4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
124 static inline int rt_policy(int policy
)
126 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
131 static inline int task_has_rt_policy(struct task_struct
*p
)
133 return rt_policy(p
->policy
);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array
{
140 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
141 struct list_head queue
[MAX_RT_PRIO
];
144 struct rt_bandwidth
{
145 /* nests inside the rq lock: */
146 spinlock_t rt_runtime_lock
;
149 struct hrtimer rt_period_timer
;
152 static struct rt_bandwidth def_rt_bandwidth
;
154 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
156 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
158 struct rt_bandwidth
*rt_b
=
159 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
165 now
= hrtimer_cb_get_time(timer
);
166 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
171 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
174 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
178 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
180 rt_b
->rt_period
= ns_to_ktime(period
);
181 rt_b
->rt_runtime
= runtime
;
183 spin_lock_init(&rt_b
->rt_runtime_lock
);
185 hrtimer_init(&rt_b
->rt_period_timer
,
186 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
187 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime
>= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
199 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
202 if (hrtimer_active(&rt_b
->rt_period_timer
))
205 spin_lock(&rt_b
->rt_runtime_lock
);
210 if (hrtimer_active(&rt_b
->rt_period_timer
))
213 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
214 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
216 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
217 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
218 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
219 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
220 HRTIMER_MODE_ABS_PINNED
, 0);
222 spin_unlock(&rt_b
->rt_runtime_lock
);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
228 hrtimer_cancel(&rt_b
->rt_period_timer
);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex
);
238 #ifdef CONFIG_GROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups
);
246 /* task group related information */
248 #ifdef CONFIG_CGROUP_SCHED
249 struct cgroup_subsys_state css
;
252 #ifdef CONFIG_USER_SCHED
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity
**se
;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq
**cfs_rq
;
261 unsigned long shares
;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity
**rt_se
;
266 struct rt_rq
**rt_rq
;
268 struct rt_bandwidth rt_bandwidth
;
272 struct list_head list
;
274 struct task_group
*parent
;
275 struct list_head siblings
;
276 struct list_head children
;
279 #ifdef CONFIG_USER_SCHED
281 /* Helper function to pass uid information to create_sched_user() */
282 void set_tg_uid(struct user_struct
*user
)
284 user
->tg
->uid
= user
->uid
;
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group
;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU(struct cfs_rq
, init_tg_cfs_rq
) ____cacheline_aligned_in_smp
;
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
303 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock
);
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group
.children
);
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 #ifdef CONFIG_USER_SCHED
323 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
324 #else /* !CONFIG_USER_SCHED */
325 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
326 #endif /* CONFIG_USER_SCHED */
329 * A weight of 0 or 1 can cause arithmetics problems.
330 * A weight of a cfs_rq is the sum of weights of which entities
331 * are queued on this cfs_rq, so a weight of a entity should not be
332 * too large, so as the shares value of a task group.
333 * (The default weight is 1024 - so there's no practical
334 * limitation from this.)
337 #define MAX_SHARES (1UL << 18)
339 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
342 /* Default task group.
343 * Every task in system belong to this group at bootup.
345 struct task_group init_task_group
;
347 /* return group to which a task belongs */
348 static inline struct task_group
*task_group(struct task_struct
*p
)
350 struct task_group
*tg
;
352 #ifdef CONFIG_USER_SCHED
354 tg
= __task_cred(p
)->user
->tg
;
356 #elif defined(CONFIG_CGROUP_SCHED)
357 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
358 struct task_group
, css
);
360 tg
= &init_task_group
;
365 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
366 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
370 p
->se
.parent
= task_group(p
)->se
[cpu
];
373 #ifdef CONFIG_RT_GROUP_SCHED
374 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
375 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
382 static int root_task_group_empty(void)
388 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
389 static inline struct task_group
*task_group(struct task_struct
*p
)
394 #endif /* CONFIG_GROUP_SCHED */
396 /* CFS-related fields in a runqueue */
398 struct load_weight load
;
399 unsigned long nr_running
;
404 struct rb_root tasks_timeline
;
405 struct rb_node
*rb_leftmost
;
407 struct list_head tasks
;
408 struct list_head
*balance_iterator
;
411 * 'curr' points to currently running entity on this cfs_rq.
412 * It is set to NULL otherwise (i.e when none are currently running).
414 struct sched_entity
*curr
, *next
, *last
;
416 unsigned int nr_spread_over
;
418 #ifdef CONFIG_FAIR_GROUP_SCHED
419 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
422 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
423 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
424 * (like users, containers etc.)
426 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
427 * list is used during load balance.
429 struct list_head leaf_cfs_rq_list
;
430 struct task_group
*tg
; /* group that "owns" this runqueue */
434 * the part of load.weight contributed by tasks
436 unsigned long task_weight
;
439 * h_load = weight * f(tg)
441 * Where f(tg) is the recursive weight fraction assigned to
444 unsigned long h_load
;
447 * this cpu's part of tg->shares
449 unsigned long shares
;
452 * load.weight at the time we set shares
454 unsigned long rq_weight
;
459 /* Real-Time classes' related field in a runqueue: */
461 struct rt_prio_array active
;
462 unsigned long rt_nr_running
;
463 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
465 int curr
; /* highest queued rt task prio */
467 int next
; /* next highest */
472 unsigned long rt_nr_migratory
;
473 unsigned long rt_nr_total
;
475 struct plist_head pushable_tasks
;
480 /* Nests inside the rq lock: */
481 spinlock_t rt_runtime_lock
;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 unsigned long rt_nr_boosted
;
487 struct list_head leaf_rt_rq_list
;
488 struct task_group
*tg
;
489 struct sched_rt_entity
*rt_se
;
496 * We add the notion of a root-domain which will be used to define per-domain
497 * variables. Each exclusive cpuset essentially defines an island domain by
498 * fully partitioning the member cpus from any other cpuset. Whenever a new
499 * exclusive cpuset is created, we also create and attach a new root-domain
506 cpumask_var_t online
;
509 * The "RT overload" flag: it gets set if a CPU has more than
510 * one runnable RT task.
512 cpumask_var_t rto_mask
;
515 struct cpupri cpupri
;
517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
519 * Preferred wake up cpu nominated by sched_mc balance that will be
520 * used when most cpus are idle in the system indicating overall very
521 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
523 unsigned int sched_mc_preferred_wakeup_cpu
;
528 * By default the system creates a single root-domain with all cpus as
529 * members (mimicking the global state we have today).
531 static struct root_domain def_root_domain
;
536 * This is the main, per-CPU runqueue data structure.
538 * Locking rule: those places that want to lock multiple runqueues
539 * (such as the load balancing or the thread migration code), lock
540 * acquire operations must be ordered by ascending &runqueue.
547 * nr_running and cpu_load should be in the same cacheline because
548 * remote CPUs use both these fields when doing load calculation.
550 unsigned long nr_running
;
551 #define CPU_LOAD_IDX_MAX 5
552 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
554 unsigned long last_tick_seen
;
555 unsigned char in_nohz_recently
;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load
;
559 unsigned long nr_load_updates
;
561 u64 nr_migrations_in
;
566 #ifdef CONFIG_FAIR_GROUP_SCHED
567 /* list of leaf cfs_rq on this cpu: */
568 struct list_head leaf_cfs_rq_list
;
570 #ifdef CONFIG_RT_GROUP_SCHED
571 struct list_head leaf_rt_rq_list
;
575 * This is part of a global counter where only the total sum
576 * over all CPUs matters. A task can increase this counter on
577 * one CPU and if it got migrated afterwards it may decrease
578 * it on another CPU. Always updated under the runqueue lock:
580 unsigned long nr_uninterruptible
;
582 struct task_struct
*curr
, *idle
;
583 unsigned long next_balance
;
584 struct mm_struct
*prev_mm
;
591 struct root_domain
*rd
;
592 struct sched_domain
*sd
;
594 unsigned char idle_at_tick
;
595 /* For active balancing */
599 /* cpu of this runqueue: */
603 unsigned long avg_load_per_task
;
605 struct task_struct
*migration_thread
;
606 struct list_head migration_queue
;
612 /* calc_load related fields */
613 unsigned long calc_load_update
;
614 long calc_load_active
;
616 #ifdef CONFIG_SCHED_HRTICK
618 int hrtick_csd_pending
;
619 struct call_single_data hrtick_csd
;
621 struct hrtimer hrtick_timer
;
624 #ifdef CONFIG_SCHEDSTATS
626 struct sched_info rq_sched_info
;
627 unsigned long long rq_cpu_time
;
628 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
630 /* sys_sched_yield() stats */
631 unsigned int yld_count
;
633 /* schedule() stats */
634 unsigned int sched_switch
;
635 unsigned int sched_count
;
636 unsigned int sched_goidle
;
638 /* try_to_wake_up() stats */
639 unsigned int ttwu_count
;
640 unsigned int ttwu_local
;
643 unsigned int bkl_count
;
647 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
649 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
651 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
654 static inline int cpu_of(struct rq
*rq
)
664 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
665 * See detach_destroy_domains: synchronize_sched for details.
667 * The domain tree of any CPU may only be accessed from within
668 * preempt-disabled sections.
670 #define for_each_domain(cpu, __sd) \
671 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
673 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
674 #define this_rq() (&__get_cpu_var(runqueues))
675 #define task_rq(p) cpu_rq(task_cpu(p))
676 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
677 #define raw_rq() (&__raw_get_cpu_var(runqueues))
679 inline void update_rq_clock(struct rq
*rq
)
681 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
685 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
687 #ifdef CONFIG_SCHED_DEBUG
688 # define const_debug __read_mostly
690 # define const_debug static const
696 * Returns true if the current cpu runqueue is locked.
697 * This interface allows printk to be called with the runqueue lock
698 * held and know whether or not it is OK to wake up the klogd.
700 int runqueue_is_locked(void)
703 struct rq
*rq
= cpu_rq(cpu
);
706 ret
= spin_is_locked(&rq
->lock
);
712 * Debugging: various feature bits
715 #define SCHED_FEAT(name, enabled) \
716 __SCHED_FEAT_##name ,
719 #include "sched_features.h"
724 #define SCHED_FEAT(name, enabled) \
725 (1UL << __SCHED_FEAT_##name) * enabled |
727 const_debug
unsigned int sysctl_sched_features
=
728 #include "sched_features.h"
733 #ifdef CONFIG_SCHED_DEBUG
734 #define SCHED_FEAT(name, enabled) \
737 static __read_mostly
char *sched_feat_names
[] = {
738 #include "sched_features.h"
744 static int sched_feat_show(struct seq_file
*m
, void *v
)
748 for (i
= 0; sched_feat_names
[i
]; i
++) {
749 if (!(sysctl_sched_features
& (1UL << i
)))
751 seq_printf(m
, "%s ", sched_feat_names
[i
]);
759 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
760 size_t cnt
, loff_t
*ppos
)
770 if (copy_from_user(&buf
, ubuf
, cnt
))
775 if (strncmp(buf
, "NO_", 3) == 0) {
780 for (i
= 0; sched_feat_names
[i
]; i
++) {
781 int len
= strlen(sched_feat_names
[i
]);
783 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
785 sysctl_sched_features
&= ~(1UL << i
);
787 sysctl_sched_features
|= (1UL << i
);
792 if (!sched_feat_names
[i
])
800 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
802 return single_open(filp
, sched_feat_show
, NULL
);
805 static struct file_operations sched_feat_fops
= {
806 .open
= sched_feat_open
,
807 .write
= sched_feat_write
,
810 .release
= single_release
,
813 static __init
int sched_init_debug(void)
815 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
820 late_initcall(sched_init_debug
);
824 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
827 * Number of tasks to iterate in a single balance run.
828 * Limited because this is done with IRQs disabled.
830 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
833 * ratelimit for updating the group shares.
836 unsigned int sysctl_sched_shares_ratelimit
= 250000;
839 * Inject some fuzzyness into changing the per-cpu group shares
840 * this avoids remote rq-locks at the expense of fairness.
843 unsigned int sysctl_sched_shares_thresh
= 4;
846 * period over which we average the RT time consumption, measured
851 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
854 * period over which we measure -rt task cpu usage in us.
857 unsigned int sysctl_sched_rt_period
= 1000000;
859 static __read_mostly
int scheduler_running
;
862 * part of the period that we allow rt tasks to run in us.
865 int sysctl_sched_rt_runtime
= 950000;
867 static inline u64
global_rt_period(void)
869 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
872 static inline u64
global_rt_runtime(void)
874 if (sysctl_sched_rt_runtime
< 0)
877 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
880 #ifndef prepare_arch_switch
881 # define prepare_arch_switch(next) do { } while (0)
883 #ifndef finish_arch_switch
884 # define finish_arch_switch(prev) do { } while (0)
887 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
889 return rq
->curr
== p
;
892 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
893 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
895 return task_current(rq
, p
);
898 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
902 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
904 #ifdef CONFIG_DEBUG_SPINLOCK
905 /* this is a valid case when another task releases the spinlock */
906 rq
->lock
.owner
= current
;
909 * If we are tracking spinlock dependencies then we have to
910 * fix up the runqueue lock - which gets 'carried over' from
913 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
915 spin_unlock_irq(&rq
->lock
);
918 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
919 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
924 return task_current(rq
, p
);
928 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
932 * We can optimise this out completely for !SMP, because the
933 * SMP rebalancing from interrupt is the only thing that cares
938 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
939 spin_unlock_irq(&rq
->lock
);
941 spin_unlock(&rq
->lock
);
945 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
949 * After ->oncpu is cleared, the task can be moved to a different CPU.
950 * We must ensure this doesn't happen until the switch is completely
956 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
960 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
963 * __task_rq_lock - lock the runqueue a given task resides on.
964 * Must be called interrupts disabled.
966 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
970 struct rq
*rq
= task_rq(p
);
971 spin_lock(&rq
->lock
);
972 if (likely(rq
== task_rq(p
)))
974 spin_unlock(&rq
->lock
);
979 * task_rq_lock - lock the runqueue a given task resides on and disable
980 * interrupts. Note the ordering: we can safely lookup the task_rq without
981 * explicitly disabling preemption.
983 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
989 local_irq_save(*flags
);
991 spin_lock(&rq
->lock
);
992 if (likely(rq
== task_rq(p
)))
994 spin_unlock_irqrestore(&rq
->lock
, *flags
);
998 void task_rq_unlock_wait(struct task_struct
*p
)
1000 struct rq
*rq
= task_rq(p
);
1002 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1003 spin_unlock_wait(&rq
->lock
);
1006 static void __task_rq_unlock(struct rq
*rq
)
1007 __releases(rq
->lock
)
1009 spin_unlock(&rq
->lock
);
1012 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1013 __releases(rq
->lock
)
1015 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1019 * this_rq_lock - lock this runqueue and disable interrupts.
1021 static struct rq
*this_rq_lock(void)
1022 __acquires(rq
->lock
)
1026 local_irq_disable();
1028 spin_lock(&rq
->lock
);
1033 #ifdef CONFIG_SCHED_HRTICK
1035 * Use HR-timers to deliver accurate preemption points.
1037 * Its all a bit involved since we cannot program an hrt while holding the
1038 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1041 * When we get rescheduled we reprogram the hrtick_timer outside of the
1047 * - enabled by features
1048 * - hrtimer is actually high res
1050 static inline int hrtick_enabled(struct rq
*rq
)
1052 if (!sched_feat(HRTICK
))
1054 if (!cpu_active(cpu_of(rq
)))
1056 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1059 static void hrtick_clear(struct rq
*rq
)
1061 if (hrtimer_active(&rq
->hrtick_timer
))
1062 hrtimer_cancel(&rq
->hrtick_timer
);
1066 * High-resolution timer tick.
1067 * Runs from hardirq context with interrupts disabled.
1069 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1071 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1073 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1075 spin_lock(&rq
->lock
);
1076 update_rq_clock(rq
);
1077 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1078 spin_unlock(&rq
->lock
);
1080 return HRTIMER_NORESTART
;
1085 * called from hardirq (IPI) context
1087 static void __hrtick_start(void *arg
)
1089 struct rq
*rq
= arg
;
1091 spin_lock(&rq
->lock
);
1092 hrtimer_restart(&rq
->hrtick_timer
);
1093 rq
->hrtick_csd_pending
= 0;
1094 spin_unlock(&rq
->lock
);
1098 * Called to set the hrtick timer state.
1100 * called with rq->lock held and irqs disabled
1102 static void hrtick_start(struct rq
*rq
, u64 delay
)
1104 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1105 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1107 hrtimer_set_expires(timer
, time
);
1109 if (rq
== this_rq()) {
1110 hrtimer_restart(timer
);
1111 } else if (!rq
->hrtick_csd_pending
) {
1112 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1113 rq
->hrtick_csd_pending
= 1;
1118 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1120 int cpu
= (int)(long)hcpu
;
1123 case CPU_UP_CANCELED
:
1124 case CPU_UP_CANCELED_FROZEN
:
1125 case CPU_DOWN_PREPARE
:
1126 case CPU_DOWN_PREPARE_FROZEN
:
1128 case CPU_DEAD_FROZEN
:
1129 hrtick_clear(cpu_rq(cpu
));
1136 static __init
void init_hrtick(void)
1138 hotcpu_notifier(hotplug_hrtick
, 0);
1142 * Called to set the hrtick timer state.
1144 * called with rq->lock held and irqs disabled
1146 static void hrtick_start(struct rq
*rq
, u64 delay
)
1148 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1149 HRTIMER_MODE_REL_PINNED
, 0);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq
*rq
)
1160 rq
->hrtick_csd_pending
= 0;
1162 rq
->hrtick_csd
.flags
= 0;
1163 rq
->hrtick_csd
.func
= __hrtick_start
;
1164 rq
->hrtick_csd
.info
= rq
;
1167 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1168 rq
->hrtick_timer
.function
= hrtick
;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq
*rq
)
1175 static inline void init_rq_hrtick(struct rq
*rq
)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct
*p
)
1201 assert_spin_locked(&task_rq(p
)->lock
);
1203 if (test_tsk_need_resched(p
))
1206 set_tsk_need_resched(p
);
1209 if (cpu
== smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p
))
1215 smp_send_reschedule(cpu
);
1218 static void resched_cpu(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1221 unsigned long flags
;
1223 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1225 resched_task(cpu_curr(cpu
));
1226 spin_unlock_irqrestore(&rq
->lock
, flags
);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu
)
1242 struct rq
*rq
= cpu_rq(cpu
);
1244 if (cpu
== smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq
->curr
!= rq
->idle
)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_need_resched(rq
->idle
);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq
->idle
))
1267 smp_send_reschedule(cpu
);
1269 #endif /* CONFIG_NO_HZ */
1271 static u64
sched_avg_period(void)
1273 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1276 static void sched_avg_update(struct rq
*rq
)
1278 s64 period
= sched_avg_period();
1280 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1281 rq
->age_stamp
+= period
;
1286 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1288 rq
->rt_avg
+= rt_delta
;
1289 sched_avg_update(rq
);
1292 #else /* !CONFIG_SMP */
1293 static void resched_task(struct task_struct
*p
)
1295 assert_spin_locked(&task_rq(p
)->lock
);
1296 set_tsk_need_resched(p
);
1299 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1302 #endif /* CONFIG_SMP */
1304 #if BITS_PER_LONG == 32
1305 # define WMULT_CONST (~0UL)
1307 # define WMULT_CONST (1UL << 32)
1310 #define WMULT_SHIFT 32
1313 * Shift right and round:
1315 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1318 * delta *= weight / lw
1320 static unsigned long
1321 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1322 struct load_weight
*lw
)
1326 if (!lw
->inv_weight
) {
1327 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1330 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1334 tmp
= (u64
)delta_exec
* weight
;
1336 * Check whether we'd overflow the 64-bit multiplication:
1338 if (unlikely(tmp
> WMULT_CONST
))
1339 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1342 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1344 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1347 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1353 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1360 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1361 * of tasks with abnormal "nice" values across CPUs the contribution that
1362 * each task makes to its run queue's load is weighted according to its
1363 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1364 * scaled version of the new time slice allocation that they receive on time
1368 #define WEIGHT_IDLEPRIO 3
1369 #define WMULT_IDLEPRIO 1431655765
1372 * Nice levels are multiplicative, with a gentle 10% change for every
1373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1375 * that remained on nice 0.
1377 * The "10% effect" is relative and cumulative: from _any_ nice level,
1378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1380 * If a task goes up by ~10% and another task goes down by ~10% then
1381 * the relative distance between them is ~25%.)
1383 static const int prio_to_weight
[40] = {
1384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1388 /* 0 */ 1024, 820, 655, 526, 423,
1389 /* 5 */ 335, 272, 215, 172, 137,
1390 /* 10 */ 110, 87, 70, 56, 45,
1391 /* 15 */ 36, 29, 23, 18, 15,
1395 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1397 * In cases where the weight does not change often, we can use the
1398 * precalculated inverse to speed up arithmetics by turning divisions
1399 * into multiplications:
1401 static const u32 prio_to_wmult
[40] = {
1402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1412 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1415 * runqueue iterator, to support SMP load-balancing between different
1416 * scheduling classes, without having to expose their internal data
1417 * structures to the load-balancing proper:
1419 struct rq_iterator
{
1421 struct task_struct
*(*start
)(void *);
1422 struct task_struct
*(*next
)(void *);
1426 static unsigned long
1427 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1428 unsigned long max_load_move
, struct sched_domain
*sd
,
1429 enum cpu_idle_type idle
, int *all_pinned
,
1430 int *this_best_prio
, struct rq_iterator
*iterator
);
1433 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1434 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1435 struct rq_iterator
*iterator
);
1438 /* Time spent by the tasks of the cpu accounting group executing in ... */
1439 enum cpuacct_stat_index
{
1440 CPUACCT_STAT_USER
, /* ... user mode */
1441 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1443 CPUACCT_STAT_NSTATS
,
1446 #ifdef CONFIG_CGROUP_CPUACCT
1447 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1448 static void cpuacct_update_stats(struct task_struct
*tsk
,
1449 enum cpuacct_stat_index idx
, cputime_t val
);
1451 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1452 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1453 enum cpuacct_stat_index idx
, cputime_t val
) {}
1456 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1458 update_load_add(&rq
->load
, load
);
1461 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1463 update_load_sub(&rq
->load
, load
);
1466 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1467 typedef int (*tg_visitor
)(struct task_group
*, void *);
1470 * Iterate the full tree, calling @down when first entering a node and @up when
1471 * leaving it for the final time.
1473 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1475 struct task_group
*parent
, *child
;
1479 parent
= &root_task_group
;
1481 ret
= (*down
)(parent
, data
);
1484 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1491 ret
= (*up
)(parent
, data
);
1496 parent
= parent
->parent
;
1505 static int tg_nop(struct task_group
*tg
, void *data
)
1512 static unsigned long source_load(int cpu
, int type
);
1513 static unsigned long target_load(int cpu
, int type
);
1514 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1516 static unsigned long cpu_avg_load_per_task(int cpu
)
1518 struct rq
*rq
= cpu_rq(cpu
);
1519 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1522 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1524 rq
->avg_load_per_task
= 0;
1526 return rq
->avg_load_per_task
;
1529 #ifdef CONFIG_FAIR_GROUP_SCHED
1531 struct update_shares_data
{
1532 unsigned long rq_weight
[NR_CPUS
];
1535 static DEFINE_PER_CPU(struct update_shares_data
, update_shares_data
);
1537 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1540 * Calculate and set the cpu's group shares.
1542 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1543 unsigned long sd_shares
,
1544 unsigned long sd_rq_weight
,
1545 struct update_shares_data
*usd
)
1547 unsigned long shares
, rq_weight
;
1550 rq_weight
= usd
->rq_weight
[cpu
];
1553 rq_weight
= NICE_0_LOAD
;
1557 * \Sum_j shares_j * rq_weight_i
1558 * shares_i = -----------------------------
1559 * \Sum_j rq_weight_j
1561 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1562 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1564 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1565 sysctl_sched_shares_thresh
) {
1566 struct rq
*rq
= cpu_rq(cpu
);
1567 unsigned long flags
;
1569 spin_lock_irqsave(&rq
->lock
, flags
);
1570 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1571 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1572 __set_se_shares(tg
->se
[cpu
], shares
);
1573 spin_unlock_irqrestore(&rq
->lock
, flags
);
1578 * Re-compute the task group their per cpu shares over the given domain.
1579 * This needs to be done in a bottom-up fashion because the rq weight of a
1580 * parent group depends on the shares of its child groups.
1582 static int tg_shares_up(struct task_group
*tg
, void *data
)
1584 unsigned long weight
, rq_weight
= 0, shares
= 0;
1585 struct update_shares_data
*usd
;
1586 struct sched_domain
*sd
= data
;
1587 unsigned long flags
;
1593 local_irq_save(flags
);
1594 usd
= &__get_cpu_var(update_shares_data
);
1596 for_each_cpu(i
, sched_domain_span(sd
)) {
1597 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1598 usd
->rq_weight
[i
] = weight
;
1601 * If there are currently no tasks on the cpu pretend there
1602 * is one of average load so that when a new task gets to
1603 * run here it will not get delayed by group starvation.
1606 weight
= NICE_0_LOAD
;
1608 rq_weight
+= weight
;
1609 shares
+= tg
->cfs_rq
[i
]->shares
;
1612 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1613 shares
= tg
->shares
;
1615 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1616 shares
= tg
->shares
;
1618 for_each_cpu(i
, sched_domain_span(sd
))
1619 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd
);
1621 local_irq_restore(flags
);
1627 * Compute the cpu's hierarchical load factor for each task group.
1628 * This needs to be done in a top-down fashion because the load of a child
1629 * group is a fraction of its parents load.
1631 static int tg_load_down(struct task_group
*tg
, void *data
)
1634 long cpu
= (long)data
;
1637 load
= cpu_rq(cpu
)->load
.weight
;
1639 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1640 load
*= tg
->cfs_rq
[cpu
]->shares
;
1641 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1644 tg
->cfs_rq
[cpu
]->h_load
= load
;
1649 static void update_shares(struct sched_domain
*sd
)
1654 if (root_task_group_empty())
1657 now
= cpu_clock(raw_smp_processor_id());
1658 elapsed
= now
- sd
->last_update
;
1660 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1661 sd
->last_update
= now
;
1662 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1666 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1668 if (root_task_group_empty())
1671 spin_unlock(&rq
->lock
);
1673 spin_lock(&rq
->lock
);
1676 static void update_h_load(long cpu
)
1678 if (root_task_group_empty())
1681 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1686 static inline void update_shares(struct sched_domain
*sd
)
1690 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1696 #ifdef CONFIG_PREEMPT
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1707 __releases(this_rq
->lock
)
1708 __acquires(busiest
->lock
)
1709 __acquires(this_rq
->lock
)
1711 spin_unlock(&this_rq
->lock
);
1712 double_rq_lock(this_rq
, busiest
);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1726 __releases(this_rq
->lock
)
1727 __acquires(busiest
->lock
)
1728 __acquires(this_rq
->lock
)
1732 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1733 if (busiest
< this_rq
) {
1734 spin_unlock(&this_rq
->lock
);
1735 spin_lock(&busiest
->lock
);
1736 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1739 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1744 #endif /* CONFIG_PREEMPT */
1747 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1749 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1751 if (unlikely(!irqs_disabled())) {
1752 /* printk() doesn't work good under rq->lock */
1753 spin_unlock(&this_rq
->lock
);
1757 return _double_lock_balance(this_rq
, busiest
);
1760 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1761 __releases(busiest
->lock
)
1763 spin_unlock(&busiest
->lock
);
1764 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1768 #ifdef CONFIG_FAIR_GROUP_SCHED
1769 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1772 cfs_rq
->shares
= shares
;
1777 static void calc_load_account_active(struct rq
*this_rq
);
1779 #include "sched_stats.h"
1780 #include "sched_idletask.c"
1781 #include "sched_fair.c"
1782 #include "sched_rt.c"
1783 #ifdef CONFIG_SCHED_DEBUG
1784 # include "sched_debug.c"
1787 #define sched_class_highest (&rt_sched_class)
1788 #define for_each_class(class) \
1789 for (class = sched_class_highest; class; class = class->next)
1791 static void inc_nr_running(struct rq
*rq
)
1796 static void dec_nr_running(struct rq
*rq
)
1801 static void set_load_weight(struct task_struct
*p
)
1803 if (task_has_rt_policy(p
)) {
1804 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1805 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1810 * SCHED_IDLE tasks get minimal weight:
1812 if (p
->policy
== SCHED_IDLE
) {
1813 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1814 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1818 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1819 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1822 static void update_avg(u64
*avg
, u64 sample
)
1824 s64 diff
= sample
- *avg
;
1828 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1831 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1833 sched_info_queued(p
);
1834 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1838 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1841 if (p
->se
.last_wakeup
) {
1842 update_avg(&p
->se
.avg_overlap
,
1843 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1844 p
->se
.last_wakeup
= 0;
1846 update_avg(&p
->se
.avg_wakeup
,
1847 sysctl_sched_wakeup_granularity
);
1851 sched_info_dequeued(p
);
1852 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1857 * __normal_prio - return the priority that is based on the static prio
1859 static inline int __normal_prio(struct task_struct
*p
)
1861 return p
->static_prio
;
1865 * Calculate the expected normal priority: i.e. priority
1866 * without taking RT-inheritance into account. Might be
1867 * boosted by interactivity modifiers. Changes upon fork,
1868 * setprio syscalls, and whenever the interactivity
1869 * estimator recalculates.
1871 static inline int normal_prio(struct task_struct
*p
)
1875 if (task_has_rt_policy(p
))
1876 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1878 prio
= __normal_prio(p
);
1883 * Calculate the current priority, i.e. the priority
1884 * taken into account by the scheduler. This value might
1885 * be boosted by RT tasks, or might be boosted by
1886 * interactivity modifiers. Will be RT if the task got
1887 * RT-boosted. If not then it returns p->normal_prio.
1889 static int effective_prio(struct task_struct
*p
)
1891 p
->normal_prio
= normal_prio(p
);
1893 * If we are RT tasks or we were boosted to RT priority,
1894 * keep the priority unchanged. Otherwise, update priority
1895 * to the normal priority:
1897 if (!rt_prio(p
->prio
))
1898 return p
->normal_prio
;
1903 * activate_task - move a task to the runqueue.
1905 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1907 if (task_contributes_to_load(p
))
1908 rq
->nr_uninterruptible
--;
1910 enqueue_task(rq
, p
, wakeup
);
1915 * deactivate_task - remove a task from the runqueue.
1917 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1919 if (task_contributes_to_load(p
))
1920 rq
->nr_uninterruptible
++;
1922 dequeue_task(rq
, p
, sleep
);
1927 * task_curr - is this task currently executing on a CPU?
1928 * @p: the task in question.
1930 inline int task_curr(const struct task_struct
*p
)
1932 return cpu_curr(task_cpu(p
)) == p
;
1935 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1937 set_task_rq(p
, cpu
);
1940 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1941 * successfuly executed on another CPU. We must ensure that updates of
1942 * per-task data have been completed by this moment.
1945 task_thread_info(p
)->cpu
= cpu
;
1949 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1950 const struct sched_class
*prev_class
,
1951 int oldprio
, int running
)
1953 if (prev_class
!= p
->sched_class
) {
1954 if (prev_class
->switched_from
)
1955 prev_class
->switched_from(rq
, p
, running
);
1956 p
->sched_class
->switched_to(rq
, p
, running
);
1958 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1963 /* Used instead of source_load when we know the type == 0 */
1964 static unsigned long weighted_cpuload(const int cpu
)
1966 return cpu_rq(cpu
)->load
.weight
;
1970 * Is this task likely cache-hot:
1973 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1978 * Buddy candidates are cache hot:
1980 if (sched_feat(CACHE_HOT_BUDDY
) &&
1981 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1982 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1985 if (p
->sched_class
!= &fair_sched_class
)
1988 if (sysctl_sched_migration_cost
== -1)
1990 if (sysctl_sched_migration_cost
== 0)
1993 delta
= now
- p
->se
.exec_start
;
1995 return delta
< (s64
)sysctl_sched_migration_cost
;
1999 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2001 int old_cpu
= task_cpu(p
);
2002 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2003 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2004 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2007 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2009 trace_sched_migrate_task(p
, new_cpu
);
2011 #ifdef CONFIG_SCHEDSTATS
2012 if (p
->se
.wait_start
)
2013 p
->se
.wait_start
-= clock_offset
;
2014 if (p
->se
.sleep_start
)
2015 p
->se
.sleep_start
-= clock_offset
;
2016 if (p
->se
.block_start
)
2017 p
->se
.block_start
-= clock_offset
;
2019 if (old_cpu
!= new_cpu
) {
2020 p
->se
.nr_migrations
++;
2021 new_rq
->nr_migrations_in
++;
2022 #ifdef CONFIG_SCHEDSTATS
2023 if (task_hot(p
, old_rq
->clock
, NULL
))
2024 schedstat_inc(p
, se
.nr_forced2_migrations
);
2026 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2029 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2030 new_cfsrq
->min_vruntime
;
2032 __set_task_cpu(p
, new_cpu
);
2035 struct migration_req
{
2036 struct list_head list
;
2038 struct task_struct
*task
;
2041 struct completion done
;
2045 * The task's runqueue lock must be held.
2046 * Returns true if you have to wait for migration thread.
2049 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2051 struct rq
*rq
= task_rq(p
);
2054 * If the task is not on a runqueue (and not running), then
2055 * it is sufficient to simply update the task's cpu field.
2057 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2058 set_task_cpu(p
, dest_cpu
);
2062 init_completion(&req
->done
);
2064 req
->dest_cpu
= dest_cpu
;
2065 list_add(&req
->list
, &rq
->migration_queue
);
2071 * wait_task_context_switch - wait for a thread to complete at least one
2074 * @p must not be current.
2076 void wait_task_context_switch(struct task_struct
*p
)
2078 unsigned long nvcsw
, nivcsw
, flags
;
2086 * The runqueue is assigned before the actual context
2087 * switch. We need to take the runqueue lock.
2089 * We could check initially without the lock but it is
2090 * very likely that we need to take the lock in every
2093 rq
= task_rq_lock(p
, &flags
);
2094 running
= task_running(rq
, p
);
2095 task_rq_unlock(rq
, &flags
);
2097 if (likely(!running
))
2100 * The switch count is incremented before the actual
2101 * context switch. We thus wait for two switches to be
2102 * sure at least one completed.
2104 if ((p
->nvcsw
- nvcsw
) > 1)
2106 if ((p
->nivcsw
- nivcsw
) > 1)
2114 * wait_task_inactive - wait for a thread to unschedule.
2116 * If @match_state is nonzero, it's the @p->state value just checked and
2117 * not expected to change. If it changes, i.e. @p might have woken up,
2118 * then return zero. When we succeed in waiting for @p to be off its CPU,
2119 * we return a positive number (its total switch count). If a second call
2120 * a short while later returns the same number, the caller can be sure that
2121 * @p has remained unscheduled the whole time.
2123 * The caller must ensure that the task *will* unschedule sometime soon,
2124 * else this function might spin for a *long* time. This function can't
2125 * be called with interrupts off, or it may introduce deadlock with
2126 * smp_call_function() if an IPI is sent by the same process we are
2127 * waiting to become inactive.
2129 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2131 unsigned long flags
;
2138 * We do the initial early heuristics without holding
2139 * any task-queue locks at all. We'll only try to get
2140 * the runqueue lock when things look like they will
2146 * If the task is actively running on another CPU
2147 * still, just relax and busy-wait without holding
2150 * NOTE! Since we don't hold any locks, it's not
2151 * even sure that "rq" stays as the right runqueue!
2152 * But we don't care, since "task_running()" will
2153 * return false if the runqueue has changed and p
2154 * is actually now running somewhere else!
2156 while (task_running(rq
, p
)) {
2157 if (match_state
&& unlikely(p
->state
!= match_state
))
2163 * Ok, time to look more closely! We need the rq
2164 * lock now, to be *sure*. If we're wrong, we'll
2165 * just go back and repeat.
2167 rq
= task_rq_lock(p
, &flags
);
2168 trace_sched_wait_task(rq
, p
);
2169 running
= task_running(rq
, p
);
2170 on_rq
= p
->se
.on_rq
;
2172 if (!match_state
|| p
->state
== match_state
)
2173 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2174 task_rq_unlock(rq
, &flags
);
2177 * If it changed from the expected state, bail out now.
2179 if (unlikely(!ncsw
))
2183 * Was it really running after all now that we
2184 * checked with the proper locks actually held?
2186 * Oops. Go back and try again..
2188 if (unlikely(running
)) {
2194 * It's not enough that it's not actively running,
2195 * it must be off the runqueue _entirely_, and not
2198 * So if it was still runnable (but just not actively
2199 * running right now), it's preempted, and we should
2200 * yield - it could be a while.
2202 if (unlikely(on_rq
)) {
2203 schedule_timeout_uninterruptible(1);
2208 * Ahh, all good. It wasn't running, and it wasn't
2209 * runnable, which means that it will never become
2210 * running in the future either. We're all done!
2219 * kick_process - kick a running thread to enter/exit the kernel
2220 * @p: the to-be-kicked thread
2222 * Cause a process which is running on another CPU to enter
2223 * kernel-mode, without any delay. (to get signals handled.)
2225 * NOTE: this function doesnt have to take the runqueue lock,
2226 * because all it wants to ensure is that the remote task enters
2227 * the kernel. If the IPI races and the task has been migrated
2228 * to another CPU then no harm is done and the purpose has been
2231 void kick_process(struct task_struct
*p
)
2237 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2238 smp_send_reschedule(cpu
);
2241 EXPORT_SYMBOL_GPL(kick_process
);
2244 * Return a low guess at the load of a migration-source cpu weighted
2245 * according to the scheduling class and "nice" value.
2247 * We want to under-estimate the load of migration sources, to
2248 * balance conservatively.
2250 static unsigned long source_load(int cpu
, int type
)
2252 struct rq
*rq
= cpu_rq(cpu
);
2253 unsigned long total
= weighted_cpuload(cpu
);
2255 if (type
== 0 || !sched_feat(LB_BIAS
))
2258 return min(rq
->cpu_load
[type
-1], total
);
2262 * Return a high guess at the load of a migration-target cpu weighted
2263 * according to the scheduling class and "nice" value.
2265 static unsigned long target_load(int cpu
, int type
)
2267 struct rq
*rq
= cpu_rq(cpu
);
2268 unsigned long total
= weighted_cpuload(cpu
);
2270 if (type
== 0 || !sched_feat(LB_BIAS
))
2273 return max(rq
->cpu_load
[type
-1], total
);
2277 * find_idlest_group finds and returns the least busy CPU group within the
2280 static struct sched_group
*
2281 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2283 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2284 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2285 int load_idx
= sd
->forkexec_idx
;
2286 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2289 unsigned long load
, avg_load
;
2293 /* Skip over this group if it has no CPUs allowed */
2294 if (!cpumask_intersects(sched_group_cpus(group
),
2298 local_group
= cpumask_test_cpu(this_cpu
,
2299 sched_group_cpus(group
));
2301 /* Tally up the load of all CPUs in the group */
2304 for_each_cpu(i
, sched_group_cpus(group
)) {
2305 /* Bias balancing toward cpus of our domain */
2307 load
= source_load(i
, load_idx
);
2309 load
= target_load(i
, load_idx
);
2314 /* Adjust by relative CPU power of the group */
2315 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2318 this_load
= avg_load
;
2320 } else if (avg_load
< min_load
) {
2321 min_load
= avg_load
;
2324 } while (group
= group
->next
, group
!= sd
->groups
);
2326 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2332 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2335 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2337 unsigned long load
, min_load
= ULONG_MAX
;
2341 /* Traverse only the allowed CPUs */
2342 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2343 load
= weighted_cpuload(i
);
2345 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2355 * sched_balance_self: balance the current task (running on cpu) in domains
2356 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2359 * Balance, ie. select the least loaded group.
2361 * Returns the target CPU number, or the same CPU if no balancing is needed.
2363 * preempt must be disabled.
2365 static int sched_balance_self(int cpu
, int flag
)
2367 struct task_struct
*t
= current
;
2368 struct sched_domain
*tmp
, *sd
= NULL
;
2370 for_each_domain(cpu
, tmp
) {
2372 * If power savings logic is enabled for a domain, stop there.
2374 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2376 if (tmp
->flags
& flag
)
2384 struct sched_group
*group
;
2385 int new_cpu
, weight
;
2387 if (!(sd
->flags
& flag
)) {
2392 group
= find_idlest_group(sd
, t
, cpu
);
2398 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2399 if (new_cpu
== -1 || new_cpu
== cpu
) {
2400 /* Now try balancing at a lower domain level of cpu */
2405 /* Now try balancing at a lower domain level of new_cpu */
2407 weight
= cpumask_weight(sched_domain_span(sd
));
2409 for_each_domain(cpu
, tmp
) {
2410 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2412 if (tmp
->flags
& flag
)
2415 /* while loop will break here if sd == NULL */
2421 #endif /* CONFIG_SMP */
2424 * task_oncpu_function_call - call a function on the cpu on which a task runs
2425 * @p: the task to evaluate
2426 * @func: the function to be called
2427 * @info: the function call argument
2429 * Calls the function @func when the task is currently running. This might
2430 * be on the current CPU, which just calls the function directly
2432 void task_oncpu_function_call(struct task_struct
*p
,
2433 void (*func
) (void *info
), void *info
)
2440 smp_call_function_single(cpu
, func
, info
, 1);
2445 * try_to_wake_up - wake up a thread
2446 * @p: the to-be-woken-up thread
2447 * @state: the mask of task states that can be woken
2448 * @sync: do a synchronous wakeup?
2450 * Put it on the run-queue if it's not already there. The "current"
2451 * thread is always on the run-queue (except when the actual
2452 * re-schedule is in progress), and as such you're allowed to do
2453 * the simpler "current->state = TASK_RUNNING" to mark yourself
2454 * runnable without the overhead of this.
2456 * returns failure only if the task is already active.
2458 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2460 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2461 unsigned long flags
;
2465 if (!sched_feat(SYNC_WAKEUPS
))
2469 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2470 struct sched_domain
*sd
;
2472 this_cpu
= raw_smp_processor_id();
2475 for_each_domain(this_cpu
, sd
) {
2476 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2485 rq
= task_rq_lock(p
, &flags
);
2486 update_rq_clock(rq
);
2487 old_state
= p
->state
;
2488 if (!(old_state
& state
))
2496 this_cpu
= smp_processor_id();
2499 if (unlikely(task_running(rq
, p
)))
2502 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2503 if (cpu
!= orig_cpu
) {
2504 set_task_cpu(p
, cpu
);
2505 task_rq_unlock(rq
, &flags
);
2506 /* might preempt at this point */
2507 rq
= task_rq_lock(p
, &flags
);
2508 old_state
= p
->state
;
2509 if (!(old_state
& state
))
2514 this_cpu
= smp_processor_id();
2518 #ifdef CONFIG_SCHEDSTATS
2519 schedstat_inc(rq
, ttwu_count
);
2520 if (cpu
== this_cpu
)
2521 schedstat_inc(rq
, ttwu_local
);
2523 struct sched_domain
*sd
;
2524 for_each_domain(this_cpu
, sd
) {
2525 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2526 schedstat_inc(sd
, ttwu_wake_remote
);
2531 #endif /* CONFIG_SCHEDSTATS */
2534 #endif /* CONFIG_SMP */
2535 schedstat_inc(p
, se
.nr_wakeups
);
2537 schedstat_inc(p
, se
.nr_wakeups_sync
);
2538 if (orig_cpu
!= cpu
)
2539 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2540 if (cpu
== this_cpu
)
2541 schedstat_inc(p
, se
.nr_wakeups_local
);
2543 schedstat_inc(p
, se
.nr_wakeups_remote
);
2544 activate_task(rq
, p
, 1);
2548 * Only attribute actual wakeups done by this task.
2550 if (!in_interrupt()) {
2551 struct sched_entity
*se
= ¤t
->se
;
2552 u64 sample
= se
->sum_exec_runtime
;
2554 if (se
->last_wakeup
)
2555 sample
-= se
->last_wakeup
;
2557 sample
-= se
->start_runtime
;
2558 update_avg(&se
->avg_wakeup
, sample
);
2560 se
->last_wakeup
= se
->sum_exec_runtime
;
2564 trace_sched_wakeup(rq
, p
, success
);
2565 check_preempt_curr(rq
, p
, sync
);
2567 p
->state
= TASK_RUNNING
;
2569 if (p
->sched_class
->task_wake_up
)
2570 p
->sched_class
->task_wake_up(rq
, p
);
2573 task_rq_unlock(rq
, &flags
);
2579 * wake_up_process - Wake up a specific process
2580 * @p: The process to be woken up.
2582 * Attempt to wake up the nominated process and move it to the set of runnable
2583 * processes. Returns 1 if the process was woken up, 0 if it was already
2586 * It may be assumed that this function implies a write memory barrier before
2587 * changing the task state if and only if any tasks are woken up.
2589 int wake_up_process(struct task_struct
*p
)
2591 return try_to_wake_up(p
, TASK_ALL
, 0);
2593 EXPORT_SYMBOL(wake_up_process
);
2595 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2597 return try_to_wake_up(p
, state
, 0);
2601 * Perform scheduler related setup for a newly forked process p.
2602 * p is forked by current.
2604 * __sched_fork() is basic setup used by init_idle() too:
2606 static void __sched_fork(struct task_struct
*p
)
2608 p
->se
.exec_start
= 0;
2609 p
->se
.sum_exec_runtime
= 0;
2610 p
->se
.prev_sum_exec_runtime
= 0;
2611 p
->se
.nr_migrations
= 0;
2612 p
->se
.last_wakeup
= 0;
2613 p
->se
.avg_overlap
= 0;
2614 p
->se
.start_runtime
= 0;
2615 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2617 #ifdef CONFIG_SCHEDSTATS
2618 p
->se
.wait_start
= 0;
2620 p
->se
.wait_count
= 0;
2623 p
->se
.sleep_start
= 0;
2624 p
->se
.sleep_max
= 0;
2625 p
->se
.sum_sleep_runtime
= 0;
2627 p
->se
.block_start
= 0;
2628 p
->se
.block_max
= 0;
2630 p
->se
.slice_max
= 0;
2632 p
->se
.nr_migrations_cold
= 0;
2633 p
->se
.nr_failed_migrations_affine
= 0;
2634 p
->se
.nr_failed_migrations_running
= 0;
2635 p
->se
.nr_failed_migrations_hot
= 0;
2636 p
->se
.nr_forced_migrations
= 0;
2637 p
->se
.nr_forced2_migrations
= 0;
2639 p
->se
.nr_wakeups
= 0;
2640 p
->se
.nr_wakeups_sync
= 0;
2641 p
->se
.nr_wakeups_migrate
= 0;
2642 p
->se
.nr_wakeups_local
= 0;
2643 p
->se
.nr_wakeups_remote
= 0;
2644 p
->se
.nr_wakeups_affine
= 0;
2645 p
->se
.nr_wakeups_affine_attempts
= 0;
2646 p
->se
.nr_wakeups_passive
= 0;
2647 p
->se
.nr_wakeups_idle
= 0;
2651 INIT_LIST_HEAD(&p
->rt
.run_list
);
2653 INIT_LIST_HEAD(&p
->se
.group_node
);
2655 #ifdef CONFIG_PREEMPT_NOTIFIERS
2656 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2660 * We mark the process as running here, but have not actually
2661 * inserted it onto the runqueue yet. This guarantees that
2662 * nobody will actually run it, and a signal or other external
2663 * event cannot wake it up and insert it on the runqueue either.
2665 p
->state
= TASK_RUNNING
;
2669 * fork()/clone()-time setup:
2671 void sched_fork(struct task_struct
*p
, int clone_flags
)
2673 int cpu
= get_cpu();
2678 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2680 set_task_cpu(p
, cpu
);
2683 * Make sure we do not leak PI boosting priority to the child.
2685 p
->prio
= current
->normal_prio
;
2688 * Revert to default priority/policy on fork if requested.
2690 if (unlikely(p
->sched_reset_on_fork
)) {
2691 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
)
2692 p
->policy
= SCHED_NORMAL
;
2694 if (p
->normal_prio
< DEFAULT_PRIO
)
2695 p
->prio
= DEFAULT_PRIO
;
2697 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2698 p
->static_prio
= NICE_TO_PRIO(0);
2703 * We don't need the reset flag anymore after the fork. It has
2704 * fulfilled its duty:
2706 p
->sched_reset_on_fork
= 0;
2709 if (!rt_prio(p
->prio
))
2710 p
->sched_class
= &fair_sched_class
;
2712 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2713 if (likely(sched_info_on()))
2714 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2716 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2719 #ifdef CONFIG_PREEMPT
2720 /* Want to start with kernel preemption disabled. */
2721 task_thread_info(p
)->preempt_count
= 1;
2723 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2729 * wake_up_new_task - wake up a newly created task for the first time.
2731 * This function will do some initial scheduler statistics housekeeping
2732 * that must be done for every newly created context, then puts the task
2733 * on the runqueue and wakes it.
2735 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2737 unsigned long flags
;
2740 rq
= task_rq_lock(p
, &flags
);
2741 BUG_ON(p
->state
!= TASK_RUNNING
);
2742 update_rq_clock(rq
);
2744 p
->prio
= effective_prio(p
);
2746 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2747 activate_task(rq
, p
, 0);
2750 * Let the scheduling class do new task startup
2751 * management (if any):
2753 p
->sched_class
->task_new(rq
, p
);
2756 trace_sched_wakeup_new(rq
, p
, 1);
2757 check_preempt_curr(rq
, p
, 0);
2759 if (p
->sched_class
->task_wake_up
)
2760 p
->sched_class
->task_wake_up(rq
, p
);
2762 task_rq_unlock(rq
, &flags
);
2765 #ifdef CONFIG_PREEMPT_NOTIFIERS
2768 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2769 * @notifier: notifier struct to register
2771 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2773 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2775 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2778 * preempt_notifier_unregister - no longer interested in preemption notifications
2779 * @notifier: notifier struct to unregister
2781 * This is safe to call from within a preemption notifier.
2783 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2785 hlist_del(¬ifier
->link
);
2787 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2789 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2791 struct preempt_notifier
*notifier
;
2792 struct hlist_node
*node
;
2794 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2795 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2799 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2800 struct task_struct
*next
)
2802 struct preempt_notifier
*notifier
;
2803 struct hlist_node
*node
;
2805 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2806 notifier
->ops
->sched_out(notifier
, next
);
2809 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2811 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2816 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2817 struct task_struct
*next
)
2821 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2824 * prepare_task_switch - prepare to switch tasks
2825 * @rq: the runqueue preparing to switch
2826 * @prev: the current task that is being switched out
2827 * @next: the task we are going to switch to.
2829 * This is called with the rq lock held and interrupts off. It must
2830 * be paired with a subsequent finish_task_switch after the context
2833 * prepare_task_switch sets up locking and calls architecture specific
2837 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2838 struct task_struct
*next
)
2840 fire_sched_out_preempt_notifiers(prev
, next
);
2841 prepare_lock_switch(rq
, next
);
2842 prepare_arch_switch(next
);
2846 * finish_task_switch - clean up after a task-switch
2847 * @rq: runqueue associated with task-switch
2848 * @prev: the thread we just switched away from.
2850 * finish_task_switch must be called after the context switch, paired
2851 * with a prepare_task_switch call before the context switch.
2852 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2853 * and do any other architecture-specific cleanup actions.
2855 * Note that we may have delayed dropping an mm in context_switch(). If
2856 * so, we finish that here outside of the runqueue lock. (Doing it
2857 * with the lock held can cause deadlocks; see schedule() for
2860 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2861 __releases(rq
->lock
)
2863 struct mm_struct
*mm
= rq
->prev_mm
;
2869 * A task struct has one reference for the use as "current".
2870 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2871 * schedule one last time. The schedule call will never return, and
2872 * the scheduled task must drop that reference.
2873 * The test for TASK_DEAD must occur while the runqueue locks are
2874 * still held, otherwise prev could be scheduled on another cpu, die
2875 * there before we look at prev->state, and then the reference would
2877 * Manfred Spraul <manfred@colorfullife.com>
2879 prev_state
= prev
->state
;
2880 finish_arch_switch(prev
);
2881 perf_counter_task_sched_in(current
, cpu_of(rq
));
2882 finish_lock_switch(rq
, prev
);
2884 fire_sched_in_preempt_notifiers(current
);
2887 if (unlikely(prev_state
== TASK_DEAD
)) {
2889 * Remove function-return probe instances associated with this
2890 * task and put them back on the free list.
2892 kprobe_flush_task(prev
);
2893 put_task_struct(prev
);
2899 /* assumes rq->lock is held */
2900 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2902 if (prev
->sched_class
->pre_schedule
)
2903 prev
->sched_class
->pre_schedule(rq
, prev
);
2906 /* rq->lock is NOT held, but preemption is disabled */
2907 static inline void post_schedule(struct rq
*rq
)
2909 if (rq
->post_schedule
) {
2910 unsigned long flags
;
2912 spin_lock_irqsave(&rq
->lock
, flags
);
2913 if (rq
->curr
->sched_class
->post_schedule
)
2914 rq
->curr
->sched_class
->post_schedule(rq
);
2915 spin_unlock_irqrestore(&rq
->lock
, flags
);
2917 rq
->post_schedule
= 0;
2923 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2927 static inline void post_schedule(struct rq
*rq
)
2934 * schedule_tail - first thing a freshly forked thread must call.
2935 * @prev: the thread we just switched away from.
2937 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2938 __releases(rq
->lock
)
2940 struct rq
*rq
= this_rq();
2942 finish_task_switch(rq
, prev
);
2945 * FIXME: do we need to worry about rq being invalidated by the
2950 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2951 /* In this case, finish_task_switch does not reenable preemption */
2954 if (current
->set_child_tid
)
2955 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2959 * context_switch - switch to the new MM and the new
2960 * thread's register state.
2963 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2964 struct task_struct
*next
)
2966 struct mm_struct
*mm
, *oldmm
;
2968 prepare_task_switch(rq
, prev
, next
);
2969 trace_sched_switch(rq
, prev
, next
);
2971 oldmm
= prev
->active_mm
;
2973 * For paravirt, this is coupled with an exit in switch_to to
2974 * combine the page table reload and the switch backend into
2977 arch_start_context_switch(prev
);
2979 if (unlikely(!mm
)) {
2980 next
->active_mm
= oldmm
;
2981 atomic_inc(&oldmm
->mm_count
);
2982 enter_lazy_tlb(oldmm
, next
);
2984 switch_mm(oldmm
, mm
, next
);
2986 if (unlikely(!prev
->mm
)) {
2987 prev
->active_mm
= NULL
;
2988 rq
->prev_mm
= oldmm
;
2991 * Since the runqueue lock will be released by the next
2992 * task (which is an invalid locking op but in the case
2993 * of the scheduler it's an obvious special-case), so we
2994 * do an early lockdep release here:
2996 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2997 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3000 /* Here we just switch the register state and the stack. */
3001 switch_to(prev
, next
, prev
);
3005 * this_rq must be evaluated again because prev may have moved
3006 * CPUs since it called schedule(), thus the 'rq' on its stack
3007 * frame will be invalid.
3009 finish_task_switch(this_rq(), prev
);
3013 * nr_running, nr_uninterruptible and nr_context_switches:
3015 * externally visible scheduler statistics: current number of runnable
3016 * threads, current number of uninterruptible-sleeping threads, total
3017 * number of context switches performed since bootup.
3019 unsigned long nr_running(void)
3021 unsigned long i
, sum
= 0;
3023 for_each_online_cpu(i
)
3024 sum
+= cpu_rq(i
)->nr_running
;
3029 unsigned long nr_uninterruptible(void)
3031 unsigned long i
, sum
= 0;
3033 for_each_possible_cpu(i
)
3034 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3037 * Since we read the counters lockless, it might be slightly
3038 * inaccurate. Do not allow it to go below zero though:
3040 if (unlikely((long)sum
< 0))
3046 unsigned long long nr_context_switches(void)
3049 unsigned long long sum
= 0;
3051 for_each_possible_cpu(i
)
3052 sum
+= cpu_rq(i
)->nr_switches
;
3057 unsigned long nr_iowait(void)
3059 unsigned long i
, sum
= 0;
3061 for_each_possible_cpu(i
)
3062 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3067 /* Variables and functions for calc_load */
3068 static atomic_long_t calc_load_tasks
;
3069 static unsigned long calc_load_update
;
3070 unsigned long avenrun
[3];
3071 EXPORT_SYMBOL(avenrun
);
3074 * get_avenrun - get the load average array
3075 * @loads: pointer to dest load array
3076 * @offset: offset to add
3077 * @shift: shift count to shift the result left
3079 * These values are estimates at best, so no need for locking.
3081 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3083 loads
[0] = (avenrun
[0] + offset
) << shift
;
3084 loads
[1] = (avenrun
[1] + offset
) << shift
;
3085 loads
[2] = (avenrun
[2] + offset
) << shift
;
3088 static unsigned long
3089 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3092 load
+= active
* (FIXED_1
- exp
);
3093 return load
>> FSHIFT
;
3097 * calc_load - update the avenrun load estimates 10 ticks after the
3098 * CPUs have updated calc_load_tasks.
3100 void calc_global_load(void)
3102 unsigned long upd
= calc_load_update
+ 10;
3105 if (time_before(jiffies
, upd
))
3108 active
= atomic_long_read(&calc_load_tasks
);
3109 active
= active
> 0 ? active
* FIXED_1
: 0;
3111 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3112 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3113 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3115 calc_load_update
+= LOAD_FREQ
;
3119 * Either called from update_cpu_load() or from a cpu going idle
3121 static void calc_load_account_active(struct rq
*this_rq
)
3123 long nr_active
, delta
;
3125 nr_active
= this_rq
->nr_running
;
3126 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3128 if (nr_active
!= this_rq
->calc_load_active
) {
3129 delta
= nr_active
- this_rq
->calc_load_active
;
3130 this_rq
->calc_load_active
= nr_active
;
3131 atomic_long_add(delta
, &calc_load_tasks
);
3136 * Externally visible per-cpu scheduler statistics:
3137 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3139 u64
cpu_nr_migrations(int cpu
)
3141 return cpu_rq(cpu
)->nr_migrations_in
;
3145 * Update rq->cpu_load[] statistics. This function is usually called every
3146 * scheduler tick (TICK_NSEC).
3148 static void update_cpu_load(struct rq
*this_rq
)
3150 unsigned long this_load
= this_rq
->load
.weight
;
3153 this_rq
->nr_load_updates
++;
3155 /* Update our load: */
3156 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3157 unsigned long old_load
, new_load
;
3159 /* scale is effectively 1 << i now, and >> i divides by scale */
3161 old_load
= this_rq
->cpu_load
[i
];
3162 new_load
= this_load
;
3164 * Round up the averaging division if load is increasing. This
3165 * prevents us from getting stuck on 9 if the load is 10, for
3168 if (new_load
> old_load
)
3169 new_load
+= scale
-1;
3170 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3173 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3174 this_rq
->calc_load_update
+= LOAD_FREQ
;
3175 calc_load_account_active(this_rq
);
3182 * double_rq_lock - safely lock two runqueues
3184 * Note this does not disable interrupts like task_rq_lock,
3185 * you need to do so manually before calling.
3187 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3188 __acquires(rq1
->lock
)
3189 __acquires(rq2
->lock
)
3191 BUG_ON(!irqs_disabled());
3193 spin_lock(&rq1
->lock
);
3194 __acquire(rq2
->lock
); /* Fake it out ;) */
3197 spin_lock(&rq1
->lock
);
3198 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3200 spin_lock(&rq2
->lock
);
3201 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3204 update_rq_clock(rq1
);
3205 update_rq_clock(rq2
);
3209 * double_rq_unlock - safely unlock two runqueues
3211 * Note this does not restore interrupts like task_rq_unlock,
3212 * you need to do so manually after calling.
3214 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3215 __releases(rq1
->lock
)
3216 __releases(rq2
->lock
)
3218 spin_unlock(&rq1
->lock
);
3220 spin_unlock(&rq2
->lock
);
3222 __release(rq2
->lock
);
3226 * If dest_cpu is allowed for this process, migrate the task to it.
3227 * This is accomplished by forcing the cpu_allowed mask to only
3228 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3229 * the cpu_allowed mask is restored.
3231 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3233 struct migration_req req
;
3234 unsigned long flags
;
3237 rq
= task_rq_lock(p
, &flags
);
3238 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3239 || unlikely(!cpu_active(dest_cpu
)))
3242 /* force the process onto the specified CPU */
3243 if (migrate_task(p
, dest_cpu
, &req
)) {
3244 /* Need to wait for migration thread (might exit: take ref). */
3245 struct task_struct
*mt
= rq
->migration_thread
;
3247 get_task_struct(mt
);
3248 task_rq_unlock(rq
, &flags
);
3249 wake_up_process(mt
);
3250 put_task_struct(mt
);
3251 wait_for_completion(&req
.done
);
3256 task_rq_unlock(rq
, &flags
);
3260 * sched_exec - execve() is a valuable balancing opportunity, because at
3261 * this point the task has the smallest effective memory and cache footprint.
3263 void sched_exec(void)
3265 int new_cpu
, this_cpu
= get_cpu();
3266 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3268 if (new_cpu
!= this_cpu
)
3269 sched_migrate_task(current
, new_cpu
);
3273 * pull_task - move a task from a remote runqueue to the local runqueue.
3274 * Both runqueues must be locked.
3276 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3277 struct rq
*this_rq
, int this_cpu
)
3279 deactivate_task(src_rq
, p
, 0);
3280 set_task_cpu(p
, this_cpu
);
3281 activate_task(this_rq
, p
, 0);
3283 * Note that idle threads have a prio of MAX_PRIO, for this test
3284 * to be always true for them.
3286 check_preempt_curr(this_rq
, p
, 0);
3290 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3293 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3294 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3297 int tsk_cache_hot
= 0;
3299 * We do not migrate tasks that are:
3300 * 1) running (obviously), or
3301 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3302 * 3) are cache-hot on their current CPU.
3304 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3305 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3310 if (task_running(rq
, p
)) {
3311 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3316 * Aggressive migration if:
3317 * 1) task is cache cold, or
3318 * 2) too many balance attempts have failed.
3321 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3322 if (!tsk_cache_hot
||
3323 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3324 #ifdef CONFIG_SCHEDSTATS
3325 if (tsk_cache_hot
) {
3326 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3327 schedstat_inc(p
, se
.nr_forced_migrations
);
3333 if (tsk_cache_hot
) {
3334 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3340 static unsigned long
3341 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3342 unsigned long max_load_move
, struct sched_domain
*sd
,
3343 enum cpu_idle_type idle
, int *all_pinned
,
3344 int *this_best_prio
, struct rq_iterator
*iterator
)
3346 int loops
= 0, pulled
= 0, pinned
= 0;
3347 struct task_struct
*p
;
3348 long rem_load_move
= max_load_move
;
3350 if (max_load_move
== 0)
3356 * Start the load-balancing iterator:
3358 p
= iterator
->start(iterator
->arg
);
3360 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3363 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3364 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3365 p
= iterator
->next(iterator
->arg
);
3369 pull_task(busiest
, p
, this_rq
, this_cpu
);
3371 rem_load_move
-= p
->se
.load
.weight
;
3373 #ifdef CONFIG_PREEMPT
3375 * NEWIDLE balancing is a source of latency, so preemptible kernels
3376 * will stop after the first task is pulled to minimize the critical
3379 if (idle
== CPU_NEWLY_IDLE
)
3384 * We only want to steal up to the prescribed amount of weighted load.
3386 if (rem_load_move
> 0) {
3387 if (p
->prio
< *this_best_prio
)
3388 *this_best_prio
= p
->prio
;
3389 p
= iterator
->next(iterator
->arg
);
3394 * Right now, this is one of only two places pull_task() is called,
3395 * so we can safely collect pull_task() stats here rather than
3396 * inside pull_task().
3398 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3401 *all_pinned
= pinned
;
3403 return max_load_move
- rem_load_move
;
3407 * move_tasks tries to move up to max_load_move weighted load from busiest to
3408 * this_rq, as part of a balancing operation within domain "sd".
3409 * Returns 1 if successful and 0 otherwise.
3411 * Called with both runqueues locked.
3413 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3414 unsigned long max_load_move
,
3415 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3418 const struct sched_class
*class = sched_class_highest
;
3419 unsigned long total_load_moved
= 0;
3420 int this_best_prio
= this_rq
->curr
->prio
;
3424 class->load_balance(this_rq
, this_cpu
, busiest
,
3425 max_load_move
- total_load_moved
,
3426 sd
, idle
, all_pinned
, &this_best_prio
);
3427 class = class->next
;
3429 #ifdef CONFIG_PREEMPT
3431 * NEWIDLE balancing is a source of latency, so preemptible
3432 * kernels will stop after the first task is pulled to minimize
3433 * the critical section.
3435 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3438 } while (class && max_load_move
> total_load_moved
);
3440 return total_load_moved
> 0;
3444 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3445 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3446 struct rq_iterator
*iterator
)
3448 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3452 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3453 pull_task(busiest
, p
, this_rq
, this_cpu
);
3455 * Right now, this is only the second place pull_task()
3456 * is called, so we can safely collect pull_task()
3457 * stats here rather than inside pull_task().
3459 schedstat_inc(sd
, lb_gained
[idle
]);
3463 p
= iterator
->next(iterator
->arg
);
3470 * move_one_task tries to move exactly one task from busiest to this_rq, as
3471 * part of active balancing operations within "domain".
3472 * Returns 1 if successful and 0 otherwise.
3474 * Called with both runqueues locked.
3476 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3477 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3479 const struct sched_class
*class;
3481 for_each_class(class) {
3482 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3488 /********** Helpers for find_busiest_group ************************/
3490 * sd_lb_stats - Structure to store the statistics of a sched_domain
3491 * during load balancing.
3493 struct sd_lb_stats
{
3494 struct sched_group
*busiest
; /* Busiest group in this sd */
3495 struct sched_group
*this; /* Local group in this sd */
3496 unsigned long total_load
; /* Total load of all groups in sd */
3497 unsigned long total_pwr
; /* Total power of all groups in sd */
3498 unsigned long avg_load
; /* Average load across all groups in sd */
3500 /** Statistics of this group */
3501 unsigned long this_load
;
3502 unsigned long this_load_per_task
;
3503 unsigned long this_nr_running
;
3505 /* Statistics of the busiest group */
3506 unsigned long max_load
;
3507 unsigned long busiest_load_per_task
;
3508 unsigned long busiest_nr_running
;
3510 int group_imb
; /* Is there imbalance in this sd */
3511 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3512 int power_savings_balance
; /* Is powersave balance needed for this sd */
3513 struct sched_group
*group_min
; /* Least loaded group in sd */
3514 struct sched_group
*group_leader
; /* Group which relieves group_min */
3515 unsigned long min_load_per_task
; /* load_per_task in group_min */
3516 unsigned long leader_nr_running
; /* Nr running of group_leader */
3517 unsigned long min_nr_running
; /* Nr running of group_min */
3522 * sg_lb_stats - stats of a sched_group required for load_balancing
3524 struct sg_lb_stats
{
3525 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3526 unsigned long group_load
; /* Total load over the CPUs of the group */
3527 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3528 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3529 unsigned long group_capacity
;
3530 int group_imb
; /* Is there an imbalance in the group ? */
3534 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3535 * @group: The group whose first cpu is to be returned.
3537 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3539 return cpumask_first(sched_group_cpus(group
));
3543 * get_sd_load_idx - Obtain the load index for a given sched domain.
3544 * @sd: The sched_domain whose load_idx is to be obtained.
3545 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3547 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3548 enum cpu_idle_type idle
)
3554 load_idx
= sd
->busy_idx
;
3557 case CPU_NEWLY_IDLE
:
3558 load_idx
= sd
->newidle_idx
;
3561 load_idx
= sd
->idle_idx
;
3569 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3571 * init_sd_power_savings_stats - Initialize power savings statistics for
3572 * the given sched_domain, during load balancing.
3574 * @sd: Sched domain whose power-savings statistics are to be initialized.
3575 * @sds: Variable containing the statistics for sd.
3576 * @idle: Idle status of the CPU at which we're performing load-balancing.
3578 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3579 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3582 * Busy processors will not participate in power savings
3585 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3586 sds
->power_savings_balance
= 0;
3588 sds
->power_savings_balance
= 1;
3589 sds
->min_nr_running
= ULONG_MAX
;
3590 sds
->leader_nr_running
= 0;
3595 * update_sd_power_savings_stats - Update the power saving stats for a
3596 * sched_domain while performing load balancing.
3598 * @group: sched_group belonging to the sched_domain under consideration.
3599 * @sds: Variable containing the statistics of the sched_domain
3600 * @local_group: Does group contain the CPU for which we're performing
3602 * @sgs: Variable containing the statistics of the group.
3604 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3605 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3608 if (!sds
->power_savings_balance
)
3612 * If the local group is idle or completely loaded
3613 * no need to do power savings balance at this domain
3615 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3616 !sds
->this_nr_running
))
3617 sds
->power_savings_balance
= 0;
3620 * If a group is already running at full capacity or idle,
3621 * don't include that group in power savings calculations
3623 if (!sds
->power_savings_balance
||
3624 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3625 !sgs
->sum_nr_running
)
3629 * Calculate the group which has the least non-idle load.
3630 * This is the group from where we need to pick up the load
3633 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3634 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3635 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3636 sds
->group_min
= group
;
3637 sds
->min_nr_running
= sgs
->sum_nr_running
;
3638 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3639 sgs
->sum_nr_running
;
3643 * Calculate the group which is almost near its
3644 * capacity but still has some space to pick up some load
3645 * from other group and save more power
3647 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3650 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3651 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3652 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3653 sds
->group_leader
= group
;
3654 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3659 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3660 * @sds: Variable containing the statistics of the sched_domain
3661 * under consideration.
3662 * @this_cpu: Cpu at which we're currently performing load-balancing.
3663 * @imbalance: Variable to store the imbalance.
3666 * Check if we have potential to perform some power-savings balance.
3667 * If yes, set the busiest group to be the least loaded group in the
3668 * sched_domain, so that it's CPUs can be put to idle.
3670 * Returns 1 if there is potential to perform power-savings balance.
3673 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3674 int this_cpu
, unsigned long *imbalance
)
3676 if (!sds
->power_savings_balance
)
3679 if (sds
->this != sds
->group_leader
||
3680 sds
->group_leader
== sds
->group_min
)
3683 *imbalance
= sds
->min_load_per_task
;
3684 sds
->busiest
= sds
->group_min
;
3686 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3687 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3688 group_first_cpu(sds
->group_leader
);
3694 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3695 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3696 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3701 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3702 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3707 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3708 int this_cpu
, unsigned long *imbalance
)
3712 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3714 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3716 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3717 unsigned long smt_gain
= sd
->smt_gain
;
3724 unsigned long scale_rt_power(int cpu
)
3726 struct rq
*rq
= cpu_rq(cpu
);
3727 u64 total
, available
;
3729 sched_avg_update(rq
);
3731 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3732 available
= total
- rq
->rt_avg
;
3734 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3735 total
= SCHED_LOAD_SCALE
;
3737 total
>>= SCHED_LOAD_SHIFT
;
3739 return div_u64(available
, total
);
3742 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3744 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3745 unsigned long power
= SCHED_LOAD_SCALE
;
3746 struct sched_group
*sdg
= sd
->groups
;
3748 /* here we could scale based on cpufreq */
3750 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3751 power
*= arch_scale_smt_power(sd
, cpu
);
3752 power
>>= SCHED_LOAD_SHIFT
;
3755 power
*= scale_rt_power(cpu
);
3756 power
>>= SCHED_LOAD_SHIFT
;
3761 sdg
->cpu_power
= power
;
3764 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3766 struct sched_domain
*child
= sd
->child
;
3767 struct sched_group
*group
, *sdg
= sd
->groups
;
3768 unsigned long power
;
3771 update_cpu_power(sd
, cpu
);
3777 group
= child
->groups
;
3779 power
+= group
->cpu_power
;
3780 group
= group
->next
;
3781 } while (group
!= child
->groups
);
3783 sdg
->cpu_power
= power
;
3787 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3788 * @group: sched_group whose statistics are to be updated.
3789 * @this_cpu: Cpu for which load balance is currently performed.
3790 * @idle: Idle status of this_cpu
3791 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3792 * @sd_idle: Idle status of the sched_domain containing group.
3793 * @local_group: Does group contain this_cpu.
3794 * @cpus: Set of cpus considered for load balancing.
3795 * @balance: Should we balance.
3796 * @sgs: variable to hold the statistics for this group.
3798 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3799 struct sched_group
*group
, int this_cpu
,
3800 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3801 int local_group
, const struct cpumask
*cpus
,
3802 int *balance
, struct sg_lb_stats
*sgs
)
3804 unsigned long load
, max_cpu_load
, min_cpu_load
;
3806 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3807 unsigned long sum_avg_load_per_task
;
3808 unsigned long avg_load_per_task
;
3811 balance_cpu
= group_first_cpu(group
);
3812 if (balance_cpu
== this_cpu
)
3813 update_group_power(sd
, this_cpu
);
3816 /* Tally up the load of all CPUs in the group */
3817 sum_avg_load_per_task
= avg_load_per_task
= 0;
3819 min_cpu_load
= ~0UL;
3821 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3822 struct rq
*rq
= cpu_rq(i
);
3824 if (*sd_idle
&& rq
->nr_running
)
3827 /* Bias balancing toward cpus of our domain */
3829 if (idle_cpu(i
) && !first_idle_cpu
) {
3834 load
= target_load(i
, load_idx
);
3836 load
= source_load(i
, load_idx
);
3837 if (load
> max_cpu_load
)
3838 max_cpu_load
= load
;
3839 if (min_cpu_load
> load
)
3840 min_cpu_load
= load
;
3843 sgs
->group_load
+= load
;
3844 sgs
->sum_nr_running
+= rq
->nr_running
;
3845 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3847 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3851 * First idle cpu or the first cpu(busiest) in this sched group
3852 * is eligible for doing load balancing at this and above
3853 * domains. In the newly idle case, we will allow all the cpu's
3854 * to do the newly idle load balance.
3856 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3857 balance_cpu
!= this_cpu
&& balance
) {
3862 /* Adjust by relative CPU power of the group */
3863 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3867 * Consider the group unbalanced when the imbalance is larger
3868 * than the average weight of two tasks.
3870 * APZ: with cgroup the avg task weight can vary wildly and
3871 * might not be a suitable number - should we keep a
3872 * normalized nr_running number somewhere that negates
3875 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3878 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3881 sgs
->group_capacity
=
3882 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3886 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3887 * @sd: sched_domain whose statistics are to be updated.
3888 * @this_cpu: Cpu for which load balance is currently performed.
3889 * @idle: Idle status of this_cpu
3890 * @sd_idle: Idle status of the sched_domain containing group.
3891 * @cpus: Set of cpus considered for load balancing.
3892 * @balance: Should we balance.
3893 * @sds: variable to hold the statistics for this sched_domain.
3895 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3896 enum cpu_idle_type idle
, int *sd_idle
,
3897 const struct cpumask
*cpus
, int *balance
,
3898 struct sd_lb_stats
*sds
)
3900 struct sched_domain
*child
= sd
->child
;
3901 struct sched_group
*group
= sd
->groups
;
3902 struct sg_lb_stats sgs
;
3903 int load_idx
, prefer_sibling
= 0;
3905 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3908 init_sd_power_savings_stats(sd
, sds
, idle
);
3909 load_idx
= get_sd_load_idx(sd
, idle
);
3914 local_group
= cpumask_test_cpu(this_cpu
,
3915 sched_group_cpus(group
));
3916 memset(&sgs
, 0, sizeof(sgs
));
3917 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3918 local_group
, cpus
, balance
, &sgs
);
3920 if (local_group
&& balance
&& !(*balance
))
3923 sds
->total_load
+= sgs
.group_load
;
3924 sds
->total_pwr
+= group
->cpu_power
;
3927 * In case the child domain prefers tasks go to siblings
3928 * first, lower the group capacity to one so that we'll try
3929 * and move all the excess tasks away.
3932 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3935 sds
->this_load
= sgs
.avg_load
;
3937 sds
->this_nr_running
= sgs
.sum_nr_running
;
3938 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3939 } else if (sgs
.avg_load
> sds
->max_load
&&
3940 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3942 sds
->max_load
= sgs
.avg_load
;
3943 sds
->busiest
= group
;
3944 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3945 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3946 sds
->group_imb
= sgs
.group_imb
;
3949 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3950 group
= group
->next
;
3951 } while (group
!= sd
->groups
);
3955 * fix_small_imbalance - Calculate the minor imbalance that exists
3956 * amongst the groups of a sched_domain, during
3958 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3959 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3960 * @imbalance: Variable to store the imbalance.
3962 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3963 int this_cpu
, unsigned long *imbalance
)
3965 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3966 unsigned int imbn
= 2;
3968 if (sds
->this_nr_running
) {
3969 sds
->this_load_per_task
/= sds
->this_nr_running
;
3970 if (sds
->busiest_load_per_task
>
3971 sds
->this_load_per_task
)
3974 sds
->this_load_per_task
=
3975 cpu_avg_load_per_task(this_cpu
);
3977 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3978 sds
->busiest_load_per_task
* imbn
) {
3979 *imbalance
= sds
->busiest_load_per_task
;
3984 * OK, we don't have enough imbalance to justify moving tasks,
3985 * however we may be able to increase total CPU power used by
3989 pwr_now
+= sds
->busiest
->cpu_power
*
3990 min(sds
->busiest_load_per_task
, sds
->max_load
);
3991 pwr_now
+= sds
->this->cpu_power
*
3992 min(sds
->this_load_per_task
, sds
->this_load
);
3993 pwr_now
/= SCHED_LOAD_SCALE
;
3995 /* Amount of load we'd subtract */
3996 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3997 sds
->busiest
->cpu_power
;
3998 if (sds
->max_load
> tmp
)
3999 pwr_move
+= sds
->busiest
->cpu_power
*
4000 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
4002 /* Amount of load we'd add */
4003 if (sds
->max_load
* sds
->busiest
->cpu_power
<
4004 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
4005 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
4006 sds
->this->cpu_power
;
4008 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
4009 sds
->this->cpu_power
;
4010 pwr_move
+= sds
->this->cpu_power
*
4011 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
4012 pwr_move
/= SCHED_LOAD_SCALE
;
4014 /* Move if we gain throughput */
4015 if (pwr_move
> pwr_now
)
4016 *imbalance
= sds
->busiest_load_per_task
;
4020 * calculate_imbalance - Calculate the amount of imbalance present within the
4021 * groups of a given sched_domain during load balance.
4022 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4023 * @this_cpu: Cpu for which currently load balance is being performed.
4024 * @imbalance: The variable to store the imbalance.
4026 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
4027 unsigned long *imbalance
)
4029 unsigned long max_pull
;
4031 * In the presence of smp nice balancing, certain scenarios can have
4032 * max load less than avg load(as we skip the groups at or below
4033 * its cpu_power, while calculating max_load..)
4035 if (sds
->max_load
< sds
->avg_load
) {
4037 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4040 /* Don't want to pull so many tasks that a group would go idle */
4041 max_pull
= min(sds
->max_load
- sds
->avg_load
,
4042 sds
->max_load
- sds
->busiest_load_per_task
);
4044 /* How much load to actually move to equalise the imbalance */
4045 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
4046 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
4050 * if *imbalance is less than the average load per runnable task
4051 * there is no gaurantee that any tasks will be moved so we'll have
4052 * a think about bumping its value to force at least one task to be
4055 if (*imbalance
< sds
->busiest_load_per_task
)
4056 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
4059 /******* find_busiest_group() helpers end here *********************/
4062 * find_busiest_group - Returns the busiest group within the sched_domain
4063 * if there is an imbalance. If there isn't an imbalance, and
4064 * the user has opted for power-savings, it returns a group whose
4065 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4066 * such a group exists.
4068 * Also calculates the amount of weighted load which should be moved
4069 * to restore balance.
4071 * @sd: The sched_domain whose busiest group is to be returned.
4072 * @this_cpu: The cpu for which load balancing is currently being performed.
4073 * @imbalance: Variable which stores amount of weighted load which should
4074 * be moved to restore balance/put a group to idle.
4075 * @idle: The idle status of this_cpu.
4076 * @sd_idle: The idleness of sd
4077 * @cpus: The set of CPUs under consideration for load-balancing.
4078 * @balance: Pointer to a variable indicating if this_cpu
4079 * is the appropriate cpu to perform load balancing at this_level.
4081 * Returns: - the busiest group if imbalance exists.
4082 * - If no imbalance and user has opted for power-savings balance,
4083 * return the least loaded group whose CPUs can be
4084 * put to idle by rebalancing its tasks onto our group.
4086 static struct sched_group
*
4087 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4088 unsigned long *imbalance
, enum cpu_idle_type idle
,
4089 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4091 struct sd_lb_stats sds
;
4093 memset(&sds
, 0, sizeof(sds
));
4096 * Compute the various statistics relavent for load balancing at
4099 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4102 /* Cases where imbalance does not exist from POV of this_cpu */
4103 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4105 * 2) There is no busy sibling group to pull from.
4106 * 3) This group is the busiest group.
4107 * 4) This group is more busy than the avg busieness at this
4109 * 5) The imbalance is within the specified limit.
4110 * 6) Any rebalance would lead to ping-pong
4112 if (balance
&& !(*balance
))
4115 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4118 if (sds
.this_load
>= sds
.max_load
)
4121 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4123 if (sds
.this_load
>= sds
.avg_load
)
4126 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4129 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4131 sds
.busiest_load_per_task
=
4132 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4135 * We're trying to get all the cpus to the average_load, so we don't
4136 * want to push ourselves above the average load, nor do we wish to
4137 * reduce the max loaded cpu below the average load, as either of these
4138 * actions would just result in more rebalancing later, and ping-pong
4139 * tasks around. Thus we look for the minimum possible imbalance.
4140 * Negative imbalances (*we* are more loaded than anyone else) will
4141 * be counted as no imbalance for these purposes -- we can't fix that
4142 * by pulling tasks to us. Be careful of negative numbers as they'll
4143 * appear as very large values with unsigned longs.
4145 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4148 /* Looks like there is an imbalance. Compute it */
4149 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4154 * There is no obvious imbalance. But check if we can do some balancing
4157 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4164 static struct sched_group
*group_of(int cpu
)
4166 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
4174 static unsigned long power_of(int cpu
)
4176 struct sched_group
*group
= group_of(cpu
);
4179 return SCHED_LOAD_SCALE
;
4181 return group
->cpu_power
;
4185 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4188 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4189 unsigned long imbalance
, const struct cpumask
*cpus
)
4191 struct rq
*busiest
= NULL
, *rq
;
4192 unsigned long max_load
= 0;
4195 for_each_cpu(i
, sched_group_cpus(group
)) {
4196 unsigned long power
= power_of(i
);
4197 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4200 if (!cpumask_test_cpu(i
, cpus
))
4204 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4207 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4210 if (wl
> max_load
) {
4220 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4221 * so long as it is large enough.
4223 #define MAX_PINNED_INTERVAL 512
4225 /* Working cpumask for load_balance and load_balance_newidle. */
4226 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4229 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4230 * tasks if there is an imbalance.
4232 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4233 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4236 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4237 struct sched_group
*group
;
4238 unsigned long imbalance
;
4240 unsigned long flags
;
4241 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4243 cpumask_setall(cpus
);
4246 * When power savings policy is enabled for the parent domain, idle
4247 * sibling can pick up load irrespective of busy siblings. In this case,
4248 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4249 * portraying it as CPU_NOT_IDLE.
4251 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4252 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4255 schedstat_inc(sd
, lb_count
[idle
]);
4259 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4266 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4270 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4272 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4276 BUG_ON(busiest
== this_rq
);
4278 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4281 if (busiest
->nr_running
> 1) {
4283 * Attempt to move tasks. If find_busiest_group has found
4284 * an imbalance but busiest->nr_running <= 1, the group is
4285 * still unbalanced. ld_moved simply stays zero, so it is
4286 * correctly treated as an imbalance.
4288 local_irq_save(flags
);
4289 double_rq_lock(this_rq
, busiest
);
4290 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4291 imbalance
, sd
, idle
, &all_pinned
);
4292 double_rq_unlock(this_rq
, busiest
);
4293 local_irq_restore(flags
);
4296 * some other cpu did the load balance for us.
4298 if (ld_moved
&& this_cpu
!= smp_processor_id())
4299 resched_cpu(this_cpu
);
4301 /* All tasks on this runqueue were pinned by CPU affinity */
4302 if (unlikely(all_pinned
)) {
4303 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4304 if (!cpumask_empty(cpus
))
4311 schedstat_inc(sd
, lb_failed
[idle
]);
4312 sd
->nr_balance_failed
++;
4314 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4316 spin_lock_irqsave(&busiest
->lock
, flags
);
4318 /* don't kick the migration_thread, if the curr
4319 * task on busiest cpu can't be moved to this_cpu
4321 if (!cpumask_test_cpu(this_cpu
,
4322 &busiest
->curr
->cpus_allowed
)) {
4323 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4325 goto out_one_pinned
;
4328 if (!busiest
->active_balance
) {
4329 busiest
->active_balance
= 1;
4330 busiest
->push_cpu
= this_cpu
;
4333 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4335 wake_up_process(busiest
->migration_thread
);
4338 * We've kicked active balancing, reset the failure
4341 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4344 sd
->nr_balance_failed
= 0;
4346 if (likely(!active_balance
)) {
4347 /* We were unbalanced, so reset the balancing interval */
4348 sd
->balance_interval
= sd
->min_interval
;
4351 * If we've begun active balancing, start to back off. This
4352 * case may not be covered by the all_pinned logic if there
4353 * is only 1 task on the busy runqueue (because we don't call
4356 if (sd
->balance_interval
< sd
->max_interval
)
4357 sd
->balance_interval
*= 2;
4360 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4361 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4367 schedstat_inc(sd
, lb_balanced
[idle
]);
4369 sd
->nr_balance_failed
= 0;
4372 /* tune up the balancing interval */
4373 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4374 (sd
->balance_interval
< sd
->max_interval
))
4375 sd
->balance_interval
*= 2;
4377 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4378 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4389 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4390 * tasks if there is an imbalance.
4392 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4393 * this_rq is locked.
4396 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4398 struct sched_group
*group
;
4399 struct rq
*busiest
= NULL
;
4400 unsigned long imbalance
;
4404 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4406 cpumask_setall(cpus
);
4409 * When power savings policy is enabled for the parent domain, idle
4410 * sibling can pick up load irrespective of busy siblings. In this case,
4411 * let the state of idle sibling percolate up as IDLE, instead of
4412 * portraying it as CPU_NOT_IDLE.
4414 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4415 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4418 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4420 update_shares_locked(this_rq
, sd
);
4421 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4422 &sd_idle
, cpus
, NULL
);
4424 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4428 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4430 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4434 BUG_ON(busiest
== this_rq
);
4436 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4439 if (busiest
->nr_running
> 1) {
4440 /* Attempt to move tasks */
4441 double_lock_balance(this_rq
, busiest
);
4442 /* this_rq->clock is already updated */
4443 update_rq_clock(busiest
);
4444 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4445 imbalance
, sd
, CPU_NEWLY_IDLE
,
4447 double_unlock_balance(this_rq
, busiest
);
4449 if (unlikely(all_pinned
)) {
4450 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4451 if (!cpumask_empty(cpus
))
4457 int active_balance
= 0;
4459 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4460 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4461 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4464 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4467 if (sd
->nr_balance_failed
++ < 2)
4471 * The only task running in a non-idle cpu can be moved to this
4472 * cpu in an attempt to completely freeup the other CPU
4473 * package. The same method used to move task in load_balance()
4474 * have been extended for load_balance_newidle() to speedup
4475 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4477 * The package power saving logic comes from
4478 * find_busiest_group(). If there are no imbalance, then
4479 * f_b_g() will return NULL. However when sched_mc={1,2} then
4480 * f_b_g() will select a group from which a running task may be
4481 * pulled to this cpu in order to make the other package idle.
4482 * If there is no opportunity to make a package idle and if
4483 * there are no imbalance, then f_b_g() will return NULL and no
4484 * action will be taken in load_balance_newidle().
4486 * Under normal task pull operation due to imbalance, there
4487 * will be more than one task in the source run queue and
4488 * move_tasks() will succeed. ld_moved will be true and this
4489 * active balance code will not be triggered.
4492 /* Lock busiest in correct order while this_rq is held */
4493 double_lock_balance(this_rq
, busiest
);
4496 * don't kick the migration_thread, if the curr
4497 * task on busiest cpu can't be moved to this_cpu
4499 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4500 double_unlock_balance(this_rq
, busiest
);
4505 if (!busiest
->active_balance
) {
4506 busiest
->active_balance
= 1;
4507 busiest
->push_cpu
= this_cpu
;
4511 double_unlock_balance(this_rq
, busiest
);
4513 * Should not call ttwu while holding a rq->lock
4515 spin_unlock(&this_rq
->lock
);
4517 wake_up_process(busiest
->migration_thread
);
4518 spin_lock(&this_rq
->lock
);
4521 sd
->nr_balance_failed
= 0;
4523 update_shares_locked(this_rq
, sd
);
4527 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4528 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4529 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4531 sd
->nr_balance_failed
= 0;
4537 * idle_balance is called by schedule() if this_cpu is about to become
4538 * idle. Attempts to pull tasks from other CPUs.
4540 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4542 struct sched_domain
*sd
;
4543 int pulled_task
= 0;
4544 unsigned long next_balance
= jiffies
+ HZ
;
4546 for_each_domain(this_cpu
, sd
) {
4547 unsigned long interval
;
4549 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4552 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4553 /* If we've pulled tasks over stop searching: */
4554 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4557 interval
= msecs_to_jiffies(sd
->balance_interval
);
4558 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4559 next_balance
= sd
->last_balance
+ interval
;
4563 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4565 * We are going idle. next_balance may be set based on
4566 * a busy processor. So reset next_balance.
4568 this_rq
->next_balance
= next_balance
;
4573 * active_load_balance is run by migration threads. It pushes running tasks
4574 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4575 * running on each physical CPU where possible, and avoids physical /
4576 * logical imbalances.
4578 * Called with busiest_rq locked.
4580 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4582 int target_cpu
= busiest_rq
->push_cpu
;
4583 struct sched_domain
*sd
;
4584 struct rq
*target_rq
;
4586 /* Is there any task to move? */
4587 if (busiest_rq
->nr_running
<= 1)
4590 target_rq
= cpu_rq(target_cpu
);
4593 * This condition is "impossible", if it occurs
4594 * we need to fix it. Originally reported by
4595 * Bjorn Helgaas on a 128-cpu setup.
4597 BUG_ON(busiest_rq
== target_rq
);
4599 /* move a task from busiest_rq to target_rq */
4600 double_lock_balance(busiest_rq
, target_rq
);
4601 update_rq_clock(busiest_rq
);
4602 update_rq_clock(target_rq
);
4604 /* Search for an sd spanning us and the target CPU. */
4605 for_each_domain(target_cpu
, sd
) {
4606 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4607 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4612 schedstat_inc(sd
, alb_count
);
4614 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4616 schedstat_inc(sd
, alb_pushed
);
4618 schedstat_inc(sd
, alb_failed
);
4620 double_unlock_balance(busiest_rq
, target_rq
);
4625 atomic_t load_balancer
;
4626 cpumask_var_t cpu_mask
;
4627 cpumask_var_t ilb_grp_nohz_mask
;
4628 } nohz ____cacheline_aligned
= {
4629 .load_balancer
= ATOMIC_INIT(-1),
4632 int get_nohz_load_balancer(void)
4634 return atomic_read(&nohz
.load_balancer
);
4637 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4639 * lowest_flag_domain - Return lowest sched_domain containing flag.
4640 * @cpu: The cpu whose lowest level of sched domain is to
4642 * @flag: The flag to check for the lowest sched_domain
4643 * for the given cpu.
4645 * Returns the lowest sched_domain of a cpu which contains the given flag.
4647 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4649 struct sched_domain
*sd
;
4651 for_each_domain(cpu
, sd
)
4652 if (sd
&& (sd
->flags
& flag
))
4659 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4660 * @cpu: The cpu whose domains we're iterating over.
4661 * @sd: variable holding the value of the power_savings_sd
4663 * @flag: The flag to filter the sched_domains to be iterated.
4665 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4666 * set, starting from the lowest sched_domain to the highest.
4668 #define for_each_flag_domain(cpu, sd, flag) \
4669 for (sd = lowest_flag_domain(cpu, flag); \
4670 (sd && (sd->flags & flag)); sd = sd->parent)
4673 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4674 * @ilb_group: group to be checked for semi-idleness
4676 * Returns: 1 if the group is semi-idle. 0 otherwise.
4678 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4679 * and atleast one non-idle CPU. This helper function checks if the given
4680 * sched_group is semi-idle or not.
4682 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4684 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4685 sched_group_cpus(ilb_group
));
4688 * A sched_group is semi-idle when it has atleast one busy cpu
4689 * and atleast one idle cpu.
4691 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4694 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4700 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4701 * @cpu: The cpu which is nominating a new idle_load_balancer.
4703 * Returns: Returns the id of the idle load balancer if it exists,
4704 * Else, returns >= nr_cpu_ids.
4706 * This algorithm picks the idle load balancer such that it belongs to a
4707 * semi-idle powersavings sched_domain. The idea is to try and avoid
4708 * completely idle packages/cores just for the purpose of idle load balancing
4709 * when there are other idle cpu's which are better suited for that job.
4711 static int find_new_ilb(int cpu
)
4713 struct sched_domain
*sd
;
4714 struct sched_group
*ilb_group
;
4717 * Have idle load balancer selection from semi-idle packages only
4718 * when power-aware load balancing is enabled
4720 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4724 * Optimize for the case when we have no idle CPUs or only one
4725 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4727 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4730 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4731 ilb_group
= sd
->groups
;
4734 if (is_semi_idle_group(ilb_group
))
4735 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4737 ilb_group
= ilb_group
->next
;
4739 } while (ilb_group
!= sd
->groups
);
4743 return cpumask_first(nohz
.cpu_mask
);
4745 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4746 static inline int find_new_ilb(int call_cpu
)
4748 return cpumask_first(nohz
.cpu_mask
);
4753 * This routine will try to nominate the ilb (idle load balancing)
4754 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4755 * load balancing on behalf of all those cpus. If all the cpus in the system
4756 * go into this tickless mode, then there will be no ilb owner (as there is
4757 * no need for one) and all the cpus will sleep till the next wakeup event
4760 * For the ilb owner, tick is not stopped. And this tick will be used
4761 * for idle load balancing. ilb owner will still be part of
4764 * While stopping the tick, this cpu will become the ilb owner if there
4765 * is no other owner. And will be the owner till that cpu becomes busy
4766 * or if all cpus in the system stop their ticks at which point
4767 * there is no need for ilb owner.
4769 * When the ilb owner becomes busy, it nominates another owner, during the
4770 * next busy scheduler_tick()
4772 int select_nohz_load_balancer(int stop_tick
)
4774 int cpu
= smp_processor_id();
4777 cpu_rq(cpu
)->in_nohz_recently
= 1;
4779 if (!cpu_active(cpu
)) {
4780 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4784 * If we are going offline and still the leader,
4787 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4793 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4795 /* time for ilb owner also to sleep */
4796 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4797 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4798 atomic_set(&nohz
.load_balancer
, -1);
4802 if (atomic_read(&nohz
.load_balancer
) == -1) {
4803 /* make me the ilb owner */
4804 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4806 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4809 if (!(sched_smt_power_savings
||
4810 sched_mc_power_savings
))
4813 * Check to see if there is a more power-efficient
4816 new_ilb
= find_new_ilb(cpu
);
4817 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4818 atomic_set(&nohz
.load_balancer
, -1);
4819 resched_cpu(new_ilb
);
4825 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4828 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4830 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4831 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4838 static DEFINE_SPINLOCK(balancing
);
4841 * It checks each scheduling domain to see if it is due to be balanced,
4842 * and initiates a balancing operation if so.
4844 * Balancing parameters are set up in arch_init_sched_domains.
4846 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4849 struct rq
*rq
= cpu_rq(cpu
);
4850 unsigned long interval
;
4851 struct sched_domain
*sd
;
4852 /* Earliest time when we have to do rebalance again */
4853 unsigned long next_balance
= jiffies
+ 60*HZ
;
4854 int update_next_balance
= 0;
4857 for_each_domain(cpu
, sd
) {
4858 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4861 interval
= sd
->balance_interval
;
4862 if (idle
!= CPU_IDLE
)
4863 interval
*= sd
->busy_factor
;
4865 /* scale ms to jiffies */
4866 interval
= msecs_to_jiffies(interval
);
4867 if (unlikely(!interval
))
4869 if (interval
> HZ
*NR_CPUS
/10)
4870 interval
= HZ
*NR_CPUS
/10;
4872 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4874 if (need_serialize
) {
4875 if (!spin_trylock(&balancing
))
4879 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4880 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4882 * We've pulled tasks over so either we're no
4883 * longer idle, or one of our SMT siblings is
4886 idle
= CPU_NOT_IDLE
;
4888 sd
->last_balance
= jiffies
;
4891 spin_unlock(&balancing
);
4893 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4894 next_balance
= sd
->last_balance
+ interval
;
4895 update_next_balance
= 1;
4899 * Stop the load balance at this level. There is another
4900 * CPU in our sched group which is doing load balancing more
4908 * next_balance will be updated only when there is a need.
4909 * When the cpu is attached to null domain for ex, it will not be
4912 if (likely(update_next_balance
))
4913 rq
->next_balance
= next_balance
;
4917 * run_rebalance_domains is triggered when needed from the scheduler tick.
4918 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4919 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4921 static void run_rebalance_domains(struct softirq_action
*h
)
4923 int this_cpu
= smp_processor_id();
4924 struct rq
*this_rq
= cpu_rq(this_cpu
);
4925 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4926 CPU_IDLE
: CPU_NOT_IDLE
;
4928 rebalance_domains(this_cpu
, idle
);
4932 * If this cpu is the owner for idle load balancing, then do the
4933 * balancing on behalf of the other idle cpus whose ticks are
4936 if (this_rq
->idle_at_tick
&&
4937 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4941 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4942 if (balance_cpu
== this_cpu
)
4946 * If this cpu gets work to do, stop the load balancing
4947 * work being done for other cpus. Next load
4948 * balancing owner will pick it up.
4953 rebalance_domains(balance_cpu
, CPU_IDLE
);
4955 rq
= cpu_rq(balance_cpu
);
4956 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4957 this_rq
->next_balance
= rq
->next_balance
;
4963 static inline int on_null_domain(int cpu
)
4965 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4969 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4971 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4972 * idle load balancing owner or decide to stop the periodic load balancing,
4973 * if the whole system is idle.
4975 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4979 * If we were in the nohz mode recently and busy at the current
4980 * scheduler tick, then check if we need to nominate new idle
4983 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4984 rq
->in_nohz_recently
= 0;
4986 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4987 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4988 atomic_set(&nohz
.load_balancer
, -1);
4991 if (atomic_read(&nohz
.load_balancer
) == -1) {
4992 int ilb
= find_new_ilb(cpu
);
4994 if (ilb
< nr_cpu_ids
)
5000 * If this cpu is idle and doing idle load balancing for all the
5001 * cpus with ticks stopped, is it time for that to stop?
5003 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
5004 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
5010 * If this cpu is idle and the idle load balancing is done by
5011 * someone else, then no need raise the SCHED_SOFTIRQ
5013 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
5014 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
5017 /* Don't need to rebalance while attached to NULL domain */
5018 if (time_after_eq(jiffies
, rq
->next_balance
) &&
5019 likely(!on_null_domain(cpu
)))
5020 raise_softirq(SCHED_SOFTIRQ
);
5023 #else /* CONFIG_SMP */
5026 * on UP we do not need to balance between CPUs:
5028 static inline void idle_balance(int cpu
, struct rq
*rq
)
5034 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
5036 EXPORT_PER_CPU_SYMBOL(kstat
);
5039 * Return any ns on the sched_clock that have not yet been accounted in
5040 * @p in case that task is currently running.
5042 * Called with task_rq_lock() held on @rq.
5044 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
5048 if (task_current(rq
, p
)) {
5049 update_rq_clock(rq
);
5050 ns
= rq
->clock
- p
->se
.exec_start
;
5058 unsigned long long task_delta_exec(struct task_struct
*p
)
5060 unsigned long flags
;
5064 rq
= task_rq_lock(p
, &flags
);
5065 ns
= do_task_delta_exec(p
, rq
);
5066 task_rq_unlock(rq
, &flags
);
5072 * Return accounted runtime for the task.
5073 * In case the task is currently running, return the runtime plus current's
5074 * pending runtime that have not been accounted yet.
5076 unsigned long long task_sched_runtime(struct task_struct
*p
)
5078 unsigned long flags
;
5082 rq
= task_rq_lock(p
, &flags
);
5083 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5084 task_rq_unlock(rq
, &flags
);
5090 * Return sum_exec_runtime for the thread group.
5091 * In case the task is currently running, return the sum plus current's
5092 * pending runtime that have not been accounted yet.
5094 * Note that the thread group might have other running tasks as well,
5095 * so the return value not includes other pending runtime that other
5096 * running tasks might have.
5098 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5100 struct task_cputime totals
;
5101 unsigned long flags
;
5105 rq
= task_rq_lock(p
, &flags
);
5106 thread_group_cputime(p
, &totals
);
5107 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5108 task_rq_unlock(rq
, &flags
);
5114 * Account user cpu time to a process.
5115 * @p: the process that the cpu time gets accounted to
5116 * @cputime: the cpu time spent in user space since the last update
5117 * @cputime_scaled: cputime scaled by cpu frequency
5119 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5120 cputime_t cputime_scaled
)
5122 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5125 /* Add user time to process. */
5126 p
->utime
= cputime_add(p
->utime
, cputime
);
5127 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5128 account_group_user_time(p
, cputime
);
5130 /* Add user time to cpustat. */
5131 tmp
= cputime_to_cputime64(cputime
);
5132 if (TASK_NICE(p
) > 0)
5133 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5135 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5137 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5138 /* Account for user time used */
5139 acct_update_integrals(p
);
5143 * Account guest cpu time to a process.
5144 * @p: the process that the cpu time gets accounted to
5145 * @cputime: the cpu time spent in virtual machine since the last update
5146 * @cputime_scaled: cputime scaled by cpu frequency
5148 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5149 cputime_t cputime_scaled
)
5152 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5154 tmp
= cputime_to_cputime64(cputime
);
5156 /* Add guest time to process. */
5157 p
->utime
= cputime_add(p
->utime
, cputime
);
5158 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5159 account_group_user_time(p
, cputime
);
5160 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5162 /* Add guest time to cpustat. */
5163 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5164 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5168 * Account system cpu time to a process.
5169 * @p: the process that the cpu time gets accounted to
5170 * @hardirq_offset: the offset to subtract from hardirq_count()
5171 * @cputime: the cpu time spent in kernel space since the last update
5172 * @cputime_scaled: cputime scaled by cpu frequency
5174 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5175 cputime_t cputime
, cputime_t cputime_scaled
)
5177 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5180 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5181 account_guest_time(p
, cputime
, cputime_scaled
);
5185 /* Add system time to process. */
5186 p
->stime
= cputime_add(p
->stime
, cputime
);
5187 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5188 account_group_system_time(p
, cputime
);
5190 /* Add system time to cpustat. */
5191 tmp
= cputime_to_cputime64(cputime
);
5192 if (hardirq_count() - hardirq_offset
)
5193 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5194 else if (softirq_count())
5195 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5197 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5199 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5201 /* Account for system time used */
5202 acct_update_integrals(p
);
5206 * Account for involuntary wait time.
5207 * @steal: the cpu time spent in involuntary wait
5209 void account_steal_time(cputime_t cputime
)
5211 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5212 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5214 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5218 * Account for idle time.
5219 * @cputime: the cpu time spent in idle wait
5221 void account_idle_time(cputime_t cputime
)
5223 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5224 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5225 struct rq
*rq
= this_rq();
5227 if (atomic_read(&rq
->nr_iowait
) > 0)
5228 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5230 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5233 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5236 * Account a single tick of cpu time.
5237 * @p: the process that the cpu time gets accounted to
5238 * @user_tick: indicates if the tick is a user or a system tick
5240 void account_process_tick(struct task_struct
*p
, int user_tick
)
5242 cputime_t one_jiffy
= jiffies_to_cputime(1);
5243 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5244 struct rq
*rq
= this_rq();
5247 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5248 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5249 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5252 account_idle_time(one_jiffy
);
5256 * Account multiple ticks of steal time.
5257 * @p: the process from which the cpu time has been stolen
5258 * @ticks: number of stolen ticks
5260 void account_steal_ticks(unsigned long ticks
)
5262 account_steal_time(jiffies_to_cputime(ticks
));
5266 * Account multiple ticks of idle time.
5267 * @ticks: number of stolen ticks
5269 void account_idle_ticks(unsigned long ticks
)
5271 account_idle_time(jiffies_to_cputime(ticks
));
5277 * Use precise platform statistics if available:
5279 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5280 cputime_t
task_utime(struct task_struct
*p
)
5285 cputime_t
task_stime(struct task_struct
*p
)
5290 cputime_t
task_utime(struct task_struct
*p
)
5292 clock_t utime
= cputime_to_clock_t(p
->utime
),
5293 total
= utime
+ cputime_to_clock_t(p
->stime
);
5297 * Use CFS's precise accounting:
5299 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5303 do_div(temp
, total
);
5305 utime
= (clock_t)temp
;
5307 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5308 return p
->prev_utime
;
5311 cputime_t
task_stime(struct task_struct
*p
)
5316 * Use CFS's precise accounting. (we subtract utime from
5317 * the total, to make sure the total observed by userspace
5318 * grows monotonically - apps rely on that):
5320 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5321 cputime_to_clock_t(task_utime(p
));
5324 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5326 return p
->prev_stime
;
5330 inline cputime_t
task_gtime(struct task_struct
*p
)
5336 * This function gets called by the timer code, with HZ frequency.
5337 * We call it with interrupts disabled.
5339 * It also gets called by the fork code, when changing the parent's
5342 void scheduler_tick(void)
5344 int cpu
= smp_processor_id();
5345 struct rq
*rq
= cpu_rq(cpu
);
5346 struct task_struct
*curr
= rq
->curr
;
5350 spin_lock(&rq
->lock
);
5351 update_rq_clock(rq
);
5352 update_cpu_load(rq
);
5353 curr
->sched_class
->task_tick(rq
, curr
, 0);
5354 spin_unlock(&rq
->lock
);
5356 perf_counter_task_tick(curr
, cpu
);
5359 rq
->idle_at_tick
= idle_cpu(cpu
);
5360 trigger_load_balance(rq
, cpu
);
5364 notrace
unsigned long get_parent_ip(unsigned long addr
)
5366 if (in_lock_functions(addr
)) {
5367 addr
= CALLER_ADDR2
;
5368 if (in_lock_functions(addr
))
5369 addr
= CALLER_ADDR3
;
5374 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5375 defined(CONFIG_PREEMPT_TRACER))
5377 void __kprobes
add_preempt_count(int val
)
5379 #ifdef CONFIG_DEBUG_PREEMPT
5383 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5386 preempt_count() += val
;
5387 #ifdef CONFIG_DEBUG_PREEMPT
5389 * Spinlock count overflowing soon?
5391 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5394 if (preempt_count() == val
)
5395 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5397 EXPORT_SYMBOL(add_preempt_count
);
5399 void __kprobes
sub_preempt_count(int val
)
5401 #ifdef CONFIG_DEBUG_PREEMPT
5405 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5408 * Is the spinlock portion underflowing?
5410 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5411 !(preempt_count() & PREEMPT_MASK
)))
5415 if (preempt_count() == val
)
5416 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5417 preempt_count() -= val
;
5419 EXPORT_SYMBOL(sub_preempt_count
);
5424 * Print scheduling while atomic bug:
5426 static noinline
void __schedule_bug(struct task_struct
*prev
)
5428 struct pt_regs
*regs
= get_irq_regs();
5430 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
5431 prev
->comm
, prev
->pid
, preempt_count());
5433 debug_show_held_locks(prev
);
5435 if (irqs_disabled())
5436 print_irqtrace_events(prev
);
5445 * Various schedule()-time debugging checks and statistics:
5447 static inline void schedule_debug(struct task_struct
*prev
)
5450 * Test if we are atomic. Since do_exit() needs to call into
5451 * schedule() atomically, we ignore that path for now.
5452 * Otherwise, whine if we are scheduling when we should not be.
5454 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5455 __schedule_bug(prev
);
5457 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5459 schedstat_inc(this_rq(), sched_count
);
5460 #ifdef CONFIG_SCHEDSTATS
5461 if (unlikely(prev
->lock_depth
>= 0)) {
5462 schedstat_inc(this_rq(), bkl_count
);
5463 schedstat_inc(prev
, sched_info
.bkl_count
);
5468 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5470 if (prev
->state
== TASK_RUNNING
) {
5471 u64 runtime
= prev
->se
.sum_exec_runtime
;
5473 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5474 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5477 * In order to avoid avg_overlap growing stale when we are
5478 * indeed overlapping and hence not getting put to sleep, grow
5479 * the avg_overlap on preemption.
5481 * We use the average preemption runtime because that
5482 * correlates to the amount of cache footprint a task can
5485 update_avg(&prev
->se
.avg_overlap
, runtime
);
5487 prev
->sched_class
->put_prev_task(rq
, prev
);
5491 * Pick up the highest-prio task:
5493 static inline struct task_struct
*
5494 pick_next_task(struct rq
*rq
)
5496 const struct sched_class
*class;
5497 struct task_struct
*p
;
5500 * Optimization: we know that if all tasks are in
5501 * the fair class we can call that function directly:
5503 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5504 p
= fair_sched_class
.pick_next_task(rq
);
5509 class = sched_class_highest
;
5511 p
= class->pick_next_task(rq
);
5515 * Will never be NULL as the idle class always
5516 * returns a non-NULL p:
5518 class = class->next
;
5523 * schedule() is the main scheduler function.
5525 asmlinkage
void __sched
schedule(void)
5527 struct task_struct
*prev
, *next
;
5528 unsigned long *switch_count
;
5534 cpu
= smp_processor_id();
5538 switch_count
= &prev
->nivcsw
;
5540 release_kernel_lock(prev
);
5541 need_resched_nonpreemptible
:
5543 schedule_debug(prev
);
5545 if (sched_feat(HRTICK
))
5548 spin_lock_irq(&rq
->lock
);
5549 update_rq_clock(rq
);
5550 clear_tsk_need_resched(prev
);
5552 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5553 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5554 prev
->state
= TASK_RUNNING
;
5556 deactivate_task(rq
, prev
, 1);
5557 switch_count
= &prev
->nvcsw
;
5560 pre_schedule(rq
, prev
);
5562 if (unlikely(!rq
->nr_running
))
5563 idle_balance(cpu
, rq
);
5565 put_prev_task(rq
, prev
);
5566 next
= pick_next_task(rq
);
5568 if (likely(prev
!= next
)) {
5569 sched_info_switch(prev
, next
);
5570 perf_counter_task_sched_out(prev
, next
, cpu
);
5576 context_switch(rq
, prev
, next
); /* unlocks the rq */
5578 * the context switch might have flipped the stack from under
5579 * us, hence refresh the local variables.
5581 cpu
= smp_processor_id();
5584 spin_unlock_irq(&rq
->lock
);
5588 if (unlikely(reacquire_kernel_lock(current
) < 0))
5589 goto need_resched_nonpreemptible
;
5591 preempt_enable_no_resched();
5595 EXPORT_SYMBOL(schedule
);
5599 * Look out! "owner" is an entirely speculative pointer
5600 * access and not reliable.
5602 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5607 if (!sched_feat(OWNER_SPIN
))
5610 #ifdef CONFIG_DEBUG_PAGEALLOC
5612 * Need to access the cpu field knowing that
5613 * DEBUG_PAGEALLOC could have unmapped it if
5614 * the mutex owner just released it and exited.
5616 if (probe_kernel_address(&owner
->cpu
, cpu
))
5623 * Even if the access succeeded (likely case),
5624 * the cpu field may no longer be valid.
5626 if (cpu
>= nr_cpumask_bits
)
5630 * We need to validate that we can do a
5631 * get_cpu() and that we have the percpu area.
5633 if (!cpu_online(cpu
))
5640 * Owner changed, break to re-assess state.
5642 if (lock
->owner
!= owner
)
5646 * Is that owner really running on that cpu?
5648 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5658 #ifdef CONFIG_PREEMPT
5660 * this is the entry point to schedule() from in-kernel preemption
5661 * off of preempt_enable. Kernel preemptions off return from interrupt
5662 * occur there and call schedule directly.
5664 asmlinkage
void __sched
preempt_schedule(void)
5666 struct thread_info
*ti
= current_thread_info();
5669 * If there is a non-zero preempt_count or interrupts are disabled,
5670 * we do not want to preempt the current task. Just return..
5672 if (likely(ti
->preempt_count
|| irqs_disabled()))
5676 add_preempt_count(PREEMPT_ACTIVE
);
5678 sub_preempt_count(PREEMPT_ACTIVE
);
5681 * Check again in case we missed a preemption opportunity
5682 * between schedule and now.
5685 } while (need_resched());
5687 EXPORT_SYMBOL(preempt_schedule
);
5690 * this is the entry point to schedule() from kernel preemption
5691 * off of irq context.
5692 * Note, that this is called and return with irqs disabled. This will
5693 * protect us against recursive calling from irq.
5695 asmlinkage
void __sched
preempt_schedule_irq(void)
5697 struct thread_info
*ti
= current_thread_info();
5699 /* Catch callers which need to be fixed */
5700 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5703 add_preempt_count(PREEMPT_ACTIVE
);
5706 local_irq_disable();
5707 sub_preempt_count(PREEMPT_ACTIVE
);
5710 * Check again in case we missed a preemption opportunity
5711 * between schedule and now.
5714 } while (need_resched());
5717 #endif /* CONFIG_PREEMPT */
5719 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5722 return try_to_wake_up(curr
->private, mode
, sync
);
5724 EXPORT_SYMBOL(default_wake_function
);
5727 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5728 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5729 * number) then we wake all the non-exclusive tasks and one exclusive task.
5731 * There are circumstances in which we can try to wake a task which has already
5732 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5733 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5735 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5736 int nr_exclusive
, int sync
, void *key
)
5738 wait_queue_t
*curr
, *next
;
5740 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5741 unsigned flags
= curr
->flags
;
5743 if (curr
->func(curr
, mode
, sync
, key
) &&
5744 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5750 * __wake_up - wake up threads blocked on a waitqueue.
5752 * @mode: which threads
5753 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5754 * @key: is directly passed to the wakeup function
5756 * It may be assumed that this function implies a write memory barrier before
5757 * changing the task state if and only if any tasks are woken up.
5759 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5760 int nr_exclusive
, void *key
)
5762 unsigned long flags
;
5764 spin_lock_irqsave(&q
->lock
, flags
);
5765 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5766 spin_unlock_irqrestore(&q
->lock
, flags
);
5768 EXPORT_SYMBOL(__wake_up
);
5771 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5773 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5775 __wake_up_common(q
, mode
, 1, 0, NULL
);
5778 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5780 __wake_up_common(q
, mode
, 1, 0, key
);
5784 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5786 * @mode: which threads
5787 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5788 * @key: opaque value to be passed to wakeup targets
5790 * The sync wakeup differs that the waker knows that it will schedule
5791 * away soon, so while the target thread will be woken up, it will not
5792 * be migrated to another CPU - ie. the two threads are 'synchronized'
5793 * with each other. This can prevent needless bouncing between CPUs.
5795 * On UP it can prevent extra preemption.
5797 * It may be assumed that this function implies a write memory barrier before
5798 * changing the task state if and only if any tasks are woken up.
5800 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5801 int nr_exclusive
, void *key
)
5803 unsigned long flags
;
5809 if (unlikely(!nr_exclusive
))
5812 spin_lock_irqsave(&q
->lock
, flags
);
5813 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5814 spin_unlock_irqrestore(&q
->lock
, flags
);
5816 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5819 * __wake_up_sync - see __wake_up_sync_key()
5821 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5823 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5825 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5828 * complete: - signals a single thread waiting on this completion
5829 * @x: holds the state of this particular completion
5831 * This will wake up a single thread waiting on this completion. Threads will be
5832 * awakened in the same order in which they were queued.
5834 * See also complete_all(), wait_for_completion() and related routines.
5836 * It may be assumed that this function implies a write memory barrier before
5837 * changing the task state if and only if any tasks are woken up.
5839 void complete(struct completion
*x
)
5841 unsigned long flags
;
5843 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5845 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5846 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5848 EXPORT_SYMBOL(complete
);
5851 * complete_all: - signals all threads waiting on this completion
5852 * @x: holds the state of this particular completion
5854 * This will wake up all threads waiting on this particular completion event.
5856 * It may be assumed that this function implies a write memory barrier before
5857 * changing the task state if and only if any tasks are woken up.
5859 void complete_all(struct completion
*x
)
5861 unsigned long flags
;
5863 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5864 x
->done
+= UINT_MAX
/2;
5865 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5866 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5868 EXPORT_SYMBOL(complete_all
);
5870 static inline long __sched
5871 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5874 DECLARE_WAITQUEUE(wait
, current
);
5876 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5877 __add_wait_queue_tail(&x
->wait
, &wait
);
5879 if (signal_pending_state(state
, current
)) {
5880 timeout
= -ERESTARTSYS
;
5883 __set_current_state(state
);
5884 spin_unlock_irq(&x
->wait
.lock
);
5885 timeout
= schedule_timeout(timeout
);
5886 spin_lock_irq(&x
->wait
.lock
);
5887 } while (!x
->done
&& timeout
);
5888 __remove_wait_queue(&x
->wait
, &wait
);
5893 return timeout
?: 1;
5897 wait_for_common(struct completion
*x
, long timeout
, int state
)
5901 spin_lock_irq(&x
->wait
.lock
);
5902 timeout
= do_wait_for_common(x
, timeout
, state
);
5903 spin_unlock_irq(&x
->wait
.lock
);
5908 * wait_for_completion: - waits for completion of a task
5909 * @x: holds the state of this particular completion
5911 * This waits to be signaled for completion of a specific task. It is NOT
5912 * interruptible and there is no timeout.
5914 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5915 * and interrupt capability. Also see complete().
5917 void __sched
wait_for_completion(struct completion
*x
)
5919 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5921 EXPORT_SYMBOL(wait_for_completion
);
5924 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5925 * @x: holds the state of this particular completion
5926 * @timeout: timeout value in jiffies
5928 * This waits for either a completion of a specific task to be signaled or for a
5929 * specified timeout to expire. The timeout is in jiffies. It is not
5932 unsigned long __sched
5933 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5935 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5937 EXPORT_SYMBOL(wait_for_completion_timeout
);
5940 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5941 * @x: holds the state of this particular completion
5943 * This waits for completion of a specific task to be signaled. It is
5946 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5948 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5949 if (t
== -ERESTARTSYS
)
5953 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5956 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5957 * @x: holds the state of this particular completion
5958 * @timeout: timeout value in jiffies
5960 * This waits for either a completion of a specific task to be signaled or for a
5961 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5963 unsigned long __sched
5964 wait_for_completion_interruptible_timeout(struct completion
*x
,
5965 unsigned long timeout
)
5967 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5969 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5972 * wait_for_completion_killable: - waits for completion of a task (killable)
5973 * @x: holds the state of this particular completion
5975 * This waits to be signaled for completion of a specific task. It can be
5976 * interrupted by a kill signal.
5978 int __sched
wait_for_completion_killable(struct completion
*x
)
5980 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5981 if (t
== -ERESTARTSYS
)
5985 EXPORT_SYMBOL(wait_for_completion_killable
);
5988 * try_wait_for_completion - try to decrement a completion without blocking
5989 * @x: completion structure
5991 * Returns: 0 if a decrement cannot be done without blocking
5992 * 1 if a decrement succeeded.
5994 * If a completion is being used as a counting completion,
5995 * attempt to decrement the counter without blocking. This
5996 * enables us to avoid waiting if the resource the completion
5997 * is protecting is not available.
5999 bool try_wait_for_completion(struct completion
*x
)
6003 spin_lock_irq(&x
->wait
.lock
);
6008 spin_unlock_irq(&x
->wait
.lock
);
6011 EXPORT_SYMBOL(try_wait_for_completion
);
6014 * completion_done - Test to see if a completion has any waiters
6015 * @x: completion structure
6017 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6018 * 1 if there are no waiters.
6021 bool completion_done(struct completion
*x
)
6025 spin_lock_irq(&x
->wait
.lock
);
6028 spin_unlock_irq(&x
->wait
.lock
);
6031 EXPORT_SYMBOL(completion_done
);
6034 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
6036 unsigned long flags
;
6039 init_waitqueue_entry(&wait
, current
);
6041 __set_current_state(state
);
6043 spin_lock_irqsave(&q
->lock
, flags
);
6044 __add_wait_queue(q
, &wait
);
6045 spin_unlock(&q
->lock
);
6046 timeout
= schedule_timeout(timeout
);
6047 spin_lock_irq(&q
->lock
);
6048 __remove_wait_queue(q
, &wait
);
6049 spin_unlock_irqrestore(&q
->lock
, flags
);
6054 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
6056 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6058 EXPORT_SYMBOL(interruptible_sleep_on
);
6061 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6063 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
6065 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6067 void __sched
sleep_on(wait_queue_head_t
*q
)
6069 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6071 EXPORT_SYMBOL(sleep_on
);
6073 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6075 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6077 EXPORT_SYMBOL(sleep_on_timeout
);
6079 #ifdef CONFIG_RT_MUTEXES
6082 * rt_mutex_setprio - set the current priority of a task
6084 * @prio: prio value (kernel-internal form)
6086 * This function changes the 'effective' priority of a task. It does
6087 * not touch ->normal_prio like __setscheduler().
6089 * Used by the rt_mutex code to implement priority inheritance logic.
6091 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6093 unsigned long flags
;
6094 int oldprio
, on_rq
, running
;
6096 const struct sched_class
*prev_class
= p
->sched_class
;
6098 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6100 rq
= task_rq_lock(p
, &flags
);
6101 update_rq_clock(rq
);
6104 on_rq
= p
->se
.on_rq
;
6105 running
= task_current(rq
, p
);
6107 dequeue_task(rq
, p
, 0);
6109 p
->sched_class
->put_prev_task(rq
, p
);
6112 p
->sched_class
= &rt_sched_class
;
6114 p
->sched_class
= &fair_sched_class
;
6119 p
->sched_class
->set_curr_task(rq
);
6121 enqueue_task(rq
, p
, 0);
6123 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6125 task_rq_unlock(rq
, &flags
);
6130 void set_user_nice(struct task_struct
*p
, long nice
)
6132 int old_prio
, delta
, on_rq
;
6133 unsigned long flags
;
6136 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6139 * We have to be careful, if called from sys_setpriority(),
6140 * the task might be in the middle of scheduling on another CPU.
6142 rq
= task_rq_lock(p
, &flags
);
6143 update_rq_clock(rq
);
6145 * The RT priorities are set via sched_setscheduler(), but we still
6146 * allow the 'normal' nice value to be set - but as expected
6147 * it wont have any effect on scheduling until the task is
6148 * SCHED_FIFO/SCHED_RR:
6150 if (task_has_rt_policy(p
)) {
6151 p
->static_prio
= NICE_TO_PRIO(nice
);
6154 on_rq
= p
->se
.on_rq
;
6156 dequeue_task(rq
, p
, 0);
6158 p
->static_prio
= NICE_TO_PRIO(nice
);
6161 p
->prio
= effective_prio(p
);
6162 delta
= p
->prio
- old_prio
;
6165 enqueue_task(rq
, p
, 0);
6167 * If the task increased its priority or is running and
6168 * lowered its priority, then reschedule its CPU:
6170 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6171 resched_task(rq
->curr
);
6174 task_rq_unlock(rq
, &flags
);
6176 EXPORT_SYMBOL(set_user_nice
);
6179 * can_nice - check if a task can reduce its nice value
6183 int can_nice(const struct task_struct
*p
, const int nice
)
6185 /* convert nice value [19,-20] to rlimit style value [1,40] */
6186 int nice_rlim
= 20 - nice
;
6188 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6189 capable(CAP_SYS_NICE
));
6192 #ifdef __ARCH_WANT_SYS_NICE
6195 * sys_nice - change the priority of the current process.
6196 * @increment: priority increment
6198 * sys_setpriority is a more generic, but much slower function that
6199 * does similar things.
6201 SYSCALL_DEFINE1(nice
, int, increment
)
6206 * Setpriority might change our priority at the same moment.
6207 * We don't have to worry. Conceptually one call occurs first
6208 * and we have a single winner.
6210 if (increment
< -40)
6215 nice
= TASK_NICE(current
) + increment
;
6221 if (increment
< 0 && !can_nice(current
, nice
))
6224 retval
= security_task_setnice(current
, nice
);
6228 set_user_nice(current
, nice
);
6235 * task_prio - return the priority value of a given task.
6236 * @p: the task in question.
6238 * This is the priority value as seen by users in /proc.
6239 * RT tasks are offset by -200. Normal tasks are centered
6240 * around 0, value goes from -16 to +15.
6242 int task_prio(const struct task_struct
*p
)
6244 return p
->prio
- MAX_RT_PRIO
;
6248 * task_nice - return the nice value of a given task.
6249 * @p: the task in question.
6251 int task_nice(const struct task_struct
*p
)
6253 return TASK_NICE(p
);
6255 EXPORT_SYMBOL(task_nice
);
6258 * idle_cpu - is a given cpu idle currently?
6259 * @cpu: the processor in question.
6261 int idle_cpu(int cpu
)
6263 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6267 * idle_task - return the idle task for a given cpu.
6268 * @cpu: the processor in question.
6270 struct task_struct
*idle_task(int cpu
)
6272 return cpu_rq(cpu
)->idle
;
6276 * find_process_by_pid - find a process with a matching PID value.
6277 * @pid: the pid in question.
6279 static struct task_struct
*find_process_by_pid(pid_t pid
)
6281 return pid
? find_task_by_vpid(pid
) : current
;
6284 /* Actually do priority change: must hold rq lock. */
6286 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6288 BUG_ON(p
->se
.on_rq
);
6291 switch (p
->policy
) {
6295 p
->sched_class
= &fair_sched_class
;
6299 p
->sched_class
= &rt_sched_class
;
6303 p
->rt_priority
= prio
;
6304 p
->normal_prio
= normal_prio(p
);
6305 /* we are holding p->pi_lock already */
6306 p
->prio
= rt_mutex_getprio(p
);
6311 * check the target process has a UID that matches the current process's
6313 static bool check_same_owner(struct task_struct
*p
)
6315 const struct cred
*cred
= current_cred(), *pcred
;
6319 pcred
= __task_cred(p
);
6320 match
= (cred
->euid
== pcred
->euid
||
6321 cred
->euid
== pcred
->uid
);
6326 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6327 struct sched_param
*param
, bool user
)
6329 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6330 unsigned long flags
;
6331 const struct sched_class
*prev_class
= p
->sched_class
;
6335 /* may grab non-irq protected spin_locks */
6336 BUG_ON(in_interrupt());
6338 /* double check policy once rq lock held */
6340 reset_on_fork
= p
->sched_reset_on_fork
;
6341 policy
= oldpolicy
= p
->policy
;
6343 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6344 policy
&= ~SCHED_RESET_ON_FORK
;
6346 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6347 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6348 policy
!= SCHED_IDLE
)
6353 * Valid priorities for SCHED_FIFO and SCHED_RR are
6354 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6355 * SCHED_BATCH and SCHED_IDLE is 0.
6357 if (param
->sched_priority
< 0 ||
6358 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6359 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6361 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6365 * Allow unprivileged RT tasks to decrease priority:
6367 if (user
&& !capable(CAP_SYS_NICE
)) {
6368 if (rt_policy(policy
)) {
6369 unsigned long rlim_rtprio
;
6371 if (!lock_task_sighand(p
, &flags
))
6373 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6374 unlock_task_sighand(p
, &flags
);
6376 /* can't set/change the rt policy */
6377 if (policy
!= p
->policy
&& !rlim_rtprio
)
6380 /* can't increase priority */
6381 if (param
->sched_priority
> p
->rt_priority
&&
6382 param
->sched_priority
> rlim_rtprio
)
6386 * Like positive nice levels, dont allow tasks to
6387 * move out of SCHED_IDLE either:
6389 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6392 /* can't change other user's priorities */
6393 if (!check_same_owner(p
))
6396 /* Normal users shall not reset the sched_reset_on_fork flag */
6397 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6402 #ifdef CONFIG_RT_GROUP_SCHED
6404 * Do not allow realtime tasks into groups that have no runtime
6407 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6408 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6412 retval
= security_task_setscheduler(p
, policy
, param
);
6418 * make sure no PI-waiters arrive (or leave) while we are
6419 * changing the priority of the task:
6421 spin_lock_irqsave(&p
->pi_lock
, flags
);
6423 * To be able to change p->policy safely, the apropriate
6424 * runqueue lock must be held.
6426 rq
= __task_rq_lock(p
);
6427 /* recheck policy now with rq lock held */
6428 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6429 policy
= oldpolicy
= -1;
6430 __task_rq_unlock(rq
);
6431 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6434 update_rq_clock(rq
);
6435 on_rq
= p
->se
.on_rq
;
6436 running
= task_current(rq
, p
);
6438 deactivate_task(rq
, p
, 0);
6440 p
->sched_class
->put_prev_task(rq
, p
);
6442 p
->sched_reset_on_fork
= reset_on_fork
;
6445 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6448 p
->sched_class
->set_curr_task(rq
);
6450 activate_task(rq
, p
, 0);
6452 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6454 __task_rq_unlock(rq
);
6455 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6457 rt_mutex_adjust_pi(p
);
6463 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6464 * @p: the task in question.
6465 * @policy: new policy.
6466 * @param: structure containing the new RT priority.
6468 * NOTE that the task may be already dead.
6470 int sched_setscheduler(struct task_struct
*p
, int policy
,
6471 struct sched_param
*param
)
6473 return __sched_setscheduler(p
, policy
, param
, true);
6475 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6478 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6479 * @p: the task in question.
6480 * @policy: new policy.
6481 * @param: structure containing the new RT priority.
6483 * Just like sched_setscheduler, only don't bother checking if the
6484 * current context has permission. For example, this is needed in
6485 * stop_machine(): we create temporary high priority worker threads,
6486 * but our caller might not have that capability.
6488 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6489 struct sched_param
*param
)
6491 return __sched_setscheduler(p
, policy
, param
, false);
6495 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6497 struct sched_param lparam
;
6498 struct task_struct
*p
;
6501 if (!param
|| pid
< 0)
6503 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6508 p
= find_process_by_pid(pid
);
6510 retval
= sched_setscheduler(p
, policy
, &lparam
);
6517 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6518 * @pid: the pid in question.
6519 * @policy: new policy.
6520 * @param: structure containing the new RT priority.
6522 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6523 struct sched_param __user
*, param
)
6525 /* negative values for policy are not valid */
6529 return do_sched_setscheduler(pid
, policy
, param
);
6533 * sys_sched_setparam - set/change the RT priority of a thread
6534 * @pid: the pid in question.
6535 * @param: structure containing the new RT priority.
6537 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6539 return do_sched_setscheduler(pid
, -1, param
);
6543 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6544 * @pid: the pid in question.
6546 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6548 struct task_struct
*p
;
6555 read_lock(&tasklist_lock
);
6556 p
= find_process_by_pid(pid
);
6558 retval
= security_task_getscheduler(p
);
6561 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6563 read_unlock(&tasklist_lock
);
6568 * sys_sched_getparam - get the RT priority of a thread
6569 * @pid: the pid in question.
6570 * @param: structure containing the RT priority.
6572 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6574 struct sched_param lp
;
6575 struct task_struct
*p
;
6578 if (!param
|| pid
< 0)
6581 read_lock(&tasklist_lock
);
6582 p
= find_process_by_pid(pid
);
6587 retval
= security_task_getscheduler(p
);
6591 lp
.sched_priority
= p
->rt_priority
;
6592 read_unlock(&tasklist_lock
);
6595 * This one might sleep, we cannot do it with a spinlock held ...
6597 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6602 read_unlock(&tasklist_lock
);
6606 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6608 cpumask_var_t cpus_allowed
, new_mask
;
6609 struct task_struct
*p
;
6613 read_lock(&tasklist_lock
);
6615 p
= find_process_by_pid(pid
);
6617 read_unlock(&tasklist_lock
);
6623 * It is not safe to call set_cpus_allowed with the
6624 * tasklist_lock held. We will bump the task_struct's
6625 * usage count and then drop tasklist_lock.
6628 read_unlock(&tasklist_lock
);
6630 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6634 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6636 goto out_free_cpus_allowed
;
6639 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6642 retval
= security_task_setscheduler(p
, 0, NULL
);
6646 cpuset_cpus_allowed(p
, cpus_allowed
);
6647 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6649 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6652 cpuset_cpus_allowed(p
, cpus_allowed
);
6653 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6655 * We must have raced with a concurrent cpuset
6656 * update. Just reset the cpus_allowed to the
6657 * cpuset's cpus_allowed
6659 cpumask_copy(new_mask
, cpus_allowed
);
6664 free_cpumask_var(new_mask
);
6665 out_free_cpus_allowed
:
6666 free_cpumask_var(cpus_allowed
);
6673 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6674 struct cpumask
*new_mask
)
6676 if (len
< cpumask_size())
6677 cpumask_clear(new_mask
);
6678 else if (len
> cpumask_size())
6679 len
= cpumask_size();
6681 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6685 * sys_sched_setaffinity - set the cpu affinity of a process
6686 * @pid: pid of the process
6687 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6688 * @user_mask_ptr: user-space pointer to the new cpu mask
6690 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6691 unsigned long __user
*, user_mask_ptr
)
6693 cpumask_var_t new_mask
;
6696 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6699 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6701 retval
= sched_setaffinity(pid
, new_mask
);
6702 free_cpumask_var(new_mask
);
6706 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6708 struct task_struct
*p
;
6712 read_lock(&tasklist_lock
);
6715 p
= find_process_by_pid(pid
);
6719 retval
= security_task_getscheduler(p
);
6723 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6726 read_unlock(&tasklist_lock
);
6733 * sys_sched_getaffinity - get the cpu affinity of a process
6734 * @pid: pid of the process
6735 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6736 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6738 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6739 unsigned long __user
*, user_mask_ptr
)
6744 if (len
< cpumask_size())
6747 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6750 ret
= sched_getaffinity(pid
, mask
);
6752 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6755 ret
= cpumask_size();
6757 free_cpumask_var(mask
);
6763 * sys_sched_yield - yield the current processor to other threads.
6765 * This function yields the current CPU to other tasks. If there are no
6766 * other threads running on this CPU then this function will return.
6768 SYSCALL_DEFINE0(sched_yield
)
6770 struct rq
*rq
= this_rq_lock();
6772 schedstat_inc(rq
, yld_count
);
6773 current
->sched_class
->yield_task(rq
);
6776 * Since we are going to call schedule() anyway, there's
6777 * no need to preempt or enable interrupts:
6779 __release(rq
->lock
);
6780 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6781 _raw_spin_unlock(&rq
->lock
);
6782 preempt_enable_no_resched();
6789 static inline int should_resched(void)
6791 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6794 static void __cond_resched(void)
6796 add_preempt_count(PREEMPT_ACTIVE
);
6798 sub_preempt_count(PREEMPT_ACTIVE
);
6801 int __sched
_cond_resched(void)
6803 if (should_resched()) {
6809 EXPORT_SYMBOL(_cond_resched
);
6812 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6813 * call schedule, and on return reacquire the lock.
6815 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6816 * operations here to prevent schedule() from being called twice (once via
6817 * spin_unlock(), once by hand).
6819 int __cond_resched_lock(spinlock_t
*lock
)
6821 int resched
= should_resched();
6824 if (spin_needbreak(lock
) || resched
) {
6835 EXPORT_SYMBOL(__cond_resched_lock
);
6837 int __sched
__cond_resched_softirq(void)
6839 BUG_ON(!in_softirq());
6841 if (should_resched()) {
6849 EXPORT_SYMBOL(__cond_resched_softirq
);
6852 * yield - yield the current processor to other threads.
6854 * This is a shortcut for kernel-space yielding - it marks the
6855 * thread runnable and calls sys_sched_yield().
6857 void __sched
yield(void)
6859 set_current_state(TASK_RUNNING
);
6862 EXPORT_SYMBOL(yield
);
6865 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6866 * that process accounting knows that this is a task in IO wait state.
6868 * But don't do that if it is a deliberate, throttling IO wait (this task
6869 * has set its backing_dev_info: the queue against which it should throttle)
6871 void __sched
io_schedule(void)
6873 struct rq
*rq
= raw_rq();
6875 delayacct_blkio_start();
6876 atomic_inc(&rq
->nr_iowait
);
6877 current
->in_iowait
= 1;
6879 current
->in_iowait
= 0;
6880 atomic_dec(&rq
->nr_iowait
);
6881 delayacct_blkio_end();
6883 EXPORT_SYMBOL(io_schedule
);
6885 long __sched
io_schedule_timeout(long timeout
)
6887 struct rq
*rq
= raw_rq();
6890 delayacct_blkio_start();
6891 atomic_inc(&rq
->nr_iowait
);
6892 current
->in_iowait
= 1;
6893 ret
= schedule_timeout(timeout
);
6894 current
->in_iowait
= 0;
6895 atomic_dec(&rq
->nr_iowait
);
6896 delayacct_blkio_end();
6901 * sys_sched_get_priority_max - return maximum RT priority.
6902 * @policy: scheduling class.
6904 * this syscall returns the maximum rt_priority that can be used
6905 * by a given scheduling class.
6907 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6914 ret
= MAX_USER_RT_PRIO
-1;
6926 * sys_sched_get_priority_min - return minimum RT priority.
6927 * @policy: scheduling class.
6929 * this syscall returns the minimum rt_priority that can be used
6930 * by a given scheduling class.
6932 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6950 * sys_sched_rr_get_interval - return the default timeslice of a process.
6951 * @pid: pid of the process.
6952 * @interval: userspace pointer to the timeslice value.
6954 * this syscall writes the default timeslice value of a given process
6955 * into the user-space timespec buffer. A value of '0' means infinity.
6957 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6958 struct timespec __user
*, interval
)
6960 struct task_struct
*p
;
6961 unsigned int time_slice
;
6969 read_lock(&tasklist_lock
);
6970 p
= find_process_by_pid(pid
);
6974 retval
= security_task_getscheduler(p
);
6979 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6980 * tasks that are on an otherwise idle runqueue:
6983 if (p
->policy
== SCHED_RR
) {
6984 time_slice
= DEF_TIMESLICE
;
6985 } else if (p
->policy
!= SCHED_FIFO
) {
6986 struct sched_entity
*se
= &p
->se
;
6987 unsigned long flags
;
6990 rq
= task_rq_lock(p
, &flags
);
6991 if (rq
->cfs
.load
.weight
)
6992 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6993 task_rq_unlock(rq
, &flags
);
6995 read_unlock(&tasklist_lock
);
6996 jiffies_to_timespec(time_slice
, &t
);
6997 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
7001 read_unlock(&tasklist_lock
);
7005 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
7007 void sched_show_task(struct task_struct
*p
)
7009 unsigned long free
= 0;
7012 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
7013 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
7014 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
7015 #if BITS_PER_LONG == 32
7016 if (state
== TASK_RUNNING
)
7017 printk(KERN_CONT
" running ");
7019 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
7021 if (state
== TASK_RUNNING
)
7022 printk(KERN_CONT
" running task ");
7024 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
7026 #ifdef CONFIG_DEBUG_STACK_USAGE
7027 free
= stack_not_used(p
);
7029 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
7030 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
7031 (unsigned long)task_thread_info(p
)->flags
);
7033 show_stack(p
, NULL
);
7036 void show_state_filter(unsigned long state_filter
)
7038 struct task_struct
*g
, *p
;
7040 #if BITS_PER_LONG == 32
7042 " task PC stack pid father\n");
7045 " task PC stack pid father\n");
7047 read_lock(&tasklist_lock
);
7048 do_each_thread(g
, p
) {
7050 * reset the NMI-timeout, listing all files on a slow
7051 * console might take alot of time:
7053 touch_nmi_watchdog();
7054 if (!state_filter
|| (p
->state
& state_filter
))
7056 } while_each_thread(g
, p
);
7058 touch_all_softlockup_watchdogs();
7060 #ifdef CONFIG_SCHED_DEBUG
7061 sysrq_sched_debug_show();
7063 read_unlock(&tasklist_lock
);
7065 * Only show locks if all tasks are dumped:
7067 if (state_filter
== -1)
7068 debug_show_all_locks();
7071 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
7073 idle
->sched_class
= &idle_sched_class
;
7077 * init_idle - set up an idle thread for a given CPU
7078 * @idle: task in question
7079 * @cpu: cpu the idle task belongs to
7081 * NOTE: this function does not set the idle thread's NEED_RESCHED
7082 * flag, to make booting more robust.
7084 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7086 struct rq
*rq
= cpu_rq(cpu
);
7087 unsigned long flags
;
7089 spin_lock_irqsave(&rq
->lock
, flags
);
7092 idle
->se
.exec_start
= sched_clock();
7094 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
7095 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7096 __set_task_cpu(idle
, cpu
);
7098 rq
->curr
= rq
->idle
= idle
;
7099 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7102 spin_unlock_irqrestore(&rq
->lock
, flags
);
7104 /* Set the preempt count _outside_ the spinlocks! */
7105 #if defined(CONFIG_PREEMPT)
7106 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7108 task_thread_info(idle
)->preempt_count
= 0;
7111 * The idle tasks have their own, simple scheduling class:
7113 idle
->sched_class
= &idle_sched_class
;
7114 ftrace_graph_init_task(idle
);
7118 * In a system that switches off the HZ timer nohz_cpu_mask
7119 * indicates which cpus entered this state. This is used
7120 * in the rcu update to wait only for active cpus. For system
7121 * which do not switch off the HZ timer nohz_cpu_mask should
7122 * always be CPU_BITS_NONE.
7124 cpumask_var_t nohz_cpu_mask
;
7127 * Increase the granularity value when there are more CPUs,
7128 * because with more CPUs the 'effective latency' as visible
7129 * to users decreases. But the relationship is not linear,
7130 * so pick a second-best guess by going with the log2 of the
7133 * This idea comes from the SD scheduler of Con Kolivas:
7135 static inline void sched_init_granularity(void)
7137 unsigned int factor
= 1 + ilog2(num_online_cpus());
7138 const unsigned long limit
= 200000000;
7140 sysctl_sched_min_granularity
*= factor
;
7141 if (sysctl_sched_min_granularity
> limit
)
7142 sysctl_sched_min_granularity
= limit
;
7144 sysctl_sched_latency
*= factor
;
7145 if (sysctl_sched_latency
> limit
)
7146 sysctl_sched_latency
= limit
;
7148 sysctl_sched_wakeup_granularity
*= factor
;
7150 sysctl_sched_shares_ratelimit
*= factor
;
7155 * This is how migration works:
7157 * 1) we queue a struct migration_req structure in the source CPU's
7158 * runqueue and wake up that CPU's migration thread.
7159 * 2) we down() the locked semaphore => thread blocks.
7160 * 3) migration thread wakes up (implicitly it forces the migrated
7161 * thread off the CPU)
7162 * 4) it gets the migration request and checks whether the migrated
7163 * task is still in the wrong runqueue.
7164 * 5) if it's in the wrong runqueue then the migration thread removes
7165 * it and puts it into the right queue.
7166 * 6) migration thread up()s the semaphore.
7167 * 7) we wake up and the migration is done.
7171 * Change a given task's CPU affinity. Migrate the thread to a
7172 * proper CPU and schedule it away if the CPU it's executing on
7173 * is removed from the allowed bitmask.
7175 * NOTE: the caller must have a valid reference to the task, the
7176 * task must not exit() & deallocate itself prematurely. The
7177 * call is not atomic; no spinlocks may be held.
7179 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7181 struct migration_req req
;
7182 unsigned long flags
;
7186 rq
= task_rq_lock(p
, &flags
);
7187 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7192 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7193 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7198 if (p
->sched_class
->set_cpus_allowed
)
7199 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7201 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7202 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7205 /* Can the task run on the task's current CPU? If so, we're done */
7206 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7209 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7210 /* Need help from migration thread: drop lock and wait. */
7211 struct task_struct
*mt
= rq
->migration_thread
;
7213 get_task_struct(mt
);
7214 task_rq_unlock(rq
, &flags
);
7215 wake_up_process(rq
->migration_thread
);
7216 put_task_struct(mt
);
7217 wait_for_completion(&req
.done
);
7218 tlb_migrate_finish(p
->mm
);
7222 task_rq_unlock(rq
, &flags
);
7226 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7229 * Move (not current) task off this cpu, onto dest cpu. We're doing
7230 * this because either it can't run here any more (set_cpus_allowed()
7231 * away from this CPU, or CPU going down), or because we're
7232 * attempting to rebalance this task on exec (sched_exec).
7234 * So we race with normal scheduler movements, but that's OK, as long
7235 * as the task is no longer on this CPU.
7237 * Returns non-zero if task was successfully migrated.
7239 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7241 struct rq
*rq_dest
, *rq_src
;
7244 if (unlikely(!cpu_active(dest_cpu
)))
7247 rq_src
= cpu_rq(src_cpu
);
7248 rq_dest
= cpu_rq(dest_cpu
);
7250 double_rq_lock(rq_src
, rq_dest
);
7251 /* Already moved. */
7252 if (task_cpu(p
) != src_cpu
)
7254 /* Affinity changed (again). */
7255 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7258 on_rq
= p
->se
.on_rq
;
7260 deactivate_task(rq_src
, p
, 0);
7262 set_task_cpu(p
, dest_cpu
);
7264 activate_task(rq_dest
, p
, 0);
7265 check_preempt_curr(rq_dest
, p
, 0);
7270 double_rq_unlock(rq_src
, rq_dest
);
7275 * migration_thread - this is a highprio system thread that performs
7276 * thread migration by bumping thread off CPU then 'pushing' onto
7279 static int migration_thread(void *data
)
7281 int cpu
= (long)data
;
7285 BUG_ON(rq
->migration_thread
!= current
);
7287 set_current_state(TASK_INTERRUPTIBLE
);
7288 while (!kthread_should_stop()) {
7289 struct migration_req
*req
;
7290 struct list_head
*head
;
7292 spin_lock_irq(&rq
->lock
);
7294 if (cpu_is_offline(cpu
)) {
7295 spin_unlock_irq(&rq
->lock
);
7299 if (rq
->active_balance
) {
7300 active_load_balance(rq
, cpu
);
7301 rq
->active_balance
= 0;
7304 head
= &rq
->migration_queue
;
7306 if (list_empty(head
)) {
7307 spin_unlock_irq(&rq
->lock
);
7309 set_current_state(TASK_INTERRUPTIBLE
);
7312 req
= list_entry(head
->next
, struct migration_req
, list
);
7313 list_del_init(head
->next
);
7315 spin_unlock(&rq
->lock
);
7316 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7319 complete(&req
->done
);
7321 __set_current_state(TASK_RUNNING
);
7326 #ifdef CONFIG_HOTPLUG_CPU
7328 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7332 local_irq_disable();
7333 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7339 * Figure out where task on dead CPU should go, use force if necessary.
7341 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7344 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7347 /* Look for allowed, online CPU in same node. */
7348 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7349 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7352 /* Any allowed, online CPU? */
7353 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7354 if (dest_cpu
< nr_cpu_ids
)
7357 /* No more Mr. Nice Guy. */
7358 if (dest_cpu
>= nr_cpu_ids
) {
7359 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7360 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7363 * Don't tell them about moving exiting tasks or
7364 * kernel threads (both mm NULL), since they never
7367 if (p
->mm
&& printk_ratelimit()) {
7368 printk(KERN_INFO
"process %d (%s) no "
7369 "longer affine to cpu%d\n",
7370 task_pid_nr(p
), p
->comm
, dead_cpu
);
7375 /* It can have affinity changed while we were choosing. */
7376 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7381 * While a dead CPU has no uninterruptible tasks queued at this point,
7382 * it might still have a nonzero ->nr_uninterruptible counter, because
7383 * for performance reasons the counter is not stricly tracking tasks to
7384 * their home CPUs. So we just add the counter to another CPU's counter,
7385 * to keep the global sum constant after CPU-down:
7387 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7389 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7390 unsigned long flags
;
7392 local_irq_save(flags
);
7393 double_rq_lock(rq_src
, rq_dest
);
7394 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7395 rq_src
->nr_uninterruptible
= 0;
7396 double_rq_unlock(rq_src
, rq_dest
);
7397 local_irq_restore(flags
);
7400 /* Run through task list and migrate tasks from the dead cpu. */
7401 static void migrate_live_tasks(int src_cpu
)
7403 struct task_struct
*p
, *t
;
7405 read_lock(&tasklist_lock
);
7407 do_each_thread(t
, p
) {
7411 if (task_cpu(p
) == src_cpu
)
7412 move_task_off_dead_cpu(src_cpu
, p
);
7413 } while_each_thread(t
, p
);
7415 read_unlock(&tasklist_lock
);
7419 * Schedules idle task to be the next runnable task on current CPU.
7420 * It does so by boosting its priority to highest possible.
7421 * Used by CPU offline code.
7423 void sched_idle_next(void)
7425 int this_cpu
= smp_processor_id();
7426 struct rq
*rq
= cpu_rq(this_cpu
);
7427 struct task_struct
*p
= rq
->idle
;
7428 unsigned long flags
;
7430 /* cpu has to be offline */
7431 BUG_ON(cpu_online(this_cpu
));
7434 * Strictly not necessary since rest of the CPUs are stopped by now
7435 * and interrupts disabled on the current cpu.
7437 spin_lock_irqsave(&rq
->lock
, flags
);
7439 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7441 update_rq_clock(rq
);
7442 activate_task(rq
, p
, 0);
7444 spin_unlock_irqrestore(&rq
->lock
, flags
);
7448 * Ensures that the idle task is using init_mm right before its cpu goes
7451 void idle_task_exit(void)
7453 struct mm_struct
*mm
= current
->active_mm
;
7455 BUG_ON(cpu_online(smp_processor_id()));
7458 switch_mm(mm
, &init_mm
, current
);
7462 /* called under rq->lock with disabled interrupts */
7463 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7465 struct rq
*rq
= cpu_rq(dead_cpu
);
7467 /* Must be exiting, otherwise would be on tasklist. */
7468 BUG_ON(!p
->exit_state
);
7470 /* Cannot have done final schedule yet: would have vanished. */
7471 BUG_ON(p
->state
== TASK_DEAD
);
7476 * Drop lock around migration; if someone else moves it,
7477 * that's OK. No task can be added to this CPU, so iteration is
7480 spin_unlock_irq(&rq
->lock
);
7481 move_task_off_dead_cpu(dead_cpu
, p
);
7482 spin_lock_irq(&rq
->lock
);
7487 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7488 static void migrate_dead_tasks(unsigned int dead_cpu
)
7490 struct rq
*rq
= cpu_rq(dead_cpu
);
7491 struct task_struct
*next
;
7494 if (!rq
->nr_running
)
7496 update_rq_clock(rq
);
7497 next
= pick_next_task(rq
);
7500 next
->sched_class
->put_prev_task(rq
, next
);
7501 migrate_dead(dead_cpu
, next
);
7507 * remove the tasks which were accounted by rq from calc_load_tasks.
7509 static void calc_global_load_remove(struct rq
*rq
)
7511 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7512 rq
->calc_load_active
= 0;
7514 #endif /* CONFIG_HOTPLUG_CPU */
7516 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7518 static struct ctl_table sd_ctl_dir
[] = {
7520 .procname
= "sched_domain",
7526 static struct ctl_table sd_ctl_root
[] = {
7528 .ctl_name
= CTL_KERN
,
7529 .procname
= "kernel",
7531 .child
= sd_ctl_dir
,
7536 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7538 struct ctl_table
*entry
=
7539 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7544 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7546 struct ctl_table
*entry
;
7549 * In the intermediate directories, both the child directory and
7550 * procname are dynamically allocated and could fail but the mode
7551 * will always be set. In the lowest directory the names are
7552 * static strings and all have proc handlers.
7554 for (entry
= *tablep
; entry
->mode
; entry
++) {
7556 sd_free_ctl_entry(&entry
->child
);
7557 if (entry
->proc_handler
== NULL
)
7558 kfree(entry
->procname
);
7566 set_table_entry(struct ctl_table
*entry
,
7567 const char *procname
, void *data
, int maxlen
,
7568 mode_t mode
, proc_handler
*proc_handler
)
7570 entry
->procname
= procname
;
7572 entry
->maxlen
= maxlen
;
7574 entry
->proc_handler
= proc_handler
;
7577 static struct ctl_table
*
7578 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7580 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7585 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7586 sizeof(long), 0644, proc_doulongvec_minmax
);
7587 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7588 sizeof(long), 0644, proc_doulongvec_minmax
);
7589 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7590 sizeof(int), 0644, proc_dointvec_minmax
);
7591 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7592 sizeof(int), 0644, proc_dointvec_minmax
);
7593 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7594 sizeof(int), 0644, proc_dointvec_minmax
);
7595 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7596 sizeof(int), 0644, proc_dointvec_minmax
);
7597 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7598 sizeof(int), 0644, proc_dointvec_minmax
);
7599 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7600 sizeof(int), 0644, proc_dointvec_minmax
);
7601 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7602 sizeof(int), 0644, proc_dointvec_minmax
);
7603 set_table_entry(&table
[9], "cache_nice_tries",
7604 &sd
->cache_nice_tries
,
7605 sizeof(int), 0644, proc_dointvec_minmax
);
7606 set_table_entry(&table
[10], "flags", &sd
->flags
,
7607 sizeof(int), 0644, proc_dointvec_minmax
);
7608 set_table_entry(&table
[11], "name", sd
->name
,
7609 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7610 /* &table[12] is terminator */
7615 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7617 struct ctl_table
*entry
, *table
;
7618 struct sched_domain
*sd
;
7619 int domain_num
= 0, i
;
7622 for_each_domain(cpu
, sd
)
7624 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7629 for_each_domain(cpu
, sd
) {
7630 snprintf(buf
, 32, "domain%d", i
);
7631 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7633 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7640 static struct ctl_table_header
*sd_sysctl_header
;
7641 static void register_sched_domain_sysctl(void)
7643 int i
, cpu_num
= num_online_cpus();
7644 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7647 WARN_ON(sd_ctl_dir
[0].child
);
7648 sd_ctl_dir
[0].child
= entry
;
7653 for_each_online_cpu(i
) {
7654 snprintf(buf
, 32, "cpu%d", i
);
7655 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7657 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7661 WARN_ON(sd_sysctl_header
);
7662 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7665 /* may be called multiple times per register */
7666 static void unregister_sched_domain_sysctl(void)
7668 if (sd_sysctl_header
)
7669 unregister_sysctl_table(sd_sysctl_header
);
7670 sd_sysctl_header
= NULL
;
7671 if (sd_ctl_dir
[0].child
)
7672 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7675 static void register_sched_domain_sysctl(void)
7678 static void unregister_sched_domain_sysctl(void)
7683 static void set_rq_online(struct rq
*rq
)
7686 const struct sched_class
*class;
7688 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7691 for_each_class(class) {
7692 if (class->rq_online
)
7693 class->rq_online(rq
);
7698 static void set_rq_offline(struct rq
*rq
)
7701 const struct sched_class
*class;
7703 for_each_class(class) {
7704 if (class->rq_offline
)
7705 class->rq_offline(rq
);
7708 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7714 * migration_call - callback that gets triggered when a CPU is added.
7715 * Here we can start up the necessary migration thread for the new CPU.
7717 static int __cpuinit
7718 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7720 struct task_struct
*p
;
7721 int cpu
= (long)hcpu
;
7722 unsigned long flags
;
7727 case CPU_UP_PREPARE
:
7728 case CPU_UP_PREPARE_FROZEN
:
7729 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7732 kthread_bind(p
, cpu
);
7733 /* Must be high prio: stop_machine expects to yield to it. */
7734 rq
= task_rq_lock(p
, &flags
);
7735 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7736 task_rq_unlock(rq
, &flags
);
7738 cpu_rq(cpu
)->migration_thread
= p
;
7739 rq
->calc_load_update
= calc_load_update
;
7743 case CPU_ONLINE_FROZEN
:
7744 /* Strictly unnecessary, as first user will wake it. */
7745 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7747 /* Update our root-domain */
7749 spin_lock_irqsave(&rq
->lock
, flags
);
7751 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7755 spin_unlock_irqrestore(&rq
->lock
, flags
);
7758 #ifdef CONFIG_HOTPLUG_CPU
7759 case CPU_UP_CANCELED
:
7760 case CPU_UP_CANCELED_FROZEN
:
7761 if (!cpu_rq(cpu
)->migration_thread
)
7763 /* Unbind it from offline cpu so it can run. Fall thru. */
7764 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7765 cpumask_any(cpu_online_mask
));
7766 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7767 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7768 cpu_rq(cpu
)->migration_thread
= NULL
;
7772 case CPU_DEAD_FROZEN
:
7773 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7774 migrate_live_tasks(cpu
);
7776 kthread_stop(rq
->migration_thread
);
7777 put_task_struct(rq
->migration_thread
);
7778 rq
->migration_thread
= NULL
;
7779 /* Idle task back to normal (off runqueue, low prio) */
7780 spin_lock_irq(&rq
->lock
);
7781 update_rq_clock(rq
);
7782 deactivate_task(rq
, rq
->idle
, 0);
7783 rq
->idle
->static_prio
= MAX_PRIO
;
7784 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7785 rq
->idle
->sched_class
= &idle_sched_class
;
7786 migrate_dead_tasks(cpu
);
7787 spin_unlock_irq(&rq
->lock
);
7789 migrate_nr_uninterruptible(rq
);
7790 BUG_ON(rq
->nr_running
!= 0);
7791 calc_global_load_remove(rq
);
7793 * No need to migrate the tasks: it was best-effort if
7794 * they didn't take sched_hotcpu_mutex. Just wake up
7797 spin_lock_irq(&rq
->lock
);
7798 while (!list_empty(&rq
->migration_queue
)) {
7799 struct migration_req
*req
;
7801 req
= list_entry(rq
->migration_queue
.next
,
7802 struct migration_req
, list
);
7803 list_del_init(&req
->list
);
7804 spin_unlock_irq(&rq
->lock
);
7805 complete(&req
->done
);
7806 spin_lock_irq(&rq
->lock
);
7808 spin_unlock_irq(&rq
->lock
);
7812 case CPU_DYING_FROZEN
:
7813 /* Update our root-domain */
7815 spin_lock_irqsave(&rq
->lock
, flags
);
7817 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7820 spin_unlock_irqrestore(&rq
->lock
, flags
);
7828 * Register at high priority so that task migration (migrate_all_tasks)
7829 * happens before everything else. This has to be lower priority than
7830 * the notifier in the perf_counter subsystem, though.
7832 static struct notifier_block __cpuinitdata migration_notifier
= {
7833 .notifier_call
= migration_call
,
7837 static int __init
migration_init(void)
7839 void *cpu
= (void *)(long)smp_processor_id();
7842 /* Start one for the boot CPU: */
7843 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7844 BUG_ON(err
== NOTIFY_BAD
);
7845 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7846 register_cpu_notifier(&migration_notifier
);
7850 early_initcall(migration_init
);
7855 #ifdef CONFIG_SCHED_DEBUG
7857 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7858 struct cpumask
*groupmask
)
7860 struct sched_group
*group
= sd
->groups
;
7863 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7864 cpumask_clear(groupmask
);
7866 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7868 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7869 printk("does not load-balance\n");
7871 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7876 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7878 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7879 printk(KERN_ERR
"ERROR: domain->span does not contain "
7882 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7883 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7887 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7891 printk(KERN_ERR
"ERROR: group is NULL\n");
7895 if (!group
->cpu_power
) {
7896 printk(KERN_CONT
"\n");
7897 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7902 if (!cpumask_weight(sched_group_cpus(group
))) {
7903 printk(KERN_CONT
"\n");
7904 printk(KERN_ERR
"ERROR: empty group\n");
7908 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7909 printk(KERN_CONT
"\n");
7910 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7914 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7916 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7918 printk(KERN_CONT
" %s", str
);
7919 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7920 printk(KERN_CONT
" (cpu_power = %d)",
7924 group
= group
->next
;
7925 } while (group
!= sd
->groups
);
7926 printk(KERN_CONT
"\n");
7928 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7929 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7932 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7933 printk(KERN_ERR
"ERROR: parent span is not a superset "
7934 "of domain->span\n");
7938 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7940 cpumask_var_t groupmask
;
7944 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7948 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7950 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7951 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7956 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7963 free_cpumask_var(groupmask
);
7965 #else /* !CONFIG_SCHED_DEBUG */
7966 # define sched_domain_debug(sd, cpu) do { } while (0)
7967 #endif /* CONFIG_SCHED_DEBUG */
7969 static int sd_degenerate(struct sched_domain
*sd
)
7971 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7974 /* Following flags need at least 2 groups */
7975 if (sd
->flags
& (SD_LOAD_BALANCE
|
7976 SD_BALANCE_NEWIDLE
|
7980 SD_SHARE_PKG_RESOURCES
)) {
7981 if (sd
->groups
!= sd
->groups
->next
)
7985 /* Following flags don't use groups */
7986 if (sd
->flags
& (SD_WAKE_IDLE
|
7995 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7997 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7999 if (sd_degenerate(parent
))
8002 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
8005 /* Does parent contain flags not in child? */
8006 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
8007 if (cflags
& SD_WAKE_AFFINE
)
8008 pflags
&= ~SD_WAKE_BALANCE
;
8009 /* Flags needing groups don't count if only 1 group in parent */
8010 if (parent
->groups
== parent
->groups
->next
) {
8011 pflags
&= ~(SD_LOAD_BALANCE
|
8012 SD_BALANCE_NEWIDLE
|
8016 SD_SHARE_PKG_RESOURCES
);
8017 if (nr_node_ids
== 1)
8018 pflags
&= ~SD_SERIALIZE
;
8020 if (~cflags
& pflags
)
8026 static void free_rootdomain(struct root_domain
*rd
)
8028 cpupri_cleanup(&rd
->cpupri
);
8030 free_cpumask_var(rd
->rto_mask
);
8031 free_cpumask_var(rd
->online
);
8032 free_cpumask_var(rd
->span
);
8036 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8038 struct root_domain
*old_rd
= NULL
;
8039 unsigned long flags
;
8041 spin_lock_irqsave(&rq
->lock
, flags
);
8046 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8049 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8052 * If we dont want to free the old_rt yet then
8053 * set old_rd to NULL to skip the freeing later
8056 if (!atomic_dec_and_test(&old_rd
->refcount
))
8060 atomic_inc(&rd
->refcount
);
8063 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8064 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8067 spin_unlock_irqrestore(&rq
->lock
, flags
);
8070 free_rootdomain(old_rd
);
8073 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8075 gfp_t gfp
= GFP_KERNEL
;
8077 memset(rd
, 0, sizeof(*rd
));
8082 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8084 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8086 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8089 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8094 free_cpumask_var(rd
->rto_mask
);
8096 free_cpumask_var(rd
->online
);
8098 free_cpumask_var(rd
->span
);
8103 static void init_defrootdomain(void)
8105 init_rootdomain(&def_root_domain
, true);
8107 atomic_set(&def_root_domain
.refcount
, 1);
8110 static struct root_domain
*alloc_rootdomain(void)
8112 struct root_domain
*rd
;
8114 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8118 if (init_rootdomain(rd
, false) != 0) {
8127 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8128 * hold the hotplug lock.
8131 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8133 struct rq
*rq
= cpu_rq(cpu
);
8134 struct sched_domain
*tmp
;
8136 /* Remove the sched domains which do not contribute to scheduling. */
8137 for (tmp
= sd
; tmp
; ) {
8138 struct sched_domain
*parent
= tmp
->parent
;
8142 if (sd_parent_degenerate(tmp
, parent
)) {
8143 tmp
->parent
= parent
->parent
;
8145 parent
->parent
->child
= tmp
;
8150 if (sd
&& sd_degenerate(sd
)) {
8156 sched_domain_debug(sd
, cpu
);
8158 rq_attach_root(rq
, rd
);
8159 rcu_assign_pointer(rq
->sd
, sd
);
8162 /* cpus with isolated domains */
8163 static cpumask_var_t cpu_isolated_map
;
8165 /* Setup the mask of cpus configured for isolated domains */
8166 static int __init
isolated_cpu_setup(char *str
)
8168 cpulist_parse(str
, cpu_isolated_map
);
8172 __setup("isolcpus=", isolated_cpu_setup
);
8175 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8176 * to a function which identifies what group(along with sched group) a CPU
8177 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8178 * (due to the fact that we keep track of groups covered with a struct cpumask).
8180 * init_sched_build_groups will build a circular linked list of the groups
8181 * covered by the given span, and will set each group's ->cpumask correctly,
8182 * and ->cpu_power to 0.
8185 init_sched_build_groups(const struct cpumask
*span
,
8186 const struct cpumask
*cpu_map
,
8187 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8188 struct sched_group
**sg
,
8189 struct cpumask
*tmpmask
),
8190 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8192 struct sched_group
*first
= NULL
, *last
= NULL
;
8195 cpumask_clear(covered
);
8197 for_each_cpu(i
, span
) {
8198 struct sched_group
*sg
;
8199 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8202 if (cpumask_test_cpu(i
, covered
))
8205 cpumask_clear(sched_group_cpus(sg
));
8208 for_each_cpu(j
, span
) {
8209 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8212 cpumask_set_cpu(j
, covered
);
8213 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8224 #define SD_NODES_PER_DOMAIN 16
8229 * find_next_best_node - find the next node to include in a sched_domain
8230 * @node: node whose sched_domain we're building
8231 * @used_nodes: nodes already in the sched_domain
8233 * Find the next node to include in a given scheduling domain. Simply
8234 * finds the closest node not already in the @used_nodes map.
8236 * Should use nodemask_t.
8238 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8240 int i
, n
, val
, min_val
, best_node
= 0;
8244 for (i
= 0; i
< nr_node_ids
; i
++) {
8245 /* Start at @node */
8246 n
= (node
+ i
) % nr_node_ids
;
8248 if (!nr_cpus_node(n
))
8251 /* Skip already used nodes */
8252 if (node_isset(n
, *used_nodes
))
8255 /* Simple min distance search */
8256 val
= node_distance(node
, n
);
8258 if (val
< min_val
) {
8264 node_set(best_node
, *used_nodes
);
8269 * sched_domain_node_span - get a cpumask for a node's sched_domain
8270 * @node: node whose cpumask we're constructing
8271 * @span: resulting cpumask
8273 * Given a node, construct a good cpumask for its sched_domain to span. It
8274 * should be one that prevents unnecessary balancing, but also spreads tasks
8277 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8279 nodemask_t used_nodes
;
8282 cpumask_clear(span
);
8283 nodes_clear(used_nodes
);
8285 cpumask_or(span
, span
, cpumask_of_node(node
));
8286 node_set(node
, used_nodes
);
8288 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8289 int next_node
= find_next_best_node(node
, &used_nodes
);
8291 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8294 #endif /* CONFIG_NUMA */
8296 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8299 * The cpus mask in sched_group and sched_domain hangs off the end.
8301 * ( See the the comments in include/linux/sched.h:struct sched_group
8302 * and struct sched_domain. )
8304 struct static_sched_group
{
8305 struct sched_group sg
;
8306 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8309 struct static_sched_domain
{
8310 struct sched_domain sd
;
8311 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8317 cpumask_var_t domainspan
;
8318 cpumask_var_t covered
;
8319 cpumask_var_t notcovered
;
8321 cpumask_var_t nodemask
;
8322 cpumask_var_t this_sibling_map
;
8323 cpumask_var_t this_core_map
;
8324 cpumask_var_t send_covered
;
8325 cpumask_var_t tmpmask
;
8326 struct sched_group
**sched_group_nodes
;
8327 struct root_domain
*rd
;
8331 sa_sched_groups
= 0,
8336 sa_this_sibling_map
,
8338 sa_sched_group_nodes
,
8348 * SMT sched-domains:
8350 #ifdef CONFIG_SCHED_SMT
8351 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8352 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8355 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8356 struct sched_group
**sg
, struct cpumask
*unused
)
8359 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8362 #endif /* CONFIG_SCHED_SMT */
8365 * multi-core sched-domains:
8367 #ifdef CONFIG_SCHED_MC
8368 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8369 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8370 #endif /* CONFIG_SCHED_MC */
8372 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8374 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8375 struct sched_group
**sg
, struct cpumask
*mask
)
8379 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8380 group
= cpumask_first(mask
);
8382 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8385 #elif defined(CONFIG_SCHED_MC)
8387 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8388 struct sched_group
**sg
, struct cpumask
*unused
)
8391 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8396 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8397 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8400 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8401 struct sched_group
**sg
, struct cpumask
*mask
)
8404 #ifdef CONFIG_SCHED_MC
8405 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8406 group
= cpumask_first(mask
);
8407 #elif defined(CONFIG_SCHED_SMT)
8408 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8409 group
= cpumask_first(mask
);
8414 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8420 * The init_sched_build_groups can't handle what we want to do with node
8421 * groups, so roll our own. Now each node has its own list of groups which
8422 * gets dynamically allocated.
8424 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8425 static struct sched_group
***sched_group_nodes_bycpu
;
8427 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8428 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8430 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8431 struct sched_group
**sg
,
8432 struct cpumask
*nodemask
)
8436 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8437 group
= cpumask_first(nodemask
);
8440 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8444 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8446 struct sched_group
*sg
= group_head
;
8452 for_each_cpu(j
, sched_group_cpus(sg
)) {
8453 struct sched_domain
*sd
;
8455 sd
= &per_cpu(phys_domains
, j
).sd
;
8456 if (j
!= group_first_cpu(sd
->groups
)) {
8458 * Only add "power" once for each
8464 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8467 } while (sg
!= group_head
);
8470 static int build_numa_sched_groups(struct s_data
*d
,
8471 const struct cpumask
*cpu_map
, int num
)
8473 struct sched_domain
*sd
;
8474 struct sched_group
*sg
, *prev
;
8477 cpumask_clear(d
->covered
);
8478 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8479 if (cpumask_empty(d
->nodemask
)) {
8480 d
->sched_group_nodes
[num
] = NULL
;
8484 sched_domain_node_span(num
, d
->domainspan
);
8485 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8487 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8490 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
8494 d
->sched_group_nodes
[num
] = sg
;
8496 for_each_cpu(j
, d
->nodemask
) {
8497 sd
= &per_cpu(node_domains
, j
).sd
;
8502 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8504 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8507 for (j
= 0; j
< nr_node_ids
; j
++) {
8508 n
= (num
+ j
) % nr_node_ids
;
8509 cpumask_complement(d
->notcovered
, d
->covered
);
8510 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8511 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8512 if (cpumask_empty(d
->tmpmask
))
8514 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8515 if (cpumask_empty(d
->tmpmask
))
8517 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8521 "Can not alloc domain group for node %d\n", j
);
8525 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8526 sg
->next
= prev
->next
;
8527 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8534 #endif /* CONFIG_NUMA */
8537 /* Free memory allocated for various sched_group structures */
8538 static void free_sched_groups(const struct cpumask
*cpu_map
,
8539 struct cpumask
*nodemask
)
8543 for_each_cpu(cpu
, cpu_map
) {
8544 struct sched_group
**sched_group_nodes
8545 = sched_group_nodes_bycpu
[cpu
];
8547 if (!sched_group_nodes
)
8550 for (i
= 0; i
< nr_node_ids
; i
++) {
8551 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8553 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8554 if (cpumask_empty(nodemask
))
8564 if (oldsg
!= sched_group_nodes
[i
])
8567 kfree(sched_group_nodes
);
8568 sched_group_nodes_bycpu
[cpu
] = NULL
;
8571 #else /* !CONFIG_NUMA */
8572 static void free_sched_groups(const struct cpumask
*cpu_map
,
8573 struct cpumask
*nodemask
)
8576 #endif /* CONFIG_NUMA */
8579 * Initialize sched groups cpu_power.
8581 * cpu_power indicates the capacity of sched group, which is used while
8582 * distributing the load between different sched groups in a sched domain.
8583 * Typically cpu_power for all the groups in a sched domain will be same unless
8584 * there are asymmetries in the topology. If there are asymmetries, group
8585 * having more cpu_power will pickup more load compared to the group having
8588 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8590 struct sched_domain
*child
;
8591 struct sched_group
*group
;
8595 WARN_ON(!sd
|| !sd
->groups
);
8597 if (cpu
!= group_first_cpu(sd
->groups
))
8602 sd
->groups
->cpu_power
= 0;
8605 power
= SCHED_LOAD_SCALE
;
8606 weight
= cpumask_weight(sched_domain_span(sd
));
8608 * SMT siblings share the power of a single core.
8609 * Usually multiple threads get a better yield out of
8610 * that one core than a single thread would have,
8611 * reflect that in sd->smt_gain.
8613 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8614 power
*= sd
->smt_gain
;
8616 power
>>= SCHED_LOAD_SHIFT
;
8618 sd
->groups
->cpu_power
+= power
;
8623 * Add cpu_power of each child group to this groups cpu_power.
8625 group
= child
->groups
;
8627 sd
->groups
->cpu_power
+= group
->cpu_power
;
8628 group
= group
->next
;
8629 } while (group
!= child
->groups
);
8633 * Initializers for schedule domains
8634 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8637 #ifdef CONFIG_SCHED_DEBUG
8638 # define SD_INIT_NAME(sd, type) sd->name = #type
8640 # define SD_INIT_NAME(sd, type) do { } while (0)
8643 #define SD_INIT(sd, type) sd_init_##type(sd)
8645 #define SD_INIT_FUNC(type) \
8646 static noinline void sd_init_##type(struct sched_domain *sd) \
8648 memset(sd, 0, sizeof(*sd)); \
8649 *sd = SD_##type##_INIT; \
8650 sd->level = SD_LV_##type; \
8651 SD_INIT_NAME(sd, type); \
8656 SD_INIT_FUNC(ALLNODES
)
8659 #ifdef CONFIG_SCHED_SMT
8660 SD_INIT_FUNC(SIBLING
)
8662 #ifdef CONFIG_SCHED_MC
8666 static int default_relax_domain_level
= -1;
8668 static int __init
setup_relax_domain_level(char *str
)
8672 val
= simple_strtoul(str
, NULL
, 0);
8673 if (val
< SD_LV_MAX
)
8674 default_relax_domain_level
= val
;
8678 __setup("relax_domain_level=", setup_relax_domain_level
);
8680 static void set_domain_attribute(struct sched_domain
*sd
,
8681 struct sched_domain_attr
*attr
)
8685 if (!attr
|| attr
->relax_domain_level
< 0) {
8686 if (default_relax_domain_level
< 0)
8689 request
= default_relax_domain_level
;
8691 request
= attr
->relax_domain_level
;
8692 if (request
< sd
->level
) {
8693 /* turn off idle balance on this domain */
8694 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8696 /* turn on idle balance on this domain */
8697 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8701 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8702 const struct cpumask
*cpu_map
)
8705 case sa_sched_groups
:
8706 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8707 d
->sched_group_nodes
= NULL
;
8709 free_rootdomain(d
->rd
); /* fall through */
8711 free_cpumask_var(d
->tmpmask
); /* fall through */
8712 case sa_send_covered
:
8713 free_cpumask_var(d
->send_covered
); /* fall through */
8714 case sa_this_core_map
:
8715 free_cpumask_var(d
->this_core_map
); /* fall through */
8716 case sa_this_sibling_map
:
8717 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8719 free_cpumask_var(d
->nodemask
); /* fall through */
8720 case sa_sched_group_nodes
:
8722 kfree(d
->sched_group_nodes
); /* fall through */
8724 free_cpumask_var(d
->notcovered
); /* fall through */
8726 free_cpumask_var(d
->covered
); /* fall through */
8728 free_cpumask_var(d
->domainspan
); /* fall through */
8735 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8736 const struct cpumask
*cpu_map
)
8739 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8741 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8742 return sa_domainspan
;
8743 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8745 /* Allocate the per-node list of sched groups */
8746 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8747 sizeof(struct sched_group
*), GFP_KERNEL
);
8748 if (!d
->sched_group_nodes
) {
8749 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8750 return sa_notcovered
;
8752 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8754 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8755 return sa_sched_group_nodes
;
8756 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8758 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8759 return sa_this_sibling_map
;
8760 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8761 return sa_this_core_map
;
8762 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8763 return sa_send_covered
;
8764 d
->rd
= alloc_rootdomain();
8766 printk(KERN_WARNING
"Cannot alloc root domain\n");
8769 return sa_rootdomain
;
8772 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8773 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8775 struct sched_domain
*sd
= NULL
;
8777 struct sched_domain
*parent
;
8780 if (cpumask_weight(cpu_map
) >
8781 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8782 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8783 SD_INIT(sd
, ALLNODES
);
8784 set_domain_attribute(sd
, attr
);
8785 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8786 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8791 sd
= &per_cpu(node_domains
, i
).sd
;
8793 set_domain_attribute(sd
, attr
);
8794 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8795 sd
->parent
= parent
;
8798 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8803 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8804 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8805 struct sched_domain
*parent
, int i
)
8807 struct sched_domain
*sd
;
8808 sd
= &per_cpu(phys_domains
, i
).sd
;
8810 set_domain_attribute(sd
, attr
);
8811 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8812 sd
->parent
= parent
;
8815 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8819 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8820 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8821 struct sched_domain
*parent
, int i
)
8823 struct sched_domain
*sd
= parent
;
8824 #ifdef CONFIG_SCHED_MC
8825 sd
= &per_cpu(core_domains
, i
).sd
;
8827 set_domain_attribute(sd
, attr
);
8828 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8829 sd
->parent
= parent
;
8831 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8836 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8837 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8838 struct sched_domain
*parent
, int i
)
8840 struct sched_domain
*sd
= parent
;
8841 #ifdef CONFIG_SCHED_SMT
8842 sd
= &per_cpu(cpu_domains
, i
).sd
;
8843 SD_INIT(sd
, SIBLING
);
8844 set_domain_attribute(sd
, attr
);
8845 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8846 sd
->parent
= parent
;
8848 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8853 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8854 const struct cpumask
*cpu_map
, int cpu
)
8857 #ifdef CONFIG_SCHED_SMT
8858 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8859 cpumask_and(d
->this_sibling_map
, cpu_map
,
8860 topology_thread_cpumask(cpu
));
8861 if (cpu
== cpumask_first(d
->this_sibling_map
))
8862 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8864 d
->send_covered
, d
->tmpmask
);
8867 #ifdef CONFIG_SCHED_MC
8868 case SD_LV_MC
: /* set up multi-core groups */
8869 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8870 if (cpu
== cpumask_first(d
->this_core_map
))
8871 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8873 d
->send_covered
, d
->tmpmask
);
8876 case SD_LV_CPU
: /* set up physical groups */
8877 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8878 if (!cpumask_empty(d
->nodemask
))
8879 init_sched_build_groups(d
->nodemask
, cpu_map
,
8881 d
->send_covered
, d
->tmpmask
);
8884 case SD_LV_ALLNODES
:
8885 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8886 d
->send_covered
, d
->tmpmask
);
8895 * Build sched domains for a given set of cpus and attach the sched domains
8896 * to the individual cpus
8898 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8899 struct sched_domain_attr
*attr
)
8901 enum s_alloc alloc_state
= sa_none
;
8903 struct sched_domain
*sd
;
8909 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8910 if (alloc_state
!= sa_rootdomain
)
8912 alloc_state
= sa_sched_groups
;
8915 * Set up domains for cpus specified by the cpu_map.
8917 for_each_cpu(i
, cpu_map
) {
8918 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8921 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8922 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8923 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8924 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8927 for_each_cpu(i
, cpu_map
) {
8928 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8929 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8932 /* Set up physical groups */
8933 for (i
= 0; i
< nr_node_ids
; i
++)
8934 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8937 /* Set up node groups */
8939 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8941 for (i
= 0; i
< nr_node_ids
; i
++)
8942 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8946 /* Calculate CPU power for physical packages and nodes */
8947 #ifdef CONFIG_SCHED_SMT
8948 for_each_cpu(i
, cpu_map
) {
8949 sd
= &per_cpu(cpu_domains
, i
).sd
;
8950 init_sched_groups_power(i
, sd
);
8953 #ifdef CONFIG_SCHED_MC
8954 for_each_cpu(i
, cpu_map
) {
8955 sd
= &per_cpu(core_domains
, i
).sd
;
8956 init_sched_groups_power(i
, sd
);
8960 for_each_cpu(i
, cpu_map
) {
8961 sd
= &per_cpu(phys_domains
, i
).sd
;
8962 init_sched_groups_power(i
, sd
);
8966 for (i
= 0; i
< nr_node_ids
; i
++)
8967 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8969 if (d
.sd_allnodes
) {
8970 struct sched_group
*sg
;
8972 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8974 init_numa_sched_groups_power(sg
);
8978 /* Attach the domains */
8979 for_each_cpu(i
, cpu_map
) {
8980 #ifdef CONFIG_SCHED_SMT
8981 sd
= &per_cpu(cpu_domains
, i
).sd
;
8982 #elif defined(CONFIG_SCHED_MC)
8983 sd
= &per_cpu(core_domains
, i
).sd
;
8985 sd
= &per_cpu(phys_domains
, i
).sd
;
8987 cpu_attach_domain(sd
, d
.rd
, i
);
8990 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8991 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8995 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8999 static int build_sched_domains(const struct cpumask
*cpu_map
)
9001 return __build_sched_domains(cpu_map
, NULL
);
9004 static struct cpumask
*doms_cur
; /* current sched domains */
9005 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
9006 static struct sched_domain_attr
*dattr_cur
;
9007 /* attribues of custom domains in 'doms_cur' */
9010 * Special case: If a kmalloc of a doms_cur partition (array of
9011 * cpumask) fails, then fallback to a single sched domain,
9012 * as determined by the single cpumask fallback_doms.
9014 static cpumask_var_t fallback_doms
;
9017 * arch_update_cpu_topology lets virtualized architectures update the
9018 * cpu core maps. It is supposed to return 1 if the topology changed
9019 * or 0 if it stayed the same.
9021 int __attribute__((weak
)) arch_update_cpu_topology(void)
9027 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9028 * For now this just excludes isolated cpus, but could be used to
9029 * exclude other special cases in the future.
9031 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9035 arch_update_cpu_topology();
9037 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
9039 doms_cur
= fallback_doms
;
9040 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
9042 err
= build_sched_domains(doms_cur
);
9043 register_sched_domain_sysctl();
9048 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9049 struct cpumask
*tmpmask
)
9051 free_sched_groups(cpu_map
, tmpmask
);
9055 * Detach sched domains from a group of cpus specified in cpu_map
9056 * These cpus will now be attached to the NULL domain
9058 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9060 /* Save because hotplug lock held. */
9061 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9064 for_each_cpu(i
, cpu_map
)
9065 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9066 synchronize_sched();
9067 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9070 /* handle null as "default" */
9071 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9072 struct sched_domain_attr
*new, int idx_new
)
9074 struct sched_domain_attr tmp
;
9081 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9082 new ? (new + idx_new
) : &tmp
,
9083 sizeof(struct sched_domain_attr
));
9087 * Partition sched domains as specified by the 'ndoms_new'
9088 * cpumasks in the array doms_new[] of cpumasks. This compares
9089 * doms_new[] to the current sched domain partitioning, doms_cur[].
9090 * It destroys each deleted domain and builds each new domain.
9092 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9093 * The masks don't intersect (don't overlap.) We should setup one
9094 * sched domain for each mask. CPUs not in any of the cpumasks will
9095 * not be load balanced. If the same cpumask appears both in the
9096 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9099 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9100 * ownership of it and will kfree it when done with it. If the caller
9101 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9102 * ndoms_new == 1, and partition_sched_domains() will fallback to
9103 * the single partition 'fallback_doms', it also forces the domains
9106 * If doms_new == NULL it will be replaced with cpu_online_mask.
9107 * ndoms_new == 0 is a special case for destroying existing domains,
9108 * and it will not create the default domain.
9110 * Call with hotplug lock held
9112 /* FIXME: Change to struct cpumask *doms_new[] */
9113 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9114 struct sched_domain_attr
*dattr_new
)
9119 mutex_lock(&sched_domains_mutex
);
9121 /* always unregister in case we don't destroy any domains */
9122 unregister_sched_domain_sysctl();
9124 /* Let architecture update cpu core mappings. */
9125 new_topology
= arch_update_cpu_topology();
9127 n
= doms_new
? ndoms_new
: 0;
9129 /* Destroy deleted domains */
9130 for (i
= 0; i
< ndoms_cur
; i
++) {
9131 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9132 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9133 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9136 /* no match - a current sched domain not in new doms_new[] */
9137 detach_destroy_domains(doms_cur
+ i
);
9142 if (doms_new
== NULL
) {
9144 doms_new
= fallback_doms
;
9145 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
9146 WARN_ON_ONCE(dattr_new
);
9149 /* Build new domains */
9150 for (i
= 0; i
< ndoms_new
; i
++) {
9151 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9152 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9153 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9156 /* no match - add a new doms_new */
9157 __build_sched_domains(doms_new
+ i
,
9158 dattr_new
? dattr_new
+ i
: NULL
);
9163 /* Remember the new sched domains */
9164 if (doms_cur
!= fallback_doms
)
9166 kfree(dattr_cur
); /* kfree(NULL) is safe */
9167 doms_cur
= doms_new
;
9168 dattr_cur
= dattr_new
;
9169 ndoms_cur
= ndoms_new
;
9171 register_sched_domain_sysctl();
9173 mutex_unlock(&sched_domains_mutex
);
9176 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9177 static void arch_reinit_sched_domains(void)
9181 /* Destroy domains first to force the rebuild */
9182 partition_sched_domains(0, NULL
, NULL
);
9184 rebuild_sched_domains();
9188 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9190 unsigned int level
= 0;
9192 if (sscanf(buf
, "%u", &level
) != 1)
9196 * level is always be positive so don't check for
9197 * level < POWERSAVINGS_BALANCE_NONE which is 0
9198 * What happens on 0 or 1 byte write,
9199 * need to check for count as well?
9202 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9206 sched_smt_power_savings
= level
;
9208 sched_mc_power_savings
= level
;
9210 arch_reinit_sched_domains();
9215 #ifdef CONFIG_SCHED_MC
9216 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9219 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9221 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9222 const char *buf
, size_t count
)
9224 return sched_power_savings_store(buf
, count
, 0);
9226 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9227 sched_mc_power_savings_show
,
9228 sched_mc_power_savings_store
);
9231 #ifdef CONFIG_SCHED_SMT
9232 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9235 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9237 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9238 const char *buf
, size_t count
)
9240 return sched_power_savings_store(buf
, count
, 1);
9242 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9243 sched_smt_power_savings_show
,
9244 sched_smt_power_savings_store
);
9247 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9251 #ifdef CONFIG_SCHED_SMT
9253 err
= sysfs_create_file(&cls
->kset
.kobj
,
9254 &attr_sched_smt_power_savings
.attr
);
9256 #ifdef CONFIG_SCHED_MC
9257 if (!err
&& mc_capable())
9258 err
= sysfs_create_file(&cls
->kset
.kobj
,
9259 &attr_sched_mc_power_savings
.attr
);
9263 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9265 #ifndef CONFIG_CPUSETS
9267 * Add online and remove offline CPUs from the scheduler domains.
9268 * When cpusets are enabled they take over this function.
9270 static int update_sched_domains(struct notifier_block
*nfb
,
9271 unsigned long action
, void *hcpu
)
9275 case CPU_ONLINE_FROZEN
:
9277 case CPU_DEAD_FROZEN
:
9278 partition_sched_domains(1, NULL
, NULL
);
9287 static int update_runtime(struct notifier_block
*nfb
,
9288 unsigned long action
, void *hcpu
)
9290 int cpu
= (int)(long)hcpu
;
9293 case CPU_DOWN_PREPARE
:
9294 case CPU_DOWN_PREPARE_FROZEN
:
9295 disable_runtime(cpu_rq(cpu
));
9298 case CPU_DOWN_FAILED
:
9299 case CPU_DOWN_FAILED_FROZEN
:
9301 case CPU_ONLINE_FROZEN
:
9302 enable_runtime(cpu_rq(cpu
));
9310 void __init
sched_init_smp(void)
9312 cpumask_var_t non_isolated_cpus
;
9314 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9316 #if defined(CONFIG_NUMA)
9317 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9319 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9322 mutex_lock(&sched_domains_mutex
);
9323 arch_init_sched_domains(cpu_online_mask
);
9324 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9325 if (cpumask_empty(non_isolated_cpus
))
9326 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9327 mutex_unlock(&sched_domains_mutex
);
9330 #ifndef CONFIG_CPUSETS
9331 /* XXX: Theoretical race here - CPU may be hotplugged now */
9332 hotcpu_notifier(update_sched_domains
, 0);
9335 /* RT runtime code needs to handle some hotplug events */
9336 hotcpu_notifier(update_runtime
, 0);
9340 /* Move init over to a non-isolated CPU */
9341 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9343 sched_init_granularity();
9344 free_cpumask_var(non_isolated_cpus
);
9346 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9347 init_sched_rt_class();
9350 void __init
sched_init_smp(void)
9352 sched_init_granularity();
9354 #endif /* CONFIG_SMP */
9356 const_debug
unsigned int sysctl_timer_migration
= 1;
9358 int in_sched_functions(unsigned long addr
)
9360 return in_lock_functions(addr
) ||
9361 (addr
>= (unsigned long)__sched_text_start
9362 && addr
< (unsigned long)__sched_text_end
);
9365 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9367 cfs_rq
->tasks_timeline
= RB_ROOT
;
9368 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9369 #ifdef CONFIG_FAIR_GROUP_SCHED
9372 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9375 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9377 struct rt_prio_array
*array
;
9380 array
= &rt_rq
->active
;
9381 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9382 INIT_LIST_HEAD(array
->queue
+ i
);
9383 __clear_bit(i
, array
->bitmap
);
9385 /* delimiter for bitsearch: */
9386 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9388 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9389 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9391 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9395 rt_rq
->rt_nr_migratory
= 0;
9396 rt_rq
->overloaded
= 0;
9397 plist_head_init(&rt_rq
->pushable_tasks
, &rq
->lock
);
9401 rt_rq
->rt_throttled
= 0;
9402 rt_rq
->rt_runtime
= 0;
9403 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9405 #ifdef CONFIG_RT_GROUP_SCHED
9406 rt_rq
->rt_nr_boosted
= 0;
9411 #ifdef CONFIG_FAIR_GROUP_SCHED
9412 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9413 struct sched_entity
*se
, int cpu
, int add
,
9414 struct sched_entity
*parent
)
9416 struct rq
*rq
= cpu_rq(cpu
);
9417 tg
->cfs_rq
[cpu
] = cfs_rq
;
9418 init_cfs_rq(cfs_rq
, rq
);
9421 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9424 /* se could be NULL for init_task_group */
9429 se
->cfs_rq
= &rq
->cfs
;
9431 se
->cfs_rq
= parent
->my_q
;
9434 se
->load
.weight
= tg
->shares
;
9435 se
->load
.inv_weight
= 0;
9436 se
->parent
= parent
;
9440 #ifdef CONFIG_RT_GROUP_SCHED
9441 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9442 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9443 struct sched_rt_entity
*parent
)
9445 struct rq
*rq
= cpu_rq(cpu
);
9447 tg
->rt_rq
[cpu
] = rt_rq
;
9448 init_rt_rq(rt_rq
, rq
);
9450 rt_rq
->rt_se
= rt_se
;
9451 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9453 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9455 tg
->rt_se
[cpu
] = rt_se
;
9460 rt_se
->rt_rq
= &rq
->rt
;
9462 rt_se
->rt_rq
= parent
->my_q
;
9464 rt_se
->my_q
= rt_rq
;
9465 rt_se
->parent
= parent
;
9466 INIT_LIST_HEAD(&rt_se
->run_list
);
9470 void __init
sched_init(void)
9473 unsigned long alloc_size
= 0, ptr
;
9475 #ifdef CONFIG_FAIR_GROUP_SCHED
9476 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9478 #ifdef CONFIG_RT_GROUP_SCHED
9479 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9481 #ifdef CONFIG_USER_SCHED
9484 #ifdef CONFIG_CPUMASK_OFFSTACK
9485 alloc_size
+= num_possible_cpus() * cpumask_size();
9488 * As sched_init() is called before page_alloc is setup,
9489 * we use alloc_bootmem().
9492 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9494 #ifdef CONFIG_FAIR_GROUP_SCHED
9495 init_task_group
.se
= (struct sched_entity
**)ptr
;
9496 ptr
+= nr_cpu_ids
* sizeof(void **);
9498 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9499 ptr
+= nr_cpu_ids
* sizeof(void **);
9501 #ifdef CONFIG_USER_SCHED
9502 root_task_group
.se
= (struct sched_entity
**)ptr
;
9503 ptr
+= nr_cpu_ids
* sizeof(void **);
9505 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9506 ptr
+= nr_cpu_ids
* sizeof(void **);
9507 #endif /* CONFIG_USER_SCHED */
9508 #endif /* CONFIG_FAIR_GROUP_SCHED */
9509 #ifdef CONFIG_RT_GROUP_SCHED
9510 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9511 ptr
+= nr_cpu_ids
* sizeof(void **);
9513 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9514 ptr
+= nr_cpu_ids
* sizeof(void **);
9516 #ifdef CONFIG_USER_SCHED
9517 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9518 ptr
+= nr_cpu_ids
* sizeof(void **);
9520 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9521 ptr
+= nr_cpu_ids
* sizeof(void **);
9522 #endif /* CONFIG_USER_SCHED */
9523 #endif /* CONFIG_RT_GROUP_SCHED */
9524 #ifdef CONFIG_CPUMASK_OFFSTACK
9525 for_each_possible_cpu(i
) {
9526 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9527 ptr
+= cpumask_size();
9529 #endif /* CONFIG_CPUMASK_OFFSTACK */
9533 init_defrootdomain();
9536 init_rt_bandwidth(&def_rt_bandwidth
,
9537 global_rt_period(), global_rt_runtime());
9539 #ifdef CONFIG_RT_GROUP_SCHED
9540 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9541 global_rt_period(), global_rt_runtime());
9542 #ifdef CONFIG_USER_SCHED
9543 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9544 global_rt_period(), RUNTIME_INF
);
9545 #endif /* CONFIG_USER_SCHED */
9546 #endif /* CONFIG_RT_GROUP_SCHED */
9548 #ifdef CONFIG_GROUP_SCHED
9549 list_add(&init_task_group
.list
, &task_groups
);
9550 INIT_LIST_HEAD(&init_task_group
.children
);
9552 #ifdef CONFIG_USER_SCHED
9553 INIT_LIST_HEAD(&root_task_group
.children
);
9554 init_task_group
.parent
= &root_task_group
;
9555 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9556 #endif /* CONFIG_USER_SCHED */
9557 #endif /* CONFIG_GROUP_SCHED */
9559 for_each_possible_cpu(i
) {
9563 spin_lock_init(&rq
->lock
);
9565 rq
->calc_load_active
= 0;
9566 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9567 init_cfs_rq(&rq
->cfs
, rq
);
9568 init_rt_rq(&rq
->rt
, rq
);
9569 #ifdef CONFIG_FAIR_GROUP_SCHED
9570 init_task_group
.shares
= init_task_group_load
;
9571 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9572 #ifdef CONFIG_CGROUP_SCHED
9574 * How much cpu bandwidth does init_task_group get?
9576 * In case of task-groups formed thr' the cgroup filesystem, it
9577 * gets 100% of the cpu resources in the system. This overall
9578 * system cpu resource is divided among the tasks of
9579 * init_task_group and its child task-groups in a fair manner,
9580 * based on each entity's (task or task-group's) weight
9581 * (se->load.weight).
9583 * In other words, if init_task_group has 10 tasks of weight
9584 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9585 * then A0's share of the cpu resource is:
9587 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9589 * We achieve this by letting init_task_group's tasks sit
9590 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9592 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9593 #elif defined CONFIG_USER_SCHED
9594 root_task_group
.shares
= NICE_0_LOAD
;
9595 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9597 * In case of task-groups formed thr' the user id of tasks,
9598 * init_task_group represents tasks belonging to root user.
9599 * Hence it forms a sibling of all subsequent groups formed.
9600 * In this case, init_task_group gets only a fraction of overall
9601 * system cpu resource, based on the weight assigned to root
9602 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9603 * by letting tasks of init_task_group sit in a separate cfs_rq
9604 * (init_tg_cfs_rq) and having one entity represent this group of
9605 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9607 init_tg_cfs_entry(&init_task_group
,
9608 &per_cpu(init_tg_cfs_rq
, i
),
9609 &per_cpu(init_sched_entity
, i
), i
, 1,
9610 root_task_group
.se
[i
]);
9613 #endif /* CONFIG_FAIR_GROUP_SCHED */
9615 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9616 #ifdef CONFIG_RT_GROUP_SCHED
9617 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9618 #ifdef CONFIG_CGROUP_SCHED
9619 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9620 #elif defined CONFIG_USER_SCHED
9621 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9622 init_tg_rt_entry(&init_task_group
,
9623 &per_cpu(init_rt_rq
, i
),
9624 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9625 root_task_group
.rt_se
[i
]);
9629 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9630 rq
->cpu_load
[j
] = 0;
9634 rq
->post_schedule
= 0;
9635 rq
->active_balance
= 0;
9636 rq
->next_balance
= jiffies
;
9640 rq
->migration_thread
= NULL
;
9641 INIT_LIST_HEAD(&rq
->migration_queue
);
9642 rq_attach_root(rq
, &def_root_domain
);
9645 atomic_set(&rq
->nr_iowait
, 0);
9648 set_load_weight(&init_task
);
9650 #ifdef CONFIG_PREEMPT_NOTIFIERS
9651 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9655 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9658 #ifdef CONFIG_RT_MUTEXES
9659 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9663 * The boot idle thread does lazy MMU switching as well:
9665 atomic_inc(&init_mm
.mm_count
);
9666 enter_lazy_tlb(&init_mm
, current
);
9669 * Make us the idle thread. Technically, schedule() should not be
9670 * called from this thread, however somewhere below it might be,
9671 * but because we are the idle thread, we just pick up running again
9672 * when this runqueue becomes "idle".
9674 init_idle(current
, smp_processor_id());
9676 calc_load_update
= jiffies
+ LOAD_FREQ
;
9679 * During early bootup we pretend to be a normal task:
9681 current
->sched_class
= &fair_sched_class
;
9683 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9684 alloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9687 alloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9688 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9690 alloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9693 perf_counter_init();
9695 scheduler_running
= 1;
9698 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9699 static inline int preempt_count_equals(int preempt_offset
)
9701 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9703 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9706 void __might_sleep(char *file
, int line
, int preempt_offset
)
9709 static unsigned long prev_jiffy
; /* ratelimiting */
9711 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9712 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9714 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9716 prev_jiffy
= jiffies
;
9719 "BUG: sleeping function called from invalid context at %s:%d\n",
9722 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9723 in_atomic(), irqs_disabled(),
9724 current
->pid
, current
->comm
);
9726 debug_show_held_locks(current
);
9727 if (irqs_disabled())
9728 print_irqtrace_events(current
);
9732 EXPORT_SYMBOL(__might_sleep
);
9735 #ifdef CONFIG_MAGIC_SYSRQ
9736 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9740 update_rq_clock(rq
);
9741 on_rq
= p
->se
.on_rq
;
9743 deactivate_task(rq
, p
, 0);
9744 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9746 activate_task(rq
, p
, 0);
9747 resched_task(rq
->curr
);
9751 void normalize_rt_tasks(void)
9753 struct task_struct
*g
, *p
;
9754 unsigned long flags
;
9757 read_lock_irqsave(&tasklist_lock
, flags
);
9758 do_each_thread(g
, p
) {
9760 * Only normalize user tasks:
9765 p
->se
.exec_start
= 0;
9766 #ifdef CONFIG_SCHEDSTATS
9767 p
->se
.wait_start
= 0;
9768 p
->se
.sleep_start
= 0;
9769 p
->se
.block_start
= 0;
9774 * Renice negative nice level userspace
9777 if (TASK_NICE(p
) < 0 && p
->mm
)
9778 set_user_nice(p
, 0);
9782 spin_lock(&p
->pi_lock
);
9783 rq
= __task_rq_lock(p
);
9785 normalize_task(rq
, p
);
9787 __task_rq_unlock(rq
);
9788 spin_unlock(&p
->pi_lock
);
9789 } while_each_thread(g
, p
);
9791 read_unlock_irqrestore(&tasklist_lock
, flags
);
9794 #endif /* CONFIG_MAGIC_SYSRQ */
9798 * These functions are only useful for the IA64 MCA handling.
9800 * They can only be called when the whole system has been
9801 * stopped - every CPU needs to be quiescent, and no scheduling
9802 * activity can take place. Using them for anything else would
9803 * be a serious bug, and as a result, they aren't even visible
9804 * under any other configuration.
9808 * curr_task - return the current task for a given cpu.
9809 * @cpu: the processor in question.
9811 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9813 struct task_struct
*curr_task(int cpu
)
9815 return cpu_curr(cpu
);
9819 * set_curr_task - set the current task for a given cpu.
9820 * @cpu: the processor in question.
9821 * @p: the task pointer to set.
9823 * Description: This function must only be used when non-maskable interrupts
9824 * are serviced on a separate stack. It allows the architecture to switch the
9825 * notion of the current task on a cpu in a non-blocking manner. This function
9826 * must be called with all CPU's synchronized, and interrupts disabled, the
9827 * and caller must save the original value of the current task (see
9828 * curr_task() above) and restore that value before reenabling interrupts and
9829 * re-starting the system.
9831 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9833 void set_curr_task(int cpu
, struct task_struct
*p
)
9840 #ifdef CONFIG_FAIR_GROUP_SCHED
9841 static void free_fair_sched_group(struct task_group
*tg
)
9845 for_each_possible_cpu(i
) {
9847 kfree(tg
->cfs_rq
[i
]);
9857 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9859 struct cfs_rq
*cfs_rq
;
9860 struct sched_entity
*se
;
9864 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9867 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9871 tg
->shares
= NICE_0_LOAD
;
9873 for_each_possible_cpu(i
) {
9876 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9877 GFP_KERNEL
, cpu_to_node(i
));
9881 se
= kzalloc_node(sizeof(struct sched_entity
),
9882 GFP_KERNEL
, cpu_to_node(i
));
9886 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9895 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9897 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9898 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9901 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9903 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9905 #else /* !CONFG_FAIR_GROUP_SCHED */
9906 static inline void free_fair_sched_group(struct task_group
*tg
)
9911 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9916 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9920 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9923 #endif /* CONFIG_FAIR_GROUP_SCHED */
9925 #ifdef CONFIG_RT_GROUP_SCHED
9926 static void free_rt_sched_group(struct task_group
*tg
)
9930 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9932 for_each_possible_cpu(i
) {
9934 kfree(tg
->rt_rq
[i
]);
9936 kfree(tg
->rt_se
[i
]);
9944 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9946 struct rt_rq
*rt_rq
;
9947 struct sched_rt_entity
*rt_se
;
9951 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9954 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9958 init_rt_bandwidth(&tg
->rt_bandwidth
,
9959 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9961 for_each_possible_cpu(i
) {
9964 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9965 GFP_KERNEL
, cpu_to_node(i
));
9969 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9970 GFP_KERNEL
, cpu_to_node(i
));
9974 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9983 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9985 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9986 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9989 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9991 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9993 #else /* !CONFIG_RT_GROUP_SCHED */
9994 static inline void free_rt_sched_group(struct task_group
*tg
)
9999 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10004 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10008 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10011 #endif /* CONFIG_RT_GROUP_SCHED */
10013 #ifdef CONFIG_GROUP_SCHED
10014 static void free_sched_group(struct task_group
*tg
)
10016 free_fair_sched_group(tg
);
10017 free_rt_sched_group(tg
);
10021 /* allocate runqueue etc for a new task group */
10022 struct task_group
*sched_create_group(struct task_group
*parent
)
10024 struct task_group
*tg
;
10025 unsigned long flags
;
10028 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10030 return ERR_PTR(-ENOMEM
);
10032 if (!alloc_fair_sched_group(tg
, parent
))
10035 if (!alloc_rt_sched_group(tg
, parent
))
10038 spin_lock_irqsave(&task_group_lock
, flags
);
10039 for_each_possible_cpu(i
) {
10040 register_fair_sched_group(tg
, i
);
10041 register_rt_sched_group(tg
, i
);
10043 list_add_rcu(&tg
->list
, &task_groups
);
10045 WARN_ON(!parent
); /* root should already exist */
10047 tg
->parent
= parent
;
10048 INIT_LIST_HEAD(&tg
->children
);
10049 list_add_rcu(&tg
->siblings
, &parent
->children
);
10050 spin_unlock_irqrestore(&task_group_lock
, flags
);
10055 free_sched_group(tg
);
10056 return ERR_PTR(-ENOMEM
);
10059 /* rcu callback to free various structures associated with a task group */
10060 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10062 /* now it should be safe to free those cfs_rqs */
10063 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10066 /* Destroy runqueue etc associated with a task group */
10067 void sched_destroy_group(struct task_group
*tg
)
10069 unsigned long flags
;
10072 spin_lock_irqsave(&task_group_lock
, flags
);
10073 for_each_possible_cpu(i
) {
10074 unregister_fair_sched_group(tg
, i
);
10075 unregister_rt_sched_group(tg
, i
);
10077 list_del_rcu(&tg
->list
);
10078 list_del_rcu(&tg
->siblings
);
10079 spin_unlock_irqrestore(&task_group_lock
, flags
);
10081 /* wait for possible concurrent references to cfs_rqs complete */
10082 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10085 /* change task's runqueue when it moves between groups.
10086 * The caller of this function should have put the task in its new group
10087 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10088 * reflect its new group.
10090 void sched_move_task(struct task_struct
*tsk
)
10092 int on_rq
, running
;
10093 unsigned long flags
;
10096 rq
= task_rq_lock(tsk
, &flags
);
10098 update_rq_clock(rq
);
10100 running
= task_current(rq
, tsk
);
10101 on_rq
= tsk
->se
.on_rq
;
10104 dequeue_task(rq
, tsk
, 0);
10105 if (unlikely(running
))
10106 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10108 set_task_rq(tsk
, task_cpu(tsk
));
10110 #ifdef CONFIG_FAIR_GROUP_SCHED
10111 if (tsk
->sched_class
->moved_group
)
10112 tsk
->sched_class
->moved_group(tsk
);
10115 if (unlikely(running
))
10116 tsk
->sched_class
->set_curr_task(rq
);
10118 enqueue_task(rq
, tsk
, 0);
10120 task_rq_unlock(rq
, &flags
);
10122 #endif /* CONFIG_GROUP_SCHED */
10124 #ifdef CONFIG_FAIR_GROUP_SCHED
10125 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10127 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10132 dequeue_entity(cfs_rq
, se
, 0);
10134 se
->load
.weight
= shares
;
10135 se
->load
.inv_weight
= 0;
10138 enqueue_entity(cfs_rq
, se
, 0);
10141 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10143 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10144 struct rq
*rq
= cfs_rq
->rq
;
10145 unsigned long flags
;
10147 spin_lock_irqsave(&rq
->lock
, flags
);
10148 __set_se_shares(se
, shares
);
10149 spin_unlock_irqrestore(&rq
->lock
, flags
);
10152 static DEFINE_MUTEX(shares_mutex
);
10154 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10157 unsigned long flags
;
10160 * We can't change the weight of the root cgroup.
10165 if (shares
< MIN_SHARES
)
10166 shares
= MIN_SHARES
;
10167 else if (shares
> MAX_SHARES
)
10168 shares
= MAX_SHARES
;
10170 mutex_lock(&shares_mutex
);
10171 if (tg
->shares
== shares
)
10174 spin_lock_irqsave(&task_group_lock
, flags
);
10175 for_each_possible_cpu(i
)
10176 unregister_fair_sched_group(tg
, i
);
10177 list_del_rcu(&tg
->siblings
);
10178 spin_unlock_irqrestore(&task_group_lock
, flags
);
10180 /* wait for any ongoing reference to this group to finish */
10181 synchronize_sched();
10184 * Now we are free to modify the group's share on each cpu
10185 * w/o tripping rebalance_share or load_balance_fair.
10187 tg
->shares
= shares
;
10188 for_each_possible_cpu(i
) {
10190 * force a rebalance
10192 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10193 set_se_shares(tg
->se
[i
], shares
);
10197 * Enable load balance activity on this group, by inserting it back on
10198 * each cpu's rq->leaf_cfs_rq_list.
10200 spin_lock_irqsave(&task_group_lock
, flags
);
10201 for_each_possible_cpu(i
)
10202 register_fair_sched_group(tg
, i
);
10203 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10204 spin_unlock_irqrestore(&task_group_lock
, flags
);
10206 mutex_unlock(&shares_mutex
);
10210 unsigned long sched_group_shares(struct task_group
*tg
)
10216 #ifdef CONFIG_RT_GROUP_SCHED
10218 * Ensure that the real time constraints are schedulable.
10220 static DEFINE_MUTEX(rt_constraints_mutex
);
10222 static unsigned long to_ratio(u64 period
, u64 runtime
)
10224 if (runtime
== RUNTIME_INF
)
10227 return div64_u64(runtime
<< 20, period
);
10230 /* Must be called with tasklist_lock held */
10231 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10233 struct task_struct
*g
, *p
;
10235 do_each_thread(g
, p
) {
10236 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10238 } while_each_thread(g
, p
);
10243 struct rt_schedulable_data
{
10244 struct task_group
*tg
;
10249 static int tg_schedulable(struct task_group
*tg
, void *data
)
10251 struct rt_schedulable_data
*d
= data
;
10252 struct task_group
*child
;
10253 unsigned long total
, sum
= 0;
10254 u64 period
, runtime
;
10256 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10257 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10260 period
= d
->rt_period
;
10261 runtime
= d
->rt_runtime
;
10264 #ifdef CONFIG_USER_SCHED
10265 if (tg
== &root_task_group
) {
10266 period
= global_rt_period();
10267 runtime
= global_rt_runtime();
10272 * Cannot have more runtime than the period.
10274 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10278 * Ensure we don't starve existing RT tasks.
10280 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10283 total
= to_ratio(period
, runtime
);
10286 * Nobody can have more than the global setting allows.
10288 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10292 * The sum of our children's runtime should not exceed our own.
10294 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10295 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10296 runtime
= child
->rt_bandwidth
.rt_runtime
;
10298 if (child
== d
->tg
) {
10299 period
= d
->rt_period
;
10300 runtime
= d
->rt_runtime
;
10303 sum
+= to_ratio(period
, runtime
);
10312 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10314 struct rt_schedulable_data data
= {
10316 .rt_period
= period
,
10317 .rt_runtime
= runtime
,
10320 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10323 static int tg_set_bandwidth(struct task_group
*tg
,
10324 u64 rt_period
, u64 rt_runtime
)
10328 mutex_lock(&rt_constraints_mutex
);
10329 read_lock(&tasklist_lock
);
10330 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10334 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10335 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10336 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10338 for_each_possible_cpu(i
) {
10339 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10341 spin_lock(&rt_rq
->rt_runtime_lock
);
10342 rt_rq
->rt_runtime
= rt_runtime
;
10343 spin_unlock(&rt_rq
->rt_runtime_lock
);
10345 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10347 read_unlock(&tasklist_lock
);
10348 mutex_unlock(&rt_constraints_mutex
);
10353 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10355 u64 rt_runtime
, rt_period
;
10357 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10358 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10359 if (rt_runtime_us
< 0)
10360 rt_runtime
= RUNTIME_INF
;
10362 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10365 long sched_group_rt_runtime(struct task_group
*tg
)
10369 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10372 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10373 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10374 return rt_runtime_us
;
10377 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10379 u64 rt_runtime
, rt_period
;
10381 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10382 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10384 if (rt_period
== 0)
10387 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10390 long sched_group_rt_period(struct task_group
*tg
)
10394 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10395 do_div(rt_period_us
, NSEC_PER_USEC
);
10396 return rt_period_us
;
10399 static int sched_rt_global_constraints(void)
10401 u64 runtime
, period
;
10404 if (sysctl_sched_rt_period
<= 0)
10407 runtime
= global_rt_runtime();
10408 period
= global_rt_period();
10411 * Sanity check on the sysctl variables.
10413 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10416 mutex_lock(&rt_constraints_mutex
);
10417 read_lock(&tasklist_lock
);
10418 ret
= __rt_schedulable(NULL
, 0, 0);
10419 read_unlock(&tasklist_lock
);
10420 mutex_unlock(&rt_constraints_mutex
);
10425 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10427 /* Don't accept realtime tasks when there is no way for them to run */
10428 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10434 #else /* !CONFIG_RT_GROUP_SCHED */
10435 static int sched_rt_global_constraints(void)
10437 unsigned long flags
;
10440 if (sysctl_sched_rt_period
<= 0)
10444 * There's always some RT tasks in the root group
10445 * -- migration, kstopmachine etc..
10447 if (sysctl_sched_rt_runtime
== 0)
10450 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10451 for_each_possible_cpu(i
) {
10452 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10454 spin_lock(&rt_rq
->rt_runtime_lock
);
10455 rt_rq
->rt_runtime
= global_rt_runtime();
10456 spin_unlock(&rt_rq
->rt_runtime_lock
);
10458 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10462 #endif /* CONFIG_RT_GROUP_SCHED */
10464 int sched_rt_handler(struct ctl_table
*table
, int write
,
10465 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10469 int old_period
, old_runtime
;
10470 static DEFINE_MUTEX(mutex
);
10472 mutex_lock(&mutex
);
10473 old_period
= sysctl_sched_rt_period
;
10474 old_runtime
= sysctl_sched_rt_runtime
;
10476 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10478 if (!ret
&& write
) {
10479 ret
= sched_rt_global_constraints();
10481 sysctl_sched_rt_period
= old_period
;
10482 sysctl_sched_rt_runtime
= old_runtime
;
10484 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10485 def_rt_bandwidth
.rt_period
=
10486 ns_to_ktime(global_rt_period());
10489 mutex_unlock(&mutex
);
10494 #ifdef CONFIG_CGROUP_SCHED
10496 /* return corresponding task_group object of a cgroup */
10497 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10499 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10500 struct task_group
, css
);
10503 static struct cgroup_subsys_state
*
10504 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10506 struct task_group
*tg
, *parent
;
10508 if (!cgrp
->parent
) {
10509 /* This is early initialization for the top cgroup */
10510 return &init_task_group
.css
;
10513 parent
= cgroup_tg(cgrp
->parent
);
10514 tg
= sched_create_group(parent
);
10516 return ERR_PTR(-ENOMEM
);
10522 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10524 struct task_group
*tg
= cgroup_tg(cgrp
);
10526 sched_destroy_group(tg
);
10530 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10531 struct task_struct
*tsk
)
10533 #ifdef CONFIG_RT_GROUP_SCHED
10534 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10537 /* We don't support RT-tasks being in separate groups */
10538 if (tsk
->sched_class
!= &fair_sched_class
)
10546 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10547 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10549 sched_move_task(tsk
);
10552 #ifdef CONFIG_FAIR_GROUP_SCHED
10553 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10556 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10559 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10561 struct task_group
*tg
= cgroup_tg(cgrp
);
10563 return (u64
) tg
->shares
;
10565 #endif /* CONFIG_FAIR_GROUP_SCHED */
10567 #ifdef CONFIG_RT_GROUP_SCHED
10568 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10571 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10574 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10576 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10579 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10582 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10585 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10587 return sched_group_rt_period(cgroup_tg(cgrp
));
10589 #endif /* CONFIG_RT_GROUP_SCHED */
10591 static struct cftype cpu_files
[] = {
10592 #ifdef CONFIG_FAIR_GROUP_SCHED
10595 .read_u64
= cpu_shares_read_u64
,
10596 .write_u64
= cpu_shares_write_u64
,
10599 #ifdef CONFIG_RT_GROUP_SCHED
10601 .name
= "rt_runtime_us",
10602 .read_s64
= cpu_rt_runtime_read
,
10603 .write_s64
= cpu_rt_runtime_write
,
10606 .name
= "rt_period_us",
10607 .read_u64
= cpu_rt_period_read_uint
,
10608 .write_u64
= cpu_rt_period_write_uint
,
10613 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10615 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10618 struct cgroup_subsys cpu_cgroup_subsys
= {
10620 .create
= cpu_cgroup_create
,
10621 .destroy
= cpu_cgroup_destroy
,
10622 .can_attach
= cpu_cgroup_can_attach
,
10623 .attach
= cpu_cgroup_attach
,
10624 .populate
= cpu_cgroup_populate
,
10625 .subsys_id
= cpu_cgroup_subsys_id
,
10629 #endif /* CONFIG_CGROUP_SCHED */
10631 #ifdef CONFIG_CGROUP_CPUACCT
10634 * CPU accounting code for task groups.
10636 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10637 * (balbir@in.ibm.com).
10640 /* track cpu usage of a group of tasks and its child groups */
10642 struct cgroup_subsys_state css
;
10643 /* cpuusage holds pointer to a u64-type object on every cpu */
10645 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10646 struct cpuacct
*parent
;
10649 struct cgroup_subsys cpuacct_subsys
;
10651 /* return cpu accounting group corresponding to this container */
10652 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10654 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10655 struct cpuacct
, css
);
10658 /* return cpu accounting group to which this task belongs */
10659 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10661 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10662 struct cpuacct
, css
);
10665 /* create a new cpu accounting group */
10666 static struct cgroup_subsys_state
*cpuacct_create(
10667 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10669 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10675 ca
->cpuusage
= alloc_percpu(u64
);
10679 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10680 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10681 goto out_free_counters
;
10684 ca
->parent
= cgroup_ca(cgrp
->parent
);
10690 percpu_counter_destroy(&ca
->cpustat
[i
]);
10691 free_percpu(ca
->cpuusage
);
10695 return ERR_PTR(-ENOMEM
);
10698 /* destroy an existing cpu accounting group */
10700 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10702 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10705 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10706 percpu_counter_destroy(&ca
->cpustat
[i
]);
10707 free_percpu(ca
->cpuusage
);
10711 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10713 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10716 #ifndef CONFIG_64BIT
10718 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10720 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10722 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10730 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10732 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10734 #ifndef CONFIG_64BIT
10736 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10738 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10740 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10746 /* return total cpu usage (in nanoseconds) of a group */
10747 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10749 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10750 u64 totalcpuusage
= 0;
10753 for_each_present_cpu(i
)
10754 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10756 return totalcpuusage
;
10759 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10762 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10771 for_each_present_cpu(i
)
10772 cpuacct_cpuusage_write(ca
, i
, 0);
10778 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10779 struct seq_file
*m
)
10781 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10785 for_each_present_cpu(i
) {
10786 percpu
= cpuacct_cpuusage_read(ca
, i
);
10787 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10789 seq_printf(m
, "\n");
10793 static const char *cpuacct_stat_desc
[] = {
10794 [CPUACCT_STAT_USER
] = "user",
10795 [CPUACCT_STAT_SYSTEM
] = "system",
10798 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10799 struct cgroup_map_cb
*cb
)
10801 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10804 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10805 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10806 val
= cputime64_to_clock_t(val
);
10807 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10812 static struct cftype files
[] = {
10815 .read_u64
= cpuusage_read
,
10816 .write_u64
= cpuusage_write
,
10819 .name
= "usage_percpu",
10820 .read_seq_string
= cpuacct_percpu_seq_read
,
10824 .read_map
= cpuacct_stats_show
,
10828 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10830 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10834 * charge this task's execution time to its accounting group.
10836 * called with rq->lock held.
10838 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10840 struct cpuacct
*ca
;
10843 if (unlikely(!cpuacct_subsys
.active
))
10846 cpu
= task_cpu(tsk
);
10852 for (; ca
; ca
= ca
->parent
) {
10853 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10854 *cpuusage
+= cputime
;
10861 * Charge the system/user time to the task's accounting group.
10863 static void cpuacct_update_stats(struct task_struct
*tsk
,
10864 enum cpuacct_stat_index idx
, cputime_t val
)
10866 struct cpuacct
*ca
;
10868 if (unlikely(!cpuacct_subsys
.active
))
10875 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10881 struct cgroup_subsys cpuacct_subsys
= {
10883 .create
= cpuacct_create
,
10884 .destroy
= cpuacct_destroy
,
10885 .populate
= cpuacct_populate
,
10886 .subsys_id
= cpuacct_subsys_id
,
10888 #endif /* CONFIG_CGROUP_CPUACCT */