case),
5154 * the cpu field may no longer be valid.
5156 if (cpu
>= nr_cpumask_bits
)
5160 * We need to validate that we can do a
5161 * get_cpu() and that we have the percpu area.
5163 if (!cpu_online(cpu
))
5170 * Owner changed, break to re-assess state.
5172 if (lock
->owner
!= owner
)
5176 * Is that owner really running on that cpu?
5178 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5188 #ifdef CONFIG_PREEMPT
5190 * this is the entry point to schedule() from in-kernel preemption
5191 * off of preempt_enable. Kernel preemptions off return from interrupt
5192 * occur there and call schedule directly.
5194 asmlinkage
void __sched
preempt_schedule(void)
5196 struct thread_info
*ti
= current_thread_info();
5199 * If there is a non-zero preempt_count or interrupts are disabled,
5200 * we do not want to preempt the current task. Just return..
5202 if (likely(ti
->preempt_count
|| irqs_disabled()))
5206 add_preempt_count(PREEMPT_ACTIVE
);
5208 sub_preempt_count(PREEMPT_ACTIVE
);
5211 * Check again in case we missed a preemption opportunity
5212 * between schedule and now.
5215 } while (need_resched());
5217 EXPORT_SYMBOL(preempt_schedule
);
5220 * this is the entry point to schedule() from kernel preemption
5221 * off of irq context.
5222 * Note, that this is called and return with irqs disabled. This will
5223 * protect us against recursive calling from irq.
5225 asmlinkage
void __sched
preempt_schedule_irq(void)
5227 struct thread_info
*ti
= current_thread_info();
5229 /* Catch callers which need to be fixed */
5230 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5233 add_preempt_count(PREEMPT_ACTIVE
);
5236 local_irq_disable();
5237 sub_preempt_count(PREEMPT_ACTIVE
);
5240 * Check again in case we missed a preemption opportunity
5241 * between schedule and now.
5244 } while (need_resched());
5247 #endif /* CONFIG_PREEMPT */
5249 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5252 return try_to_wake_up(curr
->private, mode
, sync
);
5254 EXPORT_SYMBOL(default_wake_function
);
5257 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5258 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5259 * number) then we wake all the non-exclusive tasks and one exclusive task.
5261 * There are circumstances in which we can try to wake a task which has already
5262 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5263 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5265 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5266 int nr_exclusive
, int sync
, void *key
)
5268 wait_queue_t
*curr
, *next
;
5270 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5271 unsigned flags
= curr
->flags
;
5273 if (curr
->func(curr
, mode
, sync
, key
) &&
5274 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5280 * __wake_up - wake up threads blocked on a waitqueue.
5282 * @mode: which threads
5283 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5284 * @key: is directly passed to the wakeup function
5286 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5287 int nr_exclusive
, void *key
)
5289 unsigned long flags
;
5291 spin_lock_irqsave(&q
->lock
, flags
);
5292 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5293 spin_unlock_irqrestore(&q
->lock
, flags
);
5295 EXPORT_SYMBOL(__wake_up
);
5298 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5300 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5302 __wake_up_common(q
, mode
, 1, 0, NULL
);
5305 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5307 __wake_up_common(q
, mode
, 1, 0, key
);
5311 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5313 * @mode: which threads
5314 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5315 * @key: opaque value to be passed to wakeup targets
5317 * The sync wakeup differs that the waker knows that it will schedule
5318 * away soon, so while the target thread will be woken up, it will not
5319 * be migrated to another CPU - ie. the two threads are 'synchronized'
5320 * with each other. This can prevent needless bouncing between CPUs.
5322 * On UP it can prevent extra preemption.
5324 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5325 int nr_exclusive
, void *key
)
5327 unsigned long flags
;
5333 if (unlikely(!nr_exclusive
))
5336 spin_lock_irqsave(&q
->lock
, flags
);
5337 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5338 spin_unlock_irqrestore(&q
->lock
, flags
);
5340 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5343 * __wake_up_sync - see __wake_up_sync_key()
5345 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5347 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5349 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5352 * complete: - signals a single thread waiting on this completion
5353 * @x: holds the state of this particular completion
5355 * This will wake up a single thread waiting on this completion. Threads will be
5356 * awakened in the same order in which they were queued.
5358 * See also complete_all(), wait_for_completion() and related routines.
5360 void complete(struct completion
*x
)
5362 unsigned long flags
;
5364 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5366 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5367 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5369 EXPORT_SYMBOL(complete
);
5372 * complete_all: - signals all threads waiting on this completion
5373 * @x: holds the state of this particular completion
5375 * This will wake up all threads waiting on this particular completion event.
5377 void complete_all(struct completion
*x
)
5379 unsigned long flags
;
5381 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5382 x
->done
+= UINT_MAX
/2;
5383 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5384 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5386 EXPORT_SYMBOL(complete_all
);
5388 static inline long __sched
5389 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5392 DECLARE_WAITQUEUE(wait
, current
);
5394 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5395 __add_wait_queue_tail(&x
->wait
, &wait
);
5397 if (signal_pending_state(state
, current
)) {
5398 timeout
= -ERESTARTSYS
;
5401 __set_current_state(state
);
5402 spin_unlock_irq(&x
->wait
.lock
);
5403 timeout
= schedule_timeout(timeout
);
5404 spin_lock_irq(&x
->wait
.lock
);
5405 } while (!x
->done
&& timeout
);
5406 __remove_wait_queue(&x
->wait
, &wait
);
5411 return timeout
?: 1;
5415 wait_for_common(struct completion
*x
, long timeout
, int state
)
5419 spin_lock_irq(&x
->wait
.lock
);
5420 timeout
= do_wait_for_common(x
, timeout
, state
);
5421 spin_unlock_irq(&x
->wait
.lock
);
5426 * wait_for_completion: - waits for completion of a task
5427 * @x: holds the state of this particular completion
5429 * This waits to be signaled for completion of a specific task. It is NOT
5430 * interruptible and there is no timeout.
5432 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5433 * and interrupt capability. Also see complete().
5435 void __sched
wait_for_completion(struct completion
*x
)
5437 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5439 EXPORT_SYMBOL(wait_for_completion
);
5442 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5443 * @x: holds the state of this particular completion
5444 * @timeout: timeout value in jiffies
5446 * This waits for either a completion of a specific task to be signaled or for a
5447 * specified timeout to expire. The timeout is in jiffies. It is not
5450 unsigned long __sched
5451 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5453 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5455 EXPORT_SYMBOL(wait_for_completion_timeout
);
5458 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5459 * @x: holds the state of this particular completion
5461 * This waits for completion of a specific task to be signaled. It is
5464 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5466 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5467 if (t
== -ERESTARTSYS
)
5471 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5474 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5475 * @x: holds the state of this particular completion
5476 * @timeout: timeout value in jiffies
5478 * This waits for either a completion of a specific task to be signaled or for a
5479 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5481 unsigned long __sched
5482 wait_for_completion_interruptible_timeout(struct completion
*x
,
5483 unsigned long timeout
)
5485 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5487 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5490 * wait_for_completion_killable: - waits for completion of a task (killable)
5491 * @x: holds the state of this particular completion
5493 * This waits to be signaled for completion of a specific task. It can be
5494 * interrupted by a kill signal.
5496 int __sched
wait_for_completion_killable(struct completion
*x
)
5498 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5499 if (t
== -ERESTARTSYS
)
5503 EXPORT_SYMBOL(wait_for_completion_killable
);
5506 * try_wait_for_completion - try to decrement a completion without blocking
5507 * @x: completion structure
5509 * Returns: 0 if a decrement cannot be done without blocking
5510 * 1 if a decrement succeeded.
5512 * If a completion is being used as a counting completion,
5513 * attempt to decrement the counter without blocking. This
5514 * enables us to avoid waiting if the resource the completion
5515 * is protecting is not available.
5517 bool try_wait_for_completion(struct completion
*x
)
5521 spin_lock_irq(&x
->wait
.lock
);
5526 spin_unlock_irq(&x
->wait
.lock
);
5529 EXPORT_SYMBOL(try_wait_for_completion
);
5532 * completion_done - Test to see if a completion has any waiters
5533 * @x: completion structure
5535 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5536 * 1 if there are no waiters.
5539 bool completion_done(struct completion
*x
)
5543 spin_lock_irq(&x
->wait
.lock
);
5546 spin_unlock_irq(&x
->wait
.lock
);
5549 EXPORT_SYMBOL(completion_done
);
5552 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5554 unsigned long flags
;
5557 init_waitqueue_entry(&wait
, current
);
5559 __set_current_state(state
);
5561 spin_lock_irqsave(&q
->lock
, flags
);
5562 __add_wait_queue(q
, &wait
);
5563 spin_unlock(&q
->lock
);
5564 timeout
= schedule_timeout(timeout
);
5565 spin_lock_irq(&q
->lock
);
5566 __remove_wait_queue(q
, &wait
);
5567 spin_unlock_irqrestore(&q
->lock
, flags
);
5572 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5574 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5576 EXPORT_SYMBOL(interruptible_sleep_on
);
5579 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5581 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5583 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5585 void __sched
sleep_on(wait_queue_head_t
*q
)
5587 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5589 EXPORT_SYMBOL(sleep_on
);
5591 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5593 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5595 EXPORT_SYMBOL(sleep_on_timeout
);
5597 #ifdef CONFIG_RT_MUTEXES
5600 * rt_mutex_setprio - set the current priority of a task
5602 * @prio: prio value (kernel-internal form)
5604 * This function changes the 'effective' priority of a task. It does
5605 * not touch ->normal_prio like __setscheduler().
5607 * Used by the rt_mutex code to implement priority inheritance logic.
5609 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5611 unsigned long flags
;
5612 int oldprio
, on_rq
, running
;
5614 const struct sched_class
*prev_class
= p
->sched_class
;
5616 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5618 rq
= task_rq_lock(p
, &flags
);
5619 update_rq_clock(rq
);
5622 on_rq
= p
->se
.on_rq
;
5623 running
= task_current(rq
, p
);
5625 dequeue_task(rq
, p
, 0);
5627 p
->sched_class
->put_prev_task(rq
, p
);
5630 p
->sched_class
= &rt_sched_class
;
5632 p
->sched_class
= &fair_sched_class
;
5637 p
->sched_class
->set_curr_task(rq
);
5639 enqueue_task(rq
, p
, 0);
5641 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5643 task_rq_unlock(rq
, &flags
);
5648 void set_user_nice(struct task_struct
*p
, long nice
)
5650 int old_prio
, delta
, on_rq
;
5651 unsigned long flags
;
5654 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5657 * We have to be careful, if called from sys_setpriority(),
5658 * the task might be in the middle of scheduling on another CPU.
5660 rq
= task_rq_lock(p
, &flags
);
5661 update_rq_clock(rq
);
5663 * The RT priorities are set via sched_setscheduler(), but we still
5664 * allow the 'normal' nice value to be set - but as expected
5665 * it wont have any effect on scheduling until the task is
5666 * SCHED_FIFO/SCHED_RR:
5668 if (task_has_rt_policy(p
)) {
5669 p
->static_prio
= NICE_TO_PRIO(nice
);
5672 on_rq
= p
->se
.on_rq
;
5674 dequeue_task(rq
, p
, 0);
5676 p
->static_prio
= NICE_TO_PRIO(nice
);
5679 p
->prio
= effective_prio(p
);
5680 delta
= p
->prio
- old_prio
;
5683 enqueue_task(rq
, p
, 0);
5685 * If the task increased its priority or is running and
5686 * lowered its priority, then reschedule its CPU:
5688 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5689 resched_task(rq
->curr
);
5692 task_rq_unlock(rq
, &flags
);
5694 EXPORT_SYMBOL(set_user_nice
);
5697 * can_nice - check if a task can reduce its nice value
5701 int can_nice(const struct task_struct
*p
, const int nice
)
5703 /* convert nice value [19,-20] to rlimit style value [1,40] */
5704 int nice_rlim
= 20 - nice
;
5706 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5707 capable(CAP_SYS_NICE
));
5710 #ifdef __ARCH_WANT_SYS_NICE
5713 * sys_nice - change the priority of the current process.
5714 * @increment: priority increment
5716 * sys_setpriority is a more generic, but much slower function that
5717 * does similar things.
5719 SYSCALL_DEFINE1(nice
, int, increment
)
5724 * Setpriority might change our priority at the same moment.
5725 * We don't have to worry. Conceptually one call occurs first
5726 * and we have a single winner.
5728 if (increment
< -40)
5733 nice
= TASK_NICE(current
) + increment
;
5739 if (increment
< 0 && !can_nice(current
, nice
))
5742 retval
= security_task_setnice(current
, nice
);
5746 set_user_nice(current
, nice
);
5753 * task_prio - return the priority value of a given task.
5754 * @p: the task in question.
5756 * This is the priority value as seen by users in /proc.
5757 * RT tasks are offset by -200. Normal tasks are centered
5758 * around 0, value goes from -16 to +15.
5760 int task_prio(const struct task_struct
*p
)
5762 return p
->prio
- MAX_RT_PRIO
;
5766 * task_nice - return the nice value of a given task.
5767 * @p: the task in question.
5769 int task_nice(const struct task_struct
*p
)
5771 return TASK_NICE(p
);
5773 EXPORT_SYMBOL(task_nice
);
5776 * idle_cpu - is a given cpu idle currently?
5777 * @cpu: the processor in question.
5779 int idle_cpu(int cpu
)
5781 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5785 * idle_task - return the idle task for a given cpu.
5786 * @cpu: the processor in question.
5788 struct task_struct
*idle_task(int cpu
)
5790 return cpu_rq(cpu
)->idle
;
5794 * find_process_by_pid - find a process with a matching PID value.
5795 * @pid: the pid in question.
5797 static struct task_struct
*find_process_by_pid(pid_t pid
)
5799 return pid
? find_task_by_vpid(pid
) : current
;
5802 /* Actually do priority change: must hold rq lock. */
5804 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5806 BUG_ON(p
->se
.on_rq
);
5809 switch (p
->policy
) {
5813 p
->sched_class
= &fair_sched_class
;
5817 p
->sched_class
= &rt_sched_class
;
5821 p
->rt_priority
= prio
;
5822 p
->normal_prio
= normal_prio(p
);
5823 /* we are holding p->pi_lock already */
5824 p
->prio
= rt_mutex_getprio(p
);
5829 * check the target process has a UID that matches the current process's
5831 static bool check_same_owner(struct task_struct
*p
)
5833 const struct cred
*cred
= current_cred(), *pcred
;
5837 pcred
= __task_cred(p
);
5838 match
= (cred
->euid
== pcred
->euid
||
5839 cred
->euid
== pcred
->uid
);
5844 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5845 struct sched_param
*param
, bool user
)
5847 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5848 unsigned long flags
;
5849 const struct sched_class
*prev_class
= p
->sched_class
;
5852 /* may grab non-irq protected spin_locks */
5853 BUG_ON(in_interrupt());
5855 /* double check policy once rq lock held */
5857 policy
= oldpolicy
= p
->policy
;
5858 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5859 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5860 policy
!= SCHED_IDLE
)
5863 * Valid priorities for SCHED_FIFO and SCHED_RR are
5864 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5865 * SCHED_BATCH and SCHED_IDLE is 0.
5867 if (param
->sched_priority
< 0 ||
5868 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5869 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5871 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5875 * Allow unprivileged RT tasks to decrease priority:
5877 if (user
&& !capable(CAP_SYS_NICE
)) {
5878 if (rt_policy(policy
)) {
5879 unsigned long rlim_rtprio
;
5881 if (!lock_task_sighand(p
, &flags
))
5883 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5884 unlock_task_sighand(p
, &flags
);
5886 /* can't set/change the rt policy */
5887 if (policy
!= p
->policy
&& !rlim_rtprio
)
5890 /* can't increase priority */
5891 if (param
->sched_priority
> p
->rt_priority
&&
5892 param
->sched_priority
> rlim_rtprio
)
5896 * Like positive nice levels, dont allow tasks to
5897 * move out of SCHED_IDLE either:
5899 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5902 /* can't change other user's priorities */
5903 if (!check_same_owner(p
))
5908 #ifdef CONFIG_RT_GROUP_SCHED
5910 * Do not allow realtime tasks into groups that have no runtime
5913 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5914 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5918 retval
= security_task_setscheduler(p
, policy
, param
);
5924 * make sure no PI-waiters arrive (or leave) while we are
5925 * changing the priority of the task:
5927 spin_lock_irqsave(&p
->pi_lock
, flags
);
5929 * To be able to change p->policy safely, the apropriate
5930 * runqueue lock must be held.
5932 rq
= __task_rq_lock(p
);
5933 /* recheck policy now with rq lock held */
5934 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5935 policy
= oldpolicy
= -1;
5936 __task_rq_unlock(rq
);
5937 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5940 update_rq_clock(rq
);
5941 on_rq
= p
->se
.on_rq
;
5942 running
= task_current(rq
, p
);
5944 deactivate_task(rq
, p
, 0);
5946 p
->sched_class
->put_prev_task(rq
, p
);
5949 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5952 p
->sched_class
->set_curr_task(rq
);
5954 activate_task(rq
, p
, 0);
5956 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5958 __task_rq_unlock(rq
);
5959 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5961 rt_mutex_adjust_pi(p
);
5967 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5968 * @p: the task in question.
5969 * @policy: new policy.
5970 * @param: structure containing the new RT priority.
5972 * NOTE that the task may be already dead.
5974 int sched_setscheduler(struct task_struct
*p
, int policy
,
5975 struct sched_param
*param
)
5977 return __sched_setscheduler(p
, policy
, param
, true);
5979 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5982 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5983 * @p: the task in question.
5984 * @policy: new policy.
5985 * @param: structure containing the new RT priority.
5987 * Just like sched_setscheduler, only don't bother checking if the
5988 * current context has permission. For example, this is needed in
5989 * stop_machine(): we create temporary high priority worker threads,
5990 * but our caller might not have that capability.
5992 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5993 struct sched_param
*param
)
5995 return __sched_setscheduler(p
, policy
, param
, false);
5999 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6001 struct sched_param lparam
;
6002 struct task_struct
*p
;
6005 if (!param
|| pid
< 0)
6007 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6012 p
= find_process_by_pid(pid
);
6014 retval
= sched_setscheduler(p
, policy
, &lparam
);
6021 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6022 * @pid: the pid in question.
6023 * @policy: new policy.
6024 * @param: structure containing the new RT priority.
6026 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6027 struct sched_param __user
*, param
)
6029 /* negative values for policy are not valid */
6033 return do_sched_setscheduler(pid
, policy
, param
);
6037 * sys_sched_setparam - set/change the RT priority of a thread
6038 * @pid: the pid in question.
6039 * @param: structure containing the new RT priority.
6041 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6043 return do_sched_setscheduler(pid
, -1, param
);
6047 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6048 * @pid: the pid in question.
6050 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6052 struct task_struct
*p
;
6059 read_lock(&tasklist_lock
);
6060 p
= find_process_by_pid(pid
);
6062 retval
= security_task_getscheduler(p
);
6066 read_unlock(&tasklist_lock
);
6071 * sys_sched_getscheduler - get the RT priority of a thread
6072 * @pid: the pid in question.
6073 * @param: structure containing the RT priority.
6075 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6077 struct sched_param lp
;
6078 struct task_struct
*p
;
6081 if (!param
|| pid
< 0)
6084 read_lock(&tasklist_lock
);
6085 p
= find_process_by_pid(pid
);
6090 retval
= security_task_getscheduler(p
);
6094 lp
.sched_priority
= p
->rt_priority
;
6095 read_unlock(&tasklist_lock
);
6098 * This one might sleep, we cannot do it with a spinlock held ...
6100 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6105 read_unlock(&tasklist_lock
);
6109 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6111 cpumask_var_t cpus_allowed
, new_mask
;
6112 struct task_struct
*p
;
6116 read_lock(&tasklist_lock
);
6118 p
= find_process_by_pid(pid
);
6120 read_unlock(&tasklist_lock
);
6126 * It is not safe to call set_cpus_allowed with the
6127 * tasklist_lock held. We will bump the task_struct's
6128 * usage count and then drop tasklist_lock.
6131 read_unlock(&tasklist_lock
);
6133 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6137 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6139 goto out_free_cpus_allowed
;
6142 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6145 retval
= security_task_setscheduler(p
, 0, NULL
);
6149 cpuset_cpus_allowed(p
, cpus_allowed
);
6150 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6152 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6155 cpuset_cpus_allowed(p
, cpus_allowed
);
6156 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6158 * We must have raced with a concurrent cpuset
6159 * update. Just reset the cpus_allowed to the
6160 * cpuset's cpus_allowed
6162 cpumask_copy(new_mask
, cpus_allowed
);
6167 free_cpumask_var(new_mask
);
6168 out_free_cpus_allowed
:
6169 free_cpumask_var(cpus_allowed
);
6176 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6177 struct cpumask
*new_mask
)
6179 if (len
< cpumask_size())
6180 cpumask_clear(new_mask
);
6181 else if (len
> cpumask_size())
6182 len
= cpumask_size();
6184 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6188 * sys_sched_setaffinity - set the cpu affinity of a process
6189 * @pid: pid of the process
6190 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6191 * @user_mask_ptr: user-space pointer to the new cpu mask
6193 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6194 unsigned long __user
*, user_mask_ptr
)
6196 cpumask_var_t new_mask
;
6199 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6202 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6204 retval
= sched_setaffinity(pid
, new_mask
);
6205 free_cpumask_var(new_mask
);
6209 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6211 struct task_struct
*p
;
6215 read_lock(&tasklist_lock
);
6218 p
= find_process_by_pid(pid
);
6222 retval
= security_task_getscheduler(p
);
6226 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6229 read_unlock(&tasklist_lock
);
6236 * sys_sched_getaffinity - get the cpu affinity of a process
6237 * @pid: pid of the process
6238 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6239 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6241 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6242 unsigned long __user
*, user_mask_ptr
)
6247 if (len
< cpumask_size())
6250 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6253 ret
= sched_getaffinity(pid
, mask
);
6255 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6258 ret
= cpumask_size();
6260 free_cpumask_var(mask
);
6266 * sys_sched_yield - yield the current processor to other threads.
6268 * This function yields the current CPU to other tasks. If there are no
6269 * other threads running on this CPU then this function will return.
6271 SYSCALL_DEFINE0(sched_yield
)
6273 struct rq
*rq
= this_rq_lock();
6275 schedstat_inc(rq
, yld_count
);
6276 current
->sched_class
->yield_task(rq
);
6279 * Since we are going to call schedule() anyway, there's
6280 * no need to preempt or enable interrupts:
6282 __release(rq
->lock
);
6283 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6284 _raw_spin_unlock(&rq
->lock
);
6285 preempt_enable_no_resched();
6292 static void __cond_resched(void)
6294 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6295 __might_sleep(__FILE__
, __LINE__
);
6298 * The BKS might be reacquired before we have dropped
6299 * PREEMPT_ACTIVE, which could trigger a second
6300 * cond_resched() call.
6303 add_preempt_count(PREEMPT_ACTIVE
);
6305 sub_preempt_count(PREEMPT_ACTIVE
);
6306 } while (need_resched());
6309 int __sched
_cond_resched(void)
6311 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6312 system_state
== SYSTEM_RUNNING
) {
6318 EXPORT_SYMBOL(_cond_resched
);
6321 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6322 * call schedule, and on return reacquire the lock.
6324 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6325 * operations here to prevent schedule() from being called twice (once via
6326 * spin_unlock(), once by hand).
6328 int cond_resched_lock(spinlock_t
*lock
)
6330 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6333 if (spin_needbreak(lock
) || resched
) {
6335 if (resched
&& need_resched())
6344 EXPORT_SYMBOL(cond_resched_lock
);
6346 int __sched
cond_resched_softirq(void)
6348 BUG_ON(!in_softirq());
6350 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6358 EXPORT_SYMBOL(cond_resched_softirq
);
6361 * yield - yield the current processor to other threads.
6363 * This is a shortcut for kernel-space yielding - it marks the
6364 * thread runnable and calls sys_sched_yield().
6366 void __sched
yield(void)
6368 set_current_state(TASK_RUNNING
);
6371 EXPORT_SYMBOL(yield
);
6374 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6375 * that process accounting knows that this is a task in IO wait state.
6377 * But don't do that if it is a deliberate, throttling IO wait (this task
6378 * has set its backing_dev_info: the queue against which it should throttle)
6380 void __sched
io_schedule(void)
6382 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6384 delayacct_blkio_start();
6385 atomic_inc(&rq
->nr_iowait
);
6387 atomic_dec(&rq
->nr_iowait
);
6388 delayacct_blkio_end();
6390 EXPORT_SYMBOL(io_schedule
);
6392 long __sched
io_schedule_timeout(long timeout
)
6394 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6397 delayacct_blkio_start();
6398 atomic_inc(&rq
->nr_iowait
);
6399 ret
= schedule_timeout(timeout
);
6400 atomic_dec(&rq
->nr_iowait
);
6401 delayacct_blkio_end();
6406 * sys_sched_get_priority_max - return maximum RT priority.
6407 * @policy: scheduling class.
6409 * this syscall returns the maximum rt_priority that can be used
6410 * by a given scheduling class.
6412 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6419 ret
= MAX_USER_RT_PRIO
-1;
6431 * sys_sched_get_priority_min - return minimum RT priority.
6432 * @policy: scheduling class.
6434 * this syscall returns the minimum rt_priority that can be used
6435 * by a given scheduling class.
6437 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6455 * sys_sched_rr_get_interval - return the default timeslice of a process.
6456 * @pid: pid of the process.
6457 * @interval: userspace pointer to the timeslice value.
6459 * this syscall writes the default timeslice value of a given process
6460 * into the user-space timespec buffer. A value of '0' means infinity.
6462 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6463 struct timespec __user
*, interval
)
6465 struct task_struct
*p
;
6466 unsigned int time_slice
;
6474 read_lock(&tasklist_lock
);
6475 p
= find_process_by_pid(pid
);
6479 retval
= security_task_getscheduler(p
);
6484 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6485 * tasks that are on an otherwise idle runqueue:
6488 if (p
->policy
== SCHED_RR
) {
6489 time_slice
= DEF_TIMESLICE
;
6490 } else if (p
->policy
!= SCHED_FIFO
) {
6491 struct sched_entity
*se
= &p
->se
;
6492 unsigned long flags
;
6495 rq
= task_rq_lock(p
, &flags
);
6496 if (rq
->cfs
.load
.weight
)
6497 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6498 task_rq_unlock(rq
, &flags
);
6500 read_unlock(&tasklist_lock
);
6501 jiffies_to_timespec(time_slice
, &t
);
6502 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6506 read_unlock(&tasklist_lock
);
6510 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6512 void sched_show_task(struct task_struct
*p
)
6514 unsigned long free
= 0;
6517 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6518 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6519 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6520 #if BITS_PER_LONG == 32
6521 if (state
== TASK_RUNNING
)
6522 printk(KERN_CONT
" running ");
6524 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6526 if (state
== TASK_RUNNING
)
6527 printk(KERN_CONT
" running task ");
6529 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6531 #ifdef CONFIG_DEBUG_STACK_USAGE
6532 free
= stack_not_used(p
);
6534 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6535 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6537 show_stack(p
, NULL
);
6540 void show_state_filter(unsigned long state_filter
)
6542 struct task_struct
*g
, *p
;
6544 #if BITS_PER_LONG == 32
6546 " task PC stack pid father\n");
6549 " task PC stack pid father\n");
6551 read_lock(&tasklist_lock
);
6552 do_each_thread(g
, p
) {
6554 * reset the NMI-timeout, listing all files on a slow
6555 * console might take alot of time:
6557 touch_nmi_watchdog();
6558 if (!state_filter
|| (p
->state
& state_filter
))
6560 } while_each_thread(g
, p
);
6562 touch_all_softlockup_watchdogs();
6564 #ifdef CONFIG_SCHED_DEBUG
6565 sysrq_sched_debug_show();
6567 read_unlock(&tasklist_lock
);
6569 * Only show locks if all tasks are dumped:
6571 if (state_filter
== -1)
6572 debug_show_all_locks();
6575 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6577 idle
->sched_class
= &idle_sched_class
;
6581 * init_idle - set up an idle thread for a given CPU
6582 * @idle: task in question
6583 * @cpu: cpu the idle task belongs to
6585 * NOTE: this function does not set the idle thread's NEED_RESCHED
6586 * flag, to make booting more robust.
6588 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6590 struct rq
*rq
= cpu_rq(cpu
);
6591 unsigned long flags
;
6593 spin_lock_irqsave(&rq
->lock
, flags
);
6596 idle
->se
.exec_start
= sched_clock();
6598 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6599 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6600 __set_task_cpu(idle
, cpu
);
6602 rq
->curr
= rq
->idle
= idle
;
6603 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6606 spin_unlock_irqrestore(&rq
->lock
, flags
);
6608 /* Set the preempt count _outside_ the spinlocks! */
6609 #if defined(CONFIG_PREEMPT)
6610 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6612 task_thread_info(idle
)->preempt_count
= 0;
6615 * The idle tasks have their own, simple scheduling class:
6617 idle
->sched_class
= &idle_sched_class
;
6618 ftrace_graph_init_task(idle
);
6622 * In a system that switches off the HZ timer nohz_cpu_mask
6623 * indicates which cpus entered this state. This is used
6624 * in the rcu update to wait only for active cpus. For system
6625 * which do not switch off the HZ timer nohz_cpu_mask should
6626 * always be CPU_BITS_NONE.
6628 cpumask_var_t nohz_cpu_mask
;
6631 * Increase the granularity value when there are more CPUs,
6632 * because with more CPUs the 'effective latency' as visible
6633 * to users decreases. But the relationship is not linear,
6634 * so pick a second-best guess by going with the log2 of the
6637 * This idea comes from the SD scheduler of Con Kolivas:
6639 static inline void sched_init_granularity(void)
6641 unsigned int factor
= 1 + ilog2(num_online_cpus());
6642 const unsigned long limit
= 200000000;
6644 sysctl_sched_min_granularity
*= factor
;
6645 if (sysctl_sched_min_granularity
> limit
)
6646 sysctl_sched_min_granularity
= limit
;
6648 sysctl_sched_latency
*= factor
;
6649 if (sysctl_sched_latency
> limit
)
6650 sysctl_sched_latency
= limit
;
6652 sysctl_sched_wakeup_granularity
*= factor
;
6654 sysctl_sched_shares_ratelimit
*= factor
;
6659 * This is how migration works:
6661 * 1) we queue a struct migration_req structure in the source CPU's
6662 * runqueue and wake up that CPU's migration thread.
6663 * 2) we down() the locked semaphore => thread blocks.
6664 * 3) migration thread wakes up (implicitly it forces the migrated
6665 * thread off the CPU)
6666 * 4) it gets the migration request and checks whether the migrated
6667 * task is still in the wrong runqueue.
6668 * 5) if it's in the wrong runqueue then the migration thread removes
6669 * it and puts it into the right queue.
6670 * 6) migration thread up()s the semaphore.
6671 * 7) we wake up and the migration is done.
6675 * Change a given task's CPU affinity. Migrate the thread to a
6676 * proper CPU and schedule it away if the CPU it's executing on
6677 * is removed from the allowed bitmask.
6679 * NOTE: the caller must have a valid reference to the task, the
6680 * task must not exit() & deallocate itself prematurely. The
6681 * call is not atomic; no spinlocks may be held.
6683 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6685 struct migration_req req
;
6686 unsigned long flags
;
6690 rq
= task_rq_lock(p
, &flags
);
6691 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6696 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6697 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6702 if (p
->sched_class
->set_cpus_allowed
)
6703 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6705 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6706 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6709 /* Can the task run on the task's current CPU? If so, we're done */
6710 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6713 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6714 /* Need help from migration thread: drop lock and wait. */
6715 task_rq_unlock(rq
, &flags
);
6716 wake_up_process(rq
->migration_thread
);
6717 wait_for_completion(&req
.done
);
6718 tlb_migrate_finish(p
->mm
);
6722 task_rq_unlock(rq
, &flags
);
6726 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6729 * Move (not current) task off this cpu, onto dest cpu. We're doing
6730 * this because either it can't run here any more (set_cpus_allowed()
6731 * away from this CPU, or CPU going down), or because we're
6732 * attempting to rebalance this task on exec (sched_exec).
6734 * So we race with normal scheduler movements, but that's OK, as long
6735 * as the task is no longer on this CPU.
6737 * Returns non-zero if task was successfully migrated.
6739 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6741 struct rq
*rq_dest
, *rq_src
;
6744 if (unlikely(!cpu_active(dest_cpu
)))
6747 rq_src
= cpu_rq(src_cpu
);
6748 rq_dest
= cpu_rq(dest_cpu
);
6750 double_rq_lock(rq_src
, rq_dest
);
6751 /* Already moved. */
6752 if (task_cpu(p
) != src_cpu
)
6754 /* Affinity changed (again). */
6755 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6758 on_rq
= p
->se
.on_rq
;
6760 deactivate_task(rq_src
, p
, 0);
6762 set_task_cpu(p
, dest_cpu
);
6764 activate_task(rq_dest
, p
, 0);
6765 check_preempt_curr(rq_dest
, p
, 0);
6770 double_rq_unlock(rq_src
, rq_dest
);
6775 * migration_thread - this is a highprio system thread that performs
6776 * thread migration by bumping thread off CPU then 'pushing' onto
6779 static int migration_thread(void *data
)
6781 int cpu
= (long)data
;
6785 BUG_ON(rq
->migration_thread
!= current
);
6787 set_current_state(TASK_INTERRUPTIBLE
);
6788 while (!kthread_should_stop()) {
6789 struct migration_req
*req
;
6790 struct list_head
*head
;
6792 spin_lock_irq(&rq
->lock
);
6794 if (cpu_is_offline(cpu
)) {
6795 spin_unlock_irq(&rq
->lock
);
6799 if (rq
->active_balance
) {
6800 active_load_balance(rq
, cpu
);
6801 rq
->active_balance
= 0;
6804 head
= &rq
->migration_queue
;
6806 if (list_empty(head
)) {
6807 spin_unlock_irq(&rq
->lock
);
6809 set_current_state(TASK_INTERRUPTIBLE
);
6812 req
= list_entry(head
->next
, struct migration_req
, list
);
6813 list_del_init(head
->next
);
6815 spin_unlock(&rq
->lock
);
6816 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6819 complete(&req
->done
);
6821 __set_current_state(TASK_RUNNING
);
6825 /* Wait for kthread_stop */
6826 set_current_state(TASK_INTERRUPTIBLE
);
6827 while (!kthread_should_stop()) {
6829 set_current_state(TASK_INTERRUPTIBLE
);
6831 __set_current_state(TASK_RUNNING
);
6835 #ifdef CONFIG_HOTPLUG_CPU
6837 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6841 local_irq_disable();
6842 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6848 * Figure out where task on dead CPU should go, use force if necessary.
6850 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6853 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6856 /* Look for allowed, online CPU in same node. */
6857 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6858 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6861 /* Any allowed, online CPU? */
6862 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6863 if (dest_cpu
< nr_cpu_ids
)
6866 /* No more Mr. Nice Guy. */
6867 if (dest_cpu
>= nr_cpu_ids
) {
6868 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6869 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6872 * Don't tell them about moving exiting tasks or
6873 * kernel threads (both mm NULL), since they never
6876 if (p
->mm
&& printk_ratelimit()) {
6877 printk(KERN_INFO
"process %d (%s) no "
6878 "longer affine to cpu%d\n",
6879 task_pid_nr(p
), p
->comm
, dead_cpu
);
6884 /* It can have affinity changed while we were choosing. */
6885 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6890 * While a dead CPU has no uninterruptible tasks queued at this point,
6891 * it might still have a nonzero ->nr_uninterruptible counter, because
6892 * for performance reasons the counter is not stricly tracking tasks to
6893 * their home CPUs. So we just add the counter to another CPU's counter,
6894 * to keep the global sum constant after CPU-down:
6896 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6898 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6899 unsigned long flags
;
6901 local_irq_save(flags
);
6902 double_rq_lock(rq_src
, rq_dest
);
6903 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6904 rq_src
->nr_uninterruptible
= 0;
6905 double_rq_unlock(rq_src
, rq_dest
);
6906 local_irq_restore(flags
);
6909 /* Run through task list and migrate tasks from the dead cpu. */
6910 static void migrate_live_tasks(int src_cpu
)
6912 struct task_struct
*p
, *t
;
6914 read_lock(&tasklist_lock
);
6916 do_each_thread(t
, p
) {
6920 if (task_cpu(p
) == src_cpu
)
6921 move_task_off_dead_cpu(src_cpu
, p
);
6922 } while_each_thread(t
, p
);
6924 read_unlock(&tasklist_lock
);
6928 * Schedules idle task to be the next runnable task on current CPU.
6929 * It does so by boosting its priority to highest possible.
6930 * Used by CPU offline code.
6932 void sched_idle_next(void)
6934 int this_cpu
= smp_processor_id();
6935 struct rq
*rq
= cpu_rq(this_cpu
);
6936 struct task_struct
*p
= rq
->idle
;
6937 unsigned long flags
;
6939 /* cpu has to be offline */
6940 BUG_ON(cpu_online(this_cpu
));
6943 * Strictly not necessary since rest of the CPUs are stopped by now
6944 * and interrupts disabled on the current cpu.
6946 spin_lock_irqsave(&rq
->lock
, flags
);
6948 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6950 update_rq_clock(rq
);
6951 activate_task(rq
, p
, 0);
6953 spin_unlock_irqrestore(&rq
->lock
, flags
);
6957 * Ensures that the idle task is using init_mm right before its cpu goes
6960 void idle_task_exit(void)
6962 struct mm_struct
*mm
= current
->active_mm
;
6964 BUG_ON(cpu_online(smp_processor_id()));
6967 switch_mm(mm
, &init_mm
, current
);
6971 /* called under rq->lock with disabled interrupts */
6972 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6974 struct rq
*rq
= cpu_rq(dead_cpu
);
6976 /* Must be exiting, otherwise would be on tasklist. */
6977 BUG_ON(!p
->exit_state
);
6979 /* Cannot have done final schedule yet: would have vanished. */
6980 BUG_ON(p
->state
== TASK_DEAD
);
6985 * Drop lock around migration; if someone else moves it,
6986 * that's OK. No task can be added to this CPU, so iteration is
6989 spin_unlock_irq(&rq
->lock
);
6990 move_task_off_dead_cpu(dead_cpu
, p
);
6991 spin_lock_irq(&rq
->lock
);
6996 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6997 static void migrate_dead_tasks(unsigned int dead_cpu
)
6999 struct rq
*rq
= cpu_rq(dead_cpu
);
7000 struct task_struct
*next
;
7003 if (!rq
->nr_running
)
7005 update_rq_clock(rq
);
7006 next
= pick_next_task(rq
);
7009 next
->sched_class
->put_prev_task(rq
, next
);
7010 migrate_dead(dead_cpu
, next
);
7014 #endif /* CONFIG_HOTPLUG_CPU */
7016 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7018 static struct ctl_table sd_ctl_dir
[] = {
7020 .procname
= "sched_domain",
7026 static struct ctl_table sd_ctl_root
[] = {
7028 .ctl_name
= CTL_KERN
,
7029 .procname
= "kernel",
7031 .child
= sd_ctl_dir
,
7036 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7038 struct ctl_table
*entry
=
7039 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7044 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7046 struct ctl_table
*entry
;
7049 * In the intermediate directories, both the child directory and
7050 * procname are dynamically allocated and could fail but the mode
7051 * will always be set. In the lowest directory the names are
7052 * static strings and all have proc handlers.
7054 for (entry
= *tablep
; entry
->mode
; entry
++) {
7056 sd_free_ctl_entry(&entry
->child
);
7057 if (entry
->proc_handler
== NULL
)
7058 kfree(entry
->procname
);
7066 set_table_entry(struct ctl_table
*entry
,
7067 const char *procname
, void *data
, int maxlen
,
7068 mode_t mode
, proc_handler
*proc_handler
)
7070 entry
->procname
= procname
;
7072 entry
->maxlen
= maxlen
;
7074 entry
->proc_handler
= proc_handler
;
7077 static struct ctl_table
*
7078 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7080 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7085 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7086 sizeof(long), 0644, proc_doulongvec_minmax
);
7087 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7088 sizeof(long), 0644, proc_doulongvec_minmax
);
7089 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7090 sizeof(int), 0644, proc_dointvec_minmax
);
7091 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7092 sizeof(int), 0644, proc_dointvec_minmax
);
7093 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7094 sizeof(int), 0644, proc_dointvec_minmax
);
7095 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7096 sizeof(int), 0644, proc_dointvec_minmax
);
7097 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7098 sizeof(int), 0644, proc_dointvec_minmax
);
7099 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7100 sizeof(int), 0644, proc_dointvec_minmax
);
7101 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7102 sizeof(int), 0644, proc_dointvec_minmax
);
7103 set_table_entry(&table
[9], "cache_nice_tries",
7104 &sd
->cache_nice_tries
,
7105 sizeof(int), 0644, proc_dointvec_minmax
);
7106 set_table_entry(&table
[10], "flags", &sd
->flags
,
7107 sizeof(int), 0644, proc_dointvec_minmax
);
7108 set_table_entry(&table
[11], "name", sd
->name
,
7109 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7110 /* &table[12] is terminator */
7115 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7117 struct ctl_table
*entry
, *table
;
7118 struct sched_domain
*sd
;
7119 int domain_num
= 0, i
;
7122 for_each_domain(cpu
, sd
)
7124 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7129 for_each_domain(cpu
, sd
) {
7130 snprintf(buf
, 32, "domain%d", i
);
7131 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7133 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7140 static struct ctl_table_header
*sd_sysctl_header
;
7141 static void register_sched_domain_sysctl(void)
7143 int i
, cpu_num
= num_online_cpus();
7144 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7147 WARN_ON(sd_ctl_dir
[0].child
);
7148 sd_ctl_dir
[0].child
= entry
;
7153 for_each_online_cpu(i
) {
7154 snprintf(buf
, 32, "cpu%d", i
);
7155 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7157 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7161 WARN_ON(sd_sysctl_header
);
7162 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7165 /* may be called multiple times per register */
7166 static void unregister_sched_domain_sysctl(void)
7168 if (sd_sysctl_header
)
7169 unregister_sysctl_table(sd_sysctl_header
);
7170 sd_sysctl_header
= NULL
;
7171 if (sd_ctl_dir
[0].child
)
7172 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7175 static void register_sched_domain_sysctl(void)
7178 static void unregister_sched_domain_sysctl(void)
7183 static void set_rq_online(struct rq
*rq
)
7186 const struct sched_class
*class;
7188 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7191 for_each_class(class) {
7192 if (class->rq_online
)
7193 class->rq_online(rq
);
7198 static void set_rq_offline(struct rq
*rq
)
7201 const struct sched_class
*class;
7203 for_each_class(class) {
7204 if (class->rq_offline
)
7205 class->rq_offline(rq
);
7208 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7214 * migration_call - callback that gets triggered when a CPU is added.
7215 * Here we can start up the necessary migration thread for the new CPU.
7217 static int __cpuinit
7218 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7220 struct task_struct
*p
;
7221 int cpu
= (long)hcpu
;
7222 unsigned long flags
;
7227 case CPU_UP_PREPARE
:
7228 case CPU_UP_PREPARE_FROZEN
:
7229 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7232 kthread_bind(p
, cpu
);
7233 /* Must be high prio: stop_machine expects to yield to it. */
7234 rq
= task_rq_lock(p
, &flags
);
7235 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7236 task_rq_unlock(rq
, &flags
);
7237 cpu_rq(cpu
)->migration_thread
= p
;
7241 case CPU_ONLINE_FROZEN
:
7242 /* Strictly unnecessary, as first user will wake it. */
7243 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7245 /* Update our root-domain */
7247 spin_lock_irqsave(&rq
->lock
, flags
);
7249 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7253 spin_unlock_irqrestore(&rq
->lock
, flags
);
7256 #ifdef CONFIG_HOTPLUG_CPU
7257 case CPU_UP_CANCELED
:
7258 case CPU_UP_CANCELED_FROZEN
:
7259 if (!cpu_rq(cpu
)->migration_thread
)
7261 /* Unbind it from offline cpu so it can run. Fall thru. */
7262 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7263 cpumask_any(cpu_online_mask
));
7264 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7265 cpu_rq(cpu
)->migration_thread
= NULL
;
7269 case CPU_DEAD_FROZEN
:
7270 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7271 migrate_live_tasks(cpu
);
7273 kthread_stop(rq
->migration_thread
);
7274 rq
->migration_thread
= NULL
;
7275 /* Idle task back to normal (off runqueue, low prio) */
7276 spin_lock_irq(&rq
->lock
);
7277 update_rq_clock(rq
);
7278 deactivate_task(rq
, rq
->idle
, 0);
7279 rq
->idle
->static_prio
= MAX_PRIO
;
7280 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7281 rq
->idle
->sched_class
= &idle_sched_class
;
7282 migrate_dead_tasks(cpu
);
7283 spin_unlock_irq(&rq
->lock
);
7285 migrate_nr_uninterruptible(rq
);
7286 BUG_ON(rq
->nr_running
!= 0);
7289 * No need to migrate the tasks: it was best-effort if
7290 * they didn't take sched_hotcpu_mutex. Just wake up
7293 spin_lock_irq(&rq
->lock
);
7294 while (!list_empty(&rq
->migration_queue
)) {
7295 struct migration_req
*req
;
7297 req
= list_entry(rq
->migration_queue
.next
,
7298 struct migration_req
, list
);
7299 list_del_init(&req
->list
);
7300 spin_unlock_irq(&rq
->lock
);
7301 complete(&req
->done
);
7302 spin_lock_irq(&rq
->lock
);
7304 spin_unlock_irq(&rq
->lock
);
7308 case CPU_DYING_FROZEN
:
7309 /* Update our root-domain */
7311 spin_lock_irqsave(&rq
->lock
, flags
);
7313 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7316 spin_unlock_irqrestore(&rq
->lock
, flags
);
7324 * Register at high priority so that task migration (migrate_all_tasks)
7325 * happens before everything else. This has to be lower priority than
7326 * the notifier in the perf_counter subsystem, though.
7328 static struct notifier_block __cpuinitdata migration_notifier
= {
7329 .notifier_call
= migration_call
,
7333 static int __init
migration_init(void)
7335 void *cpu
= (void *)(long)smp_processor_id();
7338 /* Start one for the boot CPU: */
7339 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7340 BUG_ON(err
== NOTIFY_BAD
);
7341 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7342 register_cpu_notifier(&migration_notifier
);
7346 early_initcall(migration_init
);
7351 #ifdef CONFIG_SCHED_DEBUG
7353 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7354 struct cpumask
*groupmask
)
7356 struct sched_group
*group
= sd
->groups
;
7359 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7360 cpumask_clear(groupmask
);
7362 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7364 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7365 printk("does not load-balance\n");
7367 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7372 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7374 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7375 printk(KERN_ERR
"ERROR: domain->span does not contain "
7378 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7379 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7383 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7387 printk(KERN_ERR
"ERROR: group is NULL\n");
7391 if (!group
->__cpu_power
) {
7392 printk(KERN_CONT
"\n");
7393 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7398 if (!cpumask_weight(sched_group_cpus(group
))) {
7399 printk(KERN_CONT
"\n");
7400 printk(KERN_ERR
"ERROR: empty group\n");
7404 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7405 printk(KERN_CONT
"\n");
7406 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7410 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7412 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7414 printk(KERN_CONT
" %s", str
);
7415 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7416 printk(KERN_CONT
" (__cpu_power = %d)",
7417 group
->__cpu_power
);
7420 group
= group
->next
;
7421 } while (group
!= sd
->groups
);
7422 printk(KERN_CONT
"\n");
7424 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7425 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7428 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7429 printk(KERN_ERR
"ERROR: parent span is not a superset "
7430 "of domain->span\n");
7434 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7436 cpumask_var_t groupmask
;
7440 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7444 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7446 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7447 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7452 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7459 free_cpumask_var(groupmask
);
7461 #else /* !CONFIG_SCHED_DEBUG */
7462 # define sched_domain_debug(sd, cpu) do { } while (0)
7463 #endif /* CONFIG_SCHED_DEBUG */
7465 static int sd_degenerate(struct sched_domain
*sd
)
7467 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7470 /* Following flags need at least 2 groups */
7471 if (sd
->flags
& (SD_LOAD_BALANCE
|
7472 SD_BALANCE_NEWIDLE
|
7476 SD_SHARE_PKG_RESOURCES
)) {
7477 if (sd
->groups
!= sd
->groups
->next
)
7481 /* Following flags don't use groups */
7482 if (sd
->flags
& (SD_WAKE_IDLE
|
7491 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7493 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7495 if (sd_degenerate(parent
))
7498 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7501 /* Does parent contain flags not in child? */
7502 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7503 if (cflags
& SD_WAKE_AFFINE
)
7504 pflags
&= ~SD_WAKE_BALANCE
;
7505 /* Flags needing groups don't count if only 1 group in parent */
7506 if (parent
->groups
== parent
->groups
->next
) {
7507 pflags
&= ~(SD_LOAD_BALANCE
|
7508 SD_BALANCE_NEWIDLE
|
7512 SD_SHARE_PKG_RESOURCES
);
7513 if (nr_node_ids
== 1)
7514 pflags
&= ~SD_SERIALIZE
;
7516 if (~cflags
& pflags
)
7522 static void free_rootdomain(struct root_domain
*rd
)
7524 cpupri_cleanup(&rd
->cpupri
);
7526 free_cpumask_var(rd
->rto_mask
);
7527 free_cpumask_var(rd
->online
);
7528 free_cpumask_var(rd
->span
);
7532 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7534 struct root_domain
*old_rd
= NULL
;
7535 unsigned long flags
;
7537 spin_lock_irqsave(&rq
->lock
, flags
);
7542 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7545 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7548 * If we dont want to free the old_rt yet then
7549 * set old_rd to NULL to skip the freeing later
7552 if (!atomic_dec_and_test(&old_rd
->refcount
))
7556 atomic_inc(&rd
->refcount
);
7559 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7560 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7563 spin_unlock_irqrestore(&rq
->lock
, flags
);
7566 free_rootdomain(old_rd
);
7569 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7571 memset(rd
, 0, sizeof(*rd
));
7574 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7575 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7576 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7577 cpupri_init(&rd
->cpupri
, true);
7581 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7583 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7585 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7588 if (cpupri_init(&rd
->cpupri
, false) != 0)
7593 free_cpumask_var(rd
->rto_mask
);
7595 free_cpumask_var(rd
->online
);
7597 free_cpumask_var(rd
->span
);
7602 static void init_defrootdomain(void)
7604 init_rootdomain(&def_root_domain
, true);
7606 atomic_set(&def_root_domain
.refcount
, 1);
7609 static struct root_domain
*alloc_rootdomain(void)
7611 struct root_domain
*rd
;
7613 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7617 if (init_rootdomain(rd
, false) != 0) {
7626 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7627 * hold the hotplug lock.
7630 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7632 struct rq
*rq
= cpu_rq(cpu
);
7633 struct sched_domain
*tmp
;
7635 /* Remove the sched domains which do not contribute to scheduling. */
7636 for (tmp
= sd
; tmp
; ) {
7637 struct sched_domain
*parent
= tmp
->parent
;
7641 if (sd_parent_degenerate(tmp
, parent
)) {
7642 tmp
->parent
= parent
->parent
;
7644 parent
->parent
->child
= tmp
;
7649 if (sd
&& sd_degenerate(sd
)) {
7655 sched_domain_debug(sd
, cpu
);
7657 rq_attach_root(rq
, rd
);
7658 rcu_assign_pointer(rq
->sd
, sd
);
7661 /* cpus with isolated domains */
7662 static cpumask_var_t cpu_isolated_map
;
7664 /* Setup the mask of cpus configured for isolated domains */
7665 static int __init
isolated_cpu_setup(char *str
)
7667 cpulist_parse(str
, cpu_isolated_map
);
7671 __setup("isolcpus=", isolated_cpu_setup
);
7674 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7675 * to a function which identifies what group(along with sched group) a CPU
7676 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7677 * (due to the fact that we keep track of groups covered with a struct cpumask).
7679 * init_sched_build_groups will build a circular linked list of the groups
7680 * covered by the given span, and will set each group's ->cpumask correctly,
7681 * and ->cpu_power to 0.
7684 init_sched_build_groups(const struct cpumask
*span
,
7685 const struct cpumask
*cpu_map
,
7686 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7687 struct sched_group
**sg
,
7688 struct cpumask
*tmpmask
),
7689 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7691 struct sched_group
*first
= NULL
, *last
= NULL
;
7694 cpumask_clear(covered
);
7696 for_each_cpu(i
, span
) {
7697 struct sched_group
*sg
;
7698 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7701 if (cpumask_test_cpu(i
, covered
))
7704 cpumask_clear(sched_group_cpus(sg
));
7705 sg
->__cpu_power
= 0;
7707 for_each_cpu(j
, span
) {
7708 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7711 cpumask_set_cpu(j
, covered
);
7712 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7723 #define SD_NODES_PER_DOMAIN 16
7728 * find_next_best_node - find the next node to include in a sched_domain
7729 * @node: node whose sched_domain we're building
7730 * @used_nodes: nodes already in the sched_domain
7732 * Find the next node to include in a given scheduling domain. Simply
7733 * finds the closest node not already in the @used_nodes map.
7735 * Should use nodemask_t.
7737 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7739 int i
, n
, val
, min_val
, best_node
= 0;
7743 for (i
= 0; i
< nr_node_ids
; i
++) {
7744 /* Start at @node */
7745 n
= (node
+ i
) % nr_node_ids
;
7747 if (!nr_cpus_node(n
))
7750 /* Skip already used nodes */
7751 if (node_isset(n
, *used_nodes
))
7754 /* Simple min distance search */
7755 val
= node_distance(node
, n
);
7757 if (val
< min_val
) {
7763 node_set(best_node
, *used_nodes
);
7768 * sched_domain_node_span - get a cpumask for a node's sched_domain
7769 * @node: node whose cpumask we're constructing
7770 * @span: resulting cpumask
7772 * Given a node, construct a good cpumask for its sched_domain to span. It
7773 * should be one that prevents unnecessary balancing, but also spreads tasks
7776 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7778 nodemask_t used_nodes
;
7781 cpumask_clear(span
);
7782 nodes_clear(used_nodes
);
7784 cpumask_or(span
, span
, cpumask_of_node(node
));
7785 node_set(node
, used_nodes
);
7787 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7788 int next_node
= find_next_best_node(node
, &used_nodes
);
7790 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7793 #endif /* CONFIG_NUMA */
7795 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7798 * The cpus mask in sched_group and sched_domain hangs off the end.
7799 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7800 * for nr_cpu_ids < CONFIG_NR_CPUS.
7802 struct static_sched_group
{
7803 struct sched_group sg
;
7804 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7807 struct static_sched_domain
{
7808 struct sched_domain sd
;
7809 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7813 * SMT sched-domains:
7815 #ifdef CONFIG_SCHED_SMT
7816 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7817 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7820 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7821 struct sched_group
**sg
, struct cpumask
*unused
)
7824 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7827 #endif /* CONFIG_SCHED_SMT */
7830 * multi-core sched-domains:
7832 #ifdef CONFIG_SCHED_MC
7833 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7834 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7835 #endif /* CONFIG_SCHED_MC */
7837 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7839 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7840 struct sched_group
**sg
, struct cpumask
*mask
)
7844 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7845 group
= cpumask_first(mask
);
7847 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7850 #elif defined(CONFIG_SCHED_MC)
7852 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7853 struct sched_group
**sg
, struct cpumask
*unused
)
7856 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7861 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7862 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7865 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7866 struct sched_group
**sg
, struct cpumask
*mask
)
7869 #ifdef CONFIG_SCHED_MC
7870 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7871 group
= cpumask_first(mask
);
7872 #elif defined(CONFIG_SCHED_SMT)
7873 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7874 group
= cpumask_first(mask
);
7879 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7885 * The init_sched_build_groups can't handle what we want to do with node
7886 * groups, so roll our own. Now each node has its own list of groups which
7887 * gets dynamically allocated.
7889 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7890 static struct sched_group
***sched_group_nodes_bycpu
;
7892 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7893 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7895 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7896 struct sched_group
**sg
,
7897 struct cpumask
*nodemask
)
7901 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7902 group
= cpumask_first(nodemask
);
7905 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7909 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7911 struct sched_group
*sg
= group_head
;
7917 for_each_cpu(j
, sched_group_cpus(sg
)) {
7918 struct sched_domain
*sd
;
7920 sd
= &per_cpu(phys_domains
, j
).sd
;
7921 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7923 * Only add "power" once for each
7929 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7932 } while (sg
!= group_head
);
7934 #endif /* CONFIG_NUMA */
7937 /* Free memory allocated for various sched_group structures */
7938 static void free_sched_groups(const struct cpumask
*cpu_map
,
7939 struct cpumask
*nodemask
)
7943 for_each_cpu(cpu
, cpu_map
) {
7944 struct sched_group
**sched_group_nodes
7945 = sched_group_nodes_bycpu
[cpu
];
7947 if (!sched_group_nodes
)
7950 for (i
= 0; i
< nr_node_ids
; i
++) {
7951 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7953 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7954 if (cpumask_empty(nodemask
))
7964 if (oldsg
!= sched_group_nodes
[i
])
7967 kfree(sched_group_nodes
);
7968 sched_group_nodes_bycpu
[cpu
] = NULL
;
7971 #else /* !CONFIG_NUMA */
7972 static void free_sched_groups(const struct cpumask
*cpu_map
,
7973 struct cpumask
*nodemask
)
7976 #endif /* CONFIG_NUMA */
7979 * Initialize sched groups cpu_power.
7981 * cpu_power indicates the capacity of sched group, which is used while
7982 * distributing the load between different sched groups in a sched domain.
7983 * Typically cpu_power for all the groups in a sched domain will be same unless
7984 * there are asymmetries in the topology. If there are asymmetries, group
7985 * having more cpu_power will pickup more load compared to the group having
7988 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7989 * the maximum number of tasks a group can handle in the presence of other idle
7990 * or lightly loaded groups in the same sched domain.
7992 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7994 struct sched_domain
*child
;
7995 struct sched_group
*group
;
7997 WARN_ON(!sd
|| !sd
->groups
);
7999 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
8004 sd
->groups
->__cpu_power
= 0;
8007 * For perf policy, if the groups in child domain share resources
8008 * (for example cores sharing some portions of the cache hierarchy
8009 * or SMT), then set this domain groups cpu_power such that each group
8010 * can handle only one task, when there are other idle groups in the
8011 * same sched domain.
8013 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8015 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8016 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8021 * add cpu_power of each child group to this groups cpu_power
8023 group
= child
->groups
;
8025 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8026 group
= group
->next
;
8027 } while (group
!= child
->groups
);
8031 * Initializers for schedule domains
8032 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8035 #ifdef CONFIG_SCHED_DEBUG
8036 # define SD_INIT_NAME(sd, type) sd->name = #type
8038 # define SD_INIT_NAME(sd, type) do { } while (0)
8041 #define SD_INIT(sd, type) sd_init_##type(sd)
8043 #define SD_INIT_FUNC(type) \
8044 static noinline void sd_init_##type(struct sched_domain *sd) \
8046 memset(sd, 0, sizeof(*sd)); \
8047 *sd = SD_##type##_INIT; \
8048 sd->level = SD_LV_##type; \
8049 SD_INIT_NAME(sd, type); \
8054 SD_INIT_FUNC(ALLNODES
)
8057 #ifdef CONFIG_SCHED_SMT
8058 SD_INIT_FUNC(SIBLING
)
8060 #ifdef CONFIG_SCHED_MC
8064 static int default_relax_domain_level
= -1;
8066 static int __init
setup_relax_domain_level(char *str
)
8070 val
= simple_strtoul(str
, NULL
, 0);
8071 if (val
< SD_LV_MAX
)
8072 default_relax_domain_level
= val
;
8076 __setup("relax_domain_level=", setup_relax_domain_level
);
8078 static void set_domain_attribute(struct sched_domain
*sd
,
8079 struct sched_domain_attr
*attr
)
8083 if (!attr
|| attr
->relax_domain_level
< 0) {
8084 if (default_relax_domain_level
< 0)
8087 request
= default_relax_domain_level
;
8089 request
= attr
->relax_domain_level
;
8090 if (request
< sd
->level
) {
8091 /* turn off idle balance on this domain */
8092 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8094 /* turn on idle balance on this domain */
8095 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8100 * Build sched domains for a given set of cpus and attach the sched domains
8101 * to the individual cpus
8103 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8104 struct sched_domain_attr
*attr
)
8106 int i
, err
= -ENOMEM
;
8107 struct root_domain
*rd
;
8108 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8111 cpumask_var_t domainspan
, covered
, notcovered
;
8112 struct sched_group
**sched_group_nodes
= NULL
;
8113 int sd_allnodes
= 0;
8115 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8117 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8118 goto free_domainspan
;
8119 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8123 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8124 goto free_notcovered
;
8125 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8127 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8128 goto free_this_sibling_map
;
8129 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8130 goto free_this_core_map
;
8131 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8132 goto free_send_covered
;
8136 * Allocate the per-node list of sched groups
8138 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8140 if (!sched_group_nodes
) {
8141 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8146 rd
= alloc_rootdomain();
8148 printk(KERN_WARNING
"Cannot alloc root domain\n");
8149 goto free_sched_groups
;
8153 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8157 * Set up domains for cpus specified by the cpu_map.
8159 for_each_cpu(i
, cpu_map
) {
8160 struct sched_domain
*sd
= NULL
, *p
;
8162 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8165 if (cpumask_weight(cpu_map
) >
8166 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8167 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8168 SD_INIT(sd
, ALLNODES
);
8169 set_domain_attribute(sd
, attr
);
8170 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8171 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8177 sd
= &per_cpu(node_domains
, i
).sd
;
8179 set_domain_attribute(sd
, attr
);
8180 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8184 cpumask_and(sched_domain_span(sd
),
8185 sched_domain_span(sd
), cpu_map
);
8189 sd
= &per_cpu(phys_domains
, i
).sd
;
8191 set_domain_attribute(sd
, attr
);
8192 cpumask_copy(sched_domain_span(sd
), nodemask
);
8196 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8198 #ifdef CONFIG_SCHED_MC
8200 sd
= &per_cpu(core_domains
, i
).sd
;
8202 set_domain_attribute(sd
, attr
);
8203 cpumask_and(sched_domain_span(sd
), cpu_map
,
8204 cpu_coregroup_mask(i
));
8207 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8210 #ifdef CONFIG_SCHED_SMT
8212 sd
= &per_cpu(cpu_domains
, i
).sd
;
8213 SD_INIT(sd
, SIBLING
);
8214 set_domain_attribute(sd
, attr
);
8215 cpumask_and(sched_domain_span(sd
),
8216 topology_thread_cpumask(i
), cpu_map
);
8219 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8223 #ifdef CONFIG_SCHED_SMT
8224 /* Set up CPU (sibling) groups */
8225 for_each_cpu(i
, cpu_map
) {
8226 cpumask_and(this_sibling_map
,
8227 topology_thread_cpumask(i
), cpu_map
);
8228 if (i
!= cpumask_first(this_sibling_map
))
8231 init_sched_build_groups(this_sibling_map
, cpu_map
,
8233 send_covered
, tmpmask
);
8237 #ifdef CONFIG_SCHED_MC
8238 /* Set up multi-core groups */
8239 for_each_cpu(i
, cpu_map
) {
8240 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8241 if (i
!= cpumask_first(this_core_map
))
8244 init_sched_build_groups(this_core_map
, cpu_map
,
8246 send_covered
, tmpmask
);
8250 /* Set up physical groups */
8251 for (i
= 0; i
< nr_node_ids
; i
++) {
8252 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8253 if (cpumask_empty(nodemask
))
8256 init_sched_build_groups(nodemask
, cpu_map
,
8258 send_covered
, tmpmask
);
8262 /* Set up node groups */
8264 init_sched_build_groups(cpu_map
, cpu_map
,
8265 &cpu_to_allnodes_group
,
8266 send_covered
, tmpmask
);
8269 for (i
= 0; i
< nr_node_ids
; i
++) {
8270 /* Set up node groups */
8271 struct sched_group
*sg
, *prev
;
8274 cpumask_clear(covered
);
8275 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8276 if (cpumask_empty(nodemask
)) {
8277 sched_group_nodes
[i
] = NULL
;
8281 sched_domain_node_span(i
, domainspan
);
8282 cpumask_and(domainspan
, domainspan
, cpu_map
);
8284 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8287 printk(KERN_WARNING
"Can not alloc domain group for "
8291 sched_group_nodes
[i
] = sg
;
8292 for_each_cpu(j
, nodemask
) {
8293 struct sched_domain
*sd
;
8295 sd
= &per_cpu(node_domains
, j
).sd
;
8298 sg
->__cpu_power
= 0;
8299 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8301 cpumask_or(covered
, covered
, nodemask
);
8304 for (j
= 0; j
< nr_node_ids
; j
++) {
8305 int n
= (i
+ j
) % nr_node_ids
;
8307 cpumask_complement(notcovered
, covered
);
8308 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8309 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8310 if (cpumask_empty(tmpmask
))
8313 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8314 if (cpumask_empty(tmpmask
))
8317 sg
= kmalloc_node(sizeof(struct sched_group
) +
8322 "Can not alloc domain group for node %d\n", j
);
8325 sg
->__cpu_power
= 0;
8326 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8327 sg
->next
= prev
->next
;
8328 cpumask_or(covered
, covered
, tmpmask
);
8335 /* Calculate CPU power for physical packages and nodes */
8336 #ifdef CONFIG_SCHED_SMT
8337 for_each_cpu(i
, cpu_map
) {
8338 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8340 init_sched_groups_power(i
, sd
);
8343 #ifdef CONFIG_SCHED_MC
8344 for_each_cpu(i
, cpu_map
) {
8345 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8347 init_sched_groups_power(i
, sd
);
8351 for_each_cpu(i
, cpu_map
) {
8352 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8354 init_sched_groups_power(i
, sd
);
8358 for (i
= 0; i
< nr_node_ids
; i
++)
8359 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8362 struct sched_group
*sg
;
8364 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8366 init_numa_sched_groups_power(sg
);
8370 /* Attach the domains */
8371 for_each_cpu(i
, cpu_map
) {
8372 struct sched_domain
*sd
;
8373 #ifdef CONFIG_SCHED_SMT
8374 sd
= &per_cpu(cpu_domains
, i
).sd
;
8375 #elif defined(CONFIG_SCHED_MC)
8376 sd
= &per_cpu(core_domains
, i
).sd
;
8378 sd
= &per_cpu(phys_domains
, i
).sd
;
8380 cpu_attach_domain(sd
, rd
, i
);
8386 free_cpumask_var(tmpmask
);
8388 free_cpumask_var(send_covered
);
8390 free_cpumask_var(this_core_map
);
8391 free_this_sibling_map
:
8392 free_cpumask_var(this_sibling_map
);
8394 free_cpumask_var(nodemask
);
8397 free_cpumask_var(notcovered
);
8399 free_cpumask_var(covered
);
8401 free_cpumask_var(domainspan
);
8408 kfree(sched_group_nodes
);
8414 free_sched_groups(cpu_map
, tmpmask
);
8415 free_rootdomain(rd
);
8420 static int build_sched_domains(const struct cpumask
*cpu_map
)
8422 return __build_sched_domains(cpu_map
, NULL
);
8425 static struct cpumask
*doms_cur
; /* current sched domains */
8426 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8427 static struct sched_domain_attr
*dattr_cur
;
8428 /* attribues of custom domains in 'doms_cur' */
8431 * Special case: If a kmalloc of a doms_cur partition (array of
8432 * cpumask) fails, then fallback to a single sched domain,
8433 * as determined by the single cpumask fallback_doms.
8435 static cpumask_var_t fallback_doms
;
8438 * arch_update_cpu_topology lets virtualized architectures update the
8439 * cpu core maps. It is supposed to return 1 if the topology changed
8440 * or 0 if it stayed the same.
8442 int __attribute__((weak
)) arch_update_cpu_topology(void)
8448 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8449 * For now this just excludes isolated cpus, but could be used to
8450 * exclude other special cases in the future.
8452 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8456 arch_update_cpu_topology();
8458 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8460 doms_cur
= fallback_doms
;
8461 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8463 err
= build_sched_domains(doms_cur
);
8464 register_sched_domain_sysctl();
8469 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8470 struct cpumask
*tmpmask
)
8472 free_sched_groups(cpu_map
, tmpmask
);
8476 * Detach sched domains from a group of cpus specified in cpu_map
8477 * These cpus will now be attached to the NULL domain
8479 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8481 /* Save because hotplug lock held. */
8482 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8485 for_each_cpu(i
, cpu_map
)
8486 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8487 synchronize_sched();
8488 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8491 /* handle null as "default" */
8492 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8493 struct sched_domain_attr
*new, int idx_new
)
8495 struct sched_domain_attr tmp
;
8502 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8503 new ? (new + idx_new
) : &tmp
,
8504 sizeof(struct sched_domain_attr
));
8508 * Partition sched domains as specified by the 'ndoms_new'
8509 * cpumasks in the array doms_new[] of cpumasks. This compares
8510 * doms_new[] to the current sched domain partitioning, doms_cur[].
8511 * It destroys each deleted domain and builds each new domain.
8513 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8514 * The masks don't intersect (don't overlap.) We should setup one
8515 * sched domain for each mask. CPUs not in any of the cpumasks will
8516 * not be load balanced. If the same cpumask appears both in the
8517 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8520 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8521 * ownership of it and will kfree it when done with it. If the caller
8522 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8523 * ndoms_new == 1, and partition_sched_domains() will fallback to
8524 * the single partition 'fallback_doms', it also forces the domains
8527 * If doms_new == NULL it will be replaced with cpu_online_mask.
8528 * ndoms_new == 0 is a special case for destroying existing domains,
8529 * and it will not create the default domain.
8531 * Call with hotplug lock held
8533 /* FIXME: Change to struct cpumask *doms_new[] */
8534 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8535 struct sched_domain_attr
*dattr_new
)
8540 mutex_lock(&sched_domains_mutex
);
8542 /* always unregister in case we don't destroy any domains */
8543 unregister_sched_domain_sysctl();
8545 /* Let architecture update cpu core mappings. */
8546 new_topology
= arch_update_cpu_topology();
8548 n
= doms_new
? ndoms_new
: 0;
8550 /* Destroy deleted domains */
8551 for (i
= 0; i
< ndoms_cur
; i
++) {
8552 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8553 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8554 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8557 /* no match - a current sched domain not in new doms_new[] */
8558 detach_destroy_domains(doms_cur
+ i
);
8563 if (doms_new
== NULL
) {
8565 doms_new
= fallback_doms
;
8566 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8567 WARN_ON_ONCE(dattr_new
);
8570 /* Build new domains */
8571 for (i
= 0; i
< ndoms_new
; i
++) {
8572 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8573 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8574 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8577 /* no match - add a new doms_new */
8578 __build_sched_domains(doms_new
+ i
,
8579 dattr_new
? dattr_new
+ i
: NULL
);
8584 /* Remember the new sched domains */
8585 if (doms_cur
!= fallback_doms
)
8587 kfree(dattr_cur
); /* kfree(NULL) is safe */
8588 doms_cur
= doms_new
;
8589 dattr_cur
= dattr_new
;
8590 ndoms_cur
= ndoms_new
;
8592 register_sched_domain_sysctl();
8594 mutex_unlock(&sched_domains_mutex
);
8597 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8598 static void arch_reinit_sched_domains(void)
8602 /* Destroy domains first to force the rebuild */
8603 partition_sched_domains(0, NULL
, NULL
);
8605 rebuild_sched_domains();
8609 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8611 unsigned int level
= 0;
8613 if (sscanf(buf
, "%u", &level
) != 1)
8617 * level is always be positive so don't check for
8618 * level < POWERSAVINGS_BALANCE_NONE which is 0
8619 * What happens on 0 or 1 byte write,
8620 * need to check for count as well?
8623 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8627 sched_smt_power_savings
= level
;
8629 sched_mc_power_savings
= level
;
8631 arch_reinit_sched_domains();
8636 #ifdef CONFIG_SCHED_MC
8637 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8640 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8642 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8643 const char *buf
, size_t count
)
8645 return sched_power_savings_store(buf
, count
, 0);
8647 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8648 sched_mc_power_savings_show
,
8649 sched_mc_power_savings_store
);
8652 #ifdef CONFIG_SCHED_SMT
8653 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8656 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8658 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8659 const char *buf
, size_t count
)
8661 return sched_power_savings_store(buf
, count
, 1);
8663 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8664 sched_smt_power_savings_show
,
8665 sched_smt_power_savings_store
);
8668 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8672 #ifdef CONFIG_SCHED_SMT
8674 err
= sysfs_create_file(&cls
->kset
.kobj
,
8675 &attr_sched_smt_power_savings
.attr
);
8677 #ifdef CONFIG_SCHED_MC
8678 if (!err
&& mc_capable())
8679 err
= sysfs_create_file(&cls
->kset
.kobj
,
8680 &attr_sched_mc_power_savings
.attr
);
8684 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8686 #ifndef CONFIG_CPUSETS
8688 * Add online and remove offline CPUs from the scheduler domains.
8689 * When cpusets are enabled they take over this function.
8691 static int update_sched_domains(struct notifier_block
*nfb
,
8692 unsigned long action
, void *hcpu
)
8696 case CPU_ONLINE_FROZEN
:
8698 case CPU_DEAD_FROZEN
:
8699 partition_sched_domains(1, NULL
, NULL
);
8708 static int update_runtime(struct notifier_block
*nfb
,
8709 unsigned long action
, void *hcpu
)
8711 int cpu
= (int)(long)hcpu
;
8714 case CPU_DOWN_PREPARE
:
8715 case CPU_DOWN_PREPARE_FROZEN
:
8716 disable_runtime(cpu_rq(cpu
));
8719 case CPU_DOWN_FAILED
:
8720 case CPU_DOWN_FAILED_FROZEN
:
8722 case CPU_ONLINE_FROZEN
:
8723 enable_runtime(cpu_rq(cpu
));
8731 void __init
sched_init_smp(void)
8733 cpumask_var_t non_isolated_cpus
;
8735 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8737 #if defined(CONFIG_NUMA)
8738 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8740 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8743 mutex_lock(&sched_domains_mutex
);
8744 arch_init_sched_domains(cpu_online_mask
);
8745 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8746 if (cpumask_empty(non_isolated_cpus
))
8747 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8748 mutex_unlock(&sched_domains_mutex
);
8751 #ifndef CONFIG_CPUSETS
8752 /* XXX: Theoretical race here - CPU may be hotplugged now */
8753 hotcpu_notifier(update_sched_domains
, 0);
8756 /* RT runtime code needs to handle some hotplug events */
8757 hotcpu_notifier(update_runtime
, 0);
8761 /* Move init over to a non-isolated CPU */
8762 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8764 sched_init_granularity();
8765 free_cpumask_var(non_isolated_cpus
);
8767 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8768 init_sched_rt_class();
8771 void __init
sched_init_smp(void)
8773 sched_init_granularity();
8775 #endif /* CONFIG_SMP */
8777 int in_sched_functions(unsigned long addr
)
8779 return in_lock_functions(addr
) ||
8780 (addr
>= (unsigned long)__sched_text_start
8781 && addr
< (unsigned long)__sched_text_end
);
8784 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8786 cfs_rq
->tasks_timeline
= RB_ROOT
;
8787 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8788 #ifdef CONFIG_FAIR_GROUP_SCHED
8791 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8794 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8796 struct rt_prio_array
*array
;
8799 array
= &rt_rq
->active
;
8800 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8801 INIT_LIST_HEAD(array
->queue
+ i
);
8802 __clear_bit(i
, array
->bitmap
);
8804 /* delimiter for bitsearch: */
8805 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8807 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8808 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8810 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8814 rt_rq
->rt_nr_migratory
= 0;
8815 rt_rq
->overloaded
= 0;
8816 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8820 rt_rq
->rt_throttled
= 0;
8821 rt_rq
->rt_runtime
= 0;
8822 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8824 #ifdef CONFIG_RT_GROUP_SCHED
8825 rt_rq
->rt_nr_boosted
= 0;
8830 #ifdef CONFIG_FAIR_GROUP_SCHED
8831 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8832 struct sched_entity
*se
, int cpu
, int add
,
8833 struct sched_entity
*parent
)
8835 struct rq
*rq
= cpu_rq(cpu
);
8836 tg
->cfs_rq
[cpu
] = cfs_rq
;
8837 init_cfs_rq(cfs_rq
, rq
);
8840 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8843 /* se could be NULL for init_task_group */
8848 se
->cfs_rq
= &rq
->cfs
;
8850 se
->cfs_rq
= parent
->my_q
;
8853 se
->load
.weight
= tg
->shares
;
8854 se
->load
.inv_weight
= 0;
8855 se
->parent
= parent
;
8859 #ifdef CONFIG_RT_GROUP_SCHED
8860 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8861 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8862 struct sched_rt_entity
*parent
)
8864 struct rq
*rq
= cpu_rq(cpu
);
8866 tg
->rt_rq
[cpu
] = rt_rq
;
8867 init_rt_rq(rt_rq
, rq
);
8869 rt_rq
->rt_se
= rt_se
;
8870 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8872 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8874 tg
->rt_se
[cpu
] = rt_se
;
8879 rt_se
->rt_rq
= &rq
->rt
;
8881 rt_se
->rt_rq
= parent
->my_q
;
8883 rt_se
->my_q
= rt_rq
;
8884 rt_se
->parent
= parent
;
8885 INIT_LIST_HEAD(&rt_se
->run_list
);
8889 void __init
sched_init(void)
8892 unsigned long alloc_size
= 0, ptr
;
8894 #ifdef CONFIG_FAIR_GROUP_SCHED
8895 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8897 #ifdef CONFIG_RT_GROUP_SCHED
8898 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8900 #ifdef CONFIG_USER_SCHED
8903 #ifdef CONFIG_CPUMASK_OFFSTACK
8904 alloc_size
+= num_possible_cpus() * cpumask_size();
8907 * As sched_init() is called before page_alloc is setup,
8908 * we use alloc_bootmem().
8911 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8913 #ifdef CONFIG_FAIR_GROUP_SCHED
8914 init_task_group
.se
= (struct sched_entity
**)ptr
;
8915 ptr
+= nr_cpu_ids
* sizeof(void **);
8917 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8918 ptr
+= nr_cpu_ids
* sizeof(void **);
8920 #ifdef CONFIG_USER_SCHED
8921 root_task_group
.se
= (struct sched_entity
**)ptr
;
8922 ptr
+= nr_cpu_ids
* sizeof(void **);
8924 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8925 ptr
+= nr_cpu_ids
* sizeof(void **);
8926 #endif /* CONFIG_USER_SCHED */
8927 #endif /* CONFIG_FAIR_GROUP_SCHED */
8928 #ifdef CONFIG_RT_GROUP_SCHED
8929 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8930 ptr
+= nr_cpu_ids
* sizeof(void **);
8932 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8933 ptr
+= nr_cpu_ids
* sizeof(void **);
8935 #ifdef CONFIG_USER_SCHED
8936 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8937 ptr
+= nr_cpu_ids
* sizeof(void **);
8939 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8940 ptr
+= nr_cpu_ids
* sizeof(void **);
8941 #endif /* CONFIG_USER_SCHED */
8942 #endif /* CONFIG_RT_GROUP_SCHED */
8943 #ifdef CONFIG_CPUMASK_OFFSTACK
8944 for_each_possible_cpu(i
) {
8945 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8946 ptr
+= cpumask_size();
8948 #endif /* CONFIG_CPUMASK_OFFSTACK */
8952 init_defrootdomain();
8955 init_rt_bandwidth(&def_rt_bandwidth
,
8956 global_rt_period(), global_rt_runtime());
8958 #ifdef CONFIG_RT_GROUP_SCHED
8959 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8960 global_rt_period(), global_rt_runtime());
8961 #ifdef CONFIG_USER_SCHED
8962 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8963 global_rt_period(), RUNTIME_INF
);
8964 #endif /* CONFIG_USER_SCHED */
8965 #endif /* CONFIG_RT_GROUP_SCHED */
8967 #ifdef CONFIG_GROUP_SCHED
8968 list_add(&init_task_group
.list
, &task_groups
);
8969 INIT_LIST_HEAD(&init_task_group
.children
);
8971 #ifdef CONFIG_USER_SCHED
8972 INIT_LIST_HEAD(&root_task_group
.children
);
8973 init_task_group
.parent
= &root_task_group
;
8974 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8975 #endif /* CONFIG_USER_SCHED */
8976 #endif /* CONFIG_GROUP_SCHED */
8978 for_each_possible_cpu(i
) {
8982 spin_lock_init(&rq
->lock
);
8984 init_cfs_rq(&rq
->cfs
, rq
);
8985 init_rt_rq(&rq
->rt
, rq
);
8986 #ifdef CONFIG_FAIR_GROUP_SCHED
8987 init_task_group
.shares
= init_task_group_load
;
8988 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8989 #ifdef CONFIG_CGROUP_SCHED
8991 * How much cpu bandwidth does init_task_group get?
8993 * In case of task-groups formed thr' the cgroup filesystem, it
8994 * gets 100% of the cpu resources in the system. This overall
8995 * system cpu resource is divided among the tasks of
8996 * init_task_group and its child task-groups in a fair manner,
8997 * based on each entity's (task or task-group's) weight
8998 * (se->load.weight).
9000 * In other words, if init_task_group has 10 tasks of weight
9001 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9002 * then A0's share of the cpu resource is:
9004 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9006 * We achieve this by letting init_task_group's tasks sit
9007 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9009 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9010 #elif defined CONFIG_USER_SCHED
9011 root_task_group
.shares
= NICE_0_LOAD
;
9012 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9014 * In case of task-groups formed thr' the user id of tasks,
9015 * init_task_group represents tasks belonging to root user.
9016 * Hence it forms a sibling of all subsequent groups formed.
9017 * In this case, init_task_group gets only a fraction of overall
9018 * system cpu resource, based on the weight assigned to root
9019 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9020 * by letting tasks of init_task_group sit in a separate cfs_rq
9021 * (init_cfs_rq) and having one entity represent this group of
9022 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9024 init_tg_cfs_entry(&init_task_group
,
9025 &per_cpu(init_cfs_rq
, i
),
9026 &per_cpu(init_sched_entity
, i
), i
, 1,
9027 root_task_group
.se
[i
]);
9030 #endif /* CONFIG_FAIR_GROUP_SCHED */
9032 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9033 #ifdef CONFIG_RT_GROUP_SCHED
9034 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9035 #ifdef CONFIG_CGROUP_SCHED
9036 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9037 #elif defined CONFIG_USER_SCHED
9038 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9039 init_tg_rt_entry(&init_task_group
,
9040 &per_cpu(init_rt_rq
, i
),
9041 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9042 root_task_group
.rt_se
[i
]);
9046 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9047 rq
->cpu_load
[j
] = 0;
9051 rq
->active_balance
= 0;
9052 rq
->next_balance
= jiffies
;
9056 rq
->migration_thread
= NULL
;
9057 INIT_LIST_HEAD(&rq
->migration_queue
);
9058 rq_attach_root(rq
, &def_root_domain
);
9061 atomic_set(&rq
->nr_iowait
, 0);
9064 set_load_weight(&init_task
);
9066 #ifdef CONFIG_PREEMPT_NOTIFIERS
9067 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9071 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9074 #ifdef CONFIG_RT_MUTEXES
9075 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9079 * The boot idle thread does lazy MMU switching as well:
9081 atomic_inc(&init_mm
.mm_count
);
9082 enter_lazy_tlb(&init_mm
, current
);
9085 * Make us the idle thread. Technically, schedule() should not be
9086 * called from this thread, however somewhere below it might be,
9087 * but because we are the idle thread, we just pick up running again
9088 * when this runqueue becomes "idle".
9090 init_idle(current
, smp_processor_id());
9092 * During early bootup we pretend to be a normal task:
9094 current
->sched_class
= &fair_sched_class
;
9096 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9097 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9100 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9102 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9105 perf_counter_init();
9107 scheduler_running
= 1;
9110 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9111 void __might_sleep(char *file
, int line
)
9114 static unsigned long prev_jiffy
; /* ratelimiting */
9116 if ((!in_atomic() && !irqs_disabled()) ||
9117 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9119 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9121 prev_jiffy
= jiffies
;
9124 "BUG: sleeping function called from invalid context at %s:%d\n",
9127 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9128 in_atomic(), irqs_disabled(),
9129 current
->pid
, current
->comm
);
9131 debug_show_held_locks(current
);
9132 if (irqs_disabled())
9133 print_irqtrace_events(current
);
9137 EXPORT_SYMBOL(__might_sleep
);
9140 #ifdef CONFIG_MAGIC_SYSRQ
9141 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9145 update_rq_clock(rq
);
9146 on_rq
= p
->se
.on_rq
;
9148 deactivate_task(rq
, p
, 0);
9149 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9151 activate_task(rq
, p
, 0);
9152 resched_task(rq
->curr
);
9156 void normalize_rt_tasks(void)
9158 struct task_struct
*g
, *p
;
9159 unsigned long flags
;
9162 read_lock_irqsave(&tasklist_lock
, flags
);
9163 do_each_thread(g
, p
) {
9165 * Only normalize user tasks:
9170 p
->se
.exec_start
= 0;
9171 #ifdef CONFIG_SCHEDSTATS
9172 p
->se
.wait_start
= 0;
9173 p
->se
.sleep_start
= 0;
9174 p
->se
.block_start
= 0;
9179 * Renice negative nice level userspace
9182 if (TASK_NICE(p
) < 0 && p
->mm
)
9183 set_user_nice(p
, 0);
9187 spin_lock(&p
->pi_lock
);
9188 rq
= __task_rq_lock(p
);
9190 normalize_task(rq
, p
);
9192 __task_rq_unlock(rq
);
9193 spin_unlock(&p
->pi_lock
);
9194 } while_each_thread(g
, p
);
9196 read_unlock_irqrestore(&tasklist_lock
, flags
);
9199 #endif /* CONFIG_MAGIC_SYSRQ */
9203 * These functions are only useful for the IA64 MCA handling.
9205 * They can only be called when the whole system has been
9206 * stopped - every CPU needs to be quiescent, and no scheduling
9207 * activity can take place. Using them for anything else would
9208 * be a serious bug, and as a result, they aren't even visible
9209 * under any other configuration.
9213 * curr_task - return the current task for a given cpu.
9214 * @cpu: the processor in question.
9216 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9218 struct task_struct
*curr_task(int cpu
)
9220 return cpu_curr(cpu
);
9224 * set_curr_task - set the current task for a given cpu.
9225 * @cpu: the processor in question.
9226 * @p: the task pointer to set.
9228 * Description: This function must only be used when non-maskable interrupts
9229 * are serviced on a separate stack. It allows the architecture to switch the
9230 * notion of the current task on a cpu in a non-blocking manner. This function
9231 * must be called with all CPU's synchronized, and interrupts disabled, the
9232 * and caller must save the original value of the current task (see
9233 * curr_task() above) and restore that value before reenabling interrupts and
9234 * re-starting the system.
9236 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9238 void set_curr_task(int cpu
, struct task_struct
*p
)
9245 #ifdef CONFIG_FAIR_GROUP_SCHED
9246 static void free_fair_sched_group(struct task_group
*tg
)
9250 for_each_possible_cpu(i
) {
9252 kfree(tg
->cfs_rq
[i
]);
9262 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9264 struct cfs_rq
*cfs_rq
;
9265 struct sched_entity
*se
;
9269 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9272 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9276 tg
->shares
= NICE_0_LOAD
;
9278 for_each_possible_cpu(i
) {
9281 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9282 GFP_KERNEL
, cpu_to_node(i
));
9286 se
= kzalloc_node(sizeof(struct sched_entity
),
9287 GFP_KERNEL
, cpu_to_node(i
));
9291 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9300 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9302 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9303 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9306 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9308 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9310 #else /* !CONFG_FAIR_GROUP_SCHED */
9311 static inline void free_fair_sched_group(struct task_group
*tg
)
9316 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9321 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9325 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9328 #endif /* CONFIG_FAIR_GROUP_SCHED */
9330 #ifdef CONFIG_RT_GROUP_SCHED
9331 static void free_rt_sched_group(struct task_group
*tg
)
9335 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9337 for_each_possible_cpu(i
) {
9339 kfree(tg
->rt_rq
[i
]);
9341 kfree(tg
->rt_se
[i
]);
9349 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9351 struct rt_rq
*rt_rq
;
9352 struct sched_rt_entity
*rt_se
;
9356 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9359 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9363 init_rt_bandwidth(&tg
->rt_bandwidth
,
9364 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9366 for_each_possible_cpu(i
) {
9369 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9370 GFP_KERNEL
, cpu_to_node(i
));
9374 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9375 GFP_KERNEL
, cpu_to_node(i
));
9379 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9388 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9390 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9391 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9394 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9396 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9398 #else /* !CONFIG_RT_GROUP_SCHED */
9399 static inline void free_rt_sched_group(struct task_group
*tg
)
9404 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9409 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9413 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9416 #endif /* CONFIG_RT_GROUP_SCHED */
9418 #ifdef CONFIG_GROUP_SCHED
9419 static void free_sched_group(struct task_group
*tg
)
9421 free_fair_sched_group(tg
);
9422 free_rt_sched_group(tg
);
9426 /* allocate runqueue etc for a new task group */
9427 struct task_group
*sched_create_group(struct task_group
*parent
)
9429 struct task_group
*tg
;
9430 unsigned long flags
;
9433 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9435 return ERR_PTR(-ENOMEM
);
9437 if (!alloc_fair_sched_group(tg
, parent
))
9440 if (!alloc_rt_sched_group(tg
, parent
))
9443 spin_lock_irqsave(&task_group_lock
, flags
);
9444 for_each_possible_cpu(i
) {
9445 register_fair_sched_group(tg
, i
);
9446 register_rt_sched_group(tg
, i
);
9448 list_add_rcu(&tg
->list
, &task_groups
);
9450 WARN_ON(!parent
); /* root should already exist */
9452 tg
->parent
= parent
;
9453 INIT_LIST_HEAD(&tg
->children
);
9454 list_add_rcu(&tg
->siblings
, &parent
->children
);
9455 spin_unlock_irqrestore(&task_group_lock
, flags
);
9460 free_sched_group(tg
);
9461 return ERR_PTR(-ENOMEM
);
9464 /* rcu callback to free various structures associated with a task group */
9465 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9467 /* now it should be safe to free those cfs_rqs */
9468 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9471 /* Destroy runqueue etc associated with a task group */
9472 void sched_destroy_group(struct task_group
*tg
)
9474 unsigned long flags
;
9477 spin_lock_irqsave(&task_group_lock
, flags
);
9478 for_each_possible_cpu(i
) {
9479 unregister_fair_sched_group(tg
, i
);
9480 unregister_rt_sched_group(tg
, i
);
9482 list_del_rcu(&tg
->list
);
9483 list_del_rcu(&tg
->siblings
);
9484 spin_unlock_irqrestore(&task_group_lock
, flags
);
9486 /* wait for possible concurrent references to cfs_rqs complete */
9487 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9490 /* change task's runqueue when it moves between groups.
9491 * The caller of this function should have put the task in its new group
9492 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9493 * reflect its new group.
9495 void sched_move_task(struct task_struct
*tsk
)
9498 unsigned long flags
;
9501 rq
= task_rq_lock(tsk
, &flags
);
9503 update_rq_clock(rq
);
9505 running
= task_current(rq
, tsk
);
9506 on_rq
= tsk
->se
.on_rq
;
9509 dequeue_task(rq
, tsk
, 0);
9510 if (unlikely(running
))
9511 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9513 set_task_rq(tsk
, task_cpu(tsk
));
9515 #ifdef CONFIG_FAIR_GROUP_SCHED
9516 if (tsk
->sched_class
->moved_group
)
9517 tsk
->sched_class
->moved_group(tsk
);
9520 if (unlikely(running
))
9521 tsk
->sched_class
->set_curr_task(rq
);
9523 enqueue_task(rq
, tsk
, 0);
9525 task_rq_unlock(rq
, &flags
);
9527 #endif /* CONFIG_GROUP_SCHED */
9529 #ifdef CONFIG_FAIR_GROUP_SCHED
9530 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9532 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9537 dequeue_entity(cfs_rq
, se
, 0);
9539 se
->load
.weight
= shares
;
9540 se
->load
.inv_weight
= 0;
9543 enqueue_entity(cfs_rq
, se
, 0);
9546 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9548 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9549 struct rq
*rq
= cfs_rq
->rq
;
9550 unsigned long flags
;
9552 spin_lock_irqsave(&rq
->lock
, flags
);
9553 __set_se_shares(se
, shares
);
9554 spin_unlock_irqrestore(&rq
->lock
, flags
);
9557 static DEFINE_MUTEX(shares_mutex
);
9559 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9562 unsigned long flags
;
9565 * We can't change the weight of the root cgroup.
9570 if (shares
< MIN_SHARES
)
9571 shares
= MIN_SHARES
;
9572 else if (shares
> MAX_SHARES
)
9573 shares
= MAX_SHARES
;
9575 mutex_lock(&shares_mutex
);
9576 if (tg
->shares
== shares
)
9579 spin_lock_irqsave(&task_group_lock
, flags
);
9580 for_each_possible_cpu(i
)
9581 unregister_fair_sched_group(tg
, i
);
9582 list_del_rcu(&tg
->siblings
);
9583 spin_unlock_irqrestore(&task_group_lock
, flags
);
9585 /* wait for any ongoing reference to this group to finish */
9586 synchronize_sched();
9589 * Now we are free to modify the group's share on each cpu
9590 * w/o tripping rebalance_share or load_balance_fair.
9592 tg
->shares
= shares
;
9593 for_each_possible_cpu(i
) {
9597 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9598 set_se_shares(tg
->se
[i
], shares
);
9602 * Enable load balance activity on this group, by inserting it back on
9603 * each cpu's rq->leaf_cfs_rq_list.
9605 spin_lock_irqsave(&task_group_lock
, flags
);
9606 for_each_possible_cpu(i
)
9607 register_fair_sched_group(tg
, i
);
9608 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9609 spin_unlock_irqrestore(&task_group_lock
, flags
);
9611 mutex_unlock(&shares_mutex
);
9615 unsigned long sched_group_shares(struct task_group
*tg
)
9621 #ifdef CONFIG_RT_GROUP_SCHED
9623 * Ensure that the real time constraints are schedulable.
9625 static DEFINE_MUTEX(rt_constraints_mutex
);
9627 static unsigned long to_ratio(u64 period
, u64 runtime
)
9629 if (runtime
== RUNTIME_INF
)
9632 return div64_u64(runtime
<< 20, period
);
9635 /* Must be called with tasklist_lock held */
9636 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9638 struct task_struct
*g
, *p
;
9640 do_each_thread(g
, p
) {
9641 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9643 } while_each_thread(g
, p
);
9648 struct rt_schedulable_data
{
9649 struct task_group
*tg
;
9654 static int tg_schedulable(struct task_group
*tg
, void *data
)
9656 struct rt_schedulable_data
*d
= data
;
9657 struct task_group
*child
;
9658 unsigned long total
, sum
= 0;
9659 u64 period
, runtime
;
9661 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9662 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9665 period
= d
->rt_period
;
9666 runtime
= d
->rt_runtime
;
9669 #ifdef CONFIG_USER_SCHED
9670 if (tg
== &root_task_group
) {
9671 period
= global_rt_period();
9672 runtime
= global_rt_runtime();
9677 * Cannot have more runtime than the period.
9679 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9683 * Ensure we don't starve existing RT tasks.
9685 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9688 total
= to_ratio(period
, runtime
);
9691 * Nobody can have more than the global setting allows.
9693 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9697 * The sum of our children's runtime should not exceed our own.
9699 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9700 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9701 runtime
= child
->rt_bandwidth
.rt_runtime
;
9703 if (child
== d
->tg
) {
9704 period
= d
->rt_period
;
9705 runtime
= d
->rt_runtime
;
9708 sum
+= to_ratio(period
, runtime
);
9717 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9719 struct rt_schedulable_data data
= {
9721 .rt_period
= period
,
9722 .rt_runtime
= runtime
,
9725 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9728 static int tg_set_bandwidth(struct task_group
*tg
,
9729 u64 rt_period
, u64 rt_runtime
)
9733 mutex_lock(&rt_constraints_mutex
);
9734 read_lock(&tasklist_lock
);
9735 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9739 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9740 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9741 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9743 for_each_possible_cpu(i
) {
9744 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9746 spin_lock(&rt_rq
->rt_runtime_lock
);
9747 rt_rq
->rt_runtime
= rt_runtime
;
9748 spin_unlock(&rt_rq
->rt_runtime_lock
);
9750 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9752 read_unlock(&tasklist_lock
);
9753 mutex_unlock(&rt_constraints_mutex
);
9758 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9760 u64 rt_runtime
, rt_period
;
9762 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9763 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9764 if (rt_runtime_us
< 0)
9765 rt_runtime
= RUNTIME_INF
;
9767 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9770 long sched_group_rt_runtime(struct task_group
*tg
)
9774 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9777 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9778 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9779 return rt_runtime_us
;
9782 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9784 u64 rt_runtime
, rt_period
;
9786 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9787 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9792 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9795 long sched_group_rt_period(struct task_group
*tg
)
9799 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9800 do_div(rt_period_us
, NSEC_PER_USEC
);
9801 return rt_period_us
;
9804 static int sched_rt_global_constraints(void)
9806 u64 runtime
, period
;
9809 if (sysctl_sched_rt_period
<= 0)
9812 runtime
= global_rt_runtime();
9813 period
= global_rt_period();
9816 * Sanity check on the sysctl variables.
9818 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9821 mutex_lock(&rt_constraints_mutex
);
9822 read_lock(&tasklist_lock
);
9823 ret
= __rt_schedulable(NULL
, 0, 0);
9824 read_unlock(&tasklist_lock
);
9825 mutex_unlock(&rt_constraints_mutex
);
9830 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9832 /* Don't accept realtime tasks when there is no way for them to run */
9833 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9839 #else /* !CONFIG_RT_GROUP_SCHED */
9840 static int sched_rt_global_constraints(void)
9842 unsigned long flags
;
9845 if (sysctl_sched_rt_period
<= 0)
9848 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9849 for_each_possible_cpu(i
) {
9850 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9852 spin_lock(&rt_rq
->rt_runtime_lock
);
9853 rt_rq
->rt_runtime
= global_rt_runtime();
9854 spin_unlock(&rt_rq
->rt_runtime_lock
);
9856 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9860 #endif /* CONFIG_RT_GROUP_SCHED */
9862 int sched_rt_handler(struct ctl_table
*table
, int write
,
9863 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9867 int old_period
, old_runtime
;
9868 static DEFINE_MUTEX(mutex
);
9871 old_period
= sysctl_sched_rt_period
;
9872 old_runtime
= sysctl_sched_rt_runtime
;
9874 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9876 if (!ret
&& write
) {
9877 ret
= sched_rt_global_constraints();
9879 sysctl_sched_rt_period
= old_period
;
9880 sysctl_sched_rt_runtime
= old_runtime
;
9882 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9883 def_rt_bandwidth
.rt_period
=
9884 ns_to_ktime(global_rt_period());
9887 mutex_unlock(&mutex
);
9892 #ifdef CONFIG_CGROUP_SCHED
9894 /* return corresponding task_group object of a cgroup */
9895 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9897 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9898 struct task_group
, css
);
9901 static struct cgroup_subsys_state
*
9902 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9904 struct task_group
*tg
, *parent
;
9906 if (!cgrp
->parent
) {
9907 /* This is early initialization for the top cgroup */
9908 return &init_task_group
.css
;
9911 parent
= cgroup_tg(cgrp
->parent
);
9912 tg
= sched_create_group(parent
);
9914 return ERR_PTR(-ENOMEM
);
9920 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9922 struct task_group
*tg
= cgroup_tg(cgrp
);
9924 sched_destroy_group(tg
);
9928 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9929 struct task_struct
*tsk
)
9931 #ifdef CONFIG_RT_GROUP_SCHED
9932 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9935 /* We don't support RT-tasks being in separate groups */
9936 if (tsk
->sched_class
!= &fair_sched_class
)
9944 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9945 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9947 sched_move_task(tsk
);
9950 #ifdef CONFIG_FAIR_GROUP_SCHED
9951 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9954 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9957 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9959 struct task_group
*tg
= cgroup_tg(cgrp
);
9961 return (u64
) tg
->shares
;
9963 #endif /* CONFIG_FAIR_GROUP_SCHED */
9965 #ifdef CONFIG_RT_GROUP_SCHED
9966 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9969 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9972 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9974 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9977 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9980 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9983 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9985 return sched_group_rt_period(cgroup_tg(cgrp
));
9987 #endif /* CONFIG_RT_GROUP_SCHED */
9989 static struct cftype cpu_files
[] = {
9990 #ifdef CONFIG_FAIR_GROUP_SCHED
9993 .read_u64
= cpu_shares_read_u64
,
9994 .write_u64
= cpu_shares_write_u64
,
9997 #ifdef CONFIG_RT_GROUP_SCHED
9999 .name
= "rt_runtime_us",
10000 .read_s64
= cpu_rt_runtime_read
,
10001 .write_s64
= cpu_rt_runtime_write
,
10004 .name
= "rt_period_us",
10005 .read_u64
= cpu_rt_period_read_uint
,
10006 .write_u64
= cpu_rt_period_write_uint
,
10011 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10013 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10016 struct cgroup_subsys cpu_cgroup_subsys
= {
10018 .create
= cpu_cgroup_create
,
10019 .destroy
= cpu_cgroup_destroy
,
10020 .can_attach
= cpu_cgroup_can_attach
,
10021 .attach
= cpu_cgroup_attach
,
10022 .populate
= cpu_cgroup_populate
,
10023 .subsys_id
= cpu_cgroup_subsys_id
,
10027 #endif /* CONFIG_CGROUP_SCHED */
10029 #ifdef CONFIG_CGROUP_CPUACCT
10032 * CPU accounting code for task groups.
10034 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10035 * (balbir@in.ibm.com).
10038 /* track cpu usage of a group of tasks and its child groups */
10040 struct cgroup_subsys_state css
;
10041 /* cpuusage holds pointer to a u64-type object on every cpu */
10043 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10044 struct cpuacct
*parent
;
10047 struct cgroup_subsys cpuacct_subsys
;
10049 /* return cpu accounting group corresponding to this container */
10050 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10052 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10053 struct cpuacct
, css
);
10056 /* return cpu accounting group to which this task belongs */
10057 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10059 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10060 struct cpuacct
, css
);
10063 /* create a new cpu accounting group */
10064 static struct cgroup_subsys_state
*cpuacct_create(
10065 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10067 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10073 ca
->cpuusage
= alloc_percpu(u64
);
10077 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10078 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10079 goto out_free_counters
;
10082 ca
->parent
= cgroup_ca(cgrp
->parent
);
10088 percpu_counter_destroy(&ca
->cpustat
[i
]);
10089 free_percpu(ca
->cpuusage
);
10093 return ERR_PTR(-ENOMEM
);
10096 /* destroy an existing cpu accounting group */
10098 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10100 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10103 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10104 percpu_counter_destroy(&ca
->cpustat
[i
]);
10105 free_percpu(ca
->cpuusage
);
10109 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10111 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10114 #ifndef CONFIG_64BIT
10116 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10118 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10120 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10128 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10130 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10132 #ifndef CONFIG_64BIT
10134 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10136 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10138 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10144 /* return total cpu usage (in nanoseconds) of a group */
10145 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10147 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10148 u64 totalcpuusage
= 0;
10151 for_each_present_cpu(i
)
10152 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10154 return totalcpuusage
;
10157 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10160 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10169 for_each_present_cpu(i
)
10170 cpuacct_cpuusage_write(ca
, i
, 0);
10176 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10177 struct seq_file
*m
)
10179 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10183 for_each_present_cpu(i
) {
10184 percpu
= cpuacct_cpuusage_read(ca
, i
);
10185 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10187 seq_printf(m
, "\n");
10191 static const char *cpuacct_stat_desc
[] = {
10192 [CPUACCT_STAT_USER
] = "user",
10193 [CPUACCT_STAT_SYSTEM
] = "system",
10196 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10197 struct cgroup_map_cb
*cb
)
10199 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10202 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10203 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10204 val
= cputime64_to_clock_t(val
);
10205 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10210 static struct cftype files
[] = {
10213 .read_u64
= cpuusage_read
,
10214 .write_u64
= cpuusage_write
,
10217 .name
= "usage_percpu",
10218 .read_seq_string
= cpuacct_percpu_seq_read
,
10222 .read_map
= cpuacct_stats_show
,
10226 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10228 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10232 * charge this task's execution time to its accounting group.
10234 * called with rq->lock held.
10236 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10238 struct cpuacct
*ca
;
10241 if (unlikely(!cpuacct_subsys
.active
))
10244 cpu
= task_cpu(tsk
);
10250 for (; ca
; ca
= ca
->parent
) {
10251 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10252 *cpuusage
+= cputime
;
10259 * Charge the system/user time to the task's accounting group.
10261 static void cpuacct_update_stats(struct task_struct
*tsk
,
10262 enum cpuacct_stat_index idx
, cputime_t val
)
10264 struct cpuacct
*ca
;
10266 if (unlikely(!cpuacct_subsys
.active
))
10273 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10279 struct cgroup_subsys cpuacct_subsys
= {
10281 .create
= cpuacct_create
,
10282 .destroy
= cpuacct_destroy
,
10283 .populate
= cpuacct_populate
,
10284 .subsys_id
= cpuacct_subsys_id
,
10286 #endif /* CONFIG_CGROUP_CPUACCT */